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 static int update_block_group(struct btrfs_trans_handle *trans,
  37                               struct btrfs_root *root,
  38                               u64 bytenr, u64 num_bytes, int alloc);
  39 static int update_reserved_bytes(struct btrfs_block_group_cache *cache,
  40                                  u64 num_bytes, int reserve, int sinfo);
  41 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  42                                 struct btrfs_root *root,
  43                                 u64 bytenr, u64 num_bytes, u64 parent,
  44                                 u64 root_objectid, u64 owner_objectid,
  45                                 u64 owner_offset, int refs_to_drop,
  46                                 struct btrfs_delayed_extent_op *extra_op);
  47 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  48                                     struct extent_buffer *leaf,
  49                                     struct btrfs_extent_item *ei);
  50 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  51                                       struct btrfs_root *root,
  52                                       u64 parent, u64 root_objectid,
  53                                       u64 flags, u64 owner, u64 offset,
  54                                       struct btrfs_key *ins, int ref_mod);
  55 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  56                                      struct btrfs_root *root,
  57                                      u64 parent, u64 root_objectid,
  58                                      u64 flags, struct btrfs_disk_key *key,
  59                                      int level, struct btrfs_key *ins);
  60 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  61                           struct btrfs_root *extent_root, u64 alloc_bytes,
  62                           u64 flags, int force);
  63 static int find_next_key(struct btrfs_path *path, int level,
  64                          struct btrfs_key *key);
  65 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
  66                             int dump_block_groups);
  67
  68 static noinline int
  69 block_group_cache_done(struct btrfs_block_group_cache *cache)
  70 {
  71         smp_mb();
  72         return cache->cached == BTRFS_CACHE_FINISHED;
  73 }
  74
  75 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
  76 {
  77         return (cache->flags & bits) == bits;
  78 }
  79
  80 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
  81 {
  82         atomic_inc(&cache->count);
  83 }
  84
  85 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
  86 {
  87         if (atomic_dec_and_test(&cache->count)) {
  88                 WARN_ON(cache->pinned > 0);
  89                 WARN_ON(cache->reserved > 0);
  90                 WARN_ON(cache->reserved_pinned > 0);
  91                 kfree(cache);
  92         }
  93 }
  94
  95 /*
  96  * this adds the block group to the fs_info rb tree for the block group
  97  * cache
  98  */
  99 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 100                                 struct btrfs_block_group_cache *block_group)
 101 {
 102         struct rb_node **p;
 103         struct rb_node *parent = NULL;
 104         struct btrfs_block_group_cache *cache;
 105
 106         spin_lock(&info->block_group_cache_lock);
 107         p = &info->block_group_cache_tree.rb_node;
 108
 109         while (*p) {
 110                 parent = *p;
 111                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 112                                  cache_node);
 113                 if (block_group->key.objectid < cache->key.objectid) {
 114                         p = &(*p)->rb_left;
 115                 } else if (block_group->key.objectid > cache->key.objectid) {
 116                         p = &(*p)->rb_right;
 117                 } else {
 118                         spin_unlock(&info->block_group_cache_lock);
 119                         return -EEXIST;
 120                 }
 121         }
 122
 123         rb_link_node(&block_group->cache_node, parent, p);
 124         rb_insert_color(&block_group->cache_node,
 125                         &info->block_group_cache_tree);
 126         spin_unlock(&info->block_group_cache_lock);
 127
 128         return 0;
 129 }
 130
 131 /*
 132  * This will return the block group at or after bytenr if contains is 0, else
 133  * it will return the block group that contains the bytenr
 134  */
 135 static struct btrfs_block_group_cache *
 136 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 137                               int contains)
 138 {
 139         struct btrfs_block_group_cache *cache, *ret = NULL;
 140         struct rb_node *n;
 141         u64 end, start;
 142
 143         spin_lock(&info->block_group_cache_lock);
 144         n = info->block_group_cache_tree.rb_node;
 145
 146         while (n) {
 147                 cache = rb_entry(n, struct btrfs_block_group_cache,
 148                                  cache_node);
 149                 end = cache->key.objectid + cache->key.offset - 1;
 150                 start = cache->key.objectid;
 151
 152                 if (bytenr < start) {
 153                         if (!contains && (!ret || start < ret->key.objectid))
 154                                 ret = cache;
 155                         n = n->rb_left;
 156                 } else if (bytenr > start) {
 157                         if (contains && bytenr <= end) {
 158                                 ret = cache;
 159                                 break;
 160                         }
 161                         n = n->rb_right;
 162                 } else {
 163                         ret = cache;
 164                         break;
 165                 }
 166         }
 167         if (ret)
 168                 btrfs_get_block_group(ret);
 169         spin_unlock(&info->block_group_cache_lock);
 170
 171         return ret;
 172 }
 173
 174 static int add_excluded_extent(struct btrfs_root *root,
 175                                u64 start, u64 num_bytes)
 176 {
 177         u64 end = start + num_bytes - 1;
 178         set_extent_bits(&root->fs_info->freed_extents[0],
 179                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 180         set_extent_bits(&root->fs_info->freed_extents[1],
 181                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 182         return 0;
 183 }
 184
 185 static void free_excluded_extents(struct btrfs_root *root,
 186                                   struct btrfs_block_group_cache *cache)
 187 {
 188         u64 start, end;
 189
 190         start = cache->key.objectid;
 191         end = start + cache->key.offset - 1;
 192
 193         clear_extent_bits(&root->fs_info->freed_extents[0],
 194                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 195         clear_extent_bits(&root->fs_info->freed_extents[1],
 196                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 197 }
 198
 199 static int exclude_super_stripes(struct btrfs_root *root,
 200                                  struct btrfs_block_group_cache *cache)
 201 {
 202         u64 bytenr;
 203         u64 *logical;
 204         int stripe_len;
 205         int i, nr, ret;
 206
 207         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 208                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 209                 cache->bytes_super += stripe_len;
 210                 ret = add_excluded_extent(root, cache->key.objectid,
 211                                           stripe_len);
 212                 BUG_ON(ret);
 213         }
 214
 215         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 216                 bytenr = btrfs_sb_offset(i);
 217                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 218                                        cache->key.objectid, bytenr,
 219                                        0, &logical, &nr, &stripe_len);
 220                 BUG_ON(ret);
 221
 222                 while (nr--) {
 223                         cache->bytes_super += stripe_len;
 224                         ret = add_excluded_extent(root, logical[nr],
 225                                                   stripe_len);
 226                         BUG_ON(ret);
 227                 }
 228
 229                 kfree(logical);
 230         }
 231         return 0;
 232 }
 233
 234 static struct btrfs_caching_control *
 235 get_caching_control(struct btrfs_block_group_cache *cache)
 236 {
 237         struct btrfs_caching_control *ctl;
 238
 239         spin_lock(&cache->lock);
 240         if (cache->cached != BTRFS_CACHE_STARTED) {
 241                 spin_unlock(&cache->lock);
 242                 return NULL;
 243         }
 244
 245         /* We're loading it the fast way, so we don't have a caching_ctl. */
 246         if (!cache->caching_ctl) {
 247                 spin_unlock(&cache->lock);
 248                 return NULL;
 249         }
 250
 251         ctl = cache->caching_ctl;
 252         atomic_inc(&ctl->count);
 253         spin_unlock(&cache->lock);
 254         return ctl;
 255 }
 256
 257 static void put_caching_control(struct btrfs_caching_control *ctl)
 258 {
 259         if (atomic_dec_and_test(&ctl->count))
 260                 kfree(ctl);
 261 }
 262
 263 /*
 264  * this is only called by cache_block_group, since we could have freed extents
 265  * we need to check the pinned_extents for any extents that can't be used yet
 266  * since their free space will be released as soon as the transaction commits.
 267  */
 268 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 269                               struct btrfs_fs_info *info, u64 start, u64 end)
 270 {
 271         u64 extent_start, extent_end, size, total_added = 0;
 272         int ret;
 273
 274         while (start < end) {
 275                 ret = find_first_extent_bit(info->pinned_extents, start,
 276                                             &extent_start, &extent_end,
 277                                             EXTENT_DIRTY | EXTENT_UPTODATE);
 278                 if (ret)
 279                         break;
 280
 281                 if (extent_start <= start) {
 282                         start = extent_end + 1;
 283                 } else if (extent_start > start && extent_start < end) {
 284                         size = extent_start - start;
 285                         total_added += size;
 286                         ret = btrfs_add_free_space(block_group, start,
 287                                                    size);
 288                         BUG_ON(ret);
 289                         start = extent_end + 1;
 290                 } else {
 291                         break;
 292                 }
 293         }
 294
 295         if (start < end) {
 296                 size = end - start;
 297                 total_added += size;
 298                 ret = btrfs_add_free_space(block_group, start, size);
 299                 BUG_ON(ret);
 300         }
 301
 302         return total_added;
 303 }
 304
 305 static int caching_kthread(void *data)
 306 {
 307         struct btrfs_block_group_cache *block_group = data;
 308         struct btrfs_fs_info *fs_info = block_group->fs_info;
 309         struct btrfs_caching_control *caching_ctl = block_group->caching_ctl;
 310         struct btrfs_root *extent_root = fs_info->extent_root;
 311         struct btrfs_path *path;
 312         struct extent_buffer *leaf;
 313         struct btrfs_key key;
 314         u64 total_found = 0;
 315         u64 last = 0;
 316         u32 nritems;
 317         int ret = 0;
 318
 319         path = btrfs_alloc_path();
 320         if (!path)
 321                 return -ENOMEM;
 322
 323         exclude_super_stripes(extent_root, block_group);
 324         spin_lock(&block_group->space_info->lock);
 325         block_group->space_info->bytes_readonly += block_group->bytes_super;
 326         spin_unlock(&block_group->space_info->lock);
 327
 328         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 329
 330         /*
 331          * We don't want to deadlock with somebody trying to allocate a new
 332          * extent for the extent root while also trying to search the extent
 333          * root to add free space.  So we skip locking and search the commit
 334          * root, since its read-only
 335          */
 336         path->skip_locking = 1;
 337         path->search_commit_root = 1;
 338         path->reada = 2;
 339
 340         key.objectid = last;
 341         key.offset = 0;
 342         key.type = BTRFS_EXTENT_ITEM_KEY;
 343 again:
 344         mutex_lock(&caching_ctl->mutex);
 345         /* need to make sure the commit_root doesn't disappear */
 346         down_read(&fs_info->extent_commit_sem);
 347
 348         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 349         if (ret < 0)
 350                 goto err;
 351
 352         leaf = path->nodes[0];
 353         nritems = btrfs_header_nritems(leaf);
 354
 355         while (1) {
 356                 smp_mb();
 357                 if (fs_info->closing > 1) {
 358                         last = (u64)-1;
 359                         break;
 360                 }
 361
 362                 if (path->slots[0] < nritems) {
 363                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 364                 } else {
 365                         ret = find_next_key(path, 0, &key);
 366                         if (ret)
 367                                 break;
 368
 369                         caching_ctl->progress = last;
 370                         btrfs_release_path(extent_root, path);
 371                         up_read(&fs_info->extent_commit_sem);
 372                         mutex_unlock(&caching_ctl->mutex);
 373                         if (btrfs_transaction_in_commit(fs_info))
 374                                 schedule_timeout(1);
 375                         else
 376                                 cond_resched();
 377                         goto again;
 378                 }
 379
 380                 if (key.objectid < block_group->key.objectid) {
 381                         path->slots[0]++;
 382                         continue;
 383                 }
 384
 385                 if (key.objectid >= block_group->key.objectid +
 386                     block_group->key.offset)
 387                         break;
 388
 389                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 390                         total_found += add_new_free_space(block_group,
 391                                                           fs_info, last,
 392                                                           key.objectid);
 393                         last = key.objectid + key.offset;
 394
 395                         if (total_found > (1024 * 1024 * 2)) {
 396                                 total_found = 0;
 397                                 wake_up(&caching_ctl->wait);
 398                         }
 399                 }
 400                 path->slots[0]++;
 401         }
 402         ret = 0;
 403
 404         total_found += add_new_free_space(block_group, fs_info, last,
 405                                           block_group->key.objectid +
 406                                           block_group->key.offset);
 407         caching_ctl->progress = (u64)-1;
 408
 409         spin_lock(&block_group->lock);
 410         block_group->caching_ctl = NULL;
 411         block_group->cached = BTRFS_CACHE_FINISHED;
 412         spin_unlock(&block_group->lock);
 413
 414 err:
 415         btrfs_free_path(path);
 416         up_read(&fs_info->extent_commit_sem);
 417
 418         free_excluded_extents(extent_root, block_group);
 419
 420         mutex_unlock(&caching_ctl->mutex);
 421         wake_up(&caching_ctl->wait);
 422
 423         put_caching_control(caching_ctl);
 424         atomic_dec(&block_group->space_info->caching_threads);
 425         btrfs_put_block_group(block_group);
 426
 427         return 0;
 428 }
 429
 430 static int cache_block_group(struct btrfs_block_group_cache *cache,
 431                              struct btrfs_trans_handle *trans,
 432                              struct btrfs_root *root,
 433                              int load_cache_only)
 434 {
 435         struct btrfs_fs_info *fs_info = cache->fs_info;
 436         struct btrfs_caching_control *caching_ctl;
 437         struct task_struct *tsk;
 438         int ret = 0;
 439
 440         smp_mb();
 441         if (cache->cached != BTRFS_CACHE_NO)
 442                 return 0;
 443
 444         /*
 445          * We can't do the read from on-disk cache during a commit since we need
 446          * to have the normal tree locking.  Also if we are currently trying to
 447          * allocate blocks for the tree root we can't do the fast caching since
 448          * we likely hold important locks.
 449          */
 450         if (!trans->transaction->in_commit &&
 451             (root && root != root->fs_info->tree_root)) {
 452                 spin_lock(&cache->lock);
 453                 if (cache->cached != BTRFS_CACHE_NO) {
 454                         spin_unlock(&cache->lock);
 455                         return 0;
 456                 }
 457                 cache->cached = BTRFS_CACHE_STARTED;
 458                 spin_unlock(&cache->lock);
 459
 460                 ret = load_free_space_cache(fs_info, cache);
 461
 462                 spin_lock(&cache->lock);
 463                 if (ret == 1) {
 464                         cache->cached = BTRFS_CACHE_FINISHED;
 465                         cache->last_byte_to_unpin = (u64)-1;
 466                 } else {
 467                         cache->cached = BTRFS_CACHE_NO;
 468                 }
 469                 spin_unlock(&cache->lock);
 470                 if (ret == 1)
 471                         return 0;
 472         }
 473
 474         if (load_cache_only)
 475                 return 0;
 476
 477         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_KERNEL);
 478         BUG_ON(!caching_ctl);
 479
 480         INIT_LIST_HEAD(&caching_ctl->list);
 481         mutex_init(&caching_ctl->mutex);
 482         init_waitqueue_head(&caching_ctl->wait);
 483         caching_ctl->block_group = cache;
 484         caching_ctl->progress = cache->key.objectid;
 485         /* one for caching kthread, one for caching block group list */
 486         atomic_set(&caching_ctl->count, 2);
 487
 488         spin_lock(&cache->lock);
 489         if (cache->cached != BTRFS_CACHE_NO) {
 490                 spin_unlock(&cache->lock);
 491                 kfree(caching_ctl);
 492                 return 0;
 493         }
 494         cache->caching_ctl = caching_ctl;
 495         cache->cached = BTRFS_CACHE_STARTED;
 496         spin_unlock(&cache->lock);
 497
 498         down_write(&fs_info->extent_commit_sem);
 499         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 500         up_write(&fs_info->extent_commit_sem);
 501
 502         atomic_inc(&cache->space_info->caching_threads);
 503         btrfs_get_block_group(cache);
 504
 505         tsk = kthread_run(caching_kthread, cache, "btrfs-cache-%llu\n",
 506                           cache->key.objectid);
 507         if (IS_ERR(tsk)) {
 508                 ret = PTR_ERR(tsk);
 509                 printk(KERN_ERR "error running thread %d\n", ret);
 510                 BUG();
 511         }
 512
 513         return ret;
 514 }
 515
 516 /*
 517  * return the block group that starts at or after bytenr
 518  */
 519 static struct btrfs_block_group_cache *
 520 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 521 {
 522         struct btrfs_block_group_cache *cache;
 523
 524         cache = block_group_cache_tree_search(info, bytenr, 0);
 525
 526         return cache;
 527 }
 528
 529 /*
 530  * return the block group that contains the given bytenr
 531  */
 532 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 533                                                  struct btrfs_fs_info *info,
 534                                                  u64 bytenr)
 535 {
 536         struct btrfs_block_group_cache *cache;
 537
 538         cache = block_group_cache_tree_search(info, bytenr, 1);
 539
 540         return cache;
 541 }
 542
 543 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 544                                                   u64 flags)
 545 {
 546         struct list_head *head = &info->space_info;
 547         struct btrfs_space_info *found;
 548
 549         flags &= BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_SYSTEM |
 550                  BTRFS_BLOCK_GROUP_METADATA;
 551
 552         rcu_read_lock();
 553         list_for_each_entry_rcu(found, head, list) {
 554                 if (found->flags & flags) {
 555                         rcu_read_unlock();
 556                         return found;
 557                 }
 558         }
 559         rcu_read_unlock();
 560         return NULL;
 561 }
 562
 563 /*
 564  * after adding space to the filesystem, we need to clear the full flags
 565  * on all the space infos.
 566  */
 567 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 568 {
 569         struct list_head *head = &info->space_info;
 570         struct btrfs_space_info *found;
 571
 572         rcu_read_lock();
 573         list_for_each_entry_rcu(found, head, list)
 574                 found->full = 0;
 575         rcu_read_unlock();
 576 }
 577
 578 static u64 div_factor(u64 num, int factor)
 579 {
 580         if (factor == 10)
 581                 return num;
 582         num *= factor;
 583         do_div(num, 10);
 584         return num;
 585 }
 586
 587 static u64 div_factor_fine(u64 num, int factor)
 588 {
 589         if (factor == 100)
 590                 return num;
 591         num *= factor;
 592         do_div(num, 100);
 593         return num;
 594 }
 595
 596 u64 btrfs_find_block_group(struct btrfs_root *root,
 597                            u64 search_start, u64 search_hint, int owner)
 598 {
 599         struct btrfs_block_group_cache *cache;
 600         u64 used;
 601         u64 last = max(search_hint, search_start);
 602         u64 group_start = 0;
 603         int full_search = 0;
 604         int factor = 9;
 605         int wrapped = 0;
 606 again:
 607         while (1) {
 608                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
 609                 if (!cache)
 610                         break;
 611
 612                 spin_lock(&cache->lock);
 613                 last = cache->key.objectid + cache->key.offset;
 614                 used = btrfs_block_group_used(&cache->item);
 615
 616                 if ((full_search || !cache->ro) &&
 617                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
 618                         if (used + cache->pinned + cache->reserved <
 619                             div_factor(cache->key.offset, factor)) {
 620                                 group_start = cache->key.objectid;
 621                                 spin_unlock(&cache->lock);
 622                                 btrfs_put_block_group(cache);
 623                                 goto found;
 624                         }
 625                 }
 626                 spin_unlock(&cache->lock);
 627                 btrfs_put_block_group(cache);
 628                 cond_resched();
 629         }
 630         if (!wrapped) {
 631                 last = search_start;
 632                 wrapped = 1;
 633                 goto again;
 634         }
 635         if (!full_search && factor < 10) {
 636                 last = search_start;
 637                 full_search = 1;
 638                 factor = 10;
 639                 goto again;
 640         }
 641 found:
 642         return group_start;
 643 }
 644
 645 /* simple helper to search for an existing extent at a given offset */
 646 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 647 {
 648         int ret;
 649         struct btrfs_key key;
 650         struct btrfs_path *path;
 651
 652         path = btrfs_alloc_path();
 653         BUG_ON(!path);
 654         key.objectid = start;
 655         key.offset = len;
 656         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 657         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 658                                 0, 0);
 659         btrfs_free_path(path);
 660         return ret;
 661 }
 662
 663 /*
 664  * helper function to lookup reference count and flags of extent.
 665  *
 666  * the head node for delayed ref is used to store the sum of all the
 667  * reference count modifications queued up in the rbtree. the head
 668  * node may also store the extent flags to set. This way you can check
 669  * to see what the reference count and extent flags would be if all of
 670  * the delayed refs are not processed.
 671  */
 672 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 673                              struct btrfs_root *root, u64 bytenr,
 674                              u64 num_bytes, u64 *refs, u64 *flags)
 675 {
 676         struct btrfs_delayed_ref_head *head;
 677         struct btrfs_delayed_ref_root *delayed_refs;
 678         struct btrfs_path *path;
 679         struct btrfs_extent_item *ei;
 680         struct extent_buffer *leaf;
 681         struct btrfs_key key;
 682         u32 item_size;
 683         u64 num_refs;
 684         u64 extent_flags;
 685         int ret;
 686
 687         path = btrfs_alloc_path();
 688         if (!path)
 689                 return -ENOMEM;
 690
 691         key.objectid = bytenr;
 692         key.type = BTRFS_EXTENT_ITEM_KEY;
 693         key.offset = num_bytes;
 694         if (!trans) {
 695                 path->skip_locking = 1;
 696                 path->search_commit_root = 1;
 697         }
 698 again:
 699         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 700                                 &key, path, 0, 0);
 701         if (ret < 0)
 702                 goto out_free;
 703
 704         if (ret == 0) {
 705                 leaf = path->nodes[0];
 706                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 707                 if (item_size >= sizeof(*ei)) {
 708                         ei = btrfs_item_ptr(leaf, path->slots[0],
 709                                             struct btrfs_extent_item);
 710                         num_refs = btrfs_extent_refs(leaf, ei);
 711                         extent_flags = btrfs_extent_flags(leaf, ei);
 712                 } else {
 713 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 714                         struct btrfs_extent_item_v0 *ei0;
 715                         BUG_ON(item_size != sizeof(*ei0));
 716                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 717                                              struct btrfs_extent_item_v0);
 718                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 719                         /* FIXME: this isn't correct for data */
 720                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 721 #else
 722                         BUG();
 723 #endif
 724                 }
 725                 BUG_ON(num_refs == 0);
 726         } else {
 727                 num_refs = 0;
 728                 extent_flags = 0;
 729                 ret = 0;
 730         }
 731
 732         if (!trans)
 733                 goto out;
 734
 735         delayed_refs = &trans->transaction->delayed_refs;
 736         spin_lock(&delayed_refs->lock);
 737         head = btrfs_find_delayed_ref_head(trans, bytenr);
 738         if (head) {
 739                 if (!mutex_trylock(&head->mutex)) {
 740                         atomic_inc(&head->node.refs);
 741                         spin_unlock(&delayed_refs->lock);
 742
 743                         btrfs_release_path(root->fs_info->extent_root, path);
 744
 745                         mutex_lock(&head->mutex);
 746                         mutex_unlock(&head->mutex);
 747                         btrfs_put_delayed_ref(&head->node);
 748                         goto again;
 749                 }
 750                 if (head->extent_op && head->extent_op->update_flags)
 751                         extent_flags |= head->extent_op->flags_to_set;
 752                 else
 753                         BUG_ON(num_refs == 0);
 754
 755                 num_refs += head->node.ref_mod;
 756                 mutex_unlock(&head->mutex);
 757         }
 758         spin_unlock(&delayed_refs->lock);
 759 out:
 760         WARN_ON(num_refs == 0);
 761         if (refs)
 762                 *refs = num_refs;
 763         if (flags)
 764                 *flags = extent_flags;
 765 out_free:
 766         btrfs_free_path(path);
 767         return ret;
 768 }
 769
 770 /*
 771  * Back reference rules.  Back refs have three main goals:
 772  *
 773  * 1) differentiate between all holders of references to an extent so that
 774  *    when a reference is dropped we can make sure it was a valid reference
 775  *    before freeing the extent.
 776  *
 777  * 2) Provide enough information to quickly find the holders of an extent
 778  *    if we notice a given block is corrupted or bad.
 779  *
 780  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 781  *    maintenance.  This is actually the same as #2, but with a slightly
 782  *    different use case.
 783  *
 784  * There are two kinds of back refs. The implicit back refs is optimized
 785  * for pointers in non-shared tree blocks. For a given pointer in a block,
 786  * back refs of this kind provide information about the block's owner tree
 787  * and the pointer's key. These information allow us to find the block by
 788  * b-tree searching. The full back refs is for pointers in tree blocks not
 789  * referenced by their owner trees. The location of tree block is recorded
 790  * in the back refs. Actually the full back refs is generic, and can be
 791  * used in all cases the implicit back refs is used. The major shortcoming
 792  * of the full back refs is its overhead. Every time a tree block gets
 793  * COWed, we have to update back refs entry for all pointers in it.
 794  *
 795  * For a newly allocated tree block, we use implicit back refs for
 796  * pointers in it. This means most tree related operations only involve
 797  * implicit back refs. For a tree block created in old transaction, the
 798  * only way to drop a reference to it is COW it. So we can detect the
 799  * event that tree block loses its owner tree's reference and do the
 800  * back refs conversion.
 801  *
 802  * When a tree block is COW'd through a tree, there are four cases:
 803  *
 804  * The reference count of the block is one and the tree is the block's
 805  * owner tree. Nothing to do in this case.
 806  *
 807  * The reference count of the block is one and the tree is not the
 808  * block's owner tree. In this case, full back refs is used for pointers
 809  * in the block. Remove these full back refs, add implicit back refs for
 810  * every pointers in the new block.
 811  *
 812  * The reference count of the block is greater than one and the tree is
 813  * the block's owner tree. In this case, implicit back refs is used for
 814  * pointers in the block. Add full back refs for every pointers in the
 815  * block, increase lower level extents' reference counts. The original
 816  * implicit back refs are entailed to the new block.
 817  *
 818  * The reference count of the block is greater than one and the tree is
 819  * not the block's owner tree. Add implicit back refs for every pointer in
 820  * the new block, increase lower level extents' reference count.
 821  *
 822  * Back Reference Key composing:
 823  *
 824  * The key objectid corresponds to the first byte in the extent,
 825  * The key type is used to differentiate between types of back refs.
 826  * There are different meanings of the key offset for different types
 827  * of back refs.
 828  *
 829  * File extents can be referenced by:
 830  *
 831  * - multiple snapshots, subvolumes, or different generations in one subvol
 832  * - different files inside a single subvolume
 833  * - different offsets inside a file (bookend extents in file.c)
 834  *
 835  * The extent ref structure for the implicit back refs has fields for:
 836  *
 837  * - Objectid of the subvolume root
 838  * - objectid of the file holding the reference
 839  * - original offset in the file
 840  * - how many bookend extents
 841  *
 842  * The key offset for the implicit back refs is hash of the first
 843  * three fields.
 844  *
 845  * The extent ref structure for the full back refs has field for:
 846  *
 847  * - number of pointers in the tree leaf
 848  *
 849  * The key offset for the implicit back refs is the first byte of
 850  * the tree leaf
 851  *
 852  * When a file extent is allocated, The implicit back refs is used.
 853  * the fields are filled in:
 854  *
 855  *     (root_key.objectid, inode objectid, offset in file, 1)
 856  *
 857  * When a file extent is removed file truncation, we find the
 858  * corresponding implicit back refs and check the following fields:
 859  *
 860  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 861  *
 862  * Btree extents can be referenced by:
 863  *
 864  * - Different subvolumes
 865  *
 866  * Both the implicit back refs and the full back refs for tree blocks
 867  * only consist of key. The key offset for the implicit back refs is
 868  * objectid of block's owner tree. The key offset for the full back refs
 869  * is the first byte of parent block.
 870  *
 871  * When implicit back refs is used, information about the lowest key and
 872  * level of the tree block are required. These information are stored in
 873  * tree block info structure.
 874  */
 875
 876 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 877 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 878                                   struct btrfs_root *root,
 879                                   struct btrfs_path *path,
 880                                   u64 owner, u32 extra_size)
 881 {
 882         struct btrfs_extent_item *item;
 883         struct btrfs_extent_item_v0 *ei0;
 884         struct btrfs_extent_ref_v0 *ref0;
 885         struct btrfs_tree_block_info *bi;
 886         struct extent_buffer *leaf;
 887         struct btrfs_key key;
 888         struct btrfs_key found_key;
 889         u32 new_size = sizeof(*item);
 890         u64 refs;
 891         int ret;
 892
 893         leaf = path->nodes[0];
 894         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 895
 896         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 897         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 898                              struct btrfs_extent_item_v0);
 899         refs = btrfs_extent_refs_v0(leaf, ei0);
 900
 901         if (owner == (u64)-1) {
 902                 while (1) {
 903                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 904                                 ret = btrfs_next_leaf(root, path);
 905                                 if (ret < 0)
 906                                         return ret;
 907                                 BUG_ON(ret > 0);
 908                                 leaf = path->nodes[0];
 909                         }
 910                         btrfs_item_key_to_cpu(leaf, &found_key,
 911                                               path->slots[0]);
 912                         BUG_ON(key.objectid != found_key.objectid);
 913                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 914                                 path->slots[0]++;
 915                                 continue;
 916                         }
 917                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 918                                               struct btrfs_extent_ref_v0);
 919                         owner = btrfs_ref_objectid_v0(leaf, ref0);
 920                         break;
 921                 }
 922         }
 923         btrfs_release_path(root, path);
 924
 925         if (owner < BTRFS_FIRST_FREE_OBJECTID)
 926                 new_size += sizeof(*bi);
 927
 928         new_size -= sizeof(*ei0);
 929         ret = btrfs_search_slot(trans, root, &key, path,
 930                                 new_size + extra_size, 1);
 931         if (ret < 0)
 932                 return ret;
 933         BUG_ON(ret);
 934
 935         ret = btrfs_extend_item(trans, root, path, new_size);
 936         BUG_ON(ret);
 937
 938         leaf = path->nodes[0];
 939         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 940         btrfs_set_extent_refs(leaf, item, refs);
 941         /* FIXME: get real generation */
 942         btrfs_set_extent_generation(leaf, item, 0);
 943         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 944                 btrfs_set_extent_flags(leaf, item,
 945                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
 946                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
 947                 bi = (struct btrfs_tree_block_info *)(item + 1);
 948                 /* FIXME: get first key of the block */
 949                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
 950                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
 951         } else {
 952                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
 953         }
 954         btrfs_mark_buffer_dirty(leaf);
 955         return 0;
 956 }
 957 #endif
 958
 959 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
 960 {
 961         u32 high_crc = ~(u32)0;
 962         u32 low_crc = ~(u32)0;
 963         __le64 lenum;
 964
 965         lenum = cpu_to_le64(root_objectid);
 966         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
 967         lenum = cpu_to_le64(owner);
 968         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
 969         lenum = cpu_to_le64(offset);
 970         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
 971
 972         return ((u64)high_crc << 31) ^ (u64)low_crc;
 973 }
 974
 975 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
 976                                      struct btrfs_extent_data_ref *ref)
 977 {
 978         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
 979                                     btrfs_extent_data_ref_objectid(leaf, ref),
 980                                     btrfs_extent_data_ref_offset(leaf, ref));
 981 }
 982
 983 static int match_extent_data_ref(struct extent_buffer *leaf,
 984                                  struct btrfs_extent_data_ref *ref,
 985                                  u64 root_objectid, u64 owner, u64 offset)
 986 {
 987         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
 988             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
 989             btrfs_extent_data_ref_offset(leaf, ref) != offset)
 990                 return 0;
 991         return 1;
 992 }
 993
 994 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
 995                                            struct btrfs_root *root,
 996                                            struct btrfs_path *path,
 997                                            u64 bytenr, u64 parent,
 998                                            u64 root_objectid,
 999                                            u64 owner, u64 offset)
1000 {
1001         struct btrfs_key key;
1002         struct btrfs_extent_data_ref *ref;
1003         struct extent_buffer *leaf;
1004         u32 nritems;
1005         int ret;
1006         int recow;
1007         int err = -ENOENT;
1008
1009         key.objectid = bytenr;
1010         if (parent) {
1011                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1012                 key.offset = parent;
1013         } else {
1014                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1015                 key.offset = hash_extent_data_ref(root_objectid,
1016                                                   owner, offset);
1017         }
1018 again:
1019         recow = 0;
1020         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1021         if (ret < 0) {
1022                 err = ret;
1023                 goto fail;
1024         }
1025
1026         if (parent) {
1027                 if (!ret)
1028                         return 0;
1029 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1030                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1031                 btrfs_release_path(root, path);
1032                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1033                 if (ret < 0) {
1034                         err = ret;
1035                         goto fail;
1036                 }
1037                 if (!ret)
1038                         return 0;
1039 #endif
1040                 goto fail;
1041         }
1042
1043         leaf = path->nodes[0];
1044         nritems = btrfs_header_nritems(leaf);
1045         while (1) {
1046                 if (path->slots[0] >= nritems) {
1047                         ret = btrfs_next_leaf(root, path);
1048                         if (ret < 0)
1049                                 err = ret;
1050                         if (ret)
1051                                 goto fail;
1052
1053                         leaf = path->nodes[0];
1054                         nritems = btrfs_header_nritems(leaf);
1055                         recow = 1;
1056                 }
1057
1058                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1059                 if (key.objectid != bytenr ||
1060                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1061                         goto fail;
1062
1063                 ref = btrfs_item_ptr(leaf, path->slots[0],
1064                                      struct btrfs_extent_data_ref);
1065
1066                 if (match_extent_data_ref(leaf, ref, root_objectid,
1067                                           owner, offset)) {
1068                         if (recow) {
1069                                 btrfs_release_path(root, path);
1070                                 goto again;
1071                         }
1072                         err = 0;
1073                         break;
1074                 }
1075                 path->slots[0]++;
1076         }
1077 fail:
1078         return err;
1079 }
1080
1081 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1082                                            struct btrfs_root *root,
1083                                            struct btrfs_path *path,
1084                                            u64 bytenr, u64 parent,
1085                                            u64 root_objectid, u64 owner,
1086                                            u64 offset, int refs_to_add)
1087 {
1088         struct btrfs_key key;
1089         struct extent_buffer *leaf;
1090         u32 size;
1091         u32 num_refs;
1092         int ret;
1093
1094         key.objectid = bytenr;
1095         if (parent) {
1096                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1097                 key.offset = parent;
1098                 size = sizeof(struct btrfs_shared_data_ref);
1099         } else {
1100                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1101                 key.offset = hash_extent_data_ref(root_objectid,
1102                                                   owner, offset);
1103                 size = sizeof(struct btrfs_extent_data_ref);
1104         }
1105
1106         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1107         if (ret && ret != -EEXIST)
1108                 goto fail;
1109
1110         leaf = path->nodes[0];
1111         if (parent) {
1112                 struct btrfs_shared_data_ref *ref;
1113                 ref = btrfs_item_ptr(leaf, path->slots[0],
1114                                      struct btrfs_shared_data_ref);
1115                 if (ret == 0) {
1116                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1117                 } else {
1118                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1119                         num_refs += refs_to_add;
1120                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1121                 }
1122         } else {
1123                 struct btrfs_extent_data_ref *ref;
1124                 while (ret == -EEXIST) {
1125                         ref = btrfs_item_ptr(leaf, path->slots[0],
1126                                              struct btrfs_extent_data_ref);
1127                         if (match_extent_data_ref(leaf, ref, root_objectid,
1128                                                   owner, offset))
1129                                 break;
1130                         btrfs_release_path(root, path);
1131                         key.offset++;
1132                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1133                                                       size);
1134                         if (ret && ret != -EEXIST)
1135                                 goto fail;
1136
1137                         leaf = path->nodes[0];
1138                 }
1139                 ref = btrfs_item_ptr(leaf, path->slots[0],
1140                                      struct btrfs_extent_data_ref);
1141                 if (ret == 0) {
1142                         btrfs_set_extent_data_ref_root(leaf, ref,
1143                                                        root_objectid);
1144                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1145                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1146                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1147                 } else {
1148                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1149                         num_refs += refs_to_add;
1150                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1151                 }
1152         }
1153         btrfs_mark_buffer_dirty(leaf);
1154         ret = 0;
1155 fail:
1156         btrfs_release_path(root, path);
1157         return ret;
1158 }
1159
1160 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1161                                            struct btrfs_root *root,
1162                                            struct btrfs_path *path,
1163                                            int refs_to_drop)
1164 {
1165         struct btrfs_key key;
1166         struct btrfs_extent_data_ref *ref1 = NULL;
1167         struct btrfs_shared_data_ref *ref2 = NULL;
1168         struct extent_buffer *leaf;
1169         u32 num_refs = 0;
1170         int ret = 0;
1171
1172         leaf = path->nodes[0];
1173         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1174
1175         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1176                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1177                                       struct btrfs_extent_data_ref);
1178                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1179         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1180                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1181                                       struct btrfs_shared_data_ref);
1182                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1183 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1184         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1185                 struct btrfs_extent_ref_v0 *ref0;
1186                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1187                                       struct btrfs_extent_ref_v0);
1188                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1189 #endif
1190         } else {
1191                 BUG();
1192         }
1193
1194         BUG_ON(num_refs < refs_to_drop);
1195         num_refs -= refs_to_drop;
1196
1197         if (num_refs == 0) {
1198                 ret = btrfs_del_item(trans, root, path);
1199         } else {
1200                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1201                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1202                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1203                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1204 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1205                 else {
1206                         struct btrfs_extent_ref_v0 *ref0;
1207                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1208                                         struct btrfs_extent_ref_v0);
1209                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1210                 }
1211 #endif
1212                 btrfs_mark_buffer_dirty(leaf);
1213         }
1214         return ret;
1215 }
1216
1217 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1218                                           struct btrfs_path *path,
1219                                           struct btrfs_extent_inline_ref *iref)
1220 {
1221         struct btrfs_key key;
1222         struct extent_buffer *leaf;
1223         struct btrfs_extent_data_ref *ref1;
1224         struct btrfs_shared_data_ref *ref2;
1225         u32 num_refs = 0;
1226
1227         leaf = path->nodes[0];
1228         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1229         if (iref) {
1230                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1231                     BTRFS_EXTENT_DATA_REF_KEY) {
1232                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1233                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1234                 } else {
1235                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1236                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1237                 }
1238         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1239                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1240                                       struct btrfs_extent_data_ref);
1241                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1242         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1243                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1244                                       struct btrfs_shared_data_ref);
1245                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1246 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1247         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1248                 struct btrfs_extent_ref_v0 *ref0;
1249                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1250                                       struct btrfs_extent_ref_v0);
1251                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1252 #endif
1253         } else {
1254                 WARN_ON(1);
1255         }
1256         return num_refs;
1257 }
1258
1259 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1260                                           struct btrfs_root *root,
1261                                           struct btrfs_path *path,
1262                                           u64 bytenr, u64 parent,
1263                                           u64 root_objectid)
1264 {
1265         struct btrfs_key key;
1266         int ret;
1267
1268         key.objectid = bytenr;
1269         if (parent) {
1270                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1271                 key.offset = parent;
1272         } else {
1273                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1274                 key.offset = root_objectid;
1275         }
1276
1277         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1278         if (ret > 0)
1279                 ret = -ENOENT;
1280 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1281         if (ret == -ENOENT && parent) {
1282                 btrfs_release_path(root, path);
1283                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1284                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1285                 if (ret > 0)
1286                         ret = -ENOENT;
1287         }
1288 #endif
1289         return ret;
1290 }
1291
1292 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1293                                           struct btrfs_root *root,
1294                                           struct btrfs_path *path,
1295                                           u64 bytenr, u64 parent,
1296                                           u64 root_objectid)
1297 {
1298         struct btrfs_key key;
1299         int ret;
1300
1301         key.objectid = bytenr;
1302         if (parent) {
1303                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1304                 key.offset = parent;
1305         } else {
1306                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1307                 key.offset = root_objectid;
1308         }
1309
1310         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1311         btrfs_release_path(root, path);
1312         return ret;
1313 }
1314
1315 static inline int extent_ref_type(u64 parent, u64 owner)
1316 {
1317         int type;
1318         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1319                 if (parent > 0)
1320                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1321                 else
1322                         type = BTRFS_TREE_BLOCK_REF_KEY;
1323         } else {
1324                 if (parent > 0)
1325                         type = BTRFS_SHARED_DATA_REF_KEY;
1326                 else
1327                         type = BTRFS_EXTENT_DATA_REF_KEY;
1328         }
1329         return type;
1330 }
1331
1332 static int find_next_key(struct btrfs_path *path, int level,
1333                          struct btrfs_key *key)
1334
1335 {
1336         for (; level < BTRFS_MAX_LEVEL; level++) {
1337                 if (!path->nodes[level])
1338                         break;
1339                 if (path->slots[level] + 1 >=
1340                     btrfs_header_nritems(path->nodes[level]))
1341                         continue;
1342                 if (level == 0)
1343                         btrfs_item_key_to_cpu(path->nodes[level], key,
1344                                               path->slots[level] + 1);
1345                 else
1346                         btrfs_node_key_to_cpu(path->nodes[level], key,
1347                                               path->slots[level] + 1);
1348                 return 0;
1349         }
1350         return 1;
1351 }
1352
1353 /*
1354  * look for inline back ref. if back ref is found, *ref_ret is set
1355  * to the address of inline back ref, and 0 is returned.
