fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include <linux/percpu_counter.h>
  28 #include "hash.h"
  29 #include "tree-log.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "volumes.h"
  33 #include "raid56.h"
  34 #include "locking.h"
  35 #include "free-space-cache.h"
  36 #include "math.h"
  37 #include "sysfs.h"
  38 #include "qgroup.h"
  39
  40 #undef SCRAMBLE_DELAYED_REFS
  41
  42 /*
  43  * control flags for do_chunk_alloc's force field
  44  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  45  * if we really need one.
  46  *
  47  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  48  * if we have very few chunks already allocated.  This is
  49  * used as part of the clustering code to help make sure
  50  * we have a good pool of storage to cluster in, without
  51  * filling the FS with empty chunks
  52  *
  53  * CHUNK_ALLOC_FORCE means it must try to allocate one
  54  *
  55  */
  56 enum {
  57         CHUNK_ALLOC_NO_FORCE = 0,
  58         CHUNK_ALLOC_LIMITED = 1,
  59         CHUNK_ALLOC_FORCE = 2,
  60 };
  61
  62 /*
  63  * Control how reservations are dealt with.
  64  *
  65  * RESERVE_FREE - freeing a reservation.
  66  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  67  *   ENOSPC accounting
  68  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  69  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  70  */
  71 enum {
  72         RESERVE_FREE = 0,
  73         RESERVE_ALLOC = 1,
  74         RESERVE_ALLOC_NO_ACCOUNT = 2,
  75 };
  76
  77 static int update_block_group(struct btrfs_root *root,
  78                               u64 bytenr, u64 num_bytes, int alloc);
  79 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  80                                 struct btrfs_root *root,
  81                                 u64 bytenr, u64 num_bytes, u64 parent,
  82                                 u64 root_objectid, u64 owner_objectid,
  83                                 u64 owner_offset, int refs_to_drop,
  84                                 struct btrfs_delayed_extent_op *extra_op,
  85                                 int no_quota);
  86 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  87                                     struct extent_buffer *leaf,
  88                                     struct btrfs_extent_item *ei);
  89 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  90                                       struct btrfs_root *root,
  91                                       u64 parent, u64 root_objectid,
  92                                       u64 flags, u64 owner, u64 offset,
  93                                       struct btrfs_key *ins, int ref_mod);
  94 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  95                                      struct btrfs_root *root,
  96                                      u64 parent, u64 root_objectid,
  97                                      u64 flags, struct btrfs_disk_key *key,
  98                                      int level, struct btrfs_key *ins,
  99                                      int no_quota);
 100 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
 101                           struct btrfs_root *extent_root, u64 flags,
 102                           int force);
 103 static int find_next_key(struct btrfs_path *path, int level,
 104                          struct btrfs_key *key);
 105 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 106                             int dump_block_groups);
 107 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 108                                        u64 num_bytes, int reserve,
 109                                        int delalloc);
 110 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 111                                u64 num_bytes);
 112 int btrfs_pin_extent(struct btrfs_root *root,
 113                      u64 bytenr, u64 num_bytes, int reserved);
 114
 115 static noinline int
 116 block_group_cache_done(struct btrfs_block_group_cache *cache)
 117 {
 118         smp_mb();
 119         return cache->cached == BTRFS_CACHE_FINISHED ||
 120                 cache->cached == BTRFS_CACHE_ERROR;
 121 }
 122
 123 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 124 {
 125         return (cache->flags & bits) == bits;
 126 }
 127
 128 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 129 {
 130         atomic_inc(&cache->count);
 131 }
 132
 133 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 134 {
 135         if (atomic_dec_and_test(&cache->count)) {
 136                 WARN_ON(cache->pinned > 0);
 137                 WARN_ON(cache->reserved > 0);
 138                 kfree(cache->free_space_ctl);
 139                 kfree(cache);
 140         }
 141 }
 142
 143 /*
 144  * this adds the block group to the fs_info rb tree for the block group
 145  * cache
 146  */
 147 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 148                                 struct btrfs_block_group_cache *block_group)
 149 {
 150         struct rb_node **p;
 151         struct rb_node *parent = NULL;
 152         struct btrfs_block_group_cache *cache;
 153
 154         spin_lock(&info->block_group_cache_lock);
 155         p = &info->block_group_cache_tree.rb_node;
 156
 157         while (*p) {
 158                 parent = *p;
 159                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 160                                  cache_node);
 161                 if (block_group->key.objectid < cache->key.objectid) {
 162                         p = &(*p)->rb_left;
 163                 } else if (block_group->key.objectid > cache->key.objectid) {
 164                         p = &(*p)->rb_right;
 165                 } else {
 166                         spin_unlock(&info->block_group_cache_lock);
 167                         return -EEXIST;
 168                 }
 169         }
 170
 171         rb_link_node(&block_group->cache_node, parent, p);
 172         rb_insert_color(&block_group->cache_node,
 173                         &info->block_group_cache_tree);
 174
 175         if (info->first_logical_byte > block_group->key.objectid)
 176                 info->first_logical_byte = block_group->key.objectid;
 177
 178         spin_unlock(&info->block_group_cache_lock);
 179
 180         return 0;
 181 }
 182
 183 /*
 184  * This will return the block group at or after bytenr if contains is 0, else
 185  * it will return the block group that contains the bytenr
 186  */
 187 static struct btrfs_block_group_cache *
 188 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 189                               int contains)
 190 {
 191         struct btrfs_block_group_cache *cache, *ret = NULL;
 192         struct rb_node *n;
 193         u64 end, start;
 194
 195         spin_lock(&info->block_group_cache_lock);
 196         n = info->block_group_cache_tree.rb_node;
 197
 198         while (n) {
 199                 cache = rb_entry(n, struct btrfs_block_group_cache,
 200                                  cache_node);
 201                 end = cache->key.objectid + cache->key.offset - 1;
 202                 start = cache->key.objectid;
 203
 204                 if (bytenr < start) {
 205                         if (!contains && (!ret || start < ret->key.objectid))
 206                                 ret = cache;
 207                         n = n->rb_left;
 208                 } else if (bytenr > start) {
 209                         if (contains && bytenr <= end) {
 210                                 ret = cache;
 211                                 break;
 212                         }
 213                         n = n->rb_right;
 214                 } else {
 215                         ret = cache;
 216                         break;
 217                 }
 218         }
 219         if (ret) {
 220                 btrfs_get_block_group(ret);
 221                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 222                         info->first_logical_byte = ret->key.objectid;
 223         }
 224         spin_unlock(&info->block_group_cache_lock);
 225
 226         return ret;
 227 }
 228
 229 static int add_excluded_extent(struct btrfs_root *root,
 230                                u64 start, u64 num_bytes)
 231 {
 232         u64 end = start + num_bytes - 1;
 233         set_extent_bits(&root->fs_info->freed_extents[0],
 234                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 235         set_extent_bits(&root->fs_info->freed_extents[1],
 236                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 237         return 0;
 238 }
 239
 240 static void free_excluded_extents(struct btrfs_root *root,
 241                                   struct btrfs_block_group_cache *cache)
 242 {
 243         u64 start, end;
 244
 245         start = cache->key.objectid;
 246         end = start + cache->key.offset - 1;
 247
 248         clear_extent_bits(&root->fs_info->freed_extents[0],
 249                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 250         clear_extent_bits(&root->fs_info->freed_extents[1],
 251                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 252 }
 253
 254 static int exclude_super_stripes(struct btrfs_root *root,
 255                                  struct btrfs_block_group_cache *cache)
 256 {
 257         u64 bytenr;
 258         u64 *logical;
 259         int stripe_len;
 260         int i, nr, ret;
 261
 262         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 263                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 264                 cache->bytes_super += stripe_len;
 265                 ret = add_excluded_extent(root, cache->key.objectid,
 266                                           stripe_len);
 267                 if (ret)
 268                         return ret;
 269         }
 270
 271         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 272                 bytenr = btrfs_sb_offset(i);
 273                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 274                                        cache->key.objectid, bytenr,
 275                                        0, &logical, &nr, &stripe_len);
 276                 if (ret)
 277                         return ret;
 278
 279                 while (nr--) {
 280                         u64 start, len;
 281
 282                         if (logical[nr] > cache->key.objectid +
 283                             cache->key.offset)
 284                                 continue;
 285
 286                         if (logical[nr] + stripe_len <= cache->key.objectid)
 287                                 continue;
 288
 289                         start = logical[nr];
 290                         if (start < cache->key.objectid) {
 291                                 start = cache->key.objectid;
 292                                 len = (logical[nr] + stripe_len) - start;
 293                         } else {
 294                                 len = min_t(u64, stripe_len,
 295                                             cache->key.objectid +
 296                                             cache->key.offset - start);
 297                         }
 298
 299                         cache->bytes_super += len;
 300                         ret = add_excluded_extent(root, start, len);
 301                         if (ret) {
 302                                 kfree(logical);
 303                                 return ret;
 304                         }
 305                 }
 306
 307                 kfree(logical);
 308         }
 309         return 0;
 310 }
 311
 312 static struct btrfs_caching_control *
 313 get_caching_control(struct btrfs_block_group_cache *cache)
 314 {
 315         struct btrfs_caching_control *ctl;
 316
 317         spin_lock(&cache->lock);
 318         if (cache->cached != BTRFS_CACHE_STARTED) {
 319                 spin_unlock(&cache->lock);
 320                 return NULL;
 321         }
 322
 323         /* We're loading it the fast way, so we don't have a caching_ctl. */
 324         if (!cache->caching_ctl) {
 325                 spin_unlock(&cache->lock);
 326                 return NULL;
 327         }
 328
 329         ctl = cache->caching_ctl;
 330         atomic_inc(&ctl->count);
 331         spin_unlock(&cache->lock);
 332         return ctl;
 333 }
 334
 335 static void put_caching_control(struct btrfs_caching_control *ctl)
 336 {
 337         if (atomic_dec_and_test(&ctl->count))
 338                 kfree(ctl);
 339 }
 340
 341 /*
 342  * this is only called by cache_block_group, since we could have freed extents
 343  * we need to check the pinned_extents for any extents that can't be used yet
 344  * since their free space will be released as soon as the transaction commits.
 345  */
 346 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 347                               struct btrfs_fs_info *info, u64 start, u64 end)
 348 {
 349         u64 extent_start, extent_end, size, total_added = 0;
 350         int ret;
 351
 352         while (start < end) {
 353                 ret = find_first_extent_bit(info->pinned_extents, start,
 354                                             &extent_start, &extent_end,
 355                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 356                                             NULL);
 357                 if (ret)
 358                         break;
 359
 360                 if (extent_start <= start) {
 361                         start = extent_end + 1;
 362                 } else if (extent_start > start && extent_start < end) {
 363                         size = extent_start - start;
 364                         total_added += size;
 365                         ret = btrfs_add_free_space(block_group, start,
 366                                                    size);
 367                         BUG_ON(ret); /* -ENOMEM or logic error */
 368                         start = extent_end + 1;
 369                 } else {
 370                         break;
 371                 }
 372         }
 373
 374         if (start < end) {
 375                 size = end - start;
 376                 total_added += size;
 377                 ret = btrfs_add_free_space(block_group, start, size);
 378                 BUG_ON(ret); /* -ENOMEM or logic error */
 379         }
 380
 381         return total_added;
 382 }
 383
 384 static noinline void caching_thread(struct btrfs_work *work)
 385 {
 386         struct btrfs_block_group_cache *block_group;
 387         struct btrfs_fs_info *fs_info;
 388         struct btrfs_caching_control *caching_ctl;
 389         struct btrfs_root *extent_root;
 390         struct btrfs_path *path;
 391         struct extent_buffer *leaf;
 392         struct btrfs_key key;
 393         u64 total_found = 0;
 394         u64 last = 0;
 395         u32 nritems;
 396         int ret = -ENOMEM;
 397
 398         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 399         block_group = caching_ctl->block_group;
 400         fs_info = block_group->fs_info;
 401         extent_root = fs_info->extent_root;
 402
 403         path = btrfs_alloc_path();
 404         if (!path)
 405                 goto out;
 406
 407         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 408
 409         /*
 410          * We don't want to deadlock with somebody trying to allocate a new
 411          * extent for the extent root while also trying to search the extent
 412          * root to add free space.  So we skip locking and search the commit
 413          * root, since its read-only
 414          */
 415         path->skip_locking = 1;
 416         path->search_commit_root = 1;
 417         path->reada = 1;
 418
 419         key.objectid = last;
 420         key.offset = 0;
 421         key.type = BTRFS_EXTENT_ITEM_KEY;
 422 again:
 423         mutex_lock(&caching_ctl->mutex);
 424         /* need to make sure the commit_root doesn't disappear */
 425         down_read(&fs_info->commit_root_sem);
 426
 427 next:
 428         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 429         if (ret < 0)
 430                 goto err;
 431
 432         leaf = path->nodes[0];
 433         nritems = btrfs_header_nritems(leaf);
 434
 435         while (1) {
 436                 if (btrfs_fs_closing(fs_info) > 1) {
 437                         last = (u64)-1;
 438                         break;
 439                 }
 440
 441                 if (path->slots[0] < nritems) {
 442                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 443                 } else {
 444                         ret = find_next_key(path, 0, &key);
 445                         if (ret)
 446                                 break;
 447
 448                         if (need_resched() ||
 449                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 450                                 caching_ctl->progress = last;
 451                                 btrfs_release_path(path);
 452                                 up_read(&fs_info->commit_root_sem);
 453                                 mutex_unlock(&caching_ctl->mutex);
 454                                 cond_resched();
 455                                 goto again;
 456                         }
 457
 458                         ret = btrfs_next_leaf(extent_root, path);
 459                         if (ret < 0)
 460                                 goto err;
 461                         if (ret)
 462                                 break;
 463                         leaf = path->nodes[0];
 464                         nritems = btrfs_header_nritems(leaf);
 465                         continue;
 466                 }
 467
 468                 if (key.objectid < last) {
 469                         key.objectid = last;
 470                         key.offset = 0;
 471                         key.type = BTRFS_EXTENT_ITEM_KEY;
 472
 473                         caching_ctl->progress = last;
 474                         btrfs_release_path(path);
 475                         goto next;
 476                 }
 477
 478                 if (key.objectid < block_group->key.objectid) {
 479                         path->slots[0]++;
 480                         continue;
 481                 }
 482
 483                 if (key.objectid >= block_group->key.objectid +
 484                     block_group->key.offset)
 485                         break;
 486
 487                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 488                     key.type == BTRFS_METADATA_ITEM_KEY) {
 489                         total_found += add_new_free_space(block_group,
 490                                                           fs_info, last,
 491                                                           key.objectid);
 492                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 493                                 last = key.objectid +
 494                                         fs_info->tree_root->nodesize;
 495                         else
 496                                 last = key.objectid + key.offset;
 497
 498                         if (total_found > (1024 * 1024 * 2)) {
 499                                 total_found = 0;
 500                                 wake_up(&caching_ctl->wait);
 501                         }
 502                 }
 503                 path->slots[0]++;
 504         }
 505         ret = 0;
 506
 507         total_found += add_new_free_space(block_group, fs_info, last,
 508                                           block_group->key.objectid +
 509                                           block_group->key.offset);
 510         caching_ctl->progress = (u64)-1;
 511
 512         spin_lock(&block_group->lock);
 513         block_group->caching_ctl = NULL;
 514         block_group->cached = BTRFS_CACHE_FINISHED;
 515         spin_unlock(&block_group->lock);
 516
 517 err:
 518         btrfs_free_path(path);
 519         up_read(&fs_info->commit_root_sem);
 520
 521         free_excluded_extents(extent_root, block_group);
 522
 523         mutex_unlock(&caching_ctl->mutex);
 524 out:
 525         if (ret) {
 526                 spin_lock(&block_group->lock);
 527                 block_group->caching_ctl = NULL;
 528                 block_group->cached = BTRFS_CACHE_ERROR;
 529                 spin_unlock(&block_group->lock);
 530         }
 531         wake_up(&caching_ctl->wait);
 532
 533         put_caching_control(caching_ctl);
 534         btrfs_put_block_group(block_group);
 535 }
 536
 537 static int cache_block_group(struct btrfs_block_group_cache *cache,
 538                              int load_cache_only)
 539 {
 540         DEFINE_WAIT(wait);
 541         struct btrfs_fs_info *fs_info = cache->fs_info;
 542         struct btrfs_caching_control *caching_ctl;
 543         int ret = 0;
 544
 545         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 546         if (!caching_ctl)
 547                 return -ENOMEM;
 548
 549         INIT_LIST_HEAD(&caching_ctl->list);
 550         mutex_init(&caching_ctl->mutex);
 551         init_waitqueue_head(&caching_ctl->wait);
 552         caching_ctl->block_group = cache;
 553         caching_ctl->progress = cache->key.objectid;
 554         atomic_set(&caching_ctl->count, 1);
 555         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 556                         caching_thread, NULL, NULL);
 557
 558         spin_lock(&cache->lock);
 559         /*
 560          * This should be a rare occasion, but this could happen I think in the
 561          * case where one thread starts to load the space cache info, and then
 562          * some other thread starts a transaction commit which tries to do an
 563          * allocation while the other thread is still loading the space cache
 564          * info.  The previous loop should have kept us from choosing this block
 565          * group, but if we've moved to the state where we will wait on caching
 566          * block groups we need to first check if we're doing a fast load here,
 567          * so we can wait for it to finish, otherwise we could end up allocating
 568          * from a block group who's cache gets evicted for one reason or
 569          * another.
 570          */
 571         while (cache->cached == BTRFS_CACHE_FAST) {
 572                 struct btrfs_caching_control *ctl;
 573
 574                 ctl = cache->caching_ctl;
 575                 atomic_inc(&ctl->count);
 576                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 577                 spin_unlock(&cache->lock);
 578
 579                 schedule();
 580
 581                 finish_wait(&ctl->wait, &wait);
 582                 put_caching_control(ctl);
 583                 spin_lock(&cache->lock);
 584         }
 585
 586         if (cache->cached != BTRFS_CACHE_NO) {
 587                 spin_unlock(&cache->lock);
 588                 kfree(caching_ctl);
 589                 return 0;
 590         }
 591         WARN_ON(cache->caching_ctl);
 592         cache->caching_ctl = caching_ctl;
 593         cache->cached = BTRFS_CACHE_FAST;
 594         spin_unlock(&cache->lock);
 595
 596         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 597                 ret = load_free_space_cache(fs_info, cache);
 598
 599                 spin_lock(&cache->lock);
 600                 if (ret == 1) {
 601                         cache->caching_ctl = NULL;
 602                         cache->cached = BTRFS_CACHE_FINISHED;
 603                         cache->last_byte_to_unpin = (u64)-1;
 604                 } else {
 605                         if (load_cache_only) {
 606                                 cache->caching_ctl = NULL;
 607                                 cache->cached = BTRFS_CACHE_NO;
 608                         } else {
 609                                 cache->cached = BTRFS_CACHE_STARTED;
 610                         }
 611                 }
 612                 spin_unlock(&cache->lock);
 613                 wake_up(&caching_ctl->wait);
 614                 if (ret == 1) {
 615                         put_caching_control(caching_ctl);
 616                         free_excluded_extents(fs_info->extent_root, cache);
 617                         return 0;
 618                 }
 619         } else {
 620                 /*
 621                  * We are not going to do the fast caching, set cached to the
 622                  * appropriate value and wakeup any waiters.
 623                  */
 624                 spin_lock(&cache->lock);
 625                 if (load_cache_only) {
 626                         cache->caching_ctl = NULL;
 627                         cache->cached = BTRFS_CACHE_NO;
 628                 } else {
 629                         cache->cached = BTRFS_CACHE_STARTED;
 630                 }
 631                 spin_unlock(&cache->lock);
 632                 wake_up(&caching_ctl->wait);
 633         }
 634
 635         if (load_cache_only) {
 636                 put_caching_control(caching_ctl);
 637                 return 0;
 638         }
 639
 640         down_write(&fs_info->commit_root_sem);
 641         atomic_inc(&caching_ctl->count);
 642         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 643         up_write(&fs_info->commit_root_sem);
 644
 645         btrfs_get_block_group(cache);
 646
 647         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 648
 649         return ret;
 650 }
 651
 652 /*
 653  * return the block group that starts at or after bytenr
 654  */
 655 static struct btrfs_block_group_cache *
 656 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 657 {
 658         struct btrfs_block_group_cache *cache;
 659
 660         cache = block_group_cache_tree_search(info, bytenr, 0);
 661
 662         return cache;
 663 }
 664
 665 /*
 666  * return the block group that contains the given bytenr
 667  */
 668 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 669                                                  struct btrfs_fs_info *info,
 670                                                  u64 bytenr)
 671 {
 672         struct btrfs_block_group_cache *cache;
 673
 674         cache = block_group_cache_tree_search(info, bytenr, 1);
 675
 676         return cache;
 677 }
 678
 679 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 680                                                   u64 flags)
 681 {
 682         struct list_head *head = &info->space_info;
 683         struct btrfs_space_info *found;
 684
 685         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 686
 687         rcu_read_lock();
 688         list_for_each_entry_rcu(found, head, list) {
 689                 if (found->flags & flags) {
 690                         rcu_read_unlock();
 691                         return found;
 692                 }
 693         }
 694         rcu_read_unlock();
 695         return NULL;
 696 }
 697
 698 /*
 699  * after adding space to the filesystem, we need to clear the full flags
 700  * on all the space infos.
 701  */
 702 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 703 {
 704         struct list_head *head = &info->space_info;
 705         struct btrfs_space_info *found;
 706
 707         rcu_read_lock();
 708         list_for_each_entry_rcu(found, head, list)
 709                 found->full = 0;
 710         rcu_read_unlock();
 711 }
 712
 713 /* simple helper to search for an existing data extent at a given offset */
 714 int btrfs_lookup_data_extent(struct btrfs_root *root, u64 start, u64 len)
 715 {
 716         int ret;
 717         struct btrfs_key key;
 718         struct btrfs_path *path;
 719
 720         path = btrfs_alloc_path();
 721         if (!path)
 722                 return -ENOMEM;
 723
 724         key.objectid = start;
 725         key.offset = len;
 726         key.type = BTRFS_EXTENT_ITEM_KEY;
 727         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 728                                 0, 0);
 729         btrfs_free_path(path);
 730         return ret;
 731 }
 732
 733 /*
 734  * helper function to lookup reference count and flags of a tree block.
 735  *
 736  * the head node for delayed ref is used to store the sum of all the
 737  * reference count modifications queued up in the rbtree. the head
 738  * node may also store the extent flags to set. This way you can check
 739  * to see what the reference count and extent flags would be if all of
 740  * the delayed refs are not processed.
 741  */
 742 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 743                              struct btrfs_root *root, u64 bytenr,
 744                              u64 offset, int metadata, u64 *refs, u64 *flags)
 745 {
 746         struct btrfs_delayed_ref_head *head;
 747         struct btrfs_delayed_ref_root *delayed_refs;
 748         struct btrfs_path *path;
 749         struct btrfs_extent_item *ei;
 750         struct extent_buffer *leaf;
 751         struct btrfs_key key;
 752         u32 item_size;
 753         u64 num_refs;
 754         u64 extent_flags;
 755         int ret;
 756
 757         /*
 758          * If we don't have skinny metadata, don't bother doing anything
 759          * different
 760          */
 761         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 762                 offset = root->nodesize;
 763                 metadata = 0;
 764         }
 765
 766         path = btrfs_alloc_path();
 767         if (!path)
 768                 return -ENOMEM;
 769
 770         if (!trans) {
 771                 path->skip_locking = 1;
 772                 path->search_commit_root = 1;
 773         }
 774
 775 search_again:
 776         key.objectid = bytenr;
 777         key.offset = offset;
 778         if (metadata)
 779                 key.type = BTRFS_METADATA_ITEM_KEY;
 780         else
 781                 key.type = BTRFS_EXTENT_ITEM_KEY;
 782
 783         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 784                                 &key, path, 0, 0);
 785         if (ret < 0)
 786                 goto out_free;
 787
 788         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 789                 if (path->slots[0]) {
 790                         path->slots[0]--;
 791                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 792                                               path->slots[0]);
 793                         if (key.objectid == bytenr &&
 794                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 795                             key.offset == root->nodesize)
 796                                 ret = 0;
 797                 }
 798         }
 799
 800         if (ret == 0) {
 801                 leaf = path->nodes[0];
 802                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 803                 if (item_size >= sizeof(*ei)) {
 804                         ei = btrfs_item_ptr(leaf, path->slots[0],
 805                                             struct btrfs_extent_item);
 806                         num_refs = btrfs_extent_refs(leaf, ei);
 807                         extent_flags = btrfs_extent_flags(leaf, ei);
 808                 } else {
 809 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 810                         struct btrfs_extent_item_v0 *ei0;
 811                         BUG_ON(item_size != sizeof(*ei0));
 812                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 813                                              struct btrfs_extent_item_v0);
 814                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 815                         /* FIXME: this isn't correct for data */
 816                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 817 #else
 818                         BUG();
 819 #endif
 820                 }
 821                 BUG_ON(num_refs == 0);
 822         } else {
 823                 num_refs = 0;
 824                 extent_flags = 0;
 825                 ret = 0;
 826         }
 827
 828         if (!trans)
 829                 goto out;
 830
 831         delayed_refs = &trans->transaction->delayed_refs;
 832         spin_lock(&delayed_refs->lock);
 833         head = btrfs_find_delayed_ref_head(trans, bytenr);
 834         if (head) {
 835                 if (!mutex_trylock(&head->mutex)) {
 836                         atomic_inc(&head->node.refs);
 837                         spin_unlock(&delayed_refs->lock);
 838
 839                         btrfs_release_path(path);
 840
 841                         /*
 842                          * Mutex was contended, block until it's released and try
 843                          * again
 844                          */
 845                         mutex_lock(&head->mutex);
 846                         mutex_unlock(&head->mutex);
 847                         btrfs_put_delayed_ref(&head->node);
 848                         goto search_again;
 849                 }
 850                 spin_lock(&head->lock);
 851                 if (head->extent_op && head->extent_op->update_flags)
 852                         extent_flags |= head->extent_op->flags_to_set;
 853                 else
 854                         BUG_ON(num_refs == 0);
 855
 856                 num_refs += head->node.ref_mod;
 857                 spin_unlock(&head->lock);
 858                 mutex_unlock(&head->mutex);
 859         }
 860         spin_unlock(&delayed_refs->lock);
 861 out:
 862         WARN_ON(num_refs == 0);
 863         if (refs)
 864                 *refs = num_refs;
 865         if (flags)
 866                 *flags = extent_flags;
 867 out_free:
 868         btrfs_free_path(path);
 869         return ret;
 870 }
 871
 872 /*
 873  * Back reference rules.  Back refs have three main goals:
 874  *
 875  * 1) differentiate between all holders of references to an extent so that
 876  *    when a reference is dropped we can make sure it was a valid reference
 877  *    before freeing the extent.
 878  *
 879  * 2) Provide enough information to quickly find the holders of an extent
 880  *    if we notice a given block is corrupted or bad.
 881  *
 882  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 883  *    maintenance.  This is actually the same as #2, but with a slightly
 884  *    different use case.
 885  *
 886  * There are two kinds of back refs. The implicit back refs is optimized
 887  * for pointers in non-shared tree blocks. For a given pointer in a block,
 888  * back refs of this kind provide information about the block's owner tree
 889  * and the pointer's key. These information allow us to find the block by
 890  * b-tree searching. The full back refs is for pointers in tree blocks not
 891  * referenced by their owner trees. The location of tree block is recorded
 892  * in the back refs. Actually the full back refs is generic, and can be
 893  * used in all cases the implicit back refs is used. The major shortcoming
 894  * of the full back refs is its overhead. Every time a tree block gets
 895  * COWed, we have to update back refs entry for all pointers in it.
 896  *
 897  * For a newly allocated tree block, we use implicit back refs for
 898  * pointers in it. This means most tree related operations only involve
 899  * implicit back refs. For a tree block created in old transaction, the
 900  * only way to drop a reference to it is COW it. So we can detect the
 901  * event that tree block loses its owner tree's reference and do the
 902  * back refs conversion.
 903  *
 904  * When a tree block is COW'd through a tree, there are four cases:
 905  *
 906  * The reference count of the block is one and the tree is the block's
 907  * owner tree. Nothing to do in this case.
 908  *
 909  * The reference count of the block is one and the tree is not the
 910  * block's owner tree. In this case, full back refs is used for pointers
 911  * in the block. Remove these full back refs, add implicit back refs for
 912  * every pointers in the new block.
 913  *
 914  * The reference count of the block is greater than one and the tree is
 915  * the block's owner tree. In this case, implicit back refs is used for
 916  * pointers in the block. Add full back refs for every pointers in the
 917  * block, increase lower level extents' reference counts. The original
 918  * implicit back refs are entailed to the new block.
 919  *
 920  * The reference count of the block is greater than one and the tree is
 921  * not the block's owner tree. Add implicit back refs for every pointer in
 922  * the new block, increase lower level extents' reference count.
 923  *
 924  * Back Reference Key composing:
 925  *
 926  * The key objectid corresponds to the first byte in the extent,
 927  * The key type is used to differentiate between types of back refs.
 928  * There are different meanings of the key offset for different types
 929  * of back refs.
 930  *
 931  * File extents can be referenced by:
 932  *
 933  * - multiple snapshots, subvolumes, or different generations in one subvol
 934  * - different files inside a single subvolume
 935  * - different offsets inside a file (bookend extents in file.c)
 936  *
 937  * The extent ref structure for the implicit back refs has fields for:
 938  *
 939  * - Objectid of the subvolume root
 940  * - objectid of the file holding the reference
 941  * - original offset in the file
 942  * - how many bookend extents
 943  *
 944  * The key offset for the implicit back refs is hash of the first
 945  * three fields.
 946  *
 947  * The extent ref structure for the full back refs has field for:
 948  *
 949  * - number of pointers in the tree leaf
 950  *
 951  * The key offset for the implicit back refs is the first byte of
 952  * the tree leaf
 953  *
 954  * When a file extent is allocated, The implicit back refs is used.
 955  * the fields are filled in:
 956  *
 957  *     (root_key.objectid, inode objectid, offset in file, 1)
 958  *
 959  * When a file extent is removed file truncation, we find the
 960  * corresponding implicit back refs and check the following fields:
 961  *
 962  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 963  *
 964  * Btree extents can be referenced by:
 965  *
 966  * - Different subvolumes
 967  *
 968  * Both the implicit back refs and the full back refs for tree blocks
 969  * only consist of key. The key offset for the implicit back refs is
 970  * objectid of block's owner tree. The key offset for the full back refs
 971  * is the first byte of parent block.
 972  *
 973  * When implicit back refs is used, information about the lowest key and
 974  * level of the tree block are required. These information are stored in
 975  * tree block info structure.
 976  */
 977
 978 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 979 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 980                                   struct btrfs_root *root,
 981                                   struct btrfs_path *path,
 982                                   u64 owner, u32 extra_size)
 983 {
 984         struct btrfs_extent_item *item;
 985         struct btrfs_extent_item_v0 *ei0;
 986         struct btrfs_extent_ref_v0 *ref0;
 987         struct btrfs_tree_block_info *bi;
 988         struct extent_buffer *leaf;
 989         struct btrfs_key key;
 990         struct btrfs_key found_key;
 991         u32 new_size = sizeof(*item);
 992         u64 refs;
 993         int ret;
 994
 995         leaf = path->nodes[0];
 996         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 997
 998         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 999         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1000                              struct btrfs_extent_item_v0);
1001         refs = btrfs_extent_refs_v0(leaf, ei0);
1002
1003         if (owner == (u64)-1) {
1004                 while (1) {
1005                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1006                                 ret = btrfs_next_leaf(root, path);
1007                                 if (ret < 0)
1008                                         return ret;
1009                                 BUG_ON(ret > 0); /* Corruption */
1010                                 leaf = path->nodes[0];
1011                         }
1012                         btrfs_item_key_to_cpu(leaf, &found_key,
1013                                               path->slots[0]);
1014                         BUG_ON(key.objectid != found_key.objectid);
1015                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1016                                 path->slots[0]++;
1017                                 continue;
1018                         }
1019                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1020                                               struct btrfs_extent_ref_v0);
1021                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1022                         break;
1023                 }
1024         }
1025         btrfs_release_path(path);
1026
1027         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1028                 new_size += sizeof(*bi);
1029
1030         new_size -= sizeof(*ei0);
1031         ret = btrfs_search_slot(trans, root, &key, path,
1032                                 new_size + extra_size, 1);
1033         if (ret < 0)
1034                 return ret;
1035         BUG_ON(ret); /* Corruption */
1036
1037         btrfs_extend_item(root, path, new_size);
1038
1039         leaf = path->nodes[0];
1040         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1041         btrfs_set_extent_refs(leaf, item, refs);
1042         /* FIXME: get real generation */
1043         btrfs_set_extent_generation(leaf, item, 0);
1044         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1045                 btrfs_set_extent_flags(leaf, item,
1046                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1047                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1048                 bi = (struct btrfs_tree_block_info *)(item + 1);
1049                 /* FIXME: get first key of the block */
1050                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1051                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1052         } else {
1053                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1054         }
1055         btrfs_mark_buffer_dirty(leaf);
1056         return 0;
1057 }
1058 #endif
1059
1060 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1061 {
1062         u32 high_crc = ~(u32)0;
1063         u32 low_crc = ~(u32)0;
1064         __le64 lenum;
1065
1066         lenum = cpu_to_le64(root_objectid);
1067         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1068         lenum = cpu_to_le64(owner);
1069         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1070         lenum = cpu_to_le64(offset);
1071         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1072
1073         return ((u64)high_crc << 31) ^ (u64)low_crc;
1074 }
1075
1076 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1077                                      struct btrfs_extent_data_ref *ref)
1078 {
1079         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1080                                     btrfs_extent_data_ref_objectid(leaf, ref),
1081                                     btrfs_extent_data_ref_offset(leaf, ref));
1082 }
1083
1084 static int match_extent_data_ref(struct extent_buffer *leaf,
1085                                  struct btrfs_extent_data_ref *ref,
1086                                  u64 root_objectid, u64 owner, u64 offset)
1087 {
1088         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1089             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1090             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1091                 return 0;
1092         return 1;
1093 }
1094
1095 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1096                                            struct btrfs_root *root,
1097                                            struct btrfs_path *path,
1098                                            u64 bytenr, u64 parent,
1099                                            u64 root_objectid,
1100                                            u64 owner, u64 offset)
1101 {
1102         struct btrfs_key key;
1103         struct btrfs_extent_data_ref *ref;
1104         struct extent_buffer *leaf;
1105         u32 nritems;
1106         int ret;
1107         int recow;
1108         int err = -ENOENT;
1109
1110         key.objectid = bytenr;
1111         if (parent) {
1112                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1113                 key.offset = parent;
1114         } else {
1115                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1116                 key.offset = hash_extent_data_ref(root_objectid,
1117                                                   owner, offset);
1118         }
1119 again:
1120         recow = 0;
1121         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1122         if (ret < 0) {
1123                 err = ret;
1124                 goto fail;
1125         }
1126
1127         if (parent) {
1128                 if (!ret)
1129                         return 0;
1130 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1131                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1132                 btrfs_release_path(path);
1133                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1134                 if (ret < 0) {
1135                         err = ret;
1136                         goto fail;
1137                 }
1138                 if (!ret)
1139                         return 0;
1140 #endif
1141                 goto fail;
1142         }
1143
1144         leaf = path->nodes[0];
1145         nritems = btrfs_header_nritems(leaf);
1146         while (1) {
1147                 if (path->slots[0] >= nritems) {
1148                         ret = btrfs_next_leaf(root, path);
1149                         if (ret < 0)
1150                                 err = ret;
1151                         if (ret)
1152                                 goto fail;
1153
1154                         leaf = path->nodes[0];
1155                         nritems = btrfs_header_nritems(leaf);
1156                         recow = 1;
1157                 }
1158
1159                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1160                 if (key.objectid != bytenr ||
1161                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1162                         goto fail;
1163
1164                 ref = btrfs_item_ptr(leaf, path->slots[0],
1165                                      struct btrfs_extent_data_ref);
1166
1167                 if (match_extent_data_ref(leaf, ref, root_objectid,
1168                                           owner, offset)) {
1169                         if (recow) {
1170                                 btrfs_release_path(path);
1171                                 goto again;
1172                         }
1173                         err = 0;
1174                         break;
1175                 }
1176                 path->slots[0]++;
1177         }
1178 fail:
1179         return err;
1180 }
1181
1182 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1183                                            struct btrfs_root *root,
1184                                            struct btrfs_path *path,
1185                                            u64 bytenr, u64 parent,
1186                                            u64 root_objectid, u64 owner,
1187                                            u64 offset, int refs_to_add)
1188 {
1189         struct btrfs_key key;
1190         struct extent_buffer *leaf;
1191         u32 size;
1192         u32 num_refs;
1193         int ret;
1194
1195         key.objectid = bytenr;
1196         if (parent) {
1197                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1198                 key.offset = parent;
1199                 size = sizeof(struct btrfs_shared_data_ref);
1200         } else {
1201                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1202                 key.offset = hash_extent_data_ref(root_objectid,
1203                                                   owner, offset);
1204                 size = sizeof(struct btrfs_extent_data_ref);
1205         }
1206
1207         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1208         if (ret && ret != -EEXIST)
1209                 goto fail;
1210
1211         leaf = path->nodes[0];
1212         if (parent) {
1213                 struct btrfs_shared_data_ref *ref;
1214                 ref = btrfs_item_ptr(leaf, path->slots[0],
1215                                      struct btrfs_shared_data_ref);
1216                 if (ret == 0) {
1217                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1218                 } else {
1219                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1220                         num_refs += refs_to_add;
1221                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1222                 }
1223         } else {
1224                 struct btrfs_extent_data_ref *ref;
1225                 while (ret == -EEXIST) {
1226                         ref = btrfs_item_ptr(leaf, path->slots[0],
1227                                              struct btrfs_extent_data_ref);
1228                         if (match_extent_data_ref(leaf, ref, root_objectid,
1229                                                   owner, offset))
1230                                 break;
1231                         btrfs_release_path(path);
1232                         key.offset++;
1233                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1234                                                       size);
1235                         if (ret && ret != -EEXIST)
1236                                 goto fail;
1237
1238                         leaf = path->nodes[0];
1239                 }
1240                 ref = btrfs_item_ptr(leaf, path->slots[0],
1241                                      struct btrfs_extent_data_ref);
1242                 if (ret == 0) {
1243                         btrfs_set_extent_data_ref_root(leaf, ref,
1244                                                        root_objectid);
1245                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1246                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1247                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1248                 } else {
1249                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1250                         num_refs += refs_to_add;
1251                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1252                 }
1253         }
1254         btrfs_mark_buffer_dirty(leaf);
1255         ret = 0;
1256 fail:
1257         btrfs_release_path(path);
1258         return ret;
1259 }
1260
1261 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1262                                            struct btrfs_root *root,
1263                                            struct btrfs_path *path,
1264                                            int refs_to_drop, int *last_ref)
1265 {
1266         struct btrfs_key key;
1267         struct btrfs_extent_data_ref *ref1 = NULL;
1268         struct btrfs_shared_data_ref *ref2 = NULL;
1269         struct extent_buffer *leaf;
1270         u32 num_refs = 0;
1271         int ret = 0;
1272
1273         leaf = path->nodes[0];
1274         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1275
1276         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1277                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1278                                       struct btrfs_extent_data_ref);
1279                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1280         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1281                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1282                                       struct btrfs_shared_data_ref);
1283                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1284 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1285         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1286                 struct btrfs_extent_ref_v0 *ref0;
1287                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1288                                       struct btrfs_extent_ref_v0);
1289                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1290 #endif
1291         } else {
1292                 BUG();
1293         }
1294
1295         BUG_ON(num_refs < refs_to_drop);
1296         num_refs -= refs_to_drop;
1297
1298         if (num_refs == 0) {
1299                 ret = btrfs_del_item(trans, root, path);
1300                 *last_ref = 1;
1301         } else {
1302                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1303                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1304                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1305                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1306 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1307                 else {
1308                         struct btrfs_extent_ref_v0 *ref0;
1309                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1310                                         struct btrfs_extent_ref_v0);
1311                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1312                 }
1313 #endif
1314                 btrfs_mark_buffer_dirty(leaf);
1315         }
1316         return ret;
1317 }
1318
1319 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1320                                           struct btrfs_path *path,
1321                                           struct btrfs_extent_inline_ref *iref)
1322 {
1323         struct btrfs_key key;
1324         struct extent_buffer *leaf;
1325         struct btrfs_extent_data_ref *ref1;
1326         struct btrfs_shared_data_ref *ref2;
1327         u32 num_refs = 0;
1328
1329         leaf = path->nodes[0];
1330         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1331         if (iref) {
1332                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1333                     BTRFS_EXTENT_DATA_REF_KEY) {
1334                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1335                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1336                 } else {
1337                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1338                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1339                 }
1340         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1341                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1342                                       struct btrfs_extent_data_ref);
1343                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1344         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1345                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1346                                       struct btrfs_shared_data_ref);
1347                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1348 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1349         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1350                 struct btrfs_extent_ref_v0 *ref0;
1351                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1352                                       struct btrfs_extent_ref_v0);
1353                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1354 #endif
1355         } else {
1356                 WARN_ON(1);
1357         }
1358         return num_refs;
1359 }
1360
1361 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1362                                           struct btrfs_root *root,
1363                                           struct btrfs_path *path,
1364                                           u64 bytenr, u64 parent,
1365                                           u64 root_objectid)
1366 {
1367         struct btrfs_key key;
1368         int ret;
1369
1370         key.objectid = bytenr;
1371         if (parent) {
1372                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1373                 key.offset = parent;
1374         } else {
1375                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1376                 key.offset = root_objectid;
1377         }
1378
1379         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1380         if (ret > 0)
1381                 ret = -ENOENT;
1382 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1383         if (ret == -ENOENT && parent) {
1384                 btrfs_release_path(path);
1385                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1386                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1387                 if (ret > 0)
1388                         ret = -ENOENT;
1389         }
1390 #endif
1391         return ret;
1392 }
1393
1394 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1395                                           struct btrfs_root *root,
1396                                           struct btrfs_path *path,
1397                                           u64 bytenr, u64 parent,
1398                                           u64 root_objectid)
1399 {
1400         struct btrfs_key key;
1401         int ret;
1402
1403         key.objectid = bytenr;
1404         if (parent) {
1405                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1406                 key.offset = parent;
1407         } else {
1408                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1409                 key.offset = root_objectid;
1410         }
1411
1412         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1413         btrfs_release_path(path);
1414         return ret;
1415 }
1416
1417 static inline int extent_ref_type(u64 parent, u64 owner)
1418 {
1419         int type;
1420         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1421                 if (parent > 0)
1422                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1423                 else
1424                         type = BTRFS_TREE_BLOCK_REF_KEY;
1425         } else {
1426                 if (parent > 0)
1427                         type = BTRFS_SHARED_DATA_REF_KEY;
1428                 else
1429                         type = BTRFS_EXTENT_DATA_REF_KEY;
1430         }
1431         return type;
1432 }
1433
1434 static int find_next_key(struct btrfs_path *path, int level,
1435                          struct btrfs_key *key)
1436
1437 {
1438         for (; level < BTRFS_MAX_LEVEL; level++) {
1439                 if (!path->nodes[level])
1440                         break;
1441                 if (path->slots[level] + 1 >=
1442                     btrfs_header_nritems(path->nodes[level]))
1443                         continue;
1444                 if (level == 0)
1445                         btrfs_item_key_to_cpu(path->nodes[level], key,
1446                                               path->slots[level] + 1);
1447                 else
1448                         btrfs_node_key_to_cpu(path->nodes[level], key,
1449                                               path->slots[level] + 1);
1450                 return 0;
1451         }
1452         return 1;
1453 }
1454
1455 /*
1456  * look for inline back ref. if back ref is found, *ref_ret is set
1457  * to the address of inline back ref, and 0 is returned.
