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         int pause = 1;
3349         int nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3350
3351         block_rsv = &root->fs_info->delalloc_block_rsv;
3352         space_info = block_rsv->space_info;
3353
3354         smp_mb();
3355         reserved = space_info->bytes_reserved;
3356
3357         if (reserved == 0)
3358                 return 0;
3359
3360         max_reclaim = min(reserved, to_reclaim);
3361
3362         while (1) {
3363                 /* have the flusher threads jump in and do some IO */
3364                 smp_mb();
3365                 nr_pages = min_t(unsigned long, nr_pages,
3366                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3367                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages);
3368
3369                 spin_lock(&space_info->lock);
3370                 if (reserved > space_info->bytes_reserved)
3371                         reclaimed += reserved - space_info->bytes_reserved;
3372                 reserved = space_info->bytes_reserved;
3373                 spin_unlock(&space_info->lock);
3374
3375                 if (reserved == 0 || reclaimed >= max_reclaim)
3376                         break;
3377
3378                 if (trans && trans->transaction->blocked)
3379                         return -EAGAIN;
3380
3381                 __set_current_state(TASK_INTERRUPTIBLE);
3382                 schedule_timeout(pause);
3383                 pause <<= 1;
3384                 if (pause > HZ / 10)
3385                         pause = HZ / 10;
3386
3387         }
3388         return reclaimed >= to_reclaim;
3389 }
3390
3391 /*
3392  * Retries tells us how many times we've called reserve_metadata_bytes.  The
3393  * idea is if this is the first call (retries == 0) then we will add to our
3394  * reserved count if we can't make the allocation in order to hold our place
3395  * while we go and try and free up space.  That way for retries > 1 we don't try
3396  * and add space, we just check to see if the amount of unused space is >= the
3397  * total space, meaning that our reservation is valid.
3398  *
3399  * However if we don't intend to retry this reservation, pass -1 as retries so
3400  * that it short circuits this logic.
3401  */
3402 static int reserve_metadata_bytes(struct btrfs_trans_handle *trans,
3403                                   struct btrfs_root *root,
3404                                   struct btrfs_block_rsv *block_rsv,
3405                                   u64 orig_bytes, int flush)
3406 {
3407         struct btrfs_space_info *space_info = block_rsv->space_info;
3408         u64 unused;
3409         u64 num_bytes = orig_bytes;
3410         int retries = 0;
3411         int ret = 0;
3412         bool reserved = false;
3413         bool committed = false;
3414
3415 again:
3416         ret = -ENOSPC;
3417         if (reserved)
3418                 num_bytes = 0;
3419
3420         spin_lock(&space_info->lock);
3421         unused = space_info->bytes_used + space_info->bytes_reserved +
3422                  space_info->bytes_pinned + space_info->bytes_readonly +
3423                  space_info->bytes_may_use;
3424
3425         /*
3426          * The idea here is that we've not already over-reserved the block group
3427          * then we can go ahead and save our reservation first and then start
3428          * flushing if we need to.  Otherwise if we've already overcommitted
3429          * lets start flushing stuff first and then come back and try to make
3430          * our reservation.
3431          */
3432         if (unused <= space_info->total_bytes) {
3433                 unused = space_info->total_bytes - unused;
3434                 if (unused >= num_bytes) {
3435                         if (!reserved)
3436                                 space_info->bytes_reserved += orig_bytes;
3437                         ret = 0;
3438                 } else {
3439                         /*
3440                          * Ok set num_bytes to orig_bytes since we aren't
3441                          * overocmmitted, this way we only try and reclaim what
3442                          * we need.
3443                          */
3444                         num_bytes = orig_bytes;
3445                 }
3446         } else {
3447                 /*
3448                  * Ok we're over committed, set num_bytes to the overcommitted
3449                  * amount plus the amount of bytes that we need for this
3450                  * reservation.
3451                  */
3452                 num_bytes = unused - space_info->total_bytes +
3453                         (orig_bytes * (retries + 1));
3454         }
3455
3456         /*
3457          * Couldn't make our reservation, save our place so while we're trying
3458          * to reclaim space we can actually use it instead of somebody else
3459          * stealing it from us.
3460          */
3461         if (ret && !reserved) {
3462                 space_info->bytes_reserved += orig_bytes;
3463                 reserved = true;
3464         }
3465
3466         spin_unlock(&space_info->lock);
3467
3468         if (!ret)
3469                 return 0;
3470
3471         if (!flush)
3472                 goto out;
3473
3474         /*
3475          * We do synchronous shrinking since we don't actually unreserve
3476          * metadata until after the IO is completed.
3477          */
3478         ret = shrink_delalloc(trans, root, num_bytes, 1);
3479         if (ret > 0)
3480                 return 0;
3481         else if (ret < 0)
3482                 goto out;
3483
3484         /*
3485          * So if we were overcommitted it's possible that somebody else flushed
3486          * out enough space and we simply didn't have enough space to reclaim,
3487          * so go back around and try again.
3488          */
3489         if (retries < 2) {
3490                 retries++;
3491                 goto again;
3492         }
3493
3494         spin_lock(&space_info->lock);
3495         /*
3496          * Not enough space to be reclaimed, don't bother committing the
3497          * transaction.
3498          */
3499         if (space_info->bytes_pinned < orig_bytes)
3500                 ret = -ENOSPC;
3501         spin_unlock(&space_info->lock);
3502         if (ret)
3503                 goto out;
3504
3505         ret = -EAGAIN;
3506         if (trans || committed)
3507                 goto out;
3508
3509         ret = -ENOSPC;
3510         trans = btrfs_join_transaction(root, 1);
3511         if (IS_ERR(trans))
3512                 goto out;
3513         ret = btrfs_commit_transaction(trans, root);
3514         if (!ret) {
3515                 trans = NULL;
3516                 committed = true;
3517                 goto again;
3518         }
3519
3520 out:
3521         if (reserved) {
3522                 spin_lock(&space_info->lock);
3523                 space_info->bytes_reserved -= orig_bytes;
3524                 spin_unlock(&space_info->lock);
3525         }
3526
3527         return ret;
3528 }
3529
3530 static struct btrfs_block_rsv *get_block_rsv(struct btrfs_trans_handle *trans,
3531                                              struct btrfs_root *root)
3532 {
3533         struct btrfs_block_rsv *block_rsv;
3534         if (root->ref_cows)
3535                 block_rsv = trans->block_rsv;
3536         else
3537                 block_rsv = root->block_rsv;
3538
3539         if (!block_rsv)
3540                 block_rsv = &root->fs_info->empty_block_rsv;
3541
3542         return block_rsv;
3543 }
3544
3545 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3546                                u64 num_bytes)
3547 {
3548         int ret = -ENOSPC;
3549         spin_lock(&block_rsv->lock);
3550         if (block_rsv->reserved >= num_bytes) {
3551                 block_rsv->reserved -= num_bytes;
3552                 if (block_rsv->reserved < block_rsv->size)
3553                         block_rsv->full = 0;
3554                 ret = 0;
3555         }
3556         spin_unlock(&block_rsv->lock);
3557         return ret;
3558 }
3559
3560 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3561                                 u64 num_bytes, int update_size)
3562 {
3563         spin_lock(&block_rsv->lock);
3564         block_rsv->reserved += num_bytes;
3565         if (update_size)
3566                 block_rsv->size += num_bytes;
3567         else if (block_rsv->reserved >= block_rsv->size)
3568                 block_rsv->full = 1;
3569         spin_unlock(&block_rsv->lock);
3570 }
3571
3572 void block_rsv_release_bytes(struct btrfs_block_rsv *block_rsv,
3573                              struct btrfs_block_rsv *dest, u64 num_bytes)
3574 {
3575         struct btrfs_space_info *space_info = block_rsv->space_info;
3576
3577         spin_lock(&block_rsv->lock);
3578         if (num_bytes == (u64)-1)
3579                 num_bytes = block_rsv->size;
3580         block_rsv->size -= num_bytes;
3581         if (block_rsv->reserved >= block_rsv->size) {
3582                 num_bytes = block_rsv->reserved - block_rsv->size;
3583                 block_rsv->reserved = block_rsv->size;
3584                 block_rsv->full = 1;
3585         } else {
3586                 num_bytes = 0;
3587         }
3588         spin_unlock(&block_rsv->lock);
3589
3590         if (num_bytes > 0) {
3591                 if (dest) {
3592                         spin_lock(&dest->lock);
3593                         if (!dest->full) {
3594                                 u64 bytes_to_add;
3595
3596                                 bytes_to_add = dest->size - dest->reserved;
3597                                 bytes_to_add = min(num_bytes, bytes_to_add);
3598                                 dest->reserved += bytes_to_add;
3599                                 if (dest->reserved >= dest->size)
3600                                         dest->full = 1;
3601                                 num_bytes -= bytes_to_add;
3602                         }
3603                         spin_unlock(&dest->lock);
3604                 }
3605                 if (num_bytes) {
3606                         spin_lock(&space_info->lock);
3607                         space_info->bytes_reserved -= num_bytes;
3608                         spin_unlock(&space_info->lock);
3609                 }
3610         }
3611 }
3612
3613 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3614                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3615 {
3616         int ret;
3617
3618         ret = block_rsv_use_bytes(src, num_bytes);
3619         if (ret)
3620                 return ret;
3621
3622         block_rsv_add_bytes(dst, num_bytes, 1);
3623         return 0;
3624 }
3625
3626 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3627 {
3628         memset(rsv, 0, sizeof(*rsv));
3629         spin_lock_init(&rsv->lock);
3630         atomic_set(&rsv->usage, 1);
3631         rsv->priority = 6;
3632         INIT_LIST_HEAD(&rsv->list);
3633 }
3634
3635 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3636 {
3637         struct btrfs_block_rsv *block_rsv;
3638         struct btrfs_fs_info *fs_info = root->fs_info;
3639
3640         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3641         if (!block_rsv)
3642                 return NULL;
3643
3644         btrfs_init_block_rsv(block_rsv);
3645         block_rsv->space_info = __find_space_info(fs_info,
3646                                                   BTRFS_BLOCK_GROUP_METADATA);
3647         return block_rsv;
3648 }
3649
3650 void btrfs_free_block_rsv(struct btrfs_root *root,
3651                           struct btrfs_block_rsv *rsv)
3652 {
3653         if (rsv && atomic_dec_and_test(&rsv->usage)) {
3654                 btrfs_block_rsv_release(root, rsv, (u64)-1);
3655                 if (!rsv->durable)
3656                         kfree(rsv);
3657         }
3658 }
3659
3660 /*
3661  * make the block_rsv struct be able to capture freed space.
3662  * the captured space will re-add to the the block_rsv struct
3663  * after transaction commit
3664  */
3665 void btrfs_add_durable_block_rsv(struct btrfs_fs_info *fs_info,
3666                                  struct btrfs_block_rsv *block_rsv)
3667 {
3668         block_rsv->durable = 1;
3669         mutex_lock(&fs_info->durable_block_rsv_mutex);
3670         list_add_tail(&block_rsv->list, &fs_info->durable_block_rsv_list);
3671         mutex_unlock(&fs_info->durable_block_rsv_mutex);
3672 }
3673
3674 int btrfs_block_rsv_add(struct btrfs_trans_handle *trans,
3675                         struct btrfs_root *root,
3676                         struct btrfs_block_rsv *block_rsv,
3677                         u64 num_bytes)
3678 {
3679         int ret;
3680
3681         if (num_bytes == 0)
3682                 return 0;
3683
3684         ret = reserve_metadata_bytes(trans, root, block_rsv, num_bytes, 1);
3685         if (!ret) {
3686                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3687                 return 0;
3688         }
3689
3690         return ret;
3691 }
3692
3693 int btrfs_block_rsv_check(struct btrfs_trans_handle *trans,
3694                           struct btrfs_root *root,
3695                           struct btrfs_block_rsv *block_rsv,
3696                           u64 min_reserved, int min_factor)
3697 {
3698         u64 num_bytes = 0;
3699         int commit_trans = 0;
3700         int ret = -ENOSPC;
3701
3702         if (!block_rsv)
3703                 return 0;
3704
3705         spin_lock(&block_rsv->lock);
3706         if (min_factor > 0)
3707                 num_bytes = div_factor(block_rsv->size, min_factor);
3708         if (min_reserved > num_bytes)
3709                 num_bytes = min_reserved;
3710
3711         if (block_rsv->reserved >= num_bytes) {
3712                 ret = 0;
3713         } else {
3714                 num_bytes -= block_rsv->reserved;
3715                 if (block_rsv->durable &&
3716                     block_rsv->freed[0] + block_rsv->freed[1] >= num_bytes)
3717                         commit_trans = 1;
3718         }
3719         spin_unlock(&block_rsv->lock);
3720         if (!ret)
3721                 return 0;
3722
3723         if (block_rsv->refill_used) {
3724                 ret = reserve_metadata_bytes(trans, root, block_rsv,
3725                                              num_bytes, 0);
3726                 if (!ret) {
3727                         block_rsv_add_bytes(block_rsv, num_bytes, 0);
3728                         return 0;
3729                 }
3730         }
3731
3732         if (commit_trans) {
3733                 if (trans)
3734                         return -EAGAIN;
3735
3736                 trans = btrfs_join_transaction(root, 1);
3737                 BUG_ON(IS_ERR(trans));
3738                 ret = btrfs_commit_transaction(trans, root);
3739                 return 0;
3740         }
3741
3742         return -ENOSPC;
3743 }
3744
3745 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
3746                             struct btrfs_block_rsv *dst_rsv,
3747                             u64 num_bytes)
3748 {
3749         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3750 }
3751
3752 void btrfs_block_rsv_release(struct btrfs_root *root,
3753                              struct btrfs_block_rsv *block_rsv,
3754                              u64 num_bytes)
3755 {
3756         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3757         if (global_rsv->full || global_rsv == block_rsv ||
3758             block_rsv->space_info != global_rsv->space_info)
3759                 global_rsv = NULL;
3760         block_rsv_release_bytes(block_rsv, global_rsv, num_bytes);
3761 }
3762
3763 /*
3764  * helper to calculate size of global block reservation.
3765  * the desired value is sum of space used by extent tree,
3766  * checksum tree and root tree
3767  */
3768 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
3769 {
3770         struct btrfs_space_info *sinfo;
3771         u64 num_bytes;
3772         u64 meta_used;
3773         u64 data_used;
3774         int csum_size = btrfs_super_csum_size(&fs_info->super_copy);
3775 #if 0
3776         /*
3777          * per tree used space accounting can be inaccuracy, so we
3778          * can't rely on it.
3779          */
3780         spin_lock(&fs_info->extent_root->accounting_lock);
3781         num_bytes = btrfs_root_used(&fs_info->extent_root->root_item);
3782         spin_unlock(&fs_info->extent_root->accounting_lock);
3783
3784         spin_lock(&fs_info->csum_root->accounting_lock);
3785         num_bytes += btrfs_root_used(&fs_info->csum_root->root_item);
3786         spin_unlock(&fs_info->csum_root->accounting_lock);
3787
3788         spin_lock(&fs_info->tree_root->accounting_lock);
3789         num_bytes += btrfs_root_used(&fs_info->tree_root->root_item);
3790         spin_unlock(&fs_info->tree_root->accounting_lock);
3791 #endif
3792         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
3793         spin_lock(&sinfo->lock);
3794         data_used = sinfo->bytes_used;
3795         spin_unlock(&sinfo->lock);
3796
3797         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3798         spin_lock(&sinfo->lock);
3799         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
3800                 data_used = 0;
3801         meta_used = sinfo->bytes_used;
3802         spin_unlock(&sinfo->lock);
3803
3804         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
3805                     csum_size * 2;
3806         num_bytes += div64_u64(data_used + meta_used, 50);
3807
3808         if (num_bytes * 3 > meta_used)
3809                 num_bytes = div64_u64(meta_used, 3);
3810
3811         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
3812 }
3813
3814 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
3815 {
3816         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
3817         struct btrfs_space_info *sinfo = block_rsv->space_info;
3818         u64 num_bytes;
3819
3820         num_bytes = calc_global_metadata_size(fs_info);
3821
3822         spin_lock(&block_rsv->lock);
3823         spin_lock(&sinfo->lock);
3824
3825         block_rsv->size = num_bytes;
3826
3827         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
3828                     sinfo->bytes_reserved + sinfo->bytes_readonly +
3829                     sinfo->bytes_may_use;
3830
3831         if (sinfo->total_bytes > num_bytes) {
3832                 num_bytes = sinfo->total_bytes - num_bytes;
3833                 block_rsv->reserved += num_bytes;
3834                 sinfo->bytes_reserved += num_bytes;
3835         }
3836
3837         if (block_rsv->reserved >= block_rsv->size) {
3838                 num_bytes = block_rsv->reserved - block_rsv->size;
3839                 sinfo->bytes_reserved -= num_bytes;
3840                 block_rsv->reserved = block_rsv->size;
3841                 block_rsv->full = 1;
3842         }
3843 #if 0
3844         printk(KERN_INFO"global block rsv size %llu reserved %llu\n",
3845                 block_rsv->size, block_rsv->reserved);
3846 #endif
3847         spin_unlock(&sinfo->lock);
3848         spin_unlock(&block_rsv->lock);
3849 }
3850
3851 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
3852 {
3853         struct btrfs_space_info *space_info;
3854
3855         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3856         fs_info->chunk_block_rsv.space_info = space_info;
3857         fs_info->chunk_block_rsv.priority = 10;
3858
3859         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3860         fs_info->global_block_rsv.space_info = space_info;
3861         fs_info->global_block_rsv.priority = 10;
3862         fs_info->global_block_rsv.refill_used = 1;
3863         fs_info->delalloc_block_rsv.space_info = space_info;
3864         fs_info->trans_block_rsv.space_info = space_info;
3865         fs_info->empty_block_rsv.space_info = space_info;
3866         fs_info->empty_block_rsv.priority = 10;
3867
3868         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
3869         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
3870         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
3871         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
3872         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
3873
3874         btrfs_add_durable_block_rsv(fs_info, &fs_info->global_block_rsv);
3875
3876         btrfs_add_durable_block_rsv(fs_info, &fs_info->delalloc_block_rsv);
3877
3878         update_global_block_rsv(fs_info);
3879 }
3880
3881 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
3882 {
3883         block_rsv_release_bytes(&fs_info->global_block_rsv, NULL, (u64)-1);
3884         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
3885         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
3886         WARN_ON(fs_info->trans_block_rsv.size > 0);
3887         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
3888         WARN_ON(fs_info->chunk_block_rsv.size > 0);
3889         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
3890 }
3891
3892 static u64 calc_trans_metadata_size(struct btrfs_root *root, int num_items)
3893 {
3894         return (root->leafsize + root->nodesize * (BTRFS_MAX_LEVEL - 1)) *
3895                 3 * num_items;
3896 }
3897
3898 int btrfs_trans_reserve_metadata(struct btrfs_trans_handle *trans,
3899                                  struct btrfs_root *root,
3900                                  int num_items)
3901 {
3902         u64 num_bytes;
3903         int ret;
3904
3905         if (num_items == 0 || root->fs_info->chunk_root == root)
3906                 return 0;
3907
3908         num_bytes = calc_trans_metadata_size(root, num_items);
3909         ret = btrfs_block_rsv_add(trans, root, &root->fs_info->trans_block_rsv,
3910                                   num_bytes);
3911         if (!ret) {
3912                 trans->bytes_reserved += num_bytes;
3913                 trans->block_rsv = &root->fs_info->trans_block_rsv;
3914         }
3915         return ret;
3916 }
3917
3918 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
3919                                   struct btrfs_root *root)
3920 {
3921         if (!trans->bytes_reserved)
3922                 return;
3923
3924         BUG_ON(trans->block_rsv != &root->fs_info->trans_block_rsv);
3925         btrfs_block_rsv_release(root, trans->block_rsv,
3926                                 trans->bytes_reserved);
3927         trans->bytes_reserved = 0;
3928 }
3929
3930 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
3931                                   struct inode *inode)
3932 {
3933         struct btrfs_root *root = BTRFS_I(inode)->root;
3934         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
3935         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
3936
3937         /*
3938          * one for deleting orphan item, one for updating inode and
3939          * two for calling btrfs_truncate_inode_items.
3940          *
3941          * btrfs_truncate_inode_items is a delete operation, it frees
3942          * more space than it uses in most cases. So two units of
3943          * metadata space should be enough for calling it many times.
3944          * If all of the metadata space is used, we can commit
3945          * transaction and use space it freed.
3946          */
3947         u64 num_bytes = calc_trans_metadata_size(root, 4);
3948         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3949 }
3950
3951 void btrfs_orphan_release_metadata(struct inode *inode)
3952 {
3953         struct btrfs_root *root = BTRFS_I(inode)->root;
3954         u64 num_bytes = calc_trans_metadata_size(root, 4);
3955         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
3956 }
3957
3958 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
3959                                 struct btrfs_pending_snapshot *pending)
3960 {
3961         struct btrfs_root *root = pending->root;
3962         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
3963         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
3964         /*
3965          * two for root back/forward refs, two for directory entries
3966          * and one for root of the snapshot.
3967          */
3968         u64 num_bytes = calc_trans_metadata_size(root, 5);
3969         dst_rsv->space_info = src_rsv->space_info;
3970         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3971 }
3972
3973 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes)
3974 {
3975         return num_bytes >>= 3;
3976 }
3977
3978 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
3979 {
3980         struct btrfs_root *root = BTRFS_I(inode)->root;
3981         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
3982         u64 to_reserve;
3983         int nr_extents;
3984         int ret;
3985
3986         if (btrfs_transaction_in_commit(root->fs_info))
3987                 schedule_timeout(1);
3988
3989         num_bytes = ALIGN(num_bytes, root->sectorsize);
3990
3991         spin_lock(&BTRFS_I(inode)->accounting_lock);
3992         nr_extents = atomic_read(&BTRFS_I(inode)->outstanding_extents) + 1;
3993         if (nr_extents > BTRFS_I(inode)->reserved_extents) {
3994                 nr_extents -= BTRFS_I(inode)->reserved_extents;
3995                 to_reserve = calc_trans_metadata_size(root, nr_extents);
3996         } else {
3997                 nr_extents = 0;
3998                 to_reserve = 0;
3999         }
4000         spin_unlock(&BTRFS_I(inode)->accounting_lock);
4001
4002         to_reserve += calc_csum_metadata_size(inode, num_bytes);
4003         ret = reserve_metadata_bytes(NULL, root, block_rsv, to_reserve, 1);
4004         if (ret)
4005                 return ret;
4006
4007         spin_lock(&BTRFS_I(inode)->accounting_lock);
4008         BTRFS_I(inode)->reserved_extents += nr_extents;
4009         atomic_inc(&BTRFS_I(inode)->outstanding_extents);
4010         spin_unlock(&BTRFS_I(inode)->accounting_lock);
4011
4012         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4013
4014         if (block_rsv->size > 512 * 1024 * 1024)
4015                 shrink_delalloc(NULL, root, to_reserve, 0);
4016
4017         return 0;
4018 }
4019
4020 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4021 {
4022         struct btrfs_root *root = BTRFS_I(inode)->root;
4023         u64 to_free;
4024         int nr_extents;
4025
4026         num_bytes = ALIGN(num_bytes, root->sectorsize);
4027         atomic_dec(&BTRFS_I(inode)->outstanding_extents);
4028
4029         spin_lock(&BTRFS_I(inode)->accounting_lock);
4030         nr_extents = atomic_read(&BTRFS_I(inode)->outstanding_extents);
4031         if (nr_extents < BTRFS_I(inode)->reserved_extents) {
4032                 nr_extents = BTRFS_I(inode)->reserved_extents - nr_extents;
4033                 BTRFS_I(inode)->reserved_extents -= nr_extents;
4034         } else {
4035                 nr_extents = 0;
4036         }
4037         spin_unlock(&BTRFS_I(inode)->accounting_lock);
4038
4039         to_free = calc_csum_metadata_size(inode, num_bytes);
4040         if (nr_extents > 0)
4041                 to_free += calc_trans_metadata_size(root, nr_extents);
4042
4043         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4044                                 to_free);
4045 }
4046
4047 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4048 {
4049         int ret;
4050
4051         ret = btrfs_check_data_free_space(inode, num_bytes);
4052         if (ret)
4053                 return ret;
4054
4055         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4056         if (ret) {
4057                 btrfs_free_reserved_data_space(inode, num_bytes);
4058                 return ret;
4059         }
4060
4061         return 0;
4062 }
4063
4064 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4065 {
4066         btrfs_delalloc_release_metadata(inode, num_bytes);
4067         btrfs_free_reserved_data_space(inode, num_bytes);
4068 }
4069
4070 static int update_block_group(struct btrfs_trans_handle *trans,
4071                               struct btrfs_root *root,
4072                               u64 bytenr, u64 num_bytes, int alloc)
4073 {
4074         struct btrfs_block_group_cache *cache = NULL;
4075         struct btrfs_fs_info *info = root->fs_info;
4076         u64 total = num_bytes;
4077         u64 old_val;
4078         u64 byte_in_group;
4079         int factor;
4080
4081         /* block accounting for super block */
4082         spin_lock(&info->delalloc_lock);
4083         old_val = btrfs_super_bytes_used(&info->super_copy);
4084         if (alloc)
4085                 old_val += num_bytes;
4086         else
4087                 old_val -= num_bytes;
4088         btrfs_set_super_bytes_used(&info->super_copy, old_val);
4089         spin_unlock(&info->delalloc_lock);
4090
4091         while (total) {
4092                 cache = btrfs_lookup_block_group(info, bytenr);
4093                 if (!cache)
4094                         return -1;
4095                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4096                                     BTRFS_BLOCK_GROUP_RAID1 |
4097                                     BTRFS_BLOCK_GROUP_RAID10))
4098                         factor = 2;
4099                 else
4100                         factor = 1;
4101                 /*
4102                  * If this block group has free space cache written out, we
4103                  * need to make sure to load it if we are removing space.  This
4104                  * is because we need the unpinning stage to actually add the
4105                  * space back to the block group, otherwise we will leak space.
4106                  */
4107                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4108                         cache_block_group(cache, trans, NULL, 1);
4109
4110                 byte_in_group = bytenr - cache->key.objectid;
4111                 WARN_ON(byte_in_group > cache->key.offset);
4112
4113                 spin_lock(&cache->space_info->lock);
4114                 spin_lock(&cache->lock);
4115
4116                 if (btrfs_super_cache_generation(&info->super_copy) != 0 &&
4117                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4118                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4119
4120                 cache->dirty = 1;
4121                 old_val = btrfs_block_group_used(&cache->item);
4122                 num_bytes = min(total, cache->key.offset - byte_in_group);
4123                 if (alloc) {
4124                         old_val += num_bytes;
4125                         btrfs_set_block_group_used(&cache->item, old_val);
4126                         cache->reserved -= num_bytes;
4127                         cache->space_info->bytes_reserved -= num_bytes;
4128                         cache->space_info->bytes_used += num_bytes;
4129                         cache->space_info->disk_used += num_bytes * factor;
4130                         spin_unlock(&cache->lock);
4131                         spin_unlock(&cache->space_info->lock);
4132                 } else {
4133                         old_val -= num_bytes;
4134                         btrfs_set_block_group_used(&cache->item, old_val);
4135                         cache->pinned += num_bytes;
4136                         cache->space_info->bytes_pinned += num_bytes;
4137                         cache->space_info->bytes_used -= num_bytes;
4138                         cache->space_info->disk_used -= num_bytes * factor;
4139                         spin_unlock(&cache->lock);
4140                         spin_unlock(&cache->space_info->lock);
4141
4142                         set_extent_dirty(info->pinned_extents,
4143                                          bytenr, bytenr + num_bytes - 1,
4144                                          GFP_NOFS | __GFP_NOFAIL);
4145                 }
4146                 btrfs_put_block_group(cache);
4147                 total -= num_bytes;
4148                 bytenr += num_bytes;
4149         }
4150         return 0;
4151 }
4152
4153 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4154 {
4155         struct btrfs_block_group_cache *cache;
4156         u64 bytenr;
4157
4158         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4159         if (!cache)
4160                 return 0;
4161
4162         bytenr = cache->key.objectid;
4163         btrfs_put_block_group(cache);
4164
4165         return bytenr;
4166 }
4167
4168 static int pin_down_extent(struct btrfs_root *root,
4169                            struct btrfs_block_group_cache *cache,
4170                            u64 bytenr, u64 num_bytes, int reserved)
4171 {
4172         spin_lock(&cache->space_info->lock);
4173         spin_lock(&cache->lock);
4174         cache->pinned += num_bytes;
4175         cache->space_info->bytes_pinned += num_bytes;
4176         if (reserved) {
4177                 cache->reserved -= num_bytes;
4178                 cache->space_info->bytes_reserved -= num_bytes;
4179         }
4180         spin_unlock(&cache->lock);
4181         spin_unlock(&cache->space_info->lock);
4182
4183         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4184                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4185         return 0;
4186 }
4187
4188 /*
4189  * this function must be called within transaction
4190  */
4191 int btrfs_pin_extent(struct btrfs_root *root,
4192                      u64 bytenr, u64 num_bytes, int reserved)
4193 {
4194         struct btrfs_block_group_cache *cache;
4195
4196         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4197         BUG_ON(!cache);
4198
4199         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4200
4201         btrfs_put_block_group(cache);
4202         return 0;
4203 }
4204
4205 /*
4206  * update size of reserved extents. this function may return -EAGAIN
4207  * if 'reserve' is true or 'sinfo' is false.
4208  */
4209 static int update_reserved_bytes(struct btrfs_block_group_cache *cache,
4210                                  u64 num_bytes, int reserve, int sinfo)
4211 {
4212         int ret = 0;
4213         if (sinfo) {
4214                 struct btrfs_space_info *space_info = cache->space_info;
4215                 spin_lock(&space_info->lock);
4216                 spin_lock(&cache->lock);
4217                 if (reserve) {
4218                         if (cache->ro) {
4219                                 ret = -EAGAIN;
4220                         } else {
4221                                 cache->reserved += num_bytes;
4222                                 space_info->bytes_reserved += num_bytes;
4223                         }
4224                 } else {
4225                         if (cache->ro)
4226                                 space_info->bytes_readonly += num_bytes;
4227                         cache->reserved -= num_bytes;
4228                         space_info->bytes_reserved -= num_bytes;
4229                 }
4230                 spin_unlock(&cache->lock);
4231                 spin_unlock(&space_info->lock);
4232         } else {
4233                 spin_lock(&cache->lock);
4234                 if (cache->ro) {
4235                         ret = -EAGAIN;
4236                 } else {
4237                         if (reserve)
4238                                 cache->reserved += num_bytes;
4239                         else
4240                                 cache->reserved -= num_bytes;
4241                 }
4242                 spin_unlock(&cache->lock);
4243         }
4244         return ret;
4245 }
4246
4247 int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4248                                 struct btrfs_root *root)
4249 {
4250         struct btrfs_fs_info *fs_info = root->fs_info;
4251         struct btrfs_caching_control *next;
4252         struct btrfs_caching_control *caching_ctl;
4253         struct btrfs_block_group_cache *cache;
4254
4255         down_write(&fs_info->extent_commit_sem);
4256
4257         list_for_each_entry_safe(caching_ctl, next,
4258                                  &fs_info->caching_block_groups, list) {
4259                 cache = caching_ctl->block_group;
4260                 if (block_group_cache_done(cache)) {
4261                         cache->last_byte_to_unpin = (u64)-1;
4262                         list_del_init(&caching_ctl->list);
4263                         put_caching_control(caching_ctl);
4264                 } else {
4265                         cache->last_byte_to_unpin = caching_ctl->progress;
4266                 }
4267         }
4268
4269         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4270                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4271         else
4272                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4273
4274         up_write(&fs_info->extent_commit_sem);
4275
4276         update_global_block_rsv(fs_info);
4277         return 0;
4278 }
4279
4280 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4281 {
4282         struct btrfs_fs_info *fs_info = root->fs_info;
4283         struct btrfs_block_group_cache *cache = NULL;
4284         u64 len;
4285
4286         while (start <= end) {
4287                 if (!cache ||
4288                     start >= cache->key.objectid + cache->key.offset) {
4289                         if (cache)
4290                                 btrfs_put_block_group(cache);
4291                         cache = btrfs_lookup_block_group(fs_info, start);
4292                         BUG_ON(!cache);
4293                 }
4294
4295                 len = cache->key.objectid + cache->key.offset - start;
4296                 len = min(len, end + 1 - start);
4297
4298                 if (start < cache->last_byte_to_unpin) {
4299                         len = min(len, cache->last_byte_to_unpin - start);
4300                         btrfs_add_free_space(cache, start, len);
4301                 }
4302
4303                 start += len;
4304
4305                 spin_lock(&cache->space_info->lock);
4306                 spin_lock(&cache->lock);
4307                 cache->pinned -= len;
4308                 cache->space_info->bytes_pinned -= len;
4309                 if (cache->ro) {
4310                         cache->space_info->bytes_readonly += len;
4311                 } else if (cache->reserved_pinned > 0) {
4312                         len = min(len, cache->reserved_pinned);
4313                         cache->reserved_pinned -= len;
4314                         cache->space_info->bytes_reserved += len;
4315                 }
4316                 spin_unlock(&cache->lock);
4317                 spin_unlock(&cache->space_info->lock);
4318         }
4319
4320         if (cache)
4321                 btrfs_put_block_group(cache);
4322         return 0;
4323 }
4324
4325 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4326                                struct btrfs_root *root)
4327 {
4328         struct btrfs_fs_info *fs_info = root->fs_info;
4329         struct extent_io_tree *unpin;
4330         struct btrfs_block_rsv *block_rsv;
4331         struct btrfs_block_rsv *next_rsv;
4332         u64 start;
4333         u64 end;
4334         int idx;
4335         int ret;
4336
4337         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4338                 unpin = &fs_info->freed_extents[1];
4339         else
4340                 unpin = &fs_info->freed_extents[0];
4341
4342         while (1) {
4343                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4344                                             EXTENT_DIRTY);
4345                 if (ret)
4346                         break;
4347
4348                 ret = btrfs_discard_extent(root, start, end + 1 - start);
4349
4350                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4351                 unpin_extent_range(root, start, end);
4352                 cond_resched();
4353         }
4354
4355         mutex_lock(&fs_info->durable_block_rsv_mutex);
4356         list_for_each_entry_safe(block_rsv, next_rsv,
4357                                  &fs_info->durable_block_rsv_list, list) {
4358
4359                 idx = trans->transid & 0x1;
4360                 if (block_rsv->freed[idx] > 0) {
4361                         block_rsv_add_bytes(block_rsv,
4362                                             block_rsv->freed[idx], 0);
4363                         block_rsv->freed[idx] = 0;
4364                 }
4365                 if (atomic_read(&block_rsv->usage) == 0) {
4366                         btrfs_block_rsv_release(root, block_rsv, (u64)-1);
4367
4368                         if (block_rsv->freed[0] == 0 &&
4369                             block_rsv->freed[1] == 0) {
4370                                 list_del_init(&block_rsv->list);
4371                                 kfree(block_rsv);
4372                         }
4373                 } else {
4374                         btrfs_block_rsv_release(root, block_rsv, 0);
4375                 }
4376         }
4377         mutex_unlock(&fs_info->durable_block_rsv_mutex);
4378
4379         return 0;
4380 }
4381
4382 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4383                                 struct btrfs_root *root,
4384                                 u64 bytenr, u64 num_bytes, u64 parent,
4385                                 u64 root_objectid, u64 owner_objectid,
4386                                 u64 owner_offset, int refs_to_drop,
4387                                 struct btrfs_delayed_extent_op *extent_op)
4388 {
4389         struct btrfs_key key;
4390         struct btrfs_path *path;
4391         struct btrfs_fs_info *info = root->fs_info;
4392         struct btrfs_root *extent_root = info->extent_root;
4393         struct extent_buffer *leaf;
4394         struct btrfs_extent_item *ei;
4395         struct btrfs_extent_inline_ref *iref;
4396         int ret;
4397         int is_data;
4398         int extent_slot = 0;
4399         int found_extent = 0;
4400         int num_to_del = 1;
4401         u32 item_size;
4402         u64 refs;
4403
4404         path = btrfs_alloc_path();
4405         if (!path)
4406                 return -ENOMEM;
4407
4408         path->reada = 1;
4409         path->leave_spinning = 1;
4410
4411         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4412         BUG_ON(!is_data && refs_to_drop != 1);
4413
4414         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4415                                     bytenr, num_bytes, parent,
4416                                     root_objectid, owner_objectid,
4417                                     owner_offset);
4418         if (ret == 0) {
4419                 extent_slot = path->slots[0];
4420                 while (extent_slot >= 0) {
4421                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4422                                               extent_slot);
4423                         if (key.objectid != bytenr)
4424                                 break;
4425                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4426                             key.offset == num_bytes) {
4427                                 found_extent = 1;
4428                                 break;
4429                         }
4430                         if (path->slots[0] - extent_slot > 5)
4431                                 break;
4432                         extent_slot--;
4433                 }
4434 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4435                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4436                 if (found_extent && item_size < sizeof(*ei))
4437                         found_extent = 0;
4438 #endif
4439                 if (!found_extent) {
4440                         BUG_ON(iref);
4441                         ret = remove_extent_backref(trans, extent_root, path,
4442                                                     NULL, refs_to_drop,
4443                                                     is_data);
4444                         BUG_ON(ret);
4445                         btrfs_release_path(extent_root, path);
4446                         path->leave_spinning = 1;
4447
4448                         key.objectid = bytenr;
4449                         key.type = BTRFS_EXTENT_ITEM_KEY;
4450                         key.offset = num_bytes;
4451
4452                         ret = btrfs_search_slot(trans, extent_root,
4453                                                 &key, path, -1, 1);
4454                         if (ret) {
4455                                 printk(KERN_ERR "umm, got %d back from search"
4456                                        ", was looking for %llu\n", ret,
4457                                        (unsigned long long)bytenr);
4458                                 btrfs_print_leaf(extent_root, path->nodes[0]);
4459                         }
4460                         BUG_ON(ret);
4461                         extent_slot = path->slots[0];
4462                 }
4463         } else {
4464                 btrfs_print_leaf(extent_root, path->nodes[0]);
4465                 WARN_ON(1);
4466                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4467                        "parent %llu root %llu  owner %llu offset %llu\n",
4468                        (unsigned long long)bytenr,
4469                        (unsigned long long)parent,
4470                        (unsigned long long)root_objectid,
4471                        (unsigned long long)owner_objectid,
4472                        (unsigned long long)owner_offset);
4473         }
4474
4475         leaf = path->nodes[0];
4476         item_size = btrfs_item_size_nr(leaf, extent_slot);
4477 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4478         if (item_size < sizeof(*ei)) {
4479                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4480                 ret = convert_extent_item_v0(trans, extent_root, path,
4481                                              owner_objectid, 0);
4482                 BUG_ON(ret < 0);
4483
4484                 btrfs_release_path(extent_root, path);
4485                 path->leave_spinning = 1;
4486
4487                 key.objectid = bytenr;
4488                 key.type = BTRFS_EXTENT_ITEM_KEY;
4489                 key.offset = num_bytes;
4490
4491                 ret = btrfs_search_slot(trans, extent_root, &key, path,
4492                                         -1, 1);
4493                 if (ret) {
4494                         printk(KERN_ERR "umm, got %d back from search"
4495                                ", was looking for %llu\n", ret,
4496                                (unsigned long long)bytenr);
4497                         btrfs_print_leaf(extent_root, path->nodes[0]);
4498                 }
4499                 BUG_ON(ret);
4500                 extent_slot = path->slots[0];
4501                 leaf = path->nodes[0];
4502                 item_size = btrfs_item_size_nr(leaf, extent_slot);
4503         }
4504 #endif
4505         BUG_ON(item_size < sizeof(*ei));
4506         ei = btrfs_item_ptr(leaf, extent_slot,
4507                             struct btrfs_extent_item);
4508         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
4509                 struct btrfs_tree_block_info *bi;
4510                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4511                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4512                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4513         }
4514
4515         refs = btrfs_extent_refs(leaf, ei);
4516         BUG_ON(refs < refs_to_drop);
4517         refs -= refs_to_drop;
4518
4519         if (refs > 0) {
4520                 if (extent_op)
4521                         __run_delayed_extent_op(extent_op, leaf, ei);
4522                 /*
4523                  * In the case of inline back ref, reference count will
4524                  * be updated by remove_extent_backref
4525                  */
4526                 if (iref) {
4527                         BUG_ON(!found_extent);
4528                 } else {
4529                         btrfs_set_extent_refs(leaf, ei, refs);
4530                         btrfs_mark_buffer_dirty(leaf);
4531                 }
4532                 if (found_extent) {
4533                         ret = remove_extent_backref(trans, extent_root, path,
4534                                                     iref, refs_to_drop,
4535                                                     is_data);
4536                         BUG_ON(ret);
4537                 }
4538         } else {
4539                 if (found_extent) {
4540                         BUG_ON(is_data && refs_to_drop !=
4541                                extent_data_ref_count(root, path, iref));
4542                         if (iref) {
4543                                 BUG_ON(path->slots[0] != extent_slot);
4544                         } else {
4545                                 BUG_ON(path->slots[0] != extent_slot + 1);
4546                                 path->slots[0] = extent_slot;
4547                                 num_to_del = 2;
4548                         }
4549                 }
4550
4551                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
4552                                       num_to_del);
4553                 BUG_ON(ret);
4554                 btrfs_release_path(extent_root, path);
4555
4556                 if (is_data) {
4557                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
4558                         BUG_ON(ret);
4559                 } else {
4560                         invalidate_mapping_pages(info->btree_inode->i_mapping,
4561                              bytenr >> PAGE_CACHE_SHIFT,
4562                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
4563                 }
4564
4565                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
4566                 BUG_ON(ret);
4567         }
4568         btrfs_free_path(path);
4569         return ret;
4570 }
4571
4572 /*
4573  * when we free an block, it is possible (and likely) that we free the last
4574  * delayed ref for that extent as well.  This searches the delayed ref tree for
4575  * a given extent, and if there are no other delayed refs to be processed, it
4576  * removes it from the tree.
4577  */
4578 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
4579                                       struct btrfs_root *root, u64 bytenr)
4580 {
4581         struct btrfs_delayed_ref_head *head;
4582         struct btrfs_delayed_ref_root *delayed_refs;
4583         struct btrfs_delayed_ref_node *ref;
4584         struct rb_node *node;
4585         int ret = 0;
4586
4587         delayed_refs = &trans->transaction->delayed_refs;
4588         spin_lock(&delayed_refs->lock);
4589         head = btrfs_find_delayed_ref_head(trans, bytenr);
4590         if (!head)
4591                 goto out;
4592
4593         node = rb_prev(&head->node.rb_node);
4594         if (!node)
4595                 goto out;
4596
4597         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
4598
4599         /* there are still entries for this ref, we can't drop it */
4600         if (ref->bytenr == bytenr)
4601                 goto out;
4602
4603         if (head->extent_op) {
4604                 if (!head->must_insert_reserved)
4605                         goto out;
4606                 kfree(head->extent_op);
4607                 head->extent_op = NULL;
4608         }
4609
4610         /*
4611          * waiting for the lock here would deadlock.  If someone else has it
4612          * locked they are already in the process of dropping it anyway
4613          */
4614         if (!mutex_trylock(&head->mutex))
4615                 goto out;
4616
4617         /*
4618          * at this point we have a head with no other entries.  Go
4619          * ahead and process it.
4620          */
4621         head->node.in_tree = 0;
4622         rb_erase(&head->node.rb_node, &delayed_refs->root);
4623
4624         delayed_refs->num_entries--;
4625
4626         /*
4627          * we don't take a ref on the node because we're removing it from the
4628          * tree, so we just steal the ref the tree was holding.
4629          */
4630         delayed_refs->num_heads--;
4631         if (list_empty(&head->cluster))
4632                 delayed_refs->num_heads_ready--;
4633
4634         list_del_init(&head->cluster);
4635         spin_unlock(&delayed_refs->lock);
4636
4637         BUG_ON(head->extent_op);
4638         if (head->must_insert_reserved)
4639                 ret = 1;
4640
4641         mutex_unlock(&head->mutex);
4642         btrfs_put_delayed_ref(&head->node);
4643         return ret;
4644 out:
4645         spin_unlock(&delayed_refs->lock);
4646         return 0;
4647 }
4648
4649 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
4650                            struct btrfs_root *root,
4651                            struct extent_buffer *buf,
4652                            u64 parent, int last_ref)
4653 {
4654         struct btrfs_block_rsv *block_rsv;
4655         struct btrfs_block_group_cache *cache = NULL;
4656         int ret;
4657
4658         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4659                 ret = btrfs_add_delayed_tree_ref(trans, buf->start, buf->len,
4660                                                 parent, root->root_key.objectid,
4661                                                 btrfs_header_level(buf),
4662                                                 BTRFS_DROP_DELAYED_REF, NULL);
4663                 BUG_ON(ret);
4664         }
4665
4666         if (!last_ref)
4667                 return;
4668
4669         block_rsv = get_block_rsv(trans, root);
4670         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
4671         if (block_rsv->space_info != cache->space_info)
4672                 goto out;
4673
4674         if (btrfs_header_generation(buf) == trans->transid) {
4675                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4676                         ret = check_ref_cleanup(trans, root, buf->start);
4677                         if (!ret)
4678                                 goto pin;
4679                 }
4680
4681                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
4682                         pin_down_extent(root, cache, buf->start, buf->len, 1);
4683                         goto pin;
4684                 }
4685
4686                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
4687
4688                 btrfs_add_free_space(cache, buf->start, buf->len);
4689                 ret = update_reserved_bytes(cache, buf->len, 0, 0);
4690                 if (ret == -EAGAIN) {
4691                         /* block group became read-only */
4692                         update_reserved_bytes(cache, buf->len, 0, 1);
4693                         goto out;
4694                 }
4695
4696                 ret = 1;
4697                 spin_lock(&block_rsv->lock);
4698                 if (block_rsv->reserved < block_rsv->size) {
4699                         block_rsv->reserved += buf->len;
4700                         ret = 0;
4701                 }
4702                 spin_unlock(&block_rsv->lock);
4703
4704                 if (ret) {
4705                         spin_lock(&cache->space_info->lock);
4706                         cache->space_info->bytes_reserved -= buf->len;
4707                         spin_unlock(&cache->space_info->lock);
4708                 }
4709                 goto out;
4710         }
4711 pin:
4712         if (block_rsv->durable && !cache->ro) {
4713                 ret = 0;
4714                 spin_lock(&cache->lock);
4715                 if (!cache->ro) {
4716                         cache->reserved_pinned += buf->len;
4717                         ret = 1;
4718                 }
4719                 spin_unlock(&cache->lock);
4720
4721                 if (ret) {
4722                         spin_lock(&block_rsv->lock);
4723                         block_rsv->freed[trans->transid & 0x1] += buf->len;
4724                         spin_unlock(&block_rsv->lock);
4725                 }
4726         }
4727 out:
4728         btrfs_put_block_group(cache);
4729 }
4730
4731 int btrfs_free_extent(struct btrfs_trans_handle *trans,
4732                       struct btrfs_root *root,
4733                       u64 bytenr, u64 num_bytes, u64 parent,
4734                       u64 root_objectid, u64 owner, u64 offset)
4735 {
4736         int ret;
4737
4738         /*
4739          * tree log blocks never actually go into the extent allocation
4740          * tree, just update pinning info and exit early.
4741          */
4742         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
4743                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
4744                 /* unlocks the pinned mutex */
4745                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
4746                 ret = 0;
4747         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
4748                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
4749                                         parent, root_objectid, (int)owner,
4750                                         BTRFS_DROP_DELAYED_REF, NULL);
4751                 BUG_ON(ret);
4752         } else {
4753                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
4754                                         parent, root_objectid, owner,
4755                                         offset, BTRFS_DROP_DELAYED_REF, NULL);
4756                 BUG_ON(ret);
4757         }
4758         return ret;
4759 }
4760
4761 static u64 stripe_align(struct btrfs_root *root, u64 val)
4762 {
4763         u64 mask = ((u64)root->stripesize - 1);
4764         u64 ret = (val + mask) & ~mask;
4765         return ret;
4766 }
4767
4768 /*
4769  * when we wait for progress in the block group caching, its because
4770  * our allocation attempt failed at least once.  So, we must sleep
4771  * and let some progress happen before we try again.
4772  *
4773  * This function will sleep at least once waiting for new free space to
4774  * show up, and then it will check the block group free space numbers
4775  * for our min num_bytes.  Another option is to have it go ahead
4776  * and look in the rbtree for a free extent of a given size, but this
4777  * is a good start.
4778  */
4779 static noinline int
4780 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
4781                                 u64 num_bytes)
4782 {
4783         struct btrfs_caching_control *caching_ctl;
4784         DEFINE_WAIT(wait);
4785
4786         caching_ctl = get_caching_control(cache);
4787         if (!caching_ctl)
4788                 return 0;
4789
4790         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
4791                    (cache->free_space >= num_bytes));
4792
4793         put_caching_control(caching_ctl);
4794         return 0;
4795 }
4796
4797 static noinline int
4798 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
4799 {
4800         struct btrfs_caching_control *caching_ctl;
4801         DEFINE_WAIT(wait);
4802
4803         caching_ctl = get_caching_control(cache);
4804         if (!caching_ctl)
4805                 return 0;
4806
4807         wait_event(caching_ctl->wait, block_group_cache_done(cache));
4808
4809         put_caching_control(caching_ctl);
4810         return 0;
4811 }
4812
4813 static int get_block_group_index(struct btrfs_block_group_cache *cache)
4814 {
4815         int index;
4816         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
4817                 index = 0;
4818         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
4819                 index = 1;
4820         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
4821                 index = 2;
4822         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
4823                 index = 3;
4824         else
4825                 index = 4;
4826         return index;
4827 }
4828
4829 enum btrfs_loop_type {
4830         LOOP_FIND_IDEAL = 0,
4831         LOOP_CACHING_NOWAIT = 1,
4832         LOOP_CACHING_WAIT = 2,
4833         LOOP_ALLOC_CHUNK = 3,
4834         LOOP_NO_EMPTY_SIZE = 4,
4835 };
4836
4837 /*
4838  * walks the btree of allocated extents and find a hole of a given size.
4839  * The key ins is changed to record the hole:
4840  * ins->objectid == block start
4841  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4842  * ins->offset == number of blocks
4843  * Any available blocks before search_start are skipped.
4844  */
4845 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
4846                                      struct btrfs_root *orig_root,
4847                                      u64 num_bytes, u64 empty_size,
4848                                      u64 search_start, u64 search_end,
4849                                      u64 hint_byte, struct btrfs_key *ins,
4850                                      int data)
4851 {
4852         int ret = 0;
4853         struct btrfs_root *root = orig_root->fs_info->extent_root;
4854         struct btrfs_free_cluster *last_ptr = NULL;
4855         struct btrfs_block_group_cache *block_group = NULL;
4856         int empty_cluster = 2 * 1024 * 1024;
4857         int allowed_chunk_alloc = 0;
4858         int done_chunk_alloc = 0;
4859         struct btrfs_space_info *space_info;
4860         int last_ptr_loop = 0;
4861         int loop = 0;
4862         int index = 0;
4863         bool found_uncached_bg = false;
4864         bool failed_cluster_refill = false;
4865         bool failed_alloc = false;
4866         bool use_cluster = true;
4867         u64 ideal_cache_percent = 0;
4868         u64 ideal_cache_offset = 0;
4869
4870         WARN_ON(num_bytes < root->sectorsize);
4871         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
4872         ins->objectid = 0;
4873         ins->offset = 0;
4874
4875         space_info = __find_space_info(root->fs_info, data);
4876         if (!space_info) {
4877                 printk(KERN_ERR "No space info for %d\n", data);
4878                 return -ENOSPC;
4879         }
4880
4881         /*
4882          * If the space info is for both data and metadata it means we have a
4883          * small filesystem and we can't use the clustering stuff.
4884          */
4885         if (btrfs_mixed_space_info(space_info))
4886                 use_cluster = false;
4887
4888         if (orig_root->ref_cows || empty_size)
4889                 allowed_chunk_alloc = 1;
4890
4891         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
4892                 last_ptr = &root->fs_info->meta_alloc_cluster;
4893                 if (!btrfs_test_opt(root, SSD))
4894                         empty_cluster = 64 * 1024;
4895         }
4896
4897         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
4898             btrfs_test_opt(root, SSD)) {
4899                 last_ptr = &root->fs_info->data_alloc_cluster;
4900         }
4901
4902         if (last_ptr) {
4903                 spin_lock(&last_ptr->lock);
4904                 if (last_ptr->block_group)
4905                         hint_byte = last_ptr->window_start;
4906                 spin_unlock(&last_ptr->lock);
4907         }
4908
4909         search_start = max(search_start, first_logical_byte(root, 0));
4910         search_start = max(search_start, hint_byte);
4911
4912         if (!last_ptr)
4913                 empty_cluster = 0;
4914
4915         if (search_start == hint_byte) {
4916 ideal_cache:
4917                 block_group = btrfs_lookup_block_group(root->fs_info,
4918                                                        search_start);
4919                 /*
4920                  * we don't want to use the block group if it doesn't match our
4921                  * allocation bits, or if its not cached.
4922                  *
4923                  * However if we are re-searching with an ideal block group
4924                  * picked out then we don't care that the block group is cached.
4925                  */
4926                 if (block_group && block_group_bits(block_group, data) &&
4927                     (block_group->cached != BTRFS_CACHE_NO ||
4928                      search_start == ideal_cache_offset)) {
4929                         down_read(&space_info->groups_sem);
4930                         if (list_empty(&block_group->list) ||
4931                             block_group->ro) {
4932                                 /*
4933                                  * someone is removing this block group,
4934                                  * we can't jump into the have_block_group
4935                                  * target because our list pointers are not
4936                                  * valid
4937                                  */
4938                                 btrfs_put_block_group(block_group);
4939                                 up_read(&space_info->groups_sem);
4940                         } else {
4941                                 index = get_block_group_index(block_group);
4942                                 goto have_block_group;
4943                         }
4944                 } else if (block_group) {
4945                         btrfs_put_block_group(block_group);
4946                 }
4947         }
4948 search:
4949         down_read(&space_info->groups_sem);
4950         list_for_each_entry(block_group, &space_info->block_groups[index],
4951                             list) {
4952                 u64 offset;
4953                 int cached;
4954
4955                 btrfs_get_block_group(block_group);
4956                 search_start = block_group->key.objectid;
4957
4958                 /*
4959                  * this can happen if we end up cycling through all the
4960                  * raid types, but we want to make sure we only allocate
4961                  * for the proper type.
4962                  */
4963                 if (!block_group_bits(block_group, data)) {
4964                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
4965                                 BTRFS_BLOCK_GROUP_RAID1 |
4966                                 BTRFS_BLOCK_GROUP_RAID10;
4967
4968                         /*
4969                          * if they asked for extra copies and this block group
4970                          * doesn't provide them, bail.  This does allow us to
4971                          * fill raid0 from raid1.
4972                          */
4973                         if ((data & extra) && !(block_group->flags & extra))
4974                                 goto loop;
4975                 }
4976
4977 have_block_group:
4978                 if (unlikely(block_group->cached == BTRFS_CACHE_NO)) {
4979                         u64 free_percent;
4980
4981                         ret = cache_block_group(block_group, trans,
4982                                                 orig_root, 1);
4983                         if (block_group->cached == BTRFS_CACHE_FINISHED)
4984                                 goto have_block_group;
4985
4986                         free_percent = btrfs_block_group_used(&block_group->item);
4987                         free_percent *= 100;
4988                         free_percent = div64_u64(free_percent,
4989                                                  block_group->key.offset);
4990                         free_percent = 100 - free_percent;
4991                         if (free_percent > ideal_cache_percent &&
4992                             likely(!block_group->ro)) {
4993                                 ideal_cache_offset = block_group->key.objectid;
4994                                 ideal_cache_percent = free_percent;
4995                         }
4996
4997                         /*
4998                          * We only want to start kthread caching if we are at
4999                          * the point where we will wait for caching to make
5000                          * progress, or if our ideal search is over and we've
5001                          * found somebody to start caching.
5002                          */
5003                         if (loop > LOOP_CACHING_NOWAIT ||
5004                             (loop > LOOP_FIND_IDEAL &&
5005                              atomic_read(&space_info->caching_threads) < 2)) {
5006                                 ret = cache_block_group(block_group, trans,
5007                                                         orig_root, 0);
5008                                 BUG_ON(ret);
5009                         }
5010                         found_uncached_bg = true;
5011
5012                         /*
5013                          * If loop is set for cached only, try the next block
5014                          * group.
5015                          */
5016                         if (loop == LOOP_FIND_IDEAL)
5017                                 goto loop;
5018                 }
5019
5020                 cached = block_group_cache_done(block_group);
5021                 if (unlikely(!cached))
5022                         found_uncached_bg = true;
5023
5024                 if (unlikely(block_group->ro))
5025                         goto loop;
5026
5027                 /*
5028                  * Ok we want to try and use the cluster allocator, so lets look
5029                  * there, unless we are on LOOP_NO_EMPTY_SIZE, since we will
5030                  * have tried the cluster allocator plenty of times at this
5031                  * point and not have found anything, so we are likely way too
5032                  * fragmented for the clustering stuff to find anything, so lets
5033                  * just skip it and let the allocator find whatever block it can
5034                  * find
5035                  */
5036                 if (last_ptr && loop < LOOP_NO_EMPTY_SIZE) {
5037                         /*
5038                          * the refill lock keeps out other
5039                          * people trying to start a new cluster
5040                          */
5041                         spin_lock(&last_ptr->refill_lock);
5042                         if (last_ptr->block_group &&
5043                             (last_ptr->block_group->ro ||
5044                             !block_group_bits(last_ptr->block_group, data))) {
5045                                 offset = 0;
5046                                 goto refill_cluster;
5047                         }
5048
5049                         offset = btrfs_alloc_from_cluster(block_group, last_ptr,
5050                                                  num_bytes, search_start);
5051                         if (offset) {
5052                                 /* we have a block, we're done */
5053                                 spin_unlock(&last_ptr->refill_lock);
5054                                 goto checks;
5055                         }
5056
5057                         spin_lock(&last_ptr->lock);
5058                         /*
5059                          * whoops, this cluster doesn't actually point to
5060                          * this block group.  Get a ref on the block
5061                          * group is does point to and try again
5062                          */
5063                         if (!last_ptr_loop && last_ptr->block_group &&
5064                             last_ptr->block_group != block_group) {
5065
5066                                 btrfs_put_block_group(block_group);
5067                                 block_group = last_ptr->block_group;
5068                                 btrfs_get_block_group(block_group);
5069                                 spin_unlock(&last_ptr->lock);
5070                                 spin_unlock(&last_ptr->refill_lock);
5071
5072                                 last_ptr_loop = 1;
5073                                 search_start = block_group->key.objectid;
5074                                 /*
5075                                  * we know this block group is properly
5076                                  * in the list because
5077                                  * btrfs_remove_block_group, drops the
5078                                  * cluster before it removes the block
5079                                  * group from the list
5080                                  */
5081                                 goto have_block_group;
5082                         }
5083                         spin_unlock(&last_ptr->lock);
5084 refill_cluster:
5085                         /*
5086                          * this cluster didn't work out, free it and
5087                          * start over
5088                          */
5089                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5090
5091                         last_ptr_loop = 0;
5092
5093                         /* allocate a cluster in this block group */
5094                         ret = btrfs_find_space_cluster(trans, root,
5095                                                block_group, last_ptr,
5096                                                offset, num_bytes,
5097                                                empty_cluster + empty_size);
5098                         if (ret == 0) {
5099                                 /*
5100                                  * now pull our allocation out of this
5101                                  * cluster
5102                                  */
5103                                 offset = btrfs_alloc_from_cluster(block_group,
5104                                                   last_ptr, num_bytes,
5105                                                   search_start);
5106                                 if (offset) {
5107                                         /* we found one, proceed */
5108                                         spin_unlock(&last_ptr->refill_lock);
5109                                         goto checks;
5110                                 }
5111                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5112                                    && !failed_cluster_refill) {
5113                                 spin_unlock(&last_ptr->refill_lock);
5114
5115                                 failed_cluster_refill = true;
5116                                 wait_block_group_cache_progress(block_group,
5117                                        num_bytes + empty_cluster + empty_size);
5118                                 goto have_block_group;
5119                         }
5120
5121                         /*
5122                          * at this point we either didn't find a cluster
5123                          * or we weren't able to allocate a block from our
5124                          * cluster.  Free the cluster we've been trying
5125                          * to use, and go to the next block group
5126                          */
5127                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5128                         spin_unlock(&last_ptr->refill_lock);
5129                         goto loop;
5130                 }
5131
5132                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5133                                                     num_bytes, empty_size);
5134                 /*
5135                  * If we didn't find a chunk, and we haven't failed on this
5136                  * block group before, and this block group is in the middle of
5137                  * caching and we are ok with waiting, then go ahead and wait
5138                  * for progress to be made, and set failed_alloc to true.
5139                  *
5140                  * If failed_alloc is true then we've already waited on this
5141                  * block group once and should move on to the next block group.
5142                  */
5143                 if (!offset && !failed_alloc && !cached &&
5144                     loop > LOOP_CACHING_NOWAIT) {
5145                         wait_block_group_cache_progress(block_group,
5146                                                 num_bytes + empty_size);
5147                         failed_alloc = true;
5148                         goto have_block_group;
5149                 } else if (!offset) {
5150                         goto loop;
5151                 }
5152 checks:
5153                 search_start = stripe_align(root, offset);
5154                 /* move on to the next group */
5155                 if (search_start + num_bytes >= search_end) {
5156                         btrfs_add_free_space(block_group, offset, num_bytes);
5157                         goto loop;
5158                 }
5159
5160                 /* move on to the next group */
5161                 if (search_start + num_bytes >
5162                     block_group->key.objectid + block_group->key.offset) {
5163                         btrfs_add_free_space(block_group, offset, num_bytes);
5164                         goto loop;
5165                 }
5166
5167                 ins->objectid = search_start;
5168                 ins->offset = num_bytes;
5169
5170                 if (offset < search_start)
5171                         btrfs_add_free_space(block_group, offset,
5172                                              search_start - offset);
5173                 BUG_ON(offset > search_start);
5174
5175                 ret = update_reserved_bytes(block_group, num_bytes, 1,
5176                                             (data & BTRFS_BLOCK_GROUP_DATA));
5177                 if (ret == -EAGAIN) {
5178                         btrfs_add_free_space(block_group, offset, num_bytes);
5179                         goto loop;
5180                 }
5181
5182                 /* we are all good, lets return */
5183                 ins->objectid = search_start;
5184                 ins->offset = num_bytes;
5185
5186                 if (offset < search_start)
5187                         btrfs_add_free_space(block_group, offset,
5188                                              search_start - offset);
5189                 BUG_ON(offset > search_start);
5190                 break;
5191 loop:
5192                 failed_cluster_refill = false;
5193                 failed_alloc = false;
5194                 BUG_ON(index != get_block_group_index(block_group));
5195                 btrfs_put_block_group(block_group);
5196         }
5197         up_read(&space_info->groups_sem);
5198
5199         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5200                 goto search;
5201
5202         /* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
5203          *                      for them to make caching progress.  Also
5204          *                      determine the best possible bg to cache
5205          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5206          *                      caching kthreads as we move along
5207          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5208          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5209          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5210          *                      again
5211          */
5212         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE &&
5213             (found_uncached_bg || empty_size || empty_cluster ||
5214              allowed_chunk_alloc)) {
5215                 index = 0;
5216                 if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
5217                         found_uncached_bg = false;
5218                         loop++;
5219                         if (!ideal_cache_percent &&
5220                             atomic_read(&space_info->caching_threads))
5221                                 goto search;
5222
5223                         /*
5224                          * 1 of the following 2 things have happened so far
5225                          *
5226                          * 1) We found an ideal block group for caching that
5227                          * is mostly full and will cache quickly, so we might
5228                          * as well wait for it.
5229                          *
5230                          * 2) We searched for cached only and we didn't find
5231                          * anything, and we didn't start any caching kthreads
5232                          * either, so chances are we will loop through and
5233                          * start a couple caching kthreads, and then come back
5234                          * around and just wait for them.  This will be slower
5235                          * because we will have 2 caching kthreads reading at
5236                          * the same time when we could have just started one
5237                          * and waited for it to get far enough to give us an
5238                          * allocation, so go ahead and go to the wait caching
5239                          * loop.
5240                          */
5241                         loop = LOOP_CACHING_WAIT;
5242                         search_start = ideal_cache_offset;
5243                         ideal_cache_percent = 0;
5244                         goto ideal_cache;
5245                 } else if (loop == LOOP_FIND_IDEAL) {
5246                         /*
5247                          * Didn't find a uncached bg, wait on anything we find
5248                          * next.
5249                          */
5250                         loop = LOOP_CACHING_WAIT;
5251                         goto search;
5252                 }
5253
5254                 if (loop < LOOP_CACHING_WAIT) {
5255                         loop++;
5256                         goto search;
5257                 }
5258
5259                 if (loop == LOOP_ALLOC_CHUNK) {
5260                         empty_size = 0;
5261                         empty_cluster = 0;
5262                 }
5263
5264                 if (allowed_chunk_alloc) {
5265                         ret = do_chunk_alloc(trans, root, num_bytes +
5266                                              2 * 1024 * 1024, data, 1);
5267                         allowed_chunk_alloc = 0;
5268                         done_chunk_alloc = 1;
5269                 } else if (!done_chunk_alloc) {
5270                         space_info->force_alloc = 1;
5271                 }
5272
5273                 if (loop < LOOP_NO_EMPTY_SIZE) {
5274                         loop++;
5275                         goto search;
5276                 }
5277                 ret = -ENOSPC;
5278         } else if (!ins->objectid) {
5279                 ret = -ENOSPC;
5280         }
5281
5282         /* we found what we needed */
5283         if (ins->objectid) {
5284                 if (!(data & BTRFS_BLOCK_GROUP_DATA))
5285                         trans->block_group = block_group->key.objectid;
5286
5287                 btrfs_put_block_group(block_group);
5288                 ret = 0;
5289         }
5290
5291         return ret;
5292 }
5293
5294 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5295                             int dump_block_groups)
5296 {
5297         struct btrfs_block_group_cache *cache;
5298         int index = 0;
5299
5300         spin_lock(&info->lock);
5301         printk(KERN_INFO "space_info has %llu free, is %sfull\n",
5302                (unsigned long long)(info->total_bytes - info->bytes_used -
5303                                     info->bytes_pinned - info->bytes_reserved -
5304                                     info->bytes_readonly),
5305                (info->full) ? "" : "not ");
5306         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5307                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5308                (unsigned long long)info->total_bytes,
5309                (unsigned long long)info->bytes_used,
5310                (unsigned long long)info->bytes_pinned,
5311                (unsigned long long)info->bytes_reserved,
5312                (unsigned long long)info->bytes_may_use,
5313                (unsigned long long)info->bytes_readonly);
5314         spin_unlock(&info->lock);
5315
5316         if (!dump_block_groups)
5317                 return;
5318
5319         down_read(&info->groups_sem);
5320 again:
5321         list_for_each_entry(cache, &info->block_groups[index], list) {
5322                 spin_lock(&cache->lock);
5323                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5324                        "%llu pinned %llu reserved\n",
5325                        (unsigned long long)cache->key.objectid,
5326                        (unsigned long long)cache->key.offset,
5327                        (unsigned long long)btrfs_block_group_used(&cache->item),
5328                        (unsigned long long)cache->pinned,
5329                        (unsigned long long)cache->reserved);
5330                 btrfs_dump_free_space(cache, bytes);
5331                 spin_unlock(&cache->lock);
5332         }
5333         if (++index < BTRFS_NR_RAID_TYPES)
5334                 goto again;
5335         up_read(&info->groups_sem);
5336 }
5337
5338 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5339                          struct btrfs_root *root,
5340                          u64 num_bytes, u64 min_alloc_size,
5341                          u64 empty_size, u64 hint_byte,
5342                          u64 search_end, struct btrfs_key *ins,
5343                          u64 data)
5344 {
5345         int ret;
5346         u64 search_start = 0;
5347
5348         data = btrfs_get_alloc_profile(root, data);
5349 again:
5350         /*
5351          * the only place that sets empty_size is btrfs_realloc_node, which
5352          * is not called recursively on allocations
5353          */
5354         if (empty_size || root->ref_cows)
5355                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5356                                      num_bytes + 2 * 1024 * 1024, data, 0);
5357
5358         WARN_ON(num_bytes < root->sectorsize);
5359         ret = find_free_extent(trans, root, num_bytes, empty_size,
5360                                search_start, search_end, hint_byte,
5361                                ins, data);
5362
5363         if (ret == -ENOSPC && num_bytes > min_alloc_size) {
5364                 num_bytes = num_bytes >> 1;
5365                 num_bytes = num_bytes & ~(root->sectorsize - 1);
5366                 num_bytes = max(num_bytes, min_alloc_size);
5367                 do_chunk_alloc(trans, root->fs_info->extent_root,
5368                                num_bytes, data, 1);
5369                 goto again;
5370         }
5371         if (ret == -ENOSPC) {
5372                 struct btrfs_space_info *sinfo;
5373
5374                 sinfo = __find_space_info(root->fs_info, data);
5375                 printk(KERN_ERR "btrfs allocation failed flags %llu, "
5376                        "wanted %llu\n", (unsigned long long)data,
5377                        (unsigned long long)num_bytes);
5378                 dump_space_info(sinfo, num_bytes, 1);
5379         }
5380
5381         return ret;
5382 }
5383
5384 int btrfs_free_reserved_extent(struct btrfs_root *root, u64 start, u64 len)
5385 {
5386         struct btrfs_block_group_cache *cache;
5387         int ret = 0;
5388
5389         cache = btrfs_lookup_block_group(root->fs_info, start);
5390         if (!cache) {
5391                 printk(KERN_ERR "Unable to find block group for %llu\n",
5392                        (unsigned long long)start);
5393                 return -ENOSPC;
5394         }
5395
5396         ret = btrfs_discard_extent(root, start, len);
5397
5398         btrfs_add_free_space(cache, start, len);
5399         update_reserved_bytes(cache, len, 0, 1);
5400         btrfs_put_block_group(cache);
5401
5402         return ret;
5403 }
5404
5405 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5406                                       struct btrfs_root *root,
5407                                       u64 parent, u64 root_objectid,
5408                                       u64 flags, u64 owner, u64 offset,
5409                                       struct btrfs_key *ins, int ref_mod)
5410 {
5411         int ret;
5412         struct btrfs_fs_info *fs_info = root->fs_info;
5413         struct btrfs_extent_item *extent_item;
5414         struct btrfs_extent_inline_ref *iref;
5415         struct btrfs_path *path;
5416         struct extent_buffer *leaf;
5417         int type;
5418         u32 size;
5419
5420         if (parent > 0)
5421                 type = BTRFS_SHARED_DATA_REF_KEY;
5422         else
5423                 type = BTRFS_EXTENT_DATA_REF_KEY;
5424
5425         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5426
5427         path = btrfs_alloc_path();
5428         BUG_ON(!path);
5429
5430         path->leave_spinning = 1;
5431         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5432                                       ins, size);
5433         BUG_ON(ret);
5434
5435         leaf = path->nodes[0];
5436         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5437                                      struct btrfs_extent_item);
5438         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
5439         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5440         btrfs_set_extent_flags(leaf, extent_item,
5441                                flags | BTRFS_EXTENT_FLAG_DATA);
5442
5443         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
5444         btrfs_set_extent_inline_ref_type(leaf, iref, type);
5445         if (parent > 0) {
5446                 struct btrfs_shared_data_ref *ref;
5447                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
5448                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5449                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
5450         } else {
5451                 struct btrfs_extent_data_ref *ref;
5452                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
5453                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
5454                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
5455                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
5456                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
5457         }
5458
5459         btrfs_mark_buffer_dirty(path->nodes[0]);
5460         btrfs_free_path(path);
5461
5462         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5463         if (ret) {
5464                 printk(KERN_ERR "btrfs update block group failed for %llu "
5465                        "%llu\n", (unsigned long long)ins->objectid,
5466                        (unsigned long long)ins->offset);
5467                 BUG();
5468         }
5469         return ret;
5470 }
5471
5472 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
5473                                      struct btrfs_root *root,
5474                                      u64 parent, u64 root_objectid,
5475                                      u64 flags, struct btrfs_disk_key *key,
5476                                      int level, struct btrfs_key *ins)
5477 {
5478         int ret;
5479         struct btrfs_fs_info *fs_info = root->fs_info;
5480         struct btrfs_extent_item *extent_item;
5481         struct btrfs_tree_block_info *block_info;
5482         struct btrfs_extent_inline_ref *iref;
5483         struct btrfs_path *path;
5484         struct extent_buffer *leaf;
5485         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
5486
5487         path = btrfs_alloc_path();
5488         BUG_ON(!path);
5489
5490         path->leave_spinning = 1;
5491         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5492                                       ins, size);
5493         BUG_ON(ret);
5494
5495         leaf = path->nodes[0];
5496         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5497                                      struct btrfs_extent_item);
5498         btrfs_set_extent_refs(leaf, extent_item, 1);
5499         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5500         btrfs_set_extent_flags(leaf, extent_item,
5501                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
5502         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
5503
5504         btrfs_set_tree_block_key(leaf, block_info, key);
5505         btrfs_set_tree_block_level(leaf, block_info, level);
5506
5507         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
5508         if (parent > 0) {
5509                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
5510                 btrfs_set_extent_inline_ref_type(leaf, iref,
5511                                                  BTRFS_SHARED_BLOCK_REF_KEY);
5512                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5513         } else {
5514                 btrfs_set_extent_inline_ref_type(leaf, iref,
5515                                                  BTRFS_TREE_BLOCK_REF_KEY);
5516                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
5517         }
5518
5519         btrfs_mark_buffer_dirty(leaf);
5520         btrfs_free_path(path);
5521
5522         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5523         if (ret) {
5524                 printk(KERN_ERR "btrfs update block group failed for %llu "
5525                        "%llu\n", (unsigned long long)ins->objectid,
5526                        (unsigned long long)ins->offset);
5527                 BUG();
5528         }
5529         return ret;
5530 }
5531
5532 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5533                                      struct btrfs_root *root,
5534                                      u64 root_objectid, u64 owner,
5535                                      u64 offset, struct btrfs_key *ins)
5536 {
5537         int ret;
5538
5539         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
5540
5541         ret = btrfs_add_delayed_data_ref(trans, ins->objectid, ins->offset,
5542                                          0, root_objectid, owner, offset,
5543                                          BTRFS_ADD_DELAYED_EXTENT, NULL);
5544         return ret;
5545 }
5546
5547 /*
5548  * this is used by the tree logging recovery code.  It records that
5549  * an extent has been allocated and makes sure to clear the free
5550  * space cache bits as well
5551  */
5552 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
5553                                    struct btrfs_root *root,
5554                                    u64 root_objectid, u64 owner, u64 offset,
5555                                    struct btrfs_key *ins)
5556 {
5557         int ret;
5558         struct btrfs_block_group_cache *block_group;
5559         struct btrfs_caching_control *caching_ctl;
5560         u64 start = ins->objectid;
5561         u64 num_bytes = ins->offset;
5562
5563         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
5564         cache_block_group(block_group, trans, NULL, 0);
5565         caching_ctl = get_caching_control(block_group);
5566
5567         if (!caching_ctl) {
5568                 BUG_ON(!block_group_cache_done(block_group));
5569                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5570                 BUG_ON(ret);
5571         } else {
5572                 mutex_lock(&caching_ctl->mutex);
5573
5574                 if (start >= caching_ctl->progress) {
5575                         ret = add_excluded_extent(root, start, num_bytes);
5576                         BUG_ON(ret);
5577                 } else if (start + num_bytes <= caching_ctl->progress) {
5578                         ret = btrfs_remove_free_space(block_group,
5579                                                       start, num_bytes);
5580                         BUG_ON(ret);
5581                 } else {
5582                         num_bytes = caching_ctl->progress - start;
5583                         ret = btrfs_remove_free_space(block_group,
5584                                                       start, num_bytes);
5585                         BUG_ON(ret);
5586
5587                         start = caching_ctl->progress;
5588                         num_bytes = ins->objectid + ins->offset -
5589                                     caching_ctl->progress;
5590                         ret = add_excluded_extent(root, start, num_bytes);
5591                         BUG_ON(ret);
5592                 }
5593
5594                 mutex_unlock(&caching_ctl->mutex);
5595                 put_caching_control(caching_ctl);
5596         }
5597
5598         ret = update_reserved_bytes(block_group, ins->offset, 1, 1);
5599         BUG_ON(ret);
5600         btrfs_put_block_group(block_group);
5601         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
5602                                          0, owner, offset, ins, 1);
5603         return ret;
5604 }
5605
5606 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
5607                                             struct btrfs_root *root,
5608                                             u64 bytenr, u32 blocksize,
5609                                             int level)
5610 {
5611         struct extent_buffer *buf;
5612
5613         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
5614         if (!buf)
5615                 return ERR_PTR(-ENOMEM);
5616         btrfs_set_header_generation(buf, trans->transid);
5617         btrfs_set_buffer_lockdep_class(buf, level);
5618         btrfs_tree_lock(buf);
5619         clean_tree_block(trans, root, buf);
5620
5621         btrfs_set_lock_blocking(buf);
5622         btrfs_set_buffer_uptodate(buf);
5623
5624         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
5625                 /*
5626                  * we allow two log transactions at a time, use different
5627                  * EXENT bit to differentiate dirty pages.
5628                  */
5629                 if (root->log_transid % 2 == 0)
5630                         set_extent_dirty(&root->dirty_log_pages, buf->start,
5631                                         buf->start + buf->len - 1, GFP_NOFS);
5632                 else
5633                         set_extent_new(&root->dirty_log_pages, buf->start,
5634                                         buf->start + buf->len - 1, GFP_NOFS);
5635         } else {
5636                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
5637                          buf->start + buf->len - 1, GFP_NOFS);
5638         }
5639         trans->blocks_used++;
5640         /* this returns a buffer locked for blocking */
5641         return buf;
5642 }
5643
5644 static struct btrfs_block_rsv *
5645 use_block_rsv(struct btrfs_trans_handle *trans,
5646               struct btrfs_root *root, u32 blocksize)
5647 {
5648         struct btrfs_block_rsv *block_rsv;
5649         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5650         int ret;
5651
5652         block_rsv = get_block_rsv(trans, root);
5653
5654         if (block_rsv->size == 0) {
5655                 ret = reserve_metadata_bytes(trans, root, block_rsv,
5656                                              blocksize, 0);
5657                 /*
5658                  * If we couldn't reserve metadata bytes try and use some from
5659                  * the global reserve.
5660                  */
5661                 if (ret && block_rsv != global_rsv) {
5662                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5663                         if (!ret)
5664                                 return global_rsv;
5665                         return ERR_PTR(ret);
5666                 } else if (ret) {
5667                         return ERR_PTR(ret);
5668                 }
5669                 return block_rsv;
5670         }
5671
5672         ret = block_rsv_use_bytes(block_rsv, blocksize);
5673         if (!ret)
5674                 return block_rsv;
5675         if (ret) {
5676                 WARN_ON(1);
5677                 ret = reserve_metadata_bytes(trans, root, block_rsv, blocksize,
5678                                              0);
5679                 if (!ret) {
5680                         spin_lock(&block_rsv->lock);
5681                         block_rsv->size += blocksize;
5682                         spin_unlock(&block_rsv->lock);
5683                         return block_rsv;
5684                 } else if (ret && block_rsv != global_rsv) {
5685                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5686                         if (!ret)
5687                                 return global_rsv;
5688                 }
5689         }
5690
5691         return ERR_PTR(-ENOSPC);
5692 }
5693
5694 static void unuse_block_rsv(struct btrfs_block_rsv *block_rsv, u32 blocksize)
5695 {
5696         block_rsv_add_bytes(block_rsv, blocksize, 0);
5697         block_rsv_release_bytes(block_rsv, NULL, 0);
5698 }
5699
5700 /*
5701  * finds a free extent and does all the dirty work required for allocation
5702  * returns the key for the extent through ins, and a tree buffer for
5703  * the first block of the extent through buf.
5704  *
5705  * returns the tree buffer or NULL.
5706  */
5707 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
5708                                         struct btrfs_root *root, u32 blocksize,
5709                                         u64 parent, u64 root_objectid,
5710                                         struct btrfs_disk_key *key, int level,
5711                                         u64 hint, u64 empty_size)
5712 {
5713         struct btrfs_key ins;
5714         struct btrfs_block_rsv *block_rsv;
5715         struct extent_buffer *buf;
5716         u64 flags = 0;
5717         int ret;
5718
5719
5720         block_rsv = use_block_rsv(trans, root, blocksize);
5721         if (IS_ERR(block_rsv))
5722                 return ERR_CAST(block_rsv);
5723
5724         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
5725                                    empty_size, hint, (u64)-1, &ins, 0);
5726         if (ret) {
5727                 unuse_block_rsv(block_rsv, blocksize);
5728                 return ERR_PTR(ret);
5729         }
5730
5731         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
5732                                     blocksize, level);
5733         BUG_ON(IS_ERR(buf));
5734
5735         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
5736                 if (parent == 0)
5737                         parent = ins.objectid;
5738                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
5739         } else
5740                 BUG_ON(parent > 0);
5741
5742         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
5743                 struct btrfs_delayed_extent_op *extent_op;
5744                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
5745                 BUG_ON(!extent_op);
5746                 if (key)
5747                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
5748                 else
5749                         memset(&extent_op->key, 0, sizeof(extent_op->key));
5750                 extent_op->flags_to_set = flags;
5751                 extent_op->update_key = 1;
5752                 extent_op->update_flags = 1;
5753                 extent_op->is_data = 0;
5754
5755                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
5756                                         ins.offset, parent, root_objectid,
5757                                         level, BTRFS_ADD_DELAYED_EXTENT,
5758                                         extent_op);
5759                 BUG_ON(ret);
5760         }
5761         return buf;
5762 }
5763
5764 struct walk_control {
5765         u64 refs[BTRFS_MAX_LEVEL];
5766         u64 flags[BTRFS_MAX_LEVEL];
5767         struct btrfs_key update_progress;
5768         int stage;
5769         int level;
5770         int shared_level;
5771         int update_ref;
5772         int keep_locks;
5773         int reada_slot;
5774         int reada_count;
5775 };
5776
5777 #define DROP_REFERENCE  1
5778 #define UPDATE_BACKREF  2
5779
5780 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
5781                                      struct btrfs_root *root,
5782                                      struct walk_control *wc,
5783                                      struct btrfs_path *path)
5784 {
5785         u64 bytenr;
5786         u64 generation;
5787         u64 refs;
5788         u64 flags;
5789         u32 nritems;
5790         u32 blocksize;
5791         struct btrfs_key key;
5792         struct extent_buffer *eb;
5793         int ret;
5794         int slot;
5795         int nread = 0;
5796
5797         if (path->slots[wc->level] < wc->reada_slot) {
5798                 wc->reada_count = wc->reada_count * 2 / 3;
5799                 wc->reada_count = max(wc->reada_count, 2);
5800         } else {
5801                 wc->reada_count = wc->reada_count * 3 / 2;
5802                 wc->reada_count = min_t(int, wc->reada_count,
5803                                         BTRFS_NODEPTRS_PER_BLOCK(root));
5804         }
5805
5806         eb = path->nodes[wc->level];
5807         nritems = btrfs_header_nritems(eb);
5808         blocksize = btrfs_level_size(root, wc->level - 1);
5809
5810         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
5811                 if (nread >= wc->reada_count)
5812                         break;
5813
5814                 cond_resched();
5815                 bytenr = btrfs_node_blockptr(eb, slot);
5816                 generation = btrfs_node_ptr_generation(eb, slot);
5817
5818                 if (slot == path->slots[wc->level])
5819                         goto reada;
5820
5821                 if (wc->stage == UPDATE_BACKREF &&
5822                     generation <= root->root_key.offset)
5823                         continue;
5824
5825                 /* We don't lock the tree block, it's OK to be racy here */
5826                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
5827                                                &refs, &flags);
5828                 BUG_ON(ret);
5829                 BUG_ON(refs == 0);
5830
5831                 if (wc->stage == DROP_REFERENCE) {
5832                         if (refs == 1)
5833                                 goto reada;
5834
5835                         if (wc->level == 1 &&
5836                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5837                                 continue;
5838                         if (!wc->update_ref ||
5839                             generation <= root->root_key.offset)
5840                                 continue;
5841                         btrfs_node_key_to_cpu(eb, &key, slot);
5842                         ret = btrfs_comp_cpu_keys(&key,
5843                                                   &wc->update_progress);
5844                         if (ret < 0)
5845                                 continue;
5846                 } else {
5847                         if (wc->level == 1 &&
5848                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5849                                 continue;
5850                 }
5851 reada:
5852                 ret = readahead_tree_block(root, bytenr, blocksize,
5853                                            generation);
5854                 if (ret)
5855                         break;
5856                 nread++;
5857         }
5858         wc->reada_slot = slot;
5859 }
5860
5861 /*
5862  * hepler to process tree block while walking down the tree.
5863  *
5864  * when wc->stage == UPDATE_BACKREF, this function updates
5865  * back refs for pointers in the block.
5866  *
5867  * NOTE: return value 1 means we should stop walking down.
5868  */
5869 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
5870                                    struct btrfs_root *root,
5871                                    struct btrfs_path *path,
5872                                    struct walk_control *wc, int lookup_info)
5873 {
5874         int level = wc->level;
5875         struct extent_buffer *eb = path->nodes[level];
5876         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5877         int ret;
5878
5879         if (wc->stage == UPDATE_BACKREF &&
5880             btrfs_header_owner(eb) != root->root_key.objectid)
5881                 return 1;
5882
5883         /*
5884          * when reference count of tree block is 1, it won't increase
5885          * again. once full backref flag is set, we never clear it.
5886          */
5887         if (lookup_info &&
5888             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
5889              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
5890                 BUG_ON(!path->locks[level]);
5891                 ret = btrfs_lookup_extent_info(trans, root,
5892                                                eb->start, eb->len,
5893                                                &wc->refs[level],
5894                                                &wc->flags[level]);
5895                 BUG_ON(ret);
5896                 BUG_ON(wc->refs[level] == 0);
5897         }
5898
5899         if (wc->stage == DROP_REFERENCE) {
5900                 if (wc->refs[level] > 1)
5901                         return 1;
5902
5903                 if (path->locks[level] && !wc->keep_locks) {
5904                         btrfs_tree_unlock(eb);
5905                         path->locks[level] = 0;
5906                 }
5907                 return 0;
5908         }
5909
5910         /* wc->stage == UPDATE_BACKREF */
5911         if (!(wc->flags[level] & flag)) {
5912                 BUG_ON(!path->locks[level]);
5913                 ret = btrfs_inc_ref(trans, root, eb, 1);
5914                 BUG_ON(ret);
5915                 ret = btrfs_dec_ref(trans, root, eb, 0);
5916                 BUG_ON(ret);
5917                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
5918                                                   eb->len, flag, 0);
5919                 BUG_ON(ret);
5920                 wc->flags[level] |= flag;
5921         }
5922
5923         /*
5924          * the block is shared by multiple trees, so it's not good to
5925          * keep the tree lock
5926          */
5927         if (path->locks[level] && level > 0) {
5928                 btrfs_tree_unlock(eb);
5929                 path->locks[level] = 0;
5930         }
5931         return 0;
5932 }
5933
5934 /*
5935  * hepler to process tree block pointer.
5936  *
5937  * when wc->stage == DROP_REFERENCE, this function checks
5938  * reference count of the block pointed to. if the block
5939  * is shared and we need update back refs for the subtree
5940  * rooted at the block, this function changes wc->stage to
5941  * UPDATE_BACKREF. if the block is shared and there is no
5942  * need to update back, this function drops the reference
5943  * to the block.
5944  *
5945  * NOTE: return value 1 means we should stop walking down.
5946  */
5947 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
5948                                  struct btrfs_root *root,
5949                                  struct btrfs_path *path,
5950                                  struct walk_control *wc, int *lookup_info)
5951 {
5952         u64 bytenr;
5953         u64 generation;
5954         u64 parent;
5955         u32 blocksize;
5956         struct btrfs_key key;
5957         struct extent_buffer *next;
5958         int level = wc->level;
5959         int reada = 0;
5960         int ret = 0;
5961
5962         generation = btrfs_node_ptr_generation(path->nodes[level],
5963                                                path->slots[level]);
5964         /*
5965          * if the lower level block was created before the snapshot
5966          * was created, we know there is no need to update back refs
5967          * for the subtree
5968          */
5969         if (wc->stage == UPDATE_BACKREF &&
5970             generation <= root->root_key.offset) {
5971                 *lookup_info = 1;
5972                 return 1;
5973         }
5974
5975         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
5976         blocksize = btrfs_level_size(root, level - 1);
5977
5978         next = btrfs_find_tree_block(root, bytenr, blocksize);
5979         if (!next) {
5980                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
5981                 if (!next)
5982                         return -ENOMEM;
5983                 reada = 1;
5984         }
5985         btrfs_tree_lock(next);
5986         btrfs_set_lock_blocking(next);
5987
5988         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
5989                                        &wc->refs[level - 1],
5990                                        &wc->flags[level - 1]);
5991         BUG_ON(ret);
5992         BUG_ON(wc->refs[level - 1] == 0);
5993         *lookup_info = 0;
5994
5995         if (wc->stage == DROP_REFERENCE) {
5996                 if (wc->refs[level - 1] > 1) {
5997                         if (level == 1 &&
5998                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5999                                 goto skip;
6000
6001                         if (!wc->update_ref ||
6002                             generation <= root->root_key.offset)
6003                                 goto skip;
6004
6005                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6006                                               path->slots[level]);
6007                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6008                         if (ret < 0)
6009                                 goto skip;
6010
6011                         wc->stage = UPDATE_BACKREF;
6012                         wc->shared_level = level - 1;
6013                 }
6014         } else {
6015                 if (level == 1 &&
6016                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6017                         goto skip;
6018         }
6019
6020         if (!btrfs_buffer_uptodate(next, generation)) {
6021                 btrfs_tree_unlock(next);
6022                 free_extent_buffer(next);
6023                 next = NULL;
6024                 *lookup_info = 1;
6025         }
6026
6027         if (!next) {
6028                 if (reada && level == 1)
6029                         reada_walk_down(trans, root, wc, path);
6030                 next = read_tree_block(root, bytenr, blocksize, generation);
6031                 btrfs_tree_lock(next);
6032                 btrfs_set_lock_blocking(next);
6033         }
6034
6035         level--;
6036         BUG_ON(level != btrfs_header_level(next));
6037         path->nodes[level] = next;
6038         path->slots[level] = 0;
6039         path->locks[level] = 1;
6040         wc->level = level;
6041         if (wc->level == 1)
6042                 wc->reada_slot = 0;
6043         return 0;
6044 skip:
6045         wc->refs[level - 1] = 0;
6046         wc->flags[level - 1] = 0;
6047         if (wc->stage == DROP_REFERENCE) {
6048                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6049                         parent = path->nodes[level]->start;
6050                 } else {
6051                         BUG_ON(root->root_key.objectid !=
6052                                btrfs_header_owner(path->nodes[level]));
6053                         parent = 0;
6054                 }
6055
6056                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6057                                         root->root_key.objectid, level - 1, 0);
6058                 BUG_ON(ret);
6059         }
6060         btrfs_tree_unlock(next);
6061         free_extent_buffer(next);
6062         *lookup_info = 1;
6063         return 1;
6064 }
6065
6066 /*
6067  * hepler to process tree block while walking up the tree.
6068  *
6069  * when wc->stage == DROP_REFERENCE, this function drops
6070  * reference count on the block.
6071  *
6072  * when wc->stage == UPDATE_BACKREF, this function changes
6073  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6074  * to UPDATE_BACKREF previously while processing the block.
6075  *
6076  * NOTE: return value 1 means we should stop walking up.
6077  */
6078 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6079                                  struct btrfs_root *root,
6080                                  struct btrfs_path *path,
6081                                  struct walk_control *wc)
6082 {
6083         int ret;
6084         int level = wc->level;
6085         struct extent_buffer *eb = path->nodes[level];
6086         u64 parent = 0;
6087
6088         if (wc->stage == UPDATE_BACKREF) {
6089                 BUG_ON(wc->shared_level < level);
6090                 if (level < wc->shared_level)
6091                         goto out;
6092
6093                 ret = find_next_key(path, level + 1, &wc->update_progress);
6094                 if (ret > 0)
6095                         wc->update_ref = 0;
6096
6097                 wc->stage = DROP_REFERENCE;
6098                 wc->shared_level = -1;
6099                 path->slots[level] = 0;
6100
6101                 /*
6102                  * check reference count again if the block isn't locked.
6103                  * we should start walking down the tree again if reference
6104                  * count is one.
6105                  */
6106                 if (!path->locks[level]) {
6107                         BUG_ON(level == 0);
6108                         btrfs_tree_lock(eb);
6109                         btrfs_set_lock_blocking(eb);
6110                         path->locks[level] = 1;
6111
6112                         ret = btrfs_lookup_extent_info(trans, root,
6113                                                        eb->start, eb->len,
6114                                                        &wc->refs[level],
6115                                                        &wc->flags[level]);
6116                         BUG_ON(ret);
6117                         BUG_ON(wc->refs[level] == 0);
6118                         if (wc->refs[level] == 1) {
6119                                 btrfs_tree_unlock(eb);
6120                                 path->locks[level] = 0;
6121                                 return 1;
6122                         }
6123                 }
6124         }
6125
6126         /* wc->stage == DROP_REFERENCE */
6127         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6128
6129         if (wc->refs[level] == 1) {
6130                 if (level == 0) {
6131                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6132                                 ret = btrfs_dec_ref(trans, root, eb, 1);
6133                         else
6134                                 ret = btrfs_dec_ref(trans, root, eb, 0);
6135                         BUG_ON(ret);
6136                 }
6137                 /* make block locked assertion in clean_tree_block happy */
6138                 if (!path->locks[level] &&
6139                     btrfs_header_generation(eb) == trans->transid) {
6140                         btrfs_tree_lock(eb);
6141                         btrfs_set_lock_blocking(eb);
6142                         path->locks[level] = 1;
6143                 }
6144                 clean_tree_block(trans, root, eb);
6145         }
6146
6147         if (eb == root->node) {
6148                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6149                         parent = eb->start;
6150                 else
6151                         BUG_ON(root->root_key.objectid !=
6152                                btrfs_header_owner(eb));
6153         } else {
6154                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6155                         parent = path->nodes[level + 1]->start;
6156                 else
6157                         BUG_ON(root->root_key.objectid !=
6158                                btrfs_header_owner(path->nodes[level + 1]));
6159         }
6160
6161         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6162 out:
6163         wc->refs[level] = 0;
6164         wc->flags[level] = 0;
6165         return 0;
6166 }
6167
6168 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6169                                    struct btrfs_root *root,
6170                                    struct btrfs_path *path,
6171                                    struct walk_control *wc)
6172 {
6173         int level = wc->level;
6174         int lookup_info = 1;
6175         int ret;
6176
6177         while (level >= 0) {
6178                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6179                 if (ret > 0)
6180                         break;
6181
6182                 if (level == 0)
6183                         break;
6184
6185                 if (path->slots[level] >=
6186                     btrfs_header_nritems(path->nodes[level]))
6187                         break;
6188
6189                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6190                 if (ret > 0) {
6191                         path->slots[level]++;
6192                         continue;
6193                 } else if (ret < 0)
6194                         return ret;
6195                 level = wc->level;
6196         }
6197         return 0;
6198 }
6199
6200 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6201                                  struct btrfs_root *root,
6202                                  struct btrfs_path *path,
6203                                  struct walk_control *wc, int max_level)
6204 {
6205         int level = wc->level;
6206         int ret;
6207
6208         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6209         while (level < max_level && path->nodes[level]) {
6210                 wc->level = level;
6211                 if (path->slots[level] + 1 <
6212                     btrfs_header_nritems(path->nodes[level])) {
6213                         path->slots[level]++;
6214                         return 0;
6215                 } else {
6216                         ret = walk_up_proc(trans, root, path, wc);
6217                         if (ret > 0)
6218                                 return 0;
6219
6220                         if (path->locks[level]) {
6221                                 btrfs_tree_unlock(path->nodes[level]);
6222                                 path->locks[level] = 0;
6223                         }
6224                         free_extent_buffer(path->nodes[level]);
6225                         path->nodes[level] = NULL;
6226                         level++;
6227                 }
6228         }
6229         return 1;
6230 }
6231
6232 /*
6233  * drop a subvolume tree.
6234  *
6235  * this function traverses the tree freeing any blocks that only
6236  * referenced by the tree.
6237  *
6238  * when a shared tree block is found. this function decreases its
6239  * reference count by one. if update_ref is true, this function
6240  * also make sure backrefs for the shared block and all lower level
6241  * blocks are properly updated.
6242  */
6243 int btrfs_drop_snapshot(struct btrfs_root *root,
6244                         struct btrfs_block_rsv *block_rsv, int update_ref)
6245 {
6246         struct btrfs_path *path;
6247         struct btrfs_trans_handle *trans;
6248         struct btrfs_root *tree_root = root->fs_info->tree_root;
6249         struct btrfs_root_item *root_item = &root->root_item;
6250         struct walk_control *wc;
6251         struct btrfs_key key;
6252         int err = 0;
6253         int ret;
6254         int level;
6255
6256         path = btrfs_alloc_path();
6257         BUG_ON(!path);
6258
6259         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6260         BUG_ON(!wc);
6261
6262         trans = btrfs_start_transaction(tree_root, 0);
6263         if (block_rsv)
6264                 trans->block_rsv = block_rsv;
6265
6266         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6267                 level = btrfs_header_level(root->node);
6268                 path->nodes[level] = btrfs_lock_root_node(root);
6269                 btrfs_set_lock_blocking(path->nodes[level]);
6270                 path->slots[level] = 0;
6271                 path->locks[level] = 1;
6272                 memset(&wc->update_progress, 0,
6273                        sizeof(wc->update_progress));
6274         } else {
6275                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6276                 memcpy(&wc->update_progress, &key,
6277                        sizeof(wc->update_progress));
6278
6279                 level = root_item->drop_level;
6280                 BUG_ON(level == 0);
6281                 path->lowest_level = level;
6282                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6283                 path->lowest_level = 0;
6284                 if (ret < 0) {
6285                         err = ret;
6286                         goto out;
6287                 }
6288                 WARN_ON(ret > 0);
6289
6290                 /*
6291                  * unlock our path, this is safe because only this
6292                  * function is allowed to delete this snapshot
6293                  */
6294                 btrfs_unlock_up_safe(path, 0);
6295
6296                 level = btrfs_header_level(root->node);
6297                 while (1) {
6298                         btrfs_tree_lock(path->nodes[level]);
6299                         btrfs_set_lock_blocking(path->nodes[level]);
6300
6301                         ret = btrfs_lookup_extent_info(trans, root,
6302                                                 path->nodes[level]->start,
6303                                                 path->nodes[level]->len,
6304                                                 &wc->refs[level],
6305                                                 &wc->flags[level]);
6306                         BUG_ON(ret);
6307                         BUG_ON(wc->refs[level] == 0);
6308
6309                         if (level == root_item->drop_level)
6310                                 break;
6311
6312                         btrfs_tree_unlock(path->nodes[level]);
6313                         WARN_ON(wc->refs[level] != 1);
6314                         level--;
6315                 }
6316         }
6317
6318         wc->level = level;
6319         wc->shared_level = -1;
6320         wc->stage = DROP_REFERENCE;
6321         wc->update_ref = update_ref;
6322         wc->keep_locks = 0;
6323         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6324
6325         while (1) {
6326                 ret = walk_down_tree(trans, root, path, wc);
6327                 if (ret < 0) {
6328                         err = ret;
6329                         break;
6330                 }
6331
6332                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
6333                 if (ret < 0) {
6334                         err = ret;
6335                         break;
6336                 }
6337
6338                 if (ret > 0) {
6339                         BUG_ON(wc->stage != DROP_REFERENCE);
6340                         break;
6341                 }
6342
6343                 if (wc->stage == DROP_REFERENCE) {
6344                         level = wc->level;
6345                         btrfs_node_key(path->nodes[level],
6346                                        &root_item->drop_progress,
6347                                        path->slots[level]);
6348                         root_item->drop_level = level;
6349                 }
6350
6351                 BUG_ON(wc->level == 0);
6352                 if (btrfs_should_end_transaction(trans, tree_root)) {
6353                         ret = btrfs_update_root(trans, tree_root,
6354                                                 &root->root_key,
6355                                                 root_item);
6356                         BUG_ON(ret);
6357
6358                         btrfs_end_transaction_throttle(trans, tree_root);
6359                         trans = btrfs_start_transaction(tree_root, 0);
6360                         if (block_rsv)
6361                                 trans->block_rsv = block_rsv;
6362                 }
6363         }
6364         btrfs_release_path(root, path);
6365         BUG_ON(err);
6366
6367         ret = btrfs_del_root(trans, tree_root, &root->root_key);
6368         BUG_ON(ret);
6369
6370         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
6371                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
6372                                            NULL, NULL);
6373                 BUG_ON(ret < 0);
6374                 if (ret > 0) {
6375                         /* if we fail to delete the orphan item this time
6376                          * around, it'll get picked up the next time.
6377                          *
6378                          * The most common failure here is just -ENOENT.
6379                          */
6380                         btrfs_del_orphan_item(trans, tree_root,
6381                                               root->root_key.objectid);
6382                 }
6383         }
6384
6385         if (root->in_radix) {
6386                 btrfs_free_fs_root(tree_root->fs_info, root);
6387         } else {
6388                 free_extent_buffer(root->node);
6389                 free_extent_buffer(root->commit_root);
6390                 kfree(root);
6391         }
6392 out:
6393         btrfs_end_transaction_throttle(trans, tree_root);
6394         kfree(wc);
6395         btrfs_free_path(path);
6396         return err;
6397 }
6398
6399 /*
6400  * drop subtree rooted at tree block 'node'.
6401  *
6402  * NOTE: this function will unlock and release tree block 'node'
6403  */
6404 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
6405                         struct btrfs_root *root,
6406                         struct extent_buffer *node,
6407                         struct extent_buffer *parent)
6408 {
6409         struct btrfs_path *path;
6410         struct walk_control *wc;
6411         int level;
6412         int parent_level;
6413         int ret = 0;
6414         int wret;
6415
6416         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
6417
6418         path = btrfs_alloc_path();
6419         BUG_ON(!path);
6420
6421         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6422         BUG_ON(!wc);
6423
6424         btrfs_assert_tree_locked(parent);
6425         parent_level = btrfs_header_level(parent);
6426         extent_buffer_get(parent);
6427         path->nodes[parent_level] = parent;
6428         path->slots[parent_level] = btrfs_header_nritems(parent);
6429
6430         btrfs_assert_tree_locked(node);
6431         level = btrfs_header_level(node);
6432         path->nodes[level] = node;
6433         path->slots[level] = 0;
6434         path->locks[level] = 1;
6435
6436         wc->refs[parent_level] = 1;
6437         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6438         wc->level = level;
6439         wc->shared_level = -1;
6440         wc->stage = DROP_REFERENCE;
6441         wc->update_ref = 0;
6442         wc->keep_locks = 1;
6443         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6444
6445         while (1) {
6446                 wret = walk_down_tree(trans, root, path, wc);
6447                 if (wret < 0) {
6448                         ret = wret;
6449                         break;
6450                 }
6451
6452                 wret = walk_up_tree(trans, root, path, wc, parent_level);
6453                 if (wret < 0)
6454                         ret = wret;
6455                 if (wret != 0)
6456                         break;
6457         }
6458
6459         kfree(wc);
6460         btrfs_free_path(path);
6461         return ret;
6462 }
6463
6464 #if 0
6465 static unsigned long calc_ra(unsigned long start, unsigned long last,
6466                              unsigned long nr)
6467 {
6468         return min(last, start + nr - 1);
6469 }
6470
6471 static noinline int relocate_inode_pages(struct inode *inode, u64 start,
6472                                          u64 len)
6473 {
6474         u64 page_start;
6475         u64 page_end;
6476         unsigned long first_index;
6477         unsigned long last_index;
6478         unsigned long i;
6479         struct page *page;
6480         struct extent_io_tree *io_tree = &BTRFS_I(inode)->io_tree;
6481         struct file_ra_state *ra;
6482         struct btrfs_ordered_extent *ordered;
6483         unsigned int total_read = 0;
6484         unsigned int total_dirty = 0;
6485         int ret = 0;
6486
6487         ra = kzalloc(sizeof(*ra), GFP_NOFS);
6488
6489         mutex_lock(&inode->i_mutex);
6490         first_index = start >> PAGE_CACHE_SHIFT;
6491         last_index = (start + len - 1) >> PAGE_CACHE_SHIFT;
6492
6493         /* make sure the dirty trick played by the caller work */
6494         ret = invalidate_inode_pages2_range(inode->i_mapping,
6495                                             first_index, last_index);
6496         if (ret)
6497                 goto out_unlock;
6498
6499         file_ra_state_init(ra, inode->i_mapping);
6500
6501         for (i = first_index ; i <= last_index; i++) {
6502                 if (total_read % ra->ra_pages == 0) {
6503                         btrfs_force_ra(inode->i_mapping, ra, NULL, i,
6504                                        calc_ra(i, last_index, ra->ra_pages));
6505                 }
6506                 total_read++;
6507 again:
6508                 if (((u64)i << PAGE_CACHE_SHIFT) > i_size_read(inode))
6509                         BUG_ON(1);
6510                 page = grab_cache_page(inode->i_mapping, i);
6511                 if (!page) {
6512                         ret = -ENOMEM;
6513                         goto out_unlock;
6514                 }
6515                 if (!PageUptodate(page)) {
6516                         btrfs_readpage(NULL, page);
6517                         lock_page(page);
6518                         if (!PageUptodate(page)) {
6519                                 unlock_page(page);
6520                                 page_cache_release(page);
6521                                 ret = -EIO;
6522                                 goto out_unlock;
6523                         }
6524                 }
6525                 wait_on_page_writeback(page);
6526
6527                 page_start = (u64)page->index << PAGE_CACHE_SHIFT;
6528                 page_end = page_start + PAGE_CACHE_SIZE - 1;
6529                 lock_extent(io_tree, page_start, page_end, GFP_NOFS);
6530
6531                 ordered = btrfs_lookup_ordered_extent(inode, page_start);
6532                 if (ordered) {
6533                         unlock_extent(io_tree, page_start, page_end, GFP_NOFS);
6534                         unlock_page(page);
6535                         page_cache_release(page);
6536                         btrfs_start_ordered_extent(inode, ordered, 1);
6537                         btrfs_put_ordered_extent(ordered);
6538                         goto again;
6539                 }
6540                 set_page_extent_mapped(page);
6541
6542                 if (i == first_index)
6543                         set_extent_bits(io_tree, page_start, page_end,
6544                                         EXTENT_BOUNDARY, GFP_NOFS);
6545                 btrfs_set_extent_delalloc(inode, page_start, page_end);
6546
6547                 set_page_dirty(page);
6548                 total_dirty++;
6549
6550                 unlock_extent(io_tree, page_start, page_end, GFP_NOFS);
6551                 unlock_page(page);
6552                 page_cache_release(page);
6553         }
6554
6555 out_unlock:
6556         kfree(ra);
6557         mutex_unlock(&inode->i_mutex);
6558         balance_dirty_pages_ratelimited_nr(inode->i_mapping, total_dirty);
6559         return ret;
6560 }
6561
6562 static noinline int relocate_data_extent(struct inode *reloc_inode,
6563                                          struct btrfs_key *extent_key,
6564                                          u64 offset)
6565 {
6566         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
6567         struct extent_map_tree *em_tree = &BTRFS_I(reloc_inode)->extent_tree;
6568         struct extent_map *em;
6569         u64 start = extent_key->objectid - offset;
6570         u64 end = start + extent_key->offset - 1;
6571
6572         em = alloc_extent_map(GFP_NOFS);
6573         BUG_ON(!em || IS_ERR(em));
6574
6575         em->start = start;
6576         em->len = extent_key->offset;
6577         em->block_len = extent_key->offset;
6578         em->block_start = extent_key->objectid;
6579         em->bdev = root->fs_info->fs_devices->latest_bdev;
6580         set_bit(EXTENT_FLAG_PINNED, &em->flags);
6581
6582         /* setup extent map to cheat btrfs_readpage */
6583         lock_extent(&BTRFS_I(reloc_inode)->io_tree, start, end, GFP_NOFS);
6584         while (1) {
6585                 int ret;
6586                 write_lock(&em_tree->lock);
6587                 ret = add_extent_mapping(em_tree, em);
6588                 write_unlock(&em_tree->lock);
6589                 if (ret != -EEXIST) {
6590                         free_extent_map(em);
6591                         break;
6592                 }
6593                 btrfs_drop_extent_cache(reloc_inode, start, end, 0);
6594         }
6595         unlock_extent(&BTRFS_I(reloc_inode)->io_tree, start, end, GFP_NOFS);
6596
6597         return relocate_inode_pages(reloc_inode, start, extent_key->offset);
6598 }
6599
6600 struct btrfs_ref_path {
6601         u64 extent_start;
6602         u64 nodes[BTRFS_MAX_LEVEL];
6603         u64 root_objectid;
6604         u64 root_generation;
6605         u64 owner_objectid;
6606         u32 num_refs;
6607         int lowest_level;
6608         int current_level;
6609         int shared_level;
6610
6611         struct btrfs_key node_keys[BTRFS_MAX_LEVEL];
6612         u64 new_nodes[BTRFS_MAX_LEVEL];
6613 };
6614
6615 struct disk_extent {
6616         u64 ram_bytes;
6617         u64 disk_bytenr;
6618         u64 disk_num_bytes;
6619         u64 offset;
6620         u64 num_bytes;
6621         u8 compression;
6622         u8 encryption;
6623         u16 other_encoding;
6624 };
6625
6626 static int is_cowonly_root(u64 root_objectid)
6627 {
6628         if (root_objectid == BTRFS_ROOT_TREE_OBJECTID ||
6629             root_objectid == BTRFS_EXTENT_TREE_OBJECTID ||
6630             root_objectid == BTRFS_CHUNK_TREE_OBJECTID ||
6631             root_objectid == BTRFS_DEV_TREE_OBJECTID ||
6632             root_objectid == BTRFS_TREE_LOG_OBJECTID ||
6633             root_objectid == BTRFS_CSUM_TREE_OBJECTID)
6634                 return 1;
6635         return 0;
6636 }
6637
6638 static noinline int __next_ref_path(struct btrfs_trans_handle *trans,
6639                                     struct btrfs_root *extent_root,
6640                                     struct btrfs_ref_path *ref_path,
6641                                     int first_time)
6642 {
6643         struct extent_buffer *leaf;
6644         struct btrfs_path *path;
6645         struct btrfs_extent_ref *ref;
6646         struct btrfs_key key;
6647         struct btrfs_key found_key;
6648         u64 bytenr;
6649         u32 nritems;
6650         int level;
6651         int ret = 1;
6652
6653         path = btrfs_alloc_path();
6654         if (!path)
6655                 return -ENOMEM;
6656
6657         if (first_time) {
6658                 ref_path->lowest_level = -1;
6659                 ref_path->current_level = -1;
6660                 ref_path->shared_level = -1;
6661                 goto walk_up;
6662         }
6663 walk_down:
6664         level = ref_path->current_level - 1;
6665         while (level >= -1) {
6666                 u64 parent;
6667                 if (level < ref_path->lowest_level)
6668                         break;
6669
6670                 if (level >= 0)
6671                         bytenr = ref_path->nodes[level];
6672                 else
6673                         bytenr = ref_path->extent_start;
6674                 BUG_ON(bytenr == 0);
6675
6676                 parent = ref_path->nodes[level + 1];
6677                 ref_path->nodes[level + 1] = 0;
6678                 ref_path->current_level = level;
6679                 BUG_ON(parent == 0);
6680
6681                 key.objectid = bytenr;
6682                 key.offset = parent + 1;
6683                 key.type = BTRFS_EXTENT_REF_KEY;
6684
6685                 ret = btrfs_search_slot(trans, extent_root, &key, path, 0, 0);
6686                 if (ret < 0)
6687                         goto out;
6688                 BUG_ON(ret == 0);
6689
6690                 leaf = path->nodes[0];
6691                 nritems = btrfs_header_nritems(leaf);
6692                 if (path->slots[0] >= nritems) {
6693                         ret = btrfs_next_leaf(extent_root, path);
6694                         if (ret < 0)
6695                                 goto out;
6696                         if (ret > 0)
6697                                 goto next;
6698                         leaf = path->nodes[0];
6699                 }
6700
6701                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6702                 if (found_key.objectid == bytenr &&
6703                     found_key.type == BTRFS_EXTENT_REF_KEY) {
6704                         if (level < ref_path->shared_level)
6705                                 ref_path->shared_level = level;
6706                         goto found;
6707                 }
6708 next:
6709                 level--;
6710                 btrfs_release_path(extent_root, path);
6711                 cond_resched();
6712         }
6713         /* reached lowest level */
6714         ret = 1;
6715         goto out;
6716 walk_up:
6717         level = ref_path->current_level;
6718         while (level < BTRFS_MAX_LEVEL - 1) {
6719                 u64 ref_objectid;
6720
6721                 if (level >= 0)
6722                         bytenr = ref_path->nodes[level];
6723                 else
6724                         bytenr = ref_path->extent_start;
6725
6726                 BUG_ON(bytenr == 0);
6727
6728                 key.objectid = bytenr;
6729                 key.offset = 0;
6730                 key.type = BTRFS_EXTENT_REF_KEY;
6731
6732                 ret = btrfs_search_slot(trans, extent_root, &key, path, 0, 0);
6733                 if (ret < 0)
6734                         goto out;
6735
6736                 leaf = path->nodes[0];
6737                 nritems = btrfs_header_nritems(leaf);
6738                 if (path->slots[0] >= nritems) {
6739                         ret = btrfs_next_leaf(extent_root, path);
6740                         if (ret < 0)
6741                                 goto out;
6742                         if (ret > 0) {
6743                                 /* the extent was freed by someone */
6744                                 if (ref_path->lowest_level == level)
6745                                         goto out;
6746                                 btrfs_release_path(extent_root, path);
6747                                 goto walk_down;
6748                         }
6749                         leaf = path->nodes[0];
6750                 }
6751
6752                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6753                 if (found_key.objectid != bytenr ||
6754                                 found_key.type != BTRFS_EXTENT_REF_KEY) {
6755                         /* the extent was freed by someone */
6756                         if (ref_path->lowest_level == level) {
6757                                 ret = 1;
6758                                 goto out;
6759                         }
6760                         btrfs_release_path(extent_root, path);
6761                         goto walk_down;
6762                 }
6763 found:
6764                 ref = btrfs_item_ptr(leaf, path->slots[0],
6765                                 struct btrfs_extent_ref);
6766                 ref_objