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 "compat.h"
  27 #include "hash.h"
  28 #include "ctree.h"
  29 #include "disk-io.h"
  30 #include "print-tree.h"
  31 #include "transaction.h"
  32 #include "volumes.h"
  33 #include "locking.h"
  34 #include "free-space-cache.h"
  35
  36 /* control flags for do_chunk_alloc's force field
  37  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  38  * if we really need one.
  39  *
  40  * CHUNK_ALLOC_FORCE means it must try to allocate one
  41  *
  42  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  43  * if we have very few chunks already allocated.  This is
  44  * used as part of the clustering code to help make sure
  45  * we have a good pool of storage to cluster in, without
  46  * filling the FS with empty chunks
  47  *
  48  */
  49 enum {
  50         CHUNK_ALLOC_NO_FORCE = 0,
  51         CHUNK_ALLOC_FORCE = 1,
  52         CHUNK_ALLOC_LIMITED = 2,
  53 };
  54
  55 static int update_block_group(struct btrfs_trans_handle *trans,
  56                               struct btrfs_root *root,
  57                               u64 bytenr, u64 num_bytes, int alloc);
  58 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  59                                 struct btrfs_root *root,
  60                                 u64 bytenr, u64 num_bytes, u64 parent,
  61                                 u64 root_objectid, u64 owner_objectid,
  62                                 u64 owner_offset, int refs_to_drop,
  63                                 struct btrfs_delayed_extent_op *extra_op);
  64 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  65                                     struct extent_buffer *leaf,
  66                                     struct btrfs_extent_item *ei);
  67 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  68                                       struct btrfs_root *root,
  69                                       u64 parent, u64 root_objectid,
  70                                       u64 flags, u64 owner, u64 offset,
  71                                       struct btrfs_key *ins, int ref_mod);
  72 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  73                                      struct btrfs_root *root,
  74                                      u64 parent, u64 root_objectid,
  75                                      u64 flags, struct btrfs_disk_key *key,
  76                                      int level, struct btrfs_key *ins);
  77 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  78                           struct btrfs_root *extent_root, u64 alloc_bytes,
  79                           u64 flags, int force);
  80 static int find_next_key(struct btrfs_path *path, int level,
  81                          struct btrfs_key *key);
  82 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
  83                             int dump_block_groups);
  84
  85 static noinline int
  86 block_group_cache_done(struct btrfs_block_group_cache *cache)
  87 {
  88         smp_mb();
  89         return cache->cached == BTRFS_CACHE_FINISHED;
  90 }
  91
  92 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
  93 {
  94         return (cache->flags & bits) == bits;
  95 }
  96
  97 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
  98 {
  99         atomic_inc(&cache->count);
 100 }
 101
 102 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 103 {
 104         if (atomic_dec_and_test(&cache->count)) {
 105                 WARN_ON(cache->pinned > 0);
 106                 WARN_ON(cache->reserved > 0);
 107                 WARN_ON(cache->reserved_pinned > 0);
 108                 kfree(cache->free_space_ctl);
 109                 kfree(cache);
 110         }
 111 }
 112
 113 /*
 114  * this adds the block group to the fs_info rb tree for the block group
 115  * cache
 116  */
 117 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 118                                 struct btrfs_block_group_cache *block_group)
 119 {
 120         struct rb_node **p;
 121         struct rb_node *parent = NULL;
 122         struct btrfs_block_group_cache *cache;
 123
 124         spin_lock(&info->block_group_cache_lock);
 125         p = &info->block_group_cache_tree.rb_node;
 126
 127         while (*p) {
 128                 parent = *p;
 129                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 130                                  cache_node);
 131                 if (block_group->key.objectid < cache->key.objectid) {
 132                         p = &(*p)->rb_left;
 133                 } else if (block_group->key.objectid > cache->key.objectid) {
 134                         p = &(*p)->rb_right;
 135                 } else {
 136                         spin_unlock(&info->block_group_cache_lock);
 137                         return -EEXIST;
 138                 }
 139         }
 140
 141         rb_link_node(&block_group->cache_node, parent, p);
 142         rb_insert_color(&block_group->cache_node,
 143                         &info->block_group_cache_tree);
 144         spin_unlock(&info->block_group_cache_lock);
 145
 146         return 0;
 147 }
 148
 149 /*
 150  * This will return the block group at or after bytenr if contains is 0, else
 151  * it will return the block group that contains the bytenr
 152  */
 153 static struct btrfs_block_group_cache *
 154 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 155                               int contains)
 156 {
 157         struct btrfs_block_group_cache *cache, *ret = NULL;
 158         struct rb_node *n;
 159         u64 end, start;
 160
 161         spin_lock(&info->block_group_cache_lock);
 162         n = info->block_group_cache_tree.rb_node;
 163
 164         while (n) {
 165                 cache = rb_entry(n, struct btrfs_block_group_cache,
 166                                  cache_node);
 167                 end = cache->key.objectid + cache->key.offset - 1;
 168                 start = cache->key.objectid;
 169
 170                 if (bytenr < start) {
 171                         if (!contains && (!ret || start < ret->key.objectid))
 172                                 ret = cache;
 173                         n = n->rb_left;
 174                 } else if (bytenr > start) {
 175                         if (contains && bytenr <= end) {
 176                                 ret = cache;
 177                                 break;
 178                         }
 179                         n = n->rb_right;
 180                 } else {
 181                         ret = cache;
 182                         break;
 183                 }
 184         }
 185         if (ret)
 186                 btrfs_get_block_group(ret);
 187         spin_unlock(&info->block_group_cache_lock);
 188
 189         return ret;
 190 }
 191
 192 static int add_excluded_extent(struct btrfs_root *root,
 193                                u64 start, u64 num_bytes)
 194 {
 195         u64 end = start + num_bytes - 1;
 196         set_extent_bits(&root->fs_info->freed_extents[0],
 197                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 198         set_extent_bits(&root->fs_info->freed_extents[1],
 199                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 200         return 0;
 201 }
 202
 203 static void free_excluded_extents(struct btrfs_root *root,
 204                                   struct btrfs_block_group_cache *cache)
 205 {
 206         u64 start, end;
 207
 208         start = cache->key.objectid;
 209         end = start + cache->key.offset - 1;
 210
 211         clear_extent_bits(&root->fs_info->freed_extents[0],
 212                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 213         clear_extent_bits(&root->fs_info->freed_extents[1],
 214                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 215 }
 216
 217 static int exclude_super_stripes(struct btrfs_root *root,
 218                                  struct btrfs_block_group_cache *cache)
 219 {
 220         u64 bytenr;
 221         u64 *logical;
 222         int stripe_len;
 223         int i, nr, ret;
 224
 225         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 226                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 227                 cache->bytes_super += stripe_len;
 228                 ret = add_excluded_extent(root, cache->key.objectid,
 229                                           stripe_len);
 230                 BUG_ON(ret);
 231         }
 232
 233         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 234                 bytenr = btrfs_sb_offset(i);
 235                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 236                                        cache->key.objectid, bytenr,
 237                                        0, &logical, &nr, &stripe_len);
 238                 BUG_ON(ret);
 239
 240                 while (nr--) {
 241                         cache->bytes_super += stripe_len;
 242                         ret = add_excluded_extent(root, logical[nr],
 243                                                   stripe_len);
 244                         BUG_ON(ret);
 245                 }
 246
 247                 kfree(logical);
 248         }
 249         return 0;
 250 }
 251
 252 static struct btrfs_caching_control *
 253 get_caching_control(struct btrfs_block_group_cache *cache)
 254 {
 255         struct btrfs_caching_control *ctl;
 256
 257         spin_lock(&cache->lock);
 258         if (cache->cached != BTRFS_CACHE_STARTED) {
 259                 spin_unlock(&cache->lock);
 260                 return NULL;
 261         }
 262
 263         /* We're loading it the fast way, so we don't have a caching_ctl. */
 264         if (!cache->caching_ctl) {
 265                 spin_unlock(&cache->lock);
 266                 return NULL;
 267         }
 268
 269         ctl = cache->caching_ctl;
 270         atomic_inc(&ctl->count);
 271         spin_unlock(&cache->lock);
 272         return ctl;
 273 }
 274
 275 static void put_caching_control(struct btrfs_caching_control *ctl)
 276 {
 277         if (atomic_dec_and_test(&ctl->count))
 278                 kfree(ctl);
 279 }
 280
 281 /*
 282  * this is only called by cache_block_group, since we could have freed extents
 283  * we need to check the pinned_extents for any extents that can't be used yet
 284  * since their free space will be released as soon as the transaction commits.
 285  */
 286 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 287                               struct btrfs_fs_info *info, u64 start, u64 end)
 288 {
 289         u64 extent_start, extent_end, size, total_added = 0;
 290         int ret;
 291
 292         while (start < end) {
 293                 ret = find_first_extent_bit(info->pinned_extents, start,
 294                                             &extent_start, &extent_end,
 295                                             EXTENT_DIRTY | EXTENT_UPTODATE);
 296                 if (ret)
 297                         break;
 298
 299                 if (extent_start <= start) {
 300                         start = extent_end + 1;
 301                 } else if (extent_start > start && extent_start < end) {
 302                         size = extent_start - start;
 303                         total_added += size;
 304                         ret = btrfs_add_free_space(block_group, start,
 305                                                    size);
 306                         BUG_ON(ret);
 307                         start = extent_end + 1;
 308                 } else {
 309                         break;
 310                 }
 311         }
 312
 313         if (start < end) {
 314                 size = end - start;
 315                 total_added += size;
 316                 ret = btrfs_add_free_space(block_group, start, size);
 317                 BUG_ON(ret);
 318         }
 319
 320         return total_added;
 321 }
 322
 323 static int caching_kthread(void *data)
 324 {
 325         struct btrfs_block_group_cache *block_group = data;
 326         struct btrfs_fs_info *fs_info = block_group->fs_info;
 327         struct btrfs_caching_control *caching_ctl = block_group->caching_ctl;
 328         struct btrfs_root *extent_root = fs_info->extent_root;
 329         struct btrfs_path *path;
 330         struct extent_buffer *leaf;
 331         struct btrfs_key key;
 332         u64 total_found = 0;
 333         u64 last = 0;
 334         u32 nritems;
 335         int ret = 0;
 336
 337         path = btrfs_alloc_path();
 338         if (!path)
 339                 return -ENOMEM;
 340
 341         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 342
 343         /*
 344          * We don't want to deadlock with somebody trying to allocate a new
 345          * extent for the extent root while also trying to search the extent
 346          * root to add free space.  So we skip locking and search the commit
 347          * root, since its read-only
 348          */
 349         path->skip_locking = 1;
 350         path->search_commit_root = 1;
 351         path->reada = 2;
 352
 353         key.objectid = last;
 354         key.offset = 0;
 355         key.type = BTRFS_EXTENT_ITEM_KEY;
 356 again:
 357         mutex_lock(&caching_ctl->mutex);
 358         /* need to make sure the commit_root doesn't disappear */
 359         down_read(&fs_info->extent_commit_sem);
 360
 361         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 362         if (ret < 0)
 363                 goto err;
 364
 365         leaf = path->nodes[0];
 366         nritems = btrfs_header_nritems(leaf);
 367
 368         while (1) {
 369                 smp_mb();
 370                 if (fs_info->closing > 1) {
 371                         last = (u64)-1;
 372                         break;
 373                 }
 374
 375                 if (path->slots[0] < nritems) {
 376                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 377                 } else {
 378                         ret = find_next_key(path, 0, &key);
 379                         if (ret)
 380                                 break;
 381
 382                         caching_ctl->progress = last;
 383                         btrfs_release_path(path);
 384                         up_read(&fs_info->extent_commit_sem);
 385                         mutex_unlock(&caching_ctl->mutex);
 386                         if (btrfs_transaction_in_commit(fs_info))
 387                                 schedule_timeout(1);
 388                         else
 389                                 cond_resched();
 390                         goto again;
 391                 }
 392
 393                 if (key.objectid < block_group->key.objectid) {
 394                         path->slots[0]++;
 395                         continue;
 396                 }
 397
 398                 if (key.objectid >= block_group->key.objectid +
 399                     block_group->key.offset)
 400                         break;
 401
 402                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 403                         total_found += add_new_free_space(block_group,
 404                                                           fs_info, last,
 405                                                           key.objectid);
 406                         last = key.objectid + key.offset;
 407
 408                         if (total_found > (1024 * 1024 * 2)) {
 409                                 total_found = 0;
 410                                 wake_up(&caching_ctl->wait);
 411                         }
 412                 }
 413                 path->slots[0]++;
 414         }
 415         ret = 0;
 416
 417         total_found += add_new_free_space(block_group, fs_info, last,
 418                                           block_group->key.objectid +
 419                                           block_group->key.offset);
 420         caching_ctl->progress = (u64)-1;
 421
 422         spin_lock(&block_group->lock);
 423         block_group->caching_ctl = NULL;
 424         block_group->cached = BTRFS_CACHE_FINISHED;
 425         spin_unlock(&block_group->lock);
 426
 427 err:
 428         btrfs_free_path(path);
 429         up_read(&fs_info->extent_commit_sem);
 430
 431         free_excluded_extents(extent_root, block_group);
 432
 433         mutex_unlock(&caching_ctl->mutex);
 434         wake_up(&caching_ctl->wait);
 435
 436         put_caching_control(caching_ctl);
 437         atomic_dec(&block_group->space_info->caching_threads);
 438         btrfs_put_block_group(block_group);
 439
 440         return 0;
 441 }
 442
 443 static int cache_block_group(struct btrfs_block_group_cache *cache,
 444                              struct btrfs_trans_handle *trans,
 445                              struct btrfs_root *root,
 446                              int load_cache_only)
 447 {
 448         struct btrfs_fs_info *fs_info = cache->fs_info;
 449         struct btrfs_caching_control *caching_ctl;
 450         struct task_struct *tsk;
 451         int ret = 0;
 452
 453         smp_mb();
 454         if (cache->cached != BTRFS_CACHE_NO)
 455                 return 0;
 456
 457         /*
 458          * We can't do the read from on-disk cache during a commit since we need
 459          * to have the normal tree locking.  Also if we are currently trying to
 460          * allocate blocks for the tree root we can't do the fast caching since
 461          * we likely hold important locks.
 462          */
 463         if (trans && (!trans->transaction->in_commit) &&
 464             (root && root != root->fs_info->tree_root)) {
 465                 spin_lock(&cache->lock);
 466                 if (cache->cached != BTRFS_CACHE_NO) {
 467                         spin_unlock(&cache->lock);
 468                         return 0;
 469                 }
 470                 cache->cached = BTRFS_CACHE_STARTED;
 471                 spin_unlock(&cache->lock);
 472
 473                 ret = load_free_space_cache(fs_info, cache);
 474
 475                 spin_lock(&cache->lock);
 476                 if (ret == 1) {
 477                         cache->cached = BTRFS_CACHE_FINISHED;
 478                         cache->last_byte_to_unpin = (u64)-1;
 479                 } else {
 480                         cache->cached = BTRFS_CACHE_NO;
 481                 }
 482                 spin_unlock(&cache->lock);
 483                 if (ret == 1) {
 484                         free_excluded_extents(fs_info->extent_root, cache);
 485                         return 0;
 486                 }
 487         }
 488
 489         if (load_cache_only)
 490                 return 0;
 491
 492         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 493         BUG_ON(!caching_ctl);
 494
 495         INIT_LIST_HEAD(&caching_ctl->list);
 496         mutex_init(&caching_ctl->mutex);
 497         init_waitqueue_head(&caching_ctl->wait);
 498         caching_ctl->block_group = cache;
 499         caching_ctl->progress = cache->key.objectid;
 500         /* one for caching kthread, one for caching block group list */
 501         atomic_set(&caching_ctl->count, 2);
 502
 503         spin_lock(&cache->lock);
 504         if (cache->cached != BTRFS_CACHE_NO) {
 505                 spin_unlock(&cache->lock);
 506                 kfree(caching_ctl);
 507                 return 0;
 508         }
 509         cache->caching_ctl = caching_ctl;
 510         cache->cached = BTRFS_CACHE_STARTED;
 511         spin_unlock(&cache->lock);
 512
 513         down_write(&fs_info->extent_commit_sem);
 514         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 515         up_write(&fs_info->extent_commit_sem);
 516
 517         atomic_inc(&cache->space_info->caching_threads);
 518         btrfs_get_block_group(cache);
 519
 520         tsk = kthread_run(caching_kthread, cache, "btrfs-cache-%llu\n",
 521                           cache->key.objectid);
 522         if (IS_ERR(tsk)) {
 523                 ret = PTR_ERR(tsk);
 524                 printk(KERN_ERR "error running thread %d\n", ret);
 525                 BUG();
 526         }
 527
 528         return ret;
 529 }
 530
 531 /*
 532  * return the block group that starts at or after bytenr
 533  */
 534 static struct btrfs_block_group_cache *
 535 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 536 {
 537         struct btrfs_block_group_cache *cache;
 538
 539         cache = block_group_cache_tree_search(info, bytenr, 0);
 540
 541         return cache;
 542 }
 543
 544 /*
 545  * return the block group that contains the given bytenr
 546  */
 547 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 548                                                  struct btrfs_fs_info *info,
 549                                                  u64 bytenr)
 550 {
 551         struct btrfs_block_group_cache *cache;
 552
 553         cache = block_group_cache_tree_search(info, bytenr, 1);
 554
 555         return cache;
 556 }
 557
 558 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 559                                                   u64 flags)
 560 {
 561         struct list_head *head = &info->space_info;
 562         struct btrfs_space_info *found;
 563
 564         flags &= BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_SYSTEM |
 565                  BTRFS_BLOCK_GROUP_METADATA;
 566
 567         rcu_read_lock();
 568         list_for_each_entry_rcu(found, head, list) {
 569                 if (found->flags & flags) {
 570                         rcu_read_unlock();
 571                         return found;
 572                 }
 573         }
 574         rcu_read_unlock();
 575         return NULL;
 576 }
 577
 578 /*
 579  * after adding space to the filesystem, we need to clear the full flags
 580  * on all the space infos.
 581  */
 582 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 583 {
 584         struct list_head *head = &info->space_info;
 585         struct btrfs_space_info *found;
 586
 587         rcu_read_lock();
 588         list_for_each_entry_rcu(found, head, list)
 589                 found->full = 0;
 590         rcu_read_unlock();
 591 }
 592
 593 static u64 div_factor(u64 num, int factor)
 594 {
 595         if (factor == 10)
 596                 return num;
 597         num *= factor;
 598         do_div(num, 10);
 599         return num;
 600 }
 601
 602 static u64 div_factor_fine(u64 num, int factor)
 603 {
 604         if (factor == 100)
 605                 return num;
 606         num *= factor;
 607         do_div(num, 100);
 608         return num;
 609 }
 610
 611 u64 btrfs_find_block_group(struct btrfs_root *root,
 612                            u64 search_start, u64 search_hint, int owner)
 613 {
 614         struct btrfs_block_group_cache *cache;
 615         u64 used;
 616         u64 last = max(search_hint, search_start);
 617         u64 group_start = 0;
 618         int full_search = 0;
 619         int factor = 9;
 620         int wrapped = 0;
 621 again:
 622         while (1) {
 623                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
 624                 if (!cache)
 625                         break;
 626
 627                 spin_lock(&cache->lock);
 628                 last = cache->key.objectid + cache->key.offset;
 629                 used = btrfs_block_group_used(&cache->item);
 630
 631                 if ((full_search || !cache->ro) &&
 632                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
 633                         if (used + cache->pinned + cache->reserved <
 634                             div_factor(cache->key.offset, factor)) {
 635                                 group_start = cache->key.objectid;
 636                                 spin_unlock(&cache->lock);
 637                                 btrfs_put_block_group(cache);
 638                                 goto found;
 639                         }
 640                 }
 641                 spin_unlock(&cache->lock);
 642                 btrfs_put_block_group(cache);
 643                 cond_resched();
 644         }
 645         if (!wrapped) {
 646                 last = search_start;
 647                 wrapped = 1;
 648                 goto again;
 649         }
 650         if (!full_search && factor < 10) {
 651                 last = search_start;
 652                 full_search = 1;
 653                 factor = 10;
 654                 goto again;
 655         }
 656 found:
 657         return group_start;
 658 }
 659
 660 /* simple helper to search for an existing extent at a given offset */
 661 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 662 {
 663         int ret;
 664         struct btrfs_key key;
 665         struct btrfs_path *path;
 666
 667         path = btrfs_alloc_path();
 668         BUG_ON(!path);
 669         key.objectid = start;
 670         key.offset = len;
 671         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 672         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 673                                 0, 0);
 674         btrfs_free_path(path);
 675         return ret;
 676 }
 677
 678 /*
 679  * helper function to lookup reference count and flags of extent.
 680  *
 681  * the head node for delayed ref is used to store the sum of all the
 682  * reference count modifications queued up in the rbtree. the head
 683  * node may also store the extent flags to set. This way you can check
 684  * to see what the reference count and extent flags would be if all of
 685  * the delayed refs are not processed.
 686  */
 687 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 688                              struct btrfs_root *root, u64 bytenr,
 689                              u64 num_bytes, u64 *refs, u64 *flags)
 690 {
 691         struct btrfs_delayed_ref_head *head;
 692         struct btrfs_delayed_ref_root *delayed_refs;
 693         struct btrfs_path *path;
 694         struct btrfs_extent_item *ei;
 695         struct extent_buffer *leaf;
 696         struct btrfs_key key;
 697         u32 item_size;
 698         u64 num_refs;
 699         u64 extent_flags;
 700         int ret;
 701
 702         path = btrfs_alloc_path();
 703         if (!path)
 704                 return -ENOMEM;
 705
 706         key.objectid = bytenr;
 707         key.type = BTRFS_EXTENT_ITEM_KEY;
 708         key.offset = num_bytes;
 709         if (!trans) {
 710                 path->skip_locking = 1;
 711                 path->search_commit_root = 1;
 712         }
 713 again:
 714         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 715                                 &key, path, 0, 0);
 716         if (ret < 0)
 717                 goto out_free;
 718
 719         if (ret == 0) {
 720                 leaf = path->nodes[0];
 721                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 722                 if (item_size >= sizeof(*ei)) {
 723                         ei = btrfs_item_ptr(leaf, path->slots[0],
 724                                             struct btrfs_extent_item);
 725                         num_refs = btrfs_extent_refs(leaf, ei);
 726                         extent_flags = btrfs_extent_flags(leaf, ei);
 727                 } else {
 728 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 729                         struct btrfs_extent_item_v0 *ei0;
 730                         BUG_ON(item_size != sizeof(*ei0));
 731                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 732                                              struct btrfs_extent_item_v0);
 733                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 734                         /* FIXME: this isn't correct for data */
 735                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 736 #else
 737                         BUG();
 738 #endif
 739                 }
 740                 BUG_ON(num_refs == 0);
 741         } else {
 742                 num_refs = 0;
 743                 extent_flags = 0;
 744                 ret = 0;
 745         }
 746
 747         if (!trans)
 748                 goto out;
 749
 750         delayed_refs = &trans->transaction->delayed_refs;
 751         spin_lock(&delayed_refs->lock);
 752         head = btrfs_find_delayed_ref_head(trans, bytenr);
 753         if (head) {
 754                 if (!mutex_trylock(&head->mutex)) {
 755                         atomic_inc(&head->node.refs);
 756                         spin_unlock(&delayed_refs->lock);
 757
 758                         btrfs_release_path(path);
 759
 760                         /*
 761                          * Mutex was contended, block until it's released and try
 762                          * again
 763                          */
 764                         mutex_lock(&head->mutex);
 765                         mutex_unlock(&head->mutex);
 766                         btrfs_put_delayed_ref(&head->node);
 767                         goto again;
 768                 }
 769                 if (head->extent_op && head->extent_op->update_flags)
 770                         extent_flags |= head->extent_op->flags_to_set;
 771                 else
 772                         BUG_ON(num_refs == 0);
 773
 774                 num_refs += head->node.ref_mod;
 775                 mutex_unlock(&head->mutex);
 776         }
 777         spin_unlock(&delayed_refs->lock);
 778 out:
 779         WARN_ON(num_refs == 0);
 780         if (refs)
 781                 *refs = num_refs;
 782         if (flags)
 783                 *flags = extent_flags;
 784 out_free:
 785         btrfs_free_path(path);
 786         return ret;
 787 }
 788
 789 /*
 790  * Back reference rules.  Back refs have three main goals:
 791  *
 792  * 1) differentiate between all holders of references to an extent so that
 793  *    when a reference is dropped we can make sure it was a valid reference
 794  *    before freeing the extent.
 795  *
 796  * 2) Provide enough information to quickly find the holders of an extent
 797  *    if we notice a given block is corrupted or bad.
 798  *
 799  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 800  *    maintenance.  This is actually the same as #2, but with a slightly
 801  *    different use case.
 802  *
 803  * There are two kinds of back refs. The implicit back refs is optimized
 804  * for pointers in non-shared tree blocks. For a given pointer in a block,
 805  * back refs of this kind provide information about the block's owner tree
 806  * and the pointer's key. These information allow us to find the block by
 807  * b-tree searching. The full back refs is for pointers in tree blocks not
 808  * referenced by their owner trees. The location of tree block is recorded
 809  * in the back refs. Actually the full back refs is generic, and can be
 810  * used in all cases the implicit back refs is used. The major shortcoming
 811  * of the full back refs is its overhead. Every time a tree block gets
 812  * COWed, we have to update back refs entry for all pointers in it.
 813  *
 814  * For a newly allocated tree block, we use implicit back refs for
 815  * pointers in it. This means most tree related operations only involve
 816  * implicit back refs. For a tree block created in old transaction, the
 817  * only way to drop a reference to it is COW it. So we can detect the
 818  * event that tree block loses its owner tree's reference and do the
 819  * back refs conversion.
 820  *
 821  * When a tree block is COW'd through a tree, there are four cases:
 822  *
 823  * The reference count of the block is one and the tree is the block's
 824  * owner tree. Nothing to do in this case.
 825  *
 826  * The reference count of the block is one and the tree is not the
 827  * block's owner tree. In this case, full back refs is used for pointers
 828  * in the block. Remove these full back refs, add implicit back refs for
 829  * every pointers in the new block.
 830  *
 831  * The reference count of the block is greater than one and the tree is
 832  * the block's owner tree. In this case, implicit back refs is used for
 833  * pointers in the block. Add full back refs for every pointers in the
 834  * block, increase lower level extents' reference counts. The original
 835  * implicit back refs are entailed to the new block.
 836  *
 837  * The reference count of the block is greater than one and the tree is
 838  * not the block's owner tree. Add implicit back refs for every pointer in
 839  * the new block, increase lower level extents' reference count.
 840  *
 841  * Back Reference Key composing:
 842  *
 843  * The key objectid corresponds to the first byte in the extent,
 844  * The key type is used to differentiate between types of back refs.
 845  * There are different meanings of the key offset for different types
 846  * of back refs.
 847  *
 848  * File extents can be referenced by:
 849  *
 850  * - multiple snapshots, subvolumes, or different generations in one subvol
 851  * - different files inside a single subvolume
 852  * - different offsets inside a file (bookend extents in file.c)
 853  *
 854  * The extent ref structure for the implicit back refs has fields for:
 855  *
 856  * - Objectid of the subvolume root
 857  * - objectid of the file holding the reference
 858  * - original offset in the file
 859  * - how many bookend extents
 860  *
 861  * The key offset for the implicit back refs is hash of the first
 862  * three fields.
 863  *
 864  * The extent ref structure for the full back refs has field for:
 865  *
 866  * - number of pointers in the tree leaf
 867  *
 868  * The key offset for the implicit back refs is the first byte of
 869  * the tree leaf
 870  *
 871  * When a file extent is allocated, The implicit back refs is used.
 872  * the fields are filled in:
 873  *
 874  *     (root_key.objectid, inode objectid, offset in file, 1)
 875  *
 876  * When a file extent is removed file truncation, we find the
 877  * corresponding implicit back refs and check the following fields:
 878  *
 879  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 880  *
 881  * Btree extents can be referenced by:
 882  *
 883  * - Different subvolumes
 884  *
 885  * Both the implicit back refs and the full back refs for tree blocks
 886  * only consist of key. The key offset for the implicit back refs is
 887  * objectid of block's owner tree. The key offset for the full back refs
 888  * is the first byte of parent block.
 889  *
 890  * When implicit back refs is used, information about the lowest key and
 891  * level of the tree block are required. These information are stored in
 892  * tree block info structure.