1458  *
1459  * if back ref isn't found, *ref_ret is set to the address where it
1460  * should be inserted, and -ENOENT is returned.
1461  *
1462  * if insert is true and there are too many inline back refs, the path
1463  * points to the extent item, and -EAGAIN is returned.
1464  *
1465  * NOTE: inline back refs are ordered in the same way that back ref
1466  *       items in the tree are ordered.
1467  */
1468 static noinline_for_stack
1469 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1470                                  struct btrfs_root *root,
1471                                  struct btrfs_path *path,
1472                                  struct btrfs_extent_inline_ref **ref_ret,
1473                                  u64 bytenr, u64 num_bytes,
1474                                  u64 parent, u64 root_objectid,
1475                                  u64 owner, u64 offset, int insert)
1476 {
1477         struct btrfs_key key;
1478         struct extent_buffer *leaf;
1479         struct btrfs_extent_item *ei;
1480         struct btrfs_extent_inline_ref *iref;
1481         u64 flags;
1482         u64 item_size;
1483         unsigned long ptr;
1484         unsigned long end;
1485         int extra_size;
1486         int type;
1487         int want;
1488         int ret;
1489         int err = 0;
1490         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1491                                                  SKINNY_METADATA);
1492
1493         key.objectid = bytenr;
1494         key.type = BTRFS_EXTENT_ITEM_KEY;
1495         key.offset = num_bytes;
1496
1497         want = extent_ref_type(parent, owner);
1498         if (insert) {
1499                 extra_size = btrfs_extent_inline_ref_size(want);
1500                 path->keep_locks = 1;
1501         } else
1502                 extra_size = -1;
1503
1504         /*
1505          * Owner is our parent level, so we can just add one to get the level
1506          * for the block we are interested in.
1507          */
1508         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1509                 key.type = BTRFS_METADATA_ITEM_KEY;
1510                 key.offset = owner;
1511         }
1512
1513 again:
1514         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1515         if (ret < 0) {
1516                 err = ret;
1517                 goto out;
1518         }
1519
1520         /*
1521          * We may be a newly converted file system which still has the old fat
1522          * extent entries for metadata, so try and see if we have one of those.
1523          */
1524         if (ret > 0 && skinny_metadata) {
1525                 skinny_metadata = false;
1526                 if (path->slots[0]) {
1527                         path->slots[0]--;
1528                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1529                                               path->slots[0]);
1530                         if (key.objectid == bytenr &&
1531                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1532                             key.offset == num_bytes)
1533                                 ret = 0;
1534                 }
1535                 if (ret) {
1536                         key.objectid = bytenr;
1537                         key.type = BTRFS_EXTENT_ITEM_KEY;
1538                         key.offset = num_bytes;
1539                         btrfs_release_path(path);
1540                         goto again;
1541                 }
1542         }
1543
1544         if (ret && !insert) {
1545                 err = -ENOENT;
1546                 goto out;
1547         } else if (WARN_ON(ret)) {
1548                 err = -EIO;
1549                 goto out;
1550         }
1551
1552         leaf = path->nodes[0];
1553         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1554 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1555         if (item_size < sizeof(*ei)) {
1556                 if (!insert) {
1557                         err = -ENOENT;
1558                         goto out;
1559                 }
1560                 ret = convert_extent_item_v0(trans, root, path, owner,
1561                                              extra_size);
1562                 if (ret < 0) {
1563                         err = ret;
1564                         goto out;
1565                 }
1566                 leaf = path->nodes[0];
1567                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1568         }
1569 #endif
1570         BUG_ON(item_size < sizeof(*ei));
1571
1572         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1573         flags = btrfs_extent_flags(leaf, ei);
1574
1575         ptr = (unsigned long)(ei + 1);
1576         end = (unsigned long)ei + item_size;
1577
1578         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1579                 ptr += sizeof(struct btrfs_tree_block_info);
1580                 BUG_ON(ptr > end);
1581         }
1582
1583         err = -ENOENT;
1584         while (1) {
1585                 if (ptr >= end) {
1586                         WARN_ON(ptr > end);
1587                         break;
1588                 }
1589                 iref = (struct btrfs_extent_inline_ref *)ptr;
1590                 type = btrfs_extent_inline_ref_type(leaf, iref);
1591                 if (want < type)
1592                         break;
1593                 if (want > type) {
1594                         ptr += btrfs_extent_inline_ref_size(type);
1595                         continue;
1596                 }
1597
1598                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1599                         struct btrfs_extent_data_ref *dref;
1600                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1601                         if (match_extent_data_ref(leaf, dref, root_objectid,
1602                                                   owner, offset)) {
1603                                 err = 0;
1604                                 break;
1605                         }
1606                         if (hash_extent_data_ref_item(leaf, dref) <
1607                             hash_extent_data_ref(root_objectid, owner, offset))
1608                                 break;
1609                 } else {
1610                         u64 ref_offset;
1611                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1612                         if (parent > 0) {
1613                                 if (parent == ref_offset) {
1614                                         err = 0;
1615                                         break;
1616                                 }
1617                                 if (ref_offset < parent)
1618                                         break;
1619                         } else {
1620                                 if (root_objectid == ref_offset) {
1621                                         err = 0;
1622                                         break;
1623                                 }
1624                                 if (ref_offset < root_objectid)
1625                                         break;
1626                         }
1627                 }
1628                 ptr += btrfs_extent_inline_ref_size(type);
1629         }
1630         if (err == -ENOENT && insert) {
1631                 if (item_size + extra_size >=
1632                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1633                         err = -EAGAIN;
1634                         goto out;
1635                 }
1636                 /*
1637                  * To add new inline back ref, we have to make sure
1638                  * there is no corresponding back ref item.
1639                  * For simplicity, we just do not add new inline back
1640                  * ref if there is any kind of item for this block
1641                  */
1642                 if (find_next_key(path, 0, &key) == 0 &&
1643                     key.objectid == bytenr &&
1644                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1645                         err = -EAGAIN;
1646                         goto out;
1647                 }
1648         }
1649         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1650 out:
1651         if (insert) {
1652                 path->keep_locks = 0;
1653                 btrfs_unlock_up_safe(path, 1);
1654         }
1655         return err;
1656 }
1657
1658 /*
1659  * helper to add new inline back ref
1660  */
1661 static noinline_for_stack
1662 void setup_inline_extent_backref(struct btrfs_root *root,
1663                                  struct btrfs_path *path,
1664                                  struct btrfs_extent_inline_ref *iref,
1665                                  u64 parent, u64 root_objectid,
1666                                  u64 owner, u64 offset, int refs_to_add,
1667                                  struct btrfs_delayed_extent_op *extent_op)
1668 {
1669         struct extent_buffer *leaf;
1670         struct btrfs_extent_item *ei;
1671         unsigned long ptr;
1672         unsigned long end;
1673         unsigned long item_offset;
1674         u64 refs;
1675         int size;
1676         int type;
1677
1678         leaf = path->nodes[0];
1679         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1680         item_offset = (unsigned long)iref - (unsigned long)ei;
1681
1682         type = extent_ref_type(parent, owner);
1683         size = btrfs_extent_inline_ref_size(type);
1684
1685         btrfs_extend_item(root, path, size);
1686
1687         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1688         refs = btrfs_extent_refs(leaf, ei);
1689         refs += refs_to_add;
1690         btrfs_set_extent_refs(leaf, ei, refs);
1691         if (extent_op)
1692                 __run_delayed_extent_op(extent_op, leaf, ei);
1693
1694         ptr = (unsigned long)ei + item_offset;
1695         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1696         if (ptr < end - size)
1697                 memmove_extent_buffer(leaf, ptr + size, ptr,
1698                                       end - size - ptr);
1699
1700         iref = (struct btrfs_extent_inline_ref *)ptr;
1701         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1702         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1703                 struct btrfs_extent_data_ref *dref;
1704                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1705                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1706                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1707                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1708                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1709         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1710                 struct btrfs_shared_data_ref *sref;
1711                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1712                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1713                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1714         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1715                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1716         } else {
1717                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1718         }
1719         btrfs_mark_buffer_dirty(leaf);
1720 }
1721
1722 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1723                                  struct btrfs_root *root,
1724                                  struct btrfs_path *path,
1725                                  struct btrfs_extent_inline_ref **ref_ret,
1726                                  u64 bytenr, u64 num_bytes, u64 parent,
1727                                  u64 root_objectid, u64 owner, u64 offset)
1728 {
1729         int ret;
1730
1731         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1732                                            bytenr, num_bytes, parent,
1733                                            root_objectid, owner, offset, 0);
1734         if (ret != -ENOENT)
1735                 return ret;
1736
1737         btrfs_release_path(path);
1738         *ref_ret = NULL;
1739
1740         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1741                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1742                                             root_objectid);
1743         } else {
1744                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1745                                              root_objectid, owner, offset);
1746         }
1747         return ret;
1748 }
1749
1750 /*
1751  * helper to update/remove inline back ref
1752  */
1753 static noinline_for_stack
1754 void update_inline_extent_backref(struct btrfs_root *root,
1755                                   struct btrfs_path *path,
1756                                   struct btrfs_extent_inline_ref *iref,
1757                                   int refs_to_mod,
1758                                   struct btrfs_delayed_extent_op *extent_op,
1759                                   int *last_ref)
1760 {
1761         struct extent_buffer *leaf;
1762         struct btrfs_extent_item *ei;
1763         struct btrfs_extent_data_ref *dref = NULL;
1764         struct btrfs_shared_data_ref *sref = NULL;
1765         unsigned long ptr;
1766         unsigned long end;
1767         u32 item_size;
1768         int size;
1769         int type;
1770         u64 refs;
1771
1772         leaf = path->nodes[0];
1773         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1774         refs = btrfs_extent_refs(leaf, ei);
1775         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1776         refs += refs_to_mod;
1777         btrfs_set_extent_refs(leaf, ei, refs);
1778         if (extent_op)
1779                 __run_delayed_extent_op(extent_op, leaf, ei);
1780
1781         type = btrfs_extent_inline_ref_type(leaf, iref);
1782
1783         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1784                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1785                 refs = btrfs_extent_data_ref_count(leaf, dref);
1786         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1787                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1788                 refs = btrfs_shared_data_ref_count(leaf, sref);
1789         } else {
1790                 refs = 1;
1791                 BUG_ON(refs_to_mod != -1);
1792         }
1793
1794         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1795         refs += refs_to_mod;
1796
1797         if (refs > 0) {
1798                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1799                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1800                 else
1801                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1802         } else {
1803                 *last_ref = 1;
1804                 size =  btrfs_extent_inline_ref_size(type);
1805                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1806                 ptr = (unsigned long)iref;
1807                 end = (unsigned long)ei + item_size;
1808                 if (ptr + size < end)
1809                         memmove_extent_buffer(leaf, ptr, ptr + size,
1810                                               end - ptr - size);
1811                 item_size -= size;
1812                 btrfs_truncate_item(root, path, item_size, 1);
1813         }
1814         btrfs_mark_buffer_dirty(leaf);
1815 }
1816
1817 static noinline_for_stack
1818 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1819                                  struct btrfs_root *root,
1820                                  struct btrfs_path *path,
1821                                  u64 bytenr, u64 num_bytes, u64 parent,
1822                                  u64 root_objectid, u64 owner,
1823                                  u64 offset, int refs_to_add,
1824                                  struct btrfs_delayed_extent_op *extent_op)
1825 {
1826         struct btrfs_extent_inline_ref *iref;
1827         int ret;
1828
1829         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1830                                            bytenr, num_bytes, parent,
1831                                            root_objectid, owner, offset, 1);
1832         if (ret == 0) {
1833                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1834                 update_inline_extent_backref(root, path, iref,
1835                                              refs_to_add, extent_op, NULL);
1836         } else if (ret == -ENOENT) {
1837                 setup_inline_extent_backref(root, path, iref, parent,
1838                                             root_objectid, owner, offset,
1839                                             refs_to_add, extent_op);
1840                 ret = 0;
1841         }
1842         return ret;
1843 }
1844
1845 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1846                                  struct btrfs_root *root,
1847                                  struct btrfs_path *path,
1848                                  u64 bytenr, u64 parent, u64 root_objectid,
1849                                  u64 owner, u64 offset, int refs_to_add)
1850 {
1851         int ret;
1852         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1853                 BUG_ON(refs_to_add != 1);
1854                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1855                                             parent, root_objectid);
1856         } else {
1857                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1858                                              parent, root_objectid,
1859                                              owner, offset, refs_to_add);
1860         }
1861         return ret;
1862 }
1863
1864 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1865                                  struct btrfs_root *root,
1866                                  struct btrfs_path *path,
1867                                  struct btrfs_extent_inline_ref *iref,
1868                                  int refs_to_drop, int is_data, int *last_ref)
1869 {
1870         int ret = 0;
1871
1872         BUG_ON(!is_data && refs_to_drop != 1);
1873         if (iref) {
1874                 update_inline_extent_backref(root, path, iref,
1875                                              -refs_to_drop, NULL, last_ref);
1876         } else if (is_data) {
1877                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop,
1878                                              last_ref);
1879         } else {
1880                 *last_ref = 1;
1881                 ret = btrfs_del_item(trans, root, path);
1882         }
1883         return ret;
1884 }
1885
1886 static int btrfs_issue_discard(struct block_device *bdev,
1887                                 u64 start, u64 len)
1888 {
1889         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1890 }
1891
1892 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1893                                 u64 num_bytes, u64 *actual_bytes)
1894 {
1895         int ret;
1896         u64 discarded_bytes = 0;
1897         struct btrfs_bio *bbio = NULL;
1898
1899
1900         /* Tell the block device(s) that the sectors can be discarded */
1901         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1902                               bytenr, &num_bytes, &bbio, 0);
1903         /* Error condition is -ENOMEM */
1904         if (!ret) {
1905                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1906                 int i;
1907
1908
1909                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1910                         if (!stripe->dev->can_discard)
1911                                 continue;
1912
1913                         ret = btrfs_issue_discard(stripe->dev->bdev,
1914                                                   stripe->physical,
1915                                                   stripe->length);
1916                         if (!ret)
1917                                 discarded_bytes += stripe->length;
1918                         else if (ret != -EOPNOTSUPP)
1919                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1920
1921                         /*
1922                          * Just in case we get back EOPNOTSUPP for some reason,
1923                          * just ignore the return value so we don't screw up
1924                          * people calling discard_extent.
1925                          */
1926                         ret = 0;
1927                 }
1928                 kfree(bbio);
1929         }
1930
1931         if (actual_bytes)
1932                 *actual_bytes = discarded_bytes;
1933
1934
1935         if (ret == -EOPNOTSUPP)
1936                 ret = 0;
1937         return ret;
1938 }
1939
1940 /* Can return -ENOMEM */
1941 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1942                          struct btrfs_root *root,
1943                          u64 bytenr, u64 num_bytes, u64 parent,
1944                          u64 root_objectid, u64 owner, u64 offset,
1945                          int no_quota)
1946 {
1947         int ret;
1948         struct btrfs_fs_info *fs_info = root->fs_info;
1949
1950         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1951                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1952
1953         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1954                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1955                                         num_bytes,
1956                                         parent, root_objectid, (int)owner,
1957                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
1958         } else {
1959                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1960                                         num_bytes,
1961                                         parent, root_objectid, owner, offset,
1962                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
1963         }
1964         return ret;
1965 }
1966
1967 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1968                                   struct btrfs_root *root,
1969                                   u64 bytenr, u64 num_bytes,
1970                                   u64 parent, u64 root_objectid,
1971                                   u64 owner, u64 offset, int refs_to_add,
1972                                   int no_quota,
1973                                   struct btrfs_delayed_extent_op *extent_op)
1974 {
1975         struct btrfs_fs_info *fs_info = root->fs_info;
1976         struct btrfs_path *path;
1977         struct extent_buffer *leaf;
1978         struct btrfs_extent_item *item;
1979         struct btrfs_key key;
1980         u64 refs;
1981         int ret;
1982         enum btrfs_qgroup_operation_type type = BTRFS_QGROUP_OPER_ADD_EXCL;
1983
1984         path = btrfs_alloc_path();
1985         if (!path)
1986                 return -ENOMEM;
1987
1988         if (!is_fstree(root_objectid) || !root->fs_info->quota_enabled)
1989                 no_quota = 1;
1990
1991         path->reada = 1;
1992         path->leave_spinning = 1;
1993         /* this will setup the path even if it fails to insert the back ref */
1994         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
1995                                            bytenr, num_bytes, parent,
1996                                            root_objectid, owner, offset,
1997                                            refs_to_add, extent_op);
1998         if ((ret < 0 && ret != -EAGAIN) || (!ret && no_quota))
1999                 goto out;
2000         /*
2001          * Ok we were able to insert an inline extent and it appears to be a new
2002          * reference, deal with the qgroup accounting.
2003          */
2004         if (!ret && !no_quota) {
2005                 ASSERT(root->fs_info->quota_enabled);
2006                 leaf = path->nodes[0];
2007                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2008                 item = btrfs_item_ptr(leaf, path->slots[0],
2009                                       struct btrfs_extent_item);
2010                 if (btrfs_extent_refs(leaf, item) > (u64)refs_to_add)
2011                         type = BTRFS_QGROUP_OPER_ADD_SHARED;
2012                 btrfs_release_path(path);
2013
2014                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
2015                                               bytenr, num_bytes, type, 0);
2016                 goto out;
2017         }
2018
2019         /*
2020          * Ok we had -EAGAIN which means we didn't have space to insert and
2021          * inline extent ref, so just update the reference count and add a
2022          * normal backref.
2023          */
2024         leaf = path->nodes[0];
2025         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2026         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2027         refs = btrfs_extent_refs(leaf, item);
2028         if (refs)
2029                 type = BTRFS_QGROUP_OPER_ADD_SHARED;
2030         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2031         if (extent_op)
2032                 __run_delayed_extent_op(extent_op, leaf, item);
2033
2034         btrfs_mark_buffer_dirty(leaf);
2035         btrfs_release_path(path);
2036
2037         if (!no_quota) {
2038                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
2039                                               bytenr, num_bytes, type, 0);
2040                 if (ret)
2041                         goto out;
2042         }
2043
2044         path->reada = 1;
2045         path->leave_spinning = 1;
2046         /* now insert the actual backref */
2047         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2048                                     path, bytenr, parent, root_objectid,
2049                                     owner, offset, refs_to_add);
2050         if (ret)
2051                 btrfs_abort_transaction(trans, root, ret);
2052 out:
2053         btrfs_free_path(path);
2054         return ret;
2055 }
2056
2057 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2058                                 struct btrfs_root *root,
2059                                 struct btrfs_delayed_ref_node *node,
2060                                 struct btrfs_delayed_extent_op *extent_op,
2061                                 int insert_reserved)
2062 {
2063         int ret = 0;
2064         struct btrfs_delayed_data_ref *ref;
2065         struct btrfs_key ins;
2066         u64 parent = 0;
2067         u64 ref_root = 0;
2068         u64 flags = 0;
2069
2070         ins.objectid = node->bytenr;
2071         ins.offset = node->num_bytes;
2072         ins.type = BTRFS_EXTENT_ITEM_KEY;
2073
2074         ref = btrfs_delayed_node_to_data_ref(node);
2075         trace_run_delayed_data_ref(node, ref, node->action);
2076
2077         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2078                 parent = ref->parent;
2079         ref_root = ref->root;
2080
2081         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2082                 if (extent_op)
2083                         flags |= extent_op->flags_to_set;
2084                 ret = alloc_reserved_file_extent(trans, root,
2085                                                  parent, ref_root, flags,
2086                                                  ref->objectid, ref->offset,
2087                                                  &ins, node->ref_mod);
2088         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2089                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2090                                              node->num_bytes, parent,
2091                                              ref_root, ref->objectid,
2092                                              ref->offset, node->ref_mod,
2093                                              node->no_quota, extent_op);
2094         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2095                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2096                                           node->num_bytes, parent,
2097                                           ref_root, ref->objectid,
2098                                           ref->offset, node->ref_mod,
2099                                           extent_op, node->no_quota);
2100         } else {
2101                 BUG();
2102         }
2103         return ret;
2104 }
2105
2106 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2107                                     struct extent_buffer *leaf,
2108                                     struct btrfs_extent_item *ei)
2109 {
2110         u64 flags = btrfs_extent_flags(leaf, ei);
2111         if (extent_op->update_flags) {
2112                 flags |= extent_op->flags_to_set;
2113                 btrfs_set_extent_flags(leaf, ei, flags);
2114         }
2115
2116         if (extent_op->update_key) {
2117                 struct btrfs_tree_block_info *bi;
2118                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2119                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2120                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2121         }
2122 }
2123
2124 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2125                                  struct btrfs_root *root,
2126                                  struct btrfs_delayed_ref_node *node,
2127                                  struct btrfs_delayed_extent_op *extent_op)
2128 {
2129         struct btrfs_key key;
2130         struct btrfs_path *path;
2131         struct btrfs_extent_item *ei;
2132         struct extent_buffer *leaf;
2133         u32 item_size;
2134         int ret;
2135         int err = 0;
2136         int metadata = !extent_op->is_data;
2137
2138         if (trans->aborted)
2139                 return 0;
2140
2141         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2142                 metadata = 0;
2143
2144         path = btrfs_alloc_path();
2145         if (!path)
2146                 return -ENOMEM;
2147
2148         key.objectid = node->bytenr;
2149
2150         if (metadata) {
2151                 key.type = BTRFS_METADATA_ITEM_KEY;
2152                 key.offset = extent_op->level;
2153         } else {
2154                 key.type = BTRFS_EXTENT_ITEM_KEY;
2155                 key.offset = node->num_bytes;
2156         }
2157
2158 again:
2159         path->reada = 1;
2160         path->leave_spinning = 1;
2161         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2162                                 path, 0, 1);
2163         if (ret < 0) {
2164                 err = ret;
2165                 goto out;
2166         }
2167         if (ret > 0) {
2168                 if (metadata) {
2169                         if (path->slots[0] > 0) {
2170                                 path->slots[0]--;
2171                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2172                                                       path->slots[0]);
2173                                 if (key.objectid == node->bytenr &&
2174                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2175                                     key.offset == node->num_bytes)
2176                                         ret = 0;
2177                         }
2178                         if (ret > 0) {
2179                                 btrfs_release_path(path);
2180                                 metadata = 0;
2181
2182                                 key.objectid = node->bytenr;
2183                                 key.offset = node->num_bytes;
2184                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2185                                 goto again;
2186                         }
2187                 } else {
2188                         err = -EIO;
2189                         goto out;
2190                 }
2191         }
2192
2193         leaf = path->nodes[0];
2194         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2195 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2196         if (item_size < sizeof(*ei)) {
2197                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2198                                              path, (u64)-1, 0);
2199                 if (ret < 0) {
2200                         err = ret;
2201                         goto out;
2202                 }
2203                 leaf = path->nodes[0];
2204                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2205         }
2206 #endif
2207         BUG_ON(item_size < sizeof(*ei));
2208         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2209         __run_delayed_extent_op(extent_op, leaf, ei);
2210
2211         btrfs_mark_buffer_dirty(leaf);
2212 out:
2213         btrfs_free_path(path);
2214         return err;
2215 }
2216
2217 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2218                                 struct btrfs_root *root,
2219                                 struct btrfs_delayed_ref_node *node,
2220                                 struct btrfs_delayed_extent_op *extent_op,
2221                                 int insert_reserved)
2222 {
2223         int ret = 0;
2224         struct btrfs_delayed_tree_ref *ref;
2225         struct btrfs_key ins;
2226         u64 parent = 0;
2227         u64 ref_root = 0;
2228         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2229                                                  SKINNY_METADATA);
2230
2231         ref = btrfs_delayed_node_to_tree_ref(node);
2232         trace_run_delayed_tree_ref(node, ref, node->action);
2233
2234         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2235                 parent = ref->parent;
2236         ref_root = ref->root;
2237
2238         ins.objectid = node->bytenr;
2239         if (skinny_metadata) {
2240                 ins.offset = ref->level;
2241                 ins.type = BTRFS_METADATA_ITEM_KEY;
2242         } else {
2243                 ins.offset = node->num_bytes;
2244                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2245         }
2246
2247         BUG_ON(node->ref_mod != 1);
2248         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2249                 BUG_ON(!extent_op || !extent_op->update_flags);
2250                 ret = alloc_reserved_tree_block(trans, root,
2251                                                 parent, ref_root,
2252                                                 extent_op->flags_to_set,
2253                                                 &extent_op->key,
2254                                                 ref->level, &ins,
2255                                                 node->no_quota);
2256         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2257                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2258                                              node->num_bytes, parent, ref_root,
2259                                              ref->level, 0, 1, node->no_quota,
2260                                              extent_op);
2261         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2262                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2263                                           node->num_bytes, parent, ref_root,
2264                                           ref->level, 0, 1, extent_op,
2265                                           node->no_quota);
2266         } else {
2267                 BUG();
2268         }
2269         return ret;
2270 }
2271
2272 /* helper function to actually process a single delayed ref entry */
2273 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2274                                struct btrfs_root *root,
2275                                struct btrfs_delayed_ref_node *node,
2276                                struct btrfs_delayed_extent_op *extent_op,
2277                                int insert_reserved)
2278 {
2279         int ret = 0;
2280
2281         if (trans->aborted) {
2282                 if (insert_reserved)
2283                         btrfs_pin_extent(root, node->bytenr,
2284                                          node->num_bytes, 1);
2285                 return 0;
2286         }
2287
2288         if (btrfs_delayed_ref_is_head(node)) {
2289                 struct btrfs_delayed_ref_head *head;
2290                 /*
2291                  * we've hit the end of the chain and we were supposed
2292                  * to insert this extent into the tree.  But, it got
2293                  * deleted before we ever needed to insert it, so all
2294                  * we have to do is clean up the accounting
2295                  */
2296                 BUG_ON(extent_op);
2297                 head = btrfs_delayed_node_to_head(node);
2298                 trace_run_delayed_ref_head(node, head, node->action);
2299
2300                 if (insert_reserved) {
2301                         btrfs_pin_extent(root, node->bytenr,
2302                                          node->num_bytes, 1);
2303                         if (head->is_data) {
2304                                 ret = btrfs_del_csums(trans, root,
2305                                                       node->bytenr,
2306                                                       node->num_bytes);
2307                         }
2308                 }
2309                 return ret;
2310         }
2311
2312         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2313             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2314                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2315                                            insert_reserved);
2316         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2317                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2318                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2319                                            insert_reserved);
2320         else
2321                 BUG();
2322         return ret;
2323 }
2324
2325 static noinline struct btrfs_delayed_ref_node *
2326 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2327 {
2328         struct rb_node *node;
2329         struct btrfs_delayed_ref_node *ref, *last = NULL;;
2330
2331         /*
2332          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2333          * this prevents ref count from going down to zero when
2334          * there still are pending delayed ref.
2335          */
2336         node = rb_first(&head->ref_root);
2337         while (node) {
2338                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2339                                 rb_node);
2340                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2341                         return ref;
2342                 else if (last == NULL)
2343                         last = ref;
2344                 node = rb_next(node);
2345         }
2346         return last;
2347 }
2348
2349 /*
2350  * Returns 0 on success or if called with an already aborted transaction.
2351  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2352  */
2353 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2354                                              struct btrfs_root *root,
2355                                              unsigned long nr)
2356 {
2357         struct btrfs_delayed_ref_root *delayed_refs;
2358         struct btrfs_delayed_ref_node *ref;
2359         struct btrfs_delayed_ref_head *locked_ref = NULL;
2360         struct btrfs_delayed_extent_op *extent_op;
2361         struct btrfs_fs_info *fs_info = root->fs_info;
2362         ktime_t start = ktime_get();
2363         int ret;
2364         unsigned long count = 0;
2365         unsigned long actual_count = 0;
2366         int must_insert_reserved = 0;
2367
2368         delayed_refs = &trans->transaction->delayed_refs;
2369         while (1) {
2370                 if (!locked_ref) {
2371                         if (count >= nr)
2372                                 break;
2373
2374                         spin_lock(&delayed_refs->lock);
2375                         locked_ref = btrfs_select_ref_head(trans);
2376                         if (!locked_ref) {
2377                                 spin_unlock(&delayed_refs->lock);
2378                                 break;
2379                         }
2380
2381                         /* grab the lock that says we are going to process
2382                          * all the refs for this head */
2383                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2384                         spin_unlock(&delayed_refs->lock);
2385                         /*
2386                          * we may have dropped the spin lock to get the head
2387                          * mutex lock, and that might have given someone else
2388                          * time to free the head.  If that's true, it has been
2389                          * removed from our list and we can move on.
2390                          */
2391                         if (ret == -EAGAIN) {
2392                                 locked_ref = NULL;
2393                                 count++;
2394                                 continue;
2395                         }
2396                 }
2397
2398                 /*
2399                  * We need to try and merge add/drops of the same ref since we
2400                  * can run into issues with relocate dropping the implicit ref
2401                  * and then it being added back again before the drop can
2402                  * finish.  If we merged anything we need to re-loop so we can
2403                  * get a good ref.
2404                  */
2405                 spin_lock(&locked_ref->lock);
2406                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2407                                          locked_ref);
2408
2409                 /*
2410                  * locked_ref is the head node, so we have to go one
2411                  * node back for any delayed ref updates
2412                  */
2413                 ref = select_delayed_ref(locked_ref);
2414
2415                 if (ref && ref->seq &&
2416                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2417                         spin_unlock(&locked_ref->lock);
2418                         btrfs_delayed_ref_unlock(locked_ref);
2419                         spin_lock(&delayed_refs->lock);
2420                         locked_ref->processing = 0;
2421                         delayed_refs->num_heads_ready++;
2422                         spin_unlock(&delayed_refs->lock);
2423                         locked_ref = NULL;
2424                         cond_resched();
2425                         count++;
2426                         continue;
2427                 }
2428
2429                 /*
2430                  * record the must insert reserved flag before we
2431                  * drop the spin lock.
2432                  */
2433                 must_insert_reserved = locked_ref->must_insert_reserved;
2434                 locked_ref->must_insert_reserved = 0;
2435
2436                 extent_op = locked_ref->extent_op;
2437                 locked_ref->extent_op = NULL;
2438
2439                 if (!ref) {
2440
2441
2442                         /* All delayed refs have been processed, Go ahead
2443                          * and send the head node to run_one_delayed_ref,
2444                          * so that any accounting fixes can happen
2445                          */
2446                         ref = &locked_ref->node;
2447
2448                         if (extent_op && must_insert_reserved) {
2449                                 btrfs_free_delayed_extent_op(extent_op);
2450                                 extent_op = NULL;
2451                         }
2452
2453                         if (extent_op) {
2454                                 spin_unlock(&locked_ref->lock);
2455                                 ret = run_delayed_extent_op(trans, root,
2456                                                             ref, extent_op);
2457                                 btrfs_free_delayed_extent_op(extent_op);
2458
2459                                 if (ret) {
2460                                         /*
2461                                          * Need to reset must_insert_reserved if
2462                                          * there was an error so the abort stuff
2463                                          * can cleanup the reserved space
2464                                          * properly.
2465                                          */
2466                                         if (must_insert_reserved)
2467                                                 locked_ref->must_insert_reserved = 1;
2468                                         locked_ref->processing = 0;
2469                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2470                                         btrfs_delayed_ref_unlock(locked_ref);
2471                                         return ret;
2472                                 }
2473                                 continue;
2474                         }
2475
2476                         /*
2477                          * Need to drop our head ref lock and re-aqcuire the
2478                          * delayed ref lock and then re-check to make sure
2479                          * nobody got added.