1356  *
1357  * if back ref isn't found, *ref_ret is set to the address where it
1358  * should be inserted, and -ENOENT is returned.
1359  *
1360  * if insert is true and there are too many inline back refs, the path
1361  * points to the extent item, and -EAGAIN is returned.
1362  *
1363  * NOTE: inline back refs are ordered in the same way that back ref
1364  *       items in the tree are ordered.
1365  */
1366 static noinline_for_stack
1367 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1368                                  struct btrfs_root *root,
1369                                  struct btrfs_path *path,
1370                                  struct btrfs_extent_inline_ref **ref_ret,
1371                                  u64 bytenr, u64 num_bytes,
1372                                  u64 parent, u64 root_objectid,
1373                                  u64 owner, u64 offset, int insert)
1374 {
1375         struct btrfs_key key;
1376         struct extent_buffer *leaf;
1377         struct btrfs_extent_item *ei;
1378         struct btrfs_extent_inline_ref *iref;
1379         u64 flags;
1380         u64 item_size;
1381         unsigned long ptr;
1382         unsigned long end;
1383         int extra_size;
1384         int type;
1385         int want;
1386         int ret;
1387         int err = 0;
1388
1389         key.objectid = bytenr;
1390         key.type = BTRFS_EXTENT_ITEM_KEY;
1391         key.offset = num_bytes;
1392
1393         want = extent_ref_type(parent, owner);
1394         if (insert) {
1395                 extra_size = btrfs_extent_inline_ref_size(want);
1396                 path->keep_locks = 1;
1397         } else
1398                 extra_size = -1;
1399         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1400         if (ret < 0) {
1401                 err = ret;
1402                 goto out;
1403         }
1404         BUG_ON(ret);
1405
1406         leaf = path->nodes[0];
1407         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1408 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1409         if (item_size < sizeof(*ei)) {
1410                 if (!insert) {
1411                         err = -ENOENT;
1412                         goto out;
1413                 }
1414                 ret = convert_extent_item_v0(trans, root, path, owner,
1415                                              extra_size);
1416                 if (ret < 0) {
1417                         err = ret;
1418                         goto out;
1419                 }
1420                 leaf = path->nodes[0];
1421                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1422         }
1423 #endif
1424         BUG_ON(item_size < sizeof(*ei));
1425
1426         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1427         flags = btrfs_extent_flags(leaf, ei);
1428
1429         ptr = (unsigned long)(ei + 1);
1430         end = (unsigned long)ei + item_size;
1431
1432         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1433                 ptr += sizeof(struct btrfs_tree_block_info);
1434                 BUG_ON(ptr > end);
1435         } else {
1436                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1437         }
1438
1439         err = -ENOENT;
1440         while (1) {
1441                 if (ptr >= end) {
1442                         WARN_ON(ptr > end);
1443                         break;
1444                 }
1445                 iref = (struct btrfs_extent_inline_ref *)ptr;
1446                 type = btrfs_extent_inline_ref_type(leaf, iref);
1447                 if (want < type)
1448                         break;
1449                 if (want > type) {
1450                         ptr += btrfs_extent_inline_ref_size(type);
1451                         continue;
1452                 }
1453
1454                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1455                         struct btrfs_extent_data_ref *dref;
1456                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1457                         if (match_extent_data_ref(leaf, dref, root_objectid,
1458                                                   owner, offset)) {
1459                                 err = 0;
1460                                 break;
1461                         }
1462                         if (hash_extent_data_ref_item(leaf, dref) <
1463                             hash_extent_data_ref(root_objectid, owner, offset))
1464                                 break;
1465                 } else {
1466                         u64 ref_offset;
1467                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1468                         if (parent > 0) {
1469                                 if (parent == ref_offset) {
1470                                         err = 0;
1471                                         break;
1472                                 }
1473                                 if (ref_offset < parent)
1474                                         break;
1475                         } else {
1476                                 if (root_objectid == ref_offset) {
1477                                         err = 0;
1478                                         break;
1479                                 }
1480                                 if (ref_offset < root_objectid)
1481                                         break;
1482                         }
1483                 }
1484                 ptr += btrfs_extent_inline_ref_size(type);
1485         }
1486         if (err == -ENOENT && insert) {
1487                 if (item_size + extra_size >=
1488                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1489                         err = -EAGAIN;
1490                         goto out;
1491                 }
1492                 /*
1493                  * To add new inline back ref, we have to make sure
1494                  * there is no corresponding back ref item.
1495                  * For simplicity, we just do not add new inline back
1496                  * ref if there is any kind of item for this block
1497                  */
1498                 if (find_next_key(path, 0, &key) == 0 &&
1499                     key.objectid == bytenr &&
1500                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1501                         err = -EAGAIN;
1502                         goto out;
1503                 }
1504         }
1505         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1506 out:
1507         if (insert) {
1508                 path->keep_locks = 0;
1509                 btrfs_unlock_up_safe(path, 1);
1510         }
1511         return err;
1512 }
1513
1514 /*
1515  * helper to add new inline back ref
1516  */
1517 static noinline_for_stack
1518 int setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1519                                 struct btrfs_root *root,
1520                                 struct btrfs_path *path,
1521                                 struct btrfs_extent_inline_ref *iref,
1522                                 u64 parent, u64 root_objectid,
1523                                 u64 owner, u64 offset, int refs_to_add,
1524                                 struct btrfs_delayed_extent_op *extent_op)
1525 {
1526         struct extent_buffer *leaf;
1527         struct btrfs_extent_item *ei;
1528         unsigned long ptr;
1529         unsigned long end;
1530         unsigned long item_offset;
1531         u64 refs;
1532         int size;
1533         int type;
1534         int ret;
1535
1536         leaf = path->nodes[0];
1537         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1538         item_offset = (unsigned long)iref - (unsigned long)ei;
1539
1540         type = extent_ref_type(parent, owner);
1541         size = btrfs_extent_inline_ref_size(type);
1542
1543         ret = btrfs_extend_item(trans, root, path, size);
1544         BUG_ON(ret);
1545
1546         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1547         refs = btrfs_extent_refs(leaf, ei);
1548         refs += refs_to_add;
1549         btrfs_set_extent_refs(leaf, ei, refs);
1550         if (extent_op)
1551                 __run_delayed_extent_op(extent_op, leaf, ei);
1552
1553         ptr = (unsigned long)ei + item_offset;
1554         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1555         if (ptr < end - size)
1556                 memmove_extent_buffer(leaf, ptr + size, ptr,
1557                                       end - size - ptr);
1558
1559         iref = (struct btrfs_extent_inline_ref *)ptr;
1560         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1561         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1562                 struct btrfs_extent_data_ref *dref;
1563                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1564                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1565                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1566                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1567                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1568         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1569                 struct btrfs_shared_data_ref *sref;
1570                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1571                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1572                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1573         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1574                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1575         } else {
1576                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1577         }
1578         btrfs_mark_buffer_dirty(leaf);
1579         return 0;
1580 }
1581
1582 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1583                                  struct btrfs_root *root,
1584                                  struct btrfs_path *path,
1585                                  struct btrfs_extent_inline_ref **ref_ret,
1586                                  u64 bytenr, u64 num_bytes, u64 parent,
1587                                  u64 root_objectid, u64 owner, u64 offset)
1588 {
1589         int ret;
1590
1591         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1592                                            bytenr, num_bytes, parent,
1593                                            root_objectid, owner, offset, 0);
1594         if (ret != -ENOENT)
1595                 return ret;
1596
1597         btrfs_release_path(root, path);
1598         *ref_ret = NULL;
1599
1600         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1601                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1602                                             root_objectid);
1603         } else {
1604                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1605                                              root_objectid, owner, offset);
1606         }
1607         return ret;
1608 }
1609
1610 /*
1611  * helper to update/remove inline back ref
1612  */
1613 static noinline_for_stack
1614 int update_inline_extent_backref(struct btrfs_trans_handle *trans,
1615                                  struct btrfs_root *root,
1616                                  struct btrfs_path *path,
1617                                  struct btrfs_extent_inline_ref *iref,
1618                                  int refs_to_mod,
1619                                  struct btrfs_delayed_extent_op *extent_op)
1620 {
1621         struct extent_buffer *leaf;
1622         struct btrfs_extent_item *ei;
1623         struct btrfs_extent_data_ref *dref = NULL;
1624         struct btrfs_shared_data_ref *sref = NULL;
1625         unsigned long ptr;
1626         unsigned long end;
1627         u32 item_size;
1628         int size;
1629         int type;
1630         int ret;
1631         u64 refs;
1632
1633         leaf = path->nodes[0];
1634         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1635         refs = btrfs_extent_refs(leaf, ei);
1636         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1637         refs += refs_to_mod;
1638         btrfs_set_extent_refs(leaf, ei, refs);
1639         if (extent_op)
1640                 __run_delayed_extent_op(extent_op, leaf, ei);
1641
1642         type = btrfs_extent_inline_ref_type(leaf, iref);
1643
1644         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1645                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1646                 refs = btrfs_extent_data_ref_count(leaf, dref);
1647         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1648                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1649                 refs = btrfs_shared_data_ref_count(leaf, sref);
1650         } else {
1651                 refs = 1;
1652                 BUG_ON(refs_to_mod != -1);
1653         }
1654
1655         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1656         refs += refs_to_mod;
1657
1658         if (refs > 0) {
1659                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1660                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1661                 else
1662                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1663         } else {
1664                 size =  btrfs_extent_inline_ref_size(type);
1665                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1666                 ptr = (unsigned long)iref;
1667                 end = (unsigned long)ei + item_size;
1668                 if (ptr + size < end)
1669                         memmove_extent_buffer(leaf, ptr, ptr + size,
1670                                               end - ptr - size);
1671                 item_size -= size;
1672                 ret = btrfs_truncate_item(trans, root, path, item_size, 1);
1673                 BUG_ON(ret);
1674         }
1675         btrfs_mark_buffer_dirty(leaf);
1676         return 0;
1677 }
1678
1679 static noinline_for_stack
1680 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1681                                  struct btrfs_root *root,
1682                                  struct btrfs_path *path,
1683                                  u64 bytenr, u64 num_bytes, u64 parent,
1684                                  u64 root_objectid, u64 owner,
1685                                  u64 offset, int refs_to_add,
1686                                  struct btrfs_delayed_extent_op *extent_op)
1687 {
1688         struct btrfs_extent_inline_ref *iref;
1689         int ret;
1690
1691         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1692                                            bytenr, num_bytes, parent,
1693                                            root_objectid, owner, offset, 1);
1694         if (ret == 0) {
1695                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1696                 ret = update_inline_extent_backref(trans, root, path, iref,
1697                                                    refs_to_add, extent_op);
1698         } else if (ret == -ENOENT) {
1699                 ret = setup_inline_extent_backref(trans, root, path, iref,
1700                                                   parent, root_objectid,
1701                                                   owner, offset, refs_to_add,
1702                                                   extent_op);
1703         }
1704         return ret;
1705 }
1706
1707 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1708                                  struct btrfs_root *root,
1709                                  struct btrfs_path *path,
1710                                  u64 bytenr, u64 parent, u64 root_objectid,
1711                                  u64 owner, u64 offset, int refs_to_add)
1712 {
1713         int ret;
1714         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1715                 BUG_ON(refs_to_add != 1);
1716                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1717                                             parent, root_objectid);
1718         } else {
1719                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1720                                              parent, root_objectid,
1721                                              owner, offset, refs_to_add);
1722         }
1723         return ret;
1724 }
1725
1726 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1727                                  struct btrfs_root *root,
1728                                  struct btrfs_path *path,
1729                                  struct btrfs_extent_inline_ref *iref,
1730                                  int refs_to_drop, int is_data)
1731 {
1732         int ret;
1733
1734         BUG_ON(!is_data && refs_to_drop != 1);
1735         if (iref) {
1736                 ret = update_inline_extent_backref(trans, root, path, iref,
1737                                                    -refs_to_drop, NULL);
1738         } else if (is_data) {
1739                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1740         } else {
1741                 ret = btrfs_del_item(trans, root, path);
1742         }
1743         return ret;
1744 }
1745
1746 static void btrfs_issue_discard(struct block_device *bdev,
1747                                 u64 start, u64 len)
1748 {
1749         blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_KERNEL,
1750                         BLKDEV_IFL_WAIT | BLKDEV_IFL_BARRIER);
1751 }
1752
1753 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1754                                 u64 num_bytes)
1755 {
1756         int ret;
1757         u64 map_length = num_bytes;
1758         struct btrfs_multi_bio *multi = NULL;
1759
1760         if (!btrfs_test_opt(root, DISCARD))
1761                 return 0;
1762
1763         /* Tell the block device(s) that the sectors can be discarded */
1764         ret = btrfs_map_block(&root->fs_info->mapping_tree, READ,
1765                               bytenr, &map_length, &multi, 0);
1766         if (!ret) {
1767                 struct btrfs_bio_stripe *stripe = multi->stripes;
1768                 int i;
1769
1770                 if (map_length > num_bytes)
1771                         map_length = num_bytes;
1772
1773                 for (i = 0; i < multi->num_stripes; i++, stripe++) {
1774                         btrfs_issue_discard(stripe->dev->bdev,
1775                                             stripe->physical,
1776                                             map_length);
1777                 }
1778                 kfree(multi);
1779         }
1780
1781         return ret;
1782 }
1783
1784 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1785                          struct btrfs_root *root,
1786                          u64 bytenr, u64 num_bytes, u64 parent,
1787                          u64 root_objectid, u64 owner, u64 offset)
1788 {
1789         int ret;
1790         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1791                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1792
1793         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1794                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
1795                                         parent, root_objectid, (int)owner,
1796                                         BTRFS_ADD_DELAYED_REF, NULL);
1797         } else {
1798                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
1799                                         parent, root_objectid, owner, offset,
1800                                         BTRFS_ADD_DELAYED_REF, NULL);
1801         }
1802         return ret;
1803 }
1804
1805 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1806                                   struct btrfs_root *root,
1807                                   u64 bytenr, u64 num_bytes,
1808                                   u64 parent, u64 root_objectid,
1809                                   u64 owner, u64 offset, int refs_to_add,
1810                                   struct btrfs_delayed_extent_op *extent_op)
1811 {
1812         struct btrfs_path *path;
1813         struct extent_buffer *leaf;
1814         struct btrfs_extent_item *item;
1815         u64 refs;
1816         int ret;
1817         int err = 0;
1818
1819         path = btrfs_alloc_path();
1820         if (!path)
1821                 return -ENOMEM;
1822
1823         path->reada = 1;
1824         path->leave_spinning = 1;
1825         /* this will setup the path even if it fails to insert the back ref */
1826         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1827                                            path, bytenr, num_bytes, parent,
1828                                            root_objectid, owner, offset,
1829                                            refs_to_add, extent_op);
1830         if (ret == 0)
1831                 goto out;
1832
1833         if (ret != -EAGAIN) {
1834                 err = ret;
1835                 goto out;
1836         }
1837
1838         leaf = path->nodes[0];
1839         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1840         refs = btrfs_extent_refs(leaf, item);
1841         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1842         if (extent_op)
1843                 __run_delayed_extent_op(extent_op, leaf, item);
1844
1845         btrfs_mark_buffer_dirty(leaf);
1846         btrfs_release_path(root->fs_info->extent_root, path);
1847
1848         path->reada = 1;
1849         path->leave_spinning = 1;
1850
1851         /* now insert the actual backref */
1852         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1853                                     path, bytenr, parent, root_objectid,
1854                                     owner, offset, refs_to_add);
1855         BUG_ON(ret);
1856 out:
1857         btrfs_free_path(path);
1858         return err;
1859 }
1860
1861 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1862                                 struct btrfs_root *root,
1863                                 struct btrfs_delayed_ref_node *node,
1864                                 struct btrfs_delayed_extent_op *extent_op,
1865                                 int insert_reserved)
1866 {
1867         int ret = 0;
1868         struct btrfs_delayed_data_ref *ref;
1869         struct btrfs_key ins;
1870         u64 parent = 0;
1871         u64 ref_root = 0;
1872         u64 flags = 0;
1873
1874         ins.objectid = node->bytenr;
1875         ins.offset = node->num_bytes;
1876         ins.type = BTRFS_EXTENT_ITEM_KEY;
1877
1878         ref = btrfs_delayed_node_to_data_ref(node);
1879         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1880                 parent = ref->parent;
1881         else
1882                 ref_root = ref->root;
1883
1884         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1885                 if (extent_op) {
1886                         BUG_ON(extent_op->update_key);
1887                         flags |= extent_op->flags_to_set;
1888                 }
1889                 ret = alloc_reserved_file_extent(trans, root,
1890                                                  parent, ref_root, flags,
1891                                                  ref->objectid, ref->offset,
1892                                                  &ins, node->ref_mod);
1893         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1894                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1895                                              node->num_bytes, parent,
1896                                              ref_root, ref->objectid,
1897                                              ref->offset, node->ref_mod,
1898                                              extent_op);
1899         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1900                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1901                                           node->num_bytes, parent,
1902                                           ref_root, ref->objectid,
1903                                           ref->offset, node->ref_mod,
1904                                           extent_op);
1905         } else {
1906                 BUG();
1907         }
1908         return ret;
1909 }
1910
1911 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1912                                     struct extent_buffer *leaf,
1913                                     struct btrfs_extent_item *ei)
1914 {
1915         u64 flags = btrfs_extent_flags(leaf, ei);
1916         if (extent_op->update_flags) {
1917                 flags |= extent_op->flags_to_set;
1918                 btrfs_set_extent_flags(leaf, ei, flags);
1919         }
1920
1921         if (extent_op->update_key) {
1922                 struct btrfs_tree_block_info *bi;
1923                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1924                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1925                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1926         }
1927 }
1928
1929 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1930                                  struct btrfs_root *root,
1931                                  struct btrfs_delayed_ref_node *node,
1932                                  struct btrfs_delayed_extent_op *extent_op)
1933 {
1934         struct btrfs_key key;
1935         struct btrfs_path *path;
1936         struct btrfs_extent_item *ei;
1937         struct extent_buffer *leaf;
1938         u32 item_size;
1939         int ret;
1940         int err = 0;
1941
1942         path = btrfs_alloc_path();
1943         if (!path)
1944                 return -ENOMEM;
1945
1946         key.objectid = node->bytenr;
1947         key.type = BTRFS_EXTENT_ITEM_KEY;
1948         key.offset = node->num_bytes;
1949
1950         path->reada = 1;
1951         path->leave_spinning = 1;
1952         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
1953                                 path, 0, 1);
1954         if (ret < 0) {
1955                 err = ret;
1956                 goto out;
1957         }
1958         if (ret > 0) {
1959                 err = -EIO;
1960                 goto out;
1961         }
1962
1963         leaf = path->nodes[0];
1964         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1965 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1966         if (item_size < sizeof(*ei)) {
1967                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
1968                                              path, (u64)-1, 0);
1969                 if (ret < 0) {
1970                         err = ret;
1971                         goto out;
1972                 }
1973                 leaf = path->nodes[0];
1974                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1975         }
1976 #endif
1977         BUG_ON(item_size < sizeof(*ei));
1978         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1979         __run_delayed_extent_op(extent_op, leaf, ei);
1980
1981         btrfs_mark_buffer_dirty(leaf);
1982 out:
1983         btrfs_free_path(path);
1984         return err;
1985 }
1986
1987 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
1988                                 struct btrfs_root *root,
1989                                 struct btrfs_delayed_ref_node *node,
1990                                 struct btrfs_delayed_extent_op *extent_op,
1991                                 int insert_reserved)
1992 {
1993         int ret = 0;
1994         struct btrfs_delayed_tree_ref *ref;
1995         struct btrfs_key ins;
1996         u64 parent = 0;
1997         u64 ref_root = 0;
1998
1999         ins.objectid = node->bytenr;
2000         ins.offset = node->num_bytes;
2001         ins.type = BTRFS_EXTENT_ITEM_KEY;
2002
2003         ref = btrfs_delayed_node_to_tree_ref(node);
2004         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2005                 parent = ref->parent;
2006         else
2007                 ref_root = ref->root;
2008
2009         BUG_ON(node->ref_mod != 1);
2010         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2011                 BUG_ON(!extent_op || !extent_op->update_flags ||
2012                        !extent_op->update_key);
2013                 ret = alloc_reserved_tree_block(trans, root,
2014                                                 parent, ref_root,
2015                                                 extent_op->flags_to_set,
2016                                                 &extent_op->key,
2017                                                 ref->level, &ins);
2018         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2019                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2020                                              node->num_bytes, parent, ref_root,
2021                                              ref->level, 0, 1, extent_op);
2022         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2023                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2024                                           node->num_bytes, parent, ref_root,
2025                                           ref->level, 0, 1, extent_op);
2026         } else {
2027                 BUG();
2028         }
2029         return ret;
2030 }
2031
2032 /* helper function to actually process a single delayed ref entry */
2033 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2034                                struct btrfs_root *root,
2035                                struct btrfs_delayed_ref_node *node,
2036                                struct btrfs_delayed_extent_op *extent_op,
2037                                int insert_reserved)
2038 {
2039         int ret;
2040         if (btrfs_delayed_ref_is_head(node)) {
2041                 struct btrfs_delayed_ref_head *head;
2042                 /*
2043                  * we've hit the end of the chain and we were supposed
2044                  * to insert this extent into the tree.  But, it got
2045                  * deleted before we ever needed to insert it, so all
2046                  * we have to do is clean up the accounting
2047                  */
2048                 BUG_ON(extent_op);
2049                 head = btrfs_delayed_node_to_head(node);
2050                 if (insert_reserved) {
2051                         btrfs_pin_extent(root, node->bytenr,
2052                                          node->num_bytes, 1);
2053                         if (head->is_data) {
2054                                 ret = btrfs_del_csums(trans, root,
2055                                                       node->bytenr,
2056                                                       node->num_bytes);
2057                                 BUG_ON(ret);
2058                         }
2059                 }
2060                 mutex_unlock(&head->mutex);
2061                 return 0;
2062         }
2063
2064         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2065             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2066                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2067                                            insert_reserved);
2068         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2069                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2070                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2071                                            insert_reserved);
2072         else
2073                 BUG();
2074         return ret;
2075 }
2076
2077 static noinline struct btrfs_delayed_ref_node *
2078 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2079 {
2080         struct rb_node *node;
2081         struct btrfs_delayed_ref_node *ref;
2082         int action = BTRFS_ADD_DELAYED_REF;
2083 again:
2084         /*
2085          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2086          * this prevents ref count from going down to zero when
2087          * there still are pending delayed ref.
2088          */
2089         node = rb_prev(&head->node.rb_node);
2090         while (1) {
2091                 if (!node)
2092                         break;
2093                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2094                                 rb_node);
2095                 if (ref->bytenr != head->node.bytenr)
2096                         break;
2097                 if (ref->action == action)
2098                         return ref;
2099                 node = rb_prev(node);
2100         }
2101         if (action == BTRFS_ADD_DELAYED_REF) {
2102                 action = BTRFS_DROP_DELAYED_REF;
2103                 goto again;
2104         }
2105         return NULL;
2106 }
2107
2108 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2109                                        struct btrfs_root *root,
2110                                        struct list_head *cluster)
2111 {
2112         struct btrfs_delayed_ref_root *delayed_refs;
2113         struct btrfs_delayed_ref_node *ref;
2114         struct btrfs_delayed_ref_head *locked_ref = NULL;
2115         struct btrfs_delayed_extent_op *extent_op;
2116         int ret;
2117         int count = 0;
2118         int must_insert_reserved = 0;
2119
2120         delayed_refs = &trans->transaction->delayed_refs;
2121         while (1) {
2122                 if (!locked_ref) {
2123                         /* pick a new head ref from the cluster list */
2124                         if (list_empty(cluster))
2125                                 break;
2126
2127                         locked_ref = list_entry(cluster->next,
2128                                      struct btrfs_delayed_ref_head, cluster);
2129
2130                         /* grab the lock that says we are going to process
2131                          * all the refs for this head */
2132                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2133
2134                         /*
2135                          * we may have dropped the spin lock to get the head
2136                          * mutex lock, and that might have given someone else
2137                          * time to free the head.  If that's true, it has been
2138                          * removed from our list and we can move on.
2139                          */
2140                         if (ret == -EAGAIN) {
2141                                 locked_ref = NULL;
2142                                 count++;
2143                                 continue;
2144                         }
2145                 }
2146
2147                 /*
2148                  * record the must insert reserved flag before we
2149                  * drop the spin lock.
2150                  */
2151                 must_insert_reserved = locked_ref->must_insert_reserved;
2152                 locked_ref->must_insert_reserved = 0;
2153
2154                 extent_op = locked_ref->extent_op;
2155                 locked_ref->extent_op = NULL;
2156
2157                 /*
2158                  * locked_ref is the head node, so we have to go one
2159                  * node back for any delayed ref updates
2160                  */
2161                 ref = select_delayed_ref(locked_ref);
2162                 if (!ref) {
2163                         /* All delayed refs have been processed, Go ahead
2164                          * and send the head node to run_one_delayed_ref,
2165                          * so that any accounting fixes can happen
2166                          */
2167                         ref = &locked_ref->node;
2168
2169                         if (extent_op && must_insert_reserved) {
2170                                 kfree(extent_op);
2171                                 extent_op = NULL;
2172                         }
2173
2174                         if (extent_op) {
2175                                 spin_unlock(&delayed_refs->lock);
2176
2177                                 ret = run_delayed_extent_op(trans, root,
2178                                                             ref, extent_op);
2179                                 BUG_ON(ret);
2180                                 kfree(extent_op);
2181
2182                                 cond_resched();
2183                                 spin_lock(&delayed_refs->lock);
2184                                 continue;
2185                         }
2186
2187                         list_del_init(&locked_ref->cluster);
2188                         locked_ref = NULL;
2189                 }
2190
2191                 ref->in_tree = 0;
2192                 rb_erase(&ref->rb_node, &delayed_refs->root);
2193                 delayed_refs->num_entries--;
2194
2195                 spin_unlock(&delayed_refs->lock);
2196
2197                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2198                                           must_insert_reserved);
2199                 BUG_ON(ret);
2200
2201                 btrfs_put_delayed_ref(ref);
2202                 kfree(extent_op);
2203                 count++;
2204
2205                 cond_resched();
2206                 spin_lock(&delayed_refs->lock);
2207         }
2208         return count;
2209 }
2210
2211 /*
2212  * this starts processing the delayed reference count updates and
2213  * extent insertions we have queued up so far.  count can be
2214  * 0, which means to process everything in the tree at the start
2215  * of the run (but not newly added entries), or it can be some target
2216  * number you'd like to process.