 893  */
 894
 895 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 896 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 897                                   struct btrfs_root *root,
 898                                   struct btrfs_path *path,
 899                                   u64 owner, u32 extra_size)
 900 {
 901         struct btrfs_extent_item *item;
 902         struct btrfs_extent_item_v0 *ei0;
 903         struct btrfs_extent_ref_v0 *ref0;
 904         struct btrfs_tree_block_info *bi;
 905         struct extent_buffer *leaf;
 906         struct btrfs_key key;
 907         struct btrfs_key found_key;
 908         u32 new_size = sizeof(*item);
 909         u64 refs;
 910         int ret;
 911
 912         leaf = path->nodes[0];
 913         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 914
 915         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 916         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 917                              struct btrfs_extent_item_v0);
 918         refs = btrfs_extent_refs_v0(leaf, ei0);
 919
 920         if (owner == (u64)-1) {
 921                 while (1) {
 922                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 923                                 ret = btrfs_next_leaf(root, path);
 924                                 if (ret < 0)
 925                                         return ret;
 926                                 BUG_ON(ret > 0);
 927                                 leaf = path->nodes[0];
 928                         }
 929                         btrfs_item_key_to_cpu(leaf, &found_key,
 930                                               path->slots[0]);
 931                         BUG_ON(key.objectid != found_key.objectid);
 932                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 933                                 path->slots[0]++;
 934                                 continue;
 935                         }
 936                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 937                                               struct btrfs_extent_ref_v0);
 938                         owner = btrfs_ref_objectid_v0(leaf, ref0);
 939                         break;
 940                 }
 941         }
 942         btrfs_release_path(path);
 943
 944         if (owner < BTRFS_FIRST_FREE_OBJECTID)
 945                 new_size += sizeof(*bi);
 946
 947         new_size -= sizeof(*ei0);
 948         ret = btrfs_search_slot(trans, root, &key, path,
 949                                 new_size + extra_size, 1);
 950         if (ret < 0)
 951                 return ret;
 952         BUG_ON(ret);
 953
 954         ret = btrfs_extend_item(trans, root, path, new_size);
 955
 956         leaf = path->nodes[0];
 957         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 958         btrfs_set_extent_refs(leaf, item, refs);
 959         /* FIXME: get real generation */
 960         btrfs_set_extent_generation(leaf, item, 0);
 961         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 962                 btrfs_set_extent_flags(leaf, item,
 963                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
 964                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
 965                 bi = (struct btrfs_tree_block_info *)(item + 1);
 966                 /* FIXME: get first key of the block */
 967                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
 968                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
 969         } else {
 970                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
 971         }
 972         btrfs_mark_buffer_dirty(leaf);
 973         return 0;
 974 }
 975 #endif
 976
 977 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
 978 {
 979         u32 high_crc = ~(u32)0;
 980         u32 low_crc = ~(u32)0;
 981         __le64 lenum;
 982
 983         lenum = cpu_to_le64(root_objectid);
 984         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
 985         lenum = cpu_to_le64(owner);
 986         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
 987         lenum = cpu_to_le64(offset);
 988         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
 989
 990         return ((u64)high_crc << 31) ^ (u64)low_crc;
 991 }
 992
 993 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
 994                                      struct btrfs_extent_data_ref *ref)
 995 {
 996         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
 997                                     btrfs_extent_data_ref_objectid(leaf, ref),
 998                                     btrfs_extent_data_ref_offset(leaf, ref));
 999 }
1000
1001 static int match_extent_data_ref(struct extent_buffer *leaf,
1002                                  struct btrfs_extent_data_ref *ref,
1003                                  u64 root_objectid, u64 owner, u64 offset)
1004 {
1005         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1006             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1007             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1008                 return 0;
1009         return 1;
1010 }
1011
1012 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1013                                            struct btrfs_root *root,
1014                                            struct btrfs_path *path,
1015                                            u64 bytenr, u64 parent,
1016                                            u64 root_objectid,
1017                                            u64 owner, u64 offset)
1018 {
1019         struct btrfs_key key;
1020         struct btrfs_extent_data_ref *ref;
1021         struct extent_buffer *leaf;
1022         u32 nritems;
1023         int ret;
1024         int recow;
1025         int err = -ENOENT;
1026
1027         key.objectid = bytenr;
1028         if (parent) {
1029                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1030                 key.offset = parent;
1031         } else {
1032                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1033                 key.offset = hash_extent_data_ref(root_objectid,
1034                                                   owner, offset);
1035         }
1036 again:
1037         recow = 0;
1038         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1039         if (ret < 0) {
1040                 err = ret;
1041                 goto fail;
1042         }
1043
1044         if (parent) {
1045                 if (!ret)
1046                         return 0;
1047 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1048                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1049                 btrfs_release_path(path);
1050                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1051                 if (ret < 0) {
1052                         err = ret;
1053                         goto fail;
1054                 }
1055                 if (!ret)
1056                         return 0;
1057 #endif
1058                 goto fail;
1059         }
1060
1061         leaf = path->nodes[0];
1062         nritems = btrfs_header_nritems(leaf);
1063         while (1) {
1064                 if (path->slots[0] >= nritems) {
1065                         ret = btrfs_next_leaf(root, path);
1066                         if (ret < 0)
1067                                 err = ret;
1068                         if (ret)
1069                                 goto fail;
1070
1071                         leaf = path->nodes[0];
1072                         nritems = btrfs_header_nritems(leaf);
1073                         recow = 1;
1074                 }
1075
1076                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1077                 if (key.objectid != bytenr ||
1078                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1079                         goto fail;
1080
1081                 ref = btrfs_item_ptr(leaf, path->slots[0],
1082                                      struct btrfs_extent_data_ref);
1083
1084                 if (match_extent_data_ref(leaf, ref, root_objectid,
1085                                           owner, offset)) {
1086                         if (recow) {
1087                                 btrfs_release_path(path);
1088                                 goto again;
1089                         }
1090                         err = 0;
1091                         break;
1092                 }
1093                 path->slots[0]++;
1094         }
1095 fail:
1096         return err;
1097 }
1098
1099 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1100                                            struct btrfs_root *root,
1101                                            struct btrfs_path *path,
1102                                            u64 bytenr, u64 parent,
1103                                            u64 root_objectid, u64 owner,
1104                                            u64 offset, int refs_to_add)
1105 {
1106         struct btrfs_key key;
1107         struct extent_buffer *leaf;
1108         u32 size;
1109         u32 num_refs;
1110         int ret;
1111
1112         key.objectid = bytenr;
1113         if (parent) {
1114                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1115                 key.offset = parent;
1116                 size = sizeof(struct btrfs_shared_data_ref);
1117         } else {
1118                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1119                 key.offset = hash_extent_data_ref(root_objectid,
1120                                                   owner, offset);
1121                 size = sizeof(struct btrfs_extent_data_ref);
1122         }
1123
1124         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1125         if (ret && ret != -EEXIST)
1126                 goto fail;
1127
1128         leaf = path->nodes[0];
1129         if (parent) {
1130                 struct btrfs_shared_data_ref *ref;
1131                 ref = btrfs_item_ptr(leaf, path->slots[0],
1132                                      struct btrfs_shared_data_ref);
1133                 if (ret == 0) {
1134                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1135                 } else {
1136                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1137                         num_refs += refs_to_add;
1138                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1139                 }
1140         } else {
1141                 struct btrfs_extent_data_ref *ref;
1142                 while (ret == -EEXIST) {
1143                         ref = btrfs_item_ptr(leaf, path->slots[0],
1144                                              struct btrfs_extent_data_ref);
1145                         if (match_extent_data_ref(leaf, ref, root_objectid,
1146                                                   owner, offset))
1147                                 break;
1148                         btrfs_release_path(path);
1149                         key.offset++;
1150                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1151                                                       size);
1152                         if (ret && ret != -EEXIST)
1153                                 goto fail;
1154
1155                         leaf = path->nodes[0];
1156                 }
1157                 ref = btrfs_item_ptr(leaf, path->slots[0],
1158                                      struct btrfs_extent_data_ref);
1159                 if (ret == 0) {
1160                         btrfs_set_extent_data_ref_root(leaf, ref,
1161                                                        root_objectid);
1162                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1163                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1164                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1165                 } else {
1166                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1167                         num_refs += refs_to_add;
1168                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1169                 }
1170         }
1171         btrfs_mark_buffer_dirty(leaf);
1172         ret = 0;
1173 fail:
1174         btrfs_release_path(path);
1175         return ret;
1176 }
1177
1178 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1179                                            struct btrfs_root *root,
1180                                            struct btrfs_path *path,
1181                                            int refs_to_drop)
1182 {
1183         struct btrfs_key key;
1184         struct btrfs_extent_data_ref *ref1 = NULL;
1185         struct btrfs_shared_data_ref *ref2 = NULL;
1186         struct extent_buffer *leaf;
1187         u32 num_refs = 0;
1188         int ret = 0;
1189
1190         leaf = path->nodes[0];
1191         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1192
1193         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1194                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1195                                       struct btrfs_extent_data_ref);
1196                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1197         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1198                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1199                                       struct btrfs_shared_data_ref);
1200                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1201 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1202         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1203                 struct btrfs_extent_ref_v0 *ref0;
1204                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1205                                       struct btrfs_extent_ref_v0);
1206                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1207 #endif
1208         } else {
1209                 BUG();
1210         }
1211
1212         BUG_ON(num_refs < refs_to_drop);
1213         num_refs -= refs_to_drop;
1214
1215         if (num_refs == 0) {
1216                 ret = btrfs_del_item(trans, root, path);
1217         } else {
1218                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1219                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1220                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1221                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1222 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1223                 else {
1224                         struct btrfs_extent_ref_v0 *ref0;
1225                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1226                                         struct btrfs_extent_ref_v0);
1227                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1228                 }
1229 #endif
1230                 btrfs_mark_buffer_dirty(leaf);
1231         }
1232         return ret;
1233 }
1234
1235 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1236                                           struct btrfs_path *path,
1237                                           struct btrfs_extent_inline_ref *iref)
1238 {
1239         struct btrfs_key key;
1240         struct extent_buffer *leaf;
1241         struct btrfs_extent_data_ref *ref1;
1242         struct btrfs_shared_data_ref *ref2;
1243         u32 num_refs = 0;
1244
1245         leaf = path->nodes[0];
1246         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1247         if (iref) {
1248                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1249                     BTRFS_EXTENT_DATA_REF_KEY) {
1250                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1251                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1252                 } else {
1253                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1254                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1255                 }
1256         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1257                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1258                                       struct btrfs_extent_data_ref);
1259                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1260         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1261                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1262                                       struct btrfs_shared_data_ref);
1263                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1264 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1265         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1266                 struct btrfs_extent_ref_v0 *ref0;
1267                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1268                                       struct btrfs_extent_ref_v0);
1269                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1270 #endif
1271         } else {
1272                 WARN_ON(1);
1273         }
1274         return num_refs;
1275 }
1276
1277 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1278                                           struct btrfs_root *root,
1279                                           struct btrfs_path *path,
1280                                           u64 bytenr, u64 parent,
1281                                           u64 root_objectid)
1282 {
1283         struct btrfs_key key;
1284         int ret;
1285
1286         key.objectid = bytenr;
1287         if (parent) {
1288                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1289                 key.offset = parent;
1290         } else {
1291                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1292                 key.offset = root_objectid;
1293         }
1294
1295         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1296         if (ret > 0)
1297                 ret = -ENOENT;
1298 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1299         if (ret == -ENOENT && parent) {
1300                 btrfs_release_path(path);
1301                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1302                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1303                 if (ret > 0)
1304                         ret = -ENOENT;
1305         }
1306 #endif
1307         return ret;
1308 }
1309
1310 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1311                                           struct btrfs_root *root,
1312                                           struct btrfs_path *path,
1313                                           u64 bytenr, u64 parent,
1314                                           u64 root_objectid)
1315 {
1316         struct btrfs_key key;
1317         int ret;
1318
1319         key.objectid = bytenr;
1320         if (parent) {
1321                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1322                 key.offset = parent;
1323         } else {
1324                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1325                 key.offset = root_objectid;
1326         }
1327
1328         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1329         btrfs_release_path(path);
1330         return ret;
1331 }
1332
1333 static inline int extent_ref_type(u64 parent, u64 owner)
1334 {
1335         int type;
1336         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1337                 if (parent > 0)
1338                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1339                 else
1340                         type = BTRFS_TREE_BLOCK_REF_KEY;
1341         } else {
1342                 if (parent > 0)
1343                         type = BTRFS_SHARED_DATA_REF_KEY;
1344                 else
1345                         type = BTRFS_EXTENT_DATA_REF_KEY;
1346         }
1347         return type;
1348 }
1349
1350 static int find_next_key(struct btrfs_path *path, int level,
1351                          struct btrfs_key *key)
1352
1353 {
1354         for (; level < BTRFS_MAX_LEVEL; level++) {
1355                 if (!path->nodes[level])
1356                         break;
1357                 if (path->slots[level] + 1 >=
1358                     btrfs_header_nritems(path->nodes[level]))
1359                         continue;
1360                 if (level == 0)
1361                         btrfs_item_key_to_cpu(path->nodes[level], key,
1362                                               path->slots[level] + 1);
1363                 else
1364                         btrfs_node_key_to_cpu(path->nodes[level], key,
1365                                               path->slots[level] + 1);
1366                 return 0;
1367         }
1368         return 1;
1369 }
1370
1371 /*
1372  * look for inline back ref. if back ref is found, *ref_ret is set
1373  * to the address of inline back ref, and 0 is returned.
1374  *
1375  * if back ref isn't found, *ref_ret is set to the address where it
1376  * should be inserted, and -ENOENT is returned.
1377  *
1378  * if insert is true and there are too many inline back refs, the path
1379  * points to the extent item, and -EAGAIN is returned.
1380  *
1381  * NOTE: inline back refs are ordered in the same way that back ref
1382  *       items in the tree are ordered.
1383  */
1384 static noinline_for_stack
1385 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1386                                  struct btrfs_root *root,
1387                                  struct btrfs_path *path,
1388                                  struct btrfs_extent_inline_ref **ref_ret,
1389                                  u64 bytenr, u64 num_bytes,
1390                                  u64 parent, u64 root_objectid,
1391                                  u64 owner, u64 offset, int insert)
1392 {
1393         struct btrfs_key key;
1394         struct extent_buffer *leaf;
1395         struct btrfs_extent_item *ei;
1396         struct btrfs_extent_inline_ref *iref;
1397         u64 flags;
1398         u64 item_size;
1399         unsigned long ptr;
1400         unsigned long end;
1401         int extra_size;
1402         int type;
1403         int want;
1404         int ret;
1405         int err = 0;
1406
1407         key.objectid = bytenr;
1408         key.type = BTRFS_EXTENT_ITEM_KEY;
1409         key.offset = num_bytes;
1410
1411         want = extent_ref_type(parent, owner);
1412         if (insert) {
1413                 extra_size = btrfs_extent_inline_ref_size(want);
1414                 path->keep_locks = 1;
1415         } else
1416                 extra_size = -1;
1417         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1418         if (ret < 0) {
1419                 err = ret;
1420                 goto out;
1421         }
1422         BUG_ON(ret);
1423
1424         leaf = path->nodes[0];
1425         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1426 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1427         if (item_size < sizeof(*ei)) {
1428                 if (!insert) {
1429                         err = -ENOENT;
1430                         goto out;
1431                 }
1432                 ret = convert_extent_item_v0(trans, root, path, owner,
1433                                              extra_size);
1434                 if (ret < 0) {
1435                         err = ret;
1436                         goto out;
1437                 }
1438                 leaf = path->nodes[0];
1439                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1440         }
1441 #endif
1442         BUG_ON(item_size < sizeof(*ei));
1443
1444         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1445         flags = btrfs_extent_flags(leaf, ei);
1446
1447         ptr = (unsigned long)(ei + 1);
1448         end = (unsigned long)ei + item_size;
1449
1450         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1451                 ptr += sizeof(struct btrfs_tree_block_info);
1452                 BUG_ON(ptr > end);
1453         } else {
1454                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1455         }
1456
1457         err = -ENOENT;
1458         while (1) {
1459                 if (ptr >= end) {
1460                         WARN_ON(ptr > end);
1461                         break;
1462                 }
1463                 iref = (struct btrfs_extent_inline_ref *)ptr;
1464                 type = btrfs_extent_inline_ref_type(leaf, iref);
1465                 if (want < type)
1466                         break;
1467                 if (want > type) {
1468                         ptr += btrfs_extent_inline_ref_size(type);
1469                         continue;
1470                 }
1471
1472                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1473                         struct btrfs_extent_data_ref *dref;
1474                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1475                         if (match_extent_data_ref(leaf, dref, root_objectid,
1476                                                   owner, offset)) {
1477                                 err = 0;
1478                                 break;
1479                         }
1480                         if (hash_extent_data_ref_item(leaf, dref) <
1481                             hash_extent_data_ref(root_objectid, owner, offset))
1482                                 break;
1483                 } else {
1484                         u64 ref_offset;
1485                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1486                         if (parent > 0) {
1487                                 if (parent == ref_offset) {
1488                                         err = 0;
1489                                         break;
1490                                 }
1491                                 if (ref_offset < parent)
1492                                         break;
1493                         } else {
1494                                 if (root_objectid == ref_offset) {
1495                                         err = 0;
1496                                         break;
1497                                 }
1498                                 if (ref_offset < root_objectid)
1499                                         break;
1500                         }
1501                 }
1502                 ptr += btrfs_extent_inline_ref_size(type);
1503         }
1504         if (err == -ENOENT && insert) {
1505                 if (item_size + extra_size >=
1506                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1507                         err = -EAGAIN;
1508                         goto out;
1509                 }
1510                 /*
1511                  * To add new inline back ref, we have to make sure
1512                  * there is no corresponding back ref item.
1513                  * For simplicity, we just do not add new inline back
1514                  * ref if there is any kind of item for this block
1515                  */
1516                 if (find_next_key(path, 0, &key) == 0 &&
1517                     key.objectid == bytenr &&
1518                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1519                         err = -EAGAIN;
1520                         goto out;
1521                 }
1522         }
1523         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1524 out:
1525         if (insert) {
1526                 path->keep_locks = 0;
1527                 btrfs_unlock_up_safe(path, 1);
1528         }
1529         return err;
1530 }
1531
1532 /*
1533  * helper to add new inline back ref
1534  */
1535 static noinline_for_stack
1536 int setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1537                                 struct btrfs_root *root,
1538                                 struct btrfs_path *path,
1539                                 struct btrfs_extent_inline_ref *iref,
1540                                 u64 parent, u64 root_objectid,
1541                                 u64 owner, u64 offset, int refs_to_add,
1542                                 struct btrfs_delayed_extent_op *extent_op)
1543 {
1544         struct extent_buffer *leaf;
1545         struct btrfs_extent_item *ei;
1546         unsigned long ptr;
1547         unsigned long end;
1548         unsigned long item_offset;
1549         u64 refs;
1550         int size;
1551         int type;
1552         int ret;
1553
1554         leaf = path->nodes[0];
1555         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1556         item_offset = (unsigned long)iref - (unsigned long)ei;
1557
1558         type = extent_ref_type(parent, owner);
1559         size = btrfs_extent_inline_ref_size(type);
1560
1561         ret = btrfs_extend_item(trans, root, path, size);
1562
1563         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1564         refs = btrfs_extent_refs(leaf, ei);
1565         refs += refs_to_add;
1566         btrfs_set_extent_refs(leaf, ei, refs);
1567         if (extent_op)
1568                 __run_delayed_extent_op(extent_op, leaf, ei);
1569
1570         ptr = (unsigned long)ei + item_offset;
1571         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1572         if (ptr < end - size)
1573                 memmove_extent_buffer(leaf, ptr + size, ptr,
1574                                       end - size - ptr);
1575
1576         iref = (struct btrfs_extent_inline_ref *)ptr;
1577         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1578         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1579                 struct btrfs_extent_data_ref *dref;
1580                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1581                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1582                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1583                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1584                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1585         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1586                 struct btrfs_shared_data_ref *sref;
1587                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1588                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1589                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1590         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1591                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1592         } else {
1593                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1594         }
1595         btrfs_mark_buffer_dirty(leaf);
1596         return 0;
1597 }
1598
1599 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1600                                  struct btrfs_root *root,
1601                                  struct btrfs_path *path,
1602                                  struct btrfs_extent_inline_ref **ref_ret,
1603                                  u64 bytenr, u64 num_bytes, u64 parent,
1604                                  u64 root_objectid, u64 owner, u64 offset)
1605 {
1606         int ret;
1607
1608         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1609                                            bytenr, num_bytes, parent,
1610                                            root_objectid, owner, offset, 0);
1611         if (ret != -ENOENT)
1612                 return ret;
1613
1614         btrfs_release_path(path);
1615         *ref_ret = NULL;
1616
1617         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1618                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1619                                             root_objectid);
1620         } else {
1621                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1622                                              root_objectid, owner, offset);
1623         }
1624         return ret;
1625 }
1626
1627 /*
1628  * helper to update/remove inline back ref
1629  */
1630 static noinline_for_stack
1631 int update_inline_extent_backref(struct btrfs_trans_handle *trans,
1632                                  struct btrfs_root *root,
1633                                  struct btrfs_path *path,
1634                                  struct btrfs_extent_inline_ref *iref,
1635                                  int refs_to_mod,
1636                                  struct btrfs_delayed_extent_op *extent_op)
1637 {
1638         struct extent_buffer *leaf;
1639         struct btrfs_extent_item *ei;
1640         struct btrfs_extent_data_ref *dref = NULL;
1641         struct btrfs_shared_data_ref *sref = NULL;
1642         unsigned long ptr;
1643         unsigned long end;
1644         u32 item_size;
1645         int size;
1646         int type;
1647         int ret;
1648         u64 refs;
1649
1650         leaf = path->nodes[0];
1651         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1652         refs = btrfs_extent_refs(leaf, ei);
1653         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1654         refs += refs_to_mod;
1655         btrfs_set_extent_refs(leaf, ei, refs);
1656         if (extent_op)
1657                 __run_delayed_extent_op(extent_op, leaf, ei);
1658
1659         type = btrfs_extent_inline_ref_type(leaf, iref);
1660
1661         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1662                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1663                 refs = btrfs_extent_data_ref_count(leaf, dref);
1664         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1665                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1666                 refs = btrfs_shared_data_ref_count(leaf, sref);
1667         } else {
1668                 refs = 1;
1669                 BUG_ON(refs_to_mod != -1);
1670         }
1671
1672         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1673         refs += refs_to_mod;
1674
1675         if (refs > 0) {
1676                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1677                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1678                 else
1679                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1680         } else {
1681                 size =  btrfs_extent_inline_ref_size(type);
1682                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1683                 ptr = (unsigned long)iref;
1684                 end = (unsigned long)ei + item_size;
1685                 if (ptr + size < end)
1686                         memmove_extent_buffer(leaf, ptr, ptr + size,
1687                                               end - ptr - size);
1688                 item_size -= size;
1689                 ret = btrfs_truncate_item(trans, root, path, item_size, 1);
1690         }
1691         btrfs_mark_buffer_dirty(leaf);
1692         return 0;
1693 }
1694
1695 static noinline_for_stack
1696 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1697                                  struct btrfs_root *root,
1698                                  struct btrfs_path *path,
1699                                  u64 bytenr, u64 num_bytes, u64 parent,
1700                                  u64 root_objectid, u64 owner,
1701                                  u64 offset, int refs_to_add,
1702                                  struct btrfs_delayed_extent_op *extent_op)
1703 {
1704         struct btrfs_extent_inline_ref *iref;
1705         int ret;
1706
1707         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1708                                            bytenr, num_bytes, parent,
1709                                            root_objectid, owner, offset, 1);
1710         if (ret == 0) {
1711                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1712                 ret = update_inline_extent_backref(trans, root, path, iref,
1713                                                    refs_to_add, extent_op);
1714         } else if (ret == -ENOENT) {
1715                 ret = setup_inline_extent_backref(trans, root, path, iref,
1716                                                   parent, root_objectid,
1717                                                   owner, offset, refs_to_add,
1718                                                   extent_op);
1719         }
1720         return ret;
1721 }
1722
1723 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1724                                  struct btrfs_root *root,
1725                                  struct btrfs_path *path,
1726                                  u64 bytenr, u64 parent, u64 root_objectid,
1727                                  u64 owner, u64 offset, int refs_to_add)
1728 {
1729         int ret;
1730         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1731                 BUG_ON(refs_to_add != 1);
1732                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1733                                             parent, root_objectid);
1734         } else {
1735                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1736                                              parent, root_objectid,
1737                                              owner, offset, refs_to_add);
1738         }
1739         return ret;
1740 }
1741
1742 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1743                                  struct btrfs_root *root,
1744                                  struct btrfs_path *path,
1745                                  struct btrfs_extent_inline_ref *iref,
1746                                  int refs_to_drop, int is_data)
1747 {
1748         int ret;
1749
1750         BUG_ON(!is_data && refs_to_drop != 1);
1751         if (iref) {
1752                 ret = update_inline_extent_backref(trans, root, path, iref,
1753                                                    -refs_to_drop, NULL);
1754         } else if (is_data) {
1755                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1756         } else {
1757                 ret = btrfs_del_item(trans, root, path);
1758         }
1759         return ret;
1760 }
1761
1762 static int btrfs_issue_discard(struct block_device *bdev,
1763                                 u64 start, u64 len)
1764 {
1765         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1766 }
1767
1768 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1769                                 u64 num_bytes, u64 *actual_bytes)
1770 {
1771         int ret;
1772         u64 discarded_bytes = 0;
1773         struct btrfs_multi_bio *multi = NULL;
1774
1775
1776         /* Tell the block device(s) that the sectors can be discarded */
1777         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1778                               bytenr, &num_bytes, &multi, 0);
1779         if (!ret) {
1780                 struct btrfs_bio_stripe *stripe = multi->stripes;
1781                 int i;
1782
1783
1784                 for (i = 0; i < multi->num_stripes; i++, stripe++) {
1785                         ret = btrfs_issue_discard(stripe->dev->bdev,
1786                                                   stripe->physical,
1787                                                   stripe->length);
1788                         if (!ret)
1789                                 discarded_bytes += stripe->length;
1790                         else if (ret != -EOPNOTSUPP)
1791                                 break;
1792                 }
1793                 kfree(multi);
1794         }
1795         if (discarded_bytes && ret == -EOPNOTSUPP)
1796                 ret = 0;
1797
1798         if (actual_bytes)
1799                 *actual_bytes = discarded_bytes;
1800
1801
1802         return ret;
1803 }
1804
1805 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1806                          struct btrfs_root *root,
1807                          u64 bytenr, u64 num_bytes, u64 parent,
1808                          u64 root_objectid, u64 owner, u64 offset)
1809 {
1810         int ret;
1811         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1812                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1813
1814         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1815                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
1816                                         parent, root_objectid, (int)owner,
1817                                         BTRFS_ADD_DELAYED_REF, NULL);
1818         } else {
1819                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
1820                                         parent, root_objectid, owner, offset,
1821                                         BTRFS_ADD_DELAYED_REF, NULL);
1822         }
1823         return ret;
1824 }
1825
1826 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1827                                   struct btrfs_root *root,
1828                                   u64 bytenr, u64 num_bytes,
1829                                   u64 parent, u64 root_objectid,
1830                                   u64 owner, u64 offset, int refs_to_add,
1831                                   struct btrfs_delayed_extent_op *extent_op)
1832 {
1833         struct btrfs_path *path;
1834         struct extent_buffer *leaf;
1835         struct btrfs_extent_item *item;
1836         u64 refs;
1837         int ret;
1838         int err = 0;
1839
1840         path = btrfs_alloc_path();
1841         if (!path)
1842                 return -ENOMEM;
1843
1844         path->reada = 1;
1845         path->leave_spinning = 1;
1846         /* this will setup the path even if it fails to insert the back ref */
1847         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1848                                            path, bytenr, num_bytes, parent,
1849                                            root_objectid, owner, offset,
1850                                            refs_to_add, extent_op);
1851         if (ret == 0)
1852                 goto out;
1853
1854         if (ret != -EAGAIN) {
1855                 err = ret;
1856                 goto out;
1857         }
1858
1859         leaf = path->nodes[0];
1860         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1861         refs = btrfs_extent_refs(leaf, item);
1862         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1863         if (extent_op)
1864                 __run_delayed_extent_op(extent_op, leaf, item);
1865
1866         btrfs_mark_buffer_dirty(leaf);
1867         btrfs_release_path(path);
1868
1869         path->reada = 1;
1870         path->leave_spinning = 1;
1871
1872         /* now insert the actual backref */
1873         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1874                                     path, bytenr, parent, root_objectid,
1875                                     owner, offset, refs_to_add);
1876         BUG_ON(ret);
1877 out:
1878         btrfs_free_path(path);
1879         return err;
1880 }
1881
1882 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1883                                 struct btrfs_root *root,
1884                                 struct btrfs_delayed_ref_node *node,
1885                                 struct btrfs_delayed_extent_op *extent_op,
1886                                 int insert_reserved)
1887 {
1888         int ret = 0;
1889         struct btrfs_delayed_data_ref *ref;
1890         struct btrfs_key ins;
1891         u64 parent = 0;
1892         u64 ref_root = 0;
1893         u64 flags = 0;
1894
1895         ins.objectid = node->bytenr;
1896         ins.offset = node->num_bytes;
1897         ins.type = BTRFS_EXTENT_ITEM_KEY;
1898
1899         ref = btrfs_delayed_node_to_data_ref(node);
1900         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1901                 parent = ref->parent;
1902         else
1903                 ref_root = ref->root;
1904
1905         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1906                 if (extent_op) {
1907                         BUG_ON(extent_op->update_key);
1908                         flags |= extent_op->flags_to_set;
1909                 }
1910                 ret = alloc_reserved_file_extent(trans, root,
1911                                                  parent, ref_root, flags,
1912                                                  ref->objectid, ref->offset,
1913                                                  &ins, node->ref_mod);
1914         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1915                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1916                                              node->num_bytes, parent,
1917                                              ref_root, ref->objectid,
1918                                              ref->offset, node->ref_mod,
1919                                              extent_op);
1920         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1921                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1922                                           node->num_bytes, parent,
1923                                           ref_root, ref->objectid,
1924                                           ref->offset, node->ref_mod,
1925                                           extent_op);
1926         } else {
1927                 BUG();
1928         }
1929         return ret;
1930 }
1931
1932 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1933                                     struct extent_buffer *leaf,
1934                                     struct btrfs_extent_item *ei)
1935 {
1936         u64 flags = btrfs_extent_flags(leaf, ei);
1937         if (extent_op->update_flags) {
1938                 flags |= extent_op->flags_to_set;
1939                 btrfs_set_extent_flags(leaf, ei, flags);
1940         }
1941
1942         if (extent_op->update_key) {
1943                 struct btrfs_tree_block_info *bi;
1944                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1945                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1946                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1947         }
1948 }
1949
1950 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1951                                  struct btrfs_root *root,
1952                                  struct btrfs_delayed_ref_node *node,
1953                                  struct btrfs_delayed_extent_op *extent_op)
1954 {
1955         struct btrfs_key key;
1956         struct btrfs_path *path;
1957         struct btrfs_extent_item *ei;
1958         struct extent_buffer *leaf;
1959         u32 item_size;
1960         int ret;
1961         int err = 0;
1962
1963         path = btrfs_alloc_path();
1964         if (!path)
1965                 return -ENOMEM;
1966
1967         key.objectid = node->bytenr;
1968         key.type = BTRFS_EXTENT_ITEM_KEY;
1969         key.offset = node->num_bytes;
1970
1971         path->reada = 1;
1972         path->leave_spinning = 1;
1973         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
1974                                 path, 0, 1);
1975         if (ret < 0) {
1976                 err = ret;
1977                 goto out;
1978         }
1979         if (ret > 0) {
1980                 err = -EIO;
1981                 goto out;
1982         }
1983
1984         leaf = path->nodes[0];
1985         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1986 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1987         if (item_size < sizeof(*ei)) {
1988                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
1989                                              path, (u64)-1, 0);
1990                 if (ret < 0) {
1991                         err = ret;
1992                         goto out;
1993                 }
1994                 leaf = path->nodes[0];
1995                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1996         }
1997 #endif
1998         BUG_ON(item_size < sizeof(*ei));
1999         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2000         __run_delayed_extent_op(extent_op, leaf, ei);
2001
2002         btrfs_mark_buffer_dirty(leaf);
2003 out:
2004         btrfs_free_path(path);
2005         return err;
2006 }
2007
2008 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2009                                 struct btrfs_root *root,
2010                                 struct btrfs_delayed_ref_node *node,
2011                                 struct btrfs_delayed_extent_op *extent_op,
2012                                 int insert_reserved)
2013 {
2014         int ret = 0;
2015         struct btrfs_delayed_tree_ref *ref;
2016         struct btrfs_key ins;
2017         u64 parent = 0;
2018         u64 ref_root = 0;
2019
2020         ins.objectid = node->bytenr;
2021         ins.offset = node->num_bytes;
2022         ins.type = BTRFS_EXTENT_ITEM_KEY;
2023
2024         ref = btrfs_delayed_node_to_tree_ref(node);
2025         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2026                 parent = ref->parent;
2027         else
2028                 ref_root = ref->root;
2029
2030         BUG_ON(node->ref_mod != 1);
2031         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2032                 BUG_ON(!extent_op || !extent_op->update_flags ||
2033                        !extent_op->update_key);
2034                 ret = alloc_reserved_tree_block(trans, root,
2035                                                 parent, ref_root,
2036                                                 extent_op->flags_to_set,
2037                                                 &extent_op->key,
2038                                                 ref->level, &ins);
2039         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2040                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2041                                              node->num_bytes, parent, ref_root,
2042                                              ref->level, 0, 1, extent_op);
2043         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2044                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2045                                           node->num_bytes, parent, ref_root,
2046                                           ref->level, 0, 1, extent_op);
2047         } else {
2048                 BUG();
2049         }
2050         return ret;
2051 }
2052
2053 /* helper function to actually process a single delayed ref entry */
2054 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2055                                struct btrfs_root *root,
2056                                struct btrfs_delayed_ref_node *node,
2057                                struct btrfs_delayed_extent_op *extent_op,
2058                                int insert_reserved)
2059 {
2060         int ret;
2061         if (btrfs_delayed_ref_is_head(node)) {
2062                 struct btrfs_delayed_ref_head *head;
2063                 /*
2064                  * we've hit the end of the chain and we were supposed
2065                  * to insert this extent into the tree.  But, it got
2066                  * deleted before we ever needed to insert it, so all
2067                  * we have to do is clean up the accounting
2068                  */
2069                 BUG_ON(extent_op);
2070                 head = btrfs_delayed_node_to_head(node);
2071                 if (insert_reserved) {
2072                         btrfs_pin_extent(root, node->bytenr,
2073                                          node->num_bytes, 1);
2074                         if (head->is_data) {
2075                                 ret = btrfs_del_csums(trans, root,
2076                                                       node->bytenr,
2077                                                       node->num_bytes);
2078                                 BUG_ON(ret);
2079                         }
2080                 }
2081                 mutex_unlock(&head->mutex);
2082                 return 0;
2083         }
2084
2085         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2086             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2087                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2088                                            insert_reserved);
2089         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2090                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2091                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2092                                            insert_reserved);
2093         else
2094                 BUG();
2095         return ret;
2096 }
2097
2098 static noinline struct btrfs_delayed_ref_node *
2099 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2100 {
2101         struct rb_node *node;
2102         struct btrfs_delayed_ref_node *ref;
2103         int action = BTRFS_ADD_DELAYED_REF;
2104 again:
2105         /*
2106          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2107          * this prevents ref count from going down to zero when
2108          * there still are pending delayed ref.