2480                          */
2481                         spin_unlock(&locked_ref->lock);
2482                         spin_lock(&delayed_refs->lock);
2483                         spin_lock(&locked_ref->lock);
2484                         if (rb_first(&locked_ref->ref_root) ||
2485                             locked_ref->extent_op) {
2486                                 spin_unlock(&locked_ref->lock);
2487                                 spin_unlock(&delayed_refs->lock);
2488                                 continue;
2489                         }
2490                         ref->in_tree = 0;
2491                         delayed_refs->num_heads--;
2492                         rb_erase(&locked_ref->href_node,
2493                                  &delayed_refs->href_root);
2494                         spin_unlock(&delayed_refs->lock);
2495                 } else {
2496                         actual_count++;
2497                         ref->in_tree = 0;
2498                         rb_erase(&ref->rb_node, &locked_ref->ref_root);
2499                 }
2500                 atomic_dec(&delayed_refs->num_entries);
2501
2502                 if (!btrfs_delayed_ref_is_head(ref)) {
2503                         /*
2504                          * when we play the delayed ref, also correct the
2505                          * ref_mod on head
2506                          */
2507                         switch (ref->action) {
2508                         case BTRFS_ADD_DELAYED_REF:
2509                         case BTRFS_ADD_DELAYED_EXTENT:
2510                                 locked_ref->node.ref_mod -= ref->ref_mod;
2511                                 break;
2512                         case BTRFS_DROP_DELAYED_REF:
2513                                 locked_ref->node.ref_mod += ref->ref_mod;
2514                                 break;
2515                         default:
2516                                 WARN_ON(1);
2517                         }
2518                 }
2519                 spin_unlock(&locked_ref->lock);
2520
2521                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2522                                           must_insert_reserved);
2523
2524                 btrfs_free_delayed_extent_op(extent_op);
2525                 if (ret) {
2526                         locked_ref->processing = 0;
2527                         btrfs_delayed_ref_unlock(locked_ref);
2528                         btrfs_put_delayed_ref(ref);
2529                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2530                         return ret;
2531                 }
2532
2533                 /*
2534                  * If this node is a head, that means all the refs in this head
2535                  * have been dealt with, and we will pick the next head to deal
2536                  * with, so we must unlock the head and drop it from the cluster
2537                  * list before we release it.
2538                  */
2539                 if (btrfs_delayed_ref_is_head(ref)) {
2540                         btrfs_delayed_ref_unlock(locked_ref);
2541                         locked_ref = NULL;
2542                 }
2543                 btrfs_put_delayed_ref(ref);
2544                 count++;
2545                 cond_resched();
2546         }
2547
2548         /*
2549          * We don't want to include ref heads since we can have empty ref heads
2550          * and those will drastically skew our runtime down since we just do
2551          * accounting, no actual extent tree updates.
2552          */
2553         if (actual_count > 0) {
2554                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2555                 u64 avg;
2556
2557                 /*
2558                  * We weigh the current average higher than our current runtime
2559                  * to avoid large swings in the average.
2560                  */
2561                 spin_lock(&delayed_refs->lock);
2562                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2563                 avg = div64_u64(avg, 4);
2564                 fs_info->avg_delayed_ref_runtime = avg;
2565                 spin_unlock(&delayed_refs->lock);
2566         }
2567         return 0;
2568 }
2569
2570 #ifdef SCRAMBLE_DELAYED_REFS
2571 /*
2572  * Normally delayed refs get processed in ascending bytenr order. This
2573  * correlates in most cases to the order added. To expose dependencies on this
2574  * order, we start to process the tree in the middle instead of the beginning
2575  */
2576 static u64 find_middle(struct rb_root *root)
2577 {
2578         struct rb_node *n = root->rb_node;
2579         struct btrfs_delayed_ref_node *entry;
2580         int alt = 1;
2581         u64 middle;
2582         u64 first = 0, last = 0;
2583
2584         n = rb_first(root);
2585         if (n) {
2586                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2587                 first = entry->bytenr;
2588         }
2589         n = rb_last(root);
2590         if (n) {
2591                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2592                 last = entry->bytenr;
2593         }
2594         n = root->rb_node;
2595
2596         while (n) {
2597                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2598                 WARN_ON(!entry->in_tree);
2599
2600                 middle = entry->bytenr;
2601
2602                 if (alt)
2603                         n = n->rb_left;
2604                 else
2605                         n = n->rb_right;
2606
2607                 alt = 1 - alt;
2608         }
2609         return middle;
2610 }
2611 #endif
2612
2613 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2614 {
2615         u64 num_bytes;
2616
2617         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2618                              sizeof(struct btrfs_extent_inline_ref));
2619         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2620                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2621
2622         /*
2623          * We don't ever fill up leaves all the way so multiply by 2 just to be
2624          * closer to what we're really going to want to ouse.
2625          */
2626         return div64_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2627 }
2628
2629 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2630                                        struct btrfs_root *root)
2631 {
2632         struct btrfs_block_rsv *global_rsv;
2633         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2634         u64 num_bytes;
2635         int ret = 0;
2636
2637         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2638         num_heads = heads_to_leaves(root, num_heads);
2639         if (num_heads > 1)
2640                 num_bytes += (num_heads - 1) * root->nodesize;
2641         num_bytes <<= 1;
2642         global_rsv = &root->fs_info->global_block_rsv;
2643
2644         /*
2645          * If we can't allocate any more chunks lets make sure we have _lots_ of
2646          * wiggle room since running delayed refs can create more delayed refs.
2647          */
2648         if (global_rsv->space_info->full)
2649                 num_bytes <<= 1;
2650
2651         spin_lock(&global_rsv->lock);
2652         if (global_rsv->reserved <= num_bytes)
2653                 ret = 1;
2654         spin_unlock(&global_rsv->lock);
2655         return ret;
2656 }
2657
2658 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2659                                        struct btrfs_root *root)
2660 {
2661         struct btrfs_fs_info *fs_info = root->fs_info;
2662         u64 num_entries =
2663                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2664         u64 avg_runtime;
2665         u64 val;
2666
2667         smp_mb();
2668         avg_runtime = fs_info->avg_delayed_ref_runtime;
2669         val = num_entries * avg_runtime;
2670         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2671                 return 1;
2672         if (val >= NSEC_PER_SEC / 2)
2673                 return 2;
2674
2675         return btrfs_check_space_for_delayed_refs(trans, root);
2676 }
2677
2678 struct async_delayed_refs {
2679         struct btrfs_root *root;
2680         int count;
2681         int error;
2682         int sync;
2683         struct completion wait;
2684         struct btrfs_work work;
2685 };
2686
2687 static void delayed_ref_async_start(struct btrfs_work *work)
2688 {
2689         struct async_delayed_refs *async;
2690         struct btrfs_trans_handle *trans;
2691         int ret;
2692
2693         async = container_of(work, struct async_delayed_refs, work);
2694
2695         trans = btrfs_join_transaction(async->root);
2696         if (IS_ERR(trans)) {
2697                 async->error = PTR_ERR(trans);
2698                 goto done;
2699         }
2700
2701         /*
2702          * trans->sync means that when we call end_transaciton, we won't
2703          * wait on delayed refs
2704          */
2705         trans->sync = true;
2706         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2707         if (ret)
2708                 async->error = ret;
2709
2710         ret = btrfs_end_transaction(trans, async->root);
2711         if (ret && !async->error)
2712                 async->error = ret;
2713 done:
2714         if (async->sync)
2715                 complete(&async->wait);
2716         else
2717                 kfree(async);
2718 }
2719
2720 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2721                                  unsigned long count, int wait)
2722 {
2723         struct async_delayed_refs *async;
2724         int ret;
2725
2726         async = kmalloc(sizeof(*async), GFP_NOFS);
2727         if (!async)
2728                 return -ENOMEM;
2729
2730         async->root = root->fs_info->tree_root;
2731         async->count = count;
2732         async->error = 0;
2733         if (wait)
2734                 async->sync = 1;
2735         else
2736                 async->sync = 0;
2737         init_completion(&async->wait);
2738
2739         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2740                         delayed_ref_async_start, NULL, NULL);
2741
2742         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2743
2744         if (wait) {
2745                 wait_for_completion(&async->wait);
2746                 ret = async->error;
2747                 kfree(async);
2748                 return ret;
2749         }
2750         return 0;
2751 }
2752
2753 /*
2754  * this starts processing the delayed reference count updates and
2755  * extent insertions we have queued up so far.  count can be
2756  * 0, which means to process everything in the tree at the start
2757  * of the run (but not newly added entries), or it can be some target
2758  * number you'd like to process.
2759  *
2760  * Returns 0 on success or if called with an aborted transaction
2761  * Returns <0 on error and aborts the transaction
2762  */
2763 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2764                            struct btrfs_root *root, unsigned long count)
2765 {
2766         struct rb_node *node;
2767         struct btrfs_delayed_ref_root *delayed_refs;
2768         struct btrfs_delayed_ref_head *head;
2769         int ret;
2770         int run_all = count == (unsigned long)-1;
2771         int run_most = 0;
2772
2773         /* We'll clean this up in btrfs_cleanup_transaction */
2774         if (trans->aborted)
2775                 return 0;
2776
2777         if (root == root->fs_info->extent_root)
2778                 root = root->fs_info->tree_root;
2779
2780         delayed_refs = &trans->transaction->delayed_refs;
2781         if (count == 0) {
2782                 count = atomic_read(&delayed_refs->num_entries) * 2;
2783                 run_most = 1;
2784         }
2785
2786 again:
2787 #ifdef SCRAMBLE_DELAYED_REFS
2788         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2789 #endif
2790         ret = __btrfs_run_delayed_refs(trans, root, count);
2791         if (ret < 0) {
2792                 btrfs_abort_transaction(trans, root, ret);
2793                 return ret;
2794         }
2795
2796         if (run_all) {
2797                 if (!list_empty(&trans->new_bgs))
2798                         btrfs_create_pending_block_groups(trans, root);
2799
2800                 spin_lock(&delayed_refs->lock);
2801                 node = rb_first(&delayed_refs->href_root);
2802                 if (!node) {
2803                         spin_unlock(&delayed_refs->lock);
2804                         goto out;
2805                 }
2806                 count = (unsigned long)-1;
2807
2808                 while (node) {
2809                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2810                                         href_node);
2811                         if (btrfs_delayed_ref_is_head(&head->node)) {
2812                                 struct btrfs_delayed_ref_node *ref;
2813
2814                                 ref = &head->node;
2815                                 atomic_inc(&ref->refs);
2816
2817                                 spin_unlock(&delayed_refs->lock);
2818                                 /*
2819                                  * Mutex was contended, block until it's
2820                                  * released and try again
2821                                  */
2822                                 mutex_lock(&head->mutex);
2823                                 mutex_unlock(&head->mutex);
2824
2825                                 btrfs_put_delayed_ref(ref);
2826                                 cond_resched();
2827                                 goto again;
2828                         } else {
2829                                 WARN_ON(1);
2830                         }
2831                         node = rb_next(node);
2832                 }
2833                 spin_unlock(&delayed_refs->lock);
2834                 cond_resched();
2835                 goto again;
2836         }
2837 out:
2838         ret = btrfs_delayed_qgroup_accounting(trans, root->fs_info);
2839         if (ret)
2840                 return ret;
2841         assert_qgroups_uptodate(trans);
2842         return 0;
2843 }
2844
2845 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2846                                 struct btrfs_root *root,
2847                                 u64 bytenr, u64 num_bytes, u64 flags,
2848                                 int level, int is_data)
2849 {
2850         struct btrfs_delayed_extent_op *extent_op;
2851         int ret;
2852
2853         extent_op = btrfs_alloc_delayed_extent_op();
2854         if (!extent_op)
2855                 return -ENOMEM;
2856
2857         extent_op->flags_to_set = flags;
2858         extent_op->update_flags = 1;
2859         extent_op->update_key = 0;
2860         extent_op->is_data = is_data ? 1 : 0;
2861         extent_op->level = level;
2862
2863         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2864                                           num_bytes, extent_op);
2865         if (ret)
2866                 btrfs_free_delayed_extent_op(extent_op);
2867         return ret;
2868 }
2869
2870 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2871                                       struct btrfs_root *root,
2872                                       struct btrfs_path *path,
2873                                       u64 objectid, u64 offset, u64 bytenr)
2874 {
2875         struct btrfs_delayed_ref_head *head;
2876         struct btrfs_delayed_ref_node *ref;
2877         struct btrfs_delayed_data_ref *data_ref;
2878         struct btrfs_delayed_ref_root *delayed_refs;
2879         struct rb_node *node;
2880         int ret = 0;
2881
2882         delayed_refs = &trans->transaction->delayed_refs;
2883         spin_lock(&delayed_refs->lock);
2884         head = btrfs_find_delayed_ref_head(trans, bytenr);
2885         if (!head) {
2886                 spin_unlock(&delayed_refs->lock);
2887                 return 0;
2888         }
2889
2890         if (!mutex_trylock(&head->mutex)) {
2891                 atomic_inc(&head->node.refs);
2892                 spin_unlock(&delayed_refs->lock);
2893
2894                 btrfs_release_path(path);
2895
2896                 /*
2897                  * Mutex was contended, block until it's released and let
2898                  * caller try again
2899                  */
2900                 mutex_lock(&head->mutex);
2901                 mutex_unlock(&head->mutex);
2902                 btrfs_put_delayed_ref(&head->node);
2903                 return -EAGAIN;
2904         }
2905         spin_unlock(&delayed_refs->lock);
2906
2907         spin_lock(&head->lock);
2908         node = rb_first(&head->ref_root);
2909         while (node) {
2910                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2911                 node = rb_next(node);
2912
2913                 /* If it's a shared ref we know a cross reference exists */
2914                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2915                         ret = 1;
2916                         break;
2917                 }
2918
2919                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2920
2921                 /*
2922                  * If our ref doesn't match the one we're currently looking at
2923                  * then we have a cross reference.
2924                  */
2925                 if (data_ref->root != root->root_key.objectid ||
2926                     data_ref->objectid != objectid ||
2927                     data_ref->offset != offset) {
2928                         ret = 1;
2929                         break;
2930                 }
2931         }
2932         spin_unlock(&head->lock);
2933         mutex_unlock(&head->mutex);
2934         return ret;
2935 }
2936
2937 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2938                                         struct btrfs_root *root,
2939                                         struct btrfs_path *path,
2940                                         u64 objectid, u64 offset, u64 bytenr)
2941 {
2942         struct btrfs_root *extent_root = root->fs_info->extent_root;
2943         struct extent_buffer *leaf;
2944         struct btrfs_extent_data_ref *ref;
2945         struct btrfs_extent_inline_ref *iref;
2946         struct btrfs_extent_item *ei;
2947         struct btrfs_key key;
2948         u32 item_size;
2949         int ret;
2950
2951         key.objectid = bytenr;
2952         key.offset = (u64)-1;
2953         key.type = BTRFS_EXTENT_ITEM_KEY;
2954
2955         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2956         if (ret < 0)
2957                 goto out;
2958         BUG_ON(ret == 0); /* Corruption */
2959
2960         ret = -ENOENT;
2961         if (path->slots[0] == 0)
2962                 goto out;
2963
2964         path->slots[0]--;
2965         leaf = path->nodes[0];
2966         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2967
2968         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2969                 goto out;
2970
2971         ret = 1;
2972         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2973 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2974         if (item_size < sizeof(*ei)) {
2975                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2976                 goto out;
2977         }
2978 #endif
2979         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2980
2981         if (item_size != sizeof(*ei) +
2982             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2983                 goto out;
2984
2985         if (btrfs_extent_generation(leaf, ei) <=
2986             btrfs_root_last_snapshot(&root->root_item))
2987                 goto out;
2988
2989         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2990         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2991             BTRFS_EXTENT_DATA_REF_KEY)
2992                 goto out;
2993
2994         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2995         if (btrfs_extent_refs(leaf, ei) !=
2996             btrfs_extent_data_ref_count(leaf, ref) ||
2997             btrfs_extent_data_ref_root(leaf, ref) !=
2998             root->root_key.objectid ||
2999             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3000             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3001                 goto out;
3002
3003         ret = 0;
3004 out:
3005         return ret;
3006 }
3007
3008 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3009                           struct btrfs_root *root,
3010                           u64 objectid, u64 offset, u64 bytenr)
3011 {
3012         struct btrfs_path *path;
3013         int ret;
3014         int ret2;
3015
3016         path = btrfs_alloc_path();
3017         if (!path)
3018                 return -ENOENT;
3019
3020         do {
3021                 ret = check_committed_ref(trans, root, path, objectid,
3022                                           offset, bytenr);
3023                 if (ret && ret != -ENOENT)
3024                         goto out;
3025
3026                 ret2 = check_delayed_ref(trans, root, path, objectid,
3027                                          offset, bytenr);
3028         } while (ret2 == -EAGAIN);
3029
3030         if (ret2 && ret2 != -ENOENT) {
3031                 ret = ret2;
3032                 goto out;
3033         }
3034
3035         if (ret != -ENOENT || ret2 != -ENOENT)
3036                 ret = 0;
3037 out:
3038         btrfs_free_path(path);
3039         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3040                 WARN_ON(ret > 0);
3041         return ret;
3042 }
3043
3044 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3045                            struct btrfs_root *root,
3046                            struct extent_buffer *buf,
3047                            int full_backref, int inc)
3048 {
3049         u64 bytenr;
3050         u64 num_bytes;
3051         u64 parent;
3052         u64 ref_root;
3053         u32 nritems;
3054         struct btrfs_key key;
3055         struct btrfs_file_extent_item *fi;
3056         int i;
3057         int level;
3058         int ret = 0;
3059         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3060                             u64, u64, u64, u64, u64, u64, int);
3061
3062
3063         if (btrfs_test_is_dummy_root(root))
3064                 return 0;
3065
3066         ref_root = btrfs_header_owner(buf);
3067         nritems = btrfs_header_nritems(buf);
3068         level = btrfs_header_level(buf);
3069
3070         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3071                 return 0;
3072
3073         if (inc)
3074                 process_func = btrfs_inc_extent_ref;
3075         else
3076                 process_func = btrfs_free_extent;
3077
3078         if (full_backref)
3079                 parent = buf->start;
3080         else
3081                 parent = 0;
3082
3083         for (i = 0; i < nritems; i++) {
3084                 if (level == 0) {
3085                         btrfs_item_key_to_cpu(buf, &key, i);
3086                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3087                                 continue;
3088                         fi = btrfs_item_ptr(buf, i,
3089                                             struct btrfs_file_extent_item);
3090                         if (btrfs_file_extent_type(buf, fi) ==
3091                             BTRFS_FILE_EXTENT_INLINE)
3092                                 continue;
3093                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3094                         if (bytenr == 0)
3095                                 continue;
3096
3097                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3098                         key.offset -= btrfs_file_extent_offset(buf, fi);
3099                         ret = process_func(trans, root, bytenr, num_bytes,
3100                                            parent, ref_root, key.objectid,
3101                                            key.offset, 1);
3102                         if (ret)
3103                                 goto fail;
3104                 } else {
3105                         bytenr = btrfs_node_blockptr(buf, i);
3106                         num_bytes = root->nodesize;
3107                         ret = process_func(trans, root, bytenr, num_bytes,
3108                                            parent, ref_root, level - 1, 0,
3109                                            1);
3110                         if (ret)
3111                                 goto fail;
3112                 }
3113         }
3114         return 0;
3115 fail:
3116         return ret;
3117 }
3118
3119 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3120                   struct extent_buffer *buf, int full_backref)
3121 {
3122         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3123 }
3124
3125 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3126                   struct extent_buffer *buf, int full_backref)
3127 {
3128         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3129 }
3130
3131 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3132                                  struct btrfs_root *root,
3133                                  struct btrfs_path *path,
3134                                  struct btrfs_block_group_cache *cache)
3135 {
3136         int ret;
3137         struct btrfs_root *extent_root = root->fs_info->extent_root;
3138         unsigned long bi;
3139         struct extent_buffer *leaf;
3140
3141         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3142         if (ret < 0)
3143                 goto fail;
3144         BUG_ON(ret); /* Corruption */
3145
3146         leaf = path->nodes[0];
3147         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3148         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3149         btrfs_mark_buffer_dirty(leaf);
3150         btrfs_release_path(path);
3151 fail:
3152         if (ret) {
3153                 btrfs_abort_transaction(trans, root, ret);
3154                 return ret;
3155         }
3156         return 0;
3157
3158 }
3159
3160 static struct btrfs_block_group_cache *
3161 next_block_group(struct btrfs_root *root,
3162                  struct btrfs_block_group_cache *cache)
3163 {
3164         struct rb_node *node;
3165         spin_lock(&root->fs_info->block_group_cache_lock);
3166         node = rb_next(&cache->cache_node);
3167         btrfs_put_block_group(cache);
3168         if (node) {
3169                 cache = rb_entry(node, struct btrfs_block_group_cache,
3170                                  cache_node);
3171                 btrfs_get_block_group(cache);
3172         } else
3173                 cache = NULL;
3174         spin_unlock(&root->fs_info->block_group_cache_lock);
3175         return cache;
3176 }
3177
3178 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3179                             struct btrfs_trans_handle *trans,
3180                             struct btrfs_path *path)
3181 {
3182         struct btrfs_root *root = block_group->fs_info->tree_root;
3183         struct inode *inode = NULL;
3184         u64 alloc_hint = 0;
3185         int dcs = BTRFS_DC_ERROR;
3186         int num_pages = 0;
3187         int retries = 0;
3188         int ret = 0;
3189
3190         /*
3191          * If this block group is smaller than 100 megs don't bother caching the
3192          * block group.
3193          */
3194         if (block_group->key.offset < (100 * 1024 * 1024)) {
3195                 spin_lock(&block_group->lock);
3196                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3197                 spin_unlock(&block_group->lock);
3198                 return 0;
3199         }
3200
3201 again:
3202         inode = lookup_free_space_inode(root, block_group, path);
3203         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3204                 ret = PTR_ERR(inode);
3205                 btrfs_release_path(path);
3206                 goto out;
3207         }
3208
3209         if (IS_ERR(inode)) {
3210                 BUG_ON(retries);
3211                 retries++;
3212
3213                 if (block_group->ro)
3214                         goto out_free;
3215
3216                 ret = create_free_space_inode(root, trans, block_group, path);
3217                 if (ret)
3218                         goto out_free;
3219                 goto again;
3220         }
3221
3222         /* We've already setup this transaction, go ahead and exit */
3223         if (block_group->cache_generation == trans->transid &&
3224             i_size_read(inode)) {
3225                 dcs = BTRFS_DC_SETUP;
3226                 goto out_put;
3227         }
3228
3229         /*
3230          * We want to set the generation to 0, that way if anything goes wrong
3231          * from here on out we know not to trust this cache when we load up next
3232          * time.
3233          */
3234         BTRFS_I(inode)->generation = 0;
3235         ret = btrfs_update_inode(trans, root, inode);
3236         WARN_ON(ret);
3237
3238         if (i_size_read(inode) > 0) {
3239                 ret = btrfs_check_trunc_cache_free_space(root,
3240                                         &root->fs_info->global_block_rsv);
3241                 if (ret)
3242                         goto out_put;
3243
3244                 ret = btrfs_truncate_free_space_cache(root, trans, inode);
3245                 if (ret)
3246                         goto out_put;
3247         }
3248
3249         spin_lock(&block_group->lock);
3250         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3251             !btrfs_test_opt(root, SPACE_CACHE) ||
3252             block_group->delalloc_bytes) {
3253                 /*
3254                  * don't bother trying to write stuff out _if_
3255                  * a) we're not cached,
3256                  * b) we're with nospace_cache mount option.
3257                  */
3258                 dcs = BTRFS_DC_WRITTEN;
3259                 spin_unlock(&block_group->lock);
3260                 goto out_put;
3261         }
3262         spin_unlock(&block_group->lock);
3263
3264         /*
3265          * Try to preallocate enough space based on how big the block group is.
3266          * Keep in mind this has to include any pinned space which could end up
3267          * taking up quite a bit since it's not folded into the other space
3268          * cache.
3269          */
3270         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3271         if (!num_pages)
3272                 num_pages = 1;
3273
3274         num_pages *= 16;
3275         num_pages *= PAGE_CACHE_SIZE;
3276
3277         ret = btrfs_check_data_free_space(inode, num_pages);
3278         if (ret)
3279                 goto out_put;
3280
3281         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3282                                               num_pages, num_pages,
3283                                               &alloc_hint);
3284         if (!ret)
3285                 dcs = BTRFS_DC_SETUP;
3286         btrfs_free_reserved_data_space(inode, num_pages);
3287
3288 out_put:
3289         iput(inode);
3290 out_free:
3291         btrfs_release_path(path);
3292 out:
3293         spin_lock(&block_group->lock);
3294         if (!ret && dcs == BTRFS_DC_SETUP)
3295                 block_group->cache_generation = trans->transid;
3296         block_group->disk_cache_state = dcs;
3297         spin_unlock(&block_group->lock);
3298
3299         return ret;
3300 }
3301
3302 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3303                                    struct btrfs_root *root)
3304 {
3305         struct btrfs_block_group_cache *cache;
3306         int err = 0;
3307         struct btrfs_path *path;
3308         u64 last = 0;
3309
3310         path = btrfs_alloc_path();
3311         if (!path)
3312                 return -ENOMEM;
3313
3314 again:
3315         while (1) {
3316                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3317                 while (cache) {
3318                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3319                                 break;
3320                         cache = next_block_group(root, cache);
3321                 }
3322                 if (!cache) {
3323                         if (last == 0)
3324                                 break;
3325                         last = 0;
3326                         continue;
3327                 }
3328                 err = cache_save_setup(cache, trans, path);
3329                 last = cache->key.objectid + cache->key.offset;
3330                 btrfs_put_block_group(cache);
3331         }
3332
3333         while (1) {
3334                 if (last == 0) {
3335                         err = btrfs_run_delayed_refs(trans, root,
3336                                                      (unsigned long)-1);
3337                         if (err) /* File system offline */
3338                                 goto out;
3339                 }
3340
3341                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3342                 while (cache) {
3343                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3344                                 btrfs_put_block_group(cache);
3345                                 goto again;
3346                         }
3347
3348                         if (cache->dirty)
3349                                 break;
3350                         cache = next_block_group(root, cache);
3351                 }
3352                 if (!cache) {
3353                         if (last == 0)
3354                                 break;
3355                         last = 0;
3356                         continue;
3357                 }
3358
3359                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3360                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3361                 cache->dirty = 0;
3362                 last = cache->key.objectid + cache->key.offset;
3363
3364                 err = write_one_cache_group(trans, root, path, cache);
3365                 btrfs_put_block_group(cache);
3366                 if (err) /* File system offline */
3367                         goto out;
3368         }
3369
3370         while (1) {
3371                 /*
3372                  * I don't think this is needed since we're just marking our
3373                  * preallocated extent as written, but just in case it can't
3374                  * hurt.
3375                  */
3376                 if (last == 0) {
3377                         err = btrfs_run_delayed_refs(trans, root,
3378                                                      (unsigned long)-1);
3379                         if (err) /* File system offline */
3380                                 goto out;
3381                 }
3382
3383                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3384                 while (cache) {
3385                         /*
3386                          * Really this shouldn't happen, but it could if we
3387                          * couldn't write the entire preallocated extent and
3388                          * splitting the extent resulted in a new block.
3389                          */
3390                         if (cache->dirty) {
3391                                 btrfs_put_block_group(cache);
3392                                 goto again;
3393                         }
3394                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3395                                 break;
3396                         cache = next_block_group(root, cache);
3397                 }
3398                 if (!cache) {
3399                         if (last == 0)
3400                                 break;
3401                         last = 0;
3402                         continue;
3403                 }
3404
3405                 err = btrfs_write_out_cache(root, trans, cache, path);
3406
3407                 /*
3408                  * If we didn't have an error then the cache state is still
3409                  * NEED_WRITE, so we can set it to WRITTEN.
3410                  */
3411                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3412                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3413                 last = cache->key.objectid + cache->key.offset;
3414                 btrfs_put_block_group(cache);
3415         }
3416 out:
3417
3418         btrfs_free_path(path);
3419         return err;
3420 }
3421
3422 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3423 {
3424         struct btrfs_block_group_cache *block_group;
3425         int readonly = 0;
3426
3427         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3428         if (!block_group || block_group->ro)
3429                 readonly = 1;
3430         if (block_group)
3431                 btrfs_put_block_group(block_group);
3432         return readonly;
3433 }
3434
3435 static const char *alloc_name(u64 flags)
3436 {
3437         switch (flags) {
3438         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3439                 return "mixed";
3440         case BTRFS_BLOCK_GROUP_METADATA:
3441                 return "metadata";
3442         case BTRFS_BLOCK_GROUP_DATA:
3443                 return "data";
3444         case BTRFS_BLOCK_GROUP_SYSTEM:
3445                 return "system";
3446         default:
3447                 WARN_ON(1);
3448                 return "invalid-combination";
3449         };
3450 }
3451
3452 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3453                              u64 total_bytes, u64 bytes_used,
3454                              struct btrfs_space_info **space_info)
3455 {
3456         struct btrfs_space_info *found;
3457         int i;
3458         int factor;
3459         int ret;
3460
3461         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3462                      BTRFS_BLOCK_GROUP_RAID10))
3463                 factor = 2;
3464         else
3465                 factor = 1;
3466
3467         found = __find_space_info(info, flags);
3468         if (found) {
3469                 spin_lock(&found->lock);
3470                 found->total_bytes += total_bytes;
3471                 found->disk_total += total_bytes * factor;
3472                 found->bytes_used += bytes_used;
3473                 found->disk_used += bytes_used * factor;
3474                 found->full = 0;
3475                 spin_unlock(&found->lock);
3476                 *space_info = found;
3477                 return 0;
3478         }
3479         found = kzalloc(sizeof(*found), GFP_NOFS);
3480         if (!found)
3481                 return -ENOMEM;
3482
3483         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3484         if (ret) {
3485                 kfree(found);
3486                 return ret;
3487         }
3488
3489         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3490                 INIT_LIST_HEAD(&found->block_groups[i]);
3491         init_rwsem(&found->groups_sem);
3492         spin_lock_init(&found->lock);
3493         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3494         found->total_bytes = total_bytes;
3495         found->disk_total = total_bytes * factor;
3496         found->bytes_used = bytes_used;
3497         found->disk_used = bytes_used * factor;
3498         found->bytes_pinned = 0;
3499         found->bytes_reserved = 0;
3500         found->bytes_readonly = 0;
3501         found->bytes_may_use = 0;
3502         found->full = 0;
3503         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3504         found->chunk_alloc = 0;
3505         found->flush = 0;
3506         init_waitqueue_head(&found->wait);
3507         INIT_LIST_HEAD(&found->ro_bgs);
3508
3509         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3510                                     info->space_info_kobj, "%s",
3511                                     alloc_name(found->flags));
3512         if (ret) {
3513                 kfree(found);
3514                 return ret;
3515         }
3516
3517         *space_info = found;
3518         list_add_rcu(&found->list, &info->space_info);
3519         if (flags & BTRFS_BLOCK_GROUP_DATA)
3520                 info->data_sinfo = found;
3521
3522         return ret;
3523 }
3524
3525 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3526 {
3527         u64 extra_flags = chunk_to_extended(flags) &
3528                                 BTRFS_EXTENDED_PROFILE_MASK;
3529
3530         write_seqlock(&fs_info->profiles_lock);
3531         if (flags & BTRFS_BLOCK_GROUP_DATA)
3532                 fs_info->avail_data_alloc_bits |= extra_flags;
3533         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3534                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3535         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3536                 fs_info->avail_system_alloc_bits |= extra_flags;
3537         write_sequnlock(&fs_info->profiles_lock);
3538 }
3539
3540 /*
3541  * returns target flags in extended format or 0 if restripe for this
3542  * chunk_type is not in progress
3543  *
3544  * should be called with either volume_mutex or balance_lock held
3545  */
3546 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3547 {
3548         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3549         u64 target = 0;
3550
3551         if (!bctl)
3552                 return 0;
3553
3554         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3555             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3556                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3557         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3558                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3559                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3560         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3561                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3562                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3563         }
3564
3565         return target;
3566 }
3567
3568 /*
3569  * @flags: available profiles in extended format (see ctree.h)
3570  *
3571  * Returns reduced profile in chunk format.  If profile changing is in
3572  * progress (either running or paused) picks the target profile (if it's
3573  * already available), otherwise falls back to plain reducing.
3574  */
3575 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3576 {
3577         u64 num_devices = root->fs_info->fs_devices->rw_devices;
3578         u64 target;
3579         u64 tmp;
3580
3581         /*
3582          * see if restripe for this chunk_type is in progress, if so
3583          * try to reduce to the target profile
3584          */
3585         spin_lock(&root->fs_info->balance_lock);
3586         target = get_restripe_target(root->fs_info, flags);
3587         if (target) {
3588                 /* pick target profile only if it's already available */
3589                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3590                         spin_unlock(&root->fs_info->balance_lock);
3591                         return extended_to_chunk(target);
3592                 }
3593         }
3594         spin_unlock(&root->fs_info->balance_lock);
3595
3596         /* First, mask out the RAID levels which aren't possible */
3597         if (num_devices == 1)
3598                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3599                            BTRFS_BLOCK_GROUP_RAID5);
3600         if (num_devices < 3)
3601                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3602         if (num_devices < 4)
3603                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3604
3605         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3606                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3607                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3608         flags &= ~tmp;
3609
3610         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3611                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3612         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3613                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3614         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3615                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3616         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3617                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3618         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3619                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3620
3621         return extended_to_chunk(flags | tmp);
3622 }
3623
3624 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
3625 {
3626         unsigned seq;
3627         u64 flags;
3628
3629         do {
3630                 flags = orig_flags;
3631                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3632
3633                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3634                         flags |= root->fs_info->avail_data_alloc_bits;
3635                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3636                         flags |= root->fs_info->avail_system_alloc_bits;
3637                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3638                         flags |= root->fs_info->avail_metadata_alloc_bits;
3639         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3640
3641         return btrfs_reduce_alloc_profile(root, flags);
3642 }
3643
3644 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3645 {
3646         u64 flags;
3647         u64 ret;
3648
3649         if (data)
3650                 flags = BTRFS_BLOCK_GROUP_DATA;
3651         else if (root == root->fs_info->chunk_root)
3652                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3653         else
3654                 flags = BTRFS_BLOCK_GROUP_METADATA;
3655
3656         ret = get_alloc_profile(root, flags);
3657         return ret;
3658 }
3659
3660 /*
3661  * This will check the space that the inode allocates from to make sure we have
3662  * enough space for bytes.
3663  */
3664 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3665 {
3666         struct btrfs_space_info *data_sinfo;
3667         struct btrfs_root *root = BTRFS_I(inode)->root;
3668         struct btrfs_fs_info *fs_info = root->fs_info;
3669         u64 used;
3670         int ret = 0, committed = 0, alloc_chunk = 1;
3671
3672         /* make sure bytes are sectorsize aligned */
3673         bytes = ALIGN(bytes, root->sectorsize);
3674
3675         if (btrfs_is_free_space_inode(inode)) {
3676                 committed = 1;
3677                 ASSERT(current->journal_info);
3678         }
3679
3680         data_sinfo = fs_info->data_sinfo;
3681         if (!data_sinfo)
3682                 goto alloc;
3683
3684 again:
3685         /* make sure we have enough space to handle the data first */
3686         spin_lock(&data_sinfo->lock);
3687         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3688                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3689                 data_sinfo->bytes_may_use;
3690
3691         if (used + bytes > data_sinfo->total_bytes) {
3692                 struct btrfs_trans_handle *trans;
3693
3694                 /*
3695                  * if we don't have enough free bytes in this space then we need
3696                  * to alloc a new chunk.
3697                  */
3698                 if (!data_sinfo->full && alloc_chunk) {
3699                         u64 alloc_target;
3700
3701                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3702                         spin_unlock(&data_sinfo->lock);
3703 alloc:
3704                         alloc_target = btrfs_get_alloc_profile(root, 1);
3705                         /*
3706                          * It is ugly that we don't call nolock join
3707                          * transaction for the free space inode case here.
3708                          * But it is safe because we only do the data space
3709                          * reservation for the free space cache in the
3710                          * transaction context, the common join transaction
3711                          * just increase the counter of the current transaction
3712                          * handler, doesn't try to acquire the trans_lock of
3713                          * the fs.
3714                          */
3715                         trans = btrfs_join_transaction(root);
3716                         if (IS_ERR(trans))
3717                                 return PTR_ERR(trans);
3718
3719                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3720                                              alloc_target,
3721                                              CHUNK_ALLOC_NO_FORCE);
3722                         btrfs_end_transaction(trans, root);
3723                         if (ret < 0) {
3724                                 if (ret != -ENOSPC)
3725                                         return ret;
3726                                 else
3727                                         goto commit_trans;
3728                         }
3729
3730                         if (!data_sinfo)
3731                                 data_sinfo = fs_info->data_sinfo;
3732
3733                         goto again;
3734                 }
3735
3736                 /*
3737                  * If we don't have enough pinned space to deal with this
3738                  * allocation don't bother committing the transaction.
3739                  */
3740                 if (percpu_counter_compare(&data_sinfo->total_bytes_pinned,
3741                                            bytes) < 0)
3742                         committed = 1;
3743                 spin_unlock(&data_sinfo->lock);
3744
3745                 /* commit the current transaction and try again */
3746 commit_trans:
3747                 if (!committed &&
3748                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3749                         committed = 1;
3750
3751                         trans = btrfs_join_transaction(root);
3752                         if (IS_ERR(trans))
3753                                 return PTR_ERR(trans);
3754                         ret = btrfs_commit_transaction(trans, root);
3755                         if (ret)
3756                                 return ret;
3757                         goto again;
3758                 }
3759
3760                 trace_btrfs_space_reservation(root->fs_info,
3761                                               "space_info:enospc",
3762                                               data_sinfo->flags, bytes, 1);
3763                 return -ENOSPC;
3764         }
3765         data_sinfo->bytes_may_use += bytes;
3766         trace_btrfs_space_reservation(root->fs_info, "space_info",
3767                                       data_sinfo->flags, bytes, 1);
3768         spin_unlock(&data_sinfo->lock);
3769
3770         return 0;
3771 }
3772
3773 /*
3774  * Called if we need to clear a data reservation for this inode.