2217  */
2218 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2219                            struct btrfs_root *root, unsigned long count)
2220 {
2221         struct rb_node *node;
2222         struct btrfs_delayed_ref_root *delayed_refs;
2223         struct btrfs_delayed_ref_node *ref;
2224         struct list_head cluster;
2225         int ret;
2226         int run_all = count == (unsigned long)-1;
2227         int run_most = 0;
2228
2229         if (root == root->fs_info->extent_root)
2230                 root = root->fs_info->tree_root;
2231
2232         delayed_refs = &trans->transaction->delayed_refs;
2233         INIT_LIST_HEAD(&cluster);
2234 again:
2235         spin_lock(&delayed_refs->lock);
2236         if (count == 0) {
2237                 count = delayed_refs->num_entries * 2;
2238                 run_most = 1;
2239         }
2240         while (1) {
2241                 if (!(run_all || run_most) &&
2242                     delayed_refs->num_heads_ready < 64)
2243                         break;
2244
2245                 /*
2246                  * go find something we can process in the rbtree.  We start at
2247                  * the beginning of the tree, and then build a cluster
2248                  * of refs to process starting at the first one we are able to
2249                  * lock
2250                  */
2251                 ret = btrfs_find_ref_cluster(trans, &cluster,
2252                                              delayed_refs->run_delayed_start);
2253                 if (ret)
2254                         break;
2255
2256                 ret = run_clustered_refs(trans, root, &cluster);
2257                 BUG_ON(ret < 0);
2258
2259                 count -= min_t(unsigned long, ret, count);
2260
2261                 if (count == 0)
2262                         break;
2263         }
2264
2265         if (run_all) {
2266                 node = rb_first(&delayed_refs->root);
2267                 if (!node)
2268                         goto out;
2269                 count = (unsigned long)-1;
2270
2271                 while (node) {
2272                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2273                                        rb_node);
2274                         if (btrfs_delayed_ref_is_head(ref)) {
2275                                 struct btrfs_delayed_ref_head *head;
2276
2277                                 head = btrfs_delayed_node_to_head(ref);
2278                                 atomic_inc(&ref->refs);
2279
2280                                 spin_unlock(&delayed_refs->lock);
2281                                 mutex_lock(&head->mutex);
2282                                 mutex_unlock(&head->mutex);
2283
2284                                 btrfs_put_delayed_ref(ref);
2285                                 cond_resched();
2286                                 goto again;
2287                         }
2288                         node = rb_next(node);
2289                 }
2290                 spin_unlock(&delayed_refs->lock);
2291                 schedule_timeout(1);
2292                 goto again;
2293         }
2294 out:
2295         spin_unlock(&delayed_refs->lock);
2296         return 0;
2297 }
2298
2299 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2300                                 struct btrfs_root *root,
2301                                 u64 bytenr, u64 num_bytes, u64 flags,
2302                                 int is_data)
2303 {
2304         struct btrfs_delayed_extent_op *extent_op;
2305         int ret;
2306
2307         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2308         if (!extent_op)
2309                 return -ENOMEM;
2310
2311         extent_op->flags_to_set = flags;
2312         extent_op->update_flags = 1;
2313         extent_op->update_key = 0;
2314         extent_op->is_data = is_data ? 1 : 0;
2315
2316         ret = btrfs_add_delayed_extent_op(trans, bytenr, num_bytes, extent_op);
2317         if (ret)
2318                 kfree(extent_op);
2319         return ret;
2320 }
2321
2322 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2323                                       struct btrfs_root *root,
2324                                       struct btrfs_path *path,
2325                                       u64 objectid, u64 offset, u64 bytenr)
2326 {
2327         struct btrfs_delayed_ref_head *head;
2328         struct btrfs_delayed_ref_node *ref;
2329         struct btrfs_delayed_data_ref *data_ref;
2330         struct btrfs_delayed_ref_root *delayed_refs;
2331         struct rb_node *node;
2332         int ret = 0;
2333
2334         ret = -ENOENT;
2335         delayed_refs = &trans->transaction->delayed_refs;
2336         spin_lock(&delayed_refs->lock);
2337         head = btrfs_find_delayed_ref_head(trans, bytenr);
2338         if (!head)
2339                 goto out;
2340
2341         if (!mutex_trylock(&head->mutex)) {
2342                 atomic_inc(&head->node.refs);
2343                 spin_unlock(&delayed_refs->lock);
2344
2345                 btrfs_release_path(root->fs_info->extent_root, path);
2346
2347                 mutex_lock(&head->mutex);
2348                 mutex_unlock(&head->mutex);
2349                 btrfs_put_delayed_ref(&head->node);
2350                 return -EAGAIN;
2351         }
2352
2353         node = rb_prev(&head->node.rb_node);
2354         if (!node)
2355                 goto out_unlock;
2356
2357         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2358
2359         if (ref->bytenr != bytenr)
2360                 goto out_unlock;
2361
2362         ret = 1;
2363         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2364                 goto out_unlock;
2365
2366         data_ref = btrfs_delayed_node_to_data_ref(ref);
2367
2368         node = rb_prev(node);
2369         if (node) {
2370                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2371                 if (ref->bytenr == bytenr)
2372                         goto out_unlock;
2373         }
2374
2375         if (data_ref->root != root->root_key.objectid ||
2376             data_ref->objectid != objectid || data_ref->offset != offset)
2377                 goto out_unlock;
2378
2379         ret = 0;
2380 out_unlock:
2381         mutex_unlock(&head->mutex);
2382 out:
2383         spin_unlock(&delayed_refs->lock);
2384         return ret;
2385 }
2386
2387 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2388                                         struct btrfs_root *root,
2389                                         struct btrfs_path *path,
2390                                         u64 objectid, u64 offset, u64 bytenr)
2391 {
2392         struct btrfs_root *extent_root = root->fs_info->extent_root;
2393         struct extent_buffer *leaf;
2394         struct btrfs_extent_data_ref *ref;
2395         struct btrfs_extent_inline_ref *iref;
2396         struct btrfs_extent_item *ei;
2397         struct btrfs_key key;
2398         u32 item_size;
2399         int ret;
2400
2401         key.objectid = bytenr;
2402         key.offset = (u64)-1;
2403         key.type = BTRFS_EXTENT_ITEM_KEY;
2404
2405         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2406         if (ret < 0)
2407                 goto out;
2408         BUG_ON(ret == 0);
2409
2410         ret = -ENOENT;
2411         if (path->slots[0] == 0)
2412                 goto out;
2413
2414         path->slots[0]--;
2415         leaf = path->nodes[0];
2416         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2417
2418         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2419                 goto out;
2420
2421         ret = 1;
2422         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2423 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2424         if (item_size < sizeof(*ei)) {
2425                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2426                 goto out;
2427         }
2428 #endif
2429         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2430
2431         if (item_size != sizeof(*ei) +
2432             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2433                 goto out;
2434
2435         if (btrfs_extent_generation(leaf, ei) <=
2436             btrfs_root_last_snapshot(&root->root_item))
2437                 goto out;
2438
2439         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2440         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2441             BTRFS_EXTENT_DATA_REF_KEY)
2442                 goto out;
2443
2444         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2445         if (btrfs_extent_refs(leaf, ei) !=
2446             btrfs_extent_data_ref_count(leaf, ref) ||
2447             btrfs_extent_data_ref_root(leaf, ref) !=
2448             root->root_key.objectid ||
2449             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2450             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2451                 goto out;
2452
2453         ret = 0;
2454 out:
2455         return ret;
2456 }
2457
2458 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2459                           struct btrfs_root *root,
2460                           u64 objectid, u64 offset, u64 bytenr)
2461 {
2462         struct btrfs_path *path;
2463         int ret;
2464         int ret2;
2465
2466         path = btrfs_alloc_path();
2467         if (!path)
2468                 return -ENOENT;
2469
2470         do {
2471                 ret = check_committed_ref(trans, root, path, objectid,
2472                                           offset, bytenr);
2473                 if (ret && ret != -ENOENT)
2474                         goto out;
2475
2476                 ret2 = check_delayed_ref(trans, root, path, objectid,
2477                                          offset, bytenr);
2478         } while (ret2 == -EAGAIN);
2479
2480         if (ret2 && ret2 != -ENOENT) {
2481                 ret = ret2;
2482                 goto out;
2483         }
2484
2485         if (ret != -ENOENT || ret2 != -ENOENT)
2486                 ret = 0;
2487 out:
2488         btrfs_free_path(path);
2489         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2490                 WARN_ON(ret > 0);
2491         return ret;
2492 }
2493
2494 #if 0
2495 int btrfs_cache_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2496                     struct extent_buffer *buf, u32 nr_extents)
2497 {
2498         struct btrfs_key key;
2499         struct btrfs_file_extent_item *fi;
2500         u64 root_gen;
2501         u32 nritems;
2502         int i;
2503         int level;
2504         int ret = 0;
2505         int shared = 0;
2506
2507         if (!root->ref_cows)
2508                 return 0;
2509
2510         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
2511                 shared = 0;
2512                 root_gen = root->root_key.offset;
2513         } else {
2514                 shared = 1;
2515                 root_gen = trans->transid - 1;
2516         }
2517
2518         level = btrfs_header_level(buf);
2519         nritems = btrfs_header_nritems(buf);
2520
2521         if (level == 0) {
2522                 struct btrfs_leaf_ref *ref;
2523                 struct btrfs_extent_info *info;
2524
2525                 ref = btrfs_alloc_leaf_ref(root, nr_extents);
2526                 if (!ref) {
2527                         ret = -ENOMEM;
2528                         goto out;
2529                 }
2530
2531                 ref->root_gen = root_gen;
2532                 ref->bytenr = buf->start;
2533                 ref->owner = btrfs_header_owner(buf);
2534                 ref->generation = btrfs_header_generation(buf);
2535                 ref->nritems = nr_extents;
2536                 info = ref->extents;
2537
2538                 for (i = 0; nr_extents > 0 && i < nritems; i++) {
2539                         u64 disk_bytenr;
2540                         btrfs_item_key_to_cpu(buf, &key, i);
2541                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2542                                 continue;
2543                         fi = btrfs_item_ptr(buf, i,
2544                                             struct btrfs_file_extent_item);
2545                         if (btrfs_file_extent_type(buf, fi) ==
2546                             BTRFS_FILE_EXTENT_INLINE)
2547                                 continue;
2548                         disk_bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2549                         if (disk_bytenr == 0)
2550                                 continue;
2551
2552                         info->bytenr = disk_bytenr;
2553                         info->num_bytes =
2554                                 btrfs_file_extent_disk_num_bytes(buf, fi);
2555                         info->objectid = key.objectid;
2556                         info->offset = key.offset;
2557                         info++;
2558                 }
2559
2560                 ret = btrfs_add_leaf_ref(root, ref, shared);
2561                 if (ret == -EEXIST && shared) {
2562                         struct btrfs_leaf_ref *old;
2563                         old = btrfs_lookup_leaf_ref(root, ref->bytenr);
2564                         BUG_ON(!old);
2565                         btrfs_remove_leaf_ref(root, old);
2566                         btrfs_free_leaf_ref(root, old);
2567                         ret = btrfs_add_leaf_ref(root, ref, shared);
2568                 }
2569                 WARN_ON(ret);
2570                 btrfs_free_leaf_ref(root, ref);
2571         }
2572 out:
2573         return ret;
2574 }
2575
2576 /* when a block goes through cow, we update the reference counts of
2577  * everything that block points to.  The internal pointers of the block
2578  * can be in just about any order, and it is likely to have clusters of
2579  * things that are close together and clusters of things that are not.
2580  *
2581  * To help reduce the seeks that come with updating all of these reference
2582  * counts, sort them by byte number before actual updates are done.
2583  *
2584  * struct refsort is used to match byte number to slot in the btree block.
2585  * we sort based on the byte number and then use the slot to actually
2586  * find the item.
2587  *
2588  * struct refsort is smaller than strcut btrfs_item and smaller than
2589  * struct btrfs_key_ptr.  Since we're currently limited to the page size
2590  * for a btree block, there's no way for a kmalloc of refsorts for a
2591  * single node to be bigger than a page.
2592  */
2593 struct refsort {
2594         u64 bytenr;
2595         u32 slot;
2596 };
2597
2598 /*
2599  * for passing into sort()
2600  */
2601 static int refsort_cmp(const void *a_void, const void *b_void)
2602 {
2603         const struct refsort *a = a_void;
2604         const struct refsort *b = b_void;
2605
2606         if (a->bytenr < b->bytenr)
2607                 return -1;
2608         if (a->bytenr > b->bytenr)
2609                 return 1;
2610         return 0;
2611 }
2612 #endif
2613
2614 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2615                            struct btrfs_root *root,
2616                            struct extent_buffer *buf,
2617                            int full_backref, int inc)
2618 {
2619         u64 bytenr;
2620         u64 num_bytes;
2621         u64 parent;
2622         u64 ref_root;
2623         u32 nritems;
2624         struct btrfs_key key;
2625         struct btrfs_file_extent_item *fi;
2626         int i;
2627         int level;
2628         int ret = 0;
2629         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2630                             u64, u64, u64, u64, u64, u64);
2631
2632         ref_root = btrfs_header_owner(buf);
2633         nritems = btrfs_header_nritems(buf);
2634         level = btrfs_header_level(buf);
2635
2636         if (!root->ref_cows && level == 0)
2637                 return 0;
2638
2639         if (inc)
2640                 process_func = btrfs_inc_extent_ref;
2641         else
2642                 process_func = btrfs_free_extent;
2643
2644         if (full_backref)
2645                 parent = buf->start;
2646         else
2647                 parent = 0;
2648
2649         for (i = 0; i < nritems; i++) {
2650                 if (level == 0) {
2651                         btrfs_item_key_to_cpu(buf, &key, i);
2652                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2653                                 continue;
2654                         fi = btrfs_item_ptr(buf, i,
2655                                             struct btrfs_file_extent_item);
2656                         if (btrfs_file_extent_type(buf, fi) ==
2657                             BTRFS_FILE_EXTENT_INLINE)
2658                                 continue;
2659                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2660                         if (bytenr == 0)
2661                                 continue;
2662
2663                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2664                         key.offset -= btrfs_file_extent_offset(buf, fi);
2665                         ret = process_func(trans, root, bytenr, num_bytes,
2666                                            parent, ref_root, key.objectid,
2667                                            key.offset);
2668                         if (ret)
2669                                 goto fail;
2670                 } else {
2671                         bytenr = btrfs_node_blockptr(buf, i);
2672                         num_bytes = btrfs_level_size(root, level - 1);
2673                         ret = process_func(trans, root, bytenr, num_bytes,
2674                                            parent, ref_root, level - 1, 0);
2675                         if (ret)
2676                                 goto fail;
2677                 }
2678         }
2679         return 0;
2680 fail:
2681         BUG();
2682         return ret;
2683 }
2684
2685 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2686                   struct extent_buffer *buf, int full_backref)
2687 {
2688         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2689 }
2690
2691 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2692                   struct extent_buffer *buf, int full_backref)
2693 {
2694         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2695 }
2696
2697 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2698                                  struct btrfs_root *root,
2699                                  struct btrfs_path *path,
2700                                  struct btrfs_block_group_cache *cache)
2701 {
2702         int ret;
2703         struct btrfs_root *extent_root = root->fs_info->extent_root;
2704         unsigned long bi;
2705         struct extent_buffer *leaf;
2706
2707         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2708         if (ret < 0)
2709                 goto fail;
2710         BUG_ON(ret);
2711
2712         leaf = path->nodes[0];
2713         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2714         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2715         btrfs_mark_buffer_dirty(leaf);
2716         btrfs_release_path(extent_root, path);
2717 fail:
2718         if (ret)
2719                 return ret;
2720         return 0;
2721
2722 }
2723
2724 static struct btrfs_block_group_cache *
2725 next_block_group(struct btrfs_root *root,
2726                  struct btrfs_block_group_cache *cache)
2727 {
2728         struct rb_node *node;
2729         spin_lock(&root->fs_info->block_group_cache_lock);
2730         node = rb_next(&cache->cache_node);
2731         btrfs_put_block_group(cache);
2732         if (node) {
2733                 cache = rb_entry(node, struct btrfs_block_group_cache,
2734                                  cache_node);
2735                 btrfs_get_block_group(cache);
2736         } else
2737                 cache = NULL;
2738         spin_unlock(&root->fs_info->block_group_cache_lock);
2739         return cache;
2740 }
2741
2742 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2743                             struct btrfs_trans_handle *trans,
2744                             struct btrfs_path *path)
2745 {
2746         struct btrfs_root *root = block_group->fs_info->tree_root;
2747         struct inode *inode = NULL;
2748         u64 alloc_hint = 0;
2749         int dcs = BTRFS_DC_ERROR;
2750         int num_pages = 0;
2751         int retries = 0;
2752         int ret = 0;
2753
2754         /*
2755          * If this block group is smaller than 100 megs don't bother caching the
2756          * block group.
2757          */
2758         if (block_group->key.offset < (100 * 1024 * 1024)) {
2759                 spin_lock(&block_group->lock);
2760                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2761                 spin_unlock(&block_group->lock);
2762                 return 0;
2763         }
2764
2765 again:
2766         inode = lookup_free_space_inode(root, block_group, path);
2767         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2768                 ret = PTR_ERR(inode);
2769                 btrfs_release_path(root, path);
2770                 goto out;
2771         }
2772
2773         if (IS_ERR(inode)) {
2774                 BUG_ON(retries);
2775                 retries++;
2776
2777                 if (block_group->ro)
2778                         goto out_free;
2779
2780                 ret = create_free_space_inode(root, trans, block_group, path);
2781                 if (ret)
2782                         goto out_free;
2783                 goto again;
2784         }
2785
2786         /*
2787          * We want to set the generation to 0, that way if anything goes wrong
2788          * from here on out we know not to trust this cache when we load up next
2789          * time.
2790          */
2791         BTRFS_I(inode)->generation = 0;
2792         ret = btrfs_update_inode(trans, root, inode);
2793         WARN_ON(ret);
2794
2795         if (i_size_read(inode) > 0) {
2796                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2797                                                       inode);
2798                 if (ret)
2799                         goto out_put;
2800         }
2801
2802         spin_lock(&block_group->lock);
2803         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2804                 /* We're not cached, don't bother trying to write stuff out */
2805                 dcs = BTRFS_DC_WRITTEN;
2806                 spin_unlock(&block_group->lock);
2807                 goto out_put;
2808         }
2809         spin_unlock(&block_group->lock);
2810
2811         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2812         if (!num_pages)
2813                 num_pages = 1;
2814
2815         /*
2816          * Just to make absolutely sure we have enough space, we're going to
2817          * preallocate 12 pages worth of space for each block group.  In
2818          * practice we ought to use at most 8, but we need extra space so we can
2819          * add our header and have a terminator between the extents and the
2820          * bitmaps.
2821          */
2822         num_pages *= 16;
2823         num_pages *= PAGE_CACHE_SIZE;
2824
2825         ret = btrfs_check_data_free_space(inode, num_pages);
2826         if (ret)
2827                 goto out_put;
2828
2829         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2830                                               num_pages, num_pages,
2831                                               &alloc_hint);
2832         if (!ret)
2833                 dcs = BTRFS_DC_SETUP;
2834         btrfs_free_reserved_data_space(inode, num_pages);
2835 out_put:
2836         iput(inode);
2837 out_free:
2838         btrfs_release_path(root, path);
2839 out:
2840         spin_lock(&block_group->lock);
2841         block_group->disk_cache_state = dcs;
2842         spin_unlock(&block_group->lock);
2843
2844         return ret;
2845 }
2846
2847 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2848                                    struct btrfs_root *root)
2849 {
2850         struct btrfs_block_group_cache *cache;
2851         int err = 0;
2852         struct btrfs_path *path;
2853         u64 last = 0;
2854
2855         path = btrfs_alloc_path();
2856         if (!path)
2857                 return -ENOMEM;
2858
2859 again:
2860         while (1) {
2861                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2862                 while (cache) {
2863                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2864                                 break;
2865                         cache = next_block_group(root, cache);
2866                 }
2867                 if (!cache) {
2868                         if (last == 0)
2869                                 break;
2870                         last = 0;
2871                         continue;
2872                 }
2873                 err = cache_save_setup(cache, trans, path);
2874                 last = cache->key.objectid + cache->key.offset;
2875                 btrfs_put_block_group(cache);
2876         }
2877
2878         while (1) {
2879                 if (last == 0) {
2880                         err = btrfs_run_delayed_refs(trans, root,
2881                                                      (unsigned long)-1);
2882                         BUG_ON(err);
2883                 }
2884
2885                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2886                 while (cache) {
2887                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2888                                 btrfs_put_block_group(cache);
2889                                 goto again;
2890                         }
2891
2892                         if (cache->dirty)
2893                                 break;
2894                         cache = next_block_group(root, cache);
2895                 }
2896                 if (!cache) {
2897                         if (last == 0)
2898                                 break;
2899                         last = 0;
2900                         continue;
2901                 }
2902
2903                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
2904                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
2905                 cache->dirty = 0;
2906                 last = cache->key.objectid + cache->key.offset;
2907
2908                 err = write_one_cache_group(trans, root, path, cache);
2909                 BUG_ON(err);
2910                 btrfs_put_block_group(cache);
2911         }
2912
2913         while (1) {
2914                 /*
2915                  * I don't think this is needed since we're just marking our
2916                  * preallocated extent as written, but just in case it can't
2917                  * hurt.
2918                  */
2919                 if (last == 0) {
2920                         err = btrfs_run_delayed_refs(trans, root,
2921                                                      (unsigned long)-1);
2922                         BUG_ON(err);
2923                 }
2924
2925                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2926                 while (cache) {
2927                         /*
2928                          * Really this shouldn't happen, but it could if we
2929                          * couldn't write the entire preallocated extent and
2930                          * splitting the extent resulted in a new block.
2931                          */
2932                         if (cache->dirty) {
2933                                 btrfs_put_block_group(cache);
2934                                 goto again;
2935                         }
2936                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2937                                 break;
2938                         cache = next_block_group(root, cache);
2939                 }
2940                 if (!cache) {
2941                         if (last == 0)
2942                                 break;
2943                         last = 0;
2944                         continue;
2945                 }
2946
2947                 btrfs_write_out_cache(root, trans, cache, path);
2948
2949                 /*
2950                  * If we didn't have an error then the cache state is still
2951                  * NEED_WRITE, so we can set it to WRITTEN.
2952                  */
2953                 if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2954                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
2955                 last = cache->key.objectid + cache->key.offset;
2956                 btrfs_put_block_group(cache);
2957         }
2958
2959         btrfs_free_path(path);
2960         return 0;
2961 }
2962
2963 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
2964 {
2965         struct btrfs_block_group_cache *block_group;
2966         int readonly = 0;
2967
2968         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
2969         if (!block_group || block_group->ro)
2970                 readonly = 1;
2971         if (block_group)
2972                 btrfs_put_block_group(block_group);
2973         return readonly;
2974 }
2975
2976 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
2977                              u64 total_bytes, u64 bytes_used,
2978                              struct btrfs_space_info **space_info)
2979 {
2980         struct btrfs_space_info *found;
2981         int i;
2982         int factor;
2983
2984         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
2985                      BTRFS_BLOCK_GROUP_RAID10))
2986                 factor = 2;
2987         else
2988                 factor = 1;
2989
2990         found = __find_space_info(info, flags);
2991         if (found) {
2992                 spin_lock(&found->lock);
2993                 found->total_bytes += total_bytes;
2994                 found->disk_total += total_bytes * factor;
2995                 found->bytes_used += bytes_used;
2996                 found->disk_used += bytes_used * factor;
2997                 found->full = 0;
2998                 spin_unlock(&found->lock);
2999                 *space_info = found;
3000                 return 0;
3001         }
3002         found = kzalloc(sizeof(*found), GFP_NOFS);
3003         if (!found)
3004                 return -ENOMEM;
3005
3006         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3007                 INIT_LIST_HEAD(&found->block_groups[i]);
3008         init_rwsem(&found->groups_sem);
3009         spin_lock_init(&found->lock);
3010         found->flags = flags & (BTRFS_BLOCK_GROUP_DATA |
3011                                 BTRFS_BLOCK_GROUP_SYSTEM |
3012                                 BTRFS_BLOCK_GROUP_METADATA);
3013         found->total_bytes = total_bytes;
3014         found->disk_total = total_bytes * factor;
3015         found->bytes_used = bytes_used;
3016         found->disk_used = bytes_used * factor;
3017         found->bytes_pinned = 0;
3018         found->bytes_reserved = 0;
3019         found->bytes_readonly = 0;
3020         found->bytes_may_use = 0;
3021         found->full = 0;
3022         found->force_alloc = 0;
3023         *space_info = found;
3024         list_add_rcu(&found->list, &info->space_info);
3025         atomic_set(&found->caching_threads, 0);
3026         return 0;
3027 }
3028
3029 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3030 {
3031         u64 extra_flags = flags & (BTRFS_BLOCK_GROUP_RAID0 |
3032                                    BTRFS_BLOCK_GROUP_RAID1 |
3033                                    BTRFS_BLOCK_GROUP_RAID10 |
3034                                    BTRFS_BLOCK_GROUP_DUP);
3035         if (extra_flags) {
3036                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3037                         fs_info->avail_data_alloc_bits |= extra_flags;
3038                 if (flags & BTRFS_BLOCK_GROUP_METADATA)
3039                         fs_info->avail_metadata_alloc_bits |= extra_flags;
3040                 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3041                         fs_info->avail_system_alloc_bits |= extra_flags;
3042         }
3043 }
3044
3045 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3046 {
3047         /*
3048          * we add in the count of missing devices because we want
3049          * to make sure that any RAID levels on a degraded FS
3050          * continue to be honored.
3051          */
3052         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3053                 root->fs_info->fs_devices->missing_devices;
3054
3055         if (num_devices == 1)
3056                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3057         if (num_devices < 4)
3058                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3059
3060         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3061             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3062                       BTRFS_BLOCK_GROUP_RAID10))) {
3063                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3064         }
3065
3066         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3067             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3068                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3069         }
3070
3071         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3072             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3073              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3074              (flags & BTRFS_BLOCK_GROUP_DUP)))
3075                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3076         return flags;
3077 }
3078
3079 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3080 {
3081         if (flags & BTRFS_BLOCK_GROUP_DATA)
3082                 flags |= root->fs_info->avail_data_alloc_bits &
3083                          root->fs_info->data_alloc_profile;
3084         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3085                 flags |= root->fs_info->avail_system_alloc_bits &
3086                          root->fs_info->system_alloc_profile;
3087         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3088                 flags |= root->fs_info->avail_metadata_alloc_bits &
3089                          root->fs_info->metadata_alloc_profile;
3090         return btrfs_reduce_alloc_profile(root, flags);
3091 }
3092
3093 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3094 {
3095         u64 flags;
3096
3097         if (data)
3098                 flags = BTRFS_BLOCK_GROUP_DATA;
3099         else if (root == root->fs_info->chunk_root)
3100                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3101         else
3102                 flags = BTRFS_BLOCK_GROUP_METADATA;
3103
3104         return get_alloc_profile(root, flags);
3105 }
3106
3107 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3108 {
3109         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3110                                                        BTRFS_BLOCK_GROUP_DATA);
3111 }
3112
3113 /*
3114  * This will check the space that the inode allocates from to make sure we have
3115  * enough space for bytes.
3116  */
3117 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3118 {
3119         struct btrfs_space_info *data_sinfo;
3120         struct btrfs_root *root = BTRFS_I(inode)->root;
3121         u64 used;
3122         int ret = 0, committed = 0, alloc_chunk = 1;
3123
3124         /* make sure bytes are sectorsize aligned */
3125         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3126
3127         if (root == root->fs_info->tree_root) {
3128                 alloc_chunk = 0;
3129                 committed = 1;
3130         }
3131
3132         data_sinfo = BTRFS_I(inode)->space_info;
3133         if (!data_sinfo)
3134                 goto alloc;
3135
3136 again:
3137         /* make sure we have enough space to handle the data first */
3138         spin_lock(&data_sinfo->lock);
3139         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3140                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3141                 data_sinfo->bytes_may_use;
3142
3143         if (used + bytes > data_sinfo->total_bytes) {
3144                 struct btrfs_trans_handle *trans;
3145
3146                 /*
3147                  * if we don't have enough free bytes in this space then we need
3148                  * to alloc a new chunk.
3149                  */
3150                 if (!data_sinfo->full && alloc_chunk) {
3151                         u64 alloc_target;
3152
3153                         data_sinfo->force_alloc = 1;
3154                         spin_unlock(&data_sinfo->lock);
3155 alloc:
3156                         alloc_target = btrfs_get_alloc_profile(root, 1);
3157                         trans = btrfs_join_transaction(root, 1);
3158                         if (IS_ERR(trans))
3159                                 return PTR_ERR(trans);
3160
3161                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3162                                              bytes + 2 * 1024 * 1024,
3163                                              alloc_target, 0);
3164                         btrfs_end_transaction(trans, root);
3165                         if (ret < 0) {
3166                                 if (ret != -ENOSPC)
3167                                         return ret;
3168                                 else
3169                                         goto commit_trans;
3170                         }
3171
3172                         if (!data_sinfo) {
3173                                 btrfs_set_inode_space_info(root, inode);
3174                                 data_sinfo = BTRFS_I(inode)->space_info;
3175                         }
3176                         goto again;
3177                 }
3178                 spin_unlock(&data_sinfo->lock);
3179
3180                 /* commit the current transaction and try again */
3181 commit_trans:
3182                 if (!committed && !root->fs_info->open_ioctl_trans) {
3183                         committed = 1;
3184                         trans = btrfs_join_transaction(root, 1);
3185                         if (IS_ERR(trans))
3186                                 return PTR_ERR(trans);
3187                         ret = btrfs_commit_transaction(trans, root);
3188                         if (ret)
3189                                 return ret;
3190                         goto again;
3191                 }
3192
3193 #if 0 /* I hope we never need this code again, just in case */
3194                 printk(KERN_ERR "no space left, need %llu, %llu bytes_used, "
3195                        "%llu bytes_reserved, " "%llu bytes_pinned, "
3196                        "%llu bytes_readonly, %llu may use %llu total\n",
3197                        (unsigned long long)bytes,
3198                        (unsigned long long)data_sinfo->bytes_used,
3199                        (unsigned long long)data_sinfo->bytes_reserved,
3200                        (unsigned long long)data_sinfo->bytes_pinned,
3201                        (unsigned long long)data_sinfo->bytes_readonly,
3202                        (unsigned long long)data_sinfo->bytes_may_use,
3203                        (unsigned long long)data_sinfo->total_bytes);
3204 #endif
3205                 return -ENOSPC;
3206         }
3207         data_sinfo->bytes_may_use += bytes;
3208         BTRFS_I(inode)->reserved_bytes += bytes;
3209         spin_unlock(&data_sinfo->lock);
3210
3211         return 0;
3212 }
3213
3214 /*
3215  * called when we are clearing an delalloc extent from the
3216  * inode's io_tree or there was an error for whatever reason
3217  * after calling btrfs_check_data_free_space
3218  */
3219 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3220 {
3221         struct btrfs_root *root = BTRFS_I(inode)->root;
3222         struct btrfs_space_info *data_sinfo;
3223
3224         /* make sure bytes are sectorsize aligned */
3225         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3226
3227         data_sinfo = BTRFS_I(inode)->space_info;
3228         spin_lock(&data_sinfo->lock);
3229         data_sinfo->bytes_may_use -= bytes;
3230         BTRFS_I(inode)->reserved_bytes -= bytes;
3231         spin_unlock(&data_sinfo->lock);
3232 }
3233
3234 static void force_metadata_allocation(struct btrfs_fs_info *info)
3235 {
3236         struct list_head *head = &info->space_info;
3237         struct btrfs_space_info *found;
3238
3239         rcu_read_lock();
3240         list_for_each_entry_rcu(found, head, list) {
3241                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3242                         found->force_alloc = 1;
3243         }
3244         rcu_read_unlock();
3245 }
3246
3247 static int should_alloc_chunk(struct btrfs_root *root,
3248                               struct btrfs_space_info *sinfo, u64 alloc_bytes)
3249 {
3250         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3251         u64 thresh;
3252
3253         if (sinfo->bytes_used + sinfo->bytes_reserved +
3254             alloc_bytes + 256 * 1024 * 1024 < num_bytes)
3255                 return 0;
3256
3257         if (sinfo->bytes_used + sinfo->bytes_reserved +
3258             alloc_bytes < div_factor(num_bytes, 8))
3259                 return 0;
3260
3261         thresh = btrfs_super_total_bytes(&root->fs_info->super_copy);
3262         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 5));
3263
3264         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 3))
3265                 return 0;
3266
3267         return 1;
3268 }
3269
3270 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3271                           struct btrfs_root *extent_root, u64 alloc_bytes,
3272                           u64 flags, int force)
3273 {
3274         struct btrfs_space_info *space_info;
3275         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3276         int ret = 0;
3277
3278         mutex_lock(&fs_info->chunk_mutex);
3279
3280         flags = btrfs_reduce_alloc_profile(extent_root, flags);
3281
3282         space_info = __find_space_info(extent_root->fs_info, flags);
3283         if (!space_info) {
3284                 ret = update_space_info(extent_root->fs_info, flags,
3285                                         0, 0, &space_info);
3286                 BUG_ON(ret);
3287         }
3288         BUG_ON(!space_info);
3289
3290         spin_lock(&space_info->lock);
3291         if (space_info->force_alloc)
3292                 force = 1;
3293         if (space_info->full) {
3294                 spin_unlock(&space_info->lock);
3295                 goto out;
3296         }
3297
3298         if (!force && !should_alloc_chunk(extent_root, space_info,
3299                                           alloc_bytes)) {
3300                 spin_unlock(&space_info->lock);
3301                 goto out;
3302         }
3303         spin_unlock(&space_info->lock);
3304
3305         /*
3306          * If we have mixed data/metadata chunks we want to make sure we keep
3307          * allocating mixed chunks instead of individual chunks.
3308          */
3309         if (btrfs_mixed_space_info(space_info))
3310                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3311
3312         /*
3313          * if we're doing a data chunk, go ahead and make sure that
3314          * we keep a reasonable number of metadata chunks allocated in the
3315          * FS as well.
3316          */
3317         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3318                 fs_info->data_chunk_allocations++;
3319                 if (!(fs_info->data_chunk_allocations %
3320                       fs_info->metadata_ratio))
3321                         force_metadata_allocation(fs_info);
3322         }
3323
3324         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3325         spin_lock(&space_info->lock);
3326         if (ret)
3327                 space_info->full = 1;
3328         else
3329                 ret = 1;
3330         space_info->force_alloc = 0;
3331         spin_unlock(&space_info->lock);
3332 out:
3333         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3334         return ret;
3335 }
3336
3337 /*
3338  * shrink metadata reservation for delalloc
3339  */
3340 static int shrink_delalloc(struct btrfs_trans_handle *trans,
3341                            struct btrfs_root *root, u64 to_reclaim, int sync)
3342 {
3343         struct btrfs_block_rsv *block_rsv;
3344         struct btrfs_space_info *space_info;
3345         u64 reserved;
3346         u64 max_reclaim;
3347         u64 reclaimed = 0;
3348         long time_left;
3349         int pause = 1;
3350         int nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3351         int loops = 0;
3352
3353         block_rsv = &root->fs_info->delalloc_block_rsv;
3354         space_info = block_rsv->space_info;
3355
3356         smp_mb();
3357         reserved = space_info->bytes_reserved;
3358
3359         if (reserved == 0)
3360                 return 0;
3361
3362         max_reclaim = min(reserved, to_reclaim);
3363
3364         while (loops < 1024) {
3365                 /* have the flusher threads jump in and do some IO */
3366                 smp_mb();
3367                 nr_pages = min_t(unsigned long, nr_pages,
3368                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3369                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages);
3370
3371                 spin_lock(&space_info->lock);
3372                 if (reserved > space_info->bytes_reserved) {
3373                         loops = 0;
3374                         reclaimed += reserved - space_info->bytes_reserved;
3375                 } else {
3376                         loops++;
3377                 }
3378                 reserved = space_info->bytes_reserved;
3379                 spin_unlock(&space_info->lock);
3380
3381                 if (reserved == 0 || reclaimed >= max_reclaim)
3382                         break;
3383
3384                 if (trans && trans->transaction->blocked)
3385                         return -EAGAIN;
3386
3387                 __set_current_state(TASK_INTERRUPTIBLE);
3388                 time_left = schedule_timeout(pause);
3389
3390                 /* We were interrupted, exit */
3391                 if (time_left)
3392                         break;
3393
3394                 pause <<= 1;
3395                 if (pause > HZ / 10)
3396                         pause = HZ / 10;
3397
3398         }
3399         return reclaimed >= to_reclaim;
3400 }
3401
3402 /*
3403  * Retries tells us how many times we've called reserve_metadata_bytes.  The
3404  * idea is if this is the first call (retries == 0) then we will add to our
3405  * reserved count if we can't make the allocation in order to hold our place
3406  * while we go and try and free up space.  That way for retries > 1 we don't try
3407  * and add space, we just check to see if the amount of unused space is >= the
3408  * total space, meaning that our reservation is valid.
3409  *
3410  * However if we don't intend to retry this reservation, pass -1 as retries so
3411  * that it short circuits this logic.
3412  */
3413 static int reserve_metadata_bytes(struct btrfs_trans_handle *trans,
3414                                   struct btrfs_root *root,
3415                                   struct btrfs_block_rsv *block_rsv,
3416                                   u64 orig_bytes, int flush)
3417 {
3418         struct btrfs_space_info *space_info = block_rsv->space_info;
3419         u64 unused;
3420         u64 num_bytes = orig_bytes;
3421         int retries = 0;
3422         int ret = 0;
3423         bool reserved = false;
3424         bool committed = false;
3425
3426 again:
3427         ret = -ENOSPC;
3428         if (reserved)
3429                 num_bytes = 0;
3430
3431         spin_lock(&space_info->lock);
3432         unused = space_info->bytes_used + space_info->bytes_reserved +
3433                  space_info->bytes_pinned + space_info->bytes_readonly +
3434                  space_info->bytes_may_use;
3435
3436         /*
3437          * The idea here is that we've not already over-reserved the block group
3438          * then we can go ahead and save our reservation first and then start
3439          * flushing if we need to.  Otherwise if we've already overcommitted
3440          * lets start flushing stuff first and then come back and try to make
3441          * our reservation.
3442          */
3443         if (unused <= space_info->total_bytes) {
3444                 unused = space_info->total_bytes - unused;
3445                 if (unused >= num_bytes) {
3446                         if (!reserved)
3447                                 space_info->bytes_reserved += orig_bytes;
3448                         ret = 0;
3449                 } else {
3450                         /*
3451                          * Ok set num_bytes to orig_bytes since we aren't
3452                          * overocmmitted, this way we only try and reclaim what
3453                          * we need.
3454                          */
3455                         num_bytes = orig_bytes;
3456                 }
3457         } else {
3458                 /*
3459                  * Ok we're over committed, set num_bytes to the overcommitted
3460                  * amount plus the amount of bytes that we need for this
3461                  * reservation.
3462                  */
3463                 num_bytes = unused - space_info->total_bytes +
3464                         (orig_bytes * (retries + 1));
3465         }
3466
3467         /*
3468          * Couldn't make our reservation, save our place so while we're trying
3469          * to reclaim space we can actually use it instead of somebody else
3470          * stealing it from us.
3471          */
3472         if (ret && !reserved) {
3473                 space_info->bytes_reserved += orig_bytes;
3474                 reserved = true;
3475         }
3476
3477         spin_unlock(&space_info->lock);
3478
3479         if (!ret)
3480                 return 0;
3481
3482         if (!flush)
3483                 goto out;
3484
3485         /*
3486          * We do synchronous shrinking since we don't actually unreserve
3487          * metadata until after the IO is completed.
3488          */
3489         ret = shrink_delalloc(trans, root, num_bytes, 1);
3490         if (ret > 0)
3491                 return 0;
3492         else if (ret < 0)
3493                 goto out;
3494
3495         /*
3496          * So if we were overcommitted it's possible that somebody else flushed
3497          * out enough space and we simply didn't have enough space to reclaim,
3498          * so go back around and try again.
3499          */
3500         if (retries < 2) {
3501                 retries++;
3502                 goto again;
3503         }
3504
3505         spin_lock(&space_info->lock);
3506         /*
3507          * Not enough space to be reclaimed, don't bother committing the
3508          * transaction.