2109          */
2110         node = rb_prev(&head->node.rb_node);
2111         while (1) {
2112                 if (!node)
2113                         break;
2114                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2115                                 rb_node);
2116                 if (ref->bytenr != head->node.bytenr)
2117                         break;
2118                 if (ref->action == action)
2119                         return ref;
2120                 node = rb_prev(node);
2121         }
2122         if (action == BTRFS_ADD_DELAYED_REF) {
2123                 action = BTRFS_DROP_DELAYED_REF;
2124                 goto again;
2125         }
2126         return NULL;
2127 }
2128
2129 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2130                                        struct btrfs_root *root,
2131                                        struct list_head *cluster)
2132 {
2133         struct btrfs_delayed_ref_root *delayed_refs;
2134         struct btrfs_delayed_ref_node *ref;
2135         struct btrfs_delayed_ref_head *locked_ref = NULL;
2136         struct btrfs_delayed_extent_op *extent_op;
2137         int ret;
2138         int count = 0;
2139         int must_insert_reserved = 0;
2140
2141         delayed_refs = &trans->transaction->delayed_refs;
2142         while (1) {
2143                 if (!locked_ref) {
2144                         /* pick a new head ref from the cluster list */
2145                         if (list_empty(cluster))
2146                                 break;
2147
2148                         locked_ref = list_entry(cluster->next,
2149                                      struct btrfs_delayed_ref_head, cluster);
2150
2151                         /* grab the lock that says we are going to process
2152                          * all the refs for this head */
2153                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2154
2155                         /*
2156                          * we may have dropped the spin lock to get the head
2157                          * mutex lock, and that might have given someone else
2158                          * time to free the head.  If that's true, it has been
2159                          * removed from our list and we can move on.
2160                          */
2161                         if (ret == -EAGAIN) {
2162                                 locked_ref = NULL;
2163                                 count++;
2164                                 continue;
2165                         }
2166                 }
2167
2168                 /*
2169                  * record the must insert reserved flag before we
2170                  * drop the spin lock.
2171                  */
2172                 must_insert_reserved = locked_ref->must_insert_reserved;
2173                 locked_ref->must_insert_reserved = 0;
2174
2175                 extent_op = locked_ref->extent_op;
2176                 locked_ref->extent_op = NULL;
2177
2178                 /*
2179                  * locked_ref is the head node, so we have to go one
2180                  * node back for any delayed ref updates
2181                  */
2182                 ref = select_delayed_ref(locked_ref);
2183                 if (!ref) {
2184                         /* All delayed refs have been processed, Go ahead
2185                          * and send the head node to run_one_delayed_ref,
2186                          * so that any accounting fixes can happen
2187                          */
2188                         ref = &locked_ref->node;
2189
2190                         if (extent_op && must_insert_reserved) {
2191                                 kfree(extent_op);
2192                                 extent_op = NULL;
2193                         }
2194
2195                         if (extent_op) {
2196                                 spin_unlock(&delayed_refs->lock);
2197
2198                                 ret = run_delayed_extent_op(trans, root,
2199                                                             ref, extent_op);
2200                                 BUG_ON(ret);
2201                                 kfree(extent_op);
2202
2203                                 cond_resched();
2204                                 spin_lock(&delayed_refs->lock);
2205                                 continue;
2206                         }
2207
2208                         list_del_init(&locked_ref->cluster);
2209                         locked_ref = NULL;
2210                 }
2211
2212                 ref->in_tree = 0;
2213                 rb_erase(&ref->rb_node, &delayed_refs->root);
2214                 delayed_refs->num_entries--;
2215
2216                 spin_unlock(&delayed_refs->lock);
2217
2218                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2219                                           must_insert_reserved);
2220                 BUG_ON(ret);
2221
2222                 btrfs_put_delayed_ref(ref);
2223                 kfree(extent_op);
2224                 count++;
2225
2226                 cond_resched();
2227                 spin_lock(&delayed_refs->lock);
2228         }
2229         return count;
2230 }
2231
2232 /*
2233  * this starts processing the delayed reference count updates and
2234  * extent insertions we have queued up so far.  count can be
2235  * 0, which means to process everything in the tree at the start
2236  * of the run (but not newly added entries), or it can be some target
2237  * number you'd like to process.
2238  */
2239 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2240                            struct btrfs_root *root, unsigned long count)
2241 {
2242         struct rb_node *node;
2243         struct btrfs_delayed_ref_root *delayed_refs;
2244         struct btrfs_delayed_ref_node *ref;
2245         struct list_head cluster;
2246         int ret;
2247         int run_all = count == (unsigned long)-1;
2248         int run_most = 0;
2249
2250         if (root == root->fs_info->extent_root)
2251                 root = root->fs_info->tree_root;
2252
2253         delayed_refs = &trans->transaction->delayed_refs;
2254         INIT_LIST_HEAD(&cluster);
2255 again:
2256         spin_lock(&delayed_refs->lock);
2257         if (count == 0) {
2258                 count = delayed_refs->num_entries * 2;
2259                 run_most = 1;
2260         }
2261         while (1) {
2262                 if (!(run_all || run_most) &&
2263                     delayed_refs->num_heads_ready < 64)
2264                         break;
2265
2266                 /*
2267                  * go find something we can process in the rbtree.  We start at
2268                  * the beginning of the tree, and then build a cluster
2269                  * of refs to process starting at the first one we are able to
2270                  * lock
2271                  */
2272                 ret = btrfs_find_ref_cluster(trans, &cluster,
2273                                              delayed_refs->run_delayed_start);
2274                 if (ret)
2275                         break;
2276
2277                 ret = run_clustered_refs(trans, root, &cluster);
2278                 BUG_ON(ret < 0);
2279
2280                 count -= min_t(unsigned long, ret, count);
2281
2282                 if (count == 0)
2283                         break;
2284         }
2285
2286         if (run_all) {
2287                 node = rb_first(&delayed_refs->root);
2288                 if (!node)
2289                         goto out;
2290                 count = (unsigned long)-1;
2291
2292                 while (node) {
2293                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2294                                        rb_node);
2295                         if (btrfs_delayed_ref_is_head(ref)) {
2296                                 struct btrfs_delayed_ref_head *head;
2297
2298                                 head = btrfs_delayed_node_to_head(ref);
2299                                 atomic_inc(&ref->refs);
2300
2301                                 spin_unlock(&delayed_refs->lock);
2302                                 /*
2303                                  * Mutex was contended, block until it's
2304                                  * released and try again
2305                                  */
2306                                 mutex_lock(&head->mutex);
2307                                 mutex_unlock(&head->mutex);
2308
2309                                 btrfs_put_delayed_ref(ref);
2310                                 cond_resched();
2311                                 goto again;
2312                         }
2313                         node = rb_next(node);
2314                 }
2315                 spin_unlock(&delayed_refs->lock);
2316                 schedule_timeout(1);
2317                 goto again;
2318         }
2319 out:
2320         spin_unlock(&delayed_refs->lock);
2321         return 0;
2322 }
2323
2324 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2325                                 struct btrfs_root *root,
2326                                 u64 bytenr, u64 num_bytes, u64 flags,
2327                                 int is_data)
2328 {
2329         struct btrfs_delayed_extent_op *extent_op;
2330         int ret;
2331
2332         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2333         if (!extent_op)
2334                 return -ENOMEM;
2335
2336         extent_op->flags_to_set = flags;
2337         extent_op->update_flags = 1;
2338         extent_op->update_key = 0;
2339         extent_op->is_data = is_data ? 1 : 0;
2340
2341         ret = btrfs_add_delayed_extent_op(trans, bytenr, num_bytes, extent_op);
2342         if (ret)
2343                 kfree(extent_op);
2344         return ret;
2345 }
2346
2347 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2348                                       struct btrfs_root *root,
2349                                       struct btrfs_path *path,
2350                                       u64 objectid, u64 offset, u64 bytenr)
2351 {
2352         struct btrfs_delayed_ref_head *head;
2353         struct btrfs_delayed_ref_node *ref;
2354         struct btrfs_delayed_data_ref *data_ref;
2355         struct btrfs_delayed_ref_root *delayed_refs;
2356         struct rb_node *node;
2357         int ret = 0;
2358
2359         ret = -ENOENT;
2360         delayed_refs = &trans->transaction->delayed_refs;
2361         spin_lock(&delayed_refs->lock);
2362         head = btrfs_find_delayed_ref_head(trans, bytenr);
2363         if (!head)
2364                 goto out;
2365
2366         if (!mutex_trylock(&head->mutex)) {
2367                 atomic_inc(&head->node.refs);
2368                 spin_unlock(&delayed_refs->lock);
2369
2370                 btrfs_release_path(path);
2371
2372                 /*
2373                  * Mutex was contended, block until it's released and let
2374                  * caller try again
2375                  */
2376                 mutex_lock(&head->mutex);
2377                 mutex_unlock(&head->mutex);
2378                 btrfs_put_delayed_ref(&head->node);
2379                 return -EAGAIN;
2380         }
2381
2382         node = rb_prev(&head->node.rb_node);
2383         if (!node)
2384                 goto out_unlock;
2385
2386         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2387
2388         if (ref->bytenr != bytenr)
2389                 goto out_unlock;
2390
2391         ret = 1;
2392         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2393                 goto out_unlock;
2394
2395         data_ref = btrfs_delayed_node_to_data_ref(ref);
2396
2397         node = rb_prev(node);
2398         if (node) {
2399                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2400                 if (ref->bytenr == bytenr)
2401                         goto out_unlock;
2402         }
2403
2404         if (data_ref->root != root->root_key.objectid ||
2405             data_ref->objectid != objectid || data_ref->offset != offset)
2406                 goto out_unlock;
2407
2408         ret = 0;
2409 out_unlock:
2410         mutex_unlock(&head->mutex);
2411 out:
2412         spin_unlock(&delayed_refs->lock);
2413         return ret;
2414 }
2415
2416 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2417                                         struct btrfs_root *root,
2418                                         struct btrfs_path *path,
2419                                         u64 objectid, u64 offset, u64 bytenr)
2420 {
2421         struct btrfs_root *extent_root = root->fs_info->extent_root;
2422         struct extent_buffer *leaf;
2423         struct btrfs_extent_data_ref *ref;
2424         struct btrfs_extent_inline_ref *iref;
2425         struct btrfs_extent_item *ei;
2426         struct btrfs_key key;
2427         u32 item_size;
2428         int ret;
2429
2430         key.objectid = bytenr;
2431         key.offset = (u64)-1;
2432         key.type = BTRFS_EXTENT_ITEM_KEY;
2433
2434         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2435         if (ret < 0)
2436                 goto out;
2437         BUG_ON(ret == 0);
2438
2439         ret = -ENOENT;
2440         if (path->slots[0] == 0)
2441                 goto out;
2442
2443         path->slots[0]--;
2444         leaf = path->nodes[0];
2445         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2446
2447         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2448                 goto out;
2449
2450         ret = 1;
2451         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2452 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2453         if (item_size < sizeof(*ei)) {
2454                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2455                 goto out;
2456         }
2457 #endif
2458         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2459
2460         if (item_size != sizeof(*ei) +
2461             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2462                 goto out;
2463
2464         if (btrfs_extent_generation(leaf, ei) <=
2465             btrfs_root_last_snapshot(&root->root_item))
2466                 goto out;
2467
2468         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2469         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2470             BTRFS_EXTENT_DATA_REF_KEY)
2471                 goto out;
2472
2473         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2474         if (btrfs_extent_refs(leaf, ei) !=
2475             btrfs_extent_data_ref_count(leaf, ref) ||
2476             btrfs_extent_data_ref_root(leaf, ref) !=
2477             root->root_key.objectid ||
2478             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2479             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2480                 goto out;
2481
2482         ret = 0;
2483 out:
2484         return ret;
2485 }
2486
2487 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2488                           struct btrfs_root *root,
2489                           u64 objectid, u64 offset, u64 bytenr)
2490 {
2491         struct btrfs_path *path;
2492         int ret;
2493         int ret2;
2494
2495         path = btrfs_alloc_path();
2496         if (!path)
2497                 return -ENOENT;
2498
2499         do {
2500                 ret = check_committed_ref(trans, root, path, objectid,
2501                                           offset, bytenr);
2502                 if (ret && ret != -ENOENT)
2503                         goto out;
2504
2505                 ret2 = check_delayed_ref(trans, root, path, objectid,
2506                                          offset, bytenr);
2507         } while (ret2 == -EAGAIN);
2508
2509         if (ret2 && ret2 != -ENOENT) {
2510                 ret = ret2;
2511                 goto out;
2512         }
2513
2514         if (ret != -ENOENT || ret2 != -ENOENT)
2515                 ret = 0;
2516 out:
2517         btrfs_free_path(path);
2518         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2519                 WARN_ON(ret > 0);
2520         return ret;
2521 }
2522
2523 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2524                            struct btrfs_root *root,
2525                            struct extent_buffer *buf,
2526                            int full_backref, int inc)
2527 {
2528         u64 bytenr;
2529         u64 num_bytes;
2530         u64 parent;
2531         u64 ref_root;
2532         u32 nritems;
2533         struct btrfs_key key;
2534         struct btrfs_file_extent_item *fi;
2535         int i;
2536         int level;
2537         int ret = 0;
2538         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2539                             u64, u64, u64, u64, u64, u64);
2540
2541         ref_root = btrfs_header_owner(buf);
2542         nritems = btrfs_header_nritems(buf);
2543         level = btrfs_header_level(buf);
2544
2545         if (!root->ref_cows && level == 0)
2546                 return 0;
2547
2548         if (inc)
2549                 process_func = btrfs_inc_extent_ref;
2550         else
2551                 process_func = btrfs_free_extent;
2552
2553         if (full_backref)
2554                 parent = buf->start;
2555         else
2556                 parent = 0;
2557
2558         for (i = 0; i < nritems; i++) {
2559                 if (level == 0) {
2560                         btrfs_item_key_to_cpu(buf, &key, i);
2561                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2562                                 continue;
2563                         fi = btrfs_item_ptr(buf, i,
2564                                             struct btrfs_file_extent_item);
2565                         if (btrfs_file_extent_type(buf, fi) ==
2566                             BTRFS_FILE_EXTENT_INLINE)
2567                                 continue;
2568                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2569                         if (bytenr == 0)
2570                                 continue;
2571
2572                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2573                         key.offset -= btrfs_file_extent_offset(buf, fi);
2574                         ret = process_func(trans, root, bytenr, num_bytes,
2575                                            parent, ref_root, key.objectid,
2576                                            key.offset);
2577                         if (ret)
2578                                 goto fail;
2579                 } else {
2580                         bytenr = btrfs_node_blockptr(buf, i);
2581                         num_bytes = btrfs_level_size(root, level - 1);
2582                         ret = process_func(trans, root, bytenr, num_bytes,
2583                                            parent, ref_root, level - 1, 0);
2584                         if (ret)
2585                                 goto fail;
2586                 }
2587         }
2588         return 0;
2589 fail:
2590         BUG();
2591         return ret;
2592 }
2593
2594 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2595                   struct extent_buffer *buf, int full_backref)
2596 {
2597         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2598 }
2599
2600 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2601                   struct extent_buffer *buf, int full_backref)
2602 {
2603         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2604 }
2605
2606 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2607                                  struct btrfs_root *root,
2608                                  struct btrfs_path *path,
2609                                  struct btrfs_block_group_cache *cache)
2610 {
2611         int ret;
2612         struct btrfs_root *extent_root = root->fs_info->extent_root;
2613         unsigned long bi;
2614         struct extent_buffer *leaf;
2615
2616         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2617         if (ret < 0)
2618                 goto fail;
2619         BUG_ON(ret);
2620
2621         leaf = path->nodes[0];
2622         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2623         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2624         btrfs_mark_buffer_dirty(leaf);
2625         btrfs_release_path(path);
2626 fail:
2627         if (ret)
2628                 return ret;
2629         return 0;
2630
2631 }
2632
2633 static struct btrfs_block_group_cache *
2634 next_block_group(struct btrfs_root *root,
2635                  struct btrfs_block_group_cache *cache)
2636 {
2637         struct rb_node *node;
2638         spin_lock(&root->fs_info->block_group_cache_lock);
2639         node = rb_next(&cache->cache_node);
2640         btrfs_put_block_group(cache);
2641         if (node) {
2642                 cache = rb_entry(node, struct btrfs_block_group_cache,
2643                                  cache_node);
2644                 btrfs_get_block_group(cache);
2645         } else
2646                 cache = NULL;
2647         spin_unlock(&root->fs_info->block_group_cache_lock);
2648         return cache;
2649 }
2650
2651 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2652                             struct btrfs_trans_handle *trans,
2653                             struct btrfs_path *path)
2654 {
2655         struct btrfs_root *root = block_group->fs_info->tree_root;
2656         struct inode *inode = NULL;
2657         u64 alloc_hint = 0;
2658         int dcs = BTRFS_DC_ERROR;
2659         int num_pages = 0;
2660         int retries = 0;
2661         int ret = 0;
2662
2663         /*
2664          * If this block group is smaller than 100 megs don't bother caching the
2665          * block group.
2666          */
2667         if (block_group->key.offset < (100 * 1024 * 1024)) {
2668                 spin_lock(&block_group->lock);
2669                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2670                 spin_unlock(&block_group->lock);
2671                 return 0;
2672         }
2673
2674 again:
2675         inode = lookup_free_space_inode(root, block_group, path);
2676         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2677                 ret = PTR_ERR(inode);
2678                 btrfs_release_path(path);
2679                 goto out;
2680         }
2681
2682         if (IS_ERR(inode)) {
2683                 BUG_ON(retries);
2684                 retries++;
2685
2686                 if (block_group->ro)
2687                         goto out_free;
2688
2689                 ret = create_free_space_inode(root, trans, block_group, path);
2690                 if (ret)
2691                         goto out_free;
2692                 goto again;
2693         }
2694
2695         /*
2696          * We want to set the generation to 0, that way if anything goes wrong
2697          * from here on out we know not to trust this cache when we load up next
2698          * time.
2699          */
2700         BTRFS_I(inode)->generation = 0;
2701         ret = btrfs_update_inode(trans, root, inode);
2702         WARN_ON(ret);
2703
2704         if (i_size_read(inode) > 0) {
2705                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2706                                                       inode);
2707                 if (ret)
2708                         goto out_put;
2709         }
2710
2711         spin_lock(&block_group->lock);
2712         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2713                 /* We're not cached, don't bother trying to write stuff out */
2714                 dcs = BTRFS_DC_WRITTEN;
2715                 spin_unlock(&block_group->lock);
2716                 goto out_put;
2717         }
2718         spin_unlock(&block_group->lock);
2719
2720         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2721         if (!num_pages)
2722                 num_pages = 1;
2723
2724         /*
2725          * Just to make absolutely sure we have enough space, we're going to
2726          * preallocate 12 pages worth of space for each block group.  In
2727          * practice we ought to use at most 8, but we need extra space so we can
2728          * add our header and have a terminator between the extents and the
2729          * bitmaps.
2730          */
2731         num_pages *= 16;
2732         num_pages *= PAGE_CACHE_SIZE;
2733
2734         ret = btrfs_check_data_free_space(inode, num_pages);
2735         if (ret)
2736                 goto out_put;
2737
2738         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2739                                               num_pages, num_pages,
2740                                               &alloc_hint);
2741         if (!ret)
2742                 dcs = BTRFS_DC_SETUP;
2743         btrfs_free_reserved_data_space(inode, num_pages);
2744 out_put:
2745         iput(inode);
2746 out_free:
2747         btrfs_release_path(path);
2748 out:
2749         spin_lock(&block_group->lock);
2750         block_group->disk_cache_state = dcs;
2751         spin_unlock(&block_group->lock);
2752
2753         return ret;
2754 }
2755
2756 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2757                                    struct btrfs_root *root)
2758 {
2759         struct btrfs_block_group_cache *cache;
2760         int err = 0;
2761         struct btrfs_path *path;
2762         u64 last = 0;
2763
2764         path = btrfs_alloc_path();
2765         if (!path)
2766                 return -ENOMEM;
2767
2768 again:
2769         while (1) {
2770                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2771                 while (cache) {
2772                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2773                                 break;
2774                         cache = next_block_group(root, cache);
2775                 }
2776                 if (!cache) {
2777                         if (last == 0)
2778                                 break;
2779                         last = 0;
2780                         continue;
2781                 }
2782                 err = cache_save_setup(cache, trans, path);
2783                 last = cache->key.objectid + cache->key.offset;
2784                 btrfs_put_block_group(cache);
2785         }
2786
2787         while (1) {
2788                 if (last == 0) {
2789                         err = btrfs_run_delayed_refs(trans, root,
2790                                                      (unsigned long)-1);
2791                         BUG_ON(err);
2792                 }
2793
2794                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2795                 while (cache) {
2796                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2797                                 btrfs_put_block_group(cache);
2798                                 goto again;
2799                         }
2800
2801                         if (cache->dirty)
2802                                 break;
2803                         cache = next_block_group(root, cache);
2804                 }
2805                 if (!cache) {
2806                         if (last == 0)
2807                                 break;
2808                         last = 0;
2809                         continue;
2810                 }
2811
2812                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
2813                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
2814                 cache->dirty = 0;
2815                 last = cache->key.objectid + cache->key.offset;
2816
2817                 err = write_one_cache_group(trans, root, path, cache);
2818                 BUG_ON(err);
2819                 btrfs_put_block_group(cache);
2820         }
2821
2822         while (1) {
2823                 /*
2824                  * I don't think this is needed since we're just marking our
2825                  * preallocated extent as written, but just in case it can't
2826                  * hurt.
2827                  */
2828                 if (last == 0) {
2829                         err = btrfs_run_delayed_refs(trans, root,
2830                                                      (unsigned long)-1);
2831                         BUG_ON(err);
2832                 }
2833
2834                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2835                 while (cache) {
2836                         /*
2837                          * Really this shouldn't happen, but it could if we
2838                          * couldn't write the entire preallocated extent and
2839                          * splitting the extent resulted in a new block.
2840                          */
2841                         if (cache->dirty) {
2842                                 btrfs_put_block_group(cache);
2843                                 goto again;
2844                         }
2845                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2846                                 break;
2847                         cache = next_block_group(root, cache);
2848                 }
2849                 if (!cache) {
2850                         if (last == 0)
2851                                 break;
2852                         last = 0;
2853                         continue;
2854                 }
2855
2856                 btrfs_write_out_cache(root, trans, cache, path);
2857
2858                 /*
2859                  * If we didn't have an error then the cache state is still
2860                  * NEED_WRITE, so we can set it to WRITTEN.