3775  */
3776 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3777 {
3778         struct btrfs_root *root = BTRFS_I(inode)->root;
3779         struct btrfs_space_info *data_sinfo;
3780
3781         /* make sure bytes are sectorsize aligned */
3782         bytes = ALIGN(bytes, root->sectorsize);
3783
3784         data_sinfo = root->fs_info->data_sinfo;
3785         spin_lock(&data_sinfo->lock);
3786         WARN_ON(data_sinfo->bytes_may_use < bytes);
3787         data_sinfo->bytes_may_use -= bytes;
3788         trace_btrfs_space_reservation(root->fs_info, "space_info",
3789                                       data_sinfo->flags, bytes, 0);
3790         spin_unlock(&data_sinfo->lock);
3791 }
3792
3793 static void force_metadata_allocation(struct btrfs_fs_info *info)
3794 {
3795         struct list_head *head = &info->space_info;
3796         struct btrfs_space_info *found;
3797
3798         rcu_read_lock();
3799         list_for_each_entry_rcu(found, head, list) {
3800                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3801                         found->force_alloc = CHUNK_ALLOC_FORCE;
3802         }
3803         rcu_read_unlock();
3804 }
3805
3806 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
3807 {
3808         return (global->size << 1);
3809 }
3810
3811 static int should_alloc_chunk(struct btrfs_root *root,
3812                               struct btrfs_space_info *sinfo, int force)
3813 {
3814         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3815         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3816         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3817         u64 thresh;
3818
3819         if (force == CHUNK_ALLOC_FORCE)
3820                 return 1;
3821
3822         /*
3823          * We need to take into account the global rsv because for all intents
3824          * and purposes it's used space.  Don't worry about locking the
3825          * global_rsv, it doesn't change except when the transaction commits.
3826          */
3827         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3828                 num_allocated += calc_global_rsv_need_space(global_rsv);
3829
3830         /*
3831          * in limited mode, we want to have some free space up to
3832          * about 1% of the FS size.
3833          */
3834         if (force == CHUNK_ALLOC_LIMITED) {
3835                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3836                 thresh = max_t(u64, 64 * 1024 * 1024,
3837                                div_factor_fine(thresh, 1));
3838
3839                 if (num_bytes - num_allocated < thresh)
3840                         return 1;
3841         }
3842
3843         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3844                 return 0;
3845         return 1;
3846 }
3847
3848 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3849 {
3850         u64 num_dev;
3851
3852         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
3853                     BTRFS_BLOCK_GROUP_RAID0 |
3854                     BTRFS_BLOCK_GROUP_RAID5 |
3855                     BTRFS_BLOCK_GROUP_RAID6))
3856                 num_dev = root->fs_info->fs_devices->rw_devices;
3857         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3858                 num_dev = 2;
3859         else
3860                 num_dev = 1;    /* DUP or single */
3861
3862         /* metadata for updaing devices and chunk tree */
3863         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3864 }
3865
3866 static void check_system_chunk(struct btrfs_trans_handle *trans,
3867                                struct btrfs_root *root, u64 type)
3868 {
3869         struct btrfs_space_info *info;
3870         u64 left;
3871         u64 thresh;
3872
3873         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3874         spin_lock(&info->lock);
3875         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3876                 info->bytes_reserved - info->bytes_readonly;
3877         spin_unlock(&info->lock);
3878
3879         thresh = get_system_chunk_thresh(root, type);
3880         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3881                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
3882                         left, thresh, type);
3883                 dump_space_info(info, 0, 0);
3884         }
3885
3886         if (left < thresh) {
3887                 u64 flags;
3888
3889                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3890                 btrfs_alloc_chunk(trans, root, flags);
3891         }
3892 }
3893
3894 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3895                           struct btrfs_root *extent_root, u64 flags, int force)
3896 {
3897         struct btrfs_space_info *space_info;
3898         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3899         int wait_for_alloc = 0;
3900         int ret = 0;
3901
3902         /* Don't re-enter if we're already allocating a chunk */
3903         if (trans->allocating_chunk)
3904                 return -ENOSPC;
3905
3906         space_info = __find_space_info(extent_root->fs_info, flags);
3907         if (!space_info) {
3908                 ret = update_space_info(extent_root->fs_info, flags,
3909                                         0, 0, &space_info);
3910                 BUG_ON(ret); /* -ENOMEM */
3911         }
3912         BUG_ON(!space_info); /* Logic error */
3913
3914 again:
3915         spin_lock(&space_info->lock);
3916         if (force < space_info->force_alloc)
3917                 force = space_info->force_alloc;
3918         if (space_info->full) {
3919                 if (should_alloc_chunk(extent_root, space_info, force))
3920                         ret = -ENOSPC;
3921                 else
3922                         ret = 0;
3923                 spin_unlock(&space_info->lock);
3924                 return ret;
3925         }
3926
3927         if (!should_alloc_chunk(extent_root, space_info, force)) {
3928                 spin_unlock(&space_info->lock);
3929                 return 0;
3930         } else if (space_info->chunk_alloc) {
3931                 wait_for_alloc = 1;
3932         } else {
3933                 space_info->chunk_alloc = 1;
3934         }
3935
3936         spin_unlock(&space_info->lock);
3937
3938         mutex_lock(&fs_info->chunk_mutex);
3939
3940         /*
3941          * The chunk_mutex is held throughout the entirety of a chunk
3942          * allocation, so once we've acquired the chunk_mutex we know that the
3943          * other guy is done and we need to recheck and see if we should
3944          * allocate.
3945          */
3946         if (wait_for_alloc) {
3947                 mutex_unlock(&fs_info->chunk_mutex);
3948                 wait_for_alloc = 0;
3949                 goto again;
3950         }
3951
3952         trans->allocating_chunk = true;
3953
3954         /*
3955          * If we have mixed data/metadata chunks we want to make sure we keep
3956          * allocating mixed chunks instead of individual chunks.
3957          */
3958         if (btrfs_mixed_space_info(space_info))
3959                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3960
3961         /*
3962          * if we're doing a data chunk, go ahead and make sure that
3963          * we keep a reasonable number of metadata chunks allocated in the
3964          * FS as well.
3965          */
3966         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3967                 fs_info->data_chunk_allocations++;
3968                 if (!(fs_info->data_chunk_allocations %
3969                       fs_info->metadata_ratio))
3970                         force_metadata_allocation(fs_info);
3971         }
3972
3973         /*
3974          * Check if we have enough space in SYSTEM chunk because we may need
3975          * to update devices.
3976          */
3977         check_system_chunk(trans, extent_root, flags);
3978
3979         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3980         trans->allocating_chunk = false;
3981
3982         spin_lock(&space_info->lock);
3983         if (ret < 0 && ret != -ENOSPC)
3984                 goto out;
3985         if (ret)
3986                 space_info->full = 1;
3987         else
3988                 ret = 1;
3989
3990         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3991 out:
3992         space_info->chunk_alloc = 0;
3993         spin_unlock(&space_info->lock);
3994         mutex_unlock(&fs_info->chunk_mutex);
3995         return ret;
3996 }
3997
3998 static int can_overcommit(struct btrfs_root *root,
3999                           struct btrfs_space_info *space_info, u64 bytes,
4000                           enum btrfs_reserve_flush_enum flush)
4001 {
4002         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4003         u64 profile = btrfs_get_alloc_profile(root, 0);
4004         u64 space_size;
4005         u64 avail;
4006         u64 used;
4007
4008         used = space_info->bytes_used + space_info->bytes_reserved +
4009                 space_info->bytes_pinned + space_info->bytes_readonly;
4010
4011         /*
4012          * We only want to allow over committing if we have lots of actual space
4013          * free, but if we don't have enough space to handle the global reserve
4014          * space then we could end up having a real enospc problem when trying
4015          * to allocate a chunk or some other such important allocation.
4016          */
4017         spin_lock(&global_rsv->lock);
4018         space_size = calc_global_rsv_need_space(global_rsv);
4019         spin_unlock(&global_rsv->lock);
4020         if (used + space_size >= space_info->total_bytes)
4021                 return 0;
4022
4023         used += space_info->bytes_may_use;
4024
4025         spin_lock(&root->fs_info->free_chunk_lock);
4026         avail = root->fs_info->free_chunk_space;
4027         spin_unlock(&root->fs_info->free_chunk_lock);
4028
4029         /*
4030          * If we have dup, raid1 or raid10 then only half of the free
4031          * space is actually useable.  For raid56, the space info used
4032          * doesn't include the parity drive, so we don't have to
4033          * change the math
4034          */
4035         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4036                        BTRFS_BLOCK_GROUP_RAID1 |
4037                        BTRFS_BLOCK_GROUP_RAID10))
4038                 avail >>= 1;
4039
4040         /*
4041          * If we aren't flushing all things, let us overcommit up to
4042          * 1/2th of the space. If we can flush, don't let us overcommit
4043          * too much, let it overcommit up to 1/8 of the space.
4044          */
4045         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4046                 avail >>= 3;
4047         else
4048                 avail >>= 1;
4049
4050         if (used + bytes < space_info->total_bytes + avail)
4051                 return 1;
4052         return 0;
4053 }
4054
4055 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4056                                          unsigned long nr_pages, int nr_items)
4057 {
4058         struct super_block *sb = root->fs_info->sb;
4059
4060         if (down_read_trylock(&sb->s_umount)) {
4061                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4062                 up_read(&sb->s_umount);
4063         } else {
4064                 /*
4065                  * We needn't worry the filesystem going from r/w to r/o though
4066                  * we don't acquire ->s_umount mutex, because the filesystem
4067                  * should guarantee the delalloc inodes list be empty after
4068                  * the filesystem is readonly(all dirty pages are written to
4069                  * the disk).
4070                  */
4071                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4072                 if (!current->journal_info)
4073                         btrfs_wait_ordered_roots(root->fs_info, nr_items);
4074         }
4075 }
4076
4077 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4078 {
4079         u64 bytes;
4080         int nr;
4081
4082         bytes = btrfs_calc_trans_metadata_size(root, 1);
4083         nr = (int)div64_u64(to_reclaim, bytes);
4084         if (!nr)
4085                 nr = 1;
4086         return nr;
4087 }
4088
4089 #define EXTENT_SIZE_PER_ITEM    (256 * 1024)
4090
4091 /*
4092  * shrink metadata reservation for delalloc
4093  */
4094 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4095                             bool wait_ordered)
4096 {
4097         struct btrfs_block_rsv *block_rsv;
4098         struct btrfs_space_info *space_info;
4099         struct btrfs_trans_handle *trans;
4100         u64 delalloc_bytes;
4101         u64 max_reclaim;
4102         long time_left;
4103         unsigned long nr_pages;
4104         int loops;
4105         int items;
4106         enum btrfs_reserve_flush_enum flush;
4107
4108         /* Calc the number of the pages we need flush for space reservation */
4109         items = calc_reclaim_items_nr(root, to_reclaim);
4110         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4111
4112         trans = (struct btrfs_trans_handle *)current->journal_info;
4113         block_rsv = &root->fs_info->delalloc_block_rsv;
4114         space_info = block_rsv->space_info;
4115
4116         delalloc_bytes = percpu_counter_sum_positive(
4117                                                 &root->fs_info->delalloc_bytes);
4118         if (delalloc_bytes == 0) {
4119                 if (trans)
4120                         return;
4121                 if (wait_ordered)
4122                         btrfs_wait_ordered_roots(root->fs_info, items);
4123                 return;
4124         }
4125
4126         loops = 0;
4127         while (delalloc_bytes && loops < 3) {
4128                 max_reclaim = min(delalloc_bytes, to_reclaim);
4129                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4130                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4131                 /*
4132                  * We need to wait for the async pages to actually start before
4133                  * we do anything.
4134                  */
4135                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4136                 if (!max_reclaim)
4137                         goto skip_async;
4138
4139                 if (max_reclaim <= nr_pages)
4140                         max_reclaim = 0;
4141                 else
4142                         max_reclaim -= nr_pages;
4143
4144                 wait_event(root->fs_info->async_submit_wait,
4145                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4146                            (int)max_reclaim);
4147 skip_async:
4148                 if (!trans)
4149                         flush = BTRFS_RESERVE_FLUSH_ALL;
4150                 else
4151                         flush = BTRFS_RESERVE_NO_FLUSH;
4152                 spin_lock(&space_info->lock);
4153                 if (can_overcommit(root, space_info, orig, flush)) {
4154                         spin_unlock(&space_info->lock);
4155                         break;
4156                 }
4157                 spin_unlock(&space_info->lock);
4158
4159                 loops++;
4160                 if (wait_ordered && !trans) {
4161                         btrfs_wait_ordered_roots(root->fs_info, items);
4162                 } else {
4163                         time_left = schedule_timeout_killable(1);
4164                         if (time_left)
4165                                 break;
4166                 }
4167                 delalloc_bytes = percpu_counter_sum_positive(
4168                                                 &root->fs_info->delalloc_bytes);
4169         }
4170 }
4171
4172 /**
4173  * maybe_commit_transaction - possibly commit the transaction if its ok to
4174  * @root - the root we're allocating for
4175  * @bytes - the number of bytes we want to reserve
4176  * @force - force the commit
4177  *
4178  * This will check to make sure that committing the transaction will actually
4179  * get us somewhere and then commit the transaction if it does.  Otherwise it
4180  * will return -ENOSPC.
4181  */
4182 static int may_commit_transaction(struct btrfs_root *root,
4183                                   struct btrfs_space_info *space_info,
4184                                   u64 bytes, int force)
4185 {
4186         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4187         struct btrfs_trans_handle *trans;
4188
4189         trans = (struct btrfs_trans_handle *)current->journal_info;
4190         if (trans)
4191                 return -EAGAIN;
4192
4193         if (force)
4194                 goto commit;
4195
4196         /* See if there is enough pinned space to make this reservation */
4197         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4198                                    bytes) >= 0)
4199                 goto commit;
4200
4201         /*
4202          * See if there is some space in the delayed insertion reservation for
4203          * this reservation.
4204          */
4205         if (space_info != delayed_rsv->space_info)
4206                 return -ENOSPC;
4207
4208         spin_lock(&delayed_rsv->lock);
4209         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4210                                    bytes - delayed_rsv->size) >= 0) {
4211                 spin_unlock(&delayed_rsv->lock);
4212                 return -ENOSPC;
4213         }
4214         spin_unlock(&delayed_rsv->lock);
4215
4216 commit:
4217         trans = btrfs_join_transaction(root);
4218         if (IS_ERR(trans))
4219                 return -ENOSPC;
4220
4221         return btrfs_commit_transaction(trans, root);
4222 }
4223
4224 enum flush_state {
4225         FLUSH_DELAYED_ITEMS_NR  =       1,
4226         FLUSH_DELAYED_ITEMS     =       2,
4227         FLUSH_DELALLOC          =       3,
4228         FLUSH_DELALLOC_WAIT     =       4,
4229         ALLOC_CHUNK             =       5,
4230         COMMIT_TRANS            =       6,
4231 };
4232
4233 static int flush_space(struct btrfs_root *root,
4234                        struct btrfs_space_info *space_info, u64 num_bytes,
4235                        u64 orig_bytes, int state)
4236 {
4237         struct btrfs_trans_handle *trans;
4238         int nr;
4239         int ret = 0;
4240
4241         switch (state) {
4242         case FLUSH_DELAYED_ITEMS_NR:
4243         case FLUSH_DELAYED_ITEMS:
4244                 if (state == FLUSH_DELAYED_ITEMS_NR)
4245                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4246                 else
4247                         nr = -1;
4248
4249                 trans = btrfs_join_transaction(root);
4250                 if (IS_ERR(trans)) {
4251                         ret = PTR_ERR(trans);
4252                         break;
4253                 }
4254                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4255                 btrfs_end_transaction(trans, root);
4256                 break;
4257         case FLUSH_DELALLOC:
4258         case FLUSH_DELALLOC_WAIT:
4259                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4260                                 state == FLUSH_DELALLOC_WAIT);
4261                 break;
4262         case ALLOC_CHUNK:
4263                 trans = btrfs_join_transaction(root);
4264                 if (IS_ERR(trans)) {
4265                         ret = PTR_ERR(trans);
4266                         break;
4267                 }
4268                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4269                                      btrfs_get_alloc_profile(root, 0),
4270                                      CHUNK_ALLOC_NO_FORCE);
4271                 btrfs_end_transaction(trans, root);
4272                 if (ret == -ENOSPC)
4273                         ret = 0;
4274                 break;
4275         case COMMIT_TRANS:
4276                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4277                 break;
4278         default:
4279                 ret = -ENOSPC;
4280                 break;
4281         }
4282
4283         return ret;
4284 }
4285
4286 static inline u64
4287 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4288                                  struct btrfs_space_info *space_info)
4289 {
4290         u64 used;
4291         u64 expected;
4292         u64 to_reclaim;
4293
4294         to_reclaim = min_t(u64, num_online_cpus() * 1024 * 1024,
4295                                 16 * 1024 * 1024);
4296         spin_lock(&space_info->lock);
4297         if (can_overcommit(root, space_info, to_reclaim,
4298                            BTRFS_RESERVE_FLUSH_ALL)) {
4299                 to_reclaim = 0;
4300                 goto out;
4301         }
4302
4303         used = space_info->bytes_used + space_info->bytes_reserved +
4304                space_info->bytes_pinned + space_info->bytes_readonly +
4305                space_info->bytes_may_use;
4306         if (can_overcommit(root, space_info, 1024 * 1024,
4307                            BTRFS_RESERVE_FLUSH_ALL))
4308                 expected = div_factor_fine(space_info->total_bytes, 95);
4309         else
4310                 expected = div_factor_fine(space_info->total_bytes, 90);
4311
4312         if (used > expected)
4313                 to_reclaim = used - expected;
4314         else
4315                 to_reclaim = 0;
4316         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4317                                      space_info->bytes_reserved);
4318 out:
4319         spin_unlock(&space_info->lock);
4320
4321         return to_reclaim;
4322 }
4323
4324 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4325                                         struct btrfs_fs_info *fs_info, u64 used)
4326 {
4327         return (used >= div_factor_fine(space_info->total_bytes, 98) &&
4328                 !btrfs_fs_closing(fs_info) &&
4329                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4330 }
4331
4332 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4333                                        struct btrfs_fs_info *fs_info,
4334                                        int flush_state)
4335 {
4336         u64 used;
4337
4338         spin_lock(&space_info->lock);
4339         /*
4340          * We run out of space and have not got any free space via flush_space,
4341          * so don't bother doing async reclaim.
4342          */
4343         if (flush_state > COMMIT_TRANS && space_info->full) {
4344                 spin_unlock(&space_info->lock);
4345                 return 0;
4346         }
4347
4348         used = space_info->bytes_used + space_info->bytes_reserved +
4349                space_info->bytes_pinned + space_info->bytes_readonly +
4350                space_info->bytes_may_use;
4351         if (need_do_async_reclaim(space_info, fs_info, used)) {
4352                 spin_unlock(&space_info->lock);
4353                 return 1;
4354         }
4355         spin_unlock(&space_info->lock);
4356
4357         return 0;
4358 }
4359
4360 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4361 {
4362         struct btrfs_fs_info *fs_info;
4363         struct btrfs_space_info *space_info;
4364         u64 to_reclaim;
4365         int flush_state;
4366
4367         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4368         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4369
4370         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4371                                                       space_info);
4372         if (!to_reclaim)
4373                 return;
4374
4375         flush_state = FLUSH_DELAYED_ITEMS_NR;
4376         do {
4377                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4378                             to_reclaim, flush_state);
4379                 flush_state++;
4380                 if (!btrfs_need_do_async_reclaim(space_info, fs_info,
4381                                                  flush_state))
4382                         return;
4383         } while (flush_state <= COMMIT_TRANS);
4384
4385         if (btrfs_need_do_async_reclaim(space_info, fs_info, flush_state))
4386                 queue_work(system_unbound_wq, work);
4387 }
4388
4389 void btrfs_init_async_reclaim_work(struct work_struct *work)
4390 {
4391         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4392 }
4393
4394 /**
4395  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4396  * @root - the root we're allocating for
4397  * @block_rsv - the block_rsv we're allocating for
4398  * @orig_bytes - the number of bytes we want
4399  * @flush - whether or not we can flush to make our reservation
4400  *
4401  * This will reserve orgi_bytes number of bytes from the space info associated
4402  * with the block_rsv.  If there is not enough space it will make an attempt to
4403  * flush out space to make room.  It will do this by flushing delalloc if
4404  * possible or committing the transaction.  If flush is 0 then no attempts to
4405  * regain reservations will be made and this will fail if there is not enough
4406  * space already.
4407  */
4408 static int reserve_metadata_bytes(struct btrfs_root *root,
4409                                   struct btrfs_block_rsv *block_rsv,
4410                                   u64 orig_bytes,
4411                                   enum btrfs_reserve_flush_enum flush)
4412 {
4413         struct btrfs_space_info *space_info = block_rsv->space_info;
4414         u64 used;
4415         u64 num_bytes = orig_bytes;
4416         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4417         int ret = 0;
4418         bool flushing = false;
4419
4420 again:
4421         ret = 0;
4422         spin_lock(&space_info->lock);
4423         /*
4424          * We only want to wait if somebody other than us is flushing and we
4425          * are actually allowed to flush all things.
4426          */
4427         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4428                space_info->flush) {
4429                 spin_unlock(&space_info->lock);
4430                 /*
4431                  * If we have a trans handle we can't wait because the flusher
4432                  * may have to commit the transaction, which would mean we would
4433                  * deadlock since we are waiting for the flusher to finish, but
4434                  * hold the current transaction open.
4435                  */
4436                 if (current->journal_info)
4437                         return -EAGAIN;
4438                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4439                 /* Must have been killed, return */
4440                 if (ret)
4441                         return -EINTR;
4442
4443                 spin_lock(&space_info->lock);
4444         }
4445
4446         ret = -ENOSPC;
4447         used = space_info->bytes_used + space_info->bytes_reserved +
4448                 space_info->bytes_pinned + space_info->bytes_readonly +
4449                 space_info->bytes_may_use;
4450
4451         /*
4452          * The idea here is that we've not already over-reserved the block group
4453          * then we can go ahead and save our reservation first and then start
4454          * flushing if we need to.  Otherwise if we've already overcommitted
4455          * lets start flushing stuff first and then come back and try to make
4456          * our reservation.
4457          */
4458         if (used <= space_info->total_bytes) {
4459                 if (used + orig_bytes <= space_info->total_bytes) {
4460                         space_info->bytes_may_use += orig_bytes;
4461                         trace_btrfs_space_reservation(root->fs_info,
4462                                 "space_info", space_info->flags, orig_bytes, 1);
4463                         ret = 0;
4464                 } else {
4465                         /*
4466                          * Ok set num_bytes to orig_bytes since we aren't
4467                          * overocmmitted, this way we only try and reclaim what
4468                          * we need.
4469                          */
4470                         num_bytes = orig_bytes;
4471                 }
4472         } else {
4473                 /*
4474                  * Ok we're over committed, set num_bytes to the overcommitted
4475                  * amount plus the amount of bytes that we need for this
4476                  * reservation.
4477                  */
4478                 num_bytes = used - space_info->total_bytes +
4479                         (orig_bytes * 2);
4480         }
4481
4482         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4483                 space_info->bytes_may_use += orig_bytes;
4484                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4485                                               space_info->flags, orig_bytes,
4486                                               1);
4487                 ret = 0;
4488         }
4489
4490         /*
4491          * Couldn't make our reservation, save our place so while we're trying
4492          * to reclaim space we can actually use it instead of somebody else
4493          * stealing it from us.
4494          *
4495          * We make the other tasks wait for the flush only when we can flush
4496          * all things.
4497          */
4498         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4499                 flushing = true;
4500                 space_info->flush = 1;
4501         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
4502                 used += orig_bytes;
4503                 /*
4504                  * We will do the space reservation dance during log replay,
4505                  * which means we won't have fs_info->fs_root set, so don't do
4506                  * the async reclaim as we will panic.
4507                  */
4508                 if (!root->fs_info->log_root_recovering &&
4509                     need_do_async_reclaim(space_info, root->fs_info, used) &&
4510                     !work_busy(&root->fs_info->async_reclaim_work))
4511                         queue_work(system_unbound_wq,
4512                                    &root->fs_info->async_reclaim_work);
4513         }
4514         spin_unlock(&space_info->lock);
4515
4516         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4517                 goto out;
4518
4519         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4520                           flush_state);
4521         flush_state++;
4522
4523         /*
4524          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4525          * would happen. So skip delalloc flush.
4526          */
4527         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4528             (flush_state == FLUSH_DELALLOC ||
4529              flush_state == FLUSH_DELALLOC_WAIT))
4530                 flush_state = ALLOC_CHUNK;
4531
4532         if (!ret)
4533                 goto again;
4534         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4535                  flush_state < COMMIT_TRANS)
4536                 goto again;
4537         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4538                  flush_state <= COMMIT_TRANS)
4539                 goto again;
4540
4541 out:
4542         if (ret == -ENOSPC &&
4543             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4544                 struct btrfs_block_rsv *global_rsv =
4545                         &root->fs_info->global_block_rsv;
4546
4547                 if (block_rsv != global_rsv &&
4548                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4549                         ret = 0;
4550         }
4551         if (ret == -ENOSPC)
4552                 trace_btrfs_space_reservation(root->fs_info,
4553                                               "space_info:enospc",
4554                                               space_info->flags, orig_bytes, 1);
4555         if (flushing) {
4556                 spin_lock(&space_info->lock);
4557                 space_info->flush = 0;
4558                 wake_up_all(&space_info->wait);
4559                 spin_unlock(&space_info->lock);
4560         }
4561         return ret;
4562 }
4563
4564 static struct btrfs_block_rsv *get_block_rsv(
4565                                         const struct btrfs_trans_handle *trans,
4566                                         const struct btrfs_root *root)
4567 {
4568         struct btrfs_block_rsv *block_rsv = NULL;
4569
4570         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4571                 block_rsv = trans->block_rsv;
4572
4573         if (root == root->fs_info->csum_root && trans->adding_csums)
4574                 block_rsv = trans->block_rsv;
4575
4576         if (root == root->fs_info->uuid_root)
4577                 block_rsv = trans->block_rsv;
4578
4579         if (!block_rsv)
4580                 block_rsv = root->block_rsv;
4581
4582         if (!block_rsv)
4583                 block_rsv = &root->fs_info->empty_block_rsv;
4584
4585         return block_rsv;
4586 }
4587
4588 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4589                                u64 num_bytes)
4590 {
4591         int ret = -ENOSPC;
4592         spin_lock(&block_rsv->lock);
4593         if (block_rsv->reserved >= num_bytes) {
4594                 block_rsv->reserved -= num_bytes;
4595                 if (block_rsv->reserved < block_rsv->size)
4596                         block_rsv->full = 0;
4597                 ret = 0;
4598         }
4599         spin_unlock(&block_rsv->lock);
4600         return ret;
4601 }
4602
4603 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4604                                 u64 num_bytes, int update_size)
4605 {
4606         spin_lock(&block_rsv->lock);
4607         block_rsv->reserved += num_bytes;
4608         if (update_size)
4609                 block_rsv->size += num_bytes;
4610         else if (block_rsv->reserved >= block_rsv->size)
4611                 block_rsv->full = 1;
4612         spin_unlock(&block_rsv->lock);
4613 }
4614
4615 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4616                              struct btrfs_block_rsv *dest, u64 num_bytes,
4617                              int min_factor)
4618 {
4619         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4620         u64 min_bytes;
4621
4622         if (global_rsv->space_info != dest->space_info)
4623                 return -ENOSPC;
4624
4625         spin_lock(&global_rsv->lock);
4626         min_bytes = div_factor(global_rsv->size, min_factor);
4627         if (global_rsv->reserved < min_bytes + num_bytes) {
4628                 spin_unlock(&global_rsv->lock);
4629                 return -ENOSPC;
4630         }
4631         global_rsv->reserved -= num_bytes;
4632         if (global_rsv->reserved < global_rsv->size)
4633                 global_rsv->full = 0;
4634         spin_unlock(&global_rsv->lock);
4635
4636         block_rsv_add_bytes(dest, num_bytes, 1);
4637         return 0;
4638 }
4639
4640 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4641                                     struct btrfs_block_rsv *block_rsv,
4642                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4643 {
4644         struct btrfs_space_info *space_info = block_rsv->space_info;
4645
4646         spin_lock(&block_rsv->lock);
4647         if (num_bytes == (u64)-1)
4648                 num_bytes = block_rsv->size;
4649         block_rsv->size -= num_bytes;
4650         if (block_rsv->reserved >= block_rsv->size) {
4651                 num_bytes = block_rsv->reserved - block_rsv->size;
4652                 block_rsv->reserved = block_rsv->size;
4653                 block_rsv->full = 1;
4654         } else {
4655                 num_bytes = 0;
4656         }
4657         spin_unlock(&block_rsv->lock);
4658
4659         if (num_bytes > 0) {
4660                 if (dest) {
4661                         spin_lock(&dest->lock);
4662                         if (!dest->full) {
4663                                 u64 bytes_to_add;
4664
4665                                 bytes_to_add = dest->size - dest->reserved;
4666                                 bytes_to_add = min(num_bytes, bytes_to_add);
4667                                 dest->reserved += bytes_to_add;
4668                                 if (dest->reserved >= dest->size)
4669                                         dest->full = 1;
4670                                 num_bytes -= bytes_to_add;
4671                         }
4672                         spin_unlock(&dest->lock);
4673                 }
4674                 if (num_bytes) {
4675                         spin_lock(&space_info->lock);
4676                         space_info->bytes_may_use -= num_bytes;
4677                         trace_btrfs_space_reservation(fs_info, "space_info",
4678                                         space_info->flags, num_bytes, 0);
4679                         spin_unlock(&space_info->lock);
4680                 }
4681         }
4682 }
4683
4684 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4685                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4686 {
4687         int ret;
4688
4689         ret = block_rsv_use_bytes(src, num_bytes);
4690         if (ret)
4691                 return ret;
4692
4693         block_rsv_add_bytes(dst, num_bytes, 1);
4694         return 0;
4695 }
4696
4697 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4698 {
4699         memset(rsv, 0, sizeof(*rsv));
4700         spin_lock_init(&rsv->lock);
4701         rsv->type = type;
4702 }
4703
4704 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4705                                               unsigned short type)
4706 {
4707         struct btrfs_block_rsv *block_rsv;
4708         struct btrfs_fs_info *fs_info = root->fs_info;
4709
4710         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4711         if (!block_rsv)
4712                 return NULL;
4713
4714         btrfs_init_block_rsv(block_rsv, type);
4715         block_rsv->space_info = __find_space_info(fs_info,
4716                                                   BTRFS_BLOCK_GROUP_METADATA);
4717         return block_rsv;
4718 }
4719
4720 void btrfs_free_block_rsv(struct btrfs_root *root,
4721                           struct btrfs_block_rsv *rsv)
4722 {
4723         if (!rsv)
4724                 return;
4725         btrfs_block_rsv_release(root, rsv, (u64)-1);
4726         kfree(rsv);
4727 }
4728
4729 int btrfs_block_rsv_add(struct btrfs_root *root,
4730                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4731                         enum btrfs_reserve_flush_enum flush)
4732 {
4733         int ret;
4734
4735         if (num_bytes == 0)
4736                 return 0;
4737
4738         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4739         if (!ret) {
4740                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4741                 return 0;
4742         }
4743
4744         return ret;
4745 }
4746
4747 int btrfs_block_rsv_check(struct btrfs_root *root,
4748                           struct btrfs_block_rsv *block_rsv, int min_factor)
4749 {
4750         u64 num_bytes = 0;
4751         int ret = -ENOSPC;
4752
4753         if (!block_rsv)
4754                 return 0;
4755
4756         spin_lock(&block_rsv->lock);
4757         num_bytes = div_factor(block_rsv->size, min_factor);
4758         if (block_rsv->reserved >= num_bytes)
4759                 ret = 0;
4760         spin_unlock(&block_rsv->lock);
4761
4762         return ret;
4763 }
4764
4765 int btrfs_block_rsv_refill(struct btrfs_root *root,
4766                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4767                            enum btrfs_reserve_flush_enum flush)
4768 {
4769         u64 num_bytes = 0;
4770         int ret = -ENOSPC;
4771
4772         if (!block_rsv)
4773                 return 0;
4774
4775         spin_lock(&block_rsv->lock);
4776         num_bytes = min_reserved;
4777         if (block_rsv->reserved >= num_bytes)
4778                 ret = 0;
4779         else
4780                 num_bytes -= block_rsv->reserved;
4781         spin_unlock(&block_rsv->lock);
4782
4783         if (!ret)
4784                 return 0;
4785
4786         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4787         if (!ret) {
4788                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4789                 return 0;
4790         }
4791
4792         return ret;
4793 }
4794
4795 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4796                             struct btrfs_block_rsv *dst_rsv,
4797                             u64 num_bytes)
4798 {
4799         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4800 }
4801
4802 void btrfs_block_rsv_release(struct btrfs_root *root,
4803                              struct btrfs_block_rsv *block_rsv,
4804                              u64 num_bytes)
4805 {
4806         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4807         if (global_rsv == block_rsv ||
4808             block_rsv->space_info != global_rsv->space_info)
4809                 global_rsv = NULL;
4810         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4811                                 num_bytes);
4812 }
4813
4814 /*
4815  * helper to calculate size of global block reservation.
4816  * the desired value is sum of space used by extent tree,
4817  * checksum tree and root tree
4818  */
4819 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4820 {
4821         struct btrfs_space_info *sinfo;
4822         u64 num_bytes;
4823         u64 meta_used;
4824         u64 data_used;
4825         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4826
4827         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4828         spin_lock(&sinfo->lock);
4829         data_used = sinfo->bytes_used;
4830         spin_unlock(&sinfo->lock);
4831
4832         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4833         spin_lock(&sinfo->lock);
4834         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4835                 data_used = 0;
4836         meta_used = sinfo->bytes_used;
4837         spin_unlock(&sinfo->lock);
4838
4839         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4840                     csum_size * 2;
4841         num_bytes += div64_u64(data_used + meta_used, 50);
4842
4843         if (num_bytes * 3 > meta_used)
4844                 num_bytes = div64_u64(meta_used, 3);
4845
4846         return ALIGN(num_bytes, fs_info->extent_root->nodesize << 10);
4847 }
4848
4849 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4850 {
4851         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4852         struct btrfs_space_info *sinfo = block_rsv->space_info;
4853         u64 num_bytes;
4854
4855         num_bytes = calc_global_metadata_size(fs_info);
4856
4857         spin_lock(&sinfo->lock);
4858         spin_lock(&block_rsv->lock);
4859
4860         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
4861
4862         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4863                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4864                     sinfo->bytes_may_use;
4865
4866         if (sinfo->total_bytes > num_bytes) {
4867                 num_bytes = sinfo->total_bytes - num_bytes;
4868                 block_rsv->reserved += num_bytes;
4869                 sinfo->bytes_may_use += num_bytes;
4870                 trace_btrfs_space_reservation(fs_info, "space_info",
4871                                       sinfo->flags, num_bytes, 1);
4872         }
4873
4874         if (block_rsv->reserved >= block_rsv->size) {
4875                 num_bytes = block_rsv->reserved - block_rsv->size;
4876                 sinfo->bytes_may_use -= num_bytes;
4877                 trace_btrfs_space_reservation(fs_info, "space_info",
4878                                       sinfo->flags, num_bytes, 0);
4879                 block_rsv->reserved = block_rsv->size;
4880                 block_rsv->full = 1;
4881         }
4882
4883         spin_unlock(&block_rsv->lock);
4884         spin_unlock(&sinfo->lock);
4885 }
4886
4887 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4888 {
4889         struct btrfs_space_info *space_info;
4890
4891         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4892         fs_info->chunk_block_rsv.space_info = space_info;
4893
4894         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4895         fs_info->global_block_rsv.space_info = space_info;
4896         fs_info->delalloc_block_rsv.space_info = space_info;
4897         fs_info->trans_block_rsv.space_info = space_info;
4898         fs_info->empty_block_rsv.space_info = space_info;
4899         fs_info->delayed_block_rsv.space_info = space_info;
4900
4901         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4902         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4903         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4904         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4905         if (fs_info->quota_root)
4906                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
4907         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4908
4909         update_global_block_rsv(fs_info);
4910 }
4911
4912 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4913 {
4914         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4915                                 (u64)-1);
4916         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4917         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4918         WARN_ON(fs_info->trans_block_rsv.size > 0);
4919         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4920         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4921         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4922         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4923         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4924 }
4925
4926 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4927                                   struct btrfs_root *root)
4928 {
4929         if (!trans->block_rsv)
4930                 return;
4931
4932         if (!trans->bytes_reserved)
4933                 return;
4934
4935         trace_btrfs_space_reservation(root->fs_info, "transaction",
4936                                       trans->transid, trans->bytes_reserved, 0);
4937         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4938         trans->bytes_reserved = 0;
4939 }
4940
4941 /* Can only return 0 or -ENOSPC */
4942 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4943                                   struct inode *inode)
4944 {
4945         struct btrfs_root *root = BTRFS_I(inode)->root;
4946         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4947         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4948
4949         /*
4950          * We need to hold space in order to delete our orphan item once we've
4951          * added it, so this takes the reservation so we can release it later
4952          * when we are truly done with the orphan item.