3509          */
3510         if (space_info->bytes_pinned < orig_bytes)
3511                 ret = -ENOSPC;
3512         spin_unlock(&space_info->lock);
3513         if (ret)
3514                 goto out;
3515
3516         ret = -EAGAIN;
3517         if (trans || committed)
3518                 goto out;
3519
3520         ret = -ENOSPC;
3521         trans = btrfs_join_transaction(root, 1);
3522         if (IS_ERR(trans))
3523                 goto out;
3524         ret = btrfs_commit_transaction(trans, root);
3525         if (!ret) {
3526                 trans = NULL;
3527                 committed = true;
3528                 goto again;
3529         }
3530
3531 out:
3532         if (reserved) {
3533                 spin_lock(&space_info->lock);
3534                 space_info->bytes_reserved -= orig_bytes;
3535                 spin_unlock(&space_info->lock);
3536         }
3537
3538         return ret;
3539 }
3540
3541 static struct btrfs_block_rsv *get_block_rsv(struct btrfs_trans_handle *trans,
3542                                              struct btrfs_root *root)
3543 {
3544         struct btrfs_block_rsv *block_rsv;
3545         if (root->ref_cows)
3546                 block_rsv = trans->block_rsv;
3547         else
3548                 block_rsv = root->block_rsv;
3549
3550         if (!block_rsv)
3551                 block_rsv = &root->fs_info->empty_block_rsv;
3552
3553         return block_rsv;
3554 }
3555
3556 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3557                                u64 num_bytes)
3558 {
3559         int ret = -ENOSPC;
3560         spin_lock(&block_rsv->lock);
3561         if (block_rsv->reserved >= num_bytes) {
3562                 block_rsv->reserved -= num_bytes;
3563                 if (block_rsv->reserved < block_rsv->size)
3564                         block_rsv->full = 0;
3565                 ret = 0;
3566         }
3567         spin_unlock(&block_rsv->lock);
3568         return ret;
3569 }
3570
3571 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3572                                 u64 num_bytes, int update_size)
3573 {
3574         spin_lock(&block_rsv->lock);
3575         block_rsv->reserved += num_bytes;
3576         if (update_size)
3577                 block_rsv->size += num_bytes;
3578         else if (block_rsv->reserved >= block_rsv->size)
3579                 block_rsv->full = 1;
3580         spin_unlock(&block_rsv->lock);
3581 }
3582
3583 void block_rsv_release_bytes(struct btrfs_block_rsv *block_rsv,
3584                              struct btrfs_block_rsv *dest, u64 num_bytes)
3585 {
3586         struct btrfs_space_info *space_info = block_rsv->space_info;
3587
3588         spin_lock(&block_rsv->lock);
3589         if (num_bytes == (u64)-1)
3590                 num_bytes = block_rsv->size;
3591         block_rsv->size -= num_bytes;
3592         if (block_rsv->reserved >= block_rsv->size) {
3593                 num_bytes = block_rsv->reserved - block_rsv->size;
3594                 block_rsv->reserved = block_rsv->size;
3595                 block_rsv->full = 1;
3596         } else {
3597                 num_bytes = 0;
3598         }
3599         spin_unlock(&block_rsv->lock);
3600
3601         if (num_bytes > 0) {
3602                 if (dest) {
3603                         spin_lock(&dest->lock);
3604                         if (!dest->full) {
3605                                 u64 bytes_to_add;
3606
3607                                 bytes_to_add = dest->size - dest->reserved;
3608                                 bytes_to_add = min(num_bytes, bytes_to_add);
3609                                 dest->reserved += bytes_to_add;
3610                                 if (dest->reserved >= dest->size)
3611                                         dest->full = 1;
3612                                 num_bytes -= bytes_to_add;
3613                         }
3614                         spin_unlock(&dest->lock);
3615                 }
3616                 if (num_bytes) {
3617                         spin_lock(&space_info->lock);
3618                         space_info->bytes_reserved -= num_bytes;
3619                         spin_unlock(&space_info->lock);
3620                 }
3621         }
3622 }
3623
3624 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3625                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3626 {
3627         int ret;
3628
3629         ret = block_rsv_use_bytes(src, num_bytes);
3630         if (ret)
3631                 return ret;
3632
3633         block_rsv_add_bytes(dst, num_bytes, 1);
3634         return 0;
3635 }
3636
3637 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3638 {
3639         memset(rsv, 0, sizeof(*rsv));
3640         spin_lock_init(&rsv->lock);
3641         atomic_set(&rsv->usage, 1);
3642         rsv->priority = 6;
3643         INIT_LIST_HEAD(&rsv->list);
3644 }
3645
3646 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3647 {
3648         struct btrfs_block_rsv *block_rsv;
3649         struct btrfs_fs_info *fs_info = root->fs_info;
3650
3651         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3652         if (!block_rsv)
3653                 return NULL;
3654
3655         btrfs_init_block_rsv(block_rsv);
3656         block_rsv->space_info = __find_space_info(fs_info,
3657                                                   BTRFS_BLOCK_GROUP_METADATA);
3658         return block_rsv;
3659 }
3660
3661 void btrfs_free_block_rsv(struct btrfs_root *root,
3662                           struct btrfs_block_rsv *rsv)
3663 {
3664         if (rsv && atomic_dec_and_test(&rsv->usage)) {
3665                 btrfs_block_rsv_release(root, rsv, (u64)-1);
3666                 if (!rsv->durable)
3667                         kfree(rsv);
3668         }
3669 }
3670
3671 /*
3672  * make the block_rsv struct be able to capture freed space.
3673  * the captured space will re-add to the the block_rsv struct
3674  * after transaction commit
3675  */
3676 void btrfs_add_durable_block_rsv(struct btrfs_fs_info *fs_info,
3677                                  struct btrfs_block_rsv *block_rsv)
3678 {
3679         block_rsv->durable = 1;
3680         mutex_lock(&fs_info->durable_block_rsv_mutex);
3681         list_add_tail(&block_rsv->list, &fs_info->durable_block_rsv_list);
3682         mutex_unlock(&fs_info->durable_block_rsv_mutex);
3683 }
3684
3685 int btrfs_block_rsv_add(struct btrfs_trans_handle *trans,
3686                         struct btrfs_root *root,
3687                         struct btrfs_block_rsv *block_rsv,
3688                         u64 num_bytes)
3689 {
3690         int ret;
3691
3692         if (num_bytes == 0)
3693                 return 0;
3694
3695         ret = reserve_metadata_bytes(trans, root, block_rsv, num_bytes, 1);
3696         if (!ret) {
3697                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3698                 return 0;
3699         }
3700
3701         return ret;
3702 }
3703
3704 int btrfs_block_rsv_check(struct btrfs_trans_handle *trans,
3705                           struct btrfs_root *root,
3706                           struct btrfs_block_rsv *block_rsv,
3707                           u64 min_reserved, int min_factor)
3708 {
3709         u64 num_bytes = 0;
3710         int commit_trans = 0;
3711         int ret = -ENOSPC;
3712
3713         if (!block_rsv)
3714                 return 0;
3715
3716         spin_lock(&block_rsv->lock);
3717         if (min_factor > 0)
3718                 num_bytes = div_factor(block_rsv->size, min_factor);
3719         if (min_reserved > num_bytes)
3720                 num_bytes = min_reserved;
3721
3722         if (block_rsv->reserved >= num_bytes) {
3723                 ret = 0;
3724         } else {
3725                 num_bytes -= block_rsv->reserved;
3726                 if (block_rsv->durable &&
3727                     block_rsv->freed[0] + block_rsv->freed[1] >= num_bytes)
3728                         commit_trans = 1;
3729         }
3730         spin_unlock(&block_rsv->lock);
3731         if (!ret)
3732                 return 0;
3733
3734         if (block_rsv->refill_used) {
3735                 ret = reserve_metadata_bytes(trans, root, block_rsv,
3736                                              num_bytes, 0);
3737                 if (!ret) {
3738                         block_rsv_add_bytes(block_rsv, num_bytes, 0);
3739                         return 0;
3740                 }
3741         }
3742
3743         if (commit_trans) {
3744                 if (trans)
3745                         return -EAGAIN;
3746
3747                 trans = btrfs_join_transaction(root, 1);
3748                 BUG_ON(IS_ERR(trans));
3749                 ret = btrfs_commit_transaction(trans, root);
3750                 return 0;
3751         }
3752
3753         return -ENOSPC;
3754 }
3755
3756 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
3757                             struct btrfs_block_rsv *dst_rsv,
3758                             u64 num_bytes)
3759 {
3760         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3761 }
3762
3763 void btrfs_block_rsv_release(struct btrfs_root *root,
3764                              struct btrfs_block_rsv *block_rsv,
3765                              u64 num_bytes)
3766 {
3767         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3768         if (global_rsv->full || global_rsv == block_rsv ||
3769             block_rsv->space_info != global_rsv->space_info)
3770                 global_rsv = NULL;
3771         block_rsv_release_bytes(block_rsv, global_rsv, num_bytes);
3772 }
3773
3774 /*
3775  * helper to calculate size of global block reservation.
3776  * the desired value is sum of space used by extent tree,
3777  * checksum tree and root tree
3778  */
3779 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
3780 {
3781         struct btrfs_space_info *sinfo;
3782         u64 num_bytes;
3783         u64 meta_used;
3784         u64 data_used;
3785         int csum_size = btrfs_super_csum_size(&fs_info->super_copy);
3786 #if 0
3787         /*
3788          * per tree used space accounting can be inaccuracy, so we
3789          * can't rely on it.
3790          */
3791         spin_lock(&fs_info->extent_root->accounting_lock);
3792         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item);
3793         spin_unlock(&fs_info->extent_root->accounting_lock);
3794
3795         spin_lock(&fs_info->csum_root->accounting_lock);
3796         num_bytes += btrfs_root_used(&fs_info->csum_root->root_item);
3797         spin_unlock(&fs_info->csum_root->accounting_lock);
3798
3799         spin_lock(&fs_info->tree_root->accounting_lock);
3800         num_bytes += btrfs_root_used(&fs_info->tree_root->root_item);
3801         spin_unlock(&fs_info->tree_root->accounting_lock);
3802 #endif
3803         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
3804         spin_lock(&sinfo->lock);
3805         data_used = sinfo->bytes_used;
3806         spin_unlock(&sinfo->lock);
3807
3808         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3809         spin_lock(&sinfo->lock);
3810         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
3811                 data_used = 0;
3812         meta_used = sinfo->bytes_used;
3813         spin_unlock(&sinfo->lock);
3814
3815         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
3816                     csum_size * 2;
3817         num_bytes += div64_u64(data_used + meta_used, 50);
3818
3819         if (num_bytes * 3 > meta_used)
3820                 num_bytes = div64_u64(meta_used, 3);
3821
3822         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
3823 }
3824
3825 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
3826 {
3827         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
3828         struct btrfs_space_info *sinfo = block_rsv->space_info;
3829         u64 num_bytes;
3830
3831         num_bytes = calc_global_metadata_size(fs_info);
3832
3833         spin_lock(&block_rsv->lock);
3834         spin_lock(&sinfo->lock);
3835
3836         block_rsv->size = num_bytes;
3837
3838         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
3839                     sinfo->bytes_reserved + sinfo->bytes_readonly +
3840                     sinfo->bytes_may_use;
3841
3842         if (sinfo->total_bytes > num_bytes) {
3843                 num_bytes = sinfo->total_bytes - num_bytes;
3844                 block_rsv->reserved += num_bytes;
3845                 sinfo->bytes_reserved += num_bytes;
3846         }
3847
3848         if (block_rsv->reserved >= block_rsv->size) {
3849                 num_bytes = block_rsv->reserved - block_rsv->size;
3850                 sinfo->bytes_reserved -= num_bytes;
3851                 block_rsv->reserved = block_rsv->size;
3852                 block_rsv->full = 1;
3853         }
3854 #if 0
3855         printk(KERN_INFO"global block rsv size %llu reserved %llu\n",
3856                 block_rsv->size, block_rsv->reserved);
3857 #endif
3858         spin_unlock(&sinfo->lock);
3859         spin_unlock(&block_rsv->lock);
3860 }
3861
3862 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
3863 {
3864         struct btrfs_space_info *space_info;
3865
3866         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3867         fs_info->chunk_block_rsv.space_info = space_info;
3868         fs_info->chunk_block_rsv.priority = 10;
3869
3870         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3871         fs_info->global_block_rsv.space_info = space_info;
3872         fs_info->global_block_rsv.priority = 10;
3873         fs_info->global_block_rsv.refill_used = 1;
3874         fs_info->delalloc_block_rsv.space_info = space_info;
3875         fs_info->trans_block_rsv.space_info = space_info;
3876         fs_info->empty_block_rsv.space_info = space_info;
3877         fs_info->empty_block_rsv.priority = 10;
3878
3879         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
3880         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
3881         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
3882         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
3883         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
3884
3885         btrfs_add_durable_block_rsv(fs_info, &fs_info->global_block_rsv);
3886
3887         btrfs_add_durable_block_rsv(fs_info, &fs_info->delalloc_block_rsv);
3888
3889         update_global_block_rsv(fs_info);
3890 }
3891
3892 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
3893 {
3894         block_rsv_release_bytes(&fs_info->global_block_rsv, NULL, (u64)-1);
3895         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
3896         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
3897         WARN_ON(fs_info->trans_block_rsv.size > 0);
3898         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
3899         WARN_ON(fs_info->chunk_block_rsv.size > 0);
3900         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
3901 }
3902
3903 static u64 calc_trans_metadata_size(struct btrfs_root *root, int num_items)
3904 {
3905         return (root->leafsize + root->nodesize * (BTRFS_MAX_LEVEL - 1)) *
3906                 3 * num_items;
3907 }
3908
3909 int btrfs_trans_reserve_metadata(struct btrfs_trans_handle *trans,
3910                                  struct btrfs_root *root,
3911                                  int num_items)
3912 {
3913         u64 num_bytes;
3914         int ret;
3915
3916         if (num_items == 0 || root->fs_info->chunk_root == root)
3917                 return 0;
3918
3919         num_bytes = calc_trans_metadata_size(root, num_items);
3920         ret = btrfs_block_rsv_add(trans, root, &root->fs_info->trans_block_rsv,
3921                                   num_bytes);
3922         if (!ret) {
3923                 trans->bytes_reserved += num_bytes;
3924                 trans->block_rsv = &root->fs_info->trans_block_rsv;
3925         }
3926         return ret;
3927 }
3928
3929 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
3930                                   struct btrfs_root *root)
3931 {
3932         if (!trans->bytes_reserved)
3933                 return;
3934
3935         BUG_ON(trans->block_rsv != &root->fs_info->trans_block_rsv);
3936         btrfs_block_rsv_release(root, trans->block_rsv,
3937                                 trans->bytes_reserved);
3938         trans->bytes_reserved = 0;
3939 }
3940
3941 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
3942                                   struct inode *inode)
3943 {
3944         struct btrfs_root *root = BTRFS_I(inode)->root;
3945         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
3946         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
3947
3948         /*
3949          * one for deleting orphan item, one for updating inode and
3950          * two for calling btrfs_truncate_inode_items.
3951          *
3952          * btrfs_truncate_inode_items is a delete operation, it frees
3953          * more space than it uses in most cases. So two units of
3954          * metadata space should be enough for calling it many times.
3955          * If all of the metadata space is used, we can commit
3956          * transaction and use space it freed.
3957          */
3958         u64 num_bytes = calc_trans_metadata_size(root, 4);
3959         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3960 }
3961
3962 void btrfs_orphan_release_metadata(struct inode *inode)
3963 {
3964         struct btrfs_root *root = BTRFS_I(inode)->root;
3965         u64 num_bytes = calc_trans_metadata_size(root, 4);
3966         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
3967 }
3968
3969 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
3970                                 struct btrfs_pending_snapshot *pending)
3971 {
3972         struct btrfs_root *root = pending->root;
3973         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
3974         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
3975         /*
3976          * two for root back/forward refs, two for directory entries
3977          * and one for root of the snapshot.
3978          */
3979         u64 num_bytes = calc_trans_metadata_size(root, 5);
3980         dst_rsv->space_info = src_rsv->space_info;
3981         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3982 }
3983
3984 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes)
3985 {
3986         return num_bytes >>= 3;
3987 }
3988
3989 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
3990 {
3991         struct btrfs_root *root = BTRFS_I(inode)->root;
3992         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
3993         u64 to_reserve;
3994         int nr_extents;
3995         int ret;
3996
3997         if (btrfs_transaction_in_commit(root->fs_info))
3998                 schedule_timeout(1);
3999
4000         num_bytes = ALIGN(num_bytes, root->sectorsize);
4001
4002         spin_lock(&BTRFS_I(inode)->accounting_lock);
4003         nr_extents = atomic_read(&BTRFS_I(inode)->outstanding_extents) + 1;
4004         if (nr_extents > BTRFS_I(inode)->reserved_extents) {
4005                 nr_extents -= BTRFS_I(inode)->reserved_extents;
4006                 to_reserve = calc_trans_metadata_size(root, nr_extents);
4007         } else {
4008                 nr_extents = 0;
4009                 to_reserve = 0;
4010         }
4011         spin_unlock(&BTRFS_I(inode)->accounting_lock);
4012
4013         to_reserve += calc_csum_metadata_size(inode, num_bytes);
4014         ret = reserve_metadata_bytes(NULL, root, block_rsv, to_reserve, 1);
4015         if (ret)
4016                 return ret;
4017
4018         spin_lock(&BTRFS_I(inode)->accounting_lock);
4019         BTRFS_I(inode)->reserved_extents += nr_extents;
4020         atomic_inc(&BTRFS_I(inode)->outstanding_extents);
4021         spin_unlock(&BTRFS_I(inode)->accounting_lock);
4022
4023         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4024
4025         if (block_rsv->size > 512 * 1024 * 1024)
4026                 shrink_delalloc(NULL, root, to_reserve, 0);
4027
4028         return 0;
4029 }
4030
4031 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4032 {
4033         struct btrfs_root *root = BTRFS_I(inode)->root;
4034         u64 to_free;
4035         int nr_extents;
4036
4037         num_bytes = ALIGN(num_bytes, root->sectorsize);
4038         atomic_dec(&BTRFS_I(inode)->outstanding_extents);
4039
4040         spin_lock(&BTRFS_I(inode)->accounting_lock);
4041         nr_extents = atomic_read(&BTRFS_I(inode)->outstanding_extents);
4042         if (nr_extents < BTRFS_I(inode)->reserved_extents) {
4043                 nr_extents = BTRFS_I(inode)->reserved_extents - nr_extents;
4044                 BTRFS_I(inode)->reserved_extents -= nr_extents;
4045         } else {
4046                 nr_extents = 0;
4047         }
4048         spin_unlock(&BTRFS_I(inode)->accounting_lock);
4049
4050         to_free = calc_csum_metadata_size(inode, num_bytes);
4051         if (nr_extents > 0)
4052                 to_free += calc_trans_metadata_size(root, nr_extents);
4053
4054         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4055                                 to_free);
4056 }
4057
4058 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4059 {
4060         int ret;
4061
4062         ret = btrfs_check_data_free_space(inode, num_bytes);
4063         if (ret)
4064                 return ret;
4065
4066         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4067         if (ret) {
4068                 btrfs_free_reserved_data_space(inode, num_bytes);
4069                 return ret;
4070         }
4071
4072         return 0;
4073 }
4074
4075 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4076 {
4077         btrfs_delalloc_release_metadata(inode, num_bytes);
4078         btrfs_free_reserved_data_space(inode, num_bytes);
4079 }
4080
4081 static int update_block_group(struct btrfs_trans_handle *trans,
4082                               struct btrfs_root *root,
4083                               u64 bytenr, u64 num_bytes, int alloc)
4084 {
4085         struct btrfs_block_group_cache *cache = NULL;
4086         struct btrfs_fs_info *info = root->fs_info;
4087         u64 total = num_bytes;
4088         u64 old_val;
4089         u64 byte_in_group;
4090         int factor;
4091
4092         /* block accounting for super block */
4093         spin_lock(&info->delalloc_lock);
4094         old_val = btrfs_super_bytes_used(&info->super_copy);
4095         if (alloc)
4096                 old_val += num_bytes;
4097         else
4098                 old_val -= num_bytes;
4099         btrfs_set_super_bytes_used(&info->super_copy, old_val);
4100         spin_unlock(&info->delalloc_lock);
4101
4102         while (total) {
4103                 cache = btrfs_lookup_block_group(info, bytenr);
4104                 if (!cache)
4105                         return -1;
4106                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4107                                     BTRFS_BLOCK_GROUP_RAID1 |
4108                                     BTRFS_BLOCK_GROUP_RAID10))
4109                         factor = 2;
4110                 else
4111                         factor = 1;
4112                 /*
4113                  * If this block group has free space cache written out, we
4114                  * need to make sure to load it if we are removing space.  This
4115                  * is because we need the unpinning stage to actually add the
4116                  * space back to the block group, otherwise we will leak space.
4117                  */
4118                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4119                         cache_block_group(cache, trans, NULL, 1);
4120
4121                 byte_in_group = bytenr - cache->key.objectid;
4122                 WARN_ON(byte_in_group > cache->key.offset);
4123
4124                 spin_lock(&cache->space_info->lock);
4125                 spin_lock(&cache->lock);
4126
4127                 if (btrfs_super_cache_generation(&info->super_copy) != 0 &&
4128                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4129                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4130
4131                 cache->dirty = 1;
4132                 old_val = btrfs_block_group_used(&cache->item);
4133                 num_bytes = min(total, cache->key.offset - byte_in_group);
4134                 if (alloc) {
4135                         old_val += num_bytes;
4136                         btrfs_set_block_group_used(&cache->item, old_val);
4137                         cache->reserved -= num_bytes;
4138                         cache->space_info->bytes_reserved -= num_bytes;
4139                         cache->space_info->bytes_used += num_bytes;
4140                         cache->space_info->disk_used += num_bytes * factor;
4141                         spin_unlock(&cache->lock);
4142                         spin_unlock(&cache->space_info->lock);
4143                 } else {
4144                         old_val -= num_bytes;
4145                         btrfs_set_block_group_used(&cache->item, old_val);
4146                         cache->pinned += num_bytes;
4147                         cache->space_info->bytes_pinned += num_bytes;
4148                         cache->space_info->bytes_used -= num_bytes;
4149                         cache->space_info->disk_used -= num_bytes * factor;
4150                         spin_unlock(&cache->lock);
4151                         spin_unlock(&cache->space_info->lock);
4152
4153                         set_extent_dirty(info->pinned_extents,
4154                                          bytenr, bytenr + num_bytes - 1,
4155                                          GFP_NOFS | __GFP_NOFAIL);
4156                 }
4157                 btrfs_put_block_group(cache);
4158                 total -= num_bytes;
4159                 bytenr += num_bytes;
4160         }
4161         return 0;
4162 }
4163
4164 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4165 {
4166         struct btrfs_block_group_cache *cache;
4167         u64 bytenr;
4168
4169         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4170         if (!cache)
4171                 return 0;
4172
4173         bytenr = cache->key.objectid;
4174         btrfs_put_block_group(cache);
4175
4176         return bytenr;
4177 }
4178
4179 static int pin_down_extent(struct btrfs_root *root,
4180                            struct btrfs_block_group_cache *cache,
4181                            u64 bytenr, u64 num_bytes, int reserved)
4182 {
4183         spin_lock(&cache->space_info->lock);
4184         spin_lock(&cache->lock);
4185         cache->pinned += num_bytes;
4186         cache->space_info->bytes_pinned += num_bytes;
4187         if (reserved) {
4188                 cache->reserved -= num_bytes;
4189                 cache->space_info->bytes_reserved -= num_bytes;
4190         }
4191         spin_unlock(&cache->lock);
4192         spin_unlock(&cache->space_info->lock);
4193
4194         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4195                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4196         return 0;
4197 }
4198
4199 /*
4200  * this function must be called within transaction
4201  */
4202 int btrfs_pin_extent(struct btrfs_root *root,
4203                      u64 bytenr, u64 num_bytes, int reserved)
4204 {
4205         struct btrfs_block_group_cache *cache;
4206
4207         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4208         BUG_ON(!cache);
4209
4210         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4211
4212         btrfs_put_block_group(cache);
4213         return 0;
4214 }
4215
4216 /*
4217  * update size of reserved extents. this function may return -EAGAIN
4218  * if 'reserve' is true or 'sinfo' is false.
4219  */
4220 static int update_reserved_bytes(struct btrfs_block_group_cache *cache,
4221                                  u64 num_bytes, int reserve, int sinfo)
4222 {
4223         int ret = 0;
4224         if (sinfo) {
4225                 struct btrfs_space_info *space_info = cache->space_info;
4226                 spin_lock(&space_info->lock);
4227                 spin_lock(&cache->lock);
4228                 if (reserve) {
4229                         if (cache->ro) {
4230                                 ret = -EAGAIN;
4231                         } else {
4232                                 cache->reserved += num_bytes;
4233                                 space_info->bytes_reserved += num_bytes;
4234                         }
4235                 } else {
4236                         if (cache->ro)
4237                                 space_info->bytes_readonly += num_bytes;
4238                         cache->reserved -= num_bytes;
4239                         space_info->bytes_reserved -= num_bytes;
4240                 }
4241                 spin_unlock(&cache->lock);
4242                 spin_unlock(&space_info->lock);
4243         } else {
4244                 spin_lock(&cache->lock);
4245                 if (cache->ro) {
4246                         ret = -EAGAIN;
4247                 } else {
4248                         if (reserve)
4249                                 cache->reserved += num_bytes;
4250                         else
4251                                 cache->reserved -= num_bytes;
4252                 }
4253                 spin_unlock(&cache->lock);
4254         }
4255         return ret;
4256 }
4257
4258 int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4259                                 struct btrfs_root *root)
4260 {
4261         struct btrfs_fs_info *fs_info = root->fs_info;
4262         struct btrfs_caching_control *next;
4263         struct btrfs_caching_control *caching_ctl;
4264         struct btrfs_block_group_cache *cache;
4265
4266         down_write(&fs_info->extent_commit_sem);
4267
4268         list_for_each_entry_safe(caching_ctl, next,
4269                                  &fs_info->caching_block_groups, list) {
4270                 cache = caching_ctl->block_group;
4271                 if (block_group_cache_done(cache)) {
4272                         cache->last_byte_to_unpin = (u64)-1;
4273                         list_del_init(&caching_ctl->list);
4274                         put_caching_control(caching_ctl);
4275                 } else {
4276                         cache->last_byte_to_unpin = caching_ctl->progress;
4277                 }
4278         }
4279
4280         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4281                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4282         else
4283                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4284
4285         up_write(&fs_info->extent_commit_sem);
4286
4287         update_global_block_rsv(fs_info);
4288         return 0;
4289 }
4290
4291 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4292 {
4293         struct btrfs_fs_info *fs_info = root->fs_info;
4294         struct btrfs_block_group_cache *cache = NULL;
4295         u64 len;
4296
4297         while (start <= end) {
4298                 if (!cache ||
4299                     start >= cache->key.objectid + cache->key.offset) {
4300                         if (cache)
4301                                 btrfs_put_block_group(cache);
4302                         cache = btrfs_lookup_block_group(fs_info, start);
4303                         BUG_ON(!cache);
4304                 }
4305
4306                 len = cache->key.objectid + cache->key.offset - start;
4307                 len = min(len, end + 1 - start);
4308
4309                 if (start < cache->last_byte_to_unpin) {
4310                         len = min(len, cache->last_byte_to_unpin - start);
4311                         btrfs_add_free_space(cache, start, len);
4312                 }
4313
4314                 start += len;
4315
4316                 spin_lock(&cache->space_info->lock);
4317                 spin_lock(&cache->lock);
4318                 cache->pinned -= len;
4319                 cache->space_info->bytes_pinned -= len;
4320                 if (cache->ro) {
4321                         cache->space_info->bytes_readonly += len;
4322                 } else if (cache->reserved_pinned > 0) {
4323                         len = min(len, cache->reserved_pinned);
4324                         cache->reserved_pinned -= len;
4325                         cache->space_info->bytes_reserved += len;
4326                 }
4327                 spin_unlock(&cache->lock);
4328                 spin_unlock(&cache->space_info->lock);
4329         }
4330
4331         if (cache)
4332                 btrfs_put_block_group(cache);
4333         return 0;
4334 }
4335
4336 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4337                                struct btrfs_root *root)
4338 {
4339         struct btrfs_fs_info *fs_info = root->fs_info;
4340         struct extent_io_tree *unpin;
4341         struct btrfs_block_rsv *block_rsv;
4342         struct btrfs_block_rsv *next_rsv;
4343         u64 start;
4344         u64 end;
4345         int idx;
4346         int ret;
4347
4348         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4349                 unpin = &fs_info->freed_extents[1];
4350         else
4351                 unpin = &fs_info->freed_extents[0];
4352
4353         while (1) {
4354                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4355                                             EXTENT_DIRTY);
4356                 if (ret)
4357                         break;
4358
4359                 ret = btrfs_discard_extent(root, start, end + 1 - start);
4360
4361                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4362                 unpin_extent_range(root, start, end);
4363                 cond_resched();
4364         }
4365
4366         mutex_lock(&fs_info->durable_block_rsv_mutex);
4367         list_for_each_entry_safe(block_rsv, next_rsv,
4368                                  &fs_info->durable_block_rsv_list, list) {
4369
4370                 idx = trans->transid & 0x1;
4371                 if (block_rsv->freed[idx] > 0) {
4372                         block_rsv_add_bytes(block_rsv,
4373                                             block_rsv->freed[idx], 0);
4374                         block_rsv->freed[idx] = 0;
4375                 }
4376                 if (atomic_read(&block_rsv->usage) == 0) {
4377                         btrfs_block_rsv_release(root, block_rsv, (u64)-1);
4378
4379                         if (block_rsv->freed[0] == 0 &&
4380                             block_rsv->freed[1] == 0) {
4381                                 list_del_init(&block_rsv->list);
4382                                 kfree(block_rsv);
4383                         }
4384                 } else {
4385                         btrfs_block_rsv_release(root, block_rsv, 0);
4386                 }
4387         }
4388         mutex_unlock(&fs_info->durable_block_rsv_mutex);
4389
4390         return 0;
4391 }
4392
4393 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4394                                 struct btrfs_root *root,
4395                                 u64 bytenr, u64 num_bytes, u64 parent,
4396                                 u64 root_objectid, u64 owner_objectid,
4397                                 u64 owner_offset, int refs_to_drop,
4398                                 struct btrfs_delayed_extent_op *extent_op)
4399 {
4400         struct btrfs_key key;
4401         struct btrfs_path *path;
4402         struct btrfs_fs_info *info = root->fs_info;
4403         struct btrfs_root *extent_root = info->extent_root;
4404         struct extent_buffer *leaf;
4405         struct btrfs_extent_item *ei;
4406         struct btrfs_extent_inline_ref *iref;
4407         int ret;
4408         int is_data;
4409         int extent_slot = 0;
4410         int found_extent = 0;
4411         int num_to_del = 1;
4412         u32 item_size;
4413         u64 refs;
4414
4415         path = btrfs_alloc_path();
4416         if (!path)
4417                 return -ENOMEM;
4418
4419         path->reada = 1;
4420         path->leave_spinning = 1;
4421
4422         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4423         BUG_ON(!is_data && refs_to_drop != 1);
4424
4425         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4426                                     bytenr, num_bytes, parent,
4427                                     root_objectid, owner_objectid,
4428                                     owner_offset);
4429         if (ret == 0) {
4430                 extent_slot = path->slots[0];
4431                 while (extent_slot >= 0) {
4432                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4433                                               extent_slot);
4434                         if (key.objectid != bytenr)
4435                                 break;
4436                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4437                             key.offset == num_bytes) {
4438                                 found_extent = 1;
4439                                 break;
4440                         }
4441                         if (path->slots[0] - extent_slot > 5)
4442                                 break;
4443                         extent_slot--;
4444                 }
4445 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4446                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4447                 if (found_extent && item_size < sizeof(*ei))
4448                         found_extent = 0;
4449 #endif
4450                 if (!found_extent) {
4451                         BUG_ON(iref);
4452                         ret = remove_extent_backref(trans, extent_root, path,
4453                                                     NULL, refs_to_drop,
4454                                                     is_data);
4455                         BUG_ON(ret);
4456                         btrfs_release_path(extent_root, path);
4457                         path->leave_spinning = 1;
4458
4459                         key.objectid = bytenr;
4460                         key.type = BTRFS_EXTENT_ITEM_KEY;
4461                         key.offset = num_bytes;
4462
4463                         ret = btrfs_search_slot(trans, extent_root,
4464                                                 &key, path, -1, 1);
4465                         if (ret) {
4466                                 printk(KERN_ERR "umm, got %d back from search"
4467                                        ", was looking for %llu\n", ret,
4468                                        (unsigned long long)bytenr);
4469                                 btrfs_print_leaf(extent_root, path->nodes[0]);
4470                         }
4471                         BUG_ON(ret);
4472                         extent_slot = path->slots[0];
4473                 }
4474         } else {
4475                 btrfs_print_leaf(extent_root, path->nodes[0]);
4476                 WARN_ON(1);
4477                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4478                        "parent %llu root %llu  owner %llu offset %llu\n",
4479                        (unsigned long long)bytenr,
4480                        (unsigned long long)parent,
4481                        (unsigned long long)root_objectid,
4482                        (unsigned long long)owner_objectid,
4483                        (unsigned long long)owner_offset);
4484         }
4485
4486         leaf = path->nodes[0];
4487         item_size = btrfs_item_size_nr(leaf, extent_slot);
4488 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4489         if (item_size < sizeof(*ei)) {
4490                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4491                 ret = convert_extent_item_v0(trans, extent_root, path,
4492                                              owner_objectid, 0);
4493                 BUG_ON(ret < 0);
4494
4495                 btrfs_release_path(extent_root, path);
4496                 path->leave_spinning = 1;
4497
4498                 key.objectid = bytenr;
4499                 key.type = BTRFS_EXTENT_ITEM_KEY;
4500                 key.offset = num_bytes;
4501
4502                 ret = btrfs_search_slot(trans, extent_root, &key, path,
4503                                         -1, 1);
4504                 if (ret) {
4505                         printk(KERN_ERR "umm, got %d back from search"
4506                                ", was looking for %llu\n", ret,
4507                                (unsigned long long)bytenr);
4508                         btrfs_print_leaf(extent_root, path->nodes[0]);
4509                 }
4510                 BUG_ON(ret);
4511                 extent_slot = path->slots[0];
4512                 leaf = path->nodes[0];
4513                 item_size = btrfs_item_size_nr(leaf, extent_slot);
4514         }
4515 #endif
4516         BUG_ON(item_size < sizeof(*ei));
4517         ei = btrfs_item_ptr(leaf, extent_slot,
4518                             struct btrfs_extent_item);
4519         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
4520                 struct btrfs_tree_block_info *bi;
4521                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4522                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4523                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4524         }
4525
4526         refs = btrfs_extent_refs(leaf, ei);
4527         BUG_ON(refs < refs_to_drop);
4528         refs -= refs_to_drop;
4529
4530         if (refs > 0) {
4531                 if (extent_op)
4532                         __run_delayed_extent_op(extent_op, leaf, ei);
4533                 /*
4534                  * In the case of inline back ref, reference count will
4535                  * be updated by remove_extent_backref
4536                  */
4537                 if (iref) {
4538                         BUG_ON(!found_extent);
4539                 } else {
4540                         btrfs_set_extent_refs(leaf, ei, refs);
4541                         btrfs_mark_buffer_dirty(leaf);
4542                 }
4543                 if (found_extent) {
4544                         ret = remove_extent_backref(trans, extent_root, path,
4545                                                     iref, refs_to_drop,
4546                                                     is_data);
4547                         BUG_ON(ret);
4548                 }
4549         } else {
4550                 if (found_extent) {
4551                         BUG_ON(is_data && refs_to_drop !=
4552                                extent_data_ref_count(root, path, iref));
4553                         if (iref) {
4554                                 BUG_ON(path->slots[0] != extent_slot);
4555                         } else {
4556                                 BUG_ON(path->slots[0] != extent_slot + 1);
4557                                 path->slots[0] = extent_slot;
4558                                 num_to_del = 2;
4559                         }
4560                 }
4561
4562                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
4563                                       num_to_del);
4564                 BUG_ON(ret);
4565                 btrfs_release_path(extent_root, path);
4566
4567                 if (is_data) {
4568                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
4569                         BUG_ON(ret);
4570                 } else {
4571                         invalidate_mapping_pages(info->btree_inode->i_mapping,
4572                              bytenr >> PAGE_CACHE_SHIFT,
4573                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
4574                 }
4575
4576                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
4577                 BUG_ON(ret);
4578         }
4579         btrfs_free_path(path);
4580         return ret;
4581 }
4582
4583 /*
4584  * when we free an block, it is possible (and likely) that we free the last
4585  * delayed ref for that extent as well.  This searches the delayed ref tree for
4586  * a given extent, and if there are no other delayed refs to be processed, it
4587  * removes it from the tree.