2861                  */
2862                 if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2863                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
2864                 last = cache->key.objectid + cache->key.offset;
2865                 btrfs_put_block_group(cache);
2866         }
2867
2868         btrfs_free_path(path);
2869         return 0;
2870 }
2871
2872 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
2873 {
2874         struct btrfs_block_group_cache *block_group;
2875         int readonly = 0;
2876
2877         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
2878         if (!block_group || block_group->ro)
2879                 readonly = 1;
2880         if (block_group)
2881                 btrfs_put_block_group(block_group);
2882         return readonly;
2883 }
2884
2885 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
2886                              u64 total_bytes, u64 bytes_used,
2887                              struct btrfs_space_info **space_info)
2888 {
2889         struct btrfs_space_info *found;
2890         int i;
2891         int factor;
2892
2893         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
2894                      BTRFS_BLOCK_GROUP_RAID10))
2895                 factor = 2;
2896         else
2897                 factor = 1;
2898
2899         found = __find_space_info(info, flags);
2900         if (found) {
2901                 spin_lock(&found->lock);
2902                 found->total_bytes += total_bytes;
2903                 found->disk_total += total_bytes * factor;
2904                 found->bytes_used += bytes_used;
2905                 found->disk_used += bytes_used * factor;
2906                 found->full = 0;
2907                 spin_unlock(&found->lock);
2908                 *space_info = found;
2909                 return 0;
2910         }
2911         found = kzalloc(sizeof(*found), GFP_NOFS);
2912         if (!found)
2913                 return -ENOMEM;
2914
2915         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
2916                 INIT_LIST_HEAD(&found->block_groups[i]);
2917         init_rwsem(&found->groups_sem);
2918         spin_lock_init(&found->lock);
2919         found->flags = flags & (BTRFS_BLOCK_GROUP_DATA |
2920                                 BTRFS_BLOCK_GROUP_SYSTEM |
2921                                 BTRFS_BLOCK_GROUP_METADATA);
2922         found->total_bytes = total_bytes;
2923         found->disk_total = total_bytes * factor;
2924         found->bytes_used = bytes_used;
2925         found->disk_used = bytes_used * factor;
2926         found->bytes_pinned = 0;
2927         found->bytes_reserved = 0;
2928         found->bytes_readonly = 0;
2929         found->bytes_may_use = 0;
2930         found->full = 0;
2931         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
2932         found->chunk_alloc = 0;
2933         *space_info = found;
2934         list_add_rcu(&found->list, &info->space_info);
2935         atomic_set(&found->caching_threads, 0);
2936         return 0;
2937 }
2938
2939 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
2940 {
2941         u64 extra_flags = flags & (BTRFS_BLOCK_GROUP_RAID0 |
2942                                    BTRFS_BLOCK_GROUP_RAID1 |
2943                                    BTRFS_BLOCK_GROUP_RAID10 |
2944                                    BTRFS_BLOCK_GROUP_DUP);
2945         if (extra_flags) {
2946                 if (flags & BTRFS_BLOCK_GROUP_DATA)
2947                         fs_info->avail_data_alloc_bits |= extra_flags;
2948                 if (flags & BTRFS_BLOCK_GROUP_METADATA)
2949                         fs_info->avail_metadata_alloc_bits |= extra_flags;
2950                 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
2951                         fs_info->avail_system_alloc_bits |= extra_flags;
2952         }
2953 }
2954
2955 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
2956 {
2957         /*
2958          * we add in the count of missing devices because we want
2959          * to make sure that any RAID levels on a degraded FS
2960          * continue to be honored.
2961          */
2962         u64 num_devices = root->fs_info->fs_devices->rw_devices +
2963                 root->fs_info->fs_devices->missing_devices;
2964
2965         if (num_devices == 1)
2966                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
2967         if (num_devices < 4)
2968                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
2969
2970         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
2971             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
2972                       BTRFS_BLOCK_GROUP_RAID10))) {
2973                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
2974         }
2975
2976         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
2977             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
2978                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
2979         }
2980
2981         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
2982             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
2983              (flags & BTRFS_BLOCK_GROUP_RAID10) |
2984              (flags & BTRFS_BLOCK_GROUP_DUP)))
2985                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
2986         return flags;
2987 }
2988
2989 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
2990 {
2991         if (flags & BTRFS_BLOCK_GROUP_DATA)
2992                 flags |= root->fs_info->avail_data_alloc_bits &
2993                          root->fs_info->data_alloc_profile;
2994         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
2995                 flags |= root->fs_info->avail_system_alloc_bits &
2996                          root->fs_info->system_alloc_profile;
2997         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
2998                 flags |= root->fs_info->avail_metadata_alloc_bits &
2999                          root->fs_info->metadata_alloc_profile;
3000         return btrfs_reduce_alloc_profile(root, flags);
3001 }
3002
3003 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3004 {
3005         u64 flags;
3006
3007         if (data)
3008                 flags = BTRFS_BLOCK_GROUP_DATA;
3009         else if (root == root->fs_info->chunk_root)
3010                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3011         else
3012                 flags = BTRFS_BLOCK_GROUP_METADATA;
3013
3014         return get_alloc_profile(root, flags);
3015 }
3016
3017 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3018 {
3019         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3020                                                        BTRFS_BLOCK_GROUP_DATA);
3021 }
3022
3023 /*
3024  * This will check the space that the inode allocates from to make sure we have
3025  * enough space for bytes.
3026  */
3027 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3028 {
3029         struct btrfs_space_info *data_sinfo;
3030         struct btrfs_root *root = BTRFS_I(inode)->root;
3031         u64 used;
3032         int ret = 0, committed = 0, alloc_chunk = 1;
3033
3034         /* make sure bytes are sectorsize aligned */
3035         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3036
3037         if (root == root->fs_info->tree_root ||
3038             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3039                 alloc_chunk = 0;
3040                 committed = 1;
3041         }
3042
3043         data_sinfo = BTRFS_I(inode)->space_info;
3044         if (!data_sinfo)
3045                 goto alloc;
3046
3047 again:
3048         /* make sure we have enough space to handle the data first */
3049         spin_lock(&data_sinfo->lock);
3050         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3051                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3052                 data_sinfo->bytes_may_use;
3053
3054         if (used + bytes > data_sinfo->total_bytes) {
3055                 struct btrfs_trans_handle *trans;
3056
3057                 /*
3058                  * if we don't have enough free bytes in this space then we need
3059                  * to alloc a new chunk.
3060                  */
3061                 if (!data_sinfo->full && alloc_chunk) {
3062                         u64 alloc_target;
3063
3064                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3065                         spin_unlock(&data_sinfo->lock);
3066 alloc:
3067                         alloc_target = btrfs_get_alloc_profile(root, 1);
3068                         trans = btrfs_join_transaction(root, 1);
3069                         if (IS_ERR(trans))
3070                                 return PTR_ERR(trans);
3071
3072                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3073                                              bytes + 2 * 1024 * 1024,
3074                                              alloc_target,
3075                                              CHUNK_ALLOC_NO_FORCE);
3076                         btrfs_end_transaction(trans, root);
3077                         if (ret < 0) {
3078                                 if (ret != -ENOSPC)
3079                                         return ret;
3080                                 else
3081                                         goto commit_trans;
3082                         }
3083
3084                         if (!data_sinfo) {
3085                                 btrfs_set_inode_space_info(root, inode);
3086                                 data_sinfo = BTRFS_I(inode)->space_info;
3087                         }
3088                         goto again;
3089                 }
3090                 spin_unlock(&data_sinfo->lock);
3091
3092                 /* commit the current transaction and try again */
3093 commit_trans:
3094                 if (!committed && !root->fs_info->open_ioctl_trans) {
3095                         committed = 1;
3096                         trans = btrfs_join_transaction(root, 1);
3097                         if (IS_ERR(trans))
3098                                 return PTR_ERR(trans);
3099                         ret = btrfs_commit_transaction(trans, root);
3100                         if (ret)
3101                                 return ret;
3102                         goto again;
3103                 }
3104
3105                 return -ENOSPC;
3106         }
3107         data_sinfo->bytes_may_use += bytes;
3108         BTRFS_I(inode)->reserved_bytes += bytes;
3109         spin_unlock(&data_sinfo->lock);
3110
3111         return 0;
3112 }
3113
3114 /*
3115  * called when we are clearing an delalloc extent from the
3116  * inode's io_tree or there was an error for whatever reason
3117  * after calling btrfs_check_data_free_space
3118  */
3119 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3120 {
3121         struct btrfs_root *root = BTRFS_I(inode)->root;
3122         struct btrfs_space_info *data_sinfo;
3123
3124         /* make sure bytes are sectorsize aligned */
3125         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3126
3127         data_sinfo = BTRFS_I(inode)->space_info;
3128         spin_lock(&data_sinfo->lock);
3129         data_sinfo->bytes_may_use -= bytes;
3130         BTRFS_I(inode)->reserved_bytes -= bytes;
3131         spin_unlock(&data_sinfo->lock);
3132 }
3133
3134 static void force_metadata_allocation(struct btrfs_fs_info *info)
3135 {
3136         struct list_head *head = &info->space_info;
3137         struct btrfs_space_info *found;
3138
3139         rcu_read_lock();
3140         list_for_each_entry_rcu(found, head, list) {
3141                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3142                         found->force_alloc = CHUNK_ALLOC_FORCE;
3143         }
3144         rcu_read_unlock();
3145 }
3146
3147 static int should_alloc_chunk(struct btrfs_root *root,
3148                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3149                               int force)
3150 {
3151         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3152         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3153         u64 thresh;
3154
3155         if (force == CHUNK_ALLOC_FORCE)
3156                 return 1;
3157
3158         /*
3159          * in limited mode, we want to have some free space up to
3160          * about 1% of the FS size.
3161          */
3162         if (force == CHUNK_ALLOC_LIMITED) {
3163                 thresh = btrfs_super_total_bytes(&root->fs_info->super_copy);
3164                 thresh = max_t(u64, 64 * 1024 * 1024,
3165                                div_factor_fine(thresh, 1));
3166
3167                 if (num_bytes - num_allocated < thresh)
3168                         return 1;
3169         }
3170
3171         /*
3172          * we have two similar checks here, one based on percentage
3173          * and once based on a hard number of 256MB.  The idea
3174          * is that if we have a good amount of free
3175          * room, don't allocate a chunk.  A good mount is
3176          * less than 80% utilized of the chunks we have allocated,
3177          * or more than 256MB free
3178          */
3179         if (num_allocated + alloc_bytes + 256 * 1024 * 1024 < num_bytes)
3180                 return 0;
3181
3182         if (num_allocated + alloc_bytes < div_factor(num_bytes, 8))
3183                 return 0;
3184
3185         thresh = btrfs_super_total_bytes(&root->fs_info->super_copy);
3186
3187         /* 256MB or 5% of the FS */
3188         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 5));
3189
3190         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 3))
3191                 return 0;
3192         return 1;
3193 }
3194
3195 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3196                           struct btrfs_root *extent_root, u64 alloc_bytes,
3197                           u64 flags, int force)
3198 {
3199         struct btrfs_space_info *space_info;
3200         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3201         int wait_for_alloc = 0;
3202         int ret = 0;
3203
3204         flags = btrfs_reduce_alloc_profile(extent_root, flags);
3205
3206         space_info = __find_space_info(extent_root->fs_info, flags);
3207         if (!space_info) {
3208                 ret = update_space_info(extent_root->fs_info, flags,
3209                                         0, 0, &space_info);
3210                 BUG_ON(ret);
3211         }
3212         BUG_ON(!space_info);
3213
3214 again:
3215         spin_lock(&space_info->lock);
3216         if (space_info->force_alloc)
3217                 force = space_info->force_alloc;
3218         if (space_info->full) {
3219                 spin_unlock(&space_info->lock);
3220                 return 0;
3221         }
3222
3223         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3224                 spin_unlock(&space_info->lock);
3225                 return 0;
3226         } else if (space_info->chunk_alloc) {
3227                 wait_for_alloc = 1;
3228         } else {
3229                 space_info->chunk_alloc = 1;
3230         }
3231
3232         spin_unlock(&space_info->lock);
3233
3234         mutex_lock(&fs_info->chunk_mutex);
3235
3236         /*
3237          * The chunk_mutex is held throughout the entirety of a chunk
3238          * allocation, so once we've acquired the chunk_mutex we know that the
3239          * other guy is done and we need to recheck and see if we should
3240          * allocate.
3241          */
3242         if (wait_for_alloc) {
3243                 mutex_unlock(&fs_info->chunk_mutex);
3244                 wait_for_alloc = 0;
3245                 goto again;
3246         }
3247
3248         /*
3249          * If we have mixed data/metadata chunks we want to make sure we keep
3250          * allocating mixed chunks instead of individual chunks.
3251          */
3252         if (btrfs_mixed_space_info(space_info))
3253                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3254
3255         /*
3256          * if we're doing a data chunk, go ahead and make sure that
3257          * we keep a reasonable number of metadata chunks allocated in the
3258          * FS as well.
3259          */
3260         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3261                 fs_info->data_chunk_allocations++;
3262                 if (!(fs_info->data_chunk_allocations %
3263                       fs_info->metadata_ratio))
3264                         force_metadata_allocation(fs_info);
3265         }
3266
3267         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3268         spin_lock(&space_info->lock);
3269         if (ret)
3270                 space_info->full = 1;
3271         else
3272                 ret = 1;
3273
3274         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3275         space_info->chunk_alloc = 0;
3276         spin_unlock(&space_info->lock);
3277         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3278         return ret;
3279 }
3280
3281 /*
3282  * shrink metadata reservation for delalloc
3283  */
3284 static int shrink_delalloc(struct btrfs_trans_handle *trans,
3285                            struct btrfs_root *root, u64 to_reclaim, int sync)
3286 {
3287         struct btrfs_block_rsv *block_rsv;
3288         struct btrfs_space_info *space_info;
3289         u64 reserved;
3290         u64 max_reclaim;
3291         u64 reclaimed = 0;
3292         long time_left;
3293         int nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3294         int loops = 0;
3295         unsigned long progress;
3296
3297         block_rsv = &root->fs_info->delalloc_block_rsv;
3298         space_info = block_rsv->space_info;
3299
3300         smp_mb();
3301         reserved = space_info->bytes_reserved;
3302         progress = space_info->reservation_progress;
3303
3304         if (reserved == 0)
3305                 return 0;
3306
3307         /* nothing to shrink - nothing to reclaim */
3308         if (root->fs_info->delalloc_bytes == 0)
3309                 return 0;
3310
3311         max_reclaim = min(reserved, to_reclaim);
3312
3313         while (loops < 1024) {
3314                 /* have the flusher threads jump in and do some IO */
3315                 smp_mb();
3316                 nr_pages = min_t(unsigned long, nr_pages,
3317                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3318                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages);
3319
3320                 spin_lock(&space_info->lock);
3321                 if (reserved > space_info->bytes_reserved)
3322                         reclaimed += reserved - space_info->bytes_reserved;
3323                 reserved = space_info->bytes_reserved;
3324                 spin_unlock(&space_info->lock);
3325
3326                 loops++;
3327
3328                 if (reserved == 0 || reclaimed >= max_reclaim)
3329                         break;
3330
3331                 if (trans && trans->transaction->blocked)
3332                         return -EAGAIN;
3333
3334                 time_left = schedule_timeout_interruptible(1);
3335
3336                 /* We were interrupted, exit */
3337                 if (time_left)
3338                         break;
3339
3340                 /* we've kicked the IO a few times, if anything has been freed,
3341                  * exit.  There is no sense in looping here for a long time
3342                  * when we really need to commit the transaction, or there are
3343                  * just too many writers without enough free space
3344                  */
3345
3346                 if (loops > 3) {
3347                         smp_mb();
3348                         if (progress != space_info->reservation_progress)
3349                                 break;
3350                 }
3351
3352         }
3353         return reclaimed >= to_reclaim;
3354 }
3355
3356 /*
3357  * Retries tells us how many times we've called reserve_metadata_bytes.  The
3358  * idea is if this is the first call (retries == 0) then we will add to our
3359  * reserved count if we can't make the allocation in order to hold our place
3360  * while we go and try and free up space.  That way for retries > 1 we don't try
3361  * and add space, we just check to see if the amount of unused space is >= the
3362  * total space, meaning that our reservation is valid.
3363  *
3364  * However if we don't intend to retry this reservation, pass -1 as retries so
3365  * that it short circuits this logic.
3366  */
3367 static int reserve_metadata_bytes(struct btrfs_trans_handle *trans,
3368                                   struct btrfs_root *root,
3369                                   struct btrfs_block_rsv *block_rsv,
3370                                   u64 orig_bytes, int flush)
3371 {
3372         struct btrfs_space_info *space_info = block_rsv->space_info;
3373         u64 unused;
3374         u64 num_bytes = orig_bytes;
3375         int retries = 0;
3376         int ret = 0;
3377         bool reserved = false;
3378         bool committed = false;
3379
3380 again:
3381         ret = -ENOSPC;
3382         if (reserved)
3383                 num_bytes = 0;
3384
3385         spin_lock(&space_info->lock);
3386         unused = space_info->bytes_used + space_info->bytes_reserved +
3387                  space_info->bytes_pinned + space_info->bytes_readonly +
3388                  space_info->bytes_may_use;
3389
3390         /*
3391          * The idea here is that we've not already over-reserved the block group
3392          * then we can go ahead and save our reservation first and then start
3393          * flushing if we need to.  Otherwise if we've already overcommitted
3394          * lets start flushing stuff first and then come back and try to make
3395          * our reservation.
3396          */
3397         if (unused <= space_info->total_bytes) {
3398                 unused = space_info->total_bytes - unused;
3399                 if (unused >= num_bytes) {
3400                         if (!reserved)
3401                                 space_info->bytes_reserved += orig_bytes;
3402                         ret = 0;
3403                 } else {
3404                         /*
3405                          * Ok set num_bytes to orig_bytes since we aren't
3406                          * overocmmitted, this way we only try and reclaim what
3407                          * we need.
3408                          */
3409                         num_bytes = orig_bytes;
3410                 }
3411         } else {
3412                 /*
3413                  * Ok we're over committed, set num_bytes to the overcommitted
3414                  * amount plus the amount of bytes that we need for this
3415                  * reservation.
3416                  */
3417                 num_bytes = unused - space_info->total_bytes +
3418                         (orig_bytes * (retries + 1));
3419         }
3420
3421         /*
3422          * Couldn't make our reservation, save our place so while we're trying
3423          * to reclaim space we can actually use it instead of somebody else
3424          * stealing it from us.
3425          */
3426         if (ret && !reserved) {
3427                 space_info->bytes_reserved += orig_bytes;
3428                 reserved = true;
3429         }
3430
3431         spin_unlock(&space_info->lock);
3432
3433         if (!ret)
3434                 return 0;
3435
3436         if (!flush)
3437                 goto out;
3438
3439         /*
3440          * We do synchronous shrinking since we don't actually unreserve
3441          * metadata until after the IO is completed.
3442          */
3443         ret = shrink_delalloc(trans, root, num_bytes, 1);
3444         if (ret > 0)
3445                 return 0;
3446         else if (ret < 0)
3447                 goto out;
3448
3449         /*
3450          * So if we were overcommitted it's possible that somebody else flushed
3451          * out enough space and we simply didn't have enough space to reclaim,
3452          * so go back around and try again.
3453          */
3454         if (retries < 2) {
3455                 retries++;
3456                 goto again;
3457         }
3458
3459         spin_lock(&space_info->lock);
3460         /*
3461          * Not enough space to be reclaimed, don't bother committing the
3462          * transaction.
3463          */
3464         if (space_info->bytes_pinned < orig_bytes)
3465                 ret = -ENOSPC;
3466         spin_unlock(&space_info->lock);
3467         if (ret)
3468                 goto out;
3469
3470         ret = -EAGAIN;
3471         if (trans || committed)
3472                 goto out;
3473
3474         ret = -ENOSPC;
3475         trans = btrfs_join_transaction(root, 1);
3476         if (IS_ERR(trans))
3477                 goto out;
3478         ret = btrfs_commit_transaction(trans, root);
3479         if (!ret) {
3480                 trans = NULL;
3481                 committed = true;
3482                 goto again;
3483         }
3484
3485 out:
3486         if (reserved) {
3487                 spin_lock(&space_info->lock);
3488                 space_info->bytes_reserved -= orig_bytes;
3489                 spin_unlock(&space_info->lock);
3490         }
3491
3492         return ret;
3493 }
3494
3495 static struct btrfs_block_rsv *get_block_rsv(struct btrfs_trans_handle *trans,
3496                                              struct btrfs_root *root)
3497 {
3498         struct btrfs_block_rsv *block_rsv;
3499         if (root->ref_cows)
3500                 block_rsv = trans->block_rsv;
3501         else
3502                 block_rsv = root->block_rsv;
3503
3504         if (!block_rsv)
3505                 block_rsv = &root->fs_info->empty_block_rsv;
3506
3507         return block_rsv;
3508 }
3509
3510 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3511                                u64 num_bytes)
3512 {
3513         int ret = -ENOSPC;
3514         spin_lock(&block_rsv->lock);
3515         if (block_rsv->reserved >= num_bytes) {
3516                 block_rsv->reserved -= num_bytes;
3517                 if (block_rsv->reserved < block_rsv->size)
3518                         block_rsv->full = 0;
3519                 ret = 0;
3520         }
3521         spin_unlock(&block_rsv->lock);
3522         return ret;
3523 }
3524
3525 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3526                                 u64 num_bytes, int update_size)
3527 {
3528         spin_lock(&block_rsv->lock);
3529         block_rsv->reserved += num_bytes;
3530         if (update_size)
3531                 block_rsv->size += num_bytes;
3532         else if (block_rsv->reserved >= block_rsv->size)
3533                 block_rsv->full = 1;
3534         spin_unlock(&block_rsv->lock);
3535 }
3536
3537 static void block_rsv_release_bytes(struct btrfs_block_rsv *block_rsv,
3538                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3539 {
3540         struct btrfs_space_info *space_info = block_rsv->space_info;
3541
3542         spin_lock(&block_rsv->lock);
3543         if (num_bytes == (u64)-1)
3544                 num_bytes = block_rsv->size;
3545         block_rsv->size -= num_bytes;
3546         if (block_rsv->reserved >= block_rsv->size) {
3547                 num_bytes = block_rsv->reserved - block_rsv->size;
3548                 block_rsv->reserved = block_rsv->size;
3549                 block_rsv->full = 1;
3550         } else {
3551                 num_bytes = 0;
3552         }
3553         spin_unlock(&block_rsv->lock);
3554
3555         if (num_bytes > 0) {
3556                 if (dest) {
3557                         spin_lock(&dest->lock);
3558                         if (!dest->full) {
3559                                 u64 bytes_to_add;
3560
3561                                 bytes_to_add = dest->size - dest->reserved;
3562                                 bytes_to_add = min(num_bytes, bytes_to_add);
3563                                 dest->reserved += bytes_to_add;
3564                                 if (dest->reserved >= dest->size)
3565                                         dest->full = 1;
3566                                 num_bytes -= bytes_to_add;
3567                         }
3568                         spin_unlock(&dest->lock);
3569                 }
3570                 if (num_bytes) {
3571                         spin_lock(&space_info->lock);
3572                         space_info->bytes_reserved -= num_bytes;
3573                         space_info->reservation_progress++;
3574                         spin_unlock(&space_info->lock);
3575                 }
3576         }
3577 }
3578
3579 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3580                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3581 {
3582         int ret;
3583
3584         ret = block_rsv_use_bytes(src, num_bytes);
3585         if (ret)
3586                 return ret;
3587
3588         block_rsv_add_bytes(dst, num_bytes, 1);
3589         return 0;
3590 }
3591
3592 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3593 {
3594         memset(rsv, 0, sizeof(*rsv));
3595         spin_lock_init(&rsv->lock);
3596         atomic_set(&rsv->usage, 1);
3597         rsv->priority = 6;
3598         INIT_LIST_HEAD(&rsv->list);
3599 }
3600
3601 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3602 {
3603         struct btrfs_block_rsv *block_rsv;
3604         struct btrfs_fs_info *fs_info = root->fs_info;
3605
3606         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3607         if (!block_rsv)
3608                 return NULL;
3609
3610         btrfs_init_block_rsv(block_rsv);
3611         block_rsv->space_info = __find_space_info(fs_info,
3612                                                   BTRFS_BLOCK_GROUP_METADATA);
3613         return block_rsv;
3614 }
3615
3616 void btrfs_free_block_rsv(struct btrfs_root *root,
3617                           struct btrfs_block_rsv *rsv)
3618 {
3619         if (rsv && atomic_dec_and_test(&rsv->usage)) {
3620                 btrfs_block_rsv_release(root, rsv, (u64)-1);
3621                 if (!rsv->durable)
3622                         kfree(rsv);
3623         }
3624 }
3625
3626 /*
3627  * make the block_rsv struct be able to capture freed space.
3628  * the captured space will re-add to the the block_rsv struct
3629  * after transaction commit
3630  */
3631 void btrfs_add_durable_block_rsv(struct btrfs_fs_info *fs_info,
3632                                  struct btrfs_block_rsv *block_rsv)
3633 {
3634         block_rsv->durable = 1;
3635         mutex_lock(&fs_info->durable_block_rsv_mutex);
3636         list_add_tail(&block_rsv->list, &fs_info->durable_block_rsv_list);
3637         mutex_unlock(&fs_info->durable_block_rsv_mutex);
3638 }
3639
3640 int btrfs_block_rsv_add(struct btrfs_trans_handle *trans,
3641                         struct btrfs_root *root,
3642                         struct btrfs_block_rsv *block_rsv,
3643                         u64 num_bytes)
3644 {
3645         int ret;
3646
3647         if (num_bytes == 0)
3648                 return 0;
3649
3650         ret = reserve_metadata_bytes(trans, root, block_rsv, num_bytes, 1);
3651         if (!ret) {
3652                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3653                 return 0;
3654         }
3655
3656         return ret;
3657 }
3658
3659 int btrfs_block_rsv_check(struct btrfs_trans_handle *trans,
3660                           struct btrfs_root *root,
3661                           struct btrfs_block_rsv *block_rsv,
3662                           u64 min_reserved, int min_factor)
3663 {
3664         u64 num_bytes = 0;
3665         int commit_trans = 0;
3666         int ret = -ENOSPC;
3667
3668         if (!block_rsv)
3669                 return 0;
3670
3671         spin_lock(&block_rsv->lock);
3672         if (min_factor > 0)
3673                 num_bytes = div_factor(block_rsv->size, min_factor);
3674         if (min_reserved > num_bytes)
3675                 num_bytes = min_reserved;
3676
3677         if (block_rsv->reserved >= num_bytes) {
3678                 ret = 0;
3679         } else {
3680                 num_bytes -= block_rsv->reserved;
3681                 if (block_rsv->durable &&
3682                     block_rsv->freed[0] + block_rsv->freed[1] >= num_bytes)
3683                         commit_trans = 1;
3684         }
3685         spin_unlock(&block_rsv->lock);
3686         if (!ret)
3687                 return 0;
3688
3689         if (block_rsv->refill_used) {
3690                 ret = reserve_metadata_bytes(trans, root, block_rsv,
3691                                              num_bytes, 0);
3692                 if (!ret) {
3693                         block_rsv_add_bytes(block_rsv, num_bytes, 0);
3694                         return 0;
3695                 }
3696         }
3697
3698         if (commit_trans) {
3699                 if (trans)
3700                         return -EAGAIN;
3701
3702                 trans = btrfs_join_transaction(root, 1);
3703                 BUG_ON(IS_ERR(trans));
3704                 ret = btrfs_commit_transaction(trans, root);
3705                 return 0;
3706         }
3707
3708         return -ENOSPC;
3709 }
3710
3711 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
3712                             struct btrfs_block_rsv *dst_rsv,
3713                             u64 num_bytes)
3714 {
3715         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3716 }
3717
3718 void btrfs_block_rsv_release(struct btrfs_root *root,
3719                              struct btrfs_block_rsv *block_rsv,
3720                              u64 num_bytes)
3721 {
3722         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3723         if (global_rsv->full || global_rsv == block_rsv ||
3724             block_rsv->space_info != global_rsv->space_info)
3725                 global_rsv = NULL;
3726         block_rsv_release_bytes(block_rsv, global_rsv, num_bytes);
3727 }
3728
3729 /*
3730  * helper to calculate size of global block reservation.
3731  * the desired value is sum of space used by extent tree,
3732  * checksum tree and root tree
3733  */
3734 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
3735 {
3736         struct btrfs_space_info *sinfo;
3737         u64 num_bytes;
3738         u64 meta_used;
3739         u64 data_used;
3740         int csum_size = btrfs_super_csum_size(&fs_info->super_copy);
3741
3742         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
3743         spin_lock(&sinfo->lock);
3744         data_used = sinfo->bytes_used;
3745         spin_unlock(&sinfo->lock);
3746
3747         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3748         spin_lock(&sinfo->lock);
3749         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
3750                 data_used = 0;
3751         meta_used = sinfo->bytes_used;
3752         spin_unlock(&sinfo->lock);
3753
3754         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
3755                     csum_size * 2;
3756         num_bytes += div64_u64(data_used + meta_used, 50);
3757
3758         if (num_bytes * 3 > meta_used)
3759                 num_bytes = div64_u64(meta_used, 3);
3760
3761         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
3762 }
3763
3764 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
3765 {
3766         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
3767         struct btrfs_space_info *sinfo = block_rsv->space_info;
3768         u64 num_bytes;
3769
3770         num_bytes = calc_global_metadata_size(fs_info);
3771
3772         spin_lock(&block_rsv->lock);
3773         spin_lock(&sinfo->lock);
3774
3775         block_rsv->size = num_bytes;
3776
3777         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
3778                     sinfo->bytes_reserved + sinfo->bytes_readonly +
3779                     sinfo->bytes_may_use;
3780
3781         if (sinfo->total_bytes > num_bytes) {
3782                 num_bytes = sinfo->total_bytes - num_bytes;
3783                 block_rsv->reserved += num_bytes;
3784                 sinfo->bytes_reserved += num_bytes;
3785         }
3786
3787         if (block_rsv->reserved >= block_rsv->size) {
3788                 num_bytes = block_rsv->reserved - block_rsv->size;
3789                 sinfo->bytes_reserved -= num_bytes;
3790                 sinfo->reservation_progress++;
3791                 block_rsv->reserved = block_rsv->size;
3792                 block_rsv->full = 1;
3793         }
3794
3795         spin_unlock(&sinfo->lock);
3796         spin_unlock(&block_rsv->lock);
3797 }
3798
3799 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
3800 {
3801         struct btrfs_space_info *space_info;
3802
3803         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3804         fs_info->chunk_block_rsv.space_info = space_info;
3805         fs_info->chunk_block_rsv.priority = 10;
3806
3807         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3808         fs_info->global_block_rsv.space_info = space_info;
3809         fs_info->global_block_rsv.priority = 10;
3810         fs_info->global_block_rsv.refill_used = 1;
3811         fs_info->delalloc_block_rsv.space_info = space_info;
3812         fs_info->trans_block_rsv.space_info = space_info;
3813         fs_info->empty_block_rsv.space_info = space_info;
3814         fs_info->empty_block_rsv.priority = 10;
3815
3816         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
3817         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
3818         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
3819         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
3820         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
3821
3822         btrfs_add_durable_block_rsv(fs_info, &fs_info->global_block_rsv);
3823
3824         btrfs_add_durable_block_rsv(fs_info, &fs_info->delalloc_block_rsv);
3825
3826         update_global_block_rsv(fs_info);
3827 }
3828
3829 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
3830 {
3831         block_rsv_release_bytes(&fs_info->global_block_rsv, NULL, (u64)-1);
3832         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
3833         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
3834         WARN_ON(fs_info->trans_block_rsv.size > 0);
3835         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
3836         WARN_ON(fs_info->chunk_block_rsv.size > 0);
3837         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
3838 }
3839
3840 int btrfs_trans_reserve_metadata(struct btrfs_trans_handle *trans,
3841                                  struct btrfs_root *root,
3842                                  int num_items)
3843 {
3844         u64 num_bytes;
3845         int ret;
3846
3847         if (num_items == 0 || root->fs_info->chunk_root == root)
3848                 return 0;
3849
3850         num_bytes = btrfs_calc_trans_metadata_size(root, num_items);
3851         ret = btrfs_block_rsv_add(trans, root, &root->fs_info->trans_block_rsv,
3852                                   num_bytes);
3853         if (!ret) {
3854                 trans->bytes_reserved += num_bytes;
3855                 trans->block_rsv = &root->fs_info->trans_block_rsv;
3856         }
3857         return ret;
3858 }
3859
3860 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
3861                                   struct btrfs_root *root)
3862 {
3863         if (!trans->bytes_reserved)
3864                 return;
3865
3866         BUG_ON(trans->block_rsv != &root->fs_info->trans_block_rsv);
3867         btrfs_block_rsv_release(root, trans->block_rsv,
3868                                 trans->bytes_reserved);
3869         trans->bytes_reserved = 0;
3870 }
3871
3872 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
3873                                   struct inode *inode)
3874 {
3875         struct btrfs_root *root = BTRFS_I(inode)->root;
3876         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
3877         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
3878
3879         /*
3880          * one for deleting orphan item, one for updating inode and
3881          * two for calling btrfs_truncate_inode_items.
3882          *
3883          * btrfs_truncate_inode_items is a delete operation, it frees
3884          * more space than it uses in most cases. So two units of
3885          * metadata space should be enough for calling it many times.
3886          * If all of the metadata space is used, we can commit
3887          * transaction and use space it freed.
3888          */
3889         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 4);
3890         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3891 }
3892
3893 void btrfs_orphan_release_metadata(struct inode *inode)
3894 {
3895         struct btrfs_root *root = BTRFS_I(inode)->root;
3896         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 4);
3897         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
3898 }
3899
3900 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
3901                                 struct btrfs_pending_snapshot *pending)
3902 {
3903         struct btrfs_root *root = pending->root;
3904         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
3905         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
3906         /*
3907          * two for root back/forward refs, two for directory entries
3908          * and one for root of the snapshot.
3909          */
3910         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
3911         dst_rsv->space_info = src_rsv->space_info;
3912         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3913 }
3914
3915 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes)
3916 {
3917         return num_bytes >>= 3;
3918 }
3919
3920 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
3921 {
3922         struct btrfs_root *root = BTRFS_I(inode)->root;
3923         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
3924         u64 to_reserve;
3925         int nr_extents;
3926         int reserved_extents;
3927         int ret;
3928
3929         if (btrfs_transaction_in_commit(root->fs_info))
3930                 schedule_timeout(1);
3931
3932         num_bytes = ALIGN(num_bytes, root->sectorsize);
3933
3934         nr_extents = atomic_read(&BTRFS_I(inode)->outstanding_extents) + 1;
3935         reserved_extents = atomic_read(&BTRFS_I(inode)->reserved_extents);
3936
3937         if (nr_extents > reserved_extents) {
3938                 nr_extents -= reserved_extents;
3939                 to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
3940         } else {
3941                 nr_extents = 0;
3942                 to_reserve = 0;
3943         }
3944
3945         to_reserve += calc_csum_metadata_size(inode, num_bytes);
3946         ret = reserve_metadata_bytes(NULL, root, block_rsv, to_reserve, 1);
3947         if (ret)
3948                 return ret;
3949
3950         atomic_add(nr_extents, &BTRFS_I(inode)->reserved_extents);
3951         atomic_inc(&BTRFS_I(inode)->outstanding_extents);
3952
3953         block_rsv_add_bytes(block_rsv, to_reserve, 1);
3954
3955         if (block_rsv->size > 512 * 1024 * 1024)
3956                 shrink_delalloc(NULL, root, to_reserve, 0);
3957
3958         return 0;
3959 }
3960
3961 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
3962 {
3963         struct btrfs_root *root = BTRFS_I(inode)->root;
3964         u64 to_free;
3965         int nr_extents;
3966         int reserved_extents;
3967
3968         num_bytes = ALIGN(num_bytes, root->sectorsize);
3969         atomic_dec(&BTRFS_I(inode)->outstanding_extents);
3970         WARN_ON(atomic_read(&BTRFS_I(inode)->outstanding_extents) < 0);
3971
3972         reserved_extents = atomic_read(&BTRFS_I(inode)->reserved_extents);
3973         do {
3974                 int old, new;
3975
3976                 nr_extents = atomic_read(&BTRFS_I(inode)->outstanding_extents);
3977                 if (nr_extents >= reserved_extents) {
3978                         nr_extents = 0;
3979                         break;
3980                 }
3981                 old = reserved_extents;
3982                 nr_extents = reserved_extents - nr_extents;
3983                 new = reserved_extents - nr_extents;
3984                 old = atomic_cmpxchg(&BTRFS_I(inode)->reserved_extents,
3985                                      reserved_extents, new);
3986                 if (likely(old == reserved_extents))
3987                         break;
3988                 reserved_extents = old;
3989         } while (1);
3990
3991         to_free = calc_csum_metadata_size(inode, num_bytes);
3992         if (nr_extents > 0)
3993                 to_free += btrfs_calc_trans_metadata_size(root, nr_extents);
3994
3995         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
3996                                 to_free);
3997 }
3998
3999 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4000 {
4001         int ret;
4002
4003         ret = btrfs_check_data_free_space(inode, num_bytes);
4004         if (ret)
4005                 return ret;
4006
4007         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4008         if (ret) {
4009                 btrfs_free_reserved_data_space(inode, num_bytes);
4010                 return ret;
4011         }
4012
4013         return 0;
4014 }
4015
4016 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4017 {
4018         btrfs_delalloc_release_metadata(inode, num_bytes);
4019         btrfs_free_reserved_data_space(inode, num_bytes);
4020 }
4021
4022 static int update_block_group(struct btrfs_trans_handle *trans,
4023                               struct btrfs_root *root,
4024                               u64 bytenr, u64 num_bytes, int alloc)
4025 {
4026         struct btrfs_block_group_cache *cache = NULL;
4027         struct btrfs_fs_info *info = root->fs_info;
4028         u64 total = num_bytes;
4029         u64 old_val;
4030         u64 byte_in_group;
4031         int factor;
4032
4033         /* block accounting for super block */
4034         spin_lock(&info->delalloc_lock);
4035         old_val = btrfs_super_bytes_used(&info->super_copy);
4036         if (alloc)
4037                 old_val += num_bytes;
4038         else
4039                 old_val -= num_bytes;
4040         btrfs_set_super_bytes_used(&info->super_copy, old_val);
4041         spin_unlock(&info->delalloc_lock);
4042
4043         while (total) {
4044                 cache = btrfs_lookup_block_group(info, bytenr);
4045                 if (!cache)
4046                         return -1;
4047                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4048                                     BTRFS_BLOCK_GROUP_RAID1 |
4049                                     BTRFS_BLOCK_GROUP_RAID10))
4050                         factor = 2;
4051                 else
4052                         factor = 1;
4053                 /*
4054                  * If this block group has free space cache written out, we
4055                  * need to make sure to load it if we are removing space.  This
4056                  * is because we need the unpinning stage to actually add the
4057                  * space back to the block group, otherwise we will leak space.
4058                  */
4059                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4060                         cache_block_group(cache, trans, NULL, 1);
4061
4062                 byte_in_group = bytenr - cache->key.objectid;
4063                 WARN_ON(byte_in_group > cache->key.offset);
4064
4065                 spin_lock(&cache->space_info->lock);
4066                 spin_lock(&cache->lock);
4067
4068                 if (btrfs_super_cache_generation(&info->super_copy) != 0 &&
4069                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4070                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4071
4072                 cache->dirty = 1;
4073                 old_val = btrfs_block_group_used(&cache->item);
4074                 num_bytes = min(total, cache->key.offset - byte_in_group);
4075                 if (alloc) {
4076                         old_val += num_bytes;
4077                         btrfs_set_block_group_used(&cache->item, old_val);
4078                         cache->reserved -= num_bytes;
4079                         cache->space_info->bytes_reserved -= num_bytes;
4080                         cache->space_info->reservation_progress++;
4081                         cache->space_info->bytes_used += num_bytes;
4082                         cache->space_info->disk_used += num_bytes * factor;
4083                         spin_unlock(&cache->lock);
4084                         spin_unlock(&cache->space_info->lock);
4085                 } else {
4086                         old_val -= num_bytes;
4087                         btrfs_set_block_group_used(&cache->item, old_val);
4088                         cache->pinned += num_bytes;
4089                         cache->space_info->bytes_pinned += num_bytes;
4090                         cache->space_info->bytes_used -= num_bytes;
4091                         cache->space_info->disk_used -= num_bytes * factor;
4092                         spin_unlock(&cache->lock);
4093                         spin_unlock(&cache->space_info->lock);
4094
4095                         set_extent_dirty(info->pinned_extents,
4096                                          bytenr, bytenr + num_bytes - 1,
4097                                          GFP_NOFS | __GFP_NOFAIL);
4098                 }
4099                 btrfs_put_block_group(cache);
4100                 total -= num_bytes;
4101                 bytenr += num_bytes;
4102         }
4103         return 0;
4104 }
4105
4106 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4107 {
4108         struct btrfs_block_group_cache *cache;
4109         u64 bytenr;
4110
4111         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4112         if (!cache)
4113                 return 0;
4114
4115         bytenr = cache->key.objectid;
4116         btrfs_put_block_group(cache);
4117
4118         return bytenr;
4119 }
4120
4121 static int pin_down_extent(struct btrfs_root *root,
4122                            struct btrfs_block_group_cache *cache,
4123                            u64 bytenr, u64 num_bytes, int reserved)
4124 {
4125         spin_lock(&cache->space_info->lock);
4126         spin_lock(&cache->lock);
4127         cache->pinned += num_bytes;
4128         cache->space_info->bytes_pinned += num_bytes;
4129         if (reserved) {
4130                 cache->reserved -= num_bytes;
4131                 cache->space_info->bytes_reserved -= num_bytes;
4132                 cache->space_info->reservation_progress++;
4133         }
4134         spin_unlock(&cache->lock);
4135         spin_unlock(&cache->space_info->lock);
4136
4137         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4138                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4139         return 0;
4140 }
4141
4142 /*
4143  * this function must be called within transaction
4144  */
4145 int btrfs_pin_extent(struct btrfs_root *root,
4146                      u64 bytenr, u64 num_bytes, int reserved)
4147 {
4148         struct btrfs_block_group_cache *cache;
4149
4150         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4151         BUG_ON(!cache);
4152
4153         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4154
4155         btrfs_put_block_group(cache);
4156         return 0;
4157 }
4158
4159 /*
4160  * update size of reserved extents. this function may return -EAGAIN
4161  * if 'reserve' is true or 'sinfo' is false.
4162  */
4163 int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4164                                 u64 num_bytes, int reserve, int sinfo)
4165 {
4166         int ret = 0;
4167         if (sinfo) {
4168                 struct btrfs_space_info *space_info = cache->space_info;
4169                 spin_lock(&space_info->lock);
4170                 spin_lock(&cache->lock);
4171                 if (reserve) {
4172                         if (cache->ro) {
4173                                 ret = -EAGAIN;
4174                         } else {
4175                                 cache->reserved += num_bytes;
4176                                 space_info->bytes_reserved += num_bytes;
4177                         }
4178                 } else {
4179                         if (cache->ro)
4180                                 space_info->bytes_readonly += num_bytes;
4181                         cache->reserved -= num_bytes;
4182                         space_info->bytes_reserved -= num_bytes;
4183                         space_info->reservation_progress++;
4184                 }
4185                 spin_unlock(&cache->lock);
4186                 spin_unlock(&space_info->lock);
4187         } else {
4188                 spin_lock(&cache->lock);
4189                 if (cache->ro) {
4190                         ret = -EAGAIN;
4191                 } else {
4192                         if (reserve)
4193                                 cache->reserved += num_bytes;
4194                         else
4195                                 cache->reserved -= num_bytes;
4196                 }
4197                 spin_unlock(&cache->lock);
4198         }
4199         return ret;
4200 }
4201
4202 int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4203                                 struct btrfs_root *root)
4204 {
4205         struct btrfs_fs_info *fs_info = root->fs_info;
4206         struct btrfs_caching_control *next;
4207         struct btrfs_caching_control *caching_ctl;
4208         struct btrfs_block_group_cache *cache;
4209
4210         down_write(&fs_info->extent_commit_sem);
4211
4212         list_for_each_entry_safe(caching_ctl, next,
4213                                  &fs_info->caching_block_groups, list) {
4214                 cache = caching_ctl->block_group;
4215                 if (block_group_cache_done(cache)) {
4216                         cache->last_byte_to_unpin = (u64)-1;
4217                         list_del_init(&caching_ctl->list);
4218                         put_caching_control(caching_ctl);
4219                 } else {
4220                         cache->last_byte_to_unpin = caching_ctl->progress;
4221                 }
4222         }
4223
4224         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4225                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4226         else
4227                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4228
4229         up_write(&fs_info->extent_commit_sem);
4230
4231         update_global_block_rsv(fs_info);
4232         return 0;
4233 }
4234
4235 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4236 {
4237         struct btrfs_fs_info *fs_info = root->fs_info;
4238         struct btrfs_block_group_cache *cache = NULL;
4239         u64 len;
4240
4241         while (start <= end) {
4242                 if (!cache ||
4243                     start >= cache->key.objectid + cache->key.offset) {
4244                         if (cache)
4245                                 btrfs_put_block_group(cache);
4246                         cache = btrfs_lookup_block_group(fs_info, start);
4247                         BUG_ON(!cache);
4248                 }
4249
4250                 len = cache->key.objectid + cache->key.offset - start;
4251                 len = min(len, end + 1 - start);
4252
4253                 if (start < cache->last_byte_to_unpin) {
4254                         len = min(len, cache->last_byte_to_unpin - start);
4255                         btrfs_add_free_space(cache, start, len);
4256                 }
4257
4258                 start += len;
4259
4260                 spin_lock(&cache->space_info->lock);
4261                 spin_lock(&cache->lock);
4262                 cache->pinned -= len;
4263                 cache->space_info->bytes_pinned -= len;
4264                 if (cache->ro) {
4265                         cache->space_info->bytes_readonly += len;
4266                 } else if (cache->reserved_pinned > 0) {
4267                         len = min(len, cache->reserved_pinned);
4268                         cache->reserved_pinned -= len;
4269                         cache->space_info->bytes_reserved += len;
4270                 }
4271                 spin_unlock(&cache->lock);
4272                 spin_unlock(&cache->space_info->lock);
4273         }
4274
4275         if (cache)
4276                 btrfs_put_block_group(cache);
4277         return 0;
4278 }
4279
4280 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4281                                struct btrfs_root *root)
4282 {
4283         struct btrfs_fs_info *fs_info = root->fs_info;
4284         struct extent_io_tree *unpin;
4285         struct btrfs_block_rsv *block_rsv;
4286         struct btrfs_block_rsv *next_rsv;
4287         u64 start;
4288         u64 end;
4289         int idx;
4290         int ret;
4291
4292         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4293                 unpin = &fs_info->freed_extents[1];
4294         else
4295                 unpin = &fs_info->freed_extents[0];
4296
4297         while (1) {
4298                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4299                                             EXTENT_DIRTY);
4300                 if (ret)
4301                         break;
4302
4303                 if (btrfs_test_opt(root, DISCARD))
4304                         ret = btrfs_discard_extent(root, start,
4305                                                    end + 1 - start, NULL);
4306
4307                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4308                 unpin_extent_range(root, start, end);
4309                 cond_resched();
4310         }
4311
4312         mutex_lock(&fs_info->durable_block_rsv_mutex);
4313         list_for_each_entry_safe(block_rsv, next_rsv,
4314                                  &fs_info->durable_block_rsv_list, list) {
4315
4316                 idx = trans->transid & 0x1;
4317                 if (block_rsv->freed[idx] > 0) {
4318                         block_rsv_add_bytes(block_rsv,
4319                                             block_rsv->freed[idx], 0);
4320                         block_rsv->freed[idx] = 0;
4321                 }
4322                 if (atomic_read(&block_rsv->usage) == 0) {
4323                         btrfs_block_rsv_release(root, block_rsv, (u64)-1);
4324
4325                         if (block_rsv->freed[0] == 0 &&
4326                             block_rsv->freed[1] == 0) {
4327                                 list_del_init(&block_rsv->list);
4328                                 kfree(block_rsv);
4329                         }
4330                 } else {
4331                         btrfs_block_rsv_release(root, block_rsv, 0);
4332                 }
4333         }
4334         mutex_unlock(&fs_info->durable_block_rsv_mutex);
4335
4336         return 0;
4337 }
4338
4339 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4340                                 struct btrfs_root *root,
4341                                 u64 bytenr, u64 num_bytes, u64 parent,
4342                                 u64 root_objectid, u64 owner_objectid,
4343                                 u64 owner_offset, int refs_to_drop,
4344                                 struct btrfs_delayed_extent_op *extent_op)
4345 {
4346         struct btrfs_key key;
4347         struct btrfs_path *path;
4348         struct btrfs_fs_info *info = root->fs_info;
4349         struct btrfs_root *extent_root = info->extent_root;
4350         struct extent_buffer *leaf;
4351         struct btrfs_extent_item *ei;
4352         struct btrfs_extent_inline_ref *iref;
4353         int ret;
4354         int is_data;
4355         int extent_slot = 0;
4356         int found_extent = 0;
4357         int num_to_del = 1;
4358         u32 item_size;
4359         u64 refs;
4360
4361         path = btrfs_alloc_path();
4362         if (!path)
4363                 return -ENOMEM;
4364
4365         path->reada = 1;
4366         path->leave_spinning = 1;
4367
4368         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4369         BUG_ON(!is_data && refs_to_drop != 1);
4370
4371         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4372                                     bytenr, num_bytes, parent,
4373                                     root_objectid, owner_objectid,
4374                                     owner_offset);
4375         if (ret == 0) {
4376                 extent_slot = path->slots[0];
4377                 while (extent_slot >= 0) {
4378                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4379                                               extent_slot);
4380                         if (key.objectid != bytenr)
4381                                 break;
4382                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4383                             key.offset == num_bytes) {
4384                                 found_extent = 1;
4385                                 break;
4386                         }
4387                         if (path->slots[0] - extent_slot > 5)
4388                                 break;
4389                         extent_slot--;
4390                 }
4391 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4392                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4393                 if (found_extent && item_size < sizeof(*ei))
4394                         found_extent = 0;
4395 #endif
4396                 if (!found_extent) {
4397                         BUG_ON(iref);
4398                         ret = remove_extent_backref(trans, extent_root, path,
4399                                                     NULL, refs_to_drop,
4400                                                     is_data);
4401                         BUG_ON(ret);
4402                         btrfs_release_path(path);
4403                         path->leave_spinning = 1;
4404
4405                         key.objectid = bytenr;
4406                         key.type = BTRFS_EXTENT_ITEM_KEY;
4407                         key.offset = num_bytes;
4408
4409                         ret = btrfs_search_slot(trans, extent_root,
4410                                                 &key, path, -1, 1);
4411                         if (ret) {
4412                                 printk(KERN_ERR "umm, got %d back from search"
4413                                        ", was looking for %llu\n", ret,
4414                                        (unsigned long long)bytenr);
4415                                 btrfs_print_leaf(extent_root, path->nodes[0]);
4416                         }
4417                         BUG_ON(ret);
4418                         extent_slot = path->slots[0];
4419                 }
4420         } else {
4421                 btrfs_print_leaf(extent_root, path->nodes[0]);
4422                 WARN_ON(1);
4423                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4424                        "parent %llu root %llu  owner %llu offset %llu\n",
4425                        (unsigned long long)bytenr,
4426                        (unsigned long long)parent,
4427                        (unsigned long long)root_objectid,
4428                        (unsigned long long)owner_objectid,
4429                        (unsigned long long)owner_offset);
4430         }
4431
4432         leaf = path->nodes[0];
4433         item_size = btrfs_item_size_nr(leaf, extent_slot);
4434 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4435         if (item_size < sizeof(*ei)) {
4436                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4437                 ret = convert_extent_item_v0(trans, extent_root, path,
4438                                              owner_objectid, 0);
4439                 BUG_ON(ret < 0);
4440
4441                 btrfs_release_path(path);
4442                 path->leave_spinning = 1;
4443
4444                 key.objectid = bytenr;
4445                 key.type = BTRFS_EXTENT_ITEM_KEY;
4446                 key.offset = num_bytes;
4447
4448                 ret = btrfs_search_slot(trans, extent_root, &key, path,
4449                                         -1, 1);
4450                 if (ret) {
4451                         printk(KERN_ERR "umm, got %d back from search"
4452                                ", was looking for %llu\n", ret,
4453                                (unsigned long long)bytenr);
4454                         btrfs_print_leaf(extent_root, path->nodes[0]);
4455                 }
4456                 BUG_ON(ret);
4457                 extent_slot = path->slots[0];
4458                 leaf = path->nodes[0];
4459                 item_size = btrfs_item_size_nr(leaf, extent_slot);
4460         }
4461 #endif
4462         BUG_ON(item_size < sizeof(*ei));
4463         ei = btrfs_item_ptr(leaf, extent_slot,
4464                             struct btrfs_extent_item);
4465         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
4466                 struct btrfs_tree_block_info *bi;
4467                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4468                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4469                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4470         }
4471
4472         refs = btrfs_extent_refs(leaf, ei);
4473         BUG_ON(refs < refs_to_drop);
4474         refs -= refs_to_drop;
4475
4476         if (refs > 0) {
4477                 if (extent_op)
4478                         __run_delayed_extent_op(extent_op, leaf, ei);
4479                 /*
4480                  * In the case of inline back ref, reference count will
4481                  * be updated by remove_extent_backref
4482                  */
4483                 if (iref) {
4484                         BUG_ON(!found_extent);
4485                 } else {
4486                         btrfs_set_extent_refs(leaf, ei, refs);
4487                         btrfs_mark_buffer_dirty(leaf);
4488                 }
4489                 if (found_extent) {
4490                         ret = remove_extent_backref(trans, extent_root, path,
4491                                                     iref, refs_to_drop,
4492                                                     is_data);
4493                         BUG_ON(ret);
4494                 }
4495         } else {
4496                 if (found_extent) {
4497                         BUG_ON(is_data && refs_to_drop !=
4498                                extent_data_ref_count(root, path, iref));
4499                         if (iref) {
4500                                 BUG_ON(path->slots[0] != extent_slot);
4501                         } else {
4502                                 BUG_ON(path->slots[0] != extent_slot + 1);
4503                                 path->slots[0] = extent_slot;
4504                                 num_to_del = 2;
4505                         }
4506                 }
4507
4508                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
4509                                       num_to_del);
4510                 BUG_ON(ret);
4511                 btrfs_release_path(path);
4512
4513                 if (is_data) {
4514                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
4515                         BUG_ON(ret);
4516                 } else {
4517                         invalidate_mapping_pages(info->btree_inode->i_mapping,
4518                              bytenr >> PAGE_CACHE_SHIFT,
4519                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
4520                 }
4521
4522                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
4523                 BUG_ON(ret);
4524         }
4525         btrfs_free_path(path);
4526         return ret;
4527 }
4528
4529 /*
4530  * when we free an block, it is possible (and likely) that we free the last
4531  * delayed ref for that extent as well.  This searches the delayed ref tree for
4532  * a given extent, and if there are no other delayed refs to be processed, it
4533  * removes it from the tree.
4534  */
4535 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
4536                                       struct btrfs_root *root, u64 bytenr)
4537 {
4538         struct btrfs_delayed_ref_head *head;
4539         struct btrfs_delayed_ref_root *delayed_refs;
4540         struct btrfs_delayed_ref_node *ref;
4541         struct rb_node *node;
4542         int ret = 0;
4543
4544         delayed_refs = &trans->transaction->delayed_refs;
4545         spin_lock(&delayed_refs->lock);
4546         head = btrfs_find_delayed_ref_head(trans, bytenr);
4547         if (!head)
4548                 goto out;
4549
4550         node = rb_prev(&head->node.rb_node);
4551         if (!node)
4552                 goto out;
4553
4554         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
4555
4556         /* there are still entries for this ref, we can't drop it */
4557         if (ref->bytenr == bytenr)
4558                 goto out;
4559
4560         if (head->extent_op) {
4561                 if (!head->must_insert_reserved)
4562                         goto out;
4563                 kfree(head->extent_op);
4564                 head->extent_op = NULL;
4565         }
4566
4567         /*
4568          * waiting for the lock here would deadlock.  If someone else has it
4569          * locked they are already in the process of dropping it anyway
4570          */
4571         if (!mutex_trylock(&head->mutex))
4572                 goto out;
4573
4574         /*
4575          * at this point we have a head with no other entries.  Go
4576          * ahead and process it.
4577          */
4578         head->node.in_tree = 0;
4579         rb_erase(&head->node.rb_node, &delayed_refs->root);
4580
4581         delayed_refs->num_entries--;
4582
4583         /*
4584          * we don't take a ref on the node because we're removing it from the
4585          * tree, so we just steal the ref the tree was holding.
4586          */
4587         delayed_refs->num_heads--;
4588         if (list_empty(&head->cluster))
4589                 delayed_refs->num_heads_ready--;
4590
4591         list_del_init(&head->cluster);
4592         spin_unlock(&delayed_refs->lock);
4593
4594         BUG_ON(head->extent_op);
4595         if (head->must_insert_reserved)
4596                 ret = 1;
4597
4598         mutex_unlock(&head->mutex);
4599         btrfs_put_delayed_ref(&head->node);
4600         return ret;
4601 out:
4602         spin_unlock(&delayed_refs->lock);
4603         return 0;
4604 }
4605
4606 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
4607                            struct btrfs_root *root,
4608                            struct extent_buffer *buf,
4609                            u64 parent, int last_ref)
4610 {
4611         struct btrfs_block_rsv *block_rsv;
4612         struct btrfs_block_group_cache *cache = NULL;
4613         int ret;
4614
4615         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4616                 ret = btrfs_add_delayed_tree_ref(trans, buf->start, buf->len,
4617                                                 parent, root->root_key.objectid,
4618                                                 btrfs_header_level(buf),
4619                                                 BTRFS_DROP_DELAYED_REF, NULL);
4620                 BUG_ON(ret);
4621         }
4622
4623         if (!last_ref)
4624                 return;
4625
4626         block_rsv = get_block_rsv(trans, root);
4627         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
4628         if (block_rsv->space_info != cache->space_info)
4629                 goto out;
4630
4631         if (btrfs_header_generation(buf) == trans->transid) {
4632                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4633                         ret = check_ref_cleanup(trans, root, buf->start);
4634                         if (!ret)
4635                                 goto pin;
4636                 }
4637
4638                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
4639                         pin_down_extent(root, cache, buf->start, buf->len, 1);
4640                         goto pin;
4641                 }
4642
4643                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
4644
4645                 btrfs_add_free_space(cache, buf->start, buf->len);
4646                 ret = btrfs_update_reserved_bytes(cache, buf->len, 0, 0);
4647                 if (ret == -EAGAIN) {
4648                         /* block group became read-only */
4649                         btrfs_update_reserved_bytes(cache, buf->len, 0, 1);
4650                         goto out;
4651                 }
4652
4653                 ret = 1;
4654                 spin_lock(&block_rsv->lock);
4655                 if (block_rsv->reserved < block_rsv->size) {
4656                         block_rsv->reserved += buf->len;
4657                         ret = 0;
4658                 }
4659                 spin_unlock(&block_rsv->lock);
4660
4661                 if (ret) {
4662                         spin_lock(&cache->space_info->lock);
4663                         cache->space_info->bytes_reserved -= buf->len;
4664                         cache->space_info->reservation_progress++;
4665                         spin_unlock(&cache->space_info->lock);
4666                 }
4667                 goto out;
4668         }
4669 pin:
4670         if (block_rsv->durable && !cache->ro) {
4671                 ret = 0;
4672                 spin_lock(&cache->lock);
4673                 if (!cache->ro) {
4674                         cache->reserved_pinned += buf->len;
4675                         ret = 1;
4676                 }
4677                 spin_unlock(&cache->lock);
4678
4679                 if (ret) {
4680                         spin_lock(&block_rsv->lock);
4681                         block_rsv->freed[trans->transid & 0x1] += buf->len;
4682                         spin_unlock(&block_rsv->lock);
4683                 }
4684         }
4685 out:
4686         /*
4687          * Deleting the buffer, clear the corrupt flag since it doesn't matter
4688          * anymore.