4953          */
4954         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4955         trace_btrfs_space_reservation(root->fs_info, "orphan",
4956                                       btrfs_ino(inode), num_bytes, 1);
4957         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4958 }
4959
4960 void btrfs_orphan_release_metadata(struct inode *inode)
4961 {
4962         struct btrfs_root *root = BTRFS_I(inode)->root;
4963         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4964         trace_btrfs_space_reservation(root->fs_info, "orphan",
4965                                       btrfs_ino(inode), num_bytes, 0);
4966         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4967 }
4968
4969 /*
4970  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
4971  * root: the root of the parent directory
4972  * rsv: block reservation
4973  * items: the number of items that we need do reservation
4974  * qgroup_reserved: used to return the reserved size in qgroup
4975  *
4976  * This function is used to reserve the space for snapshot/subvolume
4977  * creation and deletion. Those operations are different with the
4978  * common file/directory operations, they change two fs/file trees
4979  * and root tree, the number of items that the qgroup reserves is
4980  * different with the free space reservation. So we can not use
4981  * the space reseravtion mechanism in start_transaction().
4982  */
4983 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
4984                                      struct btrfs_block_rsv *rsv,
4985                                      int items,
4986                                      u64 *qgroup_reserved,
4987                                      bool use_global_rsv)
4988 {
4989         u64 num_bytes;
4990         int ret;
4991         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4992
4993         if (root->fs_info->quota_enabled) {
4994                 /* One for parent inode, two for dir entries */
4995                 num_bytes = 3 * root->nodesize;
4996                 ret = btrfs_qgroup_reserve(root, num_bytes);
4997                 if (ret)
4998                         return ret;
4999         } else {
5000                 num_bytes = 0;
5001         }
5002
5003         *qgroup_reserved = num_bytes;
5004
5005         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5006         rsv->space_info = __find_space_info(root->fs_info,
5007                                             BTRFS_BLOCK_GROUP_METADATA);
5008         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5009                                   BTRFS_RESERVE_FLUSH_ALL);
5010
5011         if (ret == -ENOSPC && use_global_rsv)
5012                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
5013
5014         if (ret) {
5015                 if (*qgroup_reserved)
5016                         btrfs_qgroup_free(root, *qgroup_reserved);
5017         }
5018
5019         return ret;
5020 }
5021
5022 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5023                                       struct btrfs_block_rsv *rsv,
5024                                       u64 qgroup_reserved)
5025 {
5026         btrfs_block_rsv_release(root, rsv, (u64)-1);
5027         if (qgroup_reserved)
5028                 btrfs_qgroup_free(root, qgroup_reserved);
5029 }
5030
5031 /**
5032  * drop_outstanding_extent - drop an outstanding extent
5033  * @inode: the inode we're dropping the extent for
5034  *
5035  * This is called when we are freeing up an outstanding extent, either called
5036  * after an error or after an extent is written.  This will return the number of
5037  * reserved extents that need to be freed.  This must be called with
5038  * BTRFS_I(inode)->lock held.
5039  */
5040 static unsigned drop_outstanding_extent(struct inode *inode)
5041 {
5042         unsigned drop_inode_space = 0;
5043         unsigned dropped_extents = 0;
5044
5045         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
5046         BTRFS_I(inode)->outstanding_extents--;
5047
5048         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5049             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5050                                &BTRFS_I(inode)->runtime_flags))
5051                 drop_inode_space = 1;
5052
5053         /*
5054          * If we have more or the same amount of outsanding extents than we have
5055          * reserved then we need to leave the reserved extents count alone.
5056          */
5057         if (BTRFS_I(inode)->outstanding_extents >=
5058             BTRFS_I(inode)->reserved_extents)
5059                 return drop_inode_space;
5060
5061         dropped_extents = BTRFS_I(inode)->reserved_extents -
5062                 BTRFS_I(inode)->outstanding_extents;
5063         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5064         return dropped_extents + drop_inode_space;
5065 }
5066
5067 /**
5068  * calc_csum_metadata_size - return the amount of metada space that must be
5069  *      reserved/free'd for the given bytes.
5070  * @inode: the inode we're manipulating
5071  * @num_bytes: the number of bytes in question
5072  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5073  *
5074  * This adjusts the number of csum_bytes in the inode and then returns the
5075  * correct amount of metadata that must either be reserved or freed.  We
5076  * calculate how many checksums we can fit into one leaf and then divide the
5077  * number of bytes that will need to be checksumed by this value to figure out
5078  * how many checksums will be required.  If we are adding bytes then the number
5079  * may go up and we will return the number of additional bytes that must be
5080  * reserved.  If it is going down we will return the number of bytes that must
5081  * be freed.
5082  *
5083  * This must be called with BTRFS_I(inode)->lock held.
5084  */
5085 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5086                                    int reserve)
5087 {
5088         struct btrfs_root *root = BTRFS_I(inode)->root;
5089         u64 csum_size;
5090         int num_csums_per_leaf;
5091         int num_csums;
5092         int old_csums;
5093
5094         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5095             BTRFS_I(inode)->csum_bytes == 0)
5096                 return 0;
5097
5098         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
5099         if (reserve)
5100                 BTRFS_I(inode)->csum_bytes += num_bytes;
5101         else
5102                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5103         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
5104         num_csums_per_leaf = (int)div64_u64(csum_size,
5105                                             sizeof(struct btrfs_csum_item) +
5106                                             sizeof(struct btrfs_disk_key));
5107         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
5108         num_csums = num_csums + num_csums_per_leaf - 1;
5109         num_csums = num_csums / num_csums_per_leaf;
5110
5111         old_csums = old_csums + num_csums_per_leaf - 1;
5112         old_csums = old_csums / num_csums_per_leaf;
5113
5114         /* No change, no need to reserve more */
5115         if (old_csums == num_csums)
5116                 return 0;
5117
5118         if (reserve)
5119                 return btrfs_calc_trans_metadata_size(root,
5120                                                       num_csums - old_csums);
5121
5122         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5123 }
5124
5125 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5126 {
5127         struct btrfs_root *root = BTRFS_I(inode)->root;
5128         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5129         u64 to_reserve = 0;
5130         u64 csum_bytes;
5131         unsigned nr_extents = 0;
5132         int extra_reserve = 0;
5133         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5134         int ret = 0;
5135         bool delalloc_lock = true;
5136         u64 to_free = 0;
5137         unsigned dropped;
5138
5139         /* If we are a free space inode we need to not flush since we will be in
5140          * the middle of a transaction commit.  We also don't need the delalloc
5141          * mutex since we won't race with anybody.  We need this mostly to make
5142          * lockdep shut its filthy mouth.
5143          */
5144         if (btrfs_is_free_space_inode(inode)) {
5145                 flush = BTRFS_RESERVE_NO_FLUSH;
5146                 delalloc_lock = false;
5147         }
5148
5149         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5150             btrfs_transaction_in_commit(root->fs_info))
5151                 schedule_timeout(1);
5152
5153         if (delalloc_lock)
5154                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5155
5156         num_bytes = ALIGN(num_bytes, root->sectorsize);
5157
5158         spin_lock(&BTRFS_I(inode)->lock);
5159         BTRFS_I(inode)->outstanding_extents++;
5160
5161         if (BTRFS_I(inode)->outstanding_extents >
5162             BTRFS_I(inode)->reserved_extents)
5163                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5164                         BTRFS_I(inode)->reserved_extents;
5165
5166         /*
5167          * Add an item to reserve for updating the inode when we complete the
5168          * delalloc io.
5169          */
5170         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5171                       &BTRFS_I(inode)->runtime_flags)) {
5172                 nr_extents++;
5173                 extra_reserve = 1;
5174         }
5175
5176         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5177         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5178         csum_bytes = BTRFS_I(inode)->csum_bytes;
5179         spin_unlock(&BTRFS_I(inode)->lock);
5180
5181         if (root->fs_info->quota_enabled) {
5182                 ret = btrfs_qgroup_reserve(root, num_bytes +
5183                                            nr_extents * root->nodesize);
5184                 if (ret)
5185                         goto out_fail;
5186         }
5187
5188         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5189         if (unlikely(ret)) {
5190                 if (root->fs_info->quota_enabled)
5191                         btrfs_qgroup_free(root, num_bytes +
5192                                                 nr_extents * root->nodesize);
5193                 goto out_fail;
5194         }
5195
5196         spin_lock(&BTRFS_I(inode)->lock);
5197         if (extra_reserve) {
5198                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5199                         &BTRFS_I(inode)->runtime_flags);
5200                 nr_extents--;
5201         }
5202         BTRFS_I(inode)->reserved_extents += nr_extents;
5203         spin_unlock(&BTRFS_I(inode)->lock);
5204
5205         if (delalloc_lock)
5206                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5207
5208         if (to_reserve)
5209                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5210                                               btrfs_ino(inode), to_reserve, 1);
5211         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5212
5213         return 0;
5214
5215 out_fail:
5216         spin_lock(&BTRFS_I(inode)->lock);
5217         dropped = drop_outstanding_extent(inode);
5218         /*
5219          * If the inodes csum_bytes is the same as the original
5220          * csum_bytes then we know we haven't raced with any free()ers
5221          * so we can just reduce our inodes csum bytes and carry on.
5222          */
5223         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5224                 calc_csum_metadata_size(inode, num_bytes, 0);
5225         } else {
5226                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5227                 u64 bytes;
5228
5229                 /*
5230                  * This is tricky, but first we need to figure out how much we
5231                  * free'd from any free-ers that occured during this
5232                  * reservation, so we reset ->csum_bytes to the csum_bytes
5233                  * before we dropped our lock, and then call the free for the
5234                  * number of bytes that were freed while we were trying our
5235                  * reservation.
5236                  */
5237                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5238                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5239                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5240
5241
5242                 /*
5243                  * Now we need to see how much we would have freed had we not
5244                  * been making this reservation and our ->csum_bytes were not
5245                  * artificially inflated.
5246                  */
5247                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5248                 bytes = csum_bytes - orig_csum_bytes;
5249                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5250
5251                 /*
5252                  * Now reset ->csum_bytes to what it should be.  If bytes is
5253                  * more than to_free then we would have free'd more space had we
5254                  * not had an artificially high ->csum_bytes, so we need to free
5255                  * the remainder.  If bytes is the same or less then we don't
5256                  * need to do anything, the other free-ers did the correct
5257                  * thing.
5258                  */
5259                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5260                 if (bytes > to_free)
5261                         to_free = bytes - to_free;
5262                 else
5263                         to_free = 0;
5264         }
5265         spin_unlock(&BTRFS_I(inode)->lock);
5266         if (dropped)
5267                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5268
5269         if (to_free) {
5270                 btrfs_block_rsv_release(root, block_rsv, to_free);
5271                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5272                                               btrfs_ino(inode), to_free, 0);
5273         }
5274         if (delalloc_lock)
5275                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5276         return ret;
5277 }
5278
5279 /**
5280  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5281  * @inode: the inode to release the reservation for
5282  * @num_bytes: the number of bytes we're releasing
5283  *
5284  * This will release the metadata reservation for an inode.  This can be called
5285  * once we complete IO for a given set of bytes to release their metadata
5286  * reservations.
5287  */
5288 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5289 {
5290         struct btrfs_root *root = BTRFS_I(inode)->root;
5291         u64 to_free = 0;
5292         unsigned dropped;
5293
5294         num_bytes = ALIGN(num_bytes, root->sectorsize);
5295         spin_lock(&BTRFS_I(inode)->lock);
5296         dropped = drop_outstanding_extent(inode);
5297
5298         if (num_bytes)
5299                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5300         spin_unlock(&BTRFS_I(inode)->lock);
5301         if (dropped > 0)
5302                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5303
5304         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5305                                       btrfs_ino(inode), to_free, 0);
5306         if (root->fs_info->quota_enabled) {
5307                 btrfs_qgroup_free(root, num_bytes +
5308                                         dropped * root->nodesize);
5309         }
5310
5311         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5312                                 to_free);
5313 }
5314
5315 /**
5316  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5317  * @inode: inode we're writing to
5318  * @num_bytes: the number of bytes we want to allocate
5319  *
5320  * This will do the following things
5321  *
5322  * o reserve space in the data space info for num_bytes
5323  * o reserve space in the metadata space info based on number of outstanding
5324  *   extents and how much csums will be needed
5325  * o add to the inodes ->delalloc_bytes
5326  * o add it to the fs_info's delalloc inodes list.
5327  *
5328  * This will return 0 for success and -ENOSPC if there is no space left.
5329  */
5330 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5331 {
5332         int ret;
5333
5334         ret = btrfs_check_data_free_space(inode, num_bytes);
5335         if (ret)
5336                 return ret;
5337
5338         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5339         if (ret) {
5340                 btrfs_free_reserved_data_space(inode, num_bytes);
5341                 return ret;
5342         }
5343
5344         return 0;
5345 }
5346
5347 /**
5348  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5349  * @inode: inode we're releasing space for
5350  * @num_bytes: the number of bytes we want to free up
5351  *
5352  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5353  * called in the case that we don't need the metadata AND data reservations
5354  * anymore.  So if there is an error or we insert an inline extent.
5355  *
5356  * This function will release the metadata space that was not used and will
5357  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5358  * list if there are no delalloc bytes left.
5359  */
5360 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5361 {
5362         btrfs_delalloc_release_metadata(inode, num_bytes);
5363         btrfs_free_reserved_data_space(inode, num_bytes);
5364 }
5365
5366 static int update_block_group(struct btrfs_root *root,
5367                               u64 bytenr, u64 num_bytes, int alloc)
5368 {
5369         struct btrfs_block_group_cache *cache = NULL;
5370         struct btrfs_fs_info *info = root->fs_info;
5371         u64 total = num_bytes;
5372         u64 old_val;
5373         u64 byte_in_group;
5374         int factor;
5375
5376         /* block accounting for super block */
5377         spin_lock(&info->delalloc_root_lock);
5378         old_val = btrfs_super_bytes_used(info->super_copy);
5379         if (alloc)
5380                 old_val += num_bytes;
5381         else
5382                 old_val -= num_bytes;
5383         btrfs_set_super_bytes_used(info->super_copy, old_val);
5384         spin_unlock(&info->delalloc_root_lock);
5385
5386         while (total) {
5387                 cache = btrfs_lookup_block_group(info, bytenr);
5388                 if (!cache)
5389                         return -ENOENT;
5390                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5391                                     BTRFS_BLOCK_GROUP_RAID1 |
5392                                     BTRFS_BLOCK_GROUP_RAID10))
5393                         factor = 2;
5394                 else
5395                         factor = 1;
5396                 /*
5397                  * If this block group has free space cache written out, we
5398                  * need to make sure to load it if we are removing space.  This
5399                  * is because we need the unpinning stage to actually add the
5400                  * space back to the block group, otherwise we will leak space.
5401                  */
5402                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5403                         cache_block_group(cache, 1);
5404
5405                 byte_in_group = bytenr - cache->key.objectid;
5406                 WARN_ON(byte_in_group > cache->key.offset);
5407
5408                 spin_lock(&cache->space_info->lock);
5409                 spin_lock(&cache->lock);
5410
5411                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5412                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5413                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5414
5415                 cache->dirty = 1;
5416                 old_val = btrfs_block_group_used(&cache->item);
5417                 num_bytes = min(total, cache->key.offset - byte_in_group);
5418                 if (alloc) {
5419                         old_val += num_bytes;
5420                         btrfs_set_block_group_used(&cache->item, old_val);
5421                         cache->reserved -= num_bytes;
5422                         cache->space_info->bytes_reserved -= num_bytes;
5423                         cache->space_info->bytes_used += num_bytes;
5424                         cache->space_info->disk_used += num_bytes * factor;
5425                         spin_unlock(&cache->lock);
5426                         spin_unlock(&cache->space_info->lock);
5427                 } else {
5428                         old_val -= num_bytes;
5429
5430                         /*
5431                          * No longer have used bytes in this block group, queue
5432                          * it for deletion.
5433                          */
5434                         if (old_val == 0) {
5435                                 spin_lock(&info->unused_bgs_lock);
5436                                 if (list_empty(&cache->bg_list)) {
5437                                         btrfs_get_block_group(cache);
5438                                         list_add_tail(&cache->bg_list,
5439                                                       &info->unused_bgs);
5440                                 }
5441                                 spin_unlock(&info->unused_bgs_lock);
5442                         }
5443                         btrfs_set_block_group_used(&cache->item, old_val);
5444                         cache->pinned += num_bytes;
5445                         cache->space_info->bytes_pinned += num_bytes;
5446                         cache->space_info->bytes_used -= num_bytes;
5447                         cache->space_info->disk_used -= num_bytes * factor;
5448                         spin_unlock(&cache->lock);
5449                         spin_unlock(&cache->space_info->lock);
5450
5451                         set_extent_dirty(info->pinned_extents,
5452                                          bytenr, bytenr + num_bytes - 1,
5453                                          GFP_NOFS | __GFP_NOFAIL);
5454                 }
5455                 btrfs_put_block_group(cache);
5456                 total -= num_bytes;
5457                 bytenr += num_bytes;
5458         }
5459         return 0;
5460 }
5461
5462 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5463 {
5464         struct btrfs_block_group_cache *cache;
5465         u64 bytenr;
5466
5467         spin_lock(&root->fs_info->block_group_cache_lock);
5468         bytenr = root->fs_info->first_logical_byte;
5469         spin_unlock(&root->fs_info->block_group_cache_lock);
5470
5471         if (bytenr < (u64)-1)
5472                 return bytenr;
5473
5474         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5475         if (!cache)
5476                 return 0;
5477
5478         bytenr = cache->key.objectid;
5479         btrfs_put_block_group(cache);
5480
5481         return bytenr;
5482 }
5483
5484 static int pin_down_extent(struct btrfs_root *root,
5485                            struct btrfs_block_group_cache *cache,
5486                            u64 bytenr, u64 num_bytes, int reserved)
5487 {
5488         spin_lock(&cache->space_info->lock);
5489         spin_lock(&cache->lock);
5490         cache->pinned += num_bytes;
5491         cache->space_info->bytes_pinned += num_bytes;
5492         if (reserved) {
5493                 cache->reserved -= num_bytes;
5494                 cache->space_info->bytes_reserved -= num_bytes;
5495         }
5496         spin_unlock(&cache->lock);
5497         spin_unlock(&cache->space_info->lock);
5498
5499         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5500                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5501         if (reserved)
5502                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
5503         return 0;
5504 }
5505
5506 /*
5507  * this function must be called within transaction
5508  */
5509 int btrfs_pin_extent(struct btrfs_root *root,
5510                      u64 bytenr, u64 num_bytes, int reserved)
5511 {
5512         struct btrfs_block_group_cache *cache;
5513
5514         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5515         BUG_ON(!cache); /* Logic error */
5516
5517         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5518
5519         btrfs_put_block_group(cache);
5520         return 0;
5521 }
5522
5523 /*
5524  * this function must be called within transaction
5525  */
5526 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5527                                     u64 bytenr, u64 num_bytes)
5528 {
5529         struct btrfs_block_group_cache *cache;
5530         int ret;
5531
5532         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5533         if (!cache)
5534                 return -EINVAL;
5535
5536         /*
5537          * pull in the free space cache (if any) so that our pin
5538          * removes the free space from the cache.  We have load_only set
5539          * to one because the slow code to read in the free extents does check
5540          * the pinned extents.
5541          */
5542         cache_block_group(cache, 1);
5543
5544         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5545
5546         /* remove us from the free space cache (if we're there at all) */
5547         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5548         btrfs_put_block_group(cache);
5549         return ret;
5550 }
5551
5552 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5553 {
5554         int ret;
5555         struct btrfs_block_group_cache *block_group;
5556         struct btrfs_caching_control *caching_ctl;
5557
5558         block_group = btrfs_lookup_block_group(root->fs_info, start);
5559         if (!block_group)
5560                 return -EINVAL;
5561
5562         cache_block_group(block_group, 0);
5563         caching_ctl = get_caching_control(block_group);
5564
5565         if (!caching_ctl) {
5566                 /* Logic error */
5567                 BUG_ON(!block_group_cache_done(block_group));
5568                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5569         } else {
5570                 mutex_lock(&caching_ctl->mutex);
5571
5572                 if (start >= caching_ctl->progress) {
5573                         ret = add_excluded_extent(root, start, num_bytes);
5574                 } else if (start + num_bytes <= caching_ctl->progress) {
5575                         ret = btrfs_remove_free_space(block_group,
5576                                                       start, num_bytes);
5577                 } else {
5578                         num_bytes = caching_ctl->progress - start;
5579                         ret = btrfs_remove_free_space(block_group,
5580                                                       start, num_bytes);
5581                         if (ret)
5582                                 goto out_lock;
5583
5584                         num_bytes = (start + num_bytes) -
5585                                 caching_ctl->progress;
5586                         start = caching_ctl->progress;
5587                         ret = add_excluded_extent(root, start, num_bytes);
5588                 }
5589 out_lock:
5590                 mutex_unlock(&caching_ctl->mutex);
5591                 put_caching_control(caching_ctl);
5592         }
5593         btrfs_put_block_group(block_group);
5594         return ret;
5595 }
5596
5597 int btrfs_exclude_logged_extents(struct btrfs_root *log,
5598                                  struct extent_buffer *eb)
5599 {
5600         struct btrfs_file_extent_item *item;
5601         struct btrfs_key key;
5602         int found_type;
5603         int i;
5604
5605         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
5606                 return 0;
5607
5608         for (i = 0; i < btrfs_header_nritems(eb); i++) {
5609                 btrfs_item_key_to_cpu(eb, &key, i);
5610                 if (key.type != BTRFS_EXTENT_DATA_KEY)
5611                         continue;
5612                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
5613                 found_type = btrfs_file_extent_type(eb, item);
5614                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
5615                         continue;
5616                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
5617                         continue;
5618                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
5619                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
5620                 __exclude_logged_extent(log, key.objectid, key.offset);
5621         }
5622
5623         return 0;
5624 }
5625
5626 /**
5627  * btrfs_update_reserved_bytes - update the block_group and space info counters
5628  * @cache:      The cache we are manipulating
5629  * @num_bytes:  The number of bytes in question
5630  * @reserve:    One of the reservation enums
5631  * @delalloc:   The blocks are allocated for the delalloc write
5632  *
5633  * This is called by the allocator when it reserves space, or by somebody who is
5634  * freeing space that was never actually used on disk.  For example if you
5635  * reserve some space for a new leaf in transaction A and before transaction A
5636  * commits you free that leaf, you call this with reserve set to 0 in order to
5637  * clear the reservation.
5638  *
5639  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5640  * ENOSPC accounting.  For data we handle the reservation through clearing the
5641  * delalloc bits in the io_tree.  We have to do this since we could end up
5642  * allocating less disk space for the amount of data we have reserved in the
5643  * case of compression.
5644  *
5645  * If this is a reservation and the block group has become read only we cannot
5646  * make the reservation and return -EAGAIN, otherwise this function always
5647  * succeeds.
5648  */
5649 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5650                                        u64 num_bytes, int reserve, int delalloc)
5651 {
5652         struct btrfs_space_info *space_info = cache->space_info;
5653         int ret = 0;
5654
5655         spin_lock(&space_info->lock);
5656         spin_lock(&cache->lock);
5657         if (reserve != RESERVE_FREE) {
5658                 if (cache->ro) {
5659                         ret = -EAGAIN;
5660                 } else {
5661                         cache->reserved += num_bytes;
5662                         space_info->bytes_reserved += num_bytes;
5663                         if (reserve == RESERVE_ALLOC) {
5664                                 trace_btrfs_space_reservation(cache->fs_info,
5665                                                 "space_info", space_info->flags,
5666                                                 num_bytes, 0);
5667                                 space_info->bytes_may_use -= num_bytes;
5668                         }
5669
5670                         if (delalloc)
5671                                 cache->delalloc_bytes += num_bytes;
5672                 }
5673         } else {
5674                 if (cache->ro)
5675                         space_info->bytes_readonly += num_bytes;
5676                 cache->reserved -= num_bytes;
5677                 space_info->bytes_reserved -= num_bytes;
5678
5679                 if (delalloc)
5680                         cache->delalloc_bytes -= num_bytes;
5681         }
5682         spin_unlock(&cache->lock);
5683         spin_unlock(&space_info->lock);
5684         return ret;
5685 }
5686
5687 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5688                                 struct btrfs_root *root)
5689 {
5690         struct btrfs_fs_info *fs_info = root->fs_info;
5691         struct btrfs_caching_control *next;
5692         struct btrfs_caching_control *caching_ctl;
5693         struct btrfs_block_group_cache *cache;
5694
5695         down_write(&fs_info->commit_root_sem);
5696
5697         list_for_each_entry_safe(caching_ctl, next,
5698                                  &fs_info->caching_block_groups, list) {
5699                 cache = caching_ctl->block_group;
5700                 if (block_group_cache_done(cache)) {
5701                         cache->last_byte_to_unpin = (u64)-1;
5702                         list_del_init(&caching_ctl->list);
5703                         put_caching_control(caching_ctl);
5704                 } else {
5705                         cache->last_byte_to_unpin = caching_ctl->progress;
5706                 }
5707         }
5708
5709         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5710                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5711         else
5712                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5713
5714         up_write(&fs_info->commit_root_sem);
5715
5716         update_global_block_rsv(fs_info);
5717 }
5718
5719 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5720 {
5721         struct btrfs_fs_info *fs_info = root->fs_info;
5722         struct btrfs_block_group_cache *cache = NULL;
5723         struct btrfs_space_info *space_info;
5724         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5725         u64 len;
5726         bool readonly;
5727
5728         while (start <= end) {
5729                 readonly = false;
5730                 if (!cache ||
5731                     start >= cache->key.objectid + cache->key.offset) {
5732                         if (cache)
5733                                 btrfs_put_block_group(cache);
5734                         cache = btrfs_lookup_block_group(fs_info, start);
5735                         BUG_ON(!cache); /* Logic error */
5736                 }
5737
5738                 len = cache->key.objectid + cache->key.offset - start;
5739                 len = min(len, end + 1 - start);
5740
5741                 if (start < cache->last_byte_to_unpin) {
5742                         len = min(len, cache->last_byte_to_unpin - start);
5743                         btrfs_add_free_space(cache, start, len);
5744                 }
5745
5746                 start += len;
5747                 space_info = cache->space_info;
5748
5749                 spin_lock(&space_info->lock);
5750                 spin_lock(&cache->lock);
5751                 cache->pinned -= len;
5752                 space_info->bytes_pinned -= len;
5753                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
5754                 if (cache->ro) {
5755                         space_info->bytes_readonly += len;
5756                         readonly = true;
5757                 }
5758                 spin_unlock(&cache->lock);
5759                 if (!readonly && global_rsv->space_info == space_info) {
5760                         spin_lock(&global_rsv->lock);
5761                         if (!global_rsv->full) {
5762                                 len = min(len, global_rsv->size -
5763                                           global_rsv->reserved);
5764                                 global_rsv->reserved += len;
5765                                 space_info->bytes_may_use += len;
5766                                 if (global_rsv->reserved >= global_rsv->size)
5767                                         global_rsv->full = 1;
5768                         }
5769                         spin_unlock(&global_rsv->lock);
5770                 }
5771                 spin_unlock(&space_info->lock);
5772         }
5773
5774         if (cache)
5775                 btrfs_put_block_group(cache);
5776         return 0;
5777 }
5778
5779 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5780                                struct btrfs_root *root)
5781 {
5782         struct btrfs_fs_info *fs_info = root->fs_info;
5783         struct extent_io_tree *unpin;
5784         u64 start;
5785         u64 end;
5786         int ret;
5787
5788         if (trans->aborted)
5789                 return 0;
5790
5791         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5792                 unpin = &fs_info->freed_extents[1];
5793         else
5794                 unpin = &fs_info->freed_extents[0];
5795
5796         while (1) {
5797                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5798                                             EXTENT_DIRTY, NULL);
5799                 if (ret)
5800                         break;
5801
5802                 if (btrfs_test_opt(root, DISCARD))
5803                         ret = btrfs_discard_extent(root, start,
5804                                                    end + 1 - start, NULL);
5805
5806                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5807                 unpin_extent_range(root, start, end);
5808                 cond_resched();
5809         }
5810
5811         return 0;
5812 }
5813
5814 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
5815                              u64 owner, u64 root_objectid)
5816 {
5817         struct btrfs_space_info *space_info;
5818         u64 flags;
5819
5820         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5821                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
5822                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
5823                 else
5824                         flags = BTRFS_BLOCK_GROUP_METADATA;
5825         } else {
5826                 flags = BTRFS_BLOCK_GROUP_DATA;
5827         }
5828
5829         space_info = __find_space_info(fs_info, flags);
5830         BUG_ON(!space_info); /* Logic bug */
5831         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
5832 }
5833
5834
5835 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5836                                 struct btrfs_root *root,
5837                                 u64 bytenr, u64 num_bytes, u64 parent,
5838                                 u64 root_objectid, u64 owner_objectid,
5839                                 u64 owner_offset, int refs_to_drop,
5840                                 struct btrfs_delayed_extent_op *extent_op,
5841                                 int no_quota)
5842 {
5843         struct btrfs_key key;
5844         struct btrfs_path *path;
5845         struct btrfs_fs_info *info = root->fs_info;
5846         struct btrfs_root *extent_root = info->extent_root;
5847         struct extent_buffer *leaf;
5848         struct btrfs_extent_item *ei;
5849         struct btrfs_extent_inline_ref *iref;
5850         int ret;
5851         int is_data;
5852         int extent_slot = 0;
5853         int found_extent = 0;
5854         int num_to_del = 1;
5855         u32 item_size;
5856         u64 refs;
5857         int last_ref = 0;
5858         enum btrfs_qgroup_operation_type type = BTRFS_QGROUP_OPER_SUB_EXCL;
5859         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
5860                                                  SKINNY_METADATA);
5861
5862         if (!info->quota_enabled || !is_fstree(root_objectid))
5863                 no_quota = 1;
5864
5865         path = btrfs_alloc_path();
5866         if (!path)
5867                 return -ENOMEM;
5868
5869         path->reada = 1;
5870         path->leave_spinning = 1;
5871
5872         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5873         BUG_ON(!is_data && refs_to_drop != 1);
5874
5875         if (is_data)
5876                 skinny_metadata = 0;
5877
5878         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5879                                     bytenr, num_bytes, parent,
5880                                     root_objectid, owner_objectid,
5881                                     owner_offset);
5882         if (ret == 0) {
5883                 extent_slot = path->slots[0];
5884                 while (extent_slot >= 0) {
5885                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5886                                               extent_slot);
5887                         if (key.objectid != bytenr)
5888                                 break;
5889                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5890                             key.offset == num_bytes) {
5891                                 found_extent = 1;
5892                                 break;
5893                         }
5894                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
5895                             key.offset == owner_objectid) {
5896                                 found_extent = 1;
5897                                 break;
5898                         }
5899                         if (path->slots[0] - extent_slot > 5)
5900                                 break;
5901                         extent_slot--;
5902                 }
5903 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5904                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5905                 if (found_extent && item_size < sizeof(*ei))
5906                         found_extent = 0;
5907 #endif
5908                 if (!found_extent) {
5909                         BUG_ON(iref);
5910                         ret = remove_extent_backref(trans, extent_root, path,
5911                                                     NULL, refs_to_drop,
5912                                                     is_data, &last_ref);
5913                         if (ret) {
5914                                 btrfs_abort_transaction(trans, extent_root, ret);
5915                                 goto out;
5916                         }
5917                         btrfs_release_path(path);
5918                         path->leave_spinning = 1;
5919
5920                         key.objectid = bytenr;
5921                         key.type = BTRFS_EXTENT_ITEM_KEY;
5922                         key.offset = num_bytes;
5923
5924                         if (!is_data && skinny_metadata) {
5925                                 key.type = BTRFS_METADATA_ITEM_KEY;
5926                                 key.offset = owner_objectid;
5927                         }
5928
5929                         ret = btrfs_search_slot(trans, extent_root,
5930                                                 &key, path, -1, 1);
5931                         if (ret > 0 && skinny_metadata && path->slots[0]) {
5932                                 /*
5933                                  * Couldn't find our skinny metadata item,
5934                                  * see if we have ye olde extent item.
5935                                  */
5936                                 path->slots[0]--;
5937                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
5938                                                       path->slots[0]);
5939                                 if (key.objectid == bytenr &&
5940                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
5941                                     key.offset == num_bytes)
5942                                         ret = 0;
5943                         }
5944
5945                         if (ret > 0 && skinny_metadata) {
5946                                 skinny_metadata = false;
5947                                 key.objectid = bytenr;
5948                                 key.type = BTRFS_EXTENT_ITEM_KEY;
5949                                 key.offset = num_bytes;
5950                                 btrfs_release_path(path);
5951                                 ret = btrfs_search_slot(trans, extent_root,
5952                                                         &key, path, -1, 1);
5953                         }
5954
5955                         if (ret) {
5956                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5957                                         ret, bytenr);
5958                                 if (ret > 0)
5959                                         btrfs_print_leaf(extent_root,
5960                                                          path->nodes[0]);
5961                         }
5962                         if (ret < 0) {
5963                                 btrfs_abort_transaction(trans, extent_root, ret);
5964                                 goto out;
5965                         }
5966                         extent_slot = path->slots[0];
5967                 }
5968         } else if (WARN_ON(ret == -ENOENT)) {
5969                 btrfs_print_leaf(extent_root, path->nodes[0]);
5970                 btrfs_err(info,
5971                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
5972                         bytenr, parent, root_objectid, owner_objectid,
5973                         owner_offset);
5974                 btrfs_abort_transaction(trans, extent_root, ret);
5975                 goto out;
5976         } else {
5977                 btrfs_abort_transaction(trans, extent_root, ret);
5978                 goto out;
5979         }
5980
5981         leaf = path->nodes[0];
5982         item_size = btrfs_item_size_nr(leaf, extent_slot);
5983 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5984         if (item_size < sizeof(*ei)) {
5985                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5986                 ret = convert_extent_item_v0(trans, extent_root, path,
5987                                              owner_objectid, 0);
5988                 if (ret < 0) {
5989                         btrfs_abort_transaction(trans, extent_root, ret);
5990                         goto out;
5991                 }
5992
5993                 btrfs_release_path(path);
5994                 path->leave_spinning = 1;
5995
5996                 key.objectid = bytenr;
5997                 key.type = BTRFS_EXTENT_ITEM_KEY;
5998                 key.offset = num_bytes;
5999
6000                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6001                                         -1, 1);
6002                 if (ret) {
6003                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6004                                 ret, bytenr);
6005                         btrfs_print_leaf(extent_root, path->nodes[0]);
6006                 }
6007                 if (ret < 0) {
6008                         btrfs_abort_transaction(trans, extent_root, ret);
6009                         goto out;
6010                 }
6011
6012                 extent_slot = path->slots[0];
6013                 leaf = path->nodes[0];
6014                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6015         }
6016 #endif
6017         BUG_ON(item_size < sizeof(*ei));
6018         ei = btrfs_item_ptr(leaf, extent_slot,
6019                             struct btrfs_extent_item);
6020         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6021             key.type == BTRFS_EXTENT_ITEM_KEY) {
6022                 struct btrfs_tree_block_info *bi;
6023                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6024                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6025                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6026         }
6027
6028         refs = btrfs_extent_refs(leaf, ei);
6029         if (refs < refs_to_drop) {
6030                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6031                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6032                 ret = -EINVAL;
6033                 btrfs_abort_transaction(trans, extent_root, ret);
6034                 goto out;
6035         }
6036         refs -= refs_to_drop;
6037
6038         if (refs > 0) {
6039                 type = BTRFS_QGROUP_OPER_SUB_SHARED;
6040                 if (extent_op)
6041                         __run_delayed_extent_op(extent_op, leaf, ei);
6042                 /*
6043                  * In the case of inline back ref, reference count will
6044                  * be updated by remove_extent_backref
6045                  */
6046                 if (iref) {
6047                         BUG_ON(!found_extent);
6048                 } else {
6049                         btrfs_set_extent_refs(leaf, ei, refs);
6050                         btrfs_mark_buffer_dirty(leaf);
6051                 }
6052                 if (found_extent) {
6053                         ret = remove_extent_backref(trans, extent_root, path,
6054                                                     iref, refs_to_drop,
6055                                                     is_data, &last_ref);
6056                         if (ret) {
6057                                 btrfs_abort_transaction(trans, extent_root, ret);
6058                                 goto out;
6059                         }
6060                 }
6061                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6062                                  root_objectid);
6063         } else {
6064                 if (found_extent) {
6065                         BUG_ON(is_data && refs_to_drop !=
6066                                extent_data_ref_count(root, path, iref));
6067                         if (iref) {
6068                                 BUG_ON(path->slots[0] != extent_slot);
6069                         } else {
6070                                 BUG_ON(path->slots[0] != extent_slot + 1);
6071                                 path->slots[0] = extent_slot;
6072                                 num_to_del = 2;
6073                         }
6074                 }
6075
6076                 last_ref = 1;
6077                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6078                                       num_to_del);
6079                 if (ret) {
6080                         btrfs_abort_transaction(trans, extent_root, ret);
6081                         goto out;
6082                 }
6083                 btrfs_release_path(path);
6084
6085                 if (is_data) {
6086                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6087                         if (ret) {
6088                                 btrfs_abort_transaction(trans, extent_root, ret);
6089                                 goto out;
6090                         }
6091                 }
6092
6093                 ret = update_block_group(root, bytenr, num_bytes, 0);
6094                 if (ret) {
6095                         btrfs_abort_transaction(trans, extent_root, ret);
6096                         goto out;
6097                 }
6098         }
6099         btrfs_release_path(path);
6100
6101         /* Deal with the quota accounting */
6102         if (!ret && last_ref && !no_quota) {
6103                 int mod_seq = 0;
6104
6105                 if (owner_objectid >= BTRFS_FIRST_FREE_OBJECTID &&
6106                     type == BTRFS_QGROUP_OPER_SUB_SHARED)
6107                         mod_seq = 1;
6108
6109                 ret = btrfs_qgroup_record_ref(trans, info, root_objectid,
6110                                               bytenr, num_bytes, type,
6111                                               mod_seq);
6112         }
6113 out:
6114         btrfs_free_path(path);
6115         return ret;
6116 }
6117
6118 /*
6119  * when we free an block, it is possible (and likely) that we free the last
6120  * delayed ref for that extent as well.  This searches the delayed ref tree for
6121  * a given extent, and if there are no other delayed refs to be processed, it
6122  * removes it from the tree.