4588  */
4589 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
4590                                       struct btrfs_root *root, u64 bytenr)
4591 {
4592         struct btrfs_delayed_ref_head *head;
4593         struct btrfs_delayed_ref_root *delayed_refs;
4594         struct btrfs_delayed_ref_node *ref;
4595         struct rb_node *node;
4596         int ret = 0;
4597
4598         delayed_refs = &trans->transaction->delayed_refs;
4599         spin_lock(&delayed_refs->lock);
4600         head = btrfs_find_delayed_ref_head(trans, bytenr);
4601         if (!head)
4602                 goto out;
4603
4604         node = rb_prev(&head->node.rb_node);
4605         if (!node)
4606                 goto out;
4607
4608         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
4609
4610         /* there are still entries for this ref, we can't drop it */
4611         if (ref->bytenr == bytenr)
4612                 goto out;
4613
4614         if (head->extent_op) {
4615                 if (!head->must_insert_reserved)
4616                         goto out;
4617                 kfree(head->extent_op);
4618                 head->extent_op = NULL;
4619         }
4620
4621         /*
4622          * waiting for the lock here would deadlock.  If someone else has it
4623          * locked they are already in the process of dropping it anyway
4624          */
4625         if (!mutex_trylock(&head->mutex))
4626                 goto out;
4627
4628         /*
4629          * at this point we have a head with no other entries.  Go
4630          * ahead and process it.
4631          */
4632         head->node.in_tree = 0;
4633         rb_erase(&head->node.rb_node, &delayed_refs->root);
4634
4635         delayed_refs->num_entries--;
4636
4637         /*
4638          * we don't take a ref on the node because we're removing it from the
4639          * tree, so we just steal the ref the tree was holding.
4640          */
4641         delayed_refs->num_heads--;
4642         if (list_empty(&head->cluster))
4643                 delayed_refs->num_heads_ready--;
4644
4645         list_del_init(&head->cluster);
4646         spin_unlock(&delayed_refs->lock);
4647
4648         BUG_ON(head->extent_op);
4649         if (head->must_insert_reserved)
4650                 ret = 1;
4651
4652         mutex_unlock(&head->mutex);
4653         btrfs_put_delayed_ref(&head->node);
4654         return ret;
4655 out:
4656         spin_unlock(&delayed_refs->lock);
4657         return 0;
4658 }
4659
4660 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
4661                            struct btrfs_root *root,
4662                            struct extent_buffer *buf,
4663                            u64 parent, int last_ref)
4664 {
4665         struct btrfs_block_rsv *block_rsv;
4666         struct btrfs_block_group_cache *cache = NULL;
4667         int ret;
4668
4669         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4670                 ret = btrfs_add_delayed_tree_ref(trans, buf->start, buf->len,
4671                                                 parent, root->root_key.objectid,
4672                                                 btrfs_header_level(buf),
4673                                                 BTRFS_DROP_DELAYED_REF, NULL);
4674                 BUG_ON(ret);
4675         }
4676
4677         if (!last_ref)
4678                 return;
4679
4680         block_rsv = get_block_rsv(trans, root);
4681         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
4682         if (block_rsv->space_info != cache->space_info)
4683                 goto out;
4684
4685         if (btrfs_header_generation(buf) == trans->transid) {
4686                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4687                         ret = check_ref_cleanup(trans, root, buf->start);
4688                         if (!ret)
4689                                 goto pin;
4690                 }
4691
4692                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
4693                         pin_down_extent(root, cache, buf->start, buf->len, 1);
4694                         goto pin;
4695                 }
4696
4697                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
4698
4699                 btrfs_add_free_space(cache, buf->start, buf->len);
4700                 ret = update_reserved_bytes(cache, buf->len, 0, 0);
4701                 if (ret == -EAGAIN) {
4702                         /* block group became read-only */
4703                         update_reserved_bytes(cache, buf->len, 0, 1);
4704                         goto out;
4705                 }
4706
4707                 ret = 1;
4708                 spin_lock(&block_rsv->lock);
4709                 if (block_rsv->reserved < block_rsv->size) {
4710                         block_rsv->reserved += buf->len;
4711                         ret = 0;
4712                 }
4713                 spin_unlock(&block_rsv->lock);
4714
4715                 if (ret) {
4716                         spin_lock(&cache->space_info->lock);
4717                         cache->space_info->bytes_reserved -= buf->len;
4718                         spin_unlock(&cache->space_info->lock);
4719                 }
4720                 goto out;
4721         }
4722 pin:
4723         if (block_rsv->durable && !cache->ro) {
4724                 ret = 0;
4725                 spin_lock(&cache->lock);
4726                 if (!cache->ro) {
4727                         cache->reserved_pinned += buf->len;
4728                         ret = 1;
4729                 }
4730                 spin_unlock(&cache->lock);
4731
4732                 if (ret) {
4733                         spin_lock(&block_rsv->lock);
4734                         block_rsv->freed[trans->transid & 0x1] += buf->len;
4735                         spin_unlock(&block_rsv->lock);
4736                 }
4737         }
4738 out:
4739         btrfs_put_block_group(cache);
4740 }
4741
4742 int btrfs_free_extent(struct btrfs_trans_handle *trans,
4743                       struct btrfs_root *root,
4744                       u64 bytenr, u64 num_bytes, u64 parent,
4745                       u64 root_objectid, u64 owner, u64 offset)
4746 {
4747         int ret;
4748
4749         /*
4750          * tree log blocks never actually go into the extent allocation
4751          * tree, just update pinning info and exit early.
4752          */
4753         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
4754                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
4755                 /* unlocks the pinned mutex */
4756                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
4757                 ret = 0;
4758         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
4759                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
4760                                         parent, root_objectid, (int)owner,
4761                                         BTRFS_DROP_DELAYED_REF, NULL);
4762                 BUG_ON(ret);
4763         } else {
4764                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
4765                                         parent, root_objectid, owner,
4766                                         offset, BTRFS_DROP_DELAYED_REF, NULL);
4767                 BUG_ON(ret);
4768         }
4769         return ret;
4770 }
4771
4772 static u64 stripe_align(struct btrfs_root *root, u64 val)
4773 {
4774         u64 mask = ((u64)root->stripesize - 1);
4775         u64 ret = (val + mask) & ~mask;
4776         return ret;
4777 }
4778
4779 /*
4780  * when we wait for progress in the block group caching, its because
4781  * our allocation attempt failed at least once.  So, we must sleep
4782  * and let some progress happen before we try again.
4783  *
4784  * This function will sleep at least once waiting for new free space to
4785  * show up, and then it will check the block group free space numbers
4786  * for our min num_bytes.  Another option is to have it go ahead
4787  * and look in the rbtree for a free extent of a given size, but this
4788  * is a good start.
4789  */
4790 static noinline int
4791 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
4792                                 u64 num_bytes)
4793 {
4794         struct btrfs_caching_control *caching_ctl;
4795         DEFINE_WAIT(wait);
4796
4797         caching_ctl = get_caching_control(cache);
4798         if (!caching_ctl)
4799                 return 0;
4800
4801         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
4802                    (cache->free_space >= num_bytes));
4803
4804         put_caching_control(caching_ctl);
4805         return 0;
4806 }
4807
4808 static noinline int
4809 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
4810 {
4811         struct btrfs_caching_control *caching_ctl;
4812         DEFINE_WAIT(wait);
4813
4814         caching_ctl = get_caching_control(cache);
4815         if (!caching_ctl)
4816                 return 0;
4817
4818         wait_event(caching_ctl->wait, block_group_cache_done(cache));
4819
4820         put_caching_control(caching_ctl);
4821         return 0;
4822 }
4823
4824 static int get_block_group_index(struct btrfs_block_group_cache *cache)
4825 {
4826         int index;
4827         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
4828                 index = 0;
4829         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
4830                 index = 1;
4831         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
4832                 index = 2;
4833         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
4834                 index = 3;
4835         else
4836                 index = 4;
4837         return index;
4838 }
4839
4840 enum btrfs_loop_type {
4841         LOOP_FIND_IDEAL = 0,
4842         LOOP_CACHING_NOWAIT = 1,
4843         LOOP_CACHING_WAIT = 2,
4844         LOOP_ALLOC_CHUNK = 3,
4845         LOOP_NO_EMPTY_SIZE = 4,
4846 };
4847
4848 /*
4849  * walks the btree of allocated extents and find a hole of a given size.
4850  * The key ins is changed to record the hole:
4851  * ins->objectid == block start
4852  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4853  * ins->offset == number of blocks
4854  * Any available blocks before search_start are skipped.
4855  */
4856 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
4857                                      struct btrfs_root *orig_root,
4858                                      u64 num_bytes, u64 empty_size,
4859                                      u64 search_start, u64 search_end,
4860                                      u64 hint_byte, struct btrfs_key *ins,
4861                                      int data)
4862 {
4863         int ret = 0;
4864         struct btrfs_root *root = orig_root->fs_info->extent_root;
4865         struct btrfs_free_cluster *last_ptr = NULL;
4866         struct btrfs_block_group_cache *block_group = NULL;
4867         int empty_cluster = 2 * 1024 * 1024;
4868         int allowed_chunk_alloc = 0;
4869         int done_chunk_alloc = 0;
4870         struct btrfs_space_info *space_info;
4871         int last_ptr_loop = 0;
4872         int loop = 0;
4873         int index = 0;
4874         bool found_uncached_bg = false;
4875         bool failed_cluster_refill = false;
4876         bool failed_alloc = false;
4877         bool use_cluster = true;
4878         u64 ideal_cache_percent = 0;
4879         u64 ideal_cache_offset = 0;
4880
4881         WARN_ON(num_bytes < root->sectorsize);
4882         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
4883         ins->objectid = 0;
4884         ins->offset = 0;
4885
4886         space_info = __find_space_info(root->fs_info, data);
4887         if (!space_info) {
4888                 printk(KERN_ERR "No space info for %d\n", data);
4889                 return -ENOSPC;
4890         }
4891
4892         /*
4893          * If the space info is for both data and metadata it means we have a
4894          * small filesystem and we can't use the clustering stuff.
4895          */
4896         if (btrfs_mixed_space_info(space_info))
4897                 use_cluster = false;
4898
4899         if (orig_root->ref_cows || empty_size)
4900                 allowed_chunk_alloc = 1;
4901
4902         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
4903                 last_ptr = &root->fs_info->meta_alloc_cluster;
4904                 if (!btrfs_test_opt(root, SSD))
4905                         empty_cluster = 64 * 1024;
4906         }
4907
4908         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
4909             btrfs_test_opt(root, SSD)) {
4910                 last_ptr = &root->fs_info->data_alloc_cluster;
4911         }
4912
4913         if (last_ptr) {
4914                 spin_lock(&last_ptr->lock);
4915                 if (last_ptr->block_group)
4916                         hint_byte = last_ptr->window_start;
4917                 spin_unlock(&last_ptr->lock);
4918         }
4919
4920         search_start = max(search_start, first_logical_byte(root, 0));
4921         search_start = max(search_start, hint_byte);
4922
4923         if (!last_ptr)
4924                 empty_cluster = 0;
4925
4926         if (search_start == hint_byte) {
4927 ideal_cache:
4928                 block_group = btrfs_lookup_block_group(root->fs_info,
4929                                                        search_start);
4930                 /*
4931                  * we don't want to use the block group if it doesn't match our
4932                  * allocation bits, or if its not cached.
4933                  *
4934                  * However if we are re-searching with an ideal block group
4935                  * picked out then we don't care that the block group is cached.
4936                  */
4937                 if (block_group && block_group_bits(block_group, data) &&
4938                     (block_group->cached != BTRFS_CACHE_NO ||
4939                      search_start == ideal_cache_offset)) {
4940                         down_read(&space_info->groups_sem);
4941                         if (list_empty(&block_group->list) ||
4942                             block_group->ro) {
4943                                 /*
4944                                  * someone is removing this block group,
4945                                  * we can't jump into the have_block_group
4946                                  * target because our list pointers are not
4947                                  * valid
4948                                  */
4949                                 btrfs_put_block_group(block_group);
4950                                 up_read(&space_info->groups_sem);
4951                         } else {
4952                                 index = get_block_group_index(block_group);
4953                                 goto have_block_group;
4954                         }
4955                 } else if (block_group) {
4956                         btrfs_put_block_group(block_group);
4957                 }
4958         }
4959 search:
4960         down_read(&space_info->groups_sem);
4961         list_for_each_entry(block_group, &space_info->block_groups[index],
4962                             list) {
4963                 u64 offset;
4964                 int cached;
4965
4966                 btrfs_get_block_group(block_group);
4967                 search_start = block_group->key.objectid;
4968
4969                 /*
4970                  * this can happen if we end up cycling through all the
4971                  * raid types, but we want to make sure we only allocate
4972                  * for the proper type.
4973                  */
4974                 if (!block_group_bits(block_group, data)) {
4975                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
4976                                 BTRFS_BLOCK_GROUP_RAID1 |
4977                                 BTRFS_BLOCK_GROUP_RAID10;
4978
4979                         /*
4980                          * if they asked for extra copies and this block group
4981                          * doesn't provide them, bail.  This does allow us to
4982                          * fill raid0 from raid1.
4983                          */
4984                         if ((data & extra) && !(block_group->flags & extra))
4985                                 goto loop;
4986                 }
4987
4988 have_block_group:
4989                 if (unlikely(block_group->cached == BTRFS_CACHE_NO)) {
4990                         u64 free_percent;
4991
4992                         ret = cache_block_group(block_group, trans,
4993                                                 orig_root, 1);
4994                         if (block_group->cached == BTRFS_CACHE_FINISHED)
4995                                 goto have_block_group;
4996
4997                         free_percent = btrfs_block_group_used(&block_group->item);
4998                         free_percent *= 100;
4999                         free_percent = div64_u64(free_percent,
5000                                                  block_group->key.offset);
5001                         free_percent = 100 - free_percent;
5002                         if (free_percent > ideal_cache_percent &&
5003                             likely(!block_group->ro)) {
5004                                 ideal_cache_offset = block_group->key.objectid;
5005                                 ideal_cache_percent = free_percent;
5006                         }
5007
5008                         /*
5009                          * We only want to start kthread caching if we are at
5010                          * the point where we will wait for caching to make
5011                          * progress, or if our ideal search is over and we've
5012                          * found somebody to start caching.
5013                          */
5014                         if (loop > LOOP_CACHING_NOWAIT ||
5015                             (loop > LOOP_FIND_IDEAL &&
5016                              atomic_read(&space_info->caching_threads) < 2)) {
5017                                 ret = cache_block_group(block_group, trans,
5018                                                         orig_root, 0);
5019                                 BUG_ON(ret);
5020                         }
5021                         found_uncached_bg = true;
5022
5023                         /*
5024                          * If loop is set for cached only, try the next block
5025                          * group.
5026                          */
5027                         if (loop == LOOP_FIND_IDEAL)
5028                                 goto loop;
5029                 }
5030
5031                 cached = block_group_cache_done(block_group);
5032                 if (unlikely(!cached))
5033                         found_uncached_bg = true;
5034
5035                 if (unlikely(block_group->ro))
5036                         goto loop;
5037
5038                 /*
5039                  * Ok we want to try and use the cluster allocator, so lets look
5040                  * there, unless we are on LOOP_NO_EMPTY_SIZE, since we will
5041                  * have tried the cluster allocator plenty of times at this
5042                  * point and not have found anything, so we are likely way too
5043                  * fragmented for the clustering stuff to find anything, so lets
5044                  * just skip it and let the allocator find whatever block it can
5045                  * find
5046                  */
5047                 if (last_ptr && loop < LOOP_NO_EMPTY_SIZE) {
5048                         /*
5049                          * the refill lock keeps out other
5050                          * people trying to start a new cluster
5051                          */
5052                         spin_lock(&last_ptr->refill_lock);
5053                         if (last_ptr->block_group &&
5054                             (last_ptr->block_group->ro ||
5055                             !block_group_bits(last_ptr->block_group, data))) {
5056                                 offset = 0;
5057                                 goto refill_cluster;
5058                         }
5059
5060                         offset = btrfs_alloc_from_cluster(block_group, last_ptr,
5061                                                  num_bytes, search_start);
5062                         if (offset) {
5063                                 /* we have a block, we're done */
5064                                 spin_unlock(&last_ptr->refill_lock);
5065                                 goto checks;
5066                         }
5067
5068                         spin_lock(&last_ptr->lock);
5069                         /*
5070                          * whoops, this cluster doesn't actually point to
5071                          * this block group.  Get a ref on the block
5072                          * group is does point to and try again
5073                          */
5074                         if (!last_ptr_loop && last_ptr->block_group &&
5075                             last_ptr->block_group != block_group) {
5076
5077                                 btrfs_put_block_group(block_group);
5078                                 block_group = last_ptr->block_group;
5079                                 btrfs_get_block_group(block_group);
5080                                 spin_unlock(&last_ptr->lock);
5081                                 spin_unlock(&last_ptr->refill_lock);
5082
5083                                 last_ptr_loop = 1;
5084                                 search_start = block_group->key.objectid;
5085                                 /*
5086                                  * we know this block group is properly
5087                                  * in the list because
5088                                  * btrfs_remove_block_group, drops the
5089                                  * cluster before it removes the block
5090                                  * group from the list
5091                                  */
5092                                 goto have_block_group;
5093                         }
5094                         spin_unlock(&last_ptr->lock);
5095 refill_cluster:
5096                         /*
5097                          * this cluster didn't work out, free it and
5098                          * start over
5099                          */
5100                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5101
5102                         last_ptr_loop = 0;
5103
5104                         /* allocate a cluster in this block group */
5105                         ret = btrfs_find_space_cluster(trans, root,
5106                                                block_group, last_ptr,
5107                                                offset, num_bytes,
5108                                                empty_cluster + empty_size);
5109                         if (ret == 0) {
5110                                 /*
5111                                  * now pull our allocation out of this
5112                                  * cluster
5113                                  */
5114                                 offset = btrfs_alloc_from_cluster(block_group,
5115                                                   last_ptr, num_bytes,
5116                                                   search_start);
5117                                 if (offset) {
5118                                         /* we found one, proceed */
5119                                         spin_unlock(&last_ptr->refill_lock);
5120                                         goto checks;
5121                                 }
5122                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5123                                    && !failed_cluster_refill) {
5124                                 spin_unlock(&last_ptr->refill_lock);
5125
5126                                 failed_cluster_refill = true;
5127                                 wait_block_group_cache_progress(block_group,
5128                                        num_bytes + empty_cluster + empty_size);
5129                                 goto have_block_group;
5130                         }
5131
5132                         /*
5133                          * at this point we either didn't find a cluster
5134                          * or we weren't able to allocate a block from our
5135                          * cluster.  Free the cluster we've been trying
5136                          * to use, and go to the next block group
5137                          */
5138                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5139                         spin_unlock(&last_ptr->refill_lock);
5140                         goto loop;
5141                 }
5142
5143                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5144                                                     num_bytes, empty_size);
5145                 /*
5146                  * If we didn't find a chunk, and we haven't failed on this
5147                  * block group before, and this block group is in the middle of
5148                  * caching and we are ok with waiting, then go ahead and wait
5149                  * for progress to be made, and set failed_alloc to true.
5150                  *
5151                  * If failed_alloc is true then we've already waited on this
5152                  * block group once and should move on to the next block group.
5153                  */
5154                 if (!offset && !failed_alloc && !cached &&
5155                     loop > LOOP_CACHING_NOWAIT) {
5156                         wait_block_group_cache_progress(block_group,
5157                                                 num_bytes + empty_size);
5158                         failed_alloc = true;
5159                         goto have_block_group;
5160                 } else if (!offset) {
5161                         goto loop;
5162                 }
5163 checks:
5164                 search_start = stripe_align(root, offset);
5165                 /* move on to the next group */
5166                 if (search_start + num_bytes >= search_end) {
5167                         btrfs_add_free_space(block_group, offset, num_bytes);
5168                         goto loop;
5169                 }
5170
5171                 /* move on to the next group */
5172                 if (search_start + num_bytes >
5173                     block_group->key.objectid + block_group->key.offset) {
5174                         btrfs_add_free_space(block_group, offset, num_bytes);
5175                         goto loop;
5176                 }
5177
5178                 ins->objectid = search_start;
5179                 ins->offset = num_bytes;
5180
5181                 if (offset < search_start)
5182                         btrfs_add_free_space(block_group, offset,
5183                                              search_start - offset);
5184                 BUG_ON(offset > search_start);
5185
5186                 ret = update_reserved_bytes(block_group, num_bytes, 1,
5187                                             (data & BTRFS_BLOCK_GROUP_DATA));
5188                 if (ret == -EAGAIN) {
5189                         btrfs_add_free_space(block_group, offset, num_bytes);
5190                         goto loop;
5191                 }
5192
5193                 /* we are all good, lets return */
5194                 ins->objectid = search_start;
5195                 ins->offset = num_bytes;
5196
5197                 if (offset < search_start)
5198                         btrfs_add_free_space(block_group, offset,
5199                                              search_start - offset);
5200                 BUG_ON(offset > search_start);
5201                 break;
5202 loop:
5203                 failed_cluster_refill = false;
5204                 failed_alloc = false;
5205                 BUG_ON(index != get_block_group_index(block_group));
5206                 btrfs_put_block_group(block_group);
5207         }
5208         up_read(&space_info->groups_sem);
5209
5210         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5211                 goto search;
5212
5213         /* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
5214          *                      for them to make caching progress.  Also
5215          *                      determine the best possible bg to cache
5216          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5217          *                      caching kthreads as we move along
5218          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5219          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5220          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5221          *                      again
5222          */
5223         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE &&
5224             (found_uncached_bg || empty_size || empty_cluster ||
5225              allowed_chunk_alloc)) {
5226                 index = 0;
5227                 if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
5228                         found_uncached_bg = false;
5229                         loop++;
5230                         if (!ideal_cache_percent &&
5231                             atomic_read(&space_info->caching_threads))
5232                                 goto search;
5233
5234                         /*
5235                          * 1 of the following 2 things have happened so far
5236                          *
5237                          * 1) We found an ideal block group for caching that
5238                          * is mostly full and will cache quickly, so we might
5239                          * as well wait for it.
5240                          *
5241                          * 2) We searched for cached only and we didn't find
5242                          * anything, and we didn't start any caching kthreads
5243                          * either, so chances are we will loop through and
5244                          * start a couple caching kthreads, and then come back
5245                          * around and just wait for them.  This will be slower
5246                          * because we will have 2 caching kthreads reading at
5247                          * the same time when we could have just started one
5248                          * and waited for it to get far enough to give us an
5249                          * allocation, so go ahead and go to the wait caching
5250                          * loop.
5251                          */
5252                         loop = LOOP_CACHING_WAIT;
5253                         search_start = ideal_cache_offset;
5254                         ideal_cache_percent = 0;
5255                         goto ideal_cache;
5256                 } else if (loop == LOOP_FIND_IDEAL) {
5257                         /*
5258                          * Didn't find a uncached bg, wait on anything we find
5259                          * next.
5260                          */
5261                         loop = LOOP_CACHING_WAIT;
5262                         goto search;
5263                 }
5264
5265                 if (loop < LOOP_CACHING_WAIT) {
5266                         loop++;
5267                         goto search;
5268                 }
5269
5270                 if (loop == LOOP_ALLOC_CHUNK) {
5271                         empty_size = 0;
5272                         empty_cluster = 0;
5273                 }
5274
5275                 if (allowed_chunk_alloc) {
5276                         ret = do_chunk_alloc(trans, root, num_bytes +
5277                                              2 * 1024 * 1024, data, 1);
5278                         allowed_chunk_alloc = 0;
5279                         done_chunk_alloc = 1;
5280                 } else if (!done_chunk_alloc) {
5281                         space_info->force_alloc = 1;
5282                 }
5283
5284                 if (loop < LOOP_NO_EMPTY_SIZE) {
5285                         loop++;
5286                         goto search;
5287                 }
5288                 ret = -ENOSPC;
5289         } else if (!ins->objectid) {
5290                 ret = -ENOSPC;
5291         }
5292
5293         /* we found what we needed */
5294         if (ins->objectid) {
5295                 if (!(data & BTRFS_BLOCK_GROUP_DATA))
5296                         trans->block_group = block_group->key.objectid;
5297
5298                 btrfs_put_block_group(block_group);
5299                 ret = 0;
5300         }
5301
5302         return ret;
5303 }
5304
5305 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5306                             int dump_block_groups)
5307 {
5308         struct btrfs_block_group_cache *cache;
5309         int index = 0;
5310
5311         spin_lock(&info->lock);
5312         printk(KERN_INFO "space_info has %llu free, is %sfull\n",
5313                (unsigned long long)(info->total_bytes - info->bytes_used -
5314                                     info->bytes_pinned - info->bytes_reserved -
5315                                     info->bytes_readonly),
5316                (info->full) ? "" : "not ");
5317         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5318                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5319                (unsigned long long)info->total_bytes,
5320                (unsigned long long)info->bytes_used,
5321                (unsigned long long)info->bytes_pinned,
5322                (unsigned long long)info->bytes_reserved,
5323                (unsigned long long)info->bytes_may_use,
5324                (unsigned long long)info->bytes_readonly);
5325         spin_unlock(&info->lock);
5326
5327         if (!dump_block_groups)
5328                 return;
5329
5330         down_read(&info->groups_sem);
5331 again:
5332         list_for_each_entry(cache, &info->block_groups[index], list) {
5333                 spin_lock(&cache->lock);
5334                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5335                        "%llu pinned %llu reserved\n",
5336                        (unsigned long long)cache->key.objectid,
5337                        (unsigned long long)cache->key.offset,
5338                        (unsigned long long)btrfs_block_group_used(&cache->item),
5339                        (unsigned long long)cache->pinned,
5340                        (unsigned long long)cache->reserved);
5341                 btrfs_dump_free_space(cache, bytes);
5342                 spin_unlock(&cache->lock);
5343         }
5344         if (++index < BTRFS_NR_RAID_TYPES)
5345                 goto again;
5346         up_read(&info->groups_sem);
5347 }
5348
5349 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5350                          struct btrfs_root *root,
5351                          u64 num_bytes, u64 min_alloc_size,
5352                          u64 empty_size, u64 hint_byte,
5353                          u64 search_end, struct btrfs_key *ins,
5354                          u64 data)
5355 {
5356         int ret;
5357         u64 search_start = 0;
5358
5359         data = btrfs_get_alloc_profile(root, data);
5360 again:
5361         /*
5362          * the only place that sets empty_size is btrfs_realloc_node, which
5363          * is not called recursively on allocations
5364          */
5365         if (empty_size || root->ref_cows)
5366                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5367                                      num_bytes + 2 * 1024 * 1024, data, 0);
5368
5369         WARN_ON(num_bytes < root->sectorsize);
5370         ret = find_free_extent(trans, root, num_bytes, empty_size,
5371                                search_start, search_end, hint_byte,
5372                                ins, data);
5373
5374         if (ret == -ENOSPC && num_bytes > min_alloc_size) {
5375                 num_bytes = num_bytes >> 1;
5376                 num_bytes = num_bytes & ~(root->sectorsize - 1);
5377                 num_bytes = max(num_bytes, min_alloc_size);
5378                 do_chunk_alloc(trans, root->fs_info->extent_root,
5379                                num_bytes, data, 1);
5380                 goto again;
5381         }
5382         if (ret == -ENOSPC) {
5383                 struct btrfs_space_info *sinfo;
5384
5385                 sinfo = __find_space_info(root->fs_info, data);
5386                 printk(KERN_ERR "btrfs allocation failed flags %llu, "
5387                        "wanted %llu\n", (unsigned long long)data,
5388                        (unsigned long long)num_bytes);
5389                 dump_space_info(sinfo, num_bytes, 1);
5390         }
5391
5392         return ret;
5393 }
5394
5395 int btrfs_free_reserved_extent(struct btrfs_root *root, u64 start, u64 len)
5396 {
5397         struct btrfs_block_group_cache *cache;
5398         int ret = 0;
5399
5400         cache = btrfs_lookup_block_group(root->fs_info, start);
5401         if (!cache) {
5402                 printk(KERN_ERR "Unable to find block group for %llu\n",
5403                        (unsigned long long)start);
5404                 return -ENOSPC;
5405         }
5406
5407         ret = btrfs_discard_extent(root, start, len);
5408
5409         btrfs_add_free_space(cache, start, len);
5410         update_reserved_bytes(cache, len, 0, 1);
5411         btrfs_put_block_group(cache);
5412
5413         return ret;
5414 }
5415
5416 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5417                                       struct btrfs_root *root,
5418                                       u64 parent, u64 root_objectid,
5419                                       u64 flags, u64 owner, u64 offset,
5420                                       struct btrfs_key *ins, int ref_mod)
5421 {
5422         int ret;
5423         struct btrfs_fs_info *fs_info = root->fs_info;
5424         struct btrfs_extent_item *extent_item;
5425         struct btrfs_extent_inline_ref *iref;
5426         struct btrfs_path *path;
5427         struct extent_buffer *leaf;
5428         int type;
5429         u32 size;
5430
5431         if (parent > 0)
5432                 type = BTRFS_SHARED_DATA_REF_KEY;
5433         else
5434                 type = BTRFS_EXTENT_DATA_REF_KEY;
5435
5436         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5437
5438         path = btrfs_alloc_path();
5439         BUG_ON(!path);
5440
5441         path->leave_spinning = 1;
5442         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5443                                       ins, size);
5444         BUG_ON(ret);
5445
5446         leaf = path->nodes[0];
5447         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5448                                      struct btrfs_extent_item);
5449         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
5450         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5451         btrfs_set_extent_flags(leaf, extent_item,
5452                                flags | BTRFS_EXTENT_FLAG_DATA);
5453
5454         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
5455         btrfs_set_extent_inline_ref_type(leaf, iref, type);
5456         if (parent > 0) {
5457                 struct btrfs_shared_data_ref *ref;
5458                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
5459                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5460                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
5461         } else {
5462                 struct btrfs_extent_data_ref *ref;
5463                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
5464                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
5465                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
5466                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
5467                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
5468         }
5469
5470         btrfs_mark_buffer_dirty(path->nodes[0]);
5471         btrfs_free_path(path);
5472
5473         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5474         if (ret) {
5475                 printk(KERN_ERR "btrfs update block group failed for %llu "
5476                        "%llu\n", (unsigned long long)ins->objectid,
5477                        (unsigned long long)ins->offset);
5478                 BUG();
5479         }
5480         return ret;
5481 }
5482
5483 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
5484                                      struct btrfs_root *root,
5485                                      u64 parent, u64 root_objectid,
5486                                      u64 flags, struct btrfs_disk_key *key,
5487                                      int level, struct btrfs_key *ins)
5488 {
5489         int ret;
5490         struct btrfs_fs_info *fs_info = root->fs_info;
5491         struct btrfs_extent_item *extent_item;
5492         struct btrfs_tree_block_info *block_info;
5493         struct btrfs_extent_inline_ref *iref;
5494         struct btrfs_path *path;
5495         struct extent_buffer *leaf;
5496         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
5497
5498         path = btrfs_alloc_path();
5499         BUG_ON(!path);
5500
5501         path->leave_spinning = 1;
5502         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5503                                       ins, size);
5504         BUG_ON(ret);
5505
5506         leaf = path->nodes[0];
5507         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5508                                      struct btrfs_extent_item);
5509         btrfs_set_extent_refs(leaf, extent_item, 1);
5510         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5511         btrfs_set_extent_flags(leaf, extent_item,
5512                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
5513         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
5514
5515         btrfs_set_tree_block_key(leaf, block_info, key);
5516         btrfs_set_tree_block_level(leaf, block_info, level);
5517
5518         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
5519         if (parent > 0) {
5520                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
5521                 btrfs_set_extent_inline_ref_type(leaf, iref,
5522                                                  BTRFS_SHARED_BLOCK_REF_KEY);
5523                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5524         } else {
5525                 btrfs_set_extent_inline_ref_type(leaf, iref,
5526                                                  BTRFS_TREE_BLOCK_REF_KEY);
5527                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
5528         }
5529
5530         btrfs_mark_buffer_dirty(leaf);
5531         btrfs_free_path(path);
5532
5533         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5534         if (ret) {
5535                 printk(KERN_ERR "btrfs update block group failed for %llu "
5536                        "%llu\n", (unsigned long long)ins->objectid,
5537                        (unsigned long long)ins->offset);
5538                 BUG();
5539         }
5540         return ret;
5541 }
5542
5543 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5544                                      struct btrfs_root *root,
5545                                      u64 root_objectid, u64 owner,
5546                                      u64 offset, struct btrfs_key *ins)
5547 {
5548         int ret;
5549
5550         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
5551
5552         ret = btrfs_add_delayed_data_ref(trans, ins->objectid, ins->offset,
5553                                          0, root_objectid, owner, offset,
5554                                          BTRFS_ADD_DELAYED_EXTENT, NULL);
5555         return ret;
5556 }
5557
5558 /*
5559  * this is used by the tree logging recovery code.  It records that
5560  * an extent has been allocated and makes sure to clear the free
5561  * space cache bits as well
5562  */
5563 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
5564                                    struct btrfs_root *root,
5565                                    u64 root_objectid, u64 owner, u64 offset,
5566                                    struct btrfs_key *ins)
5567 {
5568         int ret;
5569         struct btrfs_block_group_cache *block_group;
5570         struct btrfs_caching_control *caching_ctl;
5571         u64 start = ins->objectid;
5572         u64 num_bytes = ins->offset;
5573
5574         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
5575         cache_block_group(block_group, trans, NULL, 0);
5576         caching_ctl = get_caching_control(block_group);
5577
5578         if (!caching_ctl) {
5579                 BUG_ON(!block_group_cache_done(block_group));
5580                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5581                 BUG_ON(ret);
5582         } else {
5583                 mutex_lock(&caching_ctl->mutex);
5584
5585                 if (start >= caching_ctl->progress) {
5586                         ret = add_excluded_extent(root, start, num_bytes);
5587                         BUG_ON(ret);
5588                 } else if (start + num_bytes <= caching_ctl->progress) {
5589                         ret = btrfs_remove_free_space(block_group,
5590                                                       start, num_bytes);
5591                         BUG_ON(ret);
5592                 } else {
5593                         num_bytes = caching_ctl->progress - start;
5594                         ret = btrfs_remove_free_space(block_group,
5595                                                       start, num_bytes);
5596                         BUG_ON(ret);
5597
5598                         start = caching_ctl->progress;
5599                         num_bytes = ins->objectid + ins->offset -
5600                                     caching_ctl->progress;
5601                         ret = add_excluded_extent(root, start, num_bytes);
5602                         BUG_ON(ret);
5603                 }
5604
5605                 mutex_unlock(&caching_ctl->mutex);
5606                 put_caching_control(caching_ctl);
5607         }
5608
5609         ret = update_reserved_bytes(block_group, ins->offset, 1, 1);
5610         BUG_ON(ret);
5611         btrfs_put_block_group(block_group);
5612         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
5613                                          0, owner, offset, ins, 1);
5614         return ret;
5615 }
5616
5617 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
5618                                             struct btrfs_root *root,
5619                                             u64 bytenr, u32 blocksize,
5620                                             int level)
5621 {
5622         struct extent_buffer *buf;
5623
5624         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
5625         if (!buf)
5626                 return ERR_PTR(-ENOMEM);
5627         btrfs_set_header_generation(buf, trans->transid);
5628         btrfs_set_buffer_lockdep_class(buf, level);
5629         btrfs_tree_lock(buf);
5630         clean_tree_block(trans, root, buf);
5631
5632         btrfs_set_lock_blocking(buf);
5633         btrfs_set_buffer_uptodate(buf);
5634
5635         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
5636                 /*
5637                  * we allow two log transactions at a time, use different
5638                  * EXENT bit to differentiate dirty pages.
5639                  */
5640                 if (root->log_transid % 2 == 0)
5641                         set_extent_dirty(&root->dirty_log_pages, buf->start,
5642                                         buf->start + buf->len - 1, GFP_NOFS);
5643                 else
5644                         set_extent_new(&root->dirty_log_pages, buf->start,
5645                                         buf->start + buf->len - 1, GFP_NOFS);
5646         } else {
5647                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
5648                          buf->start + buf->len - 1, GFP_NOFS);
5649         }
5650         trans->blocks_used++;
5651         /* this returns a buffer locked for blocking */
5652         return buf;
5653 }
5654
5655 static struct btrfs_block_rsv *
5656 use_block_rsv(struct btrfs_trans_handle *trans,
5657               struct btrfs_root *root, u32 blocksize)
5658 {
5659         struct btrfs_block_rsv *block_rsv;
5660         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5661         int ret;
5662
5663         block_rsv = get_block_rsv(trans, root);
5664
5665         if (block_rsv->size == 0) {
5666                 ret = reserve_metadata_bytes(trans, root, block_rsv,
5667                                              blocksize, 0);
5668                 /*
5669                  * If we couldn't reserve metadata bytes try and use some from
5670                  * the global reserve.
5671                  */
5672                 if (ret && block_rsv != global_rsv) {
5673                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5674                         if (!ret)
5675                                 return global_rsv;
5676                         return ERR_PTR(ret);
5677                 } else if (ret) {
5678                         return ERR_PTR(ret);
5679                 }
5680                 return block_rsv;
5681         }
5682
5683         ret = block_rsv_use_bytes(block_rsv, blocksize);
5684         if (!ret)
5685                 return block_rsv;
5686         if (ret) {
5687                 WARN_ON(1);
5688                 ret = reserve_metadata_bytes(trans, root, block_rsv, blocksize,
5689                                              0);
5690                 if (!ret) {
5691                         spin_lock(&block_rsv->lock);
5692                         block_rsv->size += blocksize;
5693                         spin_unlock(&block_rsv->lock);
5694                         return block_rsv;
5695                 } else if (ret && block_rsv != global_rsv) {
5696                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5697                         if (!ret)
5698                                 return global_rsv;
5699                 }
5700         }
5701
5702         return ERR_PTR(-ENOSPC);
5703 }
5704
5705 static void unuse_block_rsv(struct btrfs_block_rsv *block_rsv, u32 blocksize)
5706 {
5707         block_rsv_add_bytes(block_rsv, blocksize, 0);
5708         block_rsv_release_bytes(block_rsv, NULL, 0);
5709 }
5710
5711 /*
5712  * finds a free extent and does all the dirty work required for allocation
5713  * returns the key for the extent through ins, and a tree buffer for
5714  * the first block of the extent through buf.
5715  *
5716  * returns the tree buffer or NULL.
5717  */
5718 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
5719                                         struct btrfs_root *root, u32 blocksize,
5720                                         u64 parent, u64 root_objectid,
5721                                         struct btrfs_disk_key *key, int level,
5722                                         u64 hint, u64 empty_size)
5723 {
5724         struct btrfs_key ins;
5725         struct btrfs_block_rsv *block_rsv;
5726         struct extent_buffer *buf;
5727         u64 flags = 0;
5728         int ret;
5729
5730
5731         block_rsv = use_block_rsv(trans, root, blocksize);
5732         if (IS_ERR(block_rsv))
5733                 return ERR_CAST(block_rsv);
5734
5735         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
5736                                    empty_size, hint, (u64)-1, &ins, 0);
5737         if (ret) {
5738                 unuse_block_rsv(block_rsv, blocksize);
5739                 return ERR_PTR(ret);
5740         }
5741
5742         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
5743                                     blocksize, level);
5744         BUG_ON(IS_ERR(buf));
5745
5746         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
5747                 if (parent == 0)
5748                         parent = ins.objectid;
5749                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
5750         } else
5751                 BUG_ON(parent > 0);
5752
5753         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
5754                 struct btrfs_delayed_extent_op *extent_op;
5755                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
5756                 BUG_ON(!extent_op);
5757                 if (key)
5758                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
5759                 else
5760                         memset(&extent_op->key, 0, sizeof(extent_op->key));
5761                 extent_op->flags_to_set = flags;
5762                 extent_op->update_key = 1;
5763                 extent_op->update_flags = 1;
5764                 extent_op->is_data = 0;
5765
5766                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
5767                                         ins.offset, parent, root_objectid,
5768                                         level, BTRFS_ADD_DELAYED_EXTENT,
5769                                         extent_op);
5770                 BUG_ON(ret);
5771         }
5772         return buf;
5773 }
5774
5775 struct walk_control {
5776         u64 refs[BTRFS_MAX_LEVEL];
5777         u64 flags[BTRFS_MAX_LEVEL];
5778         struct btrfs_key update_progress;
5779         int stage;
5780         int level;
5781         int shared_level;
5782         int update_ref;
5783         int keep_locks;
5784         int reada_slot;
5785         int reada_count;
5786 };
5787
5788 #define DROP_REFERENCE  1
5789 #define UPDATE_BACKREF  2
5790
5791 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
5792                                      struct btrfs_root *root,
5793                                      struct walk_control *wc,
5794                                      struct btrfs_path *path)
5795 {
5796         u64 bytenr;
5797         u64 generation;
5798         u64 refs;
5799         u64 flags;
5800         u32 nritems;
5801         u32 blocksize;
5802         struct btrfs_key key;
5803         struct extent_buffer *eb;
5804         int ret;
5805         int slot;
5806         int nread = 0;
5807
5808         if (path->slots[wc->level] < wc->reada_slot) {
5809                 wc->reada_count = wc->reada_count * 2 / 3;
5810                 wc->reada_count = max(wc->reada_count, 2);
5811         } else {
5812                 wc->reada_count = wc->reada_count * 3 / 2;
5813                 wc->reada_count = min_t(int, wc->reada_count,
5814                                         BTRFS_NODEPTRS_PER_BLOCK(root));
5815         }
5816
5817         eb = path->nodes[wc->level];
5818         nritems = btrfs_header_nritems(eb);
5819         blocksize = btrfs_level_size(root, wc->level - 1);
5820
5821         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
5822                 if (nread >= wc->reada_count)
5823                         break;
5824
5825                 cond_resched();
5826                 bytenr = btrfs_node_blockptr(eb, slot);
5827                 generation = btrfs_node_ptr_generation(eb, slot);
5828
5829                 if (slot == path->slots[wc->level])
5830                         goto reada;
5831
5832                 if (wc->stage == UPDATE_BACKREF &&
5833                     generation <= root->root_key.offset)
5834                         continue;
5835
5836                 /* We don't lock the tree block, it's OK to be racy here */
5837                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
5838                                                &refs, &flags);
5839                 BUG_ON(ret);
5840                 BUG_ON(refs == 0);
5841
5842                 if (wc->stage == DROP_REFERENCE) {
5843                         if (refs == 1)
5844                                 goto reada;
5845
5846                         if (wc->level == 1 &&
5847                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5848                                 continue;
5849                         if (!wc->update_ref ||
5850                             generation <= root->root_key.offset)
5851                                 continue;
5852                         btrfs_node_key_to_cpu(eb, &key, slot);
5853                         ret = btrfs_comp_cpu_keys(&key,
5854                                                   &wc->update_progress);
5855                         if (ret < 0)
5856                                 continue;
5857                 } else {
5858                         if (wc->level == 1 &&
5859                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5860                                 continue;
5861                 }
5862 reada:
5863                 ret = readahead_tree_block(root, bytenr, blocksize,
5864                                            generation);
5865                 if (ret)
5866                         break;
5867                 nread++;
5868         }
5869         wc->reada_slot = slot;
5870 }
5871
5872 /*
5873  * hepler to process tree block while walking down the tree.
5874  *
5875  * when wc->stage == UPDATE_BACKREF, this function updates
5876  * back refs for pointers in the block.
5877  *
5878  * NOTE: return value 1 means we should stop walking down.
5879  */
5880 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
5881                                    struct btrfs_root *root,
5882                                    struct btrfs_path *path,
5883                                    struct walk_control *wc, int lookup_info)
5884 {
5885         int level = wc->level;
5886         struct extent_buffer *eb = path->nodes[level];
5887         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5888         int ret;
5889
5890         if (wc->stage == UPDATE_BACKREF &&
5891             btrfs_header_owner(eb) != root->root_key.objectid)
5892                 return 1;
5893
5894         /*
5895          * when reference count of tree block is 1, it won't increase
5896          * again. once full backref flag is set, we never clear it.
5897          */
5898         if (lookup_info &&
5899             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
5900              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
5901                 BUG_ON(!path->locks[level]);
5902                 ret = btrfs_lookup_extent_info(trans, root,
5903                                                eb->start, eb->len,
5904                                                &wc->refs[level],
5905                                                &wc->flags[level]);
5906                 BUG_ON(ret);
5907                 BUG_ON(wc->refs[level] == 0);
5908         }
5909
5910         if (wc->stage == DROP_REFERENCE) {
5911                 if (wc->refs[level] > 1)
5912                         return 1;
5913
5914                 if (path->locks[level] && !wc->keep_locks) {
5915                         btrfs_tree_unlock(eb);
5916                         path->locks[level] = 0;
5917                 }
5918                 return 0;
5919         }
5920
5921         /* wc->stage == UPDATE_BACKREF */
5922         if (!(wc->flags[level] & flag)) {
5923                 BUG_ON(!path->locks[level]);
5924                 ret = btrfs_inc_ref(trans, root, eb, 1);
5925                 BUG_ON(ret);
5926                 ret = btrfs_dec_ref(trans, root, eb, 0);
5927                 BUG_ON(ret);
5928                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
5929                                                   eb->len, flag, 0);
5930                 BUG_ON(ret);
5931                 wc->flags[level] |= flag;
5932         }
5933
5934         /*
5935          * the block is shared by multiple trees, so it's not good to
5936          * keep the tree lock
5937          */
5938         if (path->locks[level] && level > 0) {
5939                 btrfs_tree_unlock(eb);
5940                 path->locks[level] = 0;
5941         }
5942         return 0;
5943 }
5944
5945 /*
5946  * hepler to process tree block pointer.
5947  *
5948  * when wc->stage == DROP_REFERENCE, this function checks
5949  * reference count of the block pointed to. if the block
5950  * is shared and we need update back refs for the subtree
5951  * rooted at the block, this function changes wc->stage to
5952  * UPDATE_BACKREF. if the block is shared and there is no
5953  * need to update back, this function drops the reference
5954  * to the block.
5955  *
5956  * NOTE: return value 1 means we should stop walking down.
5957  */
5958 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
5959                                  struct btrfs_root *root,
5960                                  struct btrfs_path *path,
5961                                  struct walk_control *wc, int *lookup_info)
5962 {
5963         u64 bytenr;
5964         u64 generation;
5965         u64 parent;
5966         u32 blocksize;
5967         struct btrfs_key key;
5968         struct extent_buffer *next;
5969         int level = wc->level;
5970         int reada = 0;
5971         int ret = 0;
5972
5973         generation = btrfs_node_ptr_generation(path->nodes[level],
5974                                                path->slots[level]);
5975         /*
5976          * if the lower level block was created before the snapshot
5977          * was created, we know there is no need to update back refs
5978          * for the subtree
5979          */
5980         if (wc->stage == UPDATE_BACKREF &&
5981             generation <= root->root_key.offset) {
5982                 *lookup_info = 1;
5983                 return 1;
5984         }
5985
5986         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
5987         blocksize = btrfs_level_size(root, level - 1);
5988
5989         next = btrfs_find_tree_block(root, bytenr, blocksize);
5990         if (!next) {
5991                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
5992                 if (!next)
5993                         return -ENOMEM;
5994                 reada = 1;
5995         }
5996         btrfs_tree_lock(next);
5997         btrfs_set_lock_blocking(next);
5998
5999         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6000                                        &wc->refs[level - 1],
6001                                        &wc->flags[level - 1]);
6002         BUG_ON(ret);
6003         BUG_ON(wc->refs[level - 1] == 0);
6004         *lookup_info = 0;
6005
6006         if (wc->stage == DROP_REFERENCE) {
6007                 if (wc->refs[level - 1] > 1) {
6008                         if (level == 1 &&
6009                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6010                                 goto skip;
6011
6012                         if (!wc->update_ref ||
6013                             generation <= root->root_key.offset)
6014                                 goto skip;
6015
6016                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6017                                               path->slots[level]);
6018                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6019                         if (ret < 0)
6020                                 goto skip;
6021
6022                         wc->stage = UPDATE_BACKREF;
6023                         wc->shared_level = level - 1;
6024                 }
6025         } else {
6026                 if (level == 1 &&
6027                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6028                         goto skip;
6029         }
6030
6031         if (!btrfs_buffer_uptodate(next, generation)) {
6032                 btrfs_tree_unlock(next);
6033                 free_extent_buffer(next);
6034                 next = NULL;
6035                 *lookup_info = 1;
6036         }
6037
6038         if (!next) {
6039                 if (reada && level == 1)
6040                         reada_walk_down(trans, root, wc, path);
6041                 next = read_tree_block(root, bytenr, blocksize, generation);
6042                 btrfs_tree_lock(next);
6043                 btrfs_set_lock_blocking(next);
6044         }
6045
6046         level--;
6047         BUG_ON(level != btrfs_header_level(next));
6048         path->nodes[level] = next;
6049         path->slots[level] = 0;
6050         path->locks[level] = 1;
6051         wc->level = level;
6052         if (wc->level == 1)
6053                 wc->reada_slot = 0;
6054         return 0;
6055 skip:
6056         wc->refs[level - 1] = 0;
6057         wc->flags[level - 1] = 0;
6058         if (wc->stage == DROP_REFERENCE) {
6059                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6060                         parent = path->nodes[level]->start;
6061                 } else {
6062                         BUG_ON(root->root_key.objectid !=
6063                                btrfs_header_owner(path->nodes[level]));
6064                         parent = 0;
6065                 }
6066
6067                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6068                                         root->root_key.objectid, level - 1, 0);
6069                 BUG_ON(ret);
6070         }
6071         btrfs_tree_unlock(next);
6072         free_extent_buffer(next);
6073         *lookup_info = 1;
6074         return 1;
6075 }
6076
6077 /*
6078  * hepler to process tree block while walking up the tree.
6079  *
6080  * when wc->stage == DROP_REFERENCE, this function drops
6081  * reference count on the block.
6082  *
6083  * when wc->stage == UPDATE_BACKREF, this function changes
6084  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6085  * to UPDATE_BACKREF previously while processing the block.
6086  *
6087  * NOTE: return value 1 means we should stop walking up.
6088  */
6089 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6090                                  struct btrfs_root *root,
6091                                  struct btrfs_path *path,
6092                                  struct walk_control *wc)
6093 {
6094         int ret;
6095         int level = wc->level;
6096         struct extent_buffer *eb = path->nodes[level];
6097         u64 parent = 0;
6098
6099         if (wc->stage == UPDATE_BACKREF) {
6100                 BUG_ON(wc->shared_level < level);
6101                 if (level < wc->shared_level)
6102                         goto out;
6103
6104                 ret = find_next_key(path, level + 1, &wc->update_progress);
6105                 if (ret > 0)
6106                         wc->update_ref = 0;
6107
6108                 wc->stage = DROP_REFERENCE;
6109                 wc->shared_level = -1;
6110                 path->slots[level] = 0;
6111
6112                 /*
6113                  * check reference count again if the block isn't locked.
6114                  * we should start walking down the tree again if reference
6115                  * count is one.
6116                  */
6117                 if (!path->locks[level]) {
6118                         BUG_ON(level == 0);
6119                         btrfs_tree_lock(eb);
6120                         btrfs_set_lock_blocking(eb);
6121                         path->locks[level] = 1;
6122
6123                         ret = btrfs_lookup_extent_info(trans, root,
6124                                                        eb->start, eb->len,
6125                                                        &wc->refs[level],
6126                                                        &wc->flags[level]);
6127                         BUG_ON(ret);
6128                         BUG_ON(wc->refs[level] == 0);
6129                         if (wc->refs[level] == 1) {
6130                                 btrfs_tree_unlock(eb);
6131                                 path->locks[level] = 0;
6132                                 return 1;
6133                         }
6134                 }
6135         }
6136
6137         /* wc->stage == DROP_REFERENCE */
6138         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6139
6140         if (wc->refs[level] == 1) {
6141                 if (level == 0) {
6142                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6143                                 ret = btrfs_dec_ref(trans, root, eb, 1);
6144                         else
6145                                 ret = btrfs_dec_ref(trans, root, eb, 0);
6146                         BUG_ON(ret);
6147                 }
6148                 /* make block locked assertion in clean_tree_block happy */
6149                 if (!path->locks[level] &&
6150                     btrfs_header_generation(eb) == trans->transid) {
6151                         btrfs_tree_lock(eb);
6152                         btrfs_set_lock_blocking(eb);
6153                         path->locks[level] = 1;
6154                 }
6155                 clean_tree_block(trans, root, eb);
6156         }
6157
6158         if (eb == root->node) {
6159                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6160                         parent = eb->start;
6161                 else
6162                         BUG_ON(root->root_key.objectid !=
6163                                btrfs_header_owner(eb));
6164         } else {
6165                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6166                         parent = path->nodes[level + 1]->start;
6167                 else
6168                         BUG_ON(root->root_key.objectid !=
6169                                btrfs_header_owner(path->nodes[level + 1]));
6170         }
6171
6172         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6173 out:
6174         wc->refs[level] = 0;
6175         wc->flags[level] = 0;
6176         return 0;
6177 }
6178
6179 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6180                                    struct btrfs_root *root,
6181                                    struct btrfs_path *path,
6182                                    struct walk_control *wc)
6183 {
6184         int level = wc->level;
6185         int lookup_info = 1;
6186         int ret;
6187
6188         while (level >= 0) {
6189                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6190                 if (ret > 0)
6191                         break;
6192
6193                 if (level == 0)
6194                         break;
6195
6196                 if (path->slots[level] >=
6197                     btrfs_header_nritems(path->nodes[level]))
6198                         break;
6199
6200                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6201                 if (ret > 0) {
6202                         path->slots[level]++;
6203                         continue;
6204                 } else if (ret < 0)
6205                         return ret;
6206                 level = wc->level;
6207         }
6208         return 0;
6209 }
6210
6211 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6212                                  struct btrfs_root *root,
6213                                  struct btrfs_path *path,
6214                                  struct walk_control *wc, int max_level)
6215 {
6216         int level = wc->level;
6217         int ret;
6218
6219         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6220         while (level < max_level && path->nodes[level]) {
6221                 wc->level = level;
6222                 if (path->slots[level] + 1 <
6223                     btrfs_header_nritems(path->nodes[level])) {
6224                         path->slots[level]++;
6225                         return 0;
6226                 } else {
6227                         ret = walk_up_proc(trans, root, path, wc);
6228                         if (ret > 0)
6229                                 return 0;
6230
6231                         if (path->locks[level]) {
6232                                 btrfs_tree_unlock(path->nodes[level]);
6233                                 path->locks[level] = 0;
6234                         }
6235                         free_extent_buffer(path->nodes[level]);
6236                         path->nodes[level] = NULL;
6237                         level++;
6238                 }
6239         }
6240         return 1;
6241 }
6242
6243 /*
6244  * drop a subvolume tree.
6245  *
6246  * this function traverses the tree freeing any blocks that only
6247  * referenced by the tree.
6248  *
6249  * when a shared tree block is found. this function decreases its
6250  * reference count by one. if update_ref is true, this function
6251  * also make sure backrefs for the shared block and all lower level
6252  * blocks are properly updated.
6253  */
6254 int btrfs_drop_snapshot(struct btrfs_root *root,
6255                         struct btrfs_block_rsv *block_rsv, int update_ref)
6256 {
6257         struct btrfs_path *path;
6258         struct btrfs_trans_handle *trans;
6259         struct btrfs_root *tree_root = root->fs_info->tree_root;
6260         struct btrfs_root_item *root_item = &root->root_item;
6261         struct walk_control *wc;
6262         struct btrfs_key key;
6263         int err = 0;
6264         int ret;
6265         int level;
6266
6267         path = btrfs_alloc_path();
6268         BUG_ON(!path);
6269
6270         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6271         BUG_ON(!wc);
6272
6273         trans = btrfs_start_transaction(tree_root, 0);
6274         if (block_rsv)
6275                 trans->block_rsv = block_rsv;
6276
6277         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6278                 level = btrfs_header_level(root->node);
6279                 path->nodes[level] = btrfs_lock_root_node(root);
6280                 btrfs_set_lock_blocking(path->nodes[level]);
6281                 path->slots[level] = 0;
6282                 path->locks[level] = 1;
6283                 memset(&wc->update_progress, 0,
6284                        sizeof(wc->update_progress));
6285         } else {
6286                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6287                 memcpy(&wc->update_progress, &key,
6288                        sizeof(wc->update_progress));
6289
6290                 level = root_item->drop_level;
6291                 BUG_ON(level == 0);
6292                 path->lowest_level = level;
6293                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6294                 path->lowest_level = 0;
6295                 if (ret < 0) {
6296                         err = ret;
6297                         goto out;
6298                 }
6299                 WARN_ON(ret > 0);
6300
6301                 /*
6302                  * unlock our path, this is safe because only this
6303                  * function is allowed to delete this snapshot
6304                  */
6305                 btrfs_unlock_up_safe(path, 0);
6306
6307                 level = btrfs_header_level(root->node);
6308                 while (1) {
6309                         btrfs_tree_lock(path->nodes[level]);
6310                         btrfs_set_lock_blocking(path->nodes[level]);
6311
6312                         ret = btrfs_lookup_extent_info(trans, root,
6313                                                 path->nodes[level]->start,
6314                                                 path->nodes[level]->len,
6315                                                 &wc->refs[level],
6316                                                 &wc->flags[level]);
6317                         BUG_ON(ret);
6318                         BUG_ON(wc->refs[level] == 0);
6319
6320                         if (level == root_item->drop_level)
6321                                 break;
6322
6323                         btrfs_tree_unlock(path->nodes[level]);
6324                         WARN_ON(wc->refs[level] != 1);
6325                         level--;
6326                 }
6327         }
6328
6329         wc->level = level;
6330         wc->shared_level = -1;
6331         wc->stage = DROP_REFERENCE;
6332         wc->update_ref = update_ref;
6333         wc->keep_locks = 0;
6334         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6335
6336         while (1) {
6337                 ret = walk_down_tree(trans, root, path, wc);
6338                 if (ret < 0) {
6339                         err = ret;
6340                         break;
6341                 }
6342
6343                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
6344                 if (ret < 0) {
6345                         err = ret;
6346                         break;
6347                 }
6348
6349                 if (ret > 0) {
6350                         BUG_ON(wc->stage != DROP_REFERENCE);
6351                         break;
6352                 }
6353
6354                 if (wc->stage == DROP_REFERENCE) {
6355                         level = wc->level;
6356                         btrfs_node_key(path->nodes[level],
6357                                        &root_item->drop_progress,
6358                                        path->slots[level]);
6359                         root_item->drop_level = level;
6360                 }
6361
6362                 BUG_ON(wc->level == 0);
6363                 if (btrfs_should_end_transaction(trans, tree_root)) {
6364                         ret = btrfs_update_root(trans, tree_root,
6365                                                 &root->root_key,
6366                                                 root_item);
6367                         BUG_ON(ret);
6368
6369                         btrfs_end_transaction_throttle(trans, tree_root);
6370                         trans = btrfs_start_transaction(tree_root, 0);
6371                         if (block_rsv)
6372                                 trans->block_rsv = block_rsv;
6373                 }
6374         }
6375         btrfs_release_path(root, path);
6376         BUG_ON(err);
6377
6378         ret = btrfs_del_root(trans, tree_root, &root->root_key);
6379         BUG_ON(ret);
6380
6381         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
6382                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
6383                                            NULL, NULL);
6384                 BUG_ON(ret < 0);
6385                 if (ret > 0) {
6386                         /* if we fail to delete the orphan item this time
6387                          * around, it'll get picked up the next time.
6388                          *
6389                          * The most common failure here is just -ENOENT.
6390                          */
6391                         btrfs_del_orphan_item(trans, tree_root,
6392                                               root->root_key.objectid);
6393                 }
6394         }
6395
6396         if (root->in_radix) {
6397                 btrfs_free_fs_root(tree_root->fs_info, root);
6398         } else {
6399                 free_extent_buffer(root->node);
6400                 free_extent_buffer(root->commit_root);
6401                 kfree(root);
6402         }
6403 out:
6404         btrfs_end_transaction_throttle(trans, tree_root);
6405         kfree(wc);
6406         btrfs_free_path(path);
6407         return err;
6408 }
6409
6410 /*
6411  * drop subtree rooted at tree block 'node'.
6412  *
6413  * NOTE: this function will unlock and release tree block 'node'
6414  */
6415 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
6416                         struct btrfs_root *root,
6417                         struct extent_buffer *node,
6418                         struct extent_buffer *parent)
6419 {
6420         struct btrfs_path *path;
6421         struct walk_control *wc;
6422         int level;
6423         int parent_level;
6424         int ret = 0;
6425         int wret;
6426
6427         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
6428
6429         path = btrfs_alloc_path();
6430         BUG_ON(!path);
6431
6432         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6433         BUG_ON(!wc);
6434
6435         btrfs_assert_tree_locked(parent);
6436         parent_level = btrfs_header_level(parent);
6437         extent_buffer_get(parent);
6438         path->nodes[parent_level] = parent;
6439         path->slots[parent_level] = btrfs_header_nritems(parent);
6440
6441         btrfs_assert_tree_locked(node);
6442         level = btrfs_header_level(node);
6443         path->nodes[level] = node;
6444         path->slots[level] = 0;
6445         path->locks[level] = 1;
6446
6447         wc->refs[parent_level] = 1;
6448         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6449         wc->level = level;
6450         wc->shared_level = -1;
6451         wc->stage = DROP_REFERENCE;
6452         wc->update_ref = 0;
6453         wc->keep_locks = 1;
6454         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6455
6456         while (1) {
6457                 wret = walk_down_tree(trans, root, path, wc);
6458                 if (wret < 0) {
6459                         ret = wret;
6460                         break;
6461                 }
6462
6463                 wret = walk_up_tree(trans, root, path, wc, parent_level);
6464                 if (wret < 0)
6465                         ret = wret;
6466                 if (wret != 0)
6467                         break;
6468         }
6469
6470         kfree(wc);
6471         btrfs_free_path(path);
6472         return ret;
6473 }
6474
6475 #if 0
6476 static unsigned long calc_ra(unsigned long start, unsigned long last,
6477                              unsigned long nr)
6478 {
6479         return min(last, start + nr - 1);
6480 }
6481
6482 static noinline int relocate_inode_pages(struct inode *inode, u64 start,
6483                                          u64 len)
6484 {
6485         u64 page_start;
6486         u64 page_end;
6487         unsigned long first_index;
6488         unsigned long last_index;
6489         unsigned long i;
6490         struct page *page;
6491         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6492         struct file_ra_state *ra;
6493         struct btrfs_ordered_extent *ordered;
6494         unsigned int total_read = 0;
6495         unsigned int total_dirty = 0;
6496         int ret = 0;
6497
6498         ra = kzalloc(sizeof(*ra), GFP_NOFS);
6499
6500         mutex_lock(&inode->i_mutex);
6501         first_index = start >> PAGE_CACHE_SHIFT;
6502         last_index = (start + len - 1) >> PAGE_CACHE_SHIFT;
6503
6504         /* make sure the dirty trick played by the caller work */
6505         ret = invalidate_inode_pages2_range(inode->i_mapping,
6506                                             first_index, last_index);
6507         if (ret)
6508                 goto out_unlock;
6509
6510         file_ra_state_init(ra, inode->i_mapping);
6511
6512         for (i = first_index ; i <= last_index; i++) {
6513                 if (total_read % ra->ra_pages == 0) {
6514                         btrfs_force_ra(inode->i_mapping, ra, NULL, i,
6515                                        calc_ra(i, last_index, ra->ra_pages));
6516                 }
6517                 total_read++;
6518 again:
6519                 if (((u64)i << PAGE_CACHE_SHIFT) > i_size_read(inode))
6520                         BUG_ON(1);
6521                 page = grab_cache_page(inode->i_mapping, i);
6522                 if (!page) {
6523                         ret = -ENOMEM;
6524                         goto out_unlock;
6525                 }
6526                 if (!PageUptodate(page)) {
6527                         btrfs_readpage(NULL, page);
6528                         lock_page(page);
6529                         if (!PageUptodate(page)) {
6530                                 unlock_page(page);
6531                                 page_cache_release(page);
6532                                 ret = -EIO;
6533                                 goto out_unlock;
6534                         }
6535                 }
6536                 wait_on_page_writeback(page);
6537
6538                 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
6539                 page_end = page_start + PAGE_CACHE_SIZE - 1;
6540                 lock_extent(io_tree, page_start, page_end, GFP_NOFS);
6541
6542                 ordered = btrfs_lookup_ordered_extent(inode, page_start);
6543                 if (ordered) {
6544                         unlock_extent(io_tree, page_start, page_end, GFP_NOFS);
6545                         unlock_page(page);
6546                         page_cache_release(page);
6547                         btrfs_start_ordered_extent(inode, ordered, 1);
6548                         btrfs_put_ordered_extent(ordered);
6549                         goto again;
6550                 }
6551                 set_page_extent_mapped(page);
6552
6553                 if (i == first_index)
6554                         set_extent_bits(io_tree, page_start, page_end,
6555                                         EXTENT_BOUNDARY, GFP_NOFS);
6556                 btrfs_set_extent_delalloc(inode, page_start, page_end);
6557
6558                 set_page_dirty(page);
6559                 total_dirty++;
6560
6561                 unlock_extent(io_tree, page_start, page_end, GFP_NOFS);
6562                 unlock_page(page);
6563                 page_cache_release(page);
6564         }
6565
6566 out_unlock:
6567         kfree(ra);
6568         mutex_unlock(&inode->i_mutex);
6569         balance_dirty_pages_ratelimited_nr(inode->i_mapping, total_dirty);
6570         return ret;
6571 }
6572
6573 static noinline int relocate_data_extent(struct inode *reloc_inode,
6574                                          struct btrfs_key *extent_key,
6575                                          u64 offset)
6576 {
6577         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
6578         struct extent_map_tree *em_tree = &BTRFS_I(reloc_inode)->extent_tree;
6579         struct extent_map *em;
6580         u64 start = extent_key->objectid - offset;
6581         u64 end = start + extent_key->offset - 1;
6582
6583         em = alloc_extent_map(GFP_NOFS);
6584         BUG_ON(!em || IS_ERR(em));
6585
6586         em->start = start;
6587         em->len = extent_key->offset;
6588         em->block_len = extent_key->offset;
6589         em->block_start = extent_key->objectid;
6590         em->bdev = root->fs_info->fs_devices->latest_bdev;
6591         set_bit(EXTENT_FLAG_PINNED, &em->flags);
6592
6593         /* setup extent map to cheat btrfs_readpage */
6594         lock_extent(&BTRFS_I(reloc_inode)->io_tree, start, end, GFP_NOFS);
6595         while (1) {
6596                 int ret;
6597                 write_lock(&em_tree->lock);
6598                 ret = add_extent_mapping(em_tree, em);
6599                 write_unlock(&em_tree->lock);
6600                 if (ret != -EEXIST) {
6601                         free_extent_map(em);
6602                         break;
6603                 }
6604                 btrfs_drop_extent_cache(reloc_inode, start, end, 0);
6605         }
6606         unlock_extent(&BTRFS_I(reloc_inode)->io_tree, start, end, GFP_NOFS);
6607
6608         return relocate_inode_pages(reloc_inode, start, extent_key->offset);
6609 }
6610
6611 struct btrfs_ref_path {
6612         u64 extent_start;
6613         u64 nodes[BTRFS_MAX_LEVEL];
6614         u64 root_objectid;
6615         u64 root_generation;
6616         u64 owner_objectid;
6617         u32 num_refs;
6618         int lowest_level;
6619         int current_level;
6620         int shared_level;
6621
6622         struct btrfs_key node_keys[BTRFS_MAX_LEVEL];
6623         u64 new_nodes[BTRFS_MAX_LEVEL];
6624 };
6625
6626 struct disk_extent {
6627         u64 ram_bytes;
6628         u64 disk_bytenr;
6629         u64 disk_num_bytes;
6630         u64 offset;
6631         u64 num_bytes;
6632         u8 compression;
6633         u8 encryption;
6634         u16 other_encoding;
6635 };
6636
6637 static int is_cowonly_root(u64 root_objectid)
6638 {
6639         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
6640             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
6641             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
6642             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
6643             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
6644             root_objectid == BTRFS_CSUM_TREE_OBJECTID)
6645                 return 1;
6646         return 0;
6647 }
6648
6649 static noinline int __next_ref_path(struct btrfs_trans_handle *trans,
6650                                     struct btrfs_root *extent_root,
6651                                     struct btrfs_ref_path *ref_path,
6652                                     int first_time)
6653 {
6654         struct extent_buffer *leaf;
6655         struct btrfs_path *path;
6656         struct btrfs_extent_ref *ref;
6657         struct btrfs_key key;
6658         struct btrfs_key found_key;
6659         u64 bytenr;
6660         u32 nritems;
6661         int level;
6662         int ret = 1;
6663
6664         path = btrfs_alloc_path();
6665         if (!path)
6666                 return -ENOMEM;
6667
6668         if (first_time) {
6669                 ref_path->lowest_level = -1;
6670                 ref_path->current_level = -1;
6671                 ref_path->shared_level = -1;
6672                 goto walk_up;
6673         }
6674 walk_down:
6675         level = ref_path->current_level - 1;
6676         while (level >= -1) {
6677                 u64 parent;
6678                 if (level < ref_path->lowest_level)
6679                         break;
6680
6681                 if (level >= 0)
6682                         bytenr = ref_path->nodes[level];
6683                 else
6684                         bytenr = ref_path->extent_start;
6685                 BUG_ON(bytenr == 0);
6686
6687                 parent = ref_path->nodes[level + 1];
6688                 ref_path->nodes[level + 1] = 0;
6689                 ref_path->current_level = level;
6690                 BUG_ON(parent == 0);
6691
6692                 key.objectid = bytenr;
6693                 key.offset = parent + 1;
6694                 key.type = BTRFS_EXTENT_REF_KEY;
6695
6696                 ret = btrfs_search_slot(trans, extent_root, &key, path, 0, 0);
6697                 if (ret < 0)
6698                         goto out;
6699                 BUG_ON(ret == 0);
6700
6701                 leaf = path->nodes[0];
6702                 nritems = btrfs_header_nritems(leaf);
6703                 if (path->slots[0] >= nritems) {
6704                         ret = btrfs_next_leaf(extent_root, path);
6705                         if (ret < 0)
6706                                 goto out;
6707                         if (ret > 0)
6708                                 goto next;
6709                         leaf = path->nodes[0];
6710                 }
6711
6712                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6713                 if (found_key.objectid == bytenr &&
6714                     found_key.type == BTRFS_EXTENT_REF_KEY) {
6715                         if (level < ref_path->shared_level)
6716                                 ref_path->shared_level = level;
6717                         goto found;
6718                 }
6719 next:
6720                 level--;
6721                 btrfs_release_path(extent_root, path);
6722                 cond_resched();
6723         }
6724         /* reached lowest level */
6725         ret = 1;
6726         goto out;
6727 walk_up:
6728         level = ref_path->current_level;
6729         while (level < BTRFS_MAX_LEVEL - 1) {
6730                 u64 ref_objectid;
6731
6732                 if (level >= 0)
6733                         bytenr = ref_path->nodes[level];
6734                 else
6735                         bytenr = ref_path->extent_start;
6736
6737                 BUG_ON(bytenr == 0);
6738
6739                 key.objectid = bytenr;
6740                 key.offset = 0;
6741                 key.type = BTRFS_EXTENT_REF_KEY;
6742
6743                 ret = btrfs_search_slot(trans, extent_root, &key, path, 0, 0);
6744                 if (ret < 0)
6745                         goto out;
6746
6747                 leaf = path->nodes[0];
6748                 nritems = btrfs_header_nritems(leaf);
6749                 if (path->slots[0] >= nritems) {
6750                         ret = btrfs_next_leaf(extent_root, path);
6751                         if (ret < 0)
6752                                 goto out;
6753                         if (ret > 0) {
6754                                 /* the extent was freed by someone */
6755                                 if (ref_path->lowest_level == level)
6756                                         goto out;
6757                                 btrfs_release_path(extent_root, path);
6758                                 goto walk_down;
6759                         }
6760                         leaf = path->nodes[0];
6761                 }
6762
6763                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6764                 if (found_key.objectid != bytenr ||
6765                                 found_key.type != BTRFS_EXTENT_REF_KEY) {
6766                         /* the extent was freed by someone */
6767                         if (ref_path->lowest_level == level) {
6768                                 ret = 1;
6769                                 goto out;
6770                         }
6771                         btrfs_release_path(extent_root, path);
6772                         goto walk_down;
6773                 }
6774 found:
6775                 ref = btrfs_item_ptr(leaf, path->slots[0],
6776                                 struct btrfs_extent_ref);
6777                 ref_objectid = btrfs_ref_objectid(leaf, ref);
6778                 if (ref_objectid < BTRFS_FIRST_FREE_OBJECTID) {
6779                         if (first_time) {
6780                                 level = (int)ref_objectid;
6781                                 BUG_ON(level >= BTRFS_MAX_LEVEL);
6782                                 ref_path->lowest_level = level;
6783                                 ref_path->current_level = level;
6784                                 ref_path->nodes[level] = bytenr;
6785                         } else {
6786                                 WARN_ON(ref_objectid != level);
6787                         }
6788                 } else {
6789                         WARN_ON(level != -1);
6790                 }
6791                 first_time = 0;
6792
6793                 if (ref_path->lowest_level == level) {
6794                         ref_path->owner_objectid = ref_objectid;
6795                         ref_path->num_refs = btrfs_ref_num_refs(leaf, ref);
6796                 }
6797
6798                 /*
6799                  * the block is tree root or the block isn't in reference
6800                  * counted tree.
6801                  */
6802                 if (found_key.objectid == found_key.offset ||
6803                     is_cowonly_root(btrfs_ref_root(leaf, ref))) {
6804                         ref_path->root_objectid = btrfs_ref_root(leaf, ref);
6805                         ref_path->root_generation =
6806                                 btrfs_ref_generation(leaf, ref);
6807                         if (level < 0) {
6808                                 /* special reference from the tree log */
6809                                 ref_path->nodes[0] = found_key.offset;
6810                                 ref_path->current_level = 0;
6811                         }
6812                         ret = 0;
6813                         goto out;
6814                 }
6815
6816                 level++;
6817                 BUG_ON(ref_path->nodes[level] != 0);
6818                 ref_path->nodes[level] = found_key.offset;
6819                 ref_path->current_level = level;
6820
6821                 /*
6822                  * the reference was created in the running transaction,
6823                  * no need to continue walking up.
6824                  */
6825                 if (btrfs_ref_generation(leaf, ref) == trans->transid) {
6826                         ref_path->root_objectid = btrfs_ref_root(leaf, ref);
6827                         ref_path->root_generation =
6828                                 btrfs_ref_generation(leaf, ref);
6829                         ret = 0;
6830                         goto out;
6831                 }
6832
6833                 btrfs_release_path(extent_root, path);
6834                 cond_resched();
6835         }
6836         /* reached max tree level, but no tree root found. */
6837         BUG();
6838 out:
6839         btrfs_free_path(path);
6840         return ret;
6841 }
6842
6843 static int btrfs_first_ref_path(struct btrfs_trans_handle *trans,
6844                                 struct btrfs_root *extent_root,
6845                                 struct btrfs_ref_path *ref_path,
6846                                 u64 extent_start)
6847 {
6848         memset(ref_path, 0, sizeof(*ref_path));
6849         ref_path->extent_start = extent_start;
6850
6851         return __next_ref_path(trans, extent_root, ref_path, 1);
6852 }
6853
6854 static int btrfs_next_ref_path(struct btrfs_trans_handle *trans,
6855                                struct btrfs_root *extent_root,
6856                                struct btrfs_ref_path *ref_path)
6857 {
6858         return __next_ref_path(trans, extent_root, ref_path, 0);
6859 }
6860
6861 static noinline int get_new_locations(struct inode *reloc_inode,
6862                                       struct btrfs_key *extent_key,
6863                                       u64 offset, int no_fragment,
6864                                       struct disk_extent **extents,
6865                                       int *nr_extents)
6866 {
6867         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
6868         struct btrfs_path *path;
6869         struct btrfs_file_extent_item *fi;
6870         struct extent_buffer *leaf;
6871         struct disk_extent *exts = *extents;
6872         struct btrfs_key found_key;
6873         u64 cur_pos;
6874         u64 last_byte;
6875         u32 nritems;
6876         int nr = 0;
6877         int max = *nr_extents;
6878         int ret;
6879
6880         WARN_ON(!no_fragment && *extents);
6881         if (!exts) {
6882                 max = 1;
6883                 exts = kmalloc(sizeof(*exts) * max, GFP_NOFS);
6884                 if (!exts)
6885                         return -ENOMEM;
6886         }
6887
6888         path = btrfs_alloc_path();
6889         BUG_ON(!path);
6890
6891         cur_pos = extent_key->objectid - offset;
6892         last_byte = extent_key->objectid + extent_key->offset;
6893         ret = btrfs_lookup_file_extent(NULL, root, path, reloc_inode->i_ino,
6894                                        cur_pos, 0);
6895         if (ret < 0)
6896                 goto out;
6897         if (ret > 0) {
6898                 ret = -ENOENT;
6899                 goto out;
6900         }
6901
6902         while (1) {
6903                 leaf = path->nodes[0];
6904                 nritems = btrfs_header_nritems(leaf);
6905                 if (path->slots[0] >= nritems) {
6906                         ret = btrfs_next_leaf(root, path);
6907                         if (ret < 0)
6908                                 goto out;
6909                         if (ret > 0)
6910                                 break;
6911                         leaf = path->nodes[0];
6912                 }
6913
6914                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6915                 if (found_key.offset != cur_pos ||
6916                     found_key.type != BTRFS_EXTENT_DATA_KEY ||
6917                     found_key.objectid != reloc_inode->i_ino)
6918                         break;
6919
6920                 fi = btrfs_item_ptr(leaf, path->slots[0],
6921                                     struct btrfs_file_extent_item);
6922                 if (btrfs_file_extent_type(leaf, fi) !=
6923                     BTRFS_FILE_EXTENT_REG ||
6924                     btrfs_file_extent_disk_bytenr(leaf, fi) == 0)
6925                         break;
6926
6927                 if (nr == max) {
6928                         struct disk_extent *old = exts;
6929                         max *= 2;
6930                         exts = kzalloc(sizeof(*exts) * max, GFP_NOFS);
6931                         memcpy(exts, old, sizeof(*exts) * nr);
6932                         if (old != *extents)
6933                                 kfree(old);
6934                 }
6935
6936                 exts[nr].disk_bytenr =
6937                         btrfs_file_extent_disk_bytenr(leaf, fi);
6938                 exts[nr].disk_num_bytes =
6939                         btrfs_file_extent_disk_num_bytes(leaf, fi);
6940                 exts[nr].offset = btrfs_file_extent_offset(leaf, fi);
6941                 exts[nr].num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
6942                 exts[nr].ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
6943                 exts[nr].compression = btrfs_file_extent_compression(leaf, fi);
6944                 exts[nr].encryption = btrfs_file_extent_encryption(leaf, fi);
6945                 exts[nr].other_encoding = btrfs_file_extent_other_encoding(leaf,
6946                                                                            fi);
6947                 BUG_ON(exts[nr].offset > 0);
6948                 BUG_ON(exts[nr].compression || exts[nr].encryption);
6949                 BUG_ON(exts[nr].num_bytes != exts[nr].disk_num_bytes);
6950
6951                 cur_pos += exts[nr].num_bytes;
6952                 nr++;
6953
6954                 if (cur_pos + offset >= last_byte)
6955                         break;
6956
6957                 if (no_fragment) {
6958                         ret = 1;
6959                         goto out;
6960                 }
6961                 path->slots[0]++;
6962         }
6963
6964         BUG_ON(cur_pos + offset > last_byte);
6965         if (cur_pos + offset < last_byte) {
6966                 ret = -ENOENT;
6967                 goto out;
6968         }
6969         ret = 0;
6970 out:
6971         btrfs_free_path(path);
6972         if (ret) {
6973                 if (exts != *extents)
6974                         kfree(exts);
6975         } else {
6976                 *extents = exts;
6977                 *nr_extents = nr;
6978         }
6979         return ret;
6980 }
6981
6982 static noinline int replace_one_extent(struct btrfs_trans_handle *trans,
6983                                         struct btrfs_root *root,
6984                                         struct btrfs_path *path,
6985                                         struct btrfs_key *extent_key,
6986                                         struct btrfs_key *leaf_key,
6987                                         struct btrfs_ref_path *ref_path,
6988                                         struct disk_extent *new_extents,
6989                                         int nr_extents)
6990 {
6991         struct extent_buffer *leaf;
6992         struct btrfs_file_extent_item *fi;
6993         struct inode *inode = NULL;
6994         struct btrfs_key key;
6995         u64 lock_start = 0;
6996         u64 lock_end = 0;
6997         u64 num_bytes;
6998         u64 ext_offset;
6999         u64 search_end = (u64)-1;
7000         u32 nritems;
7001         int nr_scaned = 0;
7002         int extent_locked = 0;
7003         int extent_type;
7004         int ret;
7005
7006         memcpy(&key, leaf_key, sizeof(key));
7007         if (ref_path->owner_objectid != BTRFS_MULTIPLE_OBJECTIDS) {
7008                 if (key.objectid < ref_path->owner_objectid ||
7009                     (key.objectid == ref_path->owner_objectid &&
7010                      key.type < BTRFS_EXTENT_DATA_KEY)) {
7011                         key.objectid = ref_path->owner_objectid;
7012                         key.type = BTRFS_EXTENT_DATA_KEY;
7013                         key.offset = 0;
7014                 }
7015         }
7016
7017         while (1) {
7018                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
7019                 if (ret < 0)
7020                         goto out;
7021
7022                 leaf = path->nodes[0];
7023                 nritems = btrfs_header_nritems(leaf);
7024 next:
7025                 if (extent_locked && ret > 0) {
7026                         /*
7027                          * the file extent item was modified by someone
7028                          * before the extent got locked.
7029                          */
7030                         unlock_extent(&BTRFS_I(inode)->io_tree, lock_start,
7031                                       lock_end, GFP_NOFS);
7032                         extent_locked = 0;
7033                 }
7034
7035                 if (path->slots[0] >= nritems) {
7036                         if (++nr_scaned > 2)
7037                                 break;
7038
7039                         BUG_ON(extent_locked);
7040                         ret = btrfs_next_leaf(root, path);
7041                         if (ret < 0)
7042                                 goto out;
7043                         if (ret > 0)
7044                                 break;
7045                         leaf = path->nodes[0];
7046                         nritems = btrfs_header_nritems(leaf);
7047                 }
7048
7049                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
7050
7051                 if (ref_path->owner_objectid != BTRFS_MULTIPLE_OBJECTIDS) {
7052                         if ((key.objectid > ref_path->owner_objectid) ||
7053                             (key.objectid == ref_path->owner_objectid &&
7054                              key.type > BTRFS_EXTENT_DATA_KEY) ||
7055                             key.offset >= search_end)
7056                                 break;
7057                 }
7058
7059                 if (inode && key.objectid != inode->i_ino) {
7060                         BUG_ON(extent_locked);
7061                         btrfs_release_path(root, path);
7062                         mutex_unlock(&inode->i_mutex);
7063                         iput(inode);
7064                         inode = NULL;
7065                         continue;
7066                 }
7067
7068                 if (key.type != BTRFS_EXTENT_DATA_KEY) {
7069                         path->slots[0]++;
7070                         ret = 1;
7071                         goto next;
7072                 }
7073                 fi = btrfs_item_ptr(leaf, path->slots[0],
7074                                     struct btrfs_file_extent_item);
7075                 extent_type = btrfs_file_extent_type(leaf, fi);
7076                 if ((extent_type != BTRFS_FILE_EXTENT_REG &&
7077                      extent_type != BTRFS_FILE_EXTENT_PREALLOC) ||
7078                     (btrfs_file_extent_disk_bytenr(leaf, fi) !=
7079                      extent_key->objectid)) {
7080                         path->slots[0]++;
7081                         ret = 1;
7082                         goto next;
7083                 }
7084
7085                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
7086                 ext_offset = btrfs_file_extent_offset(leaf, fi);
7087
7088                 if (search_end == (u64)-1) {
7089                         search_end = key.offset - ext_offset +
7090                                 btrfs_file_extent_ram_bytes(leaf, fi);
7091                 }
7092
7093                 if (!extent_locked) {
7094                         lock_start = key.offset;
7095                         lock_end = lock_start + num_bytes - 1;
7096                 } else {
7097                         if (lock_start > key.offset ||
7098                             lock_end + 1 < key.offset + num_bytes) {
7099                                 unlock_extent(&BTRFS_I(inode)->io_tree,
7100                                               lock_start, lock_end, GFP_NOFS);
7101                                 extent_locked = 0;
7102                         }
7103                 }
7104
7105                 if (!inode) {
7106                         btrfs_release_path(root, path);
7107
7108                         inode = btrfs_iget_locked(root->fs_info->sb,
7109                                                   key.objectid, root);
7110                         if (inode->i_state & I_NEW) {
7111                                 BTRFS_I(inode)->root = root;
7112                                 BTRFS_I(inode)->location.objectid =
7113                                         key.objectid;
7114                                 BTRFS_I(inode)->location.type =
7115                                         BTRFS_INODE_ITEM_KEY;
7116                                 BTRFS_I(inode)->location.offset = 0;
7117                                 btrfs_read_locked_inode(inode);
7118                                 unlock_new_inode(inode);
7119                         }
7120                         /*
7121                          * some code call btrfs_commit_transaction while
7122                          * holding the i_mutex, so we can't use mutex_lock
7123                          * here.
7124                          */
7125                         if (is_bad_inode(inode) ||
7126                             !mutex_trylock(&inode->i_mutex)) {
7127                                 iput(inode);
7128                                 inode = NULL;
7129                                 key.offset = (u64)-1;
7130                                 goto skip;
7131                         }
7132                 }
7133
7134                 if (!extent_locked) {
7135                         struct btrfs_ordered_extent *ordered;
7136
7137                         btrfs_release_path(root, path);
7138
7139                         lock_extent(&BTRFS_I(inode)->io_tree, lock_start,
7140                                     lock_end, GFP_NOFS);
7141                         ordered = btrfs_lookup_first_ordered_extent(inode,
7142                                                                     lock_end);
7143                         if (ordered &&
7144                             ordered->file_offset <= lock_end &&
7145                             ordered->file_offset + ordered->len > lock_start) {
7146                                 unlock_extent(&BTRFS_I(inode)->io_tree,
7147                                               lock_start, lock_end, GFP_NOFS);
7148                                 btrfs_start_ordered_extent(inode, ordered, 1);
7149                                 btrfs_put_ordered_extent(ordered);
7150                                 key.offset += num_bytes;
7151                                 goto skip;
7152                         }
7153                         if (ordered)
7154                                 btrfs_put_ordered_extent(ordered);
7155
7156                         extent_locked = 1;
7157                         continue;
7158                 }
7159
7160                 if (nr_extents == 1) {
7161                         /* update extent pointer in place */
7162                         btrfs_set_file_extent_disk_bytenr(leaf, fi,
7163                                                 new_extents[0].disk_bytenr);
7164                         btrfs_set_file_extent_disk_num_bytes(leaf, fi,
7165                                                 new_extents[0].disk_num_bytes);
7166                         btrfs_mark_buffer_dirty(leaf);
7167
7168                         btrfs_drop_extent_cache(inode, key.offset,
7169                                                 key.offset + num_bytes - 1, 0);
7170
7171                         ret = btrfs_inc_extent_ref(trans, root,
7172                                                 new_extents[0].disk_bytenr,
7173                                                 new_extents[0].disk_num_bytes,
7174                                                 leaf->start,
7175                                                 root->root_key.objectid,
7176                                                 trans->transid,
7177                                                 key.objectid);
7178                         BUG_ON(ret);
7179
7180                         ret = btrfs_free_extent(trans, root,
7181                                                 extent_key->objectid,
7182                                                 extent_key->offset,
7183                                                 leaf->start,
7184                                                 btrfs_header_owner(leaf),
7185                                                 btrfs_header_generation(leaf),
7186                                                 key.objectid, 0);
7187                         BUG_ON(ret);
7188
7189                         btrfs_release_path(root, path);
7190                         key.offset += num_bytes;
7191                 } else {
7192                         BUG_ON(1);
7193 #if 0
7194                         u64 alloc_hint;
7195                         u64 extent_len;
7196                         int i;
7197                         /*
7198                          * drop old extent pointer at first, then insert the
7199                          * new pointers one bye one
7200                          */
7201                         btrfs_release_path(root, path);
7202                         ret = btrfs_drop_extents(trans, root, inode, key.offset,
7203                                                  key.offset + num_bytes,
7204                                                  key.offset, &alloc_hint);
7205                         BUG_ON(ret);
7206
7207                         for (i = 0; i < nr_extents; i++) {
7208                                 if (ext_offset >= new_extents[i].num_bytes) {
7209                                         ext_offset -= new_extents[i].num_bytes;
7210                                         continue;
7211                                 }
7212                                 extent_len = min(new_extents[i].num_bytes -
7213                                                  ext_offset, num_bytes);
7214
7215                                 ret = btrfs_insert_empty_item(trans, root,
7216                                                               path, &key,
7217                                                               sizeof(*fi));
7218                                 BUG_ON(ret);
7219
7220                                 leaf = path->nodes[0];
7221                                 fi = btrfs_item_ptr(leaf, path->slots[0],
7222                                                 struct btrfs_file_extent_item);
7223                                 btrfs_set_file_extent_generation(leaf, fi,
7224                                                         trans->transid);
7225                                 btrfs_set_file_extent_type(leaf, fi,
7226                                                         BTRFS_FILE_EXTENT_REG);
7227                                 btrfs_set_file_extent_disk_bytenr(leaf, fi,
7228                                                 new_extents[i].disk_bytenr);
7229                                 btrfs_set_file_extent_disk_num_bytes(leaf, fi,
7230                                                 new_extents[i].disk_num_bytes);
7231                                 btrfs_set_file_extent_ram_bytes(leaf, fi,
7232                                                 new_extents[i].ram_bytes);
7233
7234                                 btrfs_set_file_extent_compression(leaf, fi,
7235                                                 new_extents[i].compression);
7236                                 btrfs_set_file_extent_encryption(leaf, fi,
7237                                                 new_extents[i].encryption);
7238                                 btrfs_set_file_extent_other_encoding(leaf, fi,
7239                                                 new_extents[i].other_encoding);
7240
7241                                 btrfs_set_file_extent_num_bytes(leaf, fi,
7242                                                         extent_len);
7243                                 ext_offset += new_extents[i].offset;
7244                                 btrfs_set_file_extent_offset(leaf, fi,
7245                                                         ext_offset);
7246                                 btrfs_mark_buffer_dirty(leaf);
7247
7248                                 btrfs_drop_extent_cache(inode, key.offset,
7249                                                 key.offset + extent_len - 1, 0);
7250
7251                                 ret = btrfs_inc_extent_ref(trans, root,
7252                                                 new_extents[i].disk_bytenr,
7253                                                 new_extents[i].disk_num_bytes,
7254                                                 leaf->start,
7255                                                 root->root_key.objectid,
7256                                                 trans->transid, key.objectid);
7257                                 BUG_ON(ret);
7258                                 btrfs_release_path(root, path);
7259
7260                                 inode_add_bytes(inode, extent_len);
7261
7262                                 ext_offset = 0;
7263                                 num_bytes -= extent_len;
7264                                 key.offset += extent_len;
7265
7266                                 if (num_bytes == 0)
7267                                         break;
7268                         }
7269                         BUG_ON(i >= nr_extents);
7270 #endif
7271                 }
7272
7273                 if (extent_locked) {
7274                         unlock_extent(&BTRFS_I(inode)->io_tree, lock_start,
7275                                       lock_end, GFP_NOFS);
7276                         extent_locked = 0;
7277                 }
7278 skip:
7279                 if (ref_path->owner_objectid != BTRFS_MULTIPLE_OBJECTIDS &&
7280                     key.offset >= search_end)
7281                         break;
7282
7283                 cond_resched();
7284         }
7285         ret = 0;
7286 out:
7287         btrfs_release_path(root, path);
7288         if (inode) {
7289                 mutex_unlock(&inode->i_mutex);
7290                 if (extent_locked) {
7291                         unlock_extent(&BTRFS_I(inode)->io_tree, lock_start,
7292                                       lock_end, GFP_NOFS);
7293                 }
7294                 iput(inode);
7295         }
7296         return ret;
7297 }
7298
7299 int btrfs_reloc_tree_cache_ref(struct btrfs_trans_handle *trans,
7300                                struct btrfs_root *root,
7301                                struct extent_buffer *buf, u64 orig_start)
7302 {
7303         int level;
7304         int ret;
7305
7306         BUG_ON(btrfs_header_generation(buf) != trans->transid);
7307         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7308
7309         level = btrfs_header_level(buf);
7310         if (level == 0) {
7311                 struct btrfs_leaf_ref *ref;
7312                 struct btrfs_leaf_ref *orig_ref;
7313
7314                 orig_ref = btrfs_lookup_leaf_ref(root, orig_start);
7315                 if (!orig_ref)
7316                         return -ENOENT;
7317
7318                 ref = btrfs_alloc_leaf_ref(root, orig_ref->nritems);
7319                 if (!ref) {
7320                         btrfs_free_leaf_ref(root, orig_ref);
7321                         return -ENOMEM;
7322                 }
7323
7324                 ref->nritems = orig_ref->nritems;
7325                 memcpy(ref->extents, orig_ref->extents,
7326                         sizeof(ref->extents[0]) * ref->nritems);
7327
7328                 btrfs_free_leaf_ref(root, orig_ref);
7329
7330                 ref->root_gen = trans->transid;
7331                 ref->bytenr = buf->start;
7332                 ref->owner = btrfs_header_owner(buf);
7333                 ref->generation = btrfs_header_generation(buf);
7334
7335                 ret = btrfs_add_leaf_ref(root, ref, 0);
7336                 WARN_ON(ret);
7337                 btrfs_free_leaf_ref(root, ref);
7338         }
7339         return 0;
7340 }
7341
7342 static noinline int invalidate_extent_cache(struct btrfs_root *root,
7343                                         struct extent_buffer *leaf,
7344                                         struct btrfs_block_group_cache *group,
7345                                         struct btrfs_root *target_root)
7346 {
7347         struct btrfs_key key;
7348         struct inode *inode = NULL;
7349         struct btrfs_file_extent_item *fi;
7350         struct extent_state *cached_state = NULL;
7351         u64 num_bytes;
7352         u64 skip_objectid = 0;
7353         u32 nritems;
7354         u32 i;
7355
7356         nritems = btrfs_header_nritems(leaf);
7357         for (i = 0; i < nritems; i++) {
7358                 btrfs_item_key_to_cpu(leaf, &key, i);
7359                 if (key.objectid == skip_objectid ||
7360                     key.type != BTRFS_EXTENT_DATA_KEY)
7361                         continue;
7362                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
7363                 if (btrfs_file_extent_type(leaf, fi) ==
7364                     BTRFS_FILE_EXTENT_INLINE)
7365                         continue;
7366                 if (btrfs_file_extent_disk_bytenr(leaf, fi) == 0)
7367                         continue;
7368                 if (!inode || inode->i_ino != key.objectid) {
7369                         iput(inode);
7370                         inode = btrfs_ilookup(target_root->fs_info->sb,
7371                                               key.objectid, target_root, 1);
7372                 }
7373                 if (!inode) {
7374                         skip_objectid = key.objectid;
7375                         continue;
7376                 }
7377                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
7378
7379                 lock_extent_bits(&BTRFS_I(inode)->io_tree, key.offset,
7380                                  key.offset + num_bytes - 1, 0, &cached_state,
7381                                  GFP_NOFS);
7382                 btrfs_drop_extent_cache(inode, key.offset,
7383                                         key.offset + num_bytes - 1, 1);
7384                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, key.offset,
7385                                      key.offset + num_bytes - 1, &cached_state,
7386                                      GFP_NOFS);
7387                 cond_resched();
7388         }
7389         iput(inode);
7390         return 0;
7391 }
7392
7393 static noinline int replace_extents_in_leaf(struct btrfs_trans_handle *trans,
7394                                         struct btrfs_root *root,
7395                                         struct extent_buffer *leaf,
7396                                         struct btrfs_block_group_cache *group,
7397                                         struct inode *reloc_inode)
7398 {
7399         struct btrfs_key key;
7400         struct btrfs_key extent_key;
7401         struct btrfs_file_extent_item *fi;
7402         struct btrfs_leaf_ref *ref;
7403         struct disk_extent *new_extent;
7404         u64 bytenr;
7405         u64 num_bytes;
7406         u32 nritems;
7407         u32 i;
7408         int ext_index;
7409         int nr_extent;
7410         int ret;
7411
7412         new_extent = kmalloc(sizeof(*new_extent), GFP_NOFS);
7413         BUG_ON(!new_extent);
7414
7415         ref = btrfs_lookup_leaf_ref(root, leaf->start);
7416         BUG_ON(!ref);
7417
7418         ext_index = -1;
7419         nritems = btrfs_header_nritems(leaf);
7420         for (i = 0; i < nritems; i++) {
7421                 btrfs_item_key_to_cpu(leaf, &key, i);
7422                 if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
7423                         continue;
7424                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
7425                 if (btrfs_file_extent_type(leaf, fi) ==
7426                     BTRFS_FILE_EXTENT_INLINE)
7427                         continue;
7428                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7429                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
7430                 if (bytenr == 0)
7431                         continue;
7432
7433                 ext_index++;
7434                 if (bytenr >= group->key.objectid + group->key.offset ||
7435                     bytenr + num_bytes <= group->key.objectid)
7436                         continue;
7437
7438                 extent_key.objectid = bytenr;
7439                 extent_key.offset = num_bytes;
7440                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
7441                 nr_extent = 1;
7442                 ret = get_new_locations(reloc_inode, &extent_key,
7443                                         group->key.objectid, 1,
7444                                         &new_extent, &nr_extent);
7445                 if (ret > 0)
7446                         continue;
7447                 BUG_ON(ret < 0);
7448
7449                 BUG_ON(ref->extents[ext_index].bytenr != bytenr);
7450                 BUG_ON(ref->extents[ext_index].num_bytes != num_bytes);
7451                 ref->extents[ext_index].bytenr = new_extent->disk_bytenr;
7452                 ref->extents[ext_index].num_bytes = new_extent->disk_num_bytes;
7453
7454                 btrfs_set_file_extent_disk_bytenr(leaf, fi,
7455                                                 new_extent->disk_bytenr);
7456                 btrfs_set_file_extent_disk_num_bytes(leaf, fi,
7457                                                 new_extent->disk_num_bytes);
7458                 btrfs_mark_buffer_dirty(leaf);
7459
7460                 ret = btrfs_inc_extent_ref(trans, root,
7461                                         new_extent->disk_bytenr,
7462                                         new_extent->disk_num_bytes,
7463                                         leaf->start,
7464                                         root->root_key.objectid,
7465                                         trans->transid, key.objectid);
7466                 BUG_ON(ret);
7467
7468                 ret = btrfs_free_extent(trans, root,
7469                                         bytenr, num_bytes, leaf->start,
7470                                         btrfs_header_owner(leaf),
7471                                         btrfs_header_generation(leaf),
7472                                         key.objectid, 0);
7473                 BUG_ON(ret);
7474                 cond_resched();
7475         }
7476         kfree(new_extent);
7477         BUG_ON(ext_index + 1 != ref->nritems);
7478         btrfs_free_leaf_ref(root, ref);
7479         return 0;
7480 }
7481
7482 int btrfs_free_reloc_root(struct btrfs_trans_handle *trans,
7483                           struct btrfs_root *root)
7484 {
7485         struct btrfs_root *reloc_root;
7486         int ret;
7487
7488         if (root->reloc_root) {
7489                 reloc_root = root->reloc_root;
7490                 root->reloc_root = NULL;
7491                 list_add(&reloc_root->dead_list,
7492                          &root->fs_info->dead_reloc_roots);
7493
7494                 btrfs_set_root_bytenr(&reloc_root->root_item,
7495                                       reloc_root->node->start);
7496                 btrfs_set_root_level(&root->root_item,
7497                                      btrfs_header_level(reloc_root->node));
7498                 memset(&reloc_root->root_item.drop_progress, 0,
7499                         sizeof(struct btrfs_disk_key));
7500                 reloc_root->root_item.drop_level = 0;
7501
7502                 ret = btrfs_update_root(trans, root->fs_info->tree_root,
7503                                         &reloc_root->root_key,
7504                                         &reloc_root->root_item);
7505                 BUG_ON(ret);
7506         }
7507         return 0;
7508 }
7509
7510 int btrfs_drop_dead_reloc_roots(struct btrfs_root *root)
7511 {
7512         struct btrfs_trans_handle *trans;
7513         struct btrfs_root *reloc_root;
7514         struct btrfs_root *prev_root = NULL;
7515         struct list_head dead_roots;
7516         int ret;
7517         unsigned long nr;
7518
7519         INIT_LIST_HEAD(&dead_roots);
7520         list_splice_init(&root->fs_info->dead_reloc_roots, &dead_roots);
7521
7522         while (!list_empty(&dead_roots)) {
7523                 reloc_root = list_entry(dead_roots.prev,
7524                                         struct btrfs_root, dead_list);
7525                 list_del_init(&reloc_root->dead_list);
7526
7527                 BUG_ON(reloc_root->commit_root != NULL);
7528                 while (1) {
7529                         trans = btrfs_join_transaction(root, 1);
7530                         BUG_ON(IS_ERR(trans));
7531
7532                         mutex_lock(&root->fs_info->drop_mutex);
7533                         ret = btrfs_drop_snapshot(trans, reloc_root);
7534                         if (ret != -EAGAIN)
7535                                 break;
7536                         mutex_unlock(&root->fs_info->drop_mutex);
7537
7538                         nr = trans->blocks_used;
7539                         ret = btrfs_end_transaction(trans, root);
7540                         BUG_ON(ret);
7541                         btrfs_btree_balance_dirty(root, nr);
7542                 }
7543
7544                 free_extent_buffer(reloc_root->node);
7545
7546                 ret = btrfs_del_root(trans, root->fs_info->tree_root,
7547                                      &reloc_root->root_key);
7548                 BUG_ON(ret);
7549                 mutex_unlock(&root->fs_info->drop_mutex);
7550
7551                 nr = trans->blocks_used;
7552                 ret = btrfs_end_transaction(trans, root);
7553                 BUG_ON(ret);
7554                 btrfs_btree_balance_dirty(root, nr);
7555
7556                 kfree(prev_root);
7557                 prev_root = reloc_root;
7558         }
7559         if (prev_root) {
7560                 btrfs_remove_leaf_refs(prev_root, (u64)-1, 0);
7561                 kfree(prev_root);
7562         }
7563         return 0;
7564 }
7565
7566 int btrfs_add_dead_reloc_root(struct btrfs_root *root)
7567 {
7568         list_add(&root->dead_list, &root->fs_info->dead_reloc_roots);
7569         return 0;
7570 }
7571
7572 int btrfs_cleanup_reloc_trees(struct btrfs_root *root)
7573 {
7574         struct btrfs_root *reloc_root;
7575         struct btrfs_trans_handle *trans;
7576         struct btrfs_key location;
7577         int found;
7578         int ret;
7579
7580         mutex_lock(&root->fs_info->tree_reloc_mutex);
7581         ret = btrfs_find_dead_roots(root, BTRFS_TREE_RELOC_OBJECTID, NULL);
7582         BUG_ON(ret);
7583         found = !list_empty(&root->fs_info->dead_reloc_roots);
7584         mutex_unlock(&root->fs_info->tree_reloc_mutex);
7585
7586         if (found) {
7587                 trans = btrfs_start_transaction(root, 1);
7588                 BUG_ON(!trans);
7589                 ret = btrfs_commit_transaction(trans, root);
7590                 BUG_ON(ret);
7591         }
7592
7593         location.objectid = BTRFS_DATA_RELOC_TREE_OBJECTID;
7594         location.offset = (u64)-1;
7595         location.type = BTRFS_ROOT_ITEM_KEY;
7596
7597         reloc_root = btrfs_read_fs_root_no_name(root->fs_info, &location);
7598         BUG_ON(!reloc_root);
7599         btrfs_orphan_cleanup(reloc_root);
7600         return 0;
7601 }
7602
7603 static noinline int init_reloc_tree(struct btrfs_trans_handle *trans,
7604                                     struct btrfs_root *root)
7605 {
7606         struct btrfs_root *reloc_root;
7607         struct extent_buffer *eb;
7608         struct btrfs_root_item *root_item;
7609         struct btrfs_key root_key;
7610         int ret;
7611
7612         BUG_ON(!root->ref_cows);
7613         if (root->reloc_root)
7614                 return 0;
7615
7616         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
7617         BUG_ON(!root_item);
7618
7619         ret = btrfs_copy_root(trans, root, root->commit_root,
7620                               &eb, BTRFS_TREE_RELOC_OBJECTID);
7621         BUG_ON(ret);
7622
7623         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
7624         root_key.offset = root->root_key.objectid;
7625         root_key.type = BTRFS_ROOT_ITEM_KEY;
7626
7627         memcpy(root_item, &root->root_item, sizeof(root_item));
7628         btrfs_set_root_refs(root_item, 0);
7629         btrfs_set_root_bytenr(root_item, eb->start);
7630         btrfs_set_root_level(root_item, btrfs_header_level(eb));
7631         btrfs_set_root_generation(root_item, trans->transid);
7632
7633         btrfs_tree_unlock(eb);
7634         free_extent_buffer(eb);
7635
7636         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
7637                                 &root_key, root_item);
7638         BUG_ON(ret);
7639         kfree(root_item);
7640
7641         reloc_root = btrfs_read_fs_root_no_radix(root->fs_info->tree_root,
7642                                                  &root_key);
7643         BUG_ON(!reloc_root);
7644         reloc_root->last_trans = trans->transid;
7645         reloc_root->commit_root = NULL;
7646         reloc_root->ref_tree = &root->fs_info->reloc_ref_tree;
7647
7648         root->reloc_root = reloc_root;
7649         return 0;
7650 }
7651
7652 /*
7653  * Core function of space balance.
7654  *
7655  * The idea is using reloc trees to relocate tree blocks in reference
7656  * counted roots. There is one reloc tree for each subvol, and all
7657  * reloc trees share same root key objectid. Reloc trees are snapshots
7658  * of the latest committed roots of subvols (root->commit_root).
7659  *
7660  * To relocate a tree block referenced by a subvol, there are two steps.
7661  * COW the block through subvol's reloc tree, then update block pointer
7662  * in the subvol to point to the new block. Since all reloc trees share
7663  * same root key objectid, doing special handing for tree blocks owned
7664  * by them is easy. Once a tree block has been COWed in one reloc tree,
7665  * we can use the resulting new block directly when the same block is
7666  * required to COW again through other reloc trees. By this way, relocated
7667  * tree blocks are shared between reloc trees, so they are also shared
7668  * between subvols.
7669  */
7670 static noinline int relocate_one_path(struct btrfs_trans_handle *trans,
7671                                       struct btrfs_root *root,
7672                                       struct btrfs_path *path,
7673                                       struct btrfs_key *first_key,
7674                                       struct btrfs_ref_path *ref_path,
7675                                       struct btrfs_block_group_cache *group,
7676                                       struct inode *reloc_inode)
7677 {
7678         struct btrfs_root *reloc_root;
7679         struct extent_buffer *eb = NULL;
7680         struct btrfs_key *keys;
7681         u64 *nodes;
7682         int level;
7683         int shared_level;
7684         int lowest_level = 0;
7685         int ret;
7686
7687         if (ref_path->owner_objectid < BTRFS_FIRST_FREE_OBJECTID)
7688                 lowest_level = ref_path->owner_objectid;
7689
7690         if (!root->ref_cows) {
7691                 path->lowest_level = lowest_level;
7692                 ret = btrfs_search_slot(trans, root, first_key, path, 0, 1);
7693                 BUG_ON(ret < 0);
7694                 path->lowest_level = 0;
7695                 btrfs_release_path(root, path);
7696                 return 0;
7697         }
7698
7699         mutex_lock(&root->fs_info->tree_reloc_mutex);
7700         ret = init_reloc_tree(trans, root);
7701         BUG_ON(ret);
7702         reloc_root = root->reloc_root;
7703
7704         shared_level = ref_path->shared_level;
7705         ref_path->shared_level = BTRFS_MAX_LEVEL - 1;
7706
7707         keys = ref_path->node_keys;
7708         nodes = ref_path->new_nodes;
7709         memset(&keys[shared_level + 1], 0,
7710                sizeof(*keys) * (BTRFS_MAX_LEVEL - shared_level - 1));
7711         memset(&nodes[shared_level + 1], 0,
7712                sizeof(*nodes) * (BTRFS_MAX_LEVEL - shared_level - 1));
7713
7714         if (nodes[lowest_level] == 0) {
7715                 path->lowest_level = lowest_level;
7716                 ret = btrfs_search_slot(trans, reloc_root, first_key, path,
7717                                         0, 1);
7718                 BUG_ON(ret);
7719                 for (level = lowest_level; level < BTRFS_MAX_LEVEL; level++) {
7720                         eb = path->nodes[level];
7721                         if (!eb || eb == reloc_root->node)
7722                                 break;
7723                         nodes[level] = eb->start;
7724                         if (level == 0)
7725                                 btrfs_item_key_to_cpu(eb, &keys[level], 0);
7726                         else
7727                                 btrfs_node_key_to_cpu(eb, &keys[level], 0);
7728                 }
7729                 if (nodes[0] &&
7730                     ref_path->owner_objectid >= BTRFS_FIRST_FREE_OBJECTID) {
7731                         eb = path->nodes[0];
7732                         ret = replace_extents_in_leaf(trans, reloc_root, eb,
7733                                                       group, reloc_inode);
7734                         BUG_ON(ret);
7735                 }
7736                 btrfs_release_path(reloc_root, path);
7737         } else {
7738                 ret = btrfs_merge_path(trans, reloc_root, keys, nodes,
7739                                        lowest_level);
7740                 BUG_ON(ret);
7741         }
7742
7743         /*
7744          * replace tree blocks in the fs tree with tree blocks in
7745          * the reloc tree.
7746          */
7747         ret = btrfs_merge_path(trans, root, keys, nodes, lowest_level);
7748         BUG_ON(ret < 0);
7749
7750         if (ref_path->owner_objectid >= BTRFS_FIRST_FREE_OBJECTID) {
7751                 ret = btrfs_search_slot(trans, reloc_root, first_key, path,
7752                                         0, 0);
7753                 BUG_ON(ret);
7754                 extent_buffer_get(path->nodes[0]);
7755                 eb = path->nodes[0];
7756                 btrfs_release_path(reloc_root, path);
7757                 ret = invalidate_extent_cache(reloc_root, eb, group, root);
7758                 BUG_ON(ret);
7759                 free_extent_buffer(eb);
7760         }
7761
7762         mutex_unlock(&root->fs_info->tree_reloc_mutex);
7763         path->lowest_level = 0;
7764         return 0;
7765 }
7766
7767 static noinline int relocate_tree_block(struct btrfs_trans_handle *trans,
7768                                         struct btrfs_root *root,
7769                                         struct btrfs_path *path,
7770                                         struct btrfs_key *first_key,
7771                                         struct btrfs_ref_path *ref_path)
7772 {
7773         int ret;
7774
7775         ret = relocate_one_path(trans, root, path, first_key,
7776                                 ref_path, NULL, NULL);
7777         BUG_ON(ret);
7778
7779         return 0;
7780 }
7781
7782 static noinline int del_extent_zero(struct btrfs_trans_handle *trans,
7783                                     struct btrfs_root *extent_root,
7784                                     struct btrfs_path *path,
7785                                     struct btrfs_key *extent_key)
7786 {
7787         int ret;
7788
7789         ret = btrfs_search_slot(trans, extent_root, extent_key, path, -1, 1);
7790         if (ret)
7791                 goto out;
7792         ret = btrfs_del_item(trans, extent_root, path);
7793 out:
7794         btrfs_release_path(extent_root, path);
7795         return ret;
7796 }
7797
7798 static noinline struct btrfs_root *read_ref_root(struct btrfs_fs_info *fs_info,
7799                                                 struct btrfs_ref_path *ref_path)
7800 {
7801         struct btrfs_key root_key;
7802
7803         root_key.objectid = ref_path->root_objectid;
7804         root_key.type = BTRFS_ROOT_ITEM_KEY;
7805         if (is_cowonly_root(ref_path->root_objectid))
7806                 root_key.offset = 0;
7807         else
7808                 root_key.offset = (u64)-1;
7809
7810         return btrfs_read_fs_root_no_name(fs_info, &root_key);
7811 }
7812
7813 static noinline int relocate_one_extent(struct btrfs_root *extent_root,
7814                                         struct btrfs_path *path,
7815                                         struct btrfs_key *extent_key,
7816                                         struct btrfs_block_group_cache *group,
7817                                         struct inode *reloc_inode, int pass)
7818 {
7819         struct btrfs_trans_handle *trans;
7820         struct btrfs_root *found_root;
7821         struct btrfs_ref_path *ref_path = NULL;
7822         struct disk_extent *new_extents = NULL;
7823         int nr_extents = 0;
7824         int loops;
7825         int ret;
7826         int level;
7827         struct btrfs_key first_key;
7828         u64 prev_block = 0;
7829
7830
7831         trans = btrfs_start_transaction(extent_root, 1);
7832         BUG_ON(!trans);
7833
7834         if (extent_key->objectid == 0) {
7835                 ret = del_extent_zero(trans, extent_root, path, extent_key);
7836                 goto out;
7837         }
7838
7839         ref_path = kmalloc(sizeof(*ref_path), GFP_NOFS);
7840         if (!ref_path) {
7841                 ret = -ENOMEM;
7842                 goto out;
7843         }
7844
7845         for (loops = 0; ; loops++) {
7846                 if (loops == 0) {
7847                         ret = btrfs_first_ref_path(trans, extent_root, ref_path,
7848                                                    extent_key->objectid);
7849                 } else {
7850                         ret = btrfs_next_ref_path(trans, extent_root, ref_path);
7851                 }
7852                 if (ret < 0)
7853                         goto out;
7854                 if (ret > 0)
7855                         break;
7856
7857                 if (ref_path->root_objectid == BTRFS_TREE_LOG_OBJECTID ||
7858                     ref_path->root_objectid == BTRFS_TREE_RELOC_OBJECTID)
7859                         continue;
7860
7861                 found_root = read_ref_root(extent_root->fs_info, ref_path);
7862                 BUG_ON(!found_root);
7863                 /*
7864                  * for reference counted tree, only process reference paths
7865                  * rooted at the latest committed root.
7866                  */
7867                 if (found_root->ref_cows &&
7868                     ref_path->root_generation != found_root->root_key.offset)
7869                         continue;
7870
7871                 if (ref_path->owner_objectid >= BTRFS_FIRST_FREE_OBJECTID) {
7872                         if (pass == 0) {
7873                                 /*
7874                                  * copy data extents to new locations
7875                                  */
7876                                 u64 group_start = group->key.objectid;
7877                                 ret = relocate_data_extent(reloc_inode,
7878                                                            extent_key,
7879                                                            group_start);
7880                                 if (ret < 0)
7881                                         goto out;
7882                                 break;
7883                         }
7884                         level = 0;
7885                 } else {
7886                         level = ref_path->owner_objectid;
7887                 }
7888
7889                 if (prev_block != ref_path->nodes[level]) {
7890                         struct extent_buffer *eb;
7891                         u64 block_start = ref_path->nodes[level];
7892                         u64 block_size = btrfs_level_size(found_root, level);
7893
7894                         eb = read_tree_block(found_root, block_start,
7895                                              block_size, 0);
7896                         btrfs_tree_lock(eb);
7897                         BUG_ON(level != btrfs_header_level(eb));
7898
7899                         if (level == 0)
7900                                 btrfs_item_key_to_cpu(eb, &first_key, 0);
7901                         else
7902                                 btrfs_node_key_to_cpu(eb, &first_key, 0);
7903
7904                         btrfs_tree_unlock(eb);
7905                         free_extent_buffer(eb);
7906                         prev_block = block_start;
7907                 }
7908
7909                 mutex_lock(&extent_root->fs_info->trans_mutex);
7910                 btrfs_record_root_in_trans(found_root);
7911                 mutex_unlock(&extent_root->fs_info->trans_mutex);
7912                 if (ref_path->owner_objectid >= BTRFS_FIRST_FREE_OBJECTID) {
7913                         /*
7914                          * try to update data extent references while
7915                          * keeping metadata shared between snapshots.
7916                          */
7917                         if (pass == 1) {
7918                                 ret = relocate_one_path(trans, found_root,
7919                                                 path, &first_key, ref_path,
7920                                                 group, reloc_inode);
7921                                 if (ret < 0)
7922                                         goto out;
7923                                 continue;
7924                         }
7925                         /*
7926                          * use fallback method to process the remaining
7927                          * references.
7928                          */
7929                         if (!new_extents) {
7930                                 u64 group_start = group->key.objectid;
7931                                 new_extents = kmalloc(sizeof(*new_extents),
7932                                                       GFP_NOFS);
7933                                 nr_extents = 1;
7934                                 ret = get_new_locations(reloc_inode,
7935                                                         extent_key,
7936                                                         group_start, 1,
7937                                                         &new_extents,
7938                                                         &nr_extents);
7939                                 if (ret)
7940                                         goto out;
7941                         }
7942                         ret = replace_one_extent(trans, found_root,
7943                                                 path, extent_key,
7944                                                 &first_key, ref_path,
7945                                                 new_extents, nr_extents);
7946                 } else {
7947                         ret = relocate_tree_block(trans, found_root, path,
7948                                                   &first_key, ref_path);
7949                 }
7950                 if (ret < 0)
7951                         goto out;
7952         }
7953         ret = 0;
7954 out:
7955         btrfs_end_transaction(trans, extent_root);
7956         kfree(new_extents);
7957         kfree(ref_path);
7958         return ret;
7959 }
7960 #endif
7961
7962 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7963 {
7964         u64 num_devices;
7965         u64 stripped = BTRFS_BLOCK_GROUP_RAID0 |
7966                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7967
7968         /*
7969          * we add in the count of missing devices because we want
7970          * to make sure that any RAID levels on a degraded FS
7971          * continue to be honored.
7972          */
7973         num_devices = root->fs_info->fs_devices->rw_devices +
7974                 root->fs_info->fs_devices->missing_devices;
7975
7976         if (num_devices == 1) {
7977                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7978                 stripped = flags & ~stripped;
7979
7980                 /* turn raid0 into single device chunks */
7981                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7982                         return stripped;
7983
7984                 /* turn mirroring into duplication */
7985                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7986                              BTRFS_BLOCK_GROUP_RAID10))
7987                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7988                 return flags;
7989         } else {
7990                 /* they already had raid on here, just return */
7991                 if (flags & stripped)
7992                         return flags;
7993
7994                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7995                 stripped = flags & ~stripped;
7996
7997                 /* switch duplicated blocks with raid1 */
7998                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7999                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
8000
8001                 /* turn single device chunks into raid0 */
8002                 return stripped | BTRFS_BLOCK_GROUP_RAID0;
8003         }
8004         return flags;
8005 }
8006
8007 static int set_block_group_ro(struct btrfs_block_group_cache *cache)
8008 {
8009         struct btrfs_space_info *sinfo = cache->space_info;
8010         u64 num_bytes;
8011         int ret = -ENOSPC;
8012
8013         if (cache->ro)
8014                 return 0;
8015
8016         spin_lock(&sinfo->lock);
8017         spin_lock(&cache->lock);
8018         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8019                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8020
8021         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8022             sinfo->bytes_may_use + sinfo->bytes_readonly +
8023             cache->reserved_pinned + num_bytes <= sinfo->total_bytes) {
8024                 sinfo->bytes_readonly += num_bytes;
8025                 sinfo->bytes_reserved += cache->reserved_pinned;
8026                 cache->reserved_pinned = 0;
8027                 cache->ro = 1;
8028                 ret = 0;
8029         }
8030
8031         spin_unlock(&cache->lock);
8032         spin_unlock(&sinfo->lock);
8033         return ret;
8034 }
8035
8036 int btrfs_set_block_group_ro(struct btrfs_root *root,
8037                              struct btrfs_block_group_cache *cache)
8038
8039 {
8040         struct btrfs_trans_handle *trans;
8041         u64 alloc_flags;
8042         int ret;
8043
8044         BUG_ON(cache->ro);
8045
8046         trans = btrfs_join_transaction(root, 1);
8047         BUG_ON(IS_ERR(trans));
8048
8049         alloc_flags = update_block_group_flags(root, cache->flags);
8050         if (alloc_flags != cache->flags)
8051                 do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags, 1);
8052
8053         ret = set_block_group_ro(cache);
8054         if (!ret)
8055                 goto out;
8056         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
8057         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags, 1);
8058         if (ret < 0)
8059                 goto out;
8060         ret = set_block_group_ro(cache);
8061 out:
8062         btrfs_end_transaction(trans, root);
8063         return ret;
8064 }
8065
8066 /*
8067  * helper to account the unused space of all the readonly block group in the
8068  * list. takes mirrors into account.
8069  */
8070 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
8071 {
8072         struct btrfs_block_group_cache *block_group;
8073         u64 free_bytes = 0;
8074         int factor;
8075
8076         list_for_each_entry(block_group, groups_list, list) {
8077                 spin_lock(&block_group->lock);
8078
8079                 if (!block_group->ro) {
8080                         spin_unlock(&block_group->lock);
8081                         continue;
8082                 }
8083
8084                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
8085                                           BTRFS_BLOCK_GROUP_RAID10 |
8086                                           BTRFS_BLOCK_GROUP_DUP))
8087                         factor = 2;
8088                 else
8089                         factor = 1;
8090
8091                 free_bytes += (block_group->key.offset -
8092                                btrfs_block_group_used(&block_group->item)) *
8093                                factor;
8094
8095                 spin_unlock(&block_group->lock);
8096         }
8097
8098         return free_bytes;
8099 }
8100
8101 /*
8102  * helper to account the unused space of all the readonly block group in the
8103  * space_info. takes mirrors into account.
8104  */
8105 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
8106 {
8107         int i;
8108         u64 free_bytes = 0;
8109
8110         spin_lock(&sinfo->lock);
8111
8112         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
8113                 if (!list_empty(&sinfo->block_groups[i]))
8114                         free_bytes += __btrfs_get_ro_block_group_free_space(
8115                                                 &sinfo->block_groups[i]);
8116
8117         spin_unlock(&sinfo->lock);
8118
8119         return free_bytes;
8120 }
8121
8122 int btrfs_set_block_group_rw(struct btrfs_root *root,
8123                               struct btrfs_block_group_cache *cache)
8124 {
8125         struct btrfs_space_info *sinfo = cache->space_info;
8126         u64 num_bytes;
8127
8128         BUG_ON(!cache->ro);
8129
8130         spin_lock(&sinfo->lock);
8131         spin_lock(&cache->lock);
8132         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8133                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8134         sinfo->bytes_readonly -= num_bytes;
8135         cache->ro = 0;
8136         spin_unlock(&cache->lock);
8137         spin_unlock(&sinfo->lock);
8138         return 0;
8139 }
8140
8141 /*
8142  * checks to see if its even possible to relocate this block group.
8143  *
8144  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
8145  * ok to go ahead and try.
8146  */
8147 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
8148 {
8149         struct btrfs_block_group_cache *block_group;
8150         struct btrfs_space_info *space_info;
8151         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
8152         struct btrfs_device *device;
8153         int full = 0;
8154         int ret = 0;
8155
8156         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
8157
8158         /* odd, couldn't find the block group, leave it alone */
8159         if (!block_group)
8160                 return -1;
8161
8162         /* no bytes used, we're good */
8163         if (!btrfs_block_group_used(&block_group->item))
8164                 goto out;
8165
8166         space_info = block_group->space_info;
8167         spin_lock(&space_info->lock);
8168
8169         full = space_info->full;
8170
8171         /*
8172          * if this is the last block group we have in this space, we can't
8173          * relocate it unless we're able to allocate a new chunk below.
8174          *
8175          * Otherwise, we need to make sure we have room in the space to handle
8176          * all of the extents from this block group.  If we can, we're good
8177          */
8178         if ((space_info->total_bytes != block_group->key.offset) &&
8179            (space_info->bytes_used + space_info->bytes_reserved +
8180             space_info->bytes_pinned + space_info->bytes_readonly +
8181             btrfs_block_group_used(&block_group->item) <
8182             space_info->total_bytes)) {
8183                 spin_unlock(&space_info->lock);
8184                 goto out;
8185         }
8186         spin_unlock(&space_info->lock);
8187
8188         /*
8189          * ok we don't have enough space, but maybe we have free space on our
8190          * devices to allocate new chunks for relocation, so loop through our
8191          * alloc devices and guess if we have enough space.  However, if we
8192          * were marked as full, then we know there aren't enough chunks, and we
8193          * can just return.
8194          */
8195         ret = -1;
8196         if (full)
8197                 goto out;
8198
8199         mutex_lock(&root->fs_info->chunk_mutex);
8200         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
8201                 u64 min_free = btrfs_block_group_used(&block_group->item);
8202                 u64 dev_offset;
8203
8204                 /*
8205                  * check to make sure we can actually find a chunk with enough
8206                  * space to fit our block group in.
8207                  */
8208                 if (device->total_bytes > device->bytes_used + min_free) {
8209                         ret = find_free_dev_extent(NULL, device, min_free,
8210                                                    &dev_offset, NULL);
8211                         if (!ret)
8212                                 break;
8213                         ret = -1;
8214                 }
8215         }
8216         mutex_unlock(&root->fs_info->chunk_mutex);
8217 out:
8218         btrfs_put_block_group(block_group);
8219         return ret;
8220 }
8221
8222 static int find_first_block_group(struct btrfs_root *root,
8223                 struct btrfs_path *path, struct btrfs_key *key)
8224 {
8225         int ret = 0;
8226         struct btrfs_key found_key;
8227         struct extent_buffer *leaf;
8228         int slot;
8229
8230         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
8231         if (ret < 0)
8232                 goto out;
8233
8234         while (1) {
8235                 slot = path->slots[0];
8236                 leaf = path->nodes[0];
8237                 if (slot >= btrfs_header_nritems(leaf)) {
8238                         ret = btrfs_next_leaf(root, path);
8239                         if (ret == 0)
8240                                 continue;
8241                         if (ret < 0)
8242                                 goto out;
8243                         break;
8244                 }
8245                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
8246
8247                 if (found_key.objectid >= key->objectid &&
8248                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
8249                         ret = 0;
8250                         goto out;
8251                 }
8252                 path->slots[0]++;
8253         }
8254 out:
8255         return ret;
8256 }
8257
8258 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
8259 {
8260         struct btrfs_block_group_cache *block_group;
8261         u64 last = 0;
8262
8263         while (1) {
8264                 struct inode *inode;
8265
8266                 block_group = btrfs_lookup_first_block_group(info, last);
8267                 while (block_group) {
8268                         spin_lock(&block_group->lock);
8269                         if (block_group->iref)
8270                                 break;
8271                         spin_unlock(&block_group->lock);
8272                         block_group = next_block_group(info->tree_root,
8273                                                        block_group);
8274                 }
8275                 if (!block_group) {
8276                         if (last == 0)
8277                                 break;
8278                         last = 0;
8279                         continue;
8280                 }
8281
8282                 inode = block_group->inode;
8283                 block_group->iref = 0;
8284                 block_group->inode = NULL;
8285                 spin_unlock(&block_group->lock);
8286                 iput(inode);
8287                 last = block_group->key.objectid + block_group->key.offset;
8288                 btrfs_put_block_group(block_group);
8289         }
8290 }
8291
8292 int btrfs_free_block_groups(struct btrfs_fs_info *info)
8293 {
8294         struct btrfs_block_group_cache *block_group;
8295         struct btrfs_space_info *space_info;
8296         struct btrfs_caching_control *caching_ctl;
8297         struct rb_node *n;
8298
8299         down_write(&info->extent_commit_sem);
8300         while (!list_empty(&info->caching_block_groups)) {
8301                 caching_ctl = list_entry(info->caching_block_groups.next,
8302                                          struct btrfs_caching_control, list);
8303                 list_del(&caching_ctl->list);
8304                 put_caching_control(caching_ctl);
8305         }
8306         up_write(&info->extent_commit_sem);
8307
8308         spin_lock(&info->block_group_cache_lock);
8309         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
8310                 block_group = rb_entry(n, struct btrfs_block_group_cache,
8311                                        cache_node);
8312                 rb_erase(&block_group->cache_node,
8313                          &info->block_group_cache_tree);
8314                 spin_unlock(&info->block_group_cache_lock);
8315
8316                 down_write(&block_group->space_info->groups_sem);
8317                 list_del(&block_group->list);
8318                 up_write(&block_group->space_info->groups_sem);
8319
8320                 if (block_group->cached == BTRFS_CACHE_STARTED)
8321                         wait_block_group_cache_done(block_group);
8322
8323                 btrfs_remove_free_space_cache(block_group);
8324                 btrfs_put_block_group(block_group);
8325
8326                 spin_lock(&info->block_group_cache_lock);
8327         }
8328         spin_unlock(&info->block_group_cache_lock);
8329
8330         /* now that all the block groups are freed, go through and
8331          * free all the space_info structs.  This is only called during
8332          * the final stages of unmount, and so we know nobody is
8333          * using them.  We call synchronize_rcu() once before we start,
8334          * just to be on the safe side.
8335          */
8336         synchronize_rcu();
8337
8338         release_global_block_rsv(info);
8339
8340         while(!list_empty(&info->space_info)) {
8341                 space_info = list_entry(info->space_info.next,
8342                                         struct btrfs_space_info,
8343                                         list);
8344                 if (space_info->bytes_pinned > 0 ||
8345                     space_info->bytes_reserved > 0) {
8346                         WARN_ON(1);
8347                         dump_space_info(space_info, 0, 0);
8348                 }
8349                 list_del(&space_info->list);
8350                 kfree(space_info);
8351         }
8352         return 0;
8353 }
8354
8355 static void __link_block_group(struct btrfs_space_info *space_info,
8356                                struct btrfs_block_group_cache *cache)
8357 {
8358         int index = get_block_group_index(cache);
8359
8360         down_write(&space_info->groups_sem);
8361         list_add_tail(&cache->list, &space_info->block_groups[index]);
8362         up_write(&space_info->groups_sem);
8363 }
8364
8365 int btrfs_read_block_groups(struct btrfs_root *root)
8366 {
8367         struct btrfs_path *path;
8368         int ret;
8369         struct btrfs_block_group_cache *cache;
8370         struct btrfs_fs_info *info = root->fs_info;
8371         struct btrfs_space_info *space_info;
8372         struct btrfs_key key;
8373         struct btrfs_key found_key;
8374         struct extent_buffer *leaf;
8375         int need_clear = 0;
8376         u64 cache_gen;
8377
8378         root = info->extent_root;
8379         key.objectid = 0;
8380         key.offset = 0;
8381         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
8382         path = btrfs_alloc_path();
8383         if (!path)
8384                 return -ENOMEM;
8385
8386         cache_gen = btrfs_super_cache_generation(&root->fs_info->super_copy);
8387         if (cache_gen != 0 &&
8388             btrfs_super_generation(&root->fs_info->super_copy) != cache_gen)
8389                 need_clear = 1;
8390         if (btrfs_test_opt(root, CLEAR_CACHE))
8391                 need_clear = 1;
8392         if (!btrfs_test_opt(root, SPACE_CACHE) && cache_gen)
8393                 printk(KERN_INFO "btrfs: disk space caching is enabled\n");
8394
8395         while (1) {
8396                 ret = find_first_block_group(root, path, &key);
8397                 if (ret > 0)
8398                         break;
8399                 if (ret != 0)
8400                         goto error;
8401                 leaf = path->nodes[0];
8402                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
8403                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
8404                 if (!cache) {
8405                         ret = -ENOMEM;
8406                         goto error;
8407                 }
8408
8409                 atomic_set(&cache->count, 1);
8410                 spin_lock_init(&cache->lock);
8411                 spin_lock_init(&cache->tree_lock);
8412                 cache->fs_info = info;
8413                 INIT_LIST_HEAD(&cache->list);
8414                 INIT_LIST_HEAD(&cache->cluster_list);
8415
8416                 if (need_clear)
8417                         cache->disk_cache_state = BTRFS_DC_CLEAR;
8418
8419                 /*
8420                  * we only want to have 32k of ram per block group for keeping
8421                  * track of free space, and if we pass 1/2 of that we want to
8422                  * start converting things over to using bitmaps
8423                  */
8424                 cache->extents_thresh = ((1024 * 32) / 2) /
8425                         sizeof(struct btrfs_free_space);
8426
8427                 read_extent_buffer(leaf, &cache->item,
8428                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
8429                                    sizeof(cache->item));
8430                 memcpy(&cache->key, &found_key, sizeof(found_key));
8431
8432                 key.objectid = found_key.objectid + found_key.offset;
8433                 btrfs_release_path(root, path);
8434                 cache->flags = btrfs_block_group_flags(&cache->item);
8435                 cache->sectorsize = root->sectorsize;
8436
8437                 /*
8438                  * check for two cases, either we are full, and therefore
8439                  * don't need to bother with the caching work since we won't
8440                  * find any space, or we are empty, and we can just add all
8441                  * the space in and be done with it.  This saves us _alot_ of
8442                  * time, particularly in the full case.
8443                  */
8444                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
8445                         exclude_super_stripes(root, cache);
8446                         cache->last_byte_to_unpin = (u64)-1;
8447                         cache->cached = BTRFS_CACHE_FINISHED;
8448                         free_excluded_extents(root, cache);
8449                 } else if (btrfs_block_group_used(&cache->item) == 0) {
8450                         exclude_super_stripes(root, cache);
8451                         cache->last_byte_to_unpin = (u64)-1;
8452                         cache->cached = BTRFS_CACHE_FINISHED;
8453                         add_new_free_space(cache, root->fs_info,
8454                                            found_key.objectid,
8455                                            found_key.objectid +
8456                                            found_key.offset);
8457                         free_excluded_extents(root, cache);
8458                 }
8459
8460                 ret = update_space_info(info, cache->flags, found_key.offset,
8461                                         btrfs_block_group_used(&cache->item),
8462                                         &space_info);
8463                 BUG_ON(ret);
8464                 cache->space_info = space_info;
8465                 spin_lock(&cache->space_info->lock);
8466                 cache->space_info->bytes_readonly += cache->bytes_super;
8467                 spin_unlock(&cache->space_info->lock);
8468
8469                 __link_block_group(space_info, cache);
8470
8471                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
8472                 BUG_ON(ret);
8473
8474                 set_avail_alloc_bits(root->fs_info, cache->flags);
8475                 if (btrfs_chunk_readonly(root, cache->key.objectid))
8476                         set_block_group_ro(cache);
8477         }
8478
8479         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
8480                 if (!(get_alloc_profile(root, space_info->flags) &
8481                       (BTRFS_BLOCK_GROUP_RAID10 |
8482                        BTRFS_BLOCK_GROUP_RAID1 |
8483                        BTRFS_BLOCK_GROUP_DUP)))
8484                         continue;
8485                 /*
8486                  * avoid allocating from un-mirrored block group if there are
8487                  * mirrored block groups.
8488                  */
8489                 list_for_each_entry(cache, &space_info->block_groups[3], list)
8490                         set_block_group_ro(cache);
8491                 list_for_each_entry(cache, &space_info->block_groups[4], list)
8492                         set_block_group_ro(cache);
8493         }
8494
8495         init_global_block_rsv(info);
8496         ret = 0;
8497 error:
8498         btrfs_free_path(path);
8499         return ret;
8500 }
8501
8502 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
8503                            struct btrfs_root *root, u64 bytes_used,
8504                            u64 type, u64 chunk_objectid, u64 chunk_offset,
8505                            u64 size)
8506 {
8507         int ret;
8508         struct btrfs_root *extent_root;
8509         struct btrfs_block_group_cache *cache;
8510
8511         extent_root = root->fs_info->extent_root;
8512
8513         root->fs_info->last_trans_log_full_commit = trans->transid;
8514
8515         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8516         if (!cache)
8517                 return -ENOMEM;
8518
8519         cache->key.objectid = chunk_offset;
8520         cache->key.offset = size;
8521         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8522         cache->sectorsize = root->sectorsize;
8523         cache->fs_info = root->fs_info;
8524
8525         /*
8526          * we only want to have 32k of ram per block group for keeping track
8527          * of free space, and if we pass 1/2 of that we want to start
8528          * converting things over to using bitmaps
8529          */
8530         cache->extents_thresh = ((1024 * 32) / 2) /
8531                 sizeof(struct btrfs_free_space);
8532         atomic_set(&cache->count, 1);
8533         spin_lock_init(&cache->lock);
8534         spin_lock_init(&cache->tree_lock);
8535         INIT_LIST_HEAD(&cache->list);
8536         INIT_LIST_HEAD(&cache->cluster_list);
8537
8538         btrfs_set_block_group_used(&cache->item, bytes_used);
8539         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
8540         cache->flags = type;
8541         btrfs_set_block_group_flags(&cache->item, type);
8542
8543         cache->last_byte_to_unpin = (u64)-1;
8544         cache->cached = BTRFS_CACHE_FINISHED;
8545         exclude_super_stripes(root, cache);
8546
8547         add_new_free_space(cache, root->fs_info, chunk_offset,
8548                            chunk_offset + size);
8549
8550         free_excluded_extents(root, cache);
8551
8552         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
8553                                 &cache->space_info);
8554         BUG_ON(ret);
8555
8556         spin_lock(&cache->space_info->lock);
8557         cache->space_info->bytes_readonly += cache->bytes_super;
8558         spin_unlock(&cache->space_info->lock);
8559
8560         __link_block_group(cache->space_info, cache);
8561
8562         ret = btrfs_add_block_group_cache(root->fs_info, cache);
8563         BUG_ON(ret);
8564
8565         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
8566                                 sizeof(cache->item));
8567         BUG_ON(ret);
8568
8569         set_avail_alloc_bits(extent_root->fs_info, type);
8570
8571         return 0;
8572 }
8573
8574 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
8575                              struct btrfs_root *root, u64 group_start)
8576 {
8577         struct btrfs_path *path;
8578         struct btrfs_block_group_cache *block_group;
8579         struct btrfs_free_cluster *cluster;
8580         struct btrfs_root *tree_root = root->fs_info->tree_root;
8581         struct btrfs_key key;
8582         struct inode *inode;
8583         int ret;
8584         int factor;
8585
8586         root = root->fs_info->extent_root;
8587
8588         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
8589         BUG_ON(!block_group);
8590         BUG_ON(!block_group->ro);
8591
8592         memcpy(&key, &block_group->key, sizeof(key));
8593         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
8594                                   BTRFS_BLOCK_GROUP_RAID1 |
8595                                   BTRFS_BLOCK_GROUP_RAID10))
8596                 factor = 2;
8597         else
8598                 factor = 1;
8599
8600         /* make sure this block group isn't part of an allocation cluster */
8601         cluster = &root->fs_info->data_alloc_cluster;
8602         spin_lock(&cluster->refill_lock);
8603         btrfs_return_cluster_to_free_space(block_group, cluster);
8604         spin_unlock(&cluster->refill_lock);
8605
8606         /*
8607          * make sure this block group isn't part of a metadata
8608          * allocation cluster
8609          */
8610         cluster = &root->fs_info->meta_alloc_cluster;
8611         spin_lock(&cluster->refill_lock);
8612         btrfs_return_cluster_to_free_space(block_group, cluster);
8613         spin_unlock(&cluster->refill_lock);
8614
8615         path = btrfs_alloc_path();
8616         BUG_ON(!path);
8617
8618         inode = lookup_free_space_inode(root, block_group, path);
8619         if (!IS_ERR(inode)) {
8620                 btrfs_orphan_add(trans, inode);
8621                 clear_nlink(inode);
8622                 /* One for the block groups ref */
8623                 spin_lock(&block_group->lock);
8624                 if (block_group->iref) {
8625                         block_group->iref = 0;
8626                         block_group->inode = NULL;
8627                         spin_unlock(&block_group->lock);
8628                         iput(inode);
8629                 } else {
8630                         spin_unlock(&block_group->lock);
8631                 }
8632                 /* One for our lookup ref */
8633                 iput(inode);
8634         }
8635
8636         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
8637         key.offset = block_group->key.objectid;
8638         key.type = 0;
8639
8640         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
8641         if (ret < 0)
8642                 goto out;
8643         if (ret > 0)
8644                 btrfs_release_path(tree_root, path);
8645         if (ret == 0) {
8646                 ret = btrfs_del_item(trans, tree_root, path);
8647                 if (ret)
8648                         goto out;
8649                 btrfs_release_path(tree_root, path);
8650         }
8651
8652         spin_lock(&root->fs_info->block_group_cache_lock);
8653         rb_erase(&block_group->cache_node,
8654                  &root->fs_info->block_group_cache_tree);
8655         spin_unlock(&root->fs_info->block_group_cache_lock);
8656
8657         down_write(&block_group->space_info->groups_sem);
8658         /*
8659          * we must use list_del_init so people can check to see if they
8660          * are still on the list after taking the semaphore
8661          */
8662         list_del_init(&block_group->list);
8663         up_write(&block_group->space_info->groups_sem);
8664
8665         if (block_group->cached == BTRFS_CACHE_STARTED)
8666                 wait_block_group_cache_done(block_group);
8667
8668         btrfs_remove_free_space_cache(block_group);
8669
8670         spin_lock(&block_group->space_info->lock);
8671         block_group->space_info->total_bytes -= block_group->key.offset;
8672         block_group->space_info->bytes_readonly -= block_group->key.offset;
8673         block_group->space_info->disk_total -= block_group->key.offset * factor;
8674         spin_unlock(&block_group->space_info->lock);
8675
8676         memcpy(&key, &block_group->key, sizeof(key));
8677
8678         btrfs_clear_space_info_full(root->fs_info);
8679
8680         btrfs_put_block_group(block_group);
8681         btrfs_put_block_group(block_group);
8682
8683         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8684         if (ret > 0)
8685                 ret = -EIO;
8686         if (ret < 0)
8687                 goto out;
8688
8689         ret = btrfs_del_item(trans, root, path);
8690 out:
8691         btrfs_free_path(path);
8692         return ret;
8693 }
8694
8695 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8696 {
8697         return unpin_extent_range(root, start, end);
8698 }
8699
8700 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8701                                u64 num_bytes)
8702 {
8703         return btrfs_discard_extent(root, bytenr, num_bytes);
8704 }