4689          */
4690         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
4691         btrfs_put_block_group(cache);
4692 }
4693
4694 int btrfs_free_extent(struct btrfs_trans_handle *trans,
4695                       struct btrfs_root *root,
4696                       u64 bytenr, u64 num_bytes, u64 parent,
4697                       u64 root_objectid, u64 owner, u64 offset)
4698 {
4699         int ret;
4700
4701         /*
4702          * tree log blocks never actually go into the extent allocation
4703          * tree, just update pinning info and exit early.
4704          */
4705         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
4706                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
4707                 /* unlocks the pinned mutex */
4708                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
4709                 ret = 0;
4710         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
4711                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
4712                                         parent, root_objectid, (int)owner,
4713                                         BTRFS_DROP_DELAYED_REF, NULL);
4714                 BUG_ON(ret);
4715         } else {
4716                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
4717                                         parent, root_objectid, owner,
4718                                         offset, BTRFS_DROP_DELAYED_REF, NULL);
4719                 BUG_ON(ret);
4720         }
4721         return ret;
4722 }
4723
4724 static u64 stripe_align(struct btrfs_root *root, u64 val)
4725 {
4726         u64 mask = ((u64)root->stripesize - 1);
4727         u64 ret = (val + mask) & ~mask;
4728         return ret;
4729 }
4730
4731 /*
4732  * when we wait for progress in the block group caching, its because
4733  * our allocation attempt failed at least once.  So, we must sleep
4734  * and let some progress happen before we try again.
4735  *
4736  * This function will sleep at least once waiting for new free space to
4737  * show up, and then it will check the block group free space numbers
4738  * for our min num_bytes.  Another option is to have it go ahead
4739  * and look in the rbtree for a free extent of a given size, but this
4740  * is a good start.
4741  */
4742 static noinline int
4743 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
4744                                 u64 num_bytes)
4745 {
4746         struct btrfs_caching_control *caching_ctl;
4747         DEFINE_WAIT(wait);
4748
4749         caching_ctl = get_caching_control(cache);
4750         if (!caching_ctl)
4751                 return 0;
4752
4753         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
4754                    (cache->free_space_ctl->free_space >= num_bytes));
4755
4756         put_caching_control(caching_ctl);
4757         return 0;
4758 }
4759
4760 static noinline int
4761 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
4762 {
4763         struct btrfs_caching_control *caching_ctl;
4764         DEFINE_WAIT(wait);
4765
4766         caching_ctl = get_caching_control(cache);
4767         if (!caching_ctl)
4768                 return 0;
4769
4770         wait_event(caching_ctl->wait, block_group_cache_done(cache));
4771
4772         put_caching_control(caching_ctl);
4773         return 0;
4774 }
4775
4776 static int get_block_group_index(struct btrfs_block_group_cache *cache)
4777 {
4778         int index;
4779         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
4780                 index = 0;
4781         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
4782                 index = 1;
4783         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
4784                 index = 2;
4785         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
4786                 index = 3;
4787         else
4788                 index = 4;
4789         return index;
4790 }
4791
4792 enum btrfs_loop_type {
4793         LOOP_FIND_IDEAL = 0,
4794         LOOP_CACHING_NOWAIT = 1,
4795         LOOP_CACHING_WAIT = 2,
4796         LOOP_ALLOC_CHUNK = 3,
4797         LOOP_NO_EMPTY_SIZE = 4,
4798 };
4799
4800 /*
4801  * walks the btree of allocated extents and find a hole of a given size.
4802  * The key ins is changed to record the hole:
4803  * ins->objectid == block start
4804  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4805  * ins->offset == number of blocks
4806  * Any available blocks before search_start are skipped.
4807  */
4808 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
4809                                      struct btrfs_root *orig_root,
4810                                      u64 num_bytes, u64 empty_size,
4811                                      u64 search_start, u64 search_end,
4812                                      u64 hint_byte, struct btrfs_key *ins,
4813                                      int data)
4814 {
4815         int ret = 0;
4816         struct btrfs_root *root = orig_root->fs_info->extent_root;
4817         struct btrfs_free_cluster *last_ptr = NULL;
4818         struct btrfs_block_group_cache *block_group = NULL;
4819         int empty_cluster = 2 * 1024 * 1024;
4820         int allowed_chunk_alloc = 0;
4821         int done_chunk_alloc = 0;
4822         struct btrfs_space_info *space_info;
4823         int last_ptr_loop = 0;
4824         int loop = 0;
4825         int index = 0;
4826         bool found_uncached_bg = false;
4827         bool failed_cluster_refill = false;
4828         bool failed_alloc = false;
4829         bool use_cluster = true;
4830         u64 ideal_cache_percent = 0;
4831         u64 ideal_cache_offset = 0;
4832
4833         WARN_ON(num_bytes < root->sectorsize);
4834         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
4835         ins->objectid = 0;
4836         ins->offset = 0;
4837
4838         space_info = __find_space_info(root->fs_info, data);
4839         if (!space_info) {
4840                 printk(KERN_ERR "No space info for %d\n", data);
4841                 return -ENOSPC;
4842         }
4843
4844         /*
4845          * If the space info is for both data and metadata it means we have a
4846          * small filesystem and we can't use the clustering stuff.
4847          */
4848         if (btrfs_mixed_space_info(space_info))
4849                 use_cluster = false;
4850
4851         if (orig_root->ref_cows || empty_size)
4852                 allowed_chunk_alloc = 1;
4853
4854         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
4855                 last_ptr = &root->fs_info->meta_alloc_cluster;
4856                 if (!btrfs_test_opt(root, SSD))
4857                         empty_cluster = 64 * 1024;
4858         }
4859
4860         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
4861             btrfs_test_opt(root, SSD)) {
4862                 last_ptr = &root->fs_info->data_alloc_cluster;
4863         }
4864
4865         if (last_ptr) {
4866                 spin_lock(&last_ptr->lock);
4867                 if (last_ptr->block_group)
4868                         hint_byte = last_ptr->window_start;
4869                 spin_unlock(&last_ptr->lock);
4870         }
4871
4872         search_start = max(search_start, first_logical_byte(root, 0));
4873         search_start = max(search_start, hint_byte);
4874
4875         if (!last_ptr)
4876                 empty_cluster = 0;
4877
4878         if (search_start == hint_byte) {
4879 ideal_cache:
4880                 block_group = btrfs_lookup_block_group(root->fs_info,
4881                                                        search_start);
4882                 /*
4883                  * we don't want to use the block group if it doesn't match our
4884                  * allocation bits, or if its not cached.
4885                  *
4886                  * However if we are re-searching with an ideal block group
4887                  * picked out then we don't care that the block group is cached.
4888                  */
4889                 if (block_group && block_group_bits(block_group, data) &&
4890                     (block_group->cached != BTRFS_CACHE_NO ||
4891                      search_start == ideal_cache_offset)) {
4892                         down_read(&space_info->groups_sem);
4893                         if (list_empty(&block_group->list) ||
4894                             block_group->ro) {
4895                                 /*
4896                                  * someone is removing this block group,
4897                                  * we can't jump into the have_block_group
4898                                  * target because our list pointers are not
4899                                  * valid
4900                                  */
4901                                 btrfs_put_block_group(block_group);
4902                                 up_read(&space_info->groups_sem);
4903                         } else {
4904                                 index = get_block_group_index(block_group);
4905                                 goto have_block_group;
4906                         }
4907                 } else if (block_group) {
4908                         btrfs_put_block_group(block_group);
4909                 }
4910         }
4911 search:
4912         down_read(&space_info->groups_sem);
4913         list_for_each_entry(block_group, &space_info->block_groups[index],
4914                             list) {
4915                 u64 offset;
4916                 int cached;
4917
4918                 btrfs_get_block_group(block_group);
4919                 search_start = block_group->key.objectid;
4920
4921                 /*
4922                  * this can happen if we end up cycling through all the
4923                  * raid types, but we want to make sure we only allocate
4924                  * for the proper type.
4925                  */
4926                 if (!block_group_bits(block_group, data)) {
4927                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
4928                                 BTRFS_BLOCK_GROUP_RAID1 |
4929                                 BTRFS_BLOCK_GROUP_RAID10;
4930
4931                         /*
4932                          * if they asked for extra copies and this block group
4933                          * doesn't provide them, bail.  This does allow us to
4934                          * fill raid0 from raid1.
4935                          */
4936                         if ((data & extra) && !(block_group->flags & extra))
4937                                 goto loop;
4938                 }
4939
4940 have_block_group:
4941                 if (unlikely(block_group->cached == BTRFS_CACHE_NO)) {
4942                         u64 free_percent;
4943
4944                         ret = cache_block_group(block_group, trans,
4945                                                 orig_root, 1);
4946                         if (block_group->cached == BTRFS_CACHE_FINISHED)
4947                                 goto have_block_group;
4948
4949                         free_percent = btrfs_block_group_used(&block_group->item);
4950                         free_percent *= 100;
4951                         free_percent = div64_u64(free_percent,
4952                                                  block_group->key.offset);
4953                         free_percent = 100 - free_percent;
4954                         if (free_percent > ideal_cache_percent &&
4955                             likely(!block_group->ro)) {
4956                                 ideal_cache_offset = block_group->key.objectid;
4957                                 ideal_cache_percent = free_percent;
4958                         }
4959
4960                         /*
4961                          * We only want to start kthread caching if we are at
4962                          * the point where we will wait for caching to make
4963                          * progress, or if our ideal search is over and we've
4964                          * found somebody to start caching.
4965                          */
4966                         if (loop > LOOP_CACHING_NOWAIT ||
4967                             (loop > LOOP_FIND_IDEAL &&
4968                              atomic_read(&space_info->caching_threads) < 2)) {
4969                                 ret = cache_block_group(block_group, trans,
4970                                                         orig_root, 0);
4971                                 BUG_ON(ret);
4972                         }
4973                         found_uncached_bg = true;
4974
4975                         /*
4976                          * If loop is set for cached only, try the next block
4977                          * group.
4978                          */
4979                         if (loop == LOOP_FIND_IDEAL)
4980                                 goto loop;
4981                 }
4982
4983                 cached = block_group_cache_done(block_group);
4984                 if (unlikely(!cached))
4985                         found_uncached_bg = true;
4986
4987                 if (unlikely(block_group->ro))
4988                         goto loop;
4989
4990                 /*
4991                  * Ok we want to try and use the cluster allocator, so lets look
4992                  * there, unless we are on LOOP_NO_EMPTY_SIZE, since we will
4993                  * have tried the cluster allocator plenty of times at this
4994                  * point and not have found anything, so we are likely way too
4995                  * fragmented for the clustering stuff to find anything, so lets
4996                  * just skip it and let the allocator find whatever block it can
4997                  * find
4998                  */
4999                 if (last_ptr && loop < LOOP_NO_EMPTY_SIZE) {
5000                         /*
5001                          * the refill lock keeps out other
5002                          * people trying to start a new cluster
5003                          */
5004                         spin_lock(&last_ptr->refill_lock);
5005                         if (last_ptr->block_group &&
5006                             (last_ptr->block_group->ro ||
5007                             !block_group_bits(last_ptr->block_group, data))) {
5008                                 offset = 0;
5009                                 goto refill_cluster;
5010                         }
5011
5012                         offset = btrfs_alloc_from_cluster(block_group, last_ptr,
5013                                                  num_bytes, search_start);
5014                         if (offset) {
5015                                 /* we have a block, we're done */
5016                                 spin_unlock(&last_ptr->refill_lock);
5017                                 goto checks;
5018                         }
5019
5020                         spin_lock(&last_ptr->lock);
5021                         /*
5022                          * whoops, this cluster doesn't actually point to
5023                          * this block group.  Get a ref on the block
5024                          * group is does point to and try again
5025                          */
5026                         if (!last_ptr_loop && last_ptr->block_group &&
5027                             last_ptr->block_group != block_group) {
5028
5029                                 btrfs_put_block_group(block_group);
5030                                 block_group = last_ptr->block_group;
5031                                 btrfs_get_block_group(block_group);
5032                                 spin_unlock(&last_ptr->lock);
5033                                 spin_unlock(&last_ptr->refill_lock);
5034
5035                                 last_ptr_loop = 1;
5036                                 search_start = block_group->key.objectid;
5037                                 /*
5038                                  * we know this block group is properly
5039                                  * in the list because
5040                                  * btrfs_remove_block_group, drops the
5041                                  * cluster before it removes the block
5042                                  * group from the list
5043                                  */
5044                                 goto have_block_group;
5045                         }
5046                         spin_unlock(&last_ptr->lock);
5047 refill_cluster:
5048                         /*
5049                          * this cluster didn't work out, free it and
5050                          * start over
5051                          */
5052                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5053
5054                         last_ptr_loop = 0;
5055
5056                         /* allocate a cluster in this block group */
5057                         ret = btrfs_find_space_cluster(trans, root,
5058                                                block_group, last_ptr,
5059                                                offset, num_bytes,
5060                                                empty_cluster + empty_size);
5061                         if (ret == 0) {
5062                                 /*
5063                                  * now pull our allocation out of this
5064                                  * cluster
5065                                  */
5066                                 offset = btrfs_alloc_from_cluster(block_group,
5067                                                   last_ptr, num_bytes,
5068                                                   search_start);
5069                                 if (offset) {
5070                                         /* we found one, proceed */
5071                                         spin_unlock(&last_ptr->refill_lock);
5072                                         goto checks;
5073                                 }
5074                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5075                                    && !failed_cluster_refill) {
5076                                 spin_unlock(&last_ptr->refill_lock);
5077
5078                                 failed_cluster_refill = true;
5079                                 wait_block_group_cache_progress(block_group,
5080                                        num_bytes + empty_cluster + empty_size);
5081                                 goto have_block_group;
5082                         }
5083
5084                         /*
5085                          * at this point we either didn't find a cluster
5086                          * or we weren't able to allocate a block from our
5087                          * cluster.  Free the cluster we've been trying
5088                          * to use, and go to the next block group
5089                          */
5090                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5091                         spin_unlock(&last_ptr->refill_lock);
5092                         goto loop;
5093                 }
5094
5095                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5096                                                     num_bytes, empty_size);
5097                 /*
5098                  * If we didn't find a chunk, and we haven't failed on this
5099                  * block group before, and this block group is in the middle of
5100                  * caching and we are ok with waiting, then go ahead and wait
5101                  * for progress to be made, and set failed_alloc to true.
5102                  *
5103                  * If failed_alloc is true then we've already waited on this
5104                  * block group once and should move on to the next block group.
5105                  */
5106                 if (!offset && !failed_alloc && !cached &&
5107                     loop > LOOP_CACHING_NOWAIT) {
5108                         wait_block_group_cache_progress(block_group,
5109                                                 num_bytes + empty_size);
5110                         failed_alloc = true;
5111                         goto have_block_group;
5112                 } else if (!offset) {
5113                         goto loop;
5114                 }
5115 checks:
5116                 search_start = stripe_align(root, offset);
5117                 /* move on to the next group */
5118                 if (search_start + num_bytes >= search_end) {
5119                         btrfs_add_free_space(block_group, offset, num_bytes);
5120                         goto loop;
5121                 }
5122
5123                 /* move on to the next group */
5124                 if (search_start + num_bytes >
5125                     block_group->key.objectid + block_group->key.offset) {
5126                         btrfs_add_free_space(block_group, offset, num_bytes);
5127                         goto loop;
5128                 }
5129
5130                 ins->objectid = search_start;
5131                 ins->offset = num_bytes;
5132
5133                 if (offset < search_start)
5134                         btrfs_add_free_space(block_group, offset,
5135                                              search_start - offset);
5136                 BUG_ON(offset > search_start);
5137
5138                 ret = btrfs_update_reserved_bytes(block_group, num_bytes, 1,
5139                                             (data & BTRFS_BLOCK_GROUP_DATA));
5140                 if (ret == -EAGAIN) {
5141                         btrfs_add_free_space(block_group, offset, num_bytes);
5142                         goto loop;
5143                 }
5144
5145                 /* we are all good, lets return */
5146                 ins->objectid = search_start;
5147                 ins->offset = num_bytes;
5148
5149                 if (offset < search_start)
5150                         btrfs_add_free_space(block_group, offset,
5151                                              search_start - offset);
5152                 BUG_ON(offset > search_start);
5153                 break;
5154 loop:
5155                 failed_cluster_refill = false;
5156                 failed_alloc = false;
5157                 BUG_ON(index != get_block_group_index(block_group));
5158                 btrfs_put_block_group(block_group);
5159         }
5160         up_read(&space_info->groups_sem);
5161
5162         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5163                 goto search;
5164
5165         /* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
5166          *                      for them to make caching progress.  Also
5167          *                      determine the best possible bg to cache
5168          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5169          *                      caching kthreads as we move along
5170          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5171          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5172          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5173          *                      again
5174          */
5175         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE &&
5176             (found_uncached_bg || empty_size || empty_cluster ||
5177              allowed_chunk_alloc)) {
5178                 index = 0;
5179                 if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
5180                         found_uncached_bg = false;
5181                         loop++;
5182                         if (!ideal_cache_percent &&
5183                             atomic_read(&space_info->caching_threads))
5184                                 goto search;
5185
5186                         /*
5187                          * 1 of the following 2 things have happened so far
5188                          *
5189                          * 1) We found an ideal block group for caching that
5190                          * is mostly full and will cache quickly, so we might
5191                          * as well wait for it.
5192                          *
5193                          * 2) We searched for cached only and we didn't find
5194                          * anything, and we didn't start any caching kthreads
5195                          * either, so chances are we will loop through and
5196                          * start a couple caching kthreads, and then come back
5197                          * around and just wait for them.  This will be slower
5198                          * because we will have 2 caching kthreads reading at
5199                          * the same time when we could have just started one
5200                          * and waited for it to get far enough to give us an
5201                          * allocation, so go ahead and go to the wait caching
5202                          * loop.
5203                          */
5204                         loop = LOOP_CACHING_WAIT;
5205                         search_start = ideal_cache_offset;
5206                         ideal_cache_percent = 0;
5207                         goto ideal_cache;
5208                 } else if (loop == LOOP_FIND_IDEAL) {
5209                         /*
5210                          * Didn't find a uncached bg, wait on anything we find
5211                          * next.
5212                          */
5213                         loop = LOOP_CACHING_WAIT;
5214                         goto search;
5215                 }
5216
5217                 if (loop < LOOP_CACHING_WAIT) {
5218                         loop++;
5219                         goto search;
5220                 }
5221
5222                 if (loop == LOOP_ALLOC_CHUNK) {
5223                         empty_size = 0;
5224                         empty_cluster = 0;
5225                 }
5226
5227                 if (allowed_chunk_alloc) {
5228                         ret = do_chunk_alloc(trans, root, num_bytes +
5229                                              2 * 1024 * 1024, data,
5230                                              CHUNK_ALLOC_LIMITED);
5231                         allowed_chunk_alloc = 0;
5232                         done_chunk_alloc = 1;
5233                 } else if (!done_chunk_alloc &&
5234                            space_info->force_alloc == CHUNK_ALLOC_NO_FORCE) {
5235                         space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5236                 }
5237
5238                 if (loop < LOOP_NO_EMPTY_SIZE) {
5239                         loop++;
5240                         goto search;
5241                 }
5242                 ret = -ENOSPC;
5243         } else if (!ins->objectid) {
5244                 ret = -ENOSPC;
5245         }
5246
5247         /* we found what we needed */
5248         if (ins->objectid) {
5249                 if (!(data & BTRFS_BLOCK_GROUP_DATA))
5250                         trans->block_group = block_group->key.objectid;
5251
5252                 btrfs_put_block_group(block_group);
5253                 ret = 0;
5254         }
5255
5256         return ret;
5257 }
5258
5259 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5260                             int dump_block_groups)
5261 {
5262         struct btrfs_block_group_cache *cache;
5263         int index = 0;
5264
5265         spin_lock(&info->lock);
5266         printk(KERN_INFO "space_info has %llu free, is %sfull\n",
5267                (unsigned long long)(info->total_bytes - info->bytes_used -
5268                                     info->bytes_pinned - info->bytes_reserved -
5269                                     info->bytes_readonly),
5270                (info->full) ? "" : "not ");
5271         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5272                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5273                (unsigned long long)info->total_bytes,
5274                (unsigned long long)info->bytes_used,
5275                (unsigned long long)info->bytes_pinned,
5276                (unsigned long long)info->bytes_reserved,
5277                (unsigned long long)info->bytes_may_use,
5278                (unsigned long long)info->bytes_readonly);
5279         spin_unlock(&info->lock);
5280
5281         if (!dump_block_groups)
5282                 return;
5283
5284         down_read(&info->groups_sem);
5285 again:
5286         list_for_each_entry(cache, &info->block_groups[index], list) {
5287                 spin_lock(&cache->lock);
5288                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5289                        "%llu pinned %llu reserved\n",
5290                        (unsigned long long)cache->key.objectid,
5291                        (unsigned long long)cache->key.offset,
5292                        (unsigned long long)btrfs_block_group_used(&cache->item),
5293                        (unsigned long long)cache->pinned,
5294                        (unsigned long long)cache->reserved);
5295                 btrfs_dump_free_space(cache, bytes);
5296                 spin_unlock(&cache->lock);
5297         }
5298         if (++index < BTRFS_NR_RAID_TYPES)
5299                 goto again;
5300         up_read(&info->groups_sem);
5301 }
5302
5303 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5304                          struct btrfs_root *root,
5305                          u64 num_bytes, u64 min_alloc_size,
5306                          u64 empty_size, u64 hint_byte,
5307                          u64 search_end, struct btrfs_key *ins,
5308                          u64 data)
5309 {
5310         int ret;
5311         u64 search_start = 0;
5312
5313         data = btrfs_get_alloc_profile(root, data);
5314 again:
5315         /*
5316          * the only place that sets empty_size is btrfs_realloc_node, which
5317          * is not called recursively on allocations
5318          */
5319         if (empty_size || root->ref_cows)
5320                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5321                                      num_bytes + 2 * 1024 * 1024, data,
5322                                      CHUNK_ALLOC_NO_FORCE);
5323
5324         WARN_ON(num_bytes < root->sectorsize);
5325         ret = find_free_extent(trans, root, num_bytes, empty_size,
5326                                search_start, search_end, hint_byte,
5327                                ins, data);
5328
5329         if (ret == -ENOSPC && num_bytes > min_alloc_size) {
5330                 num_bytes = num_bytes >> 1;
5331                 num_bytes = num_bytes & ~(root->sectorsize - 1);
5332                 num_bytes = max(num_bytes, min_alloc_size);
5333                 do_chunk_alloc(trans, root->fs_info->extent_root,
5334                                num_bytes, data, CHUNK_ALLOC_FORCE);
5335                 goto again;
5336         }
5337         if (ret == -ENOSPC && btrfs_test_opt(root, ENOSPC_DEBUG)) {
5338                 struct btrfs_space_info *sinfo;
5339
5340                 sinfo = __find_space_info(root->fs_info, data);
5341                 printk(KERN_ERR "btrfs allocation failed flags %llu, "
5342                        "wanted %llu\n", (unsigned long long)data,
5343                        (unsigned long long)num_bytes);
5344                 dump_space_info(sinfo, num_bytes, 1);
5345         }
5346
5347         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
5348
5349         return ret;
5350 }
5351
5352 int btrfs_free_reserved_extent(struct btrfs_root *root, u64 start, u64 len)
5353 {
5354         struct btrfs_block_group_cache *cache;
5355         int ret = 0;
5356
5357         cache = btrfs_lookup_block_group(root->fs_info, start);
5358         if (!cache) {
5359                 printk(KERN_ERR "Unable to find block group for %llu\n",
5360                        (unsigned long long)start);
5361                 return -ENOSPC;
5362         }
5363
5364         if (btrfs_test_opt(root, DISCARD))
5365                 ret = btrfs_discard_extent(root, start, len, NULL);
5366
5367         btrfs_add_free_space(cache, start, len);
5368         btrfs_update_reserved_bytes(cache, len, 0, 1);
5369         btrfs_put_block_group(cache);
5370
5371         trace_btrfs_reserved_extent_free(root, start, len);
5372
5373         return ret;
5374 }
5375
5376 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5377                                       struct btrfs_root *root,
5378                                       u64 parent, u64 root_objectid,
5379                                       u64 flags, u64 owner, u64 offset,
5380                                       struct btrfs_key *ins, int ref_mod)
5381 {
5382         int ret;
5383         struct btrfs_fs_info *fs_info = root->fs_info;
5384         struct btrfs_extent_item *extent_item;
5385         struct btrfs_extent_inline_ref *iref;
5386         struct btrfs_path *path;
5387         struct extent_buffer *leaf;
5388         int type;
5389         u32 size;
5390
5391         if (parent > 0)
5392                 type = BTRFS_SHARED_DATA_REF_KEY;
5393         else
5394                 type = BTRFS_EXTENT_DATA_REF_KEY;
5395
5396         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5397
5398         path = btrfs_alloc_path();
5399         if (!path)
5400                 return -ENOMEM;
5401
5402         path->leave_spinning = 1;
5403         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5404                                       ins, size);
5405         BUG_ON(ret);
5406
5407         leaf = path->nodes[0];
5408         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5409                                      struct btrfs_extent_item);
5410         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
5411         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5412         btrfs_set_extent_flags(leaf, extent_item,
5413                                flags | BTRFS_EXTENT_FLAG_DATA);
5414
5415         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
5416         btrfs_set_extent_inline_ref_type(leaf, iref, type);
5417         if (parent > 0) {
5418                 struct btrfs_shared_data_ref *ref;
5419                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
5420                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5421                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
5422         } else {
5423                 struct btrfs_extent_data_ref *ref;
5424                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
5425                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
5426                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
5427                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
5428                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
5429         }
5430
5431         btrfs_mark_buffer_dirty(path->nodes[0]);
5432         btrfs_free_path(path);
5433
5434         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5435         if (ret) {
5436                 printk(KERN_ERR "btrfs update block group failed for %llu "
5437                        "%llu\n", (unsigned long long)ins->objectid,
5438                        (unsigned long long)ins->offset);
5439                 BUG();
5440         }
5441         return ret;
5442 }
5443
5444 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
5445                                      struct btrfs_root *root,
5446                                      u64 parent, u64 root_objectid,
5447                                      u64 flags, struct btrfs_disk_key *key,
5448                                      int level, struct btrfs_key *ins)
5449 {
5450         int ret;
5451         struct btrfs_fs_info *fs_info = root->fs_info;
5452         struct btrfs_extent_item *extent_item;
5453         struct btrfs_tree_block_info *block_info;
5454         struct btrfs_extent_inline_ref *iref;
5455         struct btrfs_path *path;
5456         struct extent_buffer *leaf;
5457         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
5458
5459         path = btrfs_alloc_path();
5460         BUG_ON(!path);
5461
5462         path->leave_spinning = 1;
5463         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5464                                       ins, size);
5465         BUG_ON(ret);
5466
5467         leaf = path->nodes[0];
5468         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5469                                      struct btrfs_extent_item);
5470         btrfs_set_extent_refs(leaf, extent_item, 1);
5471         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5472         btrfs_set_extent_flags(leaf, extent_item,
5473                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
5474         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
5475
5476         btrfs_set_tree_block_key(leaf, block_info, key);
5477         btrfs_set_tree_block_level(leaf, block_info, level);
5478
5479         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
5480         if (parent > 0) {
5481                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
5482                 btrfs_set_extent_inline_ref_type(leaf, iref,
5483                                                  BTRFS_SHARED_BLOCK_REF_KEY);
5484                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5485         } else {
5486                 btrfs_set_extent_inline_ref_type(leaf, iref,
5487                                                  BTRFS_TREE_BLOCK_REF_KEY);
5488                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
5489         }
5490
5491         btrfs_mark_buffer_dirty(leaf);
5492         btrfs_free_path(path);
5493
5494         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5495         if (ret) {
5496                 printk(KERN_ERR "btrfs update block group failed for %llu "
5497                        "%llu\n", (unsigned long long)ins->objectid,
5498                        (unsigned long long)ins->offset);
5499                 BUG();
5500         }
5501         return ret;
5502 }
5503
5504 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5505                                      struct btrfs_root *root,
5506                                      u64 root_objectid, u64 owner,
5507                                      u64 offset, struct btrfs_key *ins)
5508 {
5509         int ret;
5510
5511         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
5512
5513         ret = btrfs_add_delayed_data_ref(trans, ins->objectid, ins->offset,
5514                                          0, root_objectid, owner, offset,
5515                                          BTRFS_ADD_DELAYED_EXTENT, NULL);
5516         return ret;
5517 }
5518
5519 /*
5520  * this is used by the tree logging recovery code.  It records that
5521  * an extent has been allocated and makes sure to clear the free
5522  * space cache bits as well
5523  */
5524 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
5525                                    struct btrfs_root *root,
5526                                    u64 root_objectid, u64 owner, u64 offset,
5527                                    struct btrfs_key *ins)
5528 {
5529         int ret;
5530         struct btrfs_block_group_cache *block_group;
5531         struct btrfs_caching_control *caching_ctl;
5532         u64 start = ins->objectid;
5533         u64 num_bytes = ins->offset;
5534
5535         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
5536         cache_block_group(block_group, trans, NULL, 0);
5537         caching_ctl = get_caching_control(block_group);
5538
5539         if (!caching_ctl) {
5540                 BUG_ON(!block_group_cache_done(block_group));
5541                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5542                 BUG_ON(ret);
5543         } else {
5544                 mutex_lock(&caching_ctl->mutex);
5545
5546                 if (start >= caching_ctl->progress) {
5547                         ret = add_excluded_extent(root, start, num_bytes);
5548                         BUG_ON(ret);
5549                 } else if (start + num_bytes <= caching_ctl->progress) {
5550                         ret = btrfs_remove_free_space(block_group,
5551                                                       start, num_bytes);
5552                         BUG_ON(ret);
5553                 } else {
5554                         num_bytes = caching_ctl->progress - start;
5555                         ret = btrfs_remove_free_space(block_group,
5556                                                       start, num_bytes);
5557                         BUG_ON(ret);
5558
5559                         start = caching_ctl->progress;
5560                         num_bytes = ins->objectid + ins->offset -
5561                                     caching_ctl->progress;
5562                         ret = add_excluded_extent(root, start, num_bytes);
5563                         BUG_ON(ret);
5564                 }
5565
5566                 mutex_unlock(&caching_ctl->mutex);
5567                 put_caching_control(caching_ctl);
5568         }
5569
5570         ret = btrfs_update_reserved_bytes(block_group, ins->offset, 1, 1);
5571         BUG_ON(ret);
5572         btrfs_put_block_group(block_group);
5573         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
5574                                          0, owner, offset, ins, 1);
5575         return ret;
5576 }
5577
5578 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
5579                                             struct btrfs_root *root,
5580                                             u64 bytenr, u32 blocksize,
5581                                             int level)
5582 {
5583         struct extent_buffer *buf;
5584
5585         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
5586         if (!buf)
5587                 return ERR_PTR(-ENOMEM);
5588         btrfs_set_header_generation(buf, trans->transid);
5589         btrfs_set_buffer_lockdep_class(buf, level);
5590         btrfs_tree_lock(buf);
5591         clean_tree_block(trans, root, buf);
5592
5593         btrfs_set_lock_blocking(buf);
5594         btrfs_set_buffer_uptodate(buf);
5595
5596         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
5597                 /*
5598                  * we allow two log transactions at a time, use different
5599                  * EXENT bit to differentiate dirty pages.
5600                  */
5601                 if (root->log_transid % 2 == 0)
5602                         set_extent_dirty(&root->dirty_log_pages, buf->start,
5603                                         buf->start + buf->len - 1, GFP_NOFS);
5604                 else
5605                         set_extent_new(&root->dirty_log_pages, buf->start,
5606                                         buf->start + buf->len - 1, GFP_NOFS);
5607         } else {
5608                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
5609                          buf->start + buf->len - 1, GFP_NOFS);
5610         }
5611         trans->blocks_used++;
5612         /* this returns a buffer locked for blocking */
5613         return buf;
5614 }
5615
5616 static struct btrfs_block_rsv *
5617 use_block_rsv(struct btrfs_trans_handle *trans,
5618               struct btrfs_root *root, u32 blocksize)
5619 {
5620         struct btrfs_block_rsv *block_rsv;
5621         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5622         int ret;
5623
5624         block_rsv = get_block_rsv(trans, root);
5625
5626         if (block_rsv->size == 0) {
5627                 ret = reserve_metadata_bytes(trans, root, block_rsv,
5628                                              blocksize, 0);
5629                 /*
5630                  * If we couldn't reserve metadata bytes try and use some from
5631                  * the global reserve.
5632                  */
5633                 if (ret && block_rsv != global_rsv) {
5634                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5635                         if (!ret)
5636                                 return global_rsv;
5637                         return ERR_PTR(ret);
5638                 } else if (ret) {
5639                         return ERR_PTR(ret);
5640                 }
5641                 return block_rsv;
5642         }
5643
5644         ret = block_rsv_use_bytes(block_rsv, blocksize);
5645         if (!ret)
5646                 return block_rsv;
5647         if (ret) {
5648                 WARN_ON(1);
5649                 ret = reserve_metadata_bytes(trans, root, block_rsv, blocksize,
5650                                              0);
5651                 if (!ret) {
5652                         spin_lock(&block_rsv->lock);
5653                         block_rsv->size += blocksize;
5654                         spin_unlock(&block_rsv->lock);
5655                         return block_rsv;
5656                 } else if (ret && block_rsv != global_rsv) {
5657                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5658                         if (!ret)
5659                                 return global_rsv;
5660                 }
5661         }
5662
5663         return ERR_PTR(-ENOSPC);
5664 }
5665
5666 static void unuse_block_rsv(struct btrfs_block_rsv *block_rsv, u32 blocksize)
5667 {
5668         block_rsv_add_bytes(block_rsv, blocksize, 0);
5669         block_rsv_release_bytes(block_rsv, NULL, 0);
5670 }
5671
5672 /*
5673  * finds a free extent and does all the dirty work required for allocation
5674  * returns the key for the extent through ins, and a tree buffer for
5675  * the first block of the extent through buf.
5676  *
5677  * returns the tree buffer or NULL.
5678  */
5679 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
5680                                         struct btrfs_root *root, u32 blocksize,
5681                                         u64 parent, u64 root_objectid,
5682                                         struct btrfs_disk_key *key, int level,
5683                                         u64 hint, u64 empty_size)
5684 {
5685         struct btrfs_key ins;
5686         struct btrfs_block_rsv *block_rsv;
5687         struct extent_buffer *buf;
5688         u64 flags = 0;
5689         int ret;
5690
5691
5692         block_rsv = use_block_rsv(trans, root, blocksize);
5693         if (IS_ERR(block_rsv))
5694                 return ERR_CAST(block_rsv);
5695
5696         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
5697                                    empty_size, hint, (u64)-1, &ins, 0);
5698         if (ret) {
5699                 unuse_block_rsv(block_rsv, blocksize);
5700                 return ERR_PTR(ret);
5701         }
5702
5703         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
5704                                     blocksize, level);
5705         BUG_ON(IS_ERR(buf));
5706
5707         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
5708                 if (parent == 0)
5709                         parent = ins.objectid;
5710                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
5711         } else
5712                 BUG_ON(parent > 0);
5713
5714         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
5715                 struct btrfs_delayed_extent_op *extent_op;
5716                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
5717                 BUG_ON(!extent_op);
5718                 if (key)
5719                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
5720                 else
5721                         memset(&extent_op->key, 0, sizeof(extent_op->key));
5722                 extent_op->flags_to_set = flags;
5723                 extent_op->update_key = 1;
5724                 extent_op->update_flags = 1;
5725                 extent_op->is_data = 0;
5726
5727                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
5728                                         ins.offset, parent, root_objectid,
5729                                         level, BTRFS_ADD_DELAYED_EXTENT,
5730                                         extent_op);
5731                 BUG_ON(ret);
5732         }
5733         return buf;
5734 }
5735
5736 struct walk_control {
5737         u64 refs[BTRFS_MAX_LEVEL];
5738         u64 flags[BTRFS_MAX_LEVEL];
5739         struct btrfs_key update_progress;
5740         int stage;
5741         int level;
5742         int shared_level;
5743         int update_ref;
5744         int keep_locks;
5745         int reada_slot;
5746         int reada_count;
5747 };
5748
5749 #define DROP_REFERENCE  1
5750 #define UPDATE_BACKREF  2
5751
5752 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
5753                                      struct btrfs_root *root,
5754                                      struct walk_control *wc,
5755                                      struct btrfs_path *path)
5756 {
5757         u64 bytenr;
5758         u64 generation;
5759         u64 refs;
5760         u64 flags;
5761         u32 nritems;
5762         u32 blocksize;
5763         struct btrfs_key key;
5764         struct extent_buffer *eb;
5765         int ret;
5766         int slot;
5767         int nread = 0;
5768
5769         if (path->slots[wc->level] < wc->reada_slot) {
5770                 wc->reada_count = wc->reada_count * 2 / 3;
5771                 wc->reada_count = max(wc->reada_count, 2);
5772         } else {
5773                 wc->reada_count = wc->reada_count * 3 / 2;
5774                 wc->reada_count = min_t(int, wc->reada_count,
5775                                         BTRFS_NODEPTRS_PER_BLOCK(root));
5776         }
5777
5778         eb = path->nodes[wc->level];
5779         nritems = btrfs_header_nritems(eb);
5780         blocksize = btrfs_level_size(root, wc->level - 1);
5781
5782         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
5783                 if (nread >= wc->reada_count)
5784                         break;
5785
5786                 cond_resched();
5787                 bytenr = btrfs_node_blockptr(eb, slot);
5788                 generation = btrfs_node_ptr_generation(eb, slot);
5789
5790                 if (slot == path->slots[wc->level])
5791                         goto reada;
5792
5793                 if (wc->stage == UPDATE_BACKREF &&
5794                     generation <= root->root_key.offset)
5795                         continue;
5796
5797                 /* We don't lock the tree block, it's OK to be racy here */
5798                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
5799                                                &refs, &flags);
5800                 BUG_ON(ret);
5801                 BUG_ON(refs == 0);
5802
5803                 if (wc->stage == DROP_REFERENCE) {
5804                         if (refs == 1)
5805                                 goto reada;
5806
5807                         if (wc->level == 1 &&
5808                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5809                                 continue;
5810                         if (!wc->update_ref ||
5811                             generation <= root->root_key.offset)
5812                                 continue;
5813                         btrfs_node_key_to_cpu(eb, &key, slot);
5814                         ret = btrfs_comp_cpu_keys(&key,
5815                                                   &wc->update_progress);
5816                         if (ret < 0)
5817                                 continue;
5818                 } else {
5819                         if (wc->level == 1 &&
5820                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5821                                 continue;
5822                 }
5823 reada:
5824                 ret = readahead_tree_block(root, bytenr, blocksize,
5825                                            generation);
5826                 if (ret)
5827                         break;
5828                 nread++;
5829         }
5830         wc->reada_slot = slot;
5831 }
5832
5833 /*
5834  * hepler to process tree block while walking down the tree.
5835  *
5836  * when wc->stage == UPDATE_BACKREF, this function updates
5837  * back refs for pointers in the block.
5838  *
5839  * NOTE: return value 1 means we should stop walking down.
5840  */
5841 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
5842                                    struct btrfs_root *root,
5843                                    struct btrfs_path *path,
5844                                    struct walk_control *wc, int lookup_info)
5845 {
5846         int level = wc->level;
5847         struct extent_buffer *eb = path->nodes[level];
5848         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5849         int ret;
5850
5851         if (wc->stage == UPDATE_BACKREF &&
5852             btrfs_header_owner(eb) != root->root_key.objectid)
5853                 return 1;
5854
5855         /*
5856          * when reference count of tree block is 1, it won't increase
5857          * again. once full backref flag is set, we never clear it.
5858          */
5859         if (lookup_info &&
5860             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
5861              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
5862                 BUG_ON(!path->locks[level]);
5863                 ret = btrfs_lookup_extent_info(trans, root,
5864                                                eb->start, eb->len,
5865                                                &wc->refs[level],
5866                                                &wc->flags[level]);
5867                 BUG_ON(ret);
5868                 BUG_ON(wc->refs[level] == 0);
5869         }
5870
5871         if (wc->stage == DROP_REFERENCE) {
5872                 if (wc->refs[level] > 1)
5873                         return 1;
5874
5875                 if (path->locks[level] && !wc->keep_locks) {
5876                         btrfs_tree_unlock(eb);
5877                         path->locks[level] = 0;
5878                 }
5879                 return 0;
5880         }
5881
5882         /* wc->stage == UPDATE_BACKREF */
5883         if (!(wc->flags[level] & flag)) {
5884                 BUG_ON(!path->locks[level]);
5885                 ret = btrfs_inc_ref(trans, root, eb, 1);
5886                 BUG_ON(ret);
5887                 ret = btrfs_dec_ref(trans, root, eb, 0);
5888                 BUG_ON(ret);
5889                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
5890                                                   eb->len, flag, 0);
5891                 BUG_ON(ret);
5892                 wc->flags[level] |= flag;
5893         }
5894
5895         /*
5896          * the block is shared by multiple trees, so it's not good to
5897          * keep the tree lock
5898          */
5899         if (path->locks[level] && level > 0) {
5900                 btrfs_tree_unlock(eb);
5901                 path->locks[level] = 0;
5902         }
5903         return 0;
5904 }
5905
5906 /*
5907  * hepler to process tree block pointer.
5908  *
5909  * when wc->stage == DROP_REFERENCE, this function checks
5910  * reference count of the block pointed to. if the block
5911  * is shared and we need update back refs for the subtree
5912  * rooted at the block, this function changes wc->stage to
5913  * UPDATE_BACKREF. if the block is shared and there is no
5914  * need to update back, this function drops the reference
5915  * to the block.
5916  *
5917  * NOTE: return value 1 means we should stop walking down.
5918  */
5919 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
5920                                  struct btrfs_root *root,
5921                                  struct btrfs_path *path,
5922                                  struct walk_control *wc, int *lookup_info)
5923 {
5924         u64 bytenr;
5925         u64 generation;
5926         u64 parent;
5927         u32 blocksize;
5928         struct btrfs_key key;
5929         struct extent_buffer *next;
5930         int level = wc->level;
5931         int reada = 0;
5932         int ret = 0;
5933
5934         generation = btrfs_node_ptr_generation(path->nodes[level],
5935                                                path->slots[level]);
5936         /*
5937          * if the lower level block was created before the snapshot
5938          * was created, we know there is no need to update back refs
5939          * for the subtree
5940          */
5941         if (wc->stage == UPDATE_BACKREF &&
5942             generation <= root->root_key.offset) {
5943                 *lookup_info = 1;
5944                 return 1;
5945         }
5946
5947         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
5948         blocksize = btrfs_level_size(root, level - 1);
5949
5950         next = btrfs_find_tree_block(root, bytenr, blocksize);
5951         if (!next) {
5952                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
5953                 if (!next)
5954                         return -ENOMEM;
5955                 reada = 1;
5956         }
5957         btrfs_tree_lock(next);
5958         btrfs_set_lock_blocking(next);
5959
5960         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
5961                                        &wc->refs[level - 1],
5962                                        &wc->flags[level - 1]);
5963         BUG_ON(ret);
5964         BUG_ON(wc->refs[level - 1] == 0);
5965         *lookup_info = 0;
5966
5967         if (wc->stage == DROP_REFERENCE) {
5968                 if (wc->refs[level - 1] > 1) {
5969                         if (level == 1 &&
5970                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5971                                 goto skip;
5972
5973                         if (!wc->update_ref ||
5974                             generation <= root->root_key.offset)
5975                                 goto skip;
5976
5977                         btrfs_node_key_to_cpu(path->nodes[level], &key,
5978                                               path->slots[level]);
5979                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
5980                         if (ret < 0)
5981                                 goto skip;
5982
5983                         wc->stage = UPDATE_BACKREF;
5984                         wc->shared_level = level - 1;
5985                 }
5986         } else {
5987                 if (level == 1 &&
5988                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5989                         goto skip;
5990         }
5991
5992         if (!btrfs_buffer_uptodate(next, generation)) {
5993                 btrfs_tree_unlock(next);
5994                 free_extent_buffer(next);
5995                 next = NULL;
5996                 *lookup_info = 1;
5997         }
5998
5999         if (!next) {
6000                 if (reada && level == 1)
6001                         reada_walk_down(trans, root, wc, path);
6002                 next = read_tree_block(root, bytenr, blocksize, generation);
6003                 if (!next)
6004                         return -EIO;
6005                 btrfs_tree_lock(next);
6006                 btrfs_set_lock_blocking(next);
6007         }
6008
6009         level--;
6010         BUG_ON(level != btrfs_header_level(next));
6011         path->nodes[level] = next;
6012         path->slots[level] = 0;
6013         path->locks[level] = 1;
6014         wc->level = level;
6015         if (wc->level == 1)
6016                 wc->reada_slot = 0;
6017         return 0;
6018 skip:
6019         wc->refs[level - 1] = 0;
6020         wc->flags[level - 1] = 0;
6021         if (wc->stage == DROP_REFERENCE) {
6022                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6023                         parent = path->nodes[level]->start;
6024                 } else {
6025                         BUG_ON(root->root_key.objectid !=
6026                                btrfs_header_owner(path->nodes[level]));
6027                         parent = 0;
6028                 }
6029
6030                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6031                                         root->root_key.objectid, level - 1, 0);
6032                 BUG_ON(ret);
6033         }
6034         btrfs_tree_unlock(next);
6035         free_extent_buffer(next);
6036         *lookup_info = 1;
6037         return 1;
6038 }
6039
6040 /*
6041  * hepler to process tree block while walking up the tree.
6042  *
6043  * when wc->stage == DROP_REFERENCE, this function drops
6044  * reference count on the block.
6045  *
6046  * when wc->stage == UPDATE_BACKREF, this function changes
6047  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6048  * to UPDATE_BACKREF previously while processing the block.
6049  *
6050  * NOTE: return value 1 means we should stop walking up.
6051  */
6052 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6053                                  struct btrfs_root *root,
6054                                  struct btrfs_path *path,
6055                                  struct walk_control *wc)
6056 {
6057         int ret;
6058         int level = wc->level;
6059         struct extent_buffer *eb = path->nodes[level];
6060         u64 parent = 0;
6061
6062         if (wc->stage == UPDATE_BACKREF) {
6063                 BUG_ON(wc->shared_level < level);
6064                 if (level < wc->shared_level)
6065                         goto out;
6066
6067                 ret = find_next_key(path, level + 1, &wc->update_progress);
6068                 if (ret > 0)
6069                         wc->update_ref = 0;
6070
6071                 wc->stage = DROP_REFERENCE;
6072                 wc->shared_level = -1;
6073                 path->slots[level] = 0;
6074
6075                 /*
6076                  * check reference count again if the block isn't locked.
6077                  * we should start walking down the tree again if reference
6078                  * count is one.
6079                  */
6080                 if (!path->locks[level]) {
6081                         BUG_ON(level == 0);
6082                         btrfs_tree_lock(eb);
6083                         btrfs_set_lock_blocking(eb);
6084                         path->locks[level] = 1;
6085
6086                         ret = btrfs_lookup_extent_info(trans, root,
6087                                                        eb->start, eb->len,
6088                                                        &wc->refs[level],
6089                                                        &wc->flags[level]);
6090                         BUG_ON(ret);
6091                         BUG_ON(wc->refs[level] == 0);
6092                         if (wc->refs[level] == 1) {
6093                                 btrfs_tree_unlock(eb);
6094                                 path->locks[level] = 0;
6095                                 return 1;
6096                         }
6097                 }
6098         }
6099
6100         /* wc->stage == DROP_REFERENCE */
6101         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6102
6103         if (wc->refs[level] == 1) {
6104                 if (level == 0) {
6105                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6106                                 ret = btrfs_dec_ref(trans, root, eb, 1);
6107                         else
6108                                 ret = btrfs_dec_ref(trans, root, eb, 0);
6109                         BUG_ON(ret);
6110                 }
6111                 /* make block locked assertion in clean_tree_block happy */
6112                 if (!path->locks[level] &&
6113                     btrfs_header_generation(eb) == trans->transid) {
6114                         btrfs_tree_lock(eb);
6115                         btrfs_set_lock_blocking(eb);
6116                         path->locks[level] = 1;
6117                 }
6118                 clean_tree_block(trans, root, eb);
6119         }
6120
6121         if (eb == root->node) {
6122                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6123                         parent = eb->start;
6124                 else
6125                         BUG_ON(root->root_key.objectid !=
6126                                btrfs_header_owner(eb));
6127         } else {
6128                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6129                         parent = path->nodes[level + 1]->start;
6130                 else
6131                         BUG_ON(root->root_key.objectid !=
6132                                btrfs_header_owner(path->nodes[level + 1]));
6133         }
6134
6135         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6136 out:
6137         wc->refs[level] = 0;
6138         wc->flags[level] = 0;
6139         return 0;
6140 }
6141
6142 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6143                                    struct btrfs_root *root,
6144                                    struct btrfs_path *path,
6145                                    struct walk_control *wc)
6146 {
6147         int level = wc->level;
6148         int lookup_info = 1;
6149         int ret;
6150
6151         while (level >= 0) {
6152                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6153                 if (ret > 0)
6154                         break;
6155
6156                 if (level == 0)
6157                         break;
6158
6159                 if (path->slots[level] >=
6160                     btrfs_header_nritems(path->nodes[level]))
6161                         break;
6162
6163                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6164                 if (ret > 0) {
6165                         path->slots[level]++;
6166                         continue;
6167                 } else if (ret < 0)
6168                         return ret;
6169                 level = wc->level;
6170         }
6171         return 0;
6172 }
6173
6174 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6175                                  struct btrfs_root *root,
6176                                  struct btrfs_path *path,
6177                                  struct walk_control *wc, int max_level)
6178 {
6179         int level = wc->level;
6180         int ret;
6181
6182         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6183         while (level < max_level && path->nodes[level]) {
6184                 wc->level = level;
6185                 if (path->slots[level] + 1 <
6186                     btrfs_header_nritems(path->nodes[level])) {
6187                         path->slots[level]++;
6188                         return 0;
6189                 } else {
6190                         ret = walk_up_proc(trans, root, path, wc);
6191                         if (ret > 0)
6192                                 return 0;
6193
6194                         if (path->locks[level]) {
6195                                 btrfs_tree_unlock(path->nodes[level]);
6196                                 path->locks[level] = 0;
6197                         }
6198                         free_extent_buffer(path->nodes[level]);
6199                         path->nodes[level] = NULL;
6200                         level++;
6201                 }
6202         }
6203         return 1;
6204 }
6205
6206 /*
6207  * drop a subvolume tree.
6208  *
6209  * this function traverses the tree freeing any blocks that only
6210  * referenced by the tree.
6211  *
6212  * when a shared tree block is found. this function decreases its
6213  * reference count by one. if update_ref is true, this function
6214  * also make sure backrefs for the shared block and all lower level
6215  * blocks are properly updated.
6216  */
6217 int btrfs_drop_snapshot(struct btrfs_root *root,
6218                         struct btrfs_block_rsv *block_rsv, int update_ref)
6219 {
6220         struct btrfs_path *path;
6221         struct btrfs_trans_handle *trans;
6222         struct btrfs_root *tree_root = root->fs_info->tree_root;
6223         struct btrfs_root_item *root_item = &root->root_item;
6224         struct walk_control *wc;
6225         struct btrfs_key key;
6226         int err = 0;
6227         int ret;
6228         int level;
6229
6230         path = btrfs_alloc_path();
6231         BUG_ON(!path);
6232
6233         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6234         BUG_ON(!wc);
6235
6236         trans = btrfs_start_transaction(tree_root, 0);
6237         BUG_ON(IS_ERR(trans));
6238
6239         if (block_rsv)
6240                 trans->block_rsv = block_rsv;
6241
6242         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6243                 level = btrfs_header_level(root->node);
6244                 path->nodes[level] = btrfs_lock_root_node(root);
6245                 btrfs_set_lock_blocking(path->nodes[level]);
6246                 path->slots[level] = 0;
6247                 path->locks[level] = 1;
6248                 memset(&wc->update_progress, 0,
6249                        sizeof(wc->update_progress));
6250         } else {
6251                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6252                 memcpy(&wc->update_progress, &key,
6253                        sizeof(wc->update_progress));
6254
6255                 level = root_item->drop_level;
6256                 BUG_ON(level == 0);
6257                 path->lowest_level = level;
6258                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6259                 path->lowest_level = 0;
6260                 if (ret < 0) {
6261                         err = ret;
6262                         goto out;
6263                 }
6264                 WARN_ON(ret > 0);
6265
6266                 /*
6267                  * unlock our path, this is safe because only this
6268                  * function is allowed to delete this snapshot
6269                  */
6270                 btrfs_unlock_up_safe(path, 0);
6271
6272                 level = btrfs_header_level(root->node);
6273                 while (1) {
6274                         btrfs_tree_lock(path->nodes[level]);
6275                         btrfs_set_lock_blocking(path->nodes[level]);
6276
6277                         ret = btrfs_lookup_extent_info(trans, root,
6278                                                 path->nodes[level]->start,
6279                                                 path->nodes[level]->len,
6280                                                 &wc->refs[level],
6281                                                 &wc->flags[level]);
6282                         BUG_ON(ret);
6283                         BUG_ON(wc->refs[level] == 0);
6284
6285                         if (level == root_item->drop_level)
6286                                 break;
6287
6288                         btrfs_tree_unlock(path->nodes[level]);
6289                         WARN_ON(wc->refs[level] != 1);
6290                         level--;
6291                 }
6292         }
6293
6294         wc->level = level;
6295         wc->shared_level = -1;
6296         wc->stage = DROP_REFERENCE;
6297         wc->update_ref = update_ref;
6298         wc->keep_locks = 0;
6299         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6300
6301         while (1) {
6302                 ret = walk_down_tree(trans, root, path, wc);
6303                 if (ret < 0) {
6304                         err = ret;
6305                         break;
6306                 }
6307
6308                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
6309                 if (ret < 0) {
6310                         err = ret;
6311                         break;
6312                 }
6313
6314                 if (ret > 0) {
6315                         BUG_ON(wc->stage != DROP_REFERENCE);
6316                         break;
6317                 }
6318
6319                 if (wc->stage == DROP_REFERENCE) {
6320                         level = wc->level;
6321                         btrfs_node_key(path->nodes[level],
6322                                        &root_item->drop_progress,
6323                                        path->slots[level]);
6324                         root_item->drop_level = level;
6325                 }
6326
6327                 BUG_ON(wc->level == 0);
6328                 if (btrfs_should_end_transaction(trans, tree_root)) {
6329                         ret = btrfs_update_root(trans, tree_root,
6330                                                 &root->root_key,
6331                                                 root_item);
6332                         BUG_ON(ret);
6333
6334                         btrfs_end_transaction_throttle(trans, tree_root);
6335                         trans = btrfs_start_transaction(tree_root, 0);
6336                         BUG_ON(IS_ERR(trans));
6337                         if (block_rsv)
6338                                 trans->block_rsv = block_rsv;
6339                 }
6340         }
6341         btrfs_release_path(path);
6342         BUG_ON(err);
6343
6344         ret = btrfs_del_root(trans, tree_root, &root->root_key);
6345         BUG_ON(ret);
6346
6347         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
6348                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
6349                                            NULL, NULL);
6350                 BUG_ON(ret < 0);
6351                 if (ret > 0) {
6352                         /* if we fail to delete the orphan item this time
6353                          * around, it'll get picked up the next time.
6354                          *
6355                          * The most common failure here is just -ENOENT.
6356                          */
6357                         btrfs_del_orphan_item(trans, tree_root,
6358                                               root->root_key.objectid);
6359                 }
6360         }
6361
6362         if (root->in_radix) {
6363                 btrfs_free_fs_root(tree_root->fs_info, root);
6364         } else {
6365                 free_extent_buffer(root->node);
6366                 free_extent_buffer(root->commit_root);
6367                 kfree(root);
6368         }
6369 out:
6370         btrfs_end_transaction_throttle(trans, tree_root);
6371         kfree(wc);
6372         btrfs_free_path(path);
6373         return err;
6374 }
6375
6376 /*
6377  * drop subtree rooted at tree block 'node'.
6378  *
6379  * NOTE: this function will unlock and release tree block 'node'
6380  */
6381 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
6382                         struct btrfs_root *root,
6383                         struct extent_buffer *node,
6384                         struct extent_buffer *parent)
6385 {
6386         struct btrfs_path *path;
6387         struct walk_control *wc;
6388         int level;
6389         int parent_level;
6390         int ret = 0;
6391         int wret;
6392
6393         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
6394
6395         path = btrfs_alloc_path();
6396         if (!path)
6397                 return -ENOMEM;
6398
6399         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6400         if (!wc) {
6401                 btrfs_free_path(path);
6402                 return -ENOMEM;
6403         }
6404
6405         btrfs_assert_tree_locked(parent);
6406         parent_level = btrfs_header_level(parent);
6407         extent_buffer_get(parent);
6408         path->nodes[parent_level] = parent;
6409         path->slots[parent_level] = btrfs_header_nritems(parent);
6410
6411         btrfs_assert_tree_locked(node);
6412         level = btrfs_header_level(node);
6413         path->nodes[level] = node;
6414         path->slots[level] = 0;
6415         path->locks[level] = 1;
6416
6417         wc->refs[parent_level] = 1;
6418         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6419         wc->level = level;
6420         wc->shared_level = -1;
6421         wc->stage = DROP_REFERENCE;
6422         wc->update_ref = 0;
6423         wc->keep_locks = 1;
6424         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6425
6426         while (1) {
6427                 wret = walk_down_tree(trans, root, path, wc);
6428                 if (wret < 0) {
6429                         ret = wret;
6430                         break;
6431                 }
6432
6433                 wret = walk_up_tree(trans, root, path, wc, parent_level);
6434                 if (wret < 0)
6435                         ret = wret;
6436                 if (wret != 0)
6437                         break;
6438         }
6439
6440         kfree(wc);
6441         btrfs_free_path(path);
6442         return ret;
6443 }
6444
6445 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
6446 {
6447         u64 num_devices;
6448         u64 stripped = BTRFS_BLOCK_GROUP_RAID0 |
6449                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
6450
6451         /*
6452          * we add in the count of missing devices because we want
6453          * to make sure that any RAID levels on a degraded FS
6454          * continue to be honored.
6455          */
6456         num_devices = root->fs_info->fs_devices->rw_devices +
6457                 root->fs_info->fs_devices->missing_devices;
6458
6459         if (num_devices == 1) {
6460                 stripped |= BTRFS_BLOCK_GROUP_DUP;
6461                 stripped = flags & ~stripped;
6462
6463                 /* turn raid0 into single device chunks */
6464                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
6465                         return stripped;
6466
6467                 /* turn mirroring into duplication */
6468                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
6469                              BTRFS_BLOCK_GROUP_RAID10))
6470                         return stripped | BTRFS_BLOCK_GROUP_DUP;
6471                 return flags;
6472         } else {
6473                 /* they already had raid on here, just return */
6474                 if (flags & stripped)
6475                         return flags;
6476
6477                 stripped |= BTRFS_BLOCK_GROUP_DUP;
6478                 stripped = flags & ~stripped;
6479
6480                 /* switch duplicated blocks with raid1 */
6481                 if (flags & BTRFS_BLOCK_GROUP_DUP)
6482                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
6483
6484                 /* turn single device chunks into raid0 */
6485                 return stripped | BTRFS_BLOCK_GROUP_RAID0;
6486         }
6487         return flags;
6488 }
6489
6490 static int set_block_group_ro(struct btrfs_block_group_cache *cache)
6491 {
6492         struct btrfs_space_info *sinfo = cache->space_info;
6493         u64 num_bytes;
6494         int ret = -ENOSPC;
6495
6496         if (cache->ro)
6497                 return 0;
6498
6499         spin_lock(&sinfo->lock);
6500         spin_lock(&cache->lock);
6501         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
6502                     cache->bytes_super - btrfs_block_group_used(&cache->item);
6503
6504         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
6505             sinfo->bytes_may_use + sinfo->bytes_readonly +
6506             cache->reserved_pinned + num_bytes <= sinfo->total_bytes) {
6507                 sinfo->bytes_readonly += num_bytes;
6508                 sinfo->bytes_reserved += cache->reserved_pinned;
6509                 cache->reserved_pinned = 0;
6510                 cache->ro = 1;
6511                 ret = 0;
6512         }
6513
6514         spin_unlock(&cache->lock);
6515         spin_unlock(&sinfo->lock);
6516         return ret;
6517 }
6518
6519 int btrfs_set_block_group_ro(struct btrfs_root *root,
6520                              struct btrfs_block_group_cache *cache)
6521
6522 {
6523         struct btrfs_trans_handle *trans;
6524         u64 alloc_flags;
6525         int ret;
6526
6527         BUG_ON(cache->ro);
6528
6529         trans = btrfs_join_transaction(root, 1);
6530         BUG_ON(IS_ERR(trans));
6531
6532         alloc_flags = update_block_group_flags(root, cache->flags);
6533         if (alloc_flags != cache->flags)
6534                 do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6535                                CHUNK_ALLOC_FORCE);
6536
6537         ret = set_block_group_ro(cache);
6538         if (!ret)
6539                 goto out;
6540         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
6541         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6542                              CHUNK_ALLOC_FORCE);
6543         if (ret < 0)
6544                 goto out;
6545         ret = set_block_group_ro(cache);
6546 out:
6547         btrfs_end_transaction(trans, root);
6548         return ret;
6549 }
6550
6551 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
6552                             struct btrfs_root *root, u64 type)
6553 {
6554         u64 alloc_flags = get_alloc_profile(root, type);
6555         return do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6556                               CHUNK_ALLOC_FORCE);
6557 }
6558
6559 /*
6560  * helper to account the unused space of all the readonly block group in the
6561  * list. takes mirrors into account.
6562  */
6563 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
6564 {
6565         struct btrfs_block_group_cache *block_group;
6566         u64 free_bytes = 0;
6567         int factor;
6568
6569         list_for_each_entry(block_group, groups_list, list) {
6570                 spin_lock(&block_group->lock);
6571
6572                 if (!block_group->ro) {
6573                         spin_unlock(&block_group->lock);
6574                         continue;
6575                 }
6576
6577                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
6578                                           BTRFS_BLOCK_GROUP_RAID10 |
6579                                           BTRFS_BLOCK_GROUP_DUP))
6580                         factor = 2;
6581                 else
6582                         factor = 1;
6583
6584                 free_bytes += (block_group->key.offset -
6585                                btrfs_block_group_used(&block_group->item)) *
6586                                factor;
6587
6588                 spin_unlock(&block_group->lock);
6589         }
6590
6591         return free_bytes;
6592 }
6593
6594 /*
6595  * helper to account the unused space of all the readonly block group in the
6596  * space_info. takes mirrors into account.
6597  */
6598 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
6599 {
6600         int i;
6601         u64 free_bytes = 0;
6602
6603         spin_lock(&sinfo->lock);
6604
6605         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
6606                 if (!list_empty(&sinfo->block_groups[i]))
6607                         free_bytes += __btrfs_get_ro_block_group_free_space(
6608                                                 &sinfo->block_groups[i]);
6609
6610         spin_unlock(&sinfo->lock);
6611
6612         return free_bytes;
6613 }
6614
6615 int btrfs_set_block_group_rw(struct btrfs_root *root,
6616                               struct btrfs_block_group_cache *cache)
6617 {
6618         struct btrfs_space_info *sinfo = cache->space_info;
6619         u64 num_bytes;
6620
6621         BUG_ON(!cache->ro);
6622
6623         spin_lock(&sinfo->lock);
6624         spin_lock(&cache->lock);
6625         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
6626                     cache->bytes_super - btrfs_block_group_used(&cache->item);
6627         sinfo->bytes_readonly -= num_bytes;
6628         cache->ro = 0;
6629         spin_unlock(&cache->lock);
6630         spin_unlock(&sinfo->lock);
6631         return 0;
6632 }
6633
6634 /*
6635  * checks to see if its even possible to relocate this block group.
6636  *
6637  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
6638  * ok to go ahead and try.
6639  */
6640 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
6641 {
6642         struct btrfs_block_group_cache *block_group;
6643         struct btrfs_space_info *space_info;
6644         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
6645         struct btrfs_device *device;
6646         int full = 0;
6647         int ret = 0;
6648
6649         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
6650
6651         /* odd, couldn't find the block group, leave it alone */
6652         if (!block_group)
6653                 return -1;
6654
6655         /* no bytes used, we're good */
6656         if (!btrfs_block_group_used(&block_group->item))
6657                 goto out;
6658
6659         space_info = block_group->space_info;
6660         spin_lock(&space_info->lock);
6661
6662         full = space_info->full;
6663
6664         /*
6665          * if this is the last block group we have in this space, we can't
6666          * relocate it unless we're able to allocate a new chunk below.
6667          *
6668          * Otherwise, we need to make sure we have room in the space to handle
6669          * all of the extents from this block group.  If we can, we're good
6670          */
6671         if ((space_info->total_bytes != block_group->key.offset) &&
6672            (space_info->bytes_used + space_info->bytes_reserved +
6673             space_info->bytes_pinned + space_info->bytes_readonly +
6674             btrfs_block_group_used(&block_group->item) <
6675             space_info->total_bytes)) {
6676                 spin_unlock(&space_info->lock);
6677                 goto out;
6678         }
6679         spin_unlock(&space_info->lock);
6680
6681         /*
6682          * ok we don't have enough space, but maybe we have free space on our
6683          * devices to allocate new chunks for relocation, so loop through our
6684          * alloc devices and guess if we have enough space.  However, if we
6685          * were marked as full, then we know there aren't enough chunks, and we
6686          * can just return.
6687          */
6688         ret = -1;
6689         if (full)
6690                 goto out;
6691
6692         mutex_lock(&root->fs_info->chunk_mutex);
6693         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
6694                 u64 min_free = btrfs_block_group_used(&block_group->item);
6695                 u64 dev_offset;
6696
6697                 /*
6698                  * check to make sure we can actually find a chunk with enough
6699                  * space to fit our block group in.
6700                  */
6701                 if (device->total_bytes > device->bytes_used + min_free) {
6702                         ret = find_free_dev_extent(NULL, device, min_free,
6703                                                    &dev_offset, NULL);
6704                         if (!ret)
6705                                 break;
6706                         ret = -1;
6707                 }
6708         }
6709         mutex_unlock(&root->fs_info->chunk_mutex);
6710 out:
6711         btrfs_put_block_group(block_group);
6712         return ret;
6713 }
6714
6715 static int find_first_block_group(struct btrfs_root *root,
6716                 struct btrfs_path *path, struct btrfs_key *key)
6717 {
6718         int ret = 0;
6719         struct btrfs_key found_key;
6720         struct extent_buffer *leaf;
6721         int slot;
6722
6723         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
6724         if (ret < 0)
6725                 goto out;
6726
6727         while (1) {
6728                 slot = path->slots[0];
6729                 leaf = path->nodes[0];
6730                 if (slot >= btrfs_header_nritems(leaf)) {
6731                         ret = btrfs_next_leaf(root, path);
6732                         if (ret == 0)
6733                                 continue;
6734                         if (ret < 0)
6735                                 goto out;
6736                         break;
6737                 }
6738                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
6739
6740                 if (found_key.objectid >= key->objectid &&
6741                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
6742                         ret = 0;
6743                         goto out;
6744                 }
6745                 path->slots[0]++;
6746         }
6747 out:
6748         return ret;
6749 }
6750
6751 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
6752 {
6753         struct btrfs_block_group_cache *block_group;
6754         u64 last = 0;
6755
6756         while (1) {
6757                 struct inode *inode;
6758
6759                 block_group = btrfs_lookup_first_block_group(info, last);
6760                 while (block_group) {
6761                         spin_lock(&block_group->lock);
6762                         if (block_group->iref)
6763                                 break;
6764                         spin_unlock(&block_group->lock);
6765                         block_group = next_block_group(info->tree_root,
6766                                                        block_group);
6767                 }
6768                 if (!block_group) {
6769                         if (last == 0)
6770                                 break;
6771                         last = 0;
6772                         continue;
6773                 }
6774
6775                 inode = block_group->inode;
6776                 block_group->iref = 0;
6777                 block_group->inode = NULL;
6778                 spin_unlock(&block_group->lock);
6779                 iput(inode);
6780                 last = block_group->key.objectid + block_group->key.offset;
6781                 btrfs_put_block_group(block_group);
6782         }
6783 }
6784
6785 int btrfs_free_block_groups(struct btrfs_fs_info *info)
6786 {
6787         struct btrfs_block_group_cache *block_group;
6788         struct btrfs_space_info *space_info;
6789         struct btrfs_caching_control *caching_ctl;
6790         struct rb_node *n;
6791
6792         down_write(&info->extent_commit_sem);
6793         while (!list_empty(&info->caching_block_groups)) {
6794                 caching_ctl = list_entry(info->caching_block_groups.next,
6795                                          struct btrfs_caching_control, list);
6796                 list_del(&caching_ctl->list);
6797                 put_caching_control(caching_ctl);
6798         }
6799         up_write(&info->extent_commit_sem);
6800
6801         spin_lock(&info->block_group_cache_lock);
6802         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
6803                 block_group = rb_entry(n, struct btrfs_block_group_cache,
6804                                        cache_node);
6805                 rb_erase(&block_group->cache_node,
6806                          &info->block_group_cache_tree);
6807                 spin_unlock(&info->block_group_cache_lock);
6808
6809                 down_write(&block_group->space_info->groups_sem);
6810                 list_del(&block_group->list);
6811                 up_write(&block_group->space_info->groups_sem);
6812
6813                 if (block_group->cached == BTRFS_CACHE_STARTED)
6814                         wait_block_group_cache_done(block_group);
6815
6816                 /*
6817                  * We haven't cached this block group, which means we could
6818                  * possibly have excluded extents on this block group.
6819                  */
6820                 if (block_group->cached == BTRFS_CACHE_NO)
6821                         free_excluded_extents(info->extent_root, block_group);
6822
6823                 btrfs_remove_free_space_cache(block_group);
6824                 btrfs_put_block_group(block_group);
6825
6826                 spin_lock(&info->block_group_cache_lock);
6827         }
6828         spin_unlock(&info->block_group_cache_lock);
6829
6830         /* now that all the block groups are freed, go through and
6831          * free all the space_info structs.  This is only called during
6832          * the final stages of unmount, and so we know nobody is
6833          * using them.  We call synchronize_rcu() once before we start,
6834          * just to be on the safe side.
6835          */
6836         synchronize_rcu();
6837
6838         release_global_block_rsv(info);
6839
6840         while(!list_empty(&info->space_info)) {
6841                 space_info = list_entry(info->space_info.next,
6842                                         struct btrfs_space_info,
6843                                         list);
6844                 if (space_info->bytes_pinned > 0 ||
6845                     space_info->bytes_reserved > 0) {
6846                         WARN_ON(1);
6847                         dump_space_info(space_info, 0, 0);
6848                 }
6849                 list_del(&space_info->list);
6850                 kfree(space_info);
6851         }
6852         return 0;
6853 }
6854
6855 static void __link_block_group(struct btrfs_space_info *space_info,
6856                                struct btrfs_block_group_cache *cache)
6857 {
6858         int index = get_block_group_index(cache);
6859
6860         down_write(&space_info->groups_sem);
6861         list_add_tail(&cache->list, &space_info->block_groups[index]);
6862         up_write(&space_info->groups_sem);
6863 }
6864
6865 int btrfs_read_block_groups(struct btrfs_root *root)
6866 {
6867         struct btrfs_path *path;
6868         int ret;
6869         struct btrfs_block_group_cache *cache;
6870         struct btrfs_fs_info *info = root->fs_info;
6871         struct btrfs_space_info *space_info;
6872         struct btrfs_key key;
6873         struct btrfs_key found_key;
6874         struct extent_buffer *leaf;
6875         int need_clear = 0;
6876         u64 cache_gen;
6877
6878         root = info->extent_root;
6879         key.objectid = 0;
6880         key.offset = 0;
6881         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
6882         path = btrfs_alloc_path();
6883         if (!path)
6884                 return -ENOMEM;
6885
6886         cache_gen = btrfs_super_cache_generation(&root->fs_info->super_copy);
6887         if (cache_gen != 0 &&
6888             btrfs_super_generation(&root->fs_info->super_copy) != cache_gen)
6889                 need_clear = 1;
6890         if (btrfs_test_opt(root, CLEAR_CACHE))
6891                 need_clear = 1;
6892         if (!btrfs_test_opt(root, SPACE_CACHE) && cache_gen)
6893                 printk(KERN_INFO "btrfs: disk space caching is enabled\n");
6894
6895         while (1) {
6896                 ret = find_first_block_group(root, path, &key);
6897                 if (ret > 0)
6898                         break;
6899                 if (ret != 0)
6900                         goto error;
6901                 leaf = path->nodes[0];
6902                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6903                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
6904                 if (!cache) {
6905                         ret = -ENOMEM;
6906                         goto error;
6907                 }
6908                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
6909                                                 GFP_NOFS);
6910                 if (!cache->free_space_ctl) {
6911                         kfree(cache);
6912                         ret = -ENOMEM;
6913                         goto error;
6914                 }
6915
6916                 atomic_set(&cache->count, 1);
6917                 spin_lock_init(&cache->lock);
6918                 cache->fs_info = info;
6919                 INIT_LIST_HEAD(&cache->list);
6920                 INIT_LIST_HEAD(&cache->cluster_list);
6921
6922                 if (need_clear)
6923                         cache->disk_cache_state = BTRFS_DC_CLEAR;
6924
6925                 read_extent_buffer(leaf, &cache->item,
6926                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
6927                                    sizeof(cache->item));
6928                 memcpy(&cache->key, &found_key, sizeof(found_key));
6929
6930                 key.objectid = found_key.objectid + found_key.offset;
6931                 btrfs_release_path(path);
6932                 cache->flags = btrfs_block_group_flags(&cache->item);
6933                 cache->sectorsize = root->sectorsize;
6934
6935                 btrfs_init_free_space_ctl(cache);
6936
6937                 /*
6938                  * We need to exclude the super stripes now so that the space
6939                  * info has super bytes accounted for, otherwise we'll think
6940                  * we have more space than we actually do.
6941                  */
6942                 exclude_super_stripes(root, cache);
6943
6944                 /*
6945                  * check for two cases, either we are full, and therefore
6946                  * don't need to bother with the caching work since we won't
6947                  * find any space, or we are empty, and we can just add all
6948                  * the space in and be done with it.  This saves us _alot_ of
6949                  * time, particularly in the full case.
6950                  */
6951                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
6952                         cache->last_byte_to_unpin = (u64)-1;
6953                         cache->cached = BTRFS_CACHE_FINISHED;
6954                         free_excluded_extents(root, cache);
6955                 } else if (btrfs_block_group_used(&cache->item) == 0) {
6956                         cache->last_byte_to_unpin = (u64)-1;
6957                         cache->cached = BTRFS_CACHE_FINISHED;
6958                         add_new_free_space(cache, root->fs_info,
6959                                            found_key.objectid,
6960                                            found_key.objectid +
6961                                            found_key.offset);
6962                         free_excluded_extents(root, cache);
6963                 }
6964
6965                 ret = update_space_info(info, cache->flags, found_key.offset,
6966                                         btrfs_block_group_used(&cache->item),
6967                                         &space_info);
6968                 BUG_ON(ret);
6969                 cache->space_info = space_info;
6970                 spin_lock(&cache->space_info->lock);
6971                 cache->space_info->bytes_readonly += cache->bytes_super;
6972                 spin_unlock(&cache->space_info->lock);
6973
6974                 __link_block_group(space_info, cache);
6975
6976                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
6977                 BUG_ON(ret);
6978
6979                 set_avail_alloc_bits(root->fs_info, cache->flags);
6980                 if (btrfs_chunk_readonly(root, cache->key.objectid))
6981                         set_block_group_ro(cache);
6982         }
6983
6984         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
6985                 if (!(get_alloc_profile(root, space_info->flags) &
6986                       (BTRFS_BLOCK_GROUP_RAID10 |
6987                        BTRFS_BLOCK_GROUP_RAID1 |
6988                        BTRFS_BLOCK_GROUP_DUP)))
6989                         continue;
6990                 /*
6991                  * avoid allocating from un-mirrored block group if there are
6992                  * mirrored block groups.
6993                  */
6994                 list_for_each_entry(cache, &space_info->block_groups[3], list)
6995                         set_block_group_ro(cache);
6996                 list_for_each_entry(cache, &space_info->block_groups[4], list)
6997                         set_block_group_ro(cache);
6998         }
6999
7000         init_global_block_rsv(info);
7001         ret = 0;
7002 error:
7003         btrfs_free_path(path);
7004         return ret;
7005 }
7006
7007 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7008                            struct btrfs_root *root, u64 bytes_used,
7009                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7010                            u64 size)
7011 {
7012         int ret;
7013         struct btrfs_root *extent_root;
7014         struct btrfs_block_group_cache *cache;
7015
7016         extent_root = root->fs_info->extent_root;
7017
7018         root->fs_info->last_trans_log_full_commit = trans->transid;
7019
7020         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7021         if (!cache)
7022                 return -ENOMEM;
7023         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7024                                         GFP_NOFS);
7025         if (!cache->free_space_ctl) {
7026                 kfree(cache);
7027                 return -ENOMEM;
7028         }
7029
7030         cache->key.objectid = chunk_offset;
7031         cache->key.offset = size;
7032         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7033         cache->sectorsize = root->sectorsize;
7034         cache->fs_info = root->fs_info;
7035
7036         atomic_set(&cache->count, 1);
7037         spin_lock_init(&cache->lock);
7038         INIT_LIST_HEAD(&cache->list);
7039         INIT_LIST_HEAD(&cache->cluster_list);
7040
7041         btrfs_init_free_space_ctl(cache);
7042
7043         btrfs_set_block_group_used(&cache->item, bytes_used);
7044         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7045         cache->flags = type;
7046         btrfs_set_block_group_flags(&cache->item, type);
7047
7048         cache->last_byte_to_unpin = (u64)-1;
7049         cache->cached = BTRFS_CACHE_FINISHED;
7050         exclude_super_stripes(root, cache);
7051
7052         add_new_free_space(cache, root->fs_info, chunk_offset,
7053                            chunk_offset + size);
7054
7055         free_excluded_extents(root, cache);
7056
7057         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7058                                 &cache->space_info);
7059         BUG_ON(ret);
7060
7061         spin_lock(&cache->space_info->lock);
7062         cache->space_info->bytes_readonly += cache->bytes_super;
7063         spin_unlock(&cache->space_info->lock);
7064
7065         __link_block_group(cache->space_info, cache);
7066
7067         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7068         BUG_ON(ret);
7069
7070         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
7071                                 sizeof(cache->item));
7072         BUG_ON(ret);
7073
7074         set_avail_alloc_bits(extent_root->fs_info, type);
7075
7076         return 0;
7077 }
7078
7079 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7080                              struct btrfs_root *root, u64 group_start)
7081 {
7082         struct btrfs_path *path;
7083         struct btrfs_block_group_cache *block_group;
7084         struct btrfs_free_cluster *cluster;
7085         struct btrfs_root *tree_root = root->fs_info->tree_root;
7086         struct btrfs_key key;
7087         struct inode *inode;
7088         int ret;
7089         int factor;
7090
7091         root = root->fs_info->extent_root;
7092
7093         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7094         BUG_ON(!block_group);
7095         BUG_ON(!block_group->ro);
7096
7097         /*
7098          * Free the reserved super bytes from this block group before
7099          * remove it.
7100          */
7101         free_excluded_extents(root, block_group);
7102
7103         memcpy(&key, &block_group->key, sizeof(key));
7104         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7105                                   BTRFS_BLOCK_GROUP_RAID1 |
7106                                   BTRFS_BLOCK_GROUP_RAID10))
7107                 factor = 2;
7108         else
7109                 factor = 1;
7110
7111         /* make sure this block group isn't part of an allocation cluster */
7112         cluster = &root->fs_info->data_alloc_cluster;
7113         spin_lock(&cluster->refill_lock);
7114         btrfs_return_cluster_to_free_space(block_group, cluster);
7115         spin_unlock(&cluster->refill_lock);
7116
7117         /*
7118          * make sure this block group isn't part of a metadata
7119          * allocation cluster
7120          */
7121         cluster = &root->fs_info->meta_alloc_cluster;
7122         spin_lock(&cluster->refill_lock);
7123         btrfs_return_cluster_to_free_space(block_group, cluster);
7124         spin_unlock(&cluster->refill_lock);
7125
7126         path = btrfs_alloc_path();
7127         BUG_ON(!path);
7128
7129         inode = lookup_free_space_inode(root, block_group, path);
7130         if (!IS_ERR(inode)) {
7131                 btrfs_orphan_add(trans, inode);
7132                 clear_nlink(inode);
7133                 /* One for the block groups ref */
7134                 spin_lock(&block_group->lock);
7135                 if (block_group->iref) {
7136                         block_group->iref = 0;
7137                         block_group->inode = NULL;
7138                         spin_unlock(&block_group->lock);
7139                         iput(inode);
7140                 } else {
7141                         spin_unlock(&block_group->lock);
7142                 }
7143                 /* One for our lookup ref */
7144                 iput(inode);
7145         }
7146
7147         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
7148         key.offset = block_group->key.objectid;
7149         key.type = 0;
7150
7151         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
7152         if (ret < 0)
7153                 goto out;
7154         if (ret > 0)
7155                 btrfs_release_path(path);
7156         if (ret == 0) {
7157                 ret = btrfs_del_item(trans, tree_root, path);
7158                 if (ret)
7159                         goto out;
7160                 btrfs_release_path(path);
7161         }
7162
7163         spin_lock(&root->fs_info->block_group_cache_lock);
7164         rb_erase(&block_group->cache_node,
7165                  &root->fs_info->block_group_cache_tree);
7166         spin_unlock(&root->fs_info->block_group_cache_lock);
7167
7168         down_write(&block_group->space_info->groups_sem);
7169         /*
7170          * we must use list_del_init so people can check to see if they
7171          * are still on the list after taking the semaphore
7172          */
7173         list_del_init(&block_group->list);
7174         up_write(&block_group->space_info->groups_sem);
7175
7176         if (block_group->cached == BTRFS_CACHE_STARTED)
7177                 wait_block_group_cache_done(block_group);
7178
7179         btrfs_remove_free_space_cache(block_group);
7180
7181         spin_lock(&block_group->space_info->lock);
7182         block_group->space_info->total_bytes -= block_group->key.offset;
7183         block_group->space_info->bytes_readonly -= block_group->key.offset;
7184         block_group->space_info->disk_total -= block_group->key.offset * factor;
7185         spin_unlock(&block_group->space_info->lock);
7186
7187         memcpy(&key, &block_group->key, sizeof(key));
7188
7189         btrfs_clear_space_info_full(root->fs_info);
7190
7191         btrfs_put_block_group(block_group);
7192         btrfs_put_block_group(block_group);
7193
7194         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
7195         if (ret > 0)
7196                 ret = -EIO;
7197         if (ret < 0)
7198                 goto out;
7199
7200         ret = btrfs_del_item(trans, root, path);
7201 out:
7202         btrfs_free_path(path);
7203         return ret;
7204 }
7205
7206 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
7207 {
7208         struct btrfs_space_info *space_info;
7209         struct btrfs_super_block *disk_super;
7210         u64 features;
7211         u64 flags;
7212         int mixed = 0;
7213         int ret;
7214
7215         disk_super = &fs_info->super_copy;
7216         if (!btrfs_super_root(disk_super))
7217                 return 1;
7218
7219         features = btrfs_super_incompat_flags(disk_super);
7220         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
7221                 mixed = 1;
7222
7223         flags = BTRFS_BLOCK_GROUP_SYSTEM;
7224         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7225         if (ret)
7226                 goto out;
7227
7228         if (mixed) {
7229                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
7230                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7231         } else {
7232                 flags = BTRFS_BLOCK_GROUP_METADATA;
7233                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7234                 if (ret)
7235                         goto out;
7236
7237                 flags = BTRFS_BLOCK_GROUP_DATA;
7238                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7239         }
7240 out:
7241         return ret;
7242 }
7243
7244 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
7245 {
7246         return unpin_extent_range(root, start, end);
7247 }
7248
7249 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
7250                                u64 num_bytes, u64 *actual_bytes)
7251 {
7252         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
7253 }
7254
7255 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
7256 {
7257         struct btrfs_fs_info *fs_info = root->fs_info;
7258         struct btrfs_block_group_cache *cache = NULL;
7259         u64 group_trimmed;
7260         u64 start;
7261         u64 end;
7262         u64 trimmed = 0;
7263         int ret = 0;
7264
7265         cache = btrfs_lookup_block_group(fs_info, range->start);
7266
7267         while (cache) {
7268                 if (cache->key.objectid >= (range->start + range->len)) {
7269                         btrfs_put_block_group(cache);
7270                         break;
7271                 }
7272
7273                 start = max(range->start, cache->key.objectid);
7274                 end = min(range->start + range->len,
7275                                 cache->key.objectid + cache->key.offset);
7276
7277                 if (end - start >= range->minlen) {
7278                         if (!block_group_cache_done(cache)) {
7279                                 ret = cache_block_group(cache, NULL, root, 0);
7280                                 if (!ret)
7281                                         wait_block_group_cache_done(cache);
7282                         }
7283                         ret = btrfs_trim_block_group(cache,
7284                                                      &group_trimmed,
7285                                                      start,
7286                                                      end,
7287                                                      range->minlen);
7288
7289                         trimmed += group_trimmed;
7290                         if (ret) {
7291                                 btrfs_put_block_group(cache);
7292                                 break;
7293                         }
7294                 }
7295
7296                 cache = next_block_group(fs_info->tree_root, cache);
7297         }
7298
7299         range->len = trimmed;
7300         return ret;
7301 }