6123  */
6124 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6125                                       struct btrfs_root *root, u64 bytenr)
6126 {
6127         struct btrfs_delayed_ref_head *head;
6128         struct btrfs_delayed_ref_root *delayed_refs;
6129         int ret = 0;
6130
6131         delayed_refs = &trans->transaction->delayed_refs;
6132         spin_lock(&delayed_refs->lock);
6133         head = btrfs_find_delayed_ref_head(trans, bytenr);
6134         if (!head)
6135                 goto out_delayed_unlock;
6136
6137         spin_lock(&head->lock);
6138         if (rb_first(&head->ref_root))
6139                 goto out;
6140
6141         if (head->extent_op) {
6142                 if (!head->must_insert_reserved)
6143                         goto out;
6144                 btrfs_free_delayed_extent_op(head->extent_op);
6145                 head->extent_op = NULL;
6146         }
6147
6148         /*
6149          * waiting for the lock here would deadlock.  If someone else has it
6150          * locked they are already in the process of dropping it anyway
6151          */
6152         if (!mutex_trylock(&head->mutex))
6153                 goto out;
6154
6155         /*
6156          * at this point we have a head with no other entries.  Go
6157          * ahead and process it.
6158          */
6159         head->node.in_tree = 0;
6160         rb_erase(&head->href_node, &delayed_refs->href_root);
6161
6162         atomic_dec(&delayed_refs->num_entries);
6163
6164         /*
6165          * we don't take a ref on the node because we're removing it from the
6166          * tree, so we just steal the ref the tree was holding.
6167          */
6168         delayed_refs->num_heads--;
6169         if (head->processing == 0)
6170                 delayed_refs->num_heads_ready--;
6171         head->processing = 0;
6172         spin_unlock(&head->lock);
6173         spin_unlock(&delayed_refs->lock);
6174
6175         BUG_ON(head->extent_op);
6176         if (head->must_insert_reserved)
6177                 ret = 1;
6178
6179         mutex_unlock(&head->mutex);
6180         btrfs_put_delayed_ref(&head->node);
6181         return ret;
6182 out:
6183         spin_unlock(&head->lock);
6184
6185 out_delayed_unlock:
6186         spin_unlock(&delayed_refs->lock);
6187         return 0;
6188 }
6189
6190 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6191                            struct btrfs_root *root,
6192                            struct extent_buffer *buf,
6193                            u64 parent, int last_ref)
6194 {
6195         struct btrfs_block_group_cache *cache = NULL;
6196         int pin = 1;
6197         int ret;
6198
6199         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6200                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6201                                         buf->start, buf->len,
6202                                         parent, root->root_key.objectid,
6203                                         btrfs_header_level(buf),
6204                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
6205                 BUG_ON(ret); /* -ENOMEM */
6206         }
6207
6208         if (!last_ref)
6209                 return;
6210
6211         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6212
6213         if (btrfs_header_generation(buf) == trans->transid) {
6214                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6215                         ret = check_ref_cleanup(trans, root, buf->start);
6216                         if (!ret)
6217                                 goto out;
6218                 }
6219
6220                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6221                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6222                         goto out;
6223                 }
6224
6225                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6226
6227                 btrfs_add_free_space(cache, buf->start, buf->len);
6228                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
6229                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6230                 pin = 0;
6231         }
6232 out:
6233         if (pin)
6234                 add_pinned_bytes(root->fs_info, buf->len,
6235                                  btrfs_header_level(buf),
6236                                  root->root_key.objectid);
6237
6238         /*
6239          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6240          * anymore.
6241          */
6242         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6243         btrfs_put_block_group(cache);
6244 }
6245
6246 /* Can return -ENOMEM */
6247 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6248                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6249                       u64 owner, u64 offset, int no_quota)
6250 {
6251         int ret;
6252         struct btrfs_fs_info *fs_info = root->fs_info;
6253
6254         if (btrfs_test_is_dummy_root(root))
6255                 return 0;
6256
6257         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6258
6259         /*
6260          * tree log blocks never actually go into the extent allocation
6261          * tree, just update pinning info and exit early.
6262          */
6263         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6264                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6265                 /* unlocks the pinned mutex */
6266                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6267                 ret = 0;
6268         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6269                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6270                                         num_bytes,
6271                                         parent, root_objectid, (int)owner,
6272                                         BTRFS_DROP_DELAYED_REF, NULL, no_quota);
6273         } else {
6274                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6275                                                 num_bytes,
6276                                                 parent, root_objectid, owner,
6277                                                 offset, BTRFS_DROP_DELAYED_REF,
6278                                                 NULL, no_quota);
6279         }
6280         return ret;
6281 }
6282
6283 /*
6284  * when we wait for progress in the block group caching, its because
6285  * our allocation attempt failed at least once.  So, we must sleep
6286  * and let some progress happen before we try again.
6287  *
6288  * This function will sleep at least once waiting for new free space to
6289  * show up, and then it will check the block group free space numbers
6290  * for our min num_bytes.  Another option is to have it go ahead
6291  * and look in the rbtree for a free extent of a given size, but this
6292  * is a good start.
6293  *
6294  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6295  * any of the information in this block group.
6296  */
6297 static noinline void
6298 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6299                                 u64 num_bytes)
6300 {
6301         struct btrfs_caching_control *caching_ctl;
6302
6303         caching_ctl = get_caching_control(cache);
6304         if (!caching_ctl)
6305                 return;
6306
6307         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6308                    (cache->free_space_ctl->free_space >= num_bytes));
6309
6310         put_caching_control(caching_ctl);
6311 }
6312
6313 static noinline int
6314 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6315 {
6316         struct btrfs_caching_control *caching_ctl;
6317         int ret = 0;
6318
6319         caching_ctl = get_caching_control(cache);
6320         if (!caching_ctl)
6321                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6322
6323         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6324         if (cache->cached == BTRFS_CACHE_ERROR)
6325                 ret = -EIO;
6326         put_caching_control(caching_ctl);
6327         return ret;
6328 }
6329
6330 int __get_raid_index(u64 flags)
6331 {
6332         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6333                 return BTRFS_RAID_RAID10;
6334         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6335                 return BTRFS_RAID_RAID1;
6336         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6337                 return BTRFS_RAID_DUP;
6338         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6339                 return BTRFS_RAID_RAID0;
6340         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6341                 return BTRFS_RAID_RAID5;
6342         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6343                 return BTRFS_RAID_RAID6;
6344
6345         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6346 }
6347
6348 int get_block_group_index(struct btrfs_block_group_cache *cache)
6349 {
6350         return __get_raid_index(cache->flags);
6351 }
6352
6353 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6354         [BTRFS_RAID_RAID10]     = "raid10",
6355         [BTRFS_RAID_RAID1]      = "raid1",
6356         [BTRFS_RAID_DUP]        = "dup",
6357         [BTRFS_RAID_RAID0]      = "raid0",
6358         [BTRFS_RAID_SINGLE]     = "single",
6359         [BTRFS_RAID_RAID5]      = "raid5",
6360         [BTRFS_RAID_RAID6]      = "raid6",
6361 };
6362
6363 static const char *get_raid_name(enum btrfs_raid_types type)
6364 {
6365         if (type >= BTRFS_NR_RAID_TYPES)
6366                 return NULL;
6367
6368         return btrfs_raid_type_names[type];
6369 }
6370
6371 enum btrfs_loop_type {
6372         LOOP_CACHING_NOWAIT = 0,
6373         LOOP_CACHING_WAIT = 1,
6374         LOOP_ALLOC_CHUNK = 2,
6375         LOOP_NO_EMPTY_SIZE = 3,
6376 };
6377
6378 static inline void
6379 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
6380                        int delalloc)
6381 {
6382         if (delalloc)
6383                 down_read(&cache->data_rwsem);
6384 }
6385
6386 static inline void
6387 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
6388                        int delalloc)
6389 {
6390         btrfs_get_block_group(cache);
6391         if (delalloc)
6392                 down_read(&cache->data_rwsem);
6393 }
6394
6395 static struct btrfs_block_group_cache *
6396 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
6397                    struct btrfs_free_cluster *cluster,
6398                    int delalloc)
6399 {
6400         struct btrfs_block_group_cache *used_bg;
6401         bool locked = false;
6402 again:
6403         spin_lock(&cluster->refill_lock);
6404         if (locked) {
6405                 if (used_bg == cluster->block_group)
6406                         return used_bg;
6407
6408                 up_read(&used_bg->data_rwsem);
6409                 btrfs_put_block_group(used_bg);
6410         }
6411
6412         used_bg = cluster->block_group;
6413         if (!used_bg)
6414                 return NULL;
6415
6416         if (used_bg == block_group)
6417                 return used_bg;
6418
6419         btrfs_get_block_group(used_bg);
6420
6421         if (!delalloc)
6422                 return used_bg;
6423
6424         if (down_read_trylock(&used_bg->data_rwsem))
6425                 return used_bg;
6426
6427         spin_unlock(&cluster->refill_lock);
6428         down_read(&used_bg->data_rwsem);
6429         locked = true;
6430         goto again;
6431 }
6432
6433 static inline void
6434 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
6435                          int delalloc)
6436 {
6437         if (delalloc)
6438                 up_read(&cache->data_rwsem);
6439         btrfs_put_block_group(cache);
6440 }
6441
6442 /*
6443  * walks the btree of allocated extents and find a hole of a given size.
6444  * The key ins is changed to record the hole:
6445  * ins->objectid == start position
6446  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6447  * ins->offset == the size of the hole.
6448  * Any available blocks before search_start are skipped.
6449  *
6450  * If there is no suitable free space, we will record the max size of
6451  * the free space extent currently.
6452  */
6453 static noinline int find_free_extent(struct btrfs_root *orig_root,
6454                                      u64 num_bytes, u64 empty_size,
6455                                      u64 hint_byte, struct btrfs_key *ins,
6456                                      u64 flags, int delalloc)
6457 {
6458         int ret = 0;
6459         struct btrfs_root *root = orig_root->fs_info->extent_root;
6460         struct btrfs_free_cluster *last_ptr = NULL;
6461         struct btrfs_block_group_cache *block_group = NULL;
6462         u64 search_start = 0;
6463         u64 max_extent_size = 0;
6464         int empty_cluster = 2 * 1024 * 1024;
6465         struct btrfs_space_info *space_info;
6466         int loop = 0;
6467         int index = __get_raid_index(flags);
6468         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6469                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6470         bool failed_cluster_refill = false;
6471         bool failed_alloc = false;
6472         bool use_cluster = true;
6473         bool have_caching_bg = false;
6474
6475         WARN_ON(num_bytes < root->sectorsize);
6476         ins->type = BTRFS_EXTENT_ITEM_KEY;
6477         ins->objectid = 0;
6478         ins->offset = 0;
6479
6480         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6481
6482         space_info = __find_space_info(root->fs_info, flags);
6483         if (!space_info) {
6484                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6485                 return -ENOSPC;
6486         }
6487
6488         /*
6489          * If the space info is for both data and metadata it means we have a
6490          * small filesystem and we can't use the clustering stuff.
6491          */
6492         if (btrfs_mixed_space_info(space_info))
6493                 use_cluster = false;
6494
6495         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6496                 last_ptr = &root->fs_info->meta_alloc_cluster;
6497                 if (!btrfs_test_opt(root, SSD))
6498                         empty_cluster = 64 * 1024;
6499         }
6500
6501         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6502             btrfs_test_opt(root, SSD)) {
6503                 last_ptr = &root->fs_info->data_alloc_cluster;
6504         }
6505
6506         if (last_ptr) {
6507                 spin_lock(&last_ptr->lock);
6508                 if (last_ptr->block_group)
6509                         hint_byte = last_ptr->window_start;
6510                 spin_unlock(&last_ptr->lock);
6511         }
6512
6513         search_start = max(search_start, first_logical_byte(root, 0));
6514         search_start = max(search_start, hint_byte);
6515
6516         if (!last_ptr)
6517                 empty_cluster = 0;
6518
6519         if (search_start == hint_byte) {
6520                 block_group = btrfs_lookup_block_group(root->fs_info,
6521                                                        search_start);
6522                 /*
6523                  * we don't want to use the block group if it doesn't match our
6524                  * allocation bits, or if its not cached.
6525                  *
6526                  * However if we are re-searching with an ideal block group
6527                  * picked out then we don't care that the block group is cached.
6528                  */
6529                 if (block_group && block_group_bits(block_group, flags) &&
6530                     block_group->cached != BTRFS_CACHE_NO) {
6531                         down_read(&space_info->groups_sem);
6532                         if (list_empty(&block_group->list) ||
6533                             block_group->ro) {
6534                                 /*
6535                                  * someone is removing this block group,
6536                                  * we can't jump into the have_block_group
6537                                  * target because our list pointers are not
6538                                  * valid
6539                                  */
6540                                 btrfs_put_block_group(block_group);
6541                                 up_read(&space_info->groups_sem);
6542                         } else {
6543                                 index = get_block_group_index(block_group);
6544                                 btrfs_lock_block_group(block_group, delalloc);
6545                                 goto have_block_group;
6546                         }
6547                 } else if (block_group) {
6548                         btrfs_put_block_group(block_group);
6549                 }
6550         }
6551 search:
6552         have_caching_bg = false;
6553         down_read(&space_info->groups_sem);
6554         list_for_each_entry(block_group, &space_info->block_groups[index],
6555                             list) {
6556                 u64 offset;
6557                 int cached;
6558
6559                 btrfs_grab_block_group(block_group, delalloc);
6560                 search_start = block_group->key.objectid;
6561
6562                 /*
6563                  * this can happen if we end up cycling through all the
6564                  * raid types, but we want to make sure we only allocate
6565                  * for the proper type.
6566                  */
6567                 if (!block_group_bits(block_group, flags)) {
6568                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6569                                 BTRFS_BLOCK_GROUP_RAID1 |
6570                                 BTRFS_BLOCK_GROUP_RAID5 |
6571                                 BTRFS_BLOCK_GROUP_RAID6 |
6572                                 BTRFS_BLOCK_GROUP_RAID10;
6573
6574                         /*
6575                          * if they asked for extra copies and this block group
6576                          * doesn't provide them, bail.  This does allow us to
6577                          * fill raid0 from raid1.
6578                          */
6579                         if ((flags & extra) && !(block_group->flags & extra))
6580                                 goto loop;
6581                 }
6582
6583 have_block_group:
6584                 cached = block_group_cache_done(block_group);
6585                 if (unlikely(!cached)) {
6586                         ret = cache_block_group(block_group, 0);
6587                         BUG_ON(ret < 0);
6588                         ret = 0;
6589                 }
6590
6591                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
6592                         goto loop;
6593                 if (unlikely(block_group->ro))
6594                         goto loop;
6595
6596                 /*
6597                  * Ok we want to try and use the cluster allocator, so
6598                  * lets look there
6599                  */
6600                 if (last_ptr) {
6601                         struct btrfs_block_group_cache *used_block_group;
6602                         unsigned long aligned_cluster;
6603                         /*
6604                          * the refill lock keeps out other
6605                          * people trying to start a new cluster
6606                          */
6607                         used_block_group = btrfs_lock_cluster(block_group,
6608                                                               last_ptr,
6609                                                               delalloc);
6610                         if (!used_block_group)
6611                                 goto refill_cluster;
6612
6613                         if (used_block_group != block_group &&
6614                             (used_block_group->ro ||
6615                              !block_group_bits(used_block_group, flags)))
6616                                 goto release_cluster;
6617
6618                         offset = btrfs_alloc_from_cluster(used_block_group,
6619                                                 last_ptr,
6620                                                 num_bytes,
6621                                                 used_block_group->key.objectid,
6622                                                 &max_extent_size);
6623                         if (offset) {
6624                                 /* we have a block, we're done */
6625                                 spin_unlock(&last_ptr->refill_lock);
6626                                 trace_btrfs_reserve_extent_cluster(root,
6627                                                 used_block_group,
6628                                                 search_start, num_bytes);
6629                                 if (used_block_group != block_group) {
6630                                         btrfs_release_block_group(block_group,
6631                                                                   delalloc);
6632                                         block_group = used_block_group;
6633                                 }
6634                                 goto checks;
6635                         }
6636
6637                         WARN_ON(last_ptr->block_group != used_block_group);
6638 release_cluster:
6639                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6640                          * set up a new clusters, so lets just skip it
6641                          * and let the allocator find whatever block
6642                          * it can find.  If we reach this point, we
6643                          * will have tried the cluster allocator
6644                          * plenty of times and not have found
6645                          * anything, so we are likely way too
6646                          * fragmented for the clustering stuff to find
6647                          * anything.
6648                          *
6649                          * However, if the cluster is taken from the
6650                          * current block group, release the cluster
6651                          * first, so that we stand a better chance of
6652                          * succeeding in the unclustered
6653                          * allocation.  */
6654                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6655                             used_block_group != block_group) {
6656                                 spin_unlock(&last_ptr->refill_lock);
6657                                 btrfs_release_block_group(used_block_group,
6658                                                           delalloc);
6659                                 goto unclustered_alloc;
6660                         }
6661
6662                         /*
6663                          * this cluster didn't work out, free it and
6664                          * start over
6665                          */
6666                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6667
6668                         if (used_block_group != block_group)
6669                                 btrfs_release_block_group(used_block_group,
6670                                                           delalloc);
6671 refill_cluster:
6672                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6673                                 spin_unlock(&last_ptr->refill_lock);
6674                                 goto unclustered_alloc;
6675                         }
6676
6677                         aligned_cluster = max_t(unsigned long,
6678                                                 empty_cluster + empty_size,
6679                                               block_group->full_stripe_len);
6680
6681                         /* allocate a cluster in this block group */
6682                         ret = btrfs_find_space_cluster(root, block_group,
6683                                                        last_ptr, search_start,
6684                                                        num_bytes,
6685                                                        aligned_cluster);
6686                         if (ret == 0) {
6687                                 /*
6688                                  * now pull our allocation out of this
6689                                  * cluster
6690                                  */
6691                                 offset = btrfs_alloc_from_cluster(block_group,
6692                                                         last_ptr,
6693                                                         num_bytes,
6694                                                         search_start,
6695                                                         &max_extent_size);
6696                                 if (offset) {
6697                                         /* we found one, proceed */
6698                                         spin_unlock(&last_ptr->refill_lock);
6699                                         trace_btrfs_reserve_extent_cluster(root,
6700                                                 block_group, search_start,
6701                                                 num_bytes);
6702                                         goto checks;
6703                                 }
6704                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6705                                    && !failed_cluster_refill) {
6706                                 spin_unlock(&last_ptr->refill_lock);
6707
6708                                 failed_cluster_refill = true;
6709                                 wait_block_group_cache_progress(block_group,
6710                                        num_bytes + empty_cluster + empty_size);
6711                                 goto have_block_group;
6712                         }
6713
6714                         /*
6715                          * at this point we either didn't find a cluster
6716                          * or we weren't able to allocate a block from our
6717                          * cluster.  Free the cluster we've been trying
6718                          * to use, and go to the next block group
6719                          */
6720                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6721                         spin_unlock(&last_ptr->refill_lock);
6722                         goto loop;
6723                 }
6724
6725 unclustered_alloc:
6726                 spin_lock(&block_group->free_space_ctl->tree_lock);
6727                 if (cached &&
6728                     block_group->free_space_ctl->free_space <
6729                     num_bytes + empty_cluster + empty_size) {
6730                         if (block_group->free_space_ctl->free_space >
6731                             max_extent_size)
6732                                 max_extent_size =
6733                                         block_group->free_space_ctl->free_space;
6734                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6735                         goto loop;
6736                 }
6737                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6738
6739                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6740                                                     num_bytes, empty_size,
6741                                                     &max_extent_size);
6742                 /*
6743                  * If we didn't find a chunk, and we haven't failed on this
6744                  * block group before, and this block group is in the middle of
6745                  * caching and we are ok with waiting, then go ahead and wait
6746                  * for progress to be made, and set failed_alloc to true.
6747                  *
6748                  * If failed_alloc is true then we've already waited on this
6749                  * block group once and should move on to the next block group.
6750                  */
6751                 if (!offset && !failed_alloc && !cached &&
6752                     loop > LOOP_CACHING_NOWAIT) {
6753                         wait_block_group_cache_progress(block_group,
6754                                                 num_bytes + empty_size);
6755                         failed_alloc = true;
6756                         goto have_block_group;
6757                 } else if (!offset) {
6758                         if (!cached)
6759                                 have_caching_bg = true;
6760                         goto loop;
6761                 }
6762 checks:
6763                 search_start = ALIGN(offset, root->stripesize);
6764
6765                 /* move on to the next group */
6766                 if (search_start + num_bytes >
6767                     block_group->key.objectid + block_group->key.offset) {
6768                         btrfs_add_free_space(block_group, offset, num_bytes);
6769                         goto loop;
6770                 }
6771
6772                 if (offset < search_start)
6773                         btrfs_add_free_space(block_group, offset,
6774                                              search_start - offset);
6775                 BUG_ON(offset > search_start);
6776
6777                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
6778                                                   alloc_type, delalloc);
6779                 if (ret == -EAGAIN) {
6780                         btrfs_add_free_space(block_group, offset, num_bytes);
6781                         goto loop;
6782                 }
6783
6784                 /* we are all good, lets return */
6785                 ins->objectid = search_start;
6786                 ins->offset = num_bytes;
6787
6788                 trace_btrfs_reserve_extent(orig_root, block_group,
6789                                            search_start, num_bytes);
6790                 btrfs_release_block_group(block_group, delalloc);
6791                 break;
6792 loop:
6793                 failed_cluster_refill = false;
6794                 failed_alloc = false;
6795                 BUG_ON(index != get_block_group_index(block_group));
6796                 btrfs_release_block_group(block_group, delalloc);
6797         }
6798         up_read(&space_info->groups_sem);
6799
6800         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6801                 goto search;
6802
6803         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6804                 goto search;
6805
6806         /*
6807          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6808          *                      caching kthreads as we move along
6809          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6810          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6811          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6812          *                      again
6813          */
6814         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6815                 index = 0;
6816                 loop++;
6817                 if (loop == LOOP_ALLOC_CHUNK) {
6818                         struct btrfs_trans_handle *trans;
6819                         int exist = 0;
6820
6821                         trans = current->journal_info;
6822                         if (trans)
6823                                 exist = 1;
6824                         else
6825                                 trans = btrfs_join_transaction(root);
6826
6827                         if (IS_ERR(trans)) {
6828                                 ret = PTR_ERR(trans);
6829                                 goto out;
6830                         }
6831
6832                         ret = do_chunk_alloc(trans, root, flags,
6833                                              CHUNK_ALLOC_FORCE);
6834                         /*
6835                          * Do not bail out on ENOSPC since we
6836                          * can do more things.
6837                          */
6838                         if (ret < 0 && ret != -ENOSPC)
6839                                 btrfs_abort_transaction(trans,
6840                                                         root, ret);
6841                         else
6842                                 ret = 0;
6843                         if (!exist)
6844                                 btrfs_end_transaction(trans, root);
6845                         if (ret)
6846                                 goto out;
6847                 }
6848
6849                 if (loop == LOOP_NO_EMPTY_SIZE) {
6850                         empty_size = 0;
6851                         empty_cluster = 0;
6852                 }
6853
6854                 goto search;
6855         } else if (!ins->objectid) {
6856                 ret = -ENOSPC;
6857         } else if (ins->objectid) {
6858                 ret = 0;
6859         }
6860 out:
6861         if (ret == -ENOSPC)
6862                 ins->offset = max_extent_size;
6863         return ret;
6864 }
6865
6866 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6867                             int dump_block_groups)
6868 {
6869         struct btrfs_block_group_cache *cache;
6870         int index = 0;
6871
6872         spin_lock(&info->lock);
6873         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
6874                info->flags,
6875                info->total_bytes - info->bytes_used - info->bytes_pinned -
6876                info->bytes_reserved - info->bytes_readonly,
6877                (info->full) ? "" : "not ");
6878         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
6879                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6880                info->total_bytes, info->bytes_used, info->bytes_pinned,
6881                info->bytes_reserved, info->bytes_may_use,
6882                info->bytes_readonly);
6883         spin_unlock(&info->lock);
6884
6885         if (!dump_block_groups)
6886                 return;
6887
6888         down_read(&info->groups_sem);
6889 again:
6890         list_for_each_entry(cache, &info->block_groups[index], list) {
6891                 spin_lock(&cache->lock);
6892                 printk(KERN_INFO "BTRFS: "
6893                            "block group %llu has %llu bytes, "
6894                            "%llu used %llu pinned %llu reserved %s\n",
6895                        cache->key.objectid, cache->key.offset,
6896                        btrfs_block_group_used(&cache->item), cache->pinned,
6897                        cache->reserved, cache->ro ? "[readonly]" : "");
6898                 btrfs_dump_free_space(cache, bytes);
6899                 spin_unlock(&cache->lock);
6900         }
6901         if (++index < BTRFS_NR_RAID_TYPES)
6902                 goto again;
6903         up_read(&info->groups_sem);
6904 }
6905
6906 int btrfs_reserve_extent(struct btrfs_root *root,
6907                          u64 num_bytes, u64 min_alloc_size,
6908                          u64 empty_size, u64 hint_byte,
6909                          struct btrfs_key *ins, int is_data, int delalloc)
6910 {
6911         bool final_tried = false;
6912         u64 flags;
6913         int ret;
6914
6915         flags = btrfs_get_alloc_profile(root, is_data);
6916 again:
6917         WARN_ON(num_bytes < root->sectorsize);
6918         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
6919                                flags, delalloc);
6920
6921         if (ret == -ENOSPC) {
6922                 if (!final_tried && ins->offset) {
6923                         num_bytes = min(num_bytes >> 1, ins->offset);
6924                         num_bytes = round_down(num_bytes, root->sectorsize);
6925                         num_bytes = max(num_bytes, min_alloc_size);
6926                         if (num_bytes == min_alloc_size)
6927                                 final_tried = true;
6928                         goto again;
6929                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6930                         struct btrfs_space_info *sinfo;
6931
6932                         sinfo = __find_space_info(root->fs_info, flags);
6933                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
6934                                 flags, num_bytes);
6935                         if (sinfo)
6936                                 dump_space_info(sinfo, num_bytes, 1);
6937                 }
6938         }
6939
6940         return ret;
6941 }
6942
6943 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6944                                         u64 start, u64 len,
6945                                         int pin, int delalloc)
6946 {
6947         struct btrfs_block_group_cache *cache;
6948         int ret = 0;
6949
6950         cache = btrfs_lookup_block_group(root->fs_info, start);
6951         if (!cache) {
6952                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
6953                         start);
6954                 return -ENOSPC;
6955         }
6956
6957         if (btrfs_test_opt(root, DISCARD))
6958                 ret = btrfs_discard_extent(root, start, len, NULL);
6959
6960         if (pin)
6961                 pin_down_extent(root, cache, start, len, 1);
6962         else {
6963                 btrfs_add_free_space(cache, start, len);
6964                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
6965         }
6966         btrfs_put_block_group(cache);
6967
6968         trace_btrfs_reserved_extent_free(root, start, len);
6969
6970         return ret;
6971 }
6972
6973 int btrfs_free_reserved_extent(struct btrfs_root *root,
6974                                u64 start, u64 len, int delalloc)
6975 {
6976         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
6977 }
6978
6979 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6980                                        u64 start, u64 len)
6981 {
6982         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
6983 }
6984
6985 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6986                                       struct btrfs_root *root,
6987                                       u64 parent, u64 root_objectid,
6988                                       u64 flags, u64 owner, u64 offset,
6989                                       struct btrfs_key *ins, int ref_mod)
6990 {
6991         int ret;
6992         struct btrfs_fs_info *fs_info = root->fs_info;
6993         struct btrfs_extent_item *extent_item;
6994         struct btrfs_extent_inline_ref *iref;
6995         struct btrfs_path *path;
6996         struct extent_buffer *leaf;
6997         int type;
6998         u32 size;
6999
7000         if (parent > 0)
7001                 type = BTRFS_SHARED_DATA_REF_KEY;
7002         else
7003                 type = BTRFS_EXTENT_DATA_REF_KEY;
7004
7005         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7006
7007         path = btrfs_alloc_path();
7008         if (!path)
7009                 return -ENOMEM;
7010
7011         path->leave_spinning = 1;
7012         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7013                                       ins, size);
7014         if (ret) {
7015                 btrfs_free_path(path);
7016                 return ret;
7017         }
7018
7019         leaf = path->nodes[0];
7020         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7021                                      struct btrfs_extent_item);
7022         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7023         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7024         btrfs_set_extent_flags(leaf, extent_item,
7025                                flags | BTRFS_EXTENT_FLAG_DATA);
7026
7027         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7028         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7029         if (parent > 0) {
7030                 struct btrfs_shared_data_ref *ref;
7031                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7032                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7033                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7034         } else {
7035                 struct btrfs_extent_data_ref *ref;
7036                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7037                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7038                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7039                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7040                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7041         }
7042
7043         btrfs_mark_buffer_dirty(path->nodes[0]);
7044         btrfs_free_path(path);
7045
7046         /* Always set parent to 0 here since its exclusive anyway. */
7047         ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
7048                                       ins->objectid, ins->offset,
7049                                       BTRFS_QGROUP_OPER_ADD_EXCL, 0);
7050         if (ret)
7051                 return ret;
7052
7053         ret = update_block_group(root, ins->objectid, ins->offset, 1);
7054         if (ret) { /* -ENOENT, logic error */
7055                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7056                         ins->objectid, ins->offset);
7057                 BUG();
7058         }
7059         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
7060         return ret;
7061 }
7062
7063 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7064                                      struct btrfs_root *root,
7065                                      u64 parent, u64 root_objectid,
7066                                      u64 flags, struct btrfs_disk_key *key,
7067                                      int level, struct btrfs_key *ins,
7068                                      int no_quota)
7069 {
7070         int ret;
7071         struct btrfs_fs_info *fs_info = root->fs_info;
7072         struct btrfs_extent_item *extent_item;
7073         struct btrfs_tree_block_info *block_info;
7074         struct btrfs_extent_inline_ref *iref;
7075         struct btrfs_path *path;
7076         struct extent_buffer *leaf;
7077         u32 size = sizeof(*extent_item) + sizeof(*iref);
7078         u64 num_bytes = ins->offset;
7079         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7080                                                  SKINNY_METADATA);
7081
7082         if (!skinny_metadata)
7083                 size += sizeof(*block_info);
7084
7085         path = btrfs_alloc_path();
7086         if (!path) {
7087                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7088                                                    root->nodesize);
7089                 return -ENOMEM;
7090         }
7091
7092         path->leave_spinning = 1;
7093         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7094                                       ins, size);
7095         if (ret) {
7096                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7097                                                    root->nodesize);
7098                 btrfs_free_path(path);
7099                 return ret;
7100         }
7101
7102         leaf = path->nodes[0];
7103         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7104                                      struct btrfs_extent_item);
7105         btrfs_set_extent_refs(leaf, extent_item, 1);
7106         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7107         btrfs_set_extent_flags(leaf, extent_item,
7108                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7109
7110         if (skinny_metadata) {
7111                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7112                 num_bytes = root->nodesize;
7113         } else {
7114                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7115                 btrfs_set_tree_block_key(leaf, block_info, key);
7116                 btrfs_set_tree_block_level(leaf, block_info, level);
7117                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7118         }
7119
7120         if (parent > 0) {
7121                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7122                 btrfs_set_extent_inline_ref_type(leaf, iref,
7123                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7124                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7125         } else {
7126                 btrfs_set_extent_inline_ref_type(leaf, iref,
7127                                                  BTRFS_TREE_BLOCK_REF_KEY);
7128                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7129         }
7130
7131         btrfs_mark_buffer_dirty(leaf);
7132         btrfs_free_path(path);
7133
7134         if (!no_quota) {
7135                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
7136                                               ins->objectid, num_bytes,
7137                                               BTRFS_QGROUP_OPER_ADD_EXCL, 0);
7138                 if (ret)
7139                         return ret;
7140         }
7141
7142         ret = update_block_group(root, ins->objectid, root->nodesize, 1);
7143         if (ret) { /* -ENOENT, logic error */
7144                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7145                         ins->objectid, ins->offset);
7146                 BUG();
7147         }
7148
7149         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
7150         return ret;
7151 }
7152
7153 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7154                                      struct btrfs_root *root,
7155                                      u64 root_objectid, u64 owner,
7156                                      u64 offset, struct btrfs_key *ins)
7157 {
7158         int ret;
7159
7160         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7161
7162         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7163                                          ins->offset, 0,
7164                                          root_objectid, owner, offset,
7165                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
7166         return ret;
7167 }
7168
7169 /*
7170  * this is used by the tree logging recovery code.  It records that
7171  * an extent has been allocated and makes sure to clear the free
7172  * space cache bits as well
7173  */
7174 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7175                                    struct btrfs_root *root,
7176                                    u64 root_objectid, u64 owner, u64 offset,
7177                                    struct btrfs_key *ins)
7178 {
7179         int ret;
7180         struct btrfs_block_group_cache *block_group;
7181
7182         /*
7183          * Mixed block groups will exclude before processing the log so we only
7184          * need to do the exlude dance if this fs isn't mixed.
7185          */
7186         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7187                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7188                 if (ret)
7189                         return ret;
7190         }
7191
7192         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
7193         if (!block_group)
7194                 return -EINVAL;
7195
7196         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
7197                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
7198         BUG_ON(ret); /* logic error */
7199         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
7200                                          0, owner, offset, ins, 1);
7201         btrfs_put_block_group(block_group);
7202         return ret;
7203 }
7204
7205 static struct extent_buffer *
7206 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7207                       u64 bytenr, u32 blocksize, int level)
7208 {
7209         struct extent_buffer *buf;
7210
7211         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
7212         if (!buf)
7213                 return ERR_PTR(-ENOMEM);
7214         btrfs_set_header_generation(buf, trans->transid);
7215         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
7216         btrfs_tree_lock(buf);
7217         clean_tree_block(trans, root, buf);
7218         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
7219
7220         btrfs_set_lock_blocking(buf);
7221         btrfs_set_buffer_uptodate(buf);
7222
7223         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
7224                 buf->log_index = root->log_transid % 2;
7225                 /*
7226                  * we allow two log transactions at a time, use different
7227                  * EXENT bit to differentiate dirty pages.
7228                  */
7229                 if (buf->log_index == 0)
7230                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7231                                         buf->start + buf->len - 1, GFP_NOFS);
7232                 else
7233                         set_extent_new(&root->dirty_log_pages, buf->start,
7234                                         buf->start + buf->len - 1, GFP_NOFS);
7235         } else {
7236                 buf->log_index = -1;
7237                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7238                          buf->start + buf->len - 1, GFP_NOFS);
7239         }
7240         trans->blocks_used++;
7241         /* this returns a buffer locked for blocking */
7242         return buf;
7243 }
7244
7245 static struct btrfs_block_rsv *
7246 use_block_rsv(struct btrfs_trans_handle *trans,
7247               struct btrfs_root *root, u32 blocksize)
7248 {
7249         struct btrfs_block_rsv *block_rsv;
7250         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7251         int ret;
7252         bool global_updated = false;
7253
7254         block_rsv = get_block_rsv(trans, root);
7255
7256         if (unlikely(block_rsv->size == 0))
7257                 goto try_reserve;
7258 again:
7259         ret = block_rsv_use_bytes(block_rsv, blocksize);
7260         if (!ret)
7261                 return block_rsv;
7262
7263         if (block_rsv->failfast)
7264                 return ERR_PTR(ret);
7265
7266         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7267                 global_updated = true;
7268                 update_global_block_rsv(root->fs_info);
7269                 goto again;
7270         }
7271
7272         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7273                 static DEFINE_RATELIMIT_STATE(_rs,
7274                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7275                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7276                 if (__ratelimit(&_rs))
7277                         WARN(1, KERN_DEBUG
7278                                 "BTRFS: block rsv returned %d\n", ret);
7279         }
7280 try_reserve:
7281         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7282                                      BTRFS_RESERVE_NO_FLUSH);
7283         if (!ret)
7284                 return block_rsv;
7285         /*
7286          * If we couldn't reserve metadata bytes try and use some from
7287          * the global reserve if its space type is the same as the global
7288          * reservation.
7289          */
7290         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7291             block_rsv->space_info == global_rsv->space_info) {
7292                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7293                 if (!ret)
7294                         return global_rsv;
7295         }
7296         return ERR_PTR(ret);
7297 }
7298
7299 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7300                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7301 {
7302         block_rsv_add_bytes(block_rsv, blocksize, 0);
7303         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7304 }
7305
7306 /*
7307  * finds a free extent and does all the dirty work required for allocation
7308  * returns the key for the extent through ins, and a tree buffer for
7309  * the first block of the extent through buf.
7310  *
7311  * returns the tree buffer or NULL.
7312  */
7313 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
7314                                         struct btrfs_root *root,
7315                                         u64 parent, u64 root_objectid,
7316                                         struct btrfs_disk_key *key, int level,
7317                                         u64 hint, u64 empty_size)
7318 {
7319         struct btrfs_key ins;
7320         struct btrfs_block_rsv *block_rsv;
7321         struct extent_buffer *buf;
7322         u64 flags = 0;
7323         int ret;
7324         u32 blocksize = root->nodesize;
7325         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7326                                                  SKINNY_METADATA);
7327
7328         if (btrfs_test_is_dummy_root(root)) {
7329                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
7330                                             blocksize, level);
7331                 if (!IS_ERR(buf))
7332                         root->alloc_bytenr += blocksize;
7333                 return buf;
7334         }
7335
7336         block_rsv = use_block_rsv(trans, root, blocksize);
7337         if (IS_ERR(block_rsv))
7338                 return ERR_CAST(block_rsv);
7339
7340         ret = btrfs_reserve_extent(root, blocksize, blocksize,
7341                                    empty_size, hint, &ins, 0, 0);
7342         if (ret) {
7343                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
7344                 return ERR_PTR(ret);
7345         }
7346
7347         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
7348                                     blocksize, level);
7349         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
7350
7351         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
7352                 if (parent == 0)
7353                         parent = ins.objectid;
7354                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
7355         } else
7356                 BUG_ON(parent > 0);
7357
7358         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
7359                 struct btrfs_delayed_extent_op *extent_op;
7360                 extent_op = btrfs_alloc_delayed_extent_op();
7361                 BUG_ON(!extent_op); /* -ENOMEM */
7362                 if (key)
7363                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
7364                 else
7365                         memset(&extent_op->key, 0, sizeof(extent_op->key));
7366                 extent_op->flags_to_set = flags;
7367                 if (skinny_metadata)
7368                         extent_op->update_key = 0;
7369                 else
7370                         extent_op->update_key = 1;
7371                 extent_op->update_flags = 1;
7372                 extent_op->is_data = 0;
7373                 extent_op->level = level;
7374
7375                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7376                                         ins.objectid,
7377                                         ins.offset, parent, root_objectid,
7378                                         level, BTRFS_ADD_DELAYED_EXTENT,
7379                                         extent_op, 0);
7380                 BUG_ON(ret); /* -ENOMEM */
7381         }
7382         return buf;
7383 }
7384
7385 struct walk_control {
7386         u64 refs[BTRFS_MAX_LEVEL];
7387         u64 flags[BTRFS_MAX_LEVEL];
7388         struct btrfs_key update_progress;
7389         int stage;
7390         int level;
7391         int shared_level;
7392         int update_ref;
7393         int keep_locks;
7394         int reada_slot;
7395         int reada_count;
7396         int for_reloc;
7397 };
7398
7399 #define DROP_REFERENCE  1
7400 #define UPDATE_BACKREF  2
7401
7402 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
7403                                      struct btrfs_root *root,
7404                                      struct walk_control *wc,
7405                                      struct btrfs_path *path)
7406 {
7407         u64 bytenr;
7408         u64 generation;
7409         u64 refs;
7410         u64 flags;
7411         u32 nritems;
7412         u32 blocksize;
7413         struct btrfs_key key;
7414         struct extent_buffer *eb;
7415         int ret;
7416         int slot;
7417         int nread = 0;
7418
7419         if (path->slots[wc->level] < wc->reada_slot) {
7420                 wc->reada_count = wc->reada_count * 2 / 3;
7421                 wc->reada_count = max(wc->reada_count, 2);
7422         } else {
7423                 wc->reada_count = wc->reada_count * 3 / 2;
7424                 wc->reada_count = min_t(int, wc->reada_count,
7425                                         BTRFS_NODEPTRS_PER_BLOCK(root));
7426         }
7427
7428         eb = path->nodes[wc->level];
7429         nritems = btrfs_header_nritems(eb);
7430         blocksize = root->nodesize;
7431
7432         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
7433                 if (nread >= wc->reada_count)
7434                         break;
7435
7436                 cond_resched();
7437                 bytenr = btrfs_node_blockptr(eb, slot);
7438                 generation = btrfs_node_ptr_generation(eb, slot);
7439
7440                 if (slot == path->slots[wc->level])
7441                         goto reada;
7442
7443                 if (wc->stage == UPDATE_BACKREF &&
7444                     generation <= root->root_key.offset)
7445                         continue;
7446
7447                 /* We don't lock the tree block, it's OK to be racy here */
7448                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7449                                                wc->level - 1, 1, &refs,
7450                                                &flags);
7451                 /* We don't care about errors in readahead. */
7452                 if (ret < 0)
7453                         continue;
7454                 BUG_ON(refs == 0);
7455
7456                 if (wc->stage == DROP_REFERENCE) {
7457                         if (refs == 1)
7458                                 goto reada;
7459
7460                         if (wc->level == 1 &&
7461                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7462                                 continue;
7463                         if (!wc->update_ref ||
7464                             generation <= root->root_key.offset)
7465                                 continue;
7466                         btrfs_node_key_to_cpu(eb, &key, slot);
7467                         ret = btrfs_comp_cpu_keys(&key,
7468                                                   &wc->update_progress);
7469                         if (ret < 0)
7470                                 continue;
7471                 } else {
7472                         if (wc->level == 1 &&
7473                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7474                                 continue;
7475                 }
7476 reada:
7477                 readahead_tree_block(root, bytenr, blocksize);
7478                 nread++;
7479         }
7480         wc->reada_slot = slot;
7481 }
7482
7483 static int account_leaf_items(struct btrfs_trans_handle *trans,
7484                               struct btrfs_root *root,
7485                               struct extent_buffer *eb)
7486 {
7487         int nr = btrfs_header_nritems(eb);
7488         int i, extent_type, ret;
7489         struct btrfs_key key;
7490         struct btrfs_file_extent_item *fi;
7491         u64 bytenr, num_bytes;
7492
7493         for (i = 0; i < nr; i++) {
7494                 btrfs_item_key_to_cpu(eb, &key, i);
7495
7496                 if (key.type != BTRFS_EXTENT_DATA_KEY)
7497                         continue;
7498
7499                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
7500                 /* filter out non qgroup-accountable extents  */
7501                 extent_type = btrfs_file_extent_type(eb, fi);
7502
7503                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
7504                         continue;
7505
7506                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
7507                 if (!bytenr)
7508                         continue;
7509
7510                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
7511
7512                 ret = btrfs_qgroup_record_ref(trans, root->fs_info,
7513                                               root->objectid,
7514                                               bytenr, num_bytes,
7515                                               BTRFS_QGROUP_OPER_SUB_SUBTREE, 0);
7516                 if (ret)
7517                         return ret;
7518         }
7519         return 0;
7520 }
7521
7522 /*
7523  * Walk up the tree from the bottom, freeing leaves and any interior
7524  * nodes which have had all slots visited. If a node (leaf or
7525  * interior) is freed, the node above it will have it's slot
7526  * incremented. The root node will never be freed.
7527  *
7528  * At the end of this function, we should have a path which has all
7529  * slots incremented to the next position for a search. If we need to
7530  * read a new node it will be NULL and the node above it will have the
7531  * correct slot selected for a later read.
7532  *
7533  * If we increment the root nodes slot counter past the number of
7534  * elements, 1 is returned to signal completion of the search.
7535  */
7536 static int adjust_slots_upwards(struct btrfs_root *root,
7537                                 struct btrfs_path *path, int root_level)
7538 {
7539         int level = 0;
7540         int nr, slot;
7541         struct extent_buffer *eb;
7542
7543         if (root_level == 0)
7544                 return 1;
7545
7546         while (level <= root_level) {
7547                 eb = path->nodes[level];
7548                 nr = btrfs_header_nritems(eb);
7549                 path->slots[level]++;
7550                 slot = path->slots[level];
7551                 if (slot >= nr || level == 0) {
7552                         /*
7553                          * Don't free the root -  we will detect this
7554                          * condition after our loop and return a
7555                          * positive value for caller to stop walking the tree.
7556                          */
7557                         if (level != root_level) {
7558                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7559                                 path->locks[level] = 0;
7560
7561                                 free_extent_buffer(eb);
7562                                 path->nodes[level] = NULL;
7563                                 path->slots[level] = 0;
7564                         }
7565                 } else {
7566                         /*
7567                          * We have a valid slot to walk back down
7568                          * from. Stop here so caller can process these
7569                          * new nodes.
7570                          */
7571                         break;
7572                 }
7573
7574                 level++;
7575         }
7576
7577         eb = path->nodes[root_level];
7578         if (path->slots[root_level] >= btrfs_header_nritems(eb))
7579                 return 1;
7580
7581         return 0;
7582 }
7583
7584 /*
7585  * root_eb is the subtree root and is locked before this function is called.
7586  */
7587 static int account_shared_subtree(struct btrfs_trans_handle *trans,
7588                                   struct btrfs_root *root,
7589                                   struct extent_buffer *root_eb,
7590                                   u64 root_gen,
7591                                   int root_level)
7592 {
7593         int ret = 0;
7594         int level;
7595         struct extent_buffer *eb = root_eb;
7596         struct btrfs_path *path = NULL;
7597
7598         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
7599         BUG_ON(root_eb == NULL);
7600
7601         if (!root->fs_info->quota_enabled)
7602                 return 0;
7603
7604         if (!extent_buffer_uptodate(root_eb)) {
7605                 ret = btrfs_read_buffer(root_eb, root_gen);
7606                 if (ret)
7607                         goto out;
7608         }
7609
7610         if (root_level == 0) {
7611                 ret = account_leaf_items(trans, root, root_eb);
7612                 goto out;
7613         }
7614
7615         path = btrfs_alloc_path();
7616         if (!path)
7617                 return -ENOMEM;
7618
7619         /*
7620          * Walk down the tree.  Missing extent blocks are filled in as
7621          * we go. Metadata is accounted every time we read a new
7622          * extent block.
7623          *
7624          * When we reach a leaf, we account for file extent items in it,
7625          * walk back up the tree (adjusting slot pointers as we go)
7626          * and restart the search process.
7627          */
7628         extent_buffer_get(root_eb); /* For path */
7629         path->nodes[root_level] = root_eb;
7630         path->slots[root_level] = 0;
7631         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
7632 walk_down:
7633         level = root_level;
7634         while (level >= 0) {
7635                 if (path->nodes[level] == NULL) {
7636                         int parent_slot;
7637                         u64 child_gen;
7638                         u64 child_bytenr;
7639
7640                         /* We need to get child blockptr/gen from
7641                          * parent before we can read it. */
7642                         eb = path->nodes[level + 1];
7643                         parent_slot = path->slots[level + 1];
7644                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
7645                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
7646
7647                         eb = read_tree_block(root, child_bytenr, child_gen);
7648                         if (!eb || !extent_buffer_uptodate(eb)) {
7649                                 ret = -EIO;
7650                                 goto out;
7651                         }
7652
7653                         path->nodes[level] = eb;
7654                         path->slots[level] = 0;
7655
7656                         btrfs_tree_read_lock(eb);
7657                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
7658                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
7659
7660                         ret = btrfs_qgroup_record_ref(trans, root->fs_info,
7661                                                 root->objectid,
7662                                                 child_bytenr,
7663                                                 root->nodesize,
7664                                                 BTRFS_QGROUP_OPER_SUB_SUBTREE,
7665                                                 0);
7666                         if (ret)
7667                                 goto out;
7668
7669                 }
7670
7671                 if (level == 0) {
7672                         ret = account_leaf_items(trans, root, path->nodes[level]);
7673                         if (ret)
7674                                 goto out;
7675
7676                         /* Nonzero return here means we completed our search */
7677                         ret = adjust_slots_upwards(root, path, root_level);
7678                         if (ret)
7679                                 break;
7680
7681                         /* Restart search with new slots */
7682                         goto walk_down;
7683                 }
7684
7685                 level--;
7686         }
7687
7688         ret = 0;
7689 out:
7690         btrfs_free_path(path);
7691
7692         return ret;
7693 }
7694
7695 /*
7696  * helper to process tree block while walking down the tree.
7697  *
7698  * when wc->stage == UPDATE_BACKREF, this function updates
7699  * back refs for pointers in the block.
7700  *
7701  * NOTE: return value 1 means we should stop walking down.
7702  */
7703 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
7704                                    struct btrfs_root *root,
7705                                    struct btrfs_path *path,
7706                                    struct walk_control *wc, int lookup_info)
7707 {
7708         int level = wc->level;
7709         struct extent_buffer *eb = path->nodes[level];
7710         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7711         int ret;
7712
7713         if (wc->stage == UPDATE_BACKREF &&
7714             btrfs_header_owner(eb) != root->root_key.objectid)
7715                 return 1;
7716
7717         /*
7718          * when reference count of tree block is 1, it won't increase
7719          * again. once full backref flag is set, we never clear it.
7720          */
7721         if (lookup_info &&
7722             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
7723              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
7724                 BUG_ON(!path->locks[level]);
7725                 ret = btrfs_lookup_extent_info(trans, root,
7726                                                eb->start, level, 1,
7727                                                &wc->refs[level],
7728                                                &wc->flags[level]);
7729                 BUG_ON(ret == -ENOMEM);
7730                 if (ret)
7731                         return ret;
7732                 BUG_ON(wc->refs[level] == 0);
7733         }
7734
7735         if (wc->stage == DROP_REFERENCE) {
7736                 if (wc->refs[level] > 1)
7737                         return 1;
7738
7739                 if (path->locks[level] && !wc->keep_locks) {
7740                         btrfs_tree_unlock_rw(eb, path->locks[level]);
7741                         path->locks[level] = 0;
7742                 }
7743                 return 0;
7744         }
7745
7746         /* wc->stage == UPDATE_BACKREF */
7747         if (!(wc->flags[level] & flag)) {
7748                 BUG_ON(!path->locks[level]);
7749                 ret = btrfs_inc_ref(trans, root, eb, 1);
7750                 BUG_ON(ret); /* -ENOMEM */
7751                 ret = btrfs_dec_ref(trans, root, eb, 0);
7752                 BUG_ON(ret); /* -ENOMEM */
7753                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
7754                                                   eb->len, flag,
7755                                                   btrfs_header_level(eb), 0);
7756                 BUG_ON(ret); /* -ENOMEM */
7757                 wc->flags[level] |= flag;
7758         }
7759
7760         /*
7761          * the block is shared by multiple trees, so it's not good to
7762          * keep the tree lock
7763          */
7764         if (path->locks[level] && level > 0) {
7765                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7766                 path->locks[level] = 0;
7767         }
7768         return 0;
7769 }
7770
7771 /*
7772  * helper to process tree block pointer.
7773  *
7774  * when wc->stage == DROP_REFERENCE, this function checks
7775  * reference count of the block pointed to. if the block
7776  * is shared and we need update back refs for the subtree
7777  * rooted at the block, this function changes wc->stage to
7778  * UPDATE_BACKREF. if the block is shared and there is no
7779  * need to update back, this function drops the reference
7780  * to the block.
7781  *
7782  * NOTE: return value 1 means we should stop walking down.
7783  */
7784 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
7785                                  struct btrfs_root *root,
7786                                  struct btrfs_path *path,
7787                                  struct walk_control *wc, int *lookup_info)
7788 {
7789         u64 bytenr;
7790         u64 generation;
7791         u64 parent;
7792         u32 blocksize;
7793         struct btrfs_key key;
7794         struct extent_buffer *next;
7795         int level = wc->level;
7796         int reada = 0;
7797         int ret = 0;
7798         bool need_account = false;
7799
7800         generation = btrfs_node_ptr_generation(path->nodes[level],
7801                                                path->slots[level]);
7802         /*
7803          * if the lower level block was created before the snapshot
7804          * was created, we know there is no need to update back refs
7805          * for the subtree
7806          */
7807         if (wc->stage == UPDATE_BACKREF &&
7808             generation <= root->root_key.offset) {
7809                 *lookup_info = 1;
7810                 return 1;
7811         }
7812
7813         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
7814         blocksize = root->nodesize;
7815
7816         next = btrfs_find_tree_block(root, bytenr);
7817         if (!next) {
7818                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
7819                 if (!next)
7820                         return -ENOMEM;
7821                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
7822                                                level - 1);
7823                 reada = 1;
7824         }
7825         btrfs_tree_lock(next);
7826         btrfs_set_lock_blocking(next);
7827
7828         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
7829                                        &wc->refs[level - 1],
7830                                        &wc->flags[level - 1]);
7831         if (ret < 0) {
7832                 btrfs_tree_unlock(next);
7833                 return ret;
7834         }
7835
7836         if (unlikely(wc->refs[level - 1] == 0)) {
7837                 btrfs_err(root->fs_info, "Missing references.");
7838                 BUG();
7839         }
7840         *lookup_info = 0;
7841
7842         if (wc->stage == DROP_REFERENCE) {
7843                 if (wc->refs[level - 1] > 1) {
7844                         need_account = true;
7845                         if (level == 1 &&
7846                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7847                                 goto skip;
7848
7849                         if (!wc->update_ref ||
7850                             generation <= root->root_key.offset)
7851                                 goto skip;
7852
7853                         btrfs_node_key_to_cpu(path->nodes[level], &key,
7854                                               path->slots[level]);
7855                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
7856                         if (ret < 0)
7857                                 goto skip;
7858
7859                         wc->stage = UPDATE_BACKREF;
7860                         wc->shared_level = level - 1;
7861                 }
7862         } else {
7863                 if (level == 1 &&
7864                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7865                         goto skip;
7866         }
7867
7868         if (!btrfs_buffer_uptodate(next, generation, 0)) {
7869                 btrfs_tree_unlock(next);
7870                 free_extent_buffer(next);
7871                 next = NULL;
7872                 *lookup_info = 1;
7873         }
7874
7875         if (!next) {
7876                 if (reada && level == 1)
7877                         reada_walk_down(trans, root, wc, path);
7878                 next = read_tree_block(root, bytenr, generation);
7879                 if (!next || !extent_buffer_uptodate(next)) {
7880                         free_extent_buffer(next);
7881                         return -EIO;
7882                 }
7883                 btrfs_tree_lock(next);
7884                 btrfs_set_lock_blocking(next);
7885         }
7886
7887         level--;
7888         BUG_ON(level != btrfs_header_level(next));
7889         path->nodes[level] = next;
7890         path->slots[level] = 0;
7891         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7892         wc->level = level;
7893         if (wc->level == 1)
7894                 wc->reada_slot = 0;
7895         return 0;
7896 skip:
7897         wc->refs[level - 1] = 0;
7898         wc->flags[level - 1] = 0;
7899         if (wc->stage == DROP_REFERENCE) {
7900                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
7901                         parent = path->nodes[level]->start;
7902                 } else {
7903                         BUG_ON(root->root_key.objectid !=
7904                                btrfs_header_owner(path->nodes[level]));
7905                         parent = 0;
7906                 }
7907
7908                 if (need_account) {
7909                         ret = account_shared_subtree(trans, root, next,
7910                                                      generation, level - 1);
7911                         if (ret) {
7912                                 printk_ratelimited(KERN_ERR "BTRFS: %s Error "
7913                                         "%d accounting shared subtree. Quota "
7914                                         "is out of sync, rescan required.\n",
7915                                         root->fs_info->sb->s_id, ret);
7916                         }
7917                 }
7918                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
7919                                 root->root_key.objectid, level - 1, 0, 0);
7920                 BUG_ON(ret); /* -ENOMEM */
7921         }
7922         btrfs_tree_unlock(next);
7923         free_extent_buffer(next);
7924         *lookup_info = 1;
7925         return 1;
7926 }
7927
7928 /*
7929  * helper to process tree block while walking up the tree.
7930  *
7931  * when wc->stage == DROP_REFERENCE, this function drops
7932  * reference count on the block.
7933  *
7934  * when wc->stage == UPDATE_BACKREF, this function changes
7935  * wc->stage back to DROP_REFERENCE if we changed wc->stage
7936  * to UPDATE_BACKREF previously while processing the block.
7937  *
7938  * NOTE: return value 1 means we should stop walking up.
7939  */
7940 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
7941                                  struct btrfs_root *root,
7942                                  struct btrfs_path *path,
7943                                  struct walk_control *wc)
7944 {
7945         int ret;
7946         int level = wc->level;
7947         struct extent_buffer *eb = path->nodes[level];
7948         u64 parent = 0;
7949
7950         if (wc->stage == UPDATE_BACKREF) {
7951                 BUG_ON(wc->shared_level < level);
7952                 if (level < wc->shared_level)
7953                         goto out;
7954
7955                 ret = find_next_key(path, level + 1, &wc->update_progress);
7956                 if (ret > 0)
7957                         wc->update_ref = 0;
7958
7959                 wc->stage = DROP_REFERENCE;
7960                 wc->shared_level = -1;
7961                 path->slots[level] = 0;
7962
7963                 /*
7964                  * check reference count again if the block isn't locked.
7965                  * we should start walking down the tree again if reference
7966                  * count is one.
7967                  */
7968                 if (!path->locks[level]) {
7969                         BUG_ON(level == 0);
7970                         btrfs_tree_lock(eb);
7971                         btrfs_set_lock_blocking(eb);
7972                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7973
7974                         ret = btrfs_lookup_extent_info(trans, root,
7975                                                        eb->start, level, 1,
7976                                                        &wc->refs[level],
7977                                                        &wc->flags[level]);
7978                         if (ret < 0) {
7979                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7980                                 path->locks[level] = 0;
7981                                 return ret;
7982                         }
7983                         BUG_ON(wc->refs[level] == 0);
7984                         if (wc->refs[level] == 1) {
7985                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7986                                 path->locks[level] = 0;
7987                                 return 1;
7988                         }
7989                 }
7990         }
7991
7992         /* wc->stage == DROP_REFERENCE */
7993         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
7994
7995         if (wc->refs[level] == 1) {
7996                 if (level == 0) {
7997                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7998                                 ret = btrfs_dec_ref(trans, root, eb, 1);
7999                         else
8000                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8001                         BUG_ON(ret); /* -ENOMEM */
8002                         ret = account_leaf_items(trans, root, eb);
8003                         if (ret) {
8004                                 printk_ratelimited(KERN_ERR "BTRFS: %s Error "
8005                                         "%d accounting leaf items. Quota "
8006                                         "is out of sync, rescan required.\n",
8007                                         root->fs_info->sb->s_id, ret);
8008                         }
8009                 }
8010                 /* make block locked assertion in clean_tree_block happy */
8011                 if (!path->locks[level] &&
8012                     btrfs_header_generation(eb) == trans->transid) {
8013                         btrfs_tree_lock(eb);
8014                         btrfs_set_lock_blocking(eb);
8015                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8016                 }
8017                 clean_tree_block(trans, root, eb);
8018         }
8019
8020         if (eb == root->node) {
8021                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8022                         parent = eb->start;
8023                 else
8024                         BUG_ON(root->root_key.objectid !=
8025                                btrfs_header_owner(eb));
8026         } else {
8027                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8028                         parent = path->nodes[level + 1]->start;
8029                 else
8030                         BUG_ON(root->root_key.objectid !=
8031                                btrfs_header_owner(path->nodes[level + 1]));
8032         }
8033
8034         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8035 out:
8036         wc->refs[level] = 0;
8037         wc->flags[level] = 0;
8038         return 0;
8039 }
8040
8041 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8042                                    struct btrfs_root *root,
8043                                    struct btrfs_path *path,
8044                                    struct walk_control *wc)
8045 {
8046         int level = wc->level;
8047         int lookup_info = 1;
8048         int ret;
8049
8050         while (level >= 0) {
8051                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8052                 if (ret > 0)
8053                         break;
8054
8055                 if (level == 0)
8056                         break;
8057
8058                 if (path->slots[level] >=
8059                     btrfs_header_nritems(path->nodes[level]))
8060                         break;
8061
8062                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8063                 if (ret > 0) {
8064                         path->slots[level]++;
8065                         continue;
8066                 } else if (ret < 0)
8067                         return ret;
8068                 level = wc->level;
8069         }
8070         return 0;
8071 }
8072
8073 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8074                                  struct btrfs_root *root,
8075                                  struct btrfs_path *path,
8076                                  struct walk_control *wc, int max_level)
8077 {
8078         int level = wc->level;
8079         int ret;
8080
8081         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8082         while (level < max_level && path->nodes[level]) {
8083                 wc->level = level;
8084                 if (path->slots[level] + 1 <
8085                     btrfs_header_nritems(path->nodes[level])) {
8086                         path->slots[level]++;
8087                         return 0;
8088                 } else {
8089                         ret = walk_up_proc(trans, root, path, wc);
8090                         if (ret > 0)
8091                                 return 0;
8092
8093                         if (path->locks[level]) {
8094                                 btrfs_tree_unlock_rw(path->nodes[level],
8095                                                      path->locks[level]);
8096                                 path->locks[level] = 0;
8097                         }
8098                         free_extent_buffer(path->nodes[level]);
8099                         path->nodes[level] = NULL;
8100                         level++;
8101                 }
8102         }
8103         return 1;
8104 }
8105
8106 /*
8107  * drop a subvolume tree.
8108  *
8109  * this function traverses the tree freeing any blocks that only
8110  * referenced by the tree.
8111  *
8112  * when a shared tree block is found. this function decreases its
8113  * reference count by one. if update_ref is true, this function
8114  * also make sure backrefs for the shared block and all lower level
8115  * blocks are properly updated.
8116  *
8117  * If called with for_reloc == 0, may exit early with -EAGAIN
8118  */
8119 int btrfs_drop_snapshot(struct btrfs_root *root,
8120                          struct btrfs_block_rsv *block_rsv, int update_ref,
8121                          int for_reloc)
8122 {
8123         struct btrfs_path *path;
8124         struct btrfs_trans_handle *trans;
8125         struct btrfs_root *tree_root = root->fs_info->tree_root;
8126         struct btrfs_root_item *root_item = &root->root_item;
8127         struct walk_control *wc;
8128         struct btrfs_key key;
8129         int err = 0;
8130         int ret;
8131         int level;
8132         bool root_dropped = false;
8133
8134         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
8135
8136         path = btrfs_alloc_path();
8137         if (!path) {
8138                 err = -ENOMEM;
8139                 goto out;
8140         }
8141
8142         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8143         if (!wc) {
8144                 btrfs_free_path(path);
8145                 err = -ENOMEM;
8146                 goto out;
8147         }
8148
8149         trans = btrfs_start_transaction(tree_root, 0);
8150         if (IS_ERR(trans)) {
8151                 err = PTR_ERR(trans);
8152                 goto out_free;
8153         }
8154
8155         if (block_rsv)
8156                 trans->block_rsv = block_rsv;
8157
8158         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8159                 level = btrfs_header_level(root->node);
8160                 path->nodes[level] = btrfs_lock_root_node(root);
8161                 btrfs_set_lock_blocking(path->nodes[level]);
8162                 path->slots[level] = 0;
8163                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8164                 memset(&wc->update_progress, 0,
8165                        sizeof(wc->update_progress));
8166         } else {
8167                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8168                 memcpy(&wc->update_progress, &key,
8169                        sizeof(wc->update_progress));
8170
8171                 level = root_item->drop_level;
8172                 BUG_ON(level == 0);
8173                 path->lowest_level = level;
8174                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8175                 path->lowest_level = 0;
8176                 if (ret < 0) {
8177                         err = ret;
8178                         goto out_end_trans;
8179                 }
8180                 WARN_ON(ret > 0);
8181
8182                 /*
8183                  * unlock our path, this is safe because only this
8184                  * function is allowed to delete this snapshot
8185                  */
8186                 btrfs_unlock_up_safe(path, 0);
8187
8188                 level = btrfs_header_level(root->node);
8189                 while (1) {
8190                         btrfs_tree_lock(path->nodes[level]);
8191                         btrfs_set_lock_blocking(path->nodes[level]);
8192                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8193
8194                         ret = btrfs_lookup_extent_info(trans, root,
8195                                                 path->nodes[level]->start,
8196                                                 level, 1, &wc->refs[level],
8197                                                 &wc->flags[level]);
8198                         if (ret < 0) {
8199                                 err = ret;
8200                                 goto out_end_trans;
8201                         }
8202                         BUG_ON(wc->refs[level] == 0);
8203
8204                         if (level == root_item->drop_level)
8205                                 break;
8206
8207                         btrfs_tree_unlock(path->nodes[level]);
8208                         path->locks[level] = 0;
8209                         WARN_ON(wc->refs[level] != 1);
8210                         level--;
8211                 }
8212         }
8213
8214         wc->level = level;
8215         wc->shared_level = -1;
8216         wc->stage = DROP_REFERENCE;
8217         wc->update_ref = update_ref;
8218         wc->keep_locks = 0;
8219         wc->for_reloc = for_reloc;
8220         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8221
8222         while (1) {
8223
8224                 ret = walk_down_tree(trans, root, path, wc);
8225                 if (ret < 0) {
8226                         err = ret;
8227                         break;
8228                 }
8229
8230                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8231                 if (ret < 0) {
8232                         err = ret;
8233                         break;
8234                 }
8235
8236                 if (ret > 0) {
8237                         BUG_ON(wc->stage != DROP_REFERENCE);
8238                         break;
8239                 }
8240
8241                 if (wc->stage == DROP_REFERENCE) {
8242                         level = wc->level;
8243                         btrfs_node_key(path->nodes[level],
8244                                        &root_item->drop_progress,
8245                                        path->slots[level]);
8246                         root_item->drop_level = level;
8247                 }
8248
8249                 BUG_ON(wc->level == 0);
8250                 if (btrfs_should_end_transaction(trans, tree_root) ||
8251                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
8252                         ret = btrfs_update_root(trans, tree_root,
8253                                                 &root->root_key,
8254                                                 root_item);
8255                         if (ret) {
8256                                 btrfs_abort_transaction(trans, tree_root, ret);
8257                                 err = ret;
8258                                 goto out_end_trans;
8259                         }
8260
8261                         /*
8262                          * Qgroup update accounting is run from
8263                          * delayed ref handling. This usually works
8264                          * out because delayed refs are normally the
8265                          * only way qgroup updates are added. However,
8266                          * we may have added updates during our tree
8267                          * walk so run qgroups here to make sure we
8268                          * don't lose any updates.
8269                          */
8270                         ret = btrfs_delayed_qgroup_accounting(trans,
8271                                                               root->fs_info);
8272                         if (ret)
8273                                 printk_ratelimited(KERN_ERR "BTRFS: Failure %d "
8274                                                    "running qgroup updates "
8275                                                    "during snapshot delete. "
8276                                                    "Quota is out of sync, "
8277                                                    "rescan required.\n", ret);
8278
8279                         btrfs_end_transaction_throttle(trans, tree_root);
8280                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
8281                                 pr_debug("BTRFS: drop snapshot early exit\n");
8282                                 err = -EAGAIN;
8283                                 goto out_free;
8284                         }
8285
8286                         trans = btrfs_start_transaction(tree_root, 0);
8287                         if (IS_ERR(trans)) {
8288                                 err = PTR_ERR(trans);
8289                                 goto out_free;
8290                         }
8291                         if (block_rsv)
8292                                 trans->block_rsv = block_rsv;
8293                 }
8294         }
8295         btrfs_release_path(path);
8296         if (err)
8297                 goto out_end_trans;
8298
8299         ret = btrfs_del_root(trans, tree_root, &root->root_key);
8300         if (ret) {
8301                 btrfs_abort_transaction(trans, tree_root, ret);
8302                 goto out_end_trans;
8303         }
8304
8305         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
8306                 ret = btrfs_find_root(tree_root, &root->root_key, path,
8307                                       NULL, NULL);
8308                 if (ret < 0) {
8309                         btrfs_abort_transaction(trans, tree_root, ret);
8310                         err = ret;
8311                         goto out_end_trans;
8312                 } else if (ret > 0) {
8313                         /* if we fail to delete the orphan item this time
8314                          * around, it'll get picked up the next time.
8315                          *
8316                          * The most common failure here is just -ENOENT.
8317                          */
8318                         btrfs_del_orphan_item(trans, tree_root,
8319                                               root->root_key.objectid);
8320                 }
8321         }
8322
8323         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
8324                 btrfs_drop_and_free_fs_root(tree_root->fs_info, root);
8325         } else {
8326                 free_extent_buffer(root->node);
8327                 free_extent_buffer(root->commit_root);
8328                 btrfs_put_fs_root(root);
8329         }
8330         root_dropped = true;
8331 out_end_trans:
8332         ret = btrfs_delayed_qgroup_accounting(trans, tree_root->fs_info);
8333         if (ret)
8334                 printk_ratelimited(KERN_ERR "BTRFS: Failure %d "
8335                                    "running qgroup updates "
8336                                    "during snapshot delete. "
8337                                    "Quota is out of sync, "
8338                                    "rescan required.\n", ret);
8339
8340         btrfs_end_transaction_throttle(trans, tree_root);
8341 out_free:
8342         kfree(wc);
8343         btrfs_free_path(path);
8344 out:
8345         /*
8346          * So if we need to stop dropping the snapshot for whatever reason we
8347          * need to make sure to add it back to the dead root list so that we
8348          * keep trying to do the work later.  This also cleans up roots if we
8349          * don't have it in the radix (like when we recover after a power fail
8350          * or unmount) so we don't leak memory.
8351          */
8352         if (!for_reloc && root_dropped == false)
8353                 btrfs_add_dead_root(root);
8354         if (err && err != -EAGAIN)
8355                 btrfs_std_error(root->fs_info, err);
8356         return err;
8357 }
8358
8359 /*
8360  * drop subtree rooted at tree block 'node'.
8361  *
8362  * NOTE: this function will unlock and release tree block 'node'
8363  * only used by relocation code
8364  */
8365 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
8366                         struct btrfs_root *root,
8367                         struct extent_buffer *node,
8368                         struct extent_buffer *parent)
8369 {
8370         struct btrfs_path *path;
8371         struct walk_control *wc;
8372         int level;
8373         int parent_level;
8374         int ret = 0;
8375         int wret;
8376
8377         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
8378
8379         path = btrfs_alloc_path();
8380         if (!path)
8381                 return -ENOMEM;
8382
8383         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8384         if (!wc) {
8385                 btrfs_free_path(path);
8386                 return -ENOMEM;
8387         }
8388
8389         btrfs_assert_tree_locked(parent);
8390         parent_level = btrfs_header_level(parent);
8391         extent_buffer_get(parent);
8392         path->nodes[parent_level] = parent;
8393         path->slots[parent_level] = btrfs_header_nritems(parent);
8394
8395         btrfs_assert_tree_locked(node);
8396         level = btrfs_header_level(node);
8397         path->nodes[level] = node;
8398         path->slots[level] = 0;
8399         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8400
8401         wc->refs[parent_level] = 1;
8402         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8403         wc->level = level;
8404         wc->shared_level = -1;
8405         wc->stage = DROP_REFERENCE;
8406         wc->update_ref = 0;
8407         wc->keep_locks = 1;
8408         wc->for_reloc = 1;
8409         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8410
8411         while (1) {
8412                 wret = walk_down_tree(trans, root, path, wc);
8413                 if (wret < 0) {
8414                         ret = wret;
8415                         break;
8416                 }
8417
8418                 wret = walk_up_tree(trans, root, path, wc, parent_level);
8419                 if (wret < 0)
8420                         ret = wret;
8421                 if (wret != 0)
8422                         break;
8423         }
8424
8425         kfree(wc);
8426         btrfs_free_path(path);
8427         return ret;
8428 }
8429
8430 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
8431 {
8432         u64 num_devices;
8433         u64 stripped;
8434
8435         /*
8436          * if restripe for this chunk_type is on pick target profile and
8437          * return, otherwise do the usual balance
8438          */
8439         stripped = get_restripe_target(root->fs_info, flags);
8440         if (stripped)
8441                 return extended_to_chunk(stripped);
8442
8443         num_devices = root->fs_info->fs_devices->rw_devices;
8444
8445         stripped = BTRFS_BLOCK_GROUP_RAID0 |
8446                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
8447                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
8448
8449         if (num_devices == 1) {
8450                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8451                 stripped = flags & ~stripped;
8452
8453                 /* turn raid0 into single device chunks */
8454                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
8455                         return stripped;
8456
8457                 /* turn mirroring into duplication */
8458                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
8459                              BTRFS_BLOCK_GROUP_RAID10))
8460                         return stripped | BTRFS_BLOCK_GROUP_DUP;
8461         } else {
8462                 /* they already had raid on here, just return */
8463                 if (flags & stripped)
8464                         return flags;
8465
8466                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8467                 stripped = flags & ~stripped;
8468
8469                 /* switch duplicated blocks with raid1 */
8470                 if (flags & BTRFS_BLOCK_GROUP_DUP)
8471                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
8472
8473                 /* this is drive concat, leave it alone */
8474         }
8475
8476         return flags;
8477 }
8478
8479 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
8480 {
8481         struct btrfs_space_info *sinfo = cache->space_info;
8482         u64 num_bytes;
8483         u64 min_allocable_bytes;
8484         int ret = -ENOSPC;
8485
8486
8487         /*
8488          * We need some metadata space and system metadata space for
8489          * allocating chunks in some corner cases until we force to set
8490          * it to be readonly.
8491          */
8492         if ((sinfo->flags &
8493              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
8494             !force)
8495                 min_allocable_bytes = 1 * 1024 * 1024;
8496         else
8497                 min_allocable_bytes = 0;
8498
8499         spin_lock(&sinfo->lock);
8500         spin_lock(&cache->lock);
8501
8502         if (cache->ro) {
8503                 ret = 0;
8504                 goto out;
8505         }
8506
8507         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8508                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8509
8510         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8511             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
8512             min_allocable_bytes <= sinfo->total_bytes) {
8513                 sinfo->bytes_readonly += num_bytes;
8514                 cache->ro = 1;
8515                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
8516                 ret = 0;
8517         }
8518 out:
8519         spin_unlock(&cache->lock);
8520         spin_unlock(&sinfo->lock);
8521         return ret;
8522 }
8523
8524 int btrfs_set_block_group_ro(struct btrfs_root *root,
8525                              struct btrfs_block_group_cache *cache)
8526
8527 {
8528         struct btrfs_trans_handle *trans;
8529         u64 alloc_flags;
8530         int ret;
8531
8532         BUG_ON(cache->ro);
8533
8534         trans = btrfs_join_transaction(root);
8535         if (IS_ERR(trans))
8536                 return PTR_ERR(trans);
8537
8538         alloc_flags = update_block_group_flags(root, cache->flags);
8539         if (alloc_flags != cache->flags) {
8540                 ret = do_chunk_alloc(trans, root, alloc_flags,
8541                                      CHUNK_ALLOC_FORCE);
8542                 if (ret < 0)
8543                         goto out;
8544         }
8545
8546         ret = set_block_group_ro(cache, 0);
8547         if (!ret)
8548                 goto out;
8549         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
8550         ret = do_chunk_alloc(trans, root, alloc_flags,
8551                              CHUNK_ALLOC_FORCE);
8552         if (ret < 0)
8553                 goto out;
8554         ret = set_block_group_ro(cache, 0);
8555 out:
8556         btrfs_end_transaction(trans, root);
8557         return ret;
8558 }
8559
8560 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
8561                             struct btrfs_root *root, u64 type)
8562 {
8563         u64 alloc_flags = get_alloc_profile(root, type);
8564         return do_chunk_alloc(trans, root, alloc_flags,
8565                               CHUNK_ALLOC_FORCE);
8566 }
8567
8568 /*
8569  * helper to account the unused space of all the readonly block group in the
8570  * space_info. takes mirrors into account.
8571  */
8572 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
8573 {
8574         struct btrfs_block_group_cache *block_group;
8575         u64 free_bytes = 0;
8576         int factor;
8577
8578         /* It's df, we don't care if it's racey */
8579         if (list_empty(&sinfo->ro_bgs))
8580                 return 0;
8581
8582         spin_lock(&sinfo->lock);
8583         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
8584                 spin_lock(&block_group->lock);
8585
8586                 if (!block_group->ro) {
8587                         spin_unlock(&block_group->lock);
8588                         continue;
8589                 }
8590
8591                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
8592                                           BTRFS_BLOCK_GROUP_RAID10 |
8593                                           BTRFS_BLOCK_GROUP_DUP))
8594                         factor = 2;
8595                 else
8596                         factor = 1;
8597
8598                 free_bytes += (block_group->key.offset -
8599                                btrfs_block_group_used(&block_group->item)) *
8600                                factor;
8601
8602                 spin_unlock(&block_group->lock);
8603         }
8604         spin_unlock(&sinfo->lock);
8605
8606         return free_bytes;
8607 }
8608
8609 void btrfs_set_block_group_rw(struct btrfs_root *root,
8610                               struct btrfs_block_group_cache *cache)
8611 {
8612         struct btrfs_space_info *sinfo = cache->space_info;
8613         u64 num_bytes;
8614
8615         BUG_ON(!cache->ro);
8616
8617         spin_lock(&sinfo->lock);
8618         spin_lock(&cache->lock);
8619         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8620                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8621         sinfo->bytes_readonly -= num_bytes;
8622         cache->ro = 0;
8623         list_del_init(&cache->ro_list);
8624         spin_unlock(&cache->lock);
8625         spin_unlock(&sinfo->lock);
8626 }
8627
8628 /*
8629  * checks to see if its even possible to relocate this block group.
8630  *
8631  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
8632  * ok to go ahead and try.
8633  */
8634 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
8635 {
8636         struct btrfs_block_group_cache *block_group;
8637         struct btrfs_space_info *space_info;
8638         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
8639         struct btrfs_device *device;
8640         struct btrfs_trans_handle *trans;
8641         u64 min_free;
8642         u64 dev_min = 1;
8643         u64 dev_nr = 0;
8644         u64 target;
8645         int index;
8646         int full = 0;
8647         int ret = 0;
8648
8649         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
8650
8651         /* odd, couldn't find the block group, leave it alone */
8652         if (!block_group)
8653                 return -1;
8654
8655         min_free = btrfs_block_group_used(&block_group->item);
8656
8657         /* no bytes used, we're good */
8658         if (!min_free)
8659                 goto out;
8660
8661         space_info = block_group->space_info;
8662         spin_lock(&space_info->lock);
8663
8664         full = space_info->full;
8665
8666         /*
8667          * if this is the last block group we have in this space, we can't
8668          * relocate it unless we're able to allocate a new chunk below.
8669          *
8670          * Otherwise, we need to make sure we have room in the space to handle
8671          * all of the extents from this block group.  If we can, we're good
8672          */
8673         if ((space_info->total_bytes != block_group->key.offset) &&
8674             (space_info->bytes_used + space_info->bytes_reserved +
8675              space_info->bytes_pinned + space_info->bytes_readonly +
8676              min_free < space_info->total_bytes)) {
8677                 spin_unlock(&space_info->lock);
8678                 goto out;
8679         }
8680         spin_unlock(&space_info->lock);
8681
8682         /*
8683          * ok we don't have enough space, but maybe we have free space on our
8684          * devices to allocate new chunks for relocation, so loop through our
8685          * alloc devices and guess if we have enough space.  if this block
8686          * group is going to be restriped, run checks against the target
8687          * profile instead of the current one.
8688          */
8689         ret = -1;
8690
8691         /*
8692          * index:
8693          *      0: raid10
8694          *      1: raid1
8695          *      2: dup
8696          *      3: raid0
8697          *      4: single
8698          */
8699         target = get_restripe_target(root->fs_info, block_group->flags);
8700         if (target) {
8701                 index = __get_raid_index(extended_to_chunk(target));
8702         } else {
8703                 /*
8704                  * this is just a balance, so if we were marked as full
8705                  * we know there is no space for a new chunk
8706                  */
8707                 if (full)
8708                         goto out;
8709
8710                 index = get_block_group_index(block_group);
8711         }
8712
8713         if (index == BTRFS_RAID_RAID10) {
8714                 dev_min = 4;
8715                 /* Divide by 2 */
8716                 min_free >>= 1;
8717         } else if (index == BTRFS_RAID_RAID1) {
8718                 dev_min = 2;
8719         } else if (index == BTRFS_RAID_DUP) {
8720                 /* Multiply by 2 */
8721                 min_free <<= 1;
8722         } else if (index == BTRFS_RAID_RAID0) {
8723                 dev_min = fs_devices->rw_devices;
8724                 do_div(min_free, dev_min);
8725         }
8726
8727         /* We need to do this so that we can look at pending chunks */
8728         trans = btrfs_join_transaction(root);
8729         if (IS_ERR(trans)) {
8730                 ret = PTR_ERR(trans);
8731                 goto out;
8732         }
8733
8734         mutex_lock(&root->fs_info->chunk_mutex);
8735         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
8736                 u64 dev_offset;
8737
8738                 /*
8739                  * check to make sure we can actually find a chunk with enough
8740                  * space to fit our block group in.
8741                  */
8742                 if (device->total_bytes > device->bytes_used + min_free &&
8743                     !device->is_tgtdev_for_dev_replace) {
8744                         ret = find_free_dev_extent(trans, device, min_free,
8745                                                    &dev_offset, NULL);
8746                         if (!ret)
8747                                 dev_nr++;
8748
8749                         if (dev_nr >= dev_min)
8750                                 break;
8751
8752                         ret = -1;
8753                 }
8754         }
8755         mutex_unlock(&root->fs_info->chunk_mutex);
8756         btrfs_end_transaction(trans, root);
8757 out:
8758         btrfs_put_block_group(block_group);
8759         return ret;
8760 }
8761
8762 static int find_first_block_group(struct btrfs_root *root,
8763                 struct btrfs_path *path, struct btrfs_key *key)
8764 {
8765         int ret = 0;
8766         struct btrfs_key found_key;
8767         struct extent_buffer *leaf;
8768         int slot;
8769
8770         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
8771         if (ret < 0)
8772                 goto out;
8773
8774         while (1) {
8775                 slot = path->slots[0];
8776                 leaf = path->nodes[0];
8777                 if (slot >= btrfs_header_nritems(leaf)) {
8778                         ret = btrfs_next_leaf(root, path);
8779                         if (ret == 0)
8780                                 continue;
8781                         if (ret < 0)
8782                                 goto out;
8783                         break;
8784                 }
8785                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
8786
8787                 if (found_key.objectid >= key->objectid &&
8788                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
8789                         ret = 0;
8790                         goto out;
8791                 }
8792                 path->slots[0]++;
8793         }
8794 out:
8795         return ret;
8796 }
8797
8798 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
8799 {
8800         struct btrfs_block_group_cache *block_group;
8801         u64 last = 0;
8802
8803         while (1) {
8804                 struct inode *inode;
8805
8806                 block_group = btrfs_lookup_first_block_group(info, last);
8807                 while (block_group) {
8808                         spin_lock(&block_group->lock);
8809                         if (block_group->iref)
8810                                 break;
8811                         spin_unlock(&block_group->lock);
8812                         block_group = next_block_group(info->tree_root,
8813                                                        block_group);
8814                 }
8815                 if (!block_group) {
8816                         if (last == 0)
8817                                 break;
8818                         last = 0;
8819                         continue;
8820                 }
8821
8822                 inode = block_group->inode;
8823                 block_group->iref = 0;
8824                 block_group->inode = NULL;
8825                 spin_unlock(&block_group->lock);
8826                 iput(inode);
8827                 last = block_group->key.objectid + block_group->key.offset;
8828                 btrfs_put_block_group(block_group);
8829         }
8830 }
8831
8832 int btrfs_free_block_groups(struct btrfs_fs_info *info)
8833 {
8834         struct btrfs_block_group_cache *block_group;
8835         struct btrfs_space_info *space_info;
8836         struct btrfs_caching_control *caching_ctl;
8837         struct rb_node *n;
8838
8839         down_write(&info->commit_root_sem);
8840         while (!list_empty(&info->caching_block_groups)) {
8841                 caching_ctl = list_entry(info->caching_block_groups.next,
8842                                          struct btrfs_caching_control, list);
8843                 list_del(&caching_ctl->list);
8844                 put_caching_control(caching_ctl);
8845         }
8846         up_write(&info->commit_root_sem);
8847
8848         spin_lock(&info->unused_bgs_lock);
8849         while (!list_empty(&info->unused_bgs)) {
8850                 block_group = list_first_entry(&info->unused_bgs,
8851                                                struct btrfs_block_group_cache,
8852                                                bg_list);
8853                 list_del_init(&block_group->bg_list);
8854                 btrfs_put_block_group(block_group);
8855         }
8856         spin_unlock(&info->unused_bgs_lock);
8857
8858         spin_lock(&info->block_group_cache_lock);
8859         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
8860                 block_group = rb_entry(n, struct btrfs_block_group_cache,
8861                                        cache_node);
8862                 rb_erase(&block_group->cache_node,
8863                          &info->block_group_cache_tree);
8864                 spin_unlock(&info->block_group_cache_lock);
8865
8866                 down_write(&block_group->space_info->groups_sem);
8867                 list_del(&block_group->list);
8868                 up_write(&block_group->space_info->groups_sem);
8869
8870                 if (block_group->cached == BTRFS_CACHE_STARTED)
8871                         wait_block_group_cache_done(block_group);
8872
8873                 /*
8874                  * We haven't cached this block group, which means we could
8875                  * possibly have excluded extents on this block group.
8876                  */
8877                 if (block_group->cached == BTRFS_CACHE_NO ||
8878                     block_group->cached == BTRFS_CACHE_ERROR)
8879                         free_excluded_extents(info->extent_root, block_group);
8880
8881                 btrfs_remove_free_space_cache(block_group);
8882                 btrfs_put_block_group(block_group);
8883
8884                 spin_lock(&info->block_group_cache_lock);
8885         }
8886         spin_unlock(&info->block_group_cache_lock);
8887
8888         /* now that all the block groups are freed, go through and
8889          * free all the space_info structs.  This is only called during
8890          * the final stages of unmount, and so we know nobody is
8891          * using them.  We call synchronize_rcu() once before we start,
8892          * just to be on the safe side.
8893          */
8894         synchronize_rcu();
8895
8896         release_global_block_rsv(info);
8897
8898         while (!list_empty(&info->space_info)) {
8899                 int i;
8900
8901                 space_info = list_entry(info->space_info.next,
8902                                         struct btrfs_space_info,
8903                                         list);
8904                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
8905                         if (WARN_ON(space_info->bytes_pinned > 0 ||
8906                             space_info->bytes_reserved > 0 ||
8907                             space_info->bytes_may_use > 0)) {
8908                                 dump_space_info(space_info, 0, 0);
8909                         }
8910                 }
8911                 list_del(&space_info->list);
8912                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
8913                         struct kobject *kobj;
8914                         kobj = space_info->block_group_kobjs[i];
8915                         space_info->block_group_kobjs[i] = NULL;
8916                         if (kobj) {
8917                                 kobject_del(kobj);
8918                                 kobject_put(kobj);
8919                         }
8920                 }
8921                 kobject_del(&space_info->kobj);
8922                 kobject_put(&space_info->kobj);
8923         }
8924         return 0;
8925 }
8926
8927 static void __link_block_group(struct btrfs_space_info *space_info,
8928                                struct btrfs_block_group_cache *cache)
8929 {
8930         int index = get_block_group_index(cache);
8931         bool first = false;
8932
8933         down_write(&space_info->groups_sem);
8934         if (list_empty(&space_info->block_groups[index]))
8935                 first = true;
8936         list_add_tail(&cache->list, &space_info->block_groups[index]);
8937         up_write(&space_info->groups_sem);
8938
8939         if (first) {
8940                 struct raid_kobject *rkobj;
8941                 int ret;
8942
8943                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
8944                 if (!rkobj)
8945                         goto out_err;
8946                 rkobj->raid_type = index;
8947                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
8948                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
8949                                   "%s", get_raid_name(index));
8950                 if (ret) {
8951                         kobject_put(&rkobj->kobj);
8952                         goto out_err;
8953                 }
8954                 space_info->block_group_kobjs[index] = &rkobj->kobj;
8955         }
8956
8957         return;
8958 out_err:
8959         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
8960 }
8961
8962 static struct btrfs_block_group_cache *
8963 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
8964 {
8965         struct btrfs_block_group_cache *cache;
8966
8967         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8968         if (!cache)
8969                 return NULL;
8970
8971         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8972                                         GFP_NOFS);
8973         if (!cache->free_space_ctl) {
8974                 kfree(cache);
8975                 return NULL;
8976         }
8977
8978         cache->key.objectid = start;
8979         cache->key.offset = size;
8980         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8981
8982         cache->sectorsize = root->sectorsize;
8983         cache->fs_info = root->fs_info;
8984         cache->full_stripe_len = btrfs_full_stripe_len(root,
8985                                                &root->fs_info->mapping_tree,
8986                                                start);
8987         atomic_set(&cache->count, 1);
8988         spin_lock_init(&cache->lock);
8989         init_rwsem(&cache->data_rwsem);
8990         INIT_LIST_HEAD(&cache->list);
8991         INIT_LIST_HEAD(&cache->cluster_list);
8992         INIT_LIST_HEAD(&cache->bg_list);
8993         INIT_LIST_HEAD(&cache->ro_list);
8994         btrfs_init_free_space_ctl(cache);
8995
8996         return cache;
8997 }
8998
8999 int btrfs_read_block_groups(struct btrfs_root *root)
9000 {
9001         struct btrfs_path *path;
9002         int ret;
9003         struct btrfs_block_group_cache *cache;
9004         struct btrfs_fs_info *info = root->fs_info;
9005         struct btrfs_space_info *space_info;
9006         struct btrfs_key key;
9007         struct btrfs_key found_key;
9008         struct extent_buffer *leaf;
9009         int need_clear = 0;
9010         u64 cache_gen;
9011
9012         root = info->extent_root;
9013         key.objectid = 0;
9014         key.offset = 0;
9015         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9016         path = btrfs_alloc_path();
9017         if (!path)
9018                 return -ENOMEM;
9019         path->reada = 1;
9020
9021         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
9022         if (btrfs_test_opt(root, SPACE_CACHE) &&
9023             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
9024                 need_clear = 1;
9025         if (btrfs_test_opt(root, CLEAR_CACHE))
9026                 need_clear = 1;
9027
9028         while (1) {
9029                 ret = find_first_block_group(root, path, &key);
9030                 if (ret > 0)
9031                         break;
9032                 if (ret != 0)
9033                         goto error;
9034
9035                 leaf = path->nodes[0];
9036                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9037
9038                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
9039                                                        found_key.offset);
9040                 if (!cache) {
9041                         ret = -ENOMEM;
9042                         goto error;
9043                 }
9044
9045                 if (need_clear) {
9046                         /*
9047                          * When we mount with old space cache, we need to
9048                          * set BTRFS_DC_CLEAR and set dirty flag.
9049                          *
9050                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9051                          *    truncate the old free space cache inode and
9052                          *    setup a new one.
9053                          * b) Setting 'dirty flag' makes sure that we flush
9054                          *    the new space cache info onto disk.
9055                          */
9056                         cache->disk_cache_state = BTRFS_DC_CLEAR;
9057                         if (btrfs_test_opt(root, SPACE_CACHE))
9058                                 cache->dirty = 1;
9059                 }
9060
9061                 read_extent_buffer(leaf, &cache->item,
9062                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9063                                    sizeof(cache->item));
9064                 cache->flags = btrfs_block_group_flags(&cache->item);
9065
9066                 key.objectid = found_key.objectid + found_key.offset;
9067                 btrfs_release_path(path);
9068
9069                 /*
9070                  * We need to exclude the super stripes now so that the space
9071                  * info has super bytes accounted for, otherwise we'll think
9072                  * we have more space than we actually do.
9073                  */
9074                 ret = exclude_super_stripes(root, cache);
9075                 if (ret) {
9076                         /*
9077                          * We may have excluded something, so call this just in
9078                          * case.
9079                          */
9080                         free_excluded_extents(root, cache);
9081                         btrfs_put_block_group(cache);
9082                         goto error;
9083                 }
9084
9085                 /*
9086                  * check for two cases, either we are full, and therefore
9087                  * don't need to bother with the caching work since we won't
9088                  * find any space, or we are empty, and we can just add all
9089                  * the space in and be done with it.  This saves us _alot_ of
9090                  * time, particularly in the full case.
9091                  */
9092                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
9093                         cache->last_byte_to_unpin = (u64)-1;
9094                         cache->cached = BTRFS_CACHE_FINISHED;
9095                         free_excluded_extents(root, cache);
9096                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9097                         cache->last_byte_to_unpin = (u64)-1;
9098                         cache->cached = BTRFS_CACHE_FINISHED;
9099                         add_new_free_space(cache, root->fs_info,
9100                                            found_key.objectid,
9101                                            found_key.objectid +
9102                                            found_key.offset);
9103                         free_excluded_extents(root, cache);
9104                 }
9105
9106                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
9107                 if (ret) {
9108                         btrfs_remove_free_space_cache(cache);
9109                         btrfs_put_block_group(cache);
9110                         goto error;
9111                 }
9112
9113                 ret = update_space_info(info, cache->flags, found_key.offset,
9114                                         btrfs_block_group_used(&cache->item),
9115                                         &space_info);
9116                 if (ret) {
9117                         btrfs_remove_free_space_cache(cache);
9118                         spin_lock(&info->block_group_cache_lock);
9119                         rb_erase(&cache->cache_node,
9120                                  &info->block_group_cache_tree);
9121                         spin_unlock(&info->block_group_cache_lock);
9122                         btrfs_put_block_group(cache);
9123                         goto error;
9124                 }
9125
9126                 cache->space_info = space_info;
9127                 spin_lock(&cache->space_info->lock);
9128                 cache->space_info->bytes_readonly += cache->bytes_super;
9129                 spin_unlock(&cache->space_info->lock);
9130
9131                 __link_block_group(space_info, cache);
9132
9133                 set_avail_alloc_bits(root->fs_info, cache->flags);
9134                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
9135                         set_block_group_ro(cache, 1);
9136                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9137                         spin_lock(&info->unused_bgs_lock);
9138                         /* Should always be true but just in case. */
9139                         if (list_empty(&cache->bg_list)) {
9140                                 btrfs_get_block_group(cache);
9141                                 list_add_tail(&cache->bg_list,
9142                                               &info->unused_bgs);
9143                         }
9144                         spin_unlock(&info->unused_bgs_lock);
9145                 }
9146         }
9147
9148         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
9149                 if (!(get_alloc_profile(root, space_info->flags) &
9150                       (BTRFS_BLOCK_GROUP_RAID10 |
9151                        BTRFS_BLOCK_GROUP_RAID1 |
9152                        BTRFS_BLOCK_GROUP_RAID5 |
9153                        BTRFS_BLOCK_GROUP_RAID6 |
9154                        BTRFS_BLOCK_GROUP_DUP)))
9155                         continue;
9156                 /*
9157                  * avoid allocating from un-mirrored block group if there are
9158                  * mirrored block groups.
9159                  */
9160                 list_for_each_entry(cache,
9161                                 &space_info->block_groups[BTRFS_RAID_RAID0],
9162                                 list)
9163                         set_block_group_ro(cache, 1);
9164                 list_for_each_entry(cache,
9165                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
9166                                 list)
9167                         set_block_group_ro(cache, 1);
9168         }
9169
9170         init_global_block_rsv(info);
9171         ret = 0;
9172 error:
9173         btrfs_free_path(path);
9174         return ret;
9175 }
9176
9177 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
9178                                        struct btrfs_root *root)
9179 {
9180         struct btrfs_block_group_cache *block_group, *tmp;
9181         struct btrfs_root *extent_root = root->fs_info->extent_root;
9182         struct btrfs_block_group_item item;
9183         struct btrfs_key key;
9184         int ret = 0;
9185
9186         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
9187                 list_del_init(&block_group->bg_list);
9188                 if (ret)
9189                         continue;
9190
9191                 spin_lock(&block_group->lock);
9192                 memcpy(&item, &block_group->item, sizeof(item));
9193                 memcpy(&key, &block_group->key, sizeof(key));
9194                 spin_unlock(&block_group->lock);
9195
9196                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
9197                                         sizeof(item));
9198                 if (ret)
9199                         btrfs_abort_transaction(trans, extent_root, ret);
9200                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
9201                                                key.objectid, key.offset);
9202                 if (ret)
9203                         btrfs_abort_transaction(trans, extent_root, ret);
9204         }
9205 }
9206
9207 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
9208                            struct btrfs_root *root, u64 bytes_used,
9209                            u64 type, u64 chunk_objectid, u64 chunk_offset,
9210                            u64 size)
9211 {
9212         int ret;
9213         struct btrfs_root *extent_root;
9214         struct btrfs_block_group_cache *cache;
9215
9216         extent_root = root->fs_info->extent_root;
9217
9218         btrfs_set_log_full_commit(root->fs_info, trans);
9219
9220         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
9221         if (!cache)
9222                 return -ENOMEM;
9223
9224         btrfs_set_block_group_used(&cache->item, bytes_used);
9225         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
9226         btrfs_set_block_group_flags(&cache->item, type);
9227
9228         cache->flags = type;
9229         cache->last_byte_to_unpin = (u64)-1;
9230         cache->cached = BTRFS_CACHE_FINISHED;
9231         ret = exclude_super_stripes(root, cache);
9232         if (ret) {
9233                 /*
9234                  * We may have excluded something, so call this just in
9235                  * case.
9236                  */
9237                 free_excluded_extents(root, cache);
9238                 btrfs_put_block_group(cache);
9239                 return ret;
9240         }
9241
9242         add_new_free_space(cache, root->fs_info, chunk_offset,
9243                            chunk_offset + size);
9244
9245         free_excluded_extents(root, cache);
9246
9247         ret = btrfs_add_block_group_cache(root->fs_info, cache);
9248         if (ret) {
9249                 btrfs_remove_free_space_cache(cache);
9250                 btrfs_put_block_group(cache);
9251                 return ret;
9252         }
9253
9254         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
9255                                 &cache->space_info);
9256         if (ret) {
9257                 btrfs_remove_free_space_cache(cache);
9258                 spin_lock(&root->fs_info->block_group_cache_lock);
9259                 rb_erase(&cache->cache_node,
9260                          &root->fs_info->block_group_cache_tree);
9261                 spin_unlock(&root->fs_info->block_group_cache_lock);
9262                 btrfs_put_block_group(cache);
9263                 return ret;
9264         }
9265         update_global_block_rsv(root->fs_info);
9266
9267         spin_lock(&cache->space_info->lock);
9268         cache->space_info->bytes_readonly += cache->bytes_super;
9269         spin_unlock(&cache->space_info->lock);
9270
9271         __link_block_group(cache->space_info, cache);
9272
9273         list_add_tail(&cache->bg_list, &trans->new_bgs);
9274
9275         set_avail_alloc_bits(extent_root->fs_info, type);
9276
9277         return 0;
9278 }
9279
9280 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
9281 {
9282         u64 extra_flags = chunk_to_extended(flags) &
9283                                 BTRFS_EXTENDED_PROFILE_MASK;
9284
9285         write_seqlock(&fs_info->profiles_lock);
9286         if (flags & BTRFS_BLOCK_GROUP_DATA)
9287                 fs_info->avail_data_alloc_bits &= ~extra_flags;
9288         if (flags & BTRFS_BLOCK_GROUP_METADATA)
9289                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
9290         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
9291                 fs_info->avail_system_alloc_bits &= ~extra_flags;
9292         write_sequnlock(&fs_info->profiles_lock);
9293 }
9294
9295 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
9296                              struct btrfs_root *root, u64 group_start)
9297 {
9298         struct btrfs_path *path;
9299         struct btrfs_block_group_cache *block_group;
9300         struct btrfs_free_cluster *cluster;
9301         struct btrfs_root *tree_root = root->fs_info->tree_root;
9302         struct btrfs_key key;
9303         struct inode *inode;
9304         struct kobject *kobj = NULL;
9305         int ret;
9306         int index;
9307         int factor;
9308
9309         root = root->fs_info->extent_root;
9310
9311         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
9312         BUG_ON(!block_group);
9313         BUG_ON(!block_group->ro);
9314
9315         /*
9316          * Free the reserved super bytes from this block group before
9317          * remove it.
9318          */
9319         free_excluded_extents(root, block_group);
9320
9321         memcpy(&key, &block_group->key, sizeof(key));
9322         index = get_block_group_index(block_group);
9323         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
9324                                   BTRFS_BLOCK_GROUP_RAID1 |
9325                                   BTRFS_BLOCK_GROUP_RAID10))
9326                 factor = 2;
9327         else
9328                 factor = 1;
9329
9330         /* make sure this block group isn't part of an allocation cluster */
9331         cluster = &root->fs_info->data_alloc_cluster;
9332         spin_lock(&cluster->refill_lock);
9333         btrfs_return_cluster_to_free_space(block_group, cluster);
9334         spin_unlock(&cluster->refill_lock);
9335
9336         /*
9337          * make sure this block group isn't part of a metadata
9338          * allocation cluster
9339          */
9340         cluster = &root->fs_info->meta_alloc_cluster;
9341         spin_lock(&cluster->refill_lock);
9342         btrfs_return_cluster_to_free_space(block_group, cluster);
9343         spin_unlock(&cluster->refill_lock);
9344
9345         path = btrfs_alloc_path();
9346         if (!path) {
9347                 ret = -ENOMEM;
9348                 goto out;
9349         }
9350
9351         inode = lookup_free_space_inode(tree_root, block_group, path);
9352         if (!IS_ERR(inode)) {
9353                 ret = btrfs_orphan_add(trans, inode);
9354                 if (ret) {
9355                         btrfs_add_delayed_iput(inode);
9356                         goto out;
9357                 }
9358                 clear_nlink(inode);
9359                 /* One for the block groups ref */
9360                 spin_lock(&block_group->lock);
9361                 if (block_group->iref) {
9362                         block_group->iref = 0;
9363                         block_group->inode = NULL;
9364                         spin_unlock(&block_group->lock);
9365                         iput(inode);
9366                 } else {
9367                         spin_unlock(&block_group->lock);
9368                 }
9369                 /* One for our lookup ref */
9370                 btrfs_add_delayed_iput(inode);
9371         }
9372
9373         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
9374         key.offset = block_group->key.objectid;
9375         key.type = 0;
9376
9377         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
9378         if (ret < 0)
9379                 goto out;
9380         if (ret > 0)
9381                 btrfs_release_path(path);
9382         if (ret == 0) {
9383                 ret = btrfs_del_item(trans, tree_root, path);
9384                 if (ret)
9385                         goto out;
9386                 btrfs_release_path(path);
9387         }
9388
9389         spin_lock(&root->fs_info->block_group_cache_lock);
9390         rb_erase(&block_group->cache_node,
9391                  &root->fs_info->block_group_cache_tree);
9392
9393         if (root->fs_info->first_logical_byte == block_group->key.objectid)
9394                 root->fs_info->first_logical_byte = (u64)-1;
9395         spin_unlock(&root->fs_info->block_group_cache_lock);
9396
9397         down_write(&block_group->space_info->groups_sem);
9398         /*
9399          * we must use list_del_init so people can check to see if they
9400          * are still on the list after taking the semaphore
9401          */
9402         list_del_init(&block_group->list);
9403         list_del_init(&block_group->ro_list);
9404         if (list_empty(&block_group->space_info->block_groups[index])) {
9405                 kobj = block_group->space_info->block_group_kobjs[index];
9406                 block_group->space_info->block_group_kobjs[index] = NULL;
9407                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
9408         }
9409         up_write(&block_group->space_info->groups_sem);
9410         if (kobj) {
9411                 kobject_del(kobj);
9412                 kobject_put(kobj);
9413         }
9414
9415         if (block_group->cached == BTRFS_CACHE_STARTED)
9416                 wait_block_group_cache_done(block_group);
9417
9418         btrfs_remove_free_space_cache(block_group);
9419
9420         spin_lock(&block_group->space_info->lock);
9421         block_group->space_info->total_bytes -= block_group->key.offset;
9422         block_group->space_info->bytes_readonly -= block_group->key.offset;
9423         block_group->space_info->disk_total -= block_group->key.offset * factor;
9424         spin_unlock(&block_group->space_info->lock);
9425
9426         memcpy(&key, &block_group->key, sizeof(key));
9427
9428         btrfs_put_block_group(block_group);
9429         btrfs_put_block_group(block_group);
9430
9431         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
9432         if (ret > 0)
9433                 ret = -EIO;
9434         if (ret < 0)
9435                 goto out;
9436
9437         ret = btrfs_del_item(trans, root, path);
9438 out:
9439         btrfs_free_path(path);
9440         return ret;
9441 }
9442
9443 /*
9444  * Process the unused_bgs list and remove any that don't have any allocated
9445  * space inside of them.
9446  */
9447 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
9448 {
9449         struct btrfs_block_group_cache *block_group;
9450         struct btrfs_space_info *space_info;
9451         struct btrfs_root *root = fs_info->extent_root;
9452         struct btrfs_trans_handle *trans;
9453         int ret = 0;
9454
9455         if (!fs_info->open)
9456                 return;
9457
9458         spin_lock(&fs_info->unused_bgs_lock);
9459         while (!list_empty(&fs_info->unused_bgs)) {
9460                 u64 start, end;
9461
9462                 block_group = list_first_entry(&fs_info->unused_bgs,
9463                                                struct btrfs_block_group_cache,
9464                                                bg_list);
9465                 space_info = block_group->space_info;
9466                 list_del_init(&block_group->bg_list);
9467                 if (ret || btrfs_mixed_space_info(space_info)) {
9468                         btrfs_put_block_group(block_group);
9469                         continue;
9470                 }
9471                 spin_unlock(&fs_info->unused_bgs_lock);
9472
9473                 /* Don't want to race with allocators so take the groups_sem */
9474                 down_write(&space_info->groups_sem);
9475                 spin_lock(&block_group->lock);
9476                 if (block_group->reserved ||
9477                     btrfs_block_group_used(&block_group->item) ||
9478                     block_group->ro) {
9479                         /*
9480                          * We want to bail if we made new allocations or have
9481                          * outstanding allocations in this block group.  We do
9482                          * the ro check in case balance is currently acting on
9483                          * this block group.
9484                          */
9485                         spin_unlock(&block_group->lock);
9486                         up_write(&space_info->groups_sem);
9487                         goto next;
9488                 }
9489                 spin_unlock(&block_group->lock);
9490
9491                 /* We don't want to force the issue, only flip if it's ok. */
9492                 ret = set_block_group_ro(block_group, 0);
9493                 up_write(&space_info->groups_sem);
9494                 if (ret < 0) {
9495                         ret = 0;
9496                         goto next;
9497                 }
9498
9499                 /*
9500                  * Want to do this before we do anything else so we can recover
9501                  * properly if we fail to join the transaction.
9502                  */
9503                 trans = btrfs_join_transaction(root);
9504                 if (IS_ERR(trans)) {
9505                         btrfs_set_block_group_rw(root, block_group);
9506                         ret = PTR_ERR(trans);
9507                         goto next;
9508                 }
9509
9510                 /*
9511                  * We could have pending pinned extents for this block group,
9512                  * just delete them, we don't care about them anymore.
9513                  */
9514                 start = block_group->key.objectid;
9515                 end = start + block_group->key.offset - 1;
9516                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
9517                                   EXTENT_DIRTY, GFP_NOFS);
9518                 if (ret) {
9519                         btrfs_set_block_group_rw(root, block_group);
9520                         goto end_trans;
9521                 }
9522                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
9523                                   EXTENT_DIRTY, GFP_NOFS);
9524                 if (ret) {
9525                         btrfs_set_block_group_rw(root, block_group);
9526                         goto end_trans;
9527                 }
9528
9529                 /* Reset pinned so btrfs_put_block_group doesn't complain */
9530                 block_group->pinned = 0;
9531
9532                 /*
9533                  * Btrfs_remove_chunk will abort the transaction if things go
9534                  * horribly wrong.
9535                  */
9536                 ret = btrfs_remove_chunk(trans, root,
9537                                          block_group->key.objectid);
9538 end_trans:
9539                 btrfs_end_transaction(trans, root);
9540 next:
9541                 btrfs_put_block_group(block_group);
9542                 spin_lock(&fs_info->unused_bgs_lock);
9543         }
9544         spin_unlock(&fs_info->unused_bgs_lock);
9545 }
9546
9547 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
9548 {
9549         struct btrfs_space_info *space_info;
9550         struct btrfs_super_block *disk_super;
9551         u64 features;
9552         u64 flags;
9553         int mixed = 0;
9554         int ret;
9555
9556         disk_super = fs_info->super_copy;
9557         if (!btrfs_super_root(disk_super))
9558                 return 1;
9559
9560         features = btrfs_super_incompat_flags(disk_super);
9561         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
9562                 mixed = 1;
9563
9564         flags = BTRFS_BLOCK_GROUP_SYSTEM;
9565         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9566         if (ret)
9567                 goto out;
9568
9569         if (mixed) {
9570                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
9571                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9572         } else {
9573                 flags = BTRFS_BLOCK_GROUP_METADATA;
9574                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9575                 if (ret)
9576                         goto out;
9577
9578                 flags = BTRFS_BLOCK_GROUP_DATA;
9579                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9580         }
9581 out:
9582         return ret;
9583 }
9584
9585 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
9586 {
9587         return unpin_extent_range(root, start, end);
9588 }
9589
9590 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
9591                                u64 num_bytes, u64 *actual_bytes)
9592 {
9593         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
9594 }
9595
9596 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
9597 {
9598         struct btrfs_fs_info *fs_info = root->fs_info;
9599         struct btrfs_block_group_cache *cache = NULL;
9600         u64 group_trimmed;
9601         u64 start;
9602         u64 end;
9603         u64 trimmed = 0;
9604         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
9605         int ret = 0;
9606
9607         /*
9608          * try to trim all FS space, our block group may start from non-zero.
9609          */
9610         if (range->len == total_bytes)
9611                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
9612         else
9613                 cache = btrfs_lookup_block_group(fs_info, range->start);
9614
9615         while (cache) {
9616                 if (cache->key.objectid >= (range->start + range->len)) {
9617                         btrfs_put_block_group(cache);
9618                         break;
9619                 }
9620
9621                 start = max(range->start, cache->key.objectid);
9622                 end = min(range->start + range->len,
9623                                 cache->key.objectid + cache->key.offset);
9624
9625                 if (end - start >= range->minlen) {
9626                         if (!block_group_cache_done(cache)) {
9627                                 ret = cache_block_group(cache, 0);
9628                                 if (ret) {
9629                                         btrfs_put_block_group(cache);
9630                                         break;
9631                                 }
9632                                 ret = wait_block_group_cache_done(cache);
9633                                 if (ret) {
9634                                         btrfs_put_block_group(cache);
9635                                         break;
9636                                 }
9637                         }
9638                         ret = btrfs_trim_block_group(cache,
9639                                                      &group_trimmed,
9640                                                      start,
9641                                                      end,
9642                                                      range->minlen);
9643
9644                         trimmed += group_trimmed;
9645                         if (ret) {
9646                                 btrfs_put_block_group(cache);
9647                                 break;
9648                         }
9649                 }
9650
9651                 cache = next_block_group(fs_info->tree_root, cache);
9652         }
9653
9654         range->len = trimmed;
9655         return ret;
9656 }
9657
9658 /*
9659  * btrfs_{start,end}_write() is similar to mnt_{want, drop}_write(),
9660  * they are used to prevent the some tasks writing data into the page cache
9661  * by nocow before the subvolume is snapshoted, but flush the data into
9662  * the disk after the snapshot creation.
9663  */
9664 void btrfs_end_nocow_write(struct btrfs_root *root)
9665 {
9666         percpu_counter_dec(&root->subv_writers->counter);
9667         /*
9668          * Make sure counter is updated before we wake up
9669          * waiters.
9670          */
9671         smp_mb();
9672         if (waitqueue_active(&root->subv_writers->wait))
9673                 wake_up(&root->subv_writers->wait);
9674 }
9675
9676 int btrfs_start_nocow_write(struct btrfs_root *root)
9677 {
9678         if (atomic_read(&root->will_be_snapshoted))
9679                 return 0;
9680
9681         percpu_counter_inc(&root->subv_writers->counter);
9682         /*
9683          * Make sure counter is updated before we check for snapshot creation.
9684          */
9685         smp_mb();
9686         if (atomic_read(&root->will_be_snapshoted)) {
9687                 btrfs_end_nocow_write(root);
9688                 return 0;
9689         }
9690         return 1;
9691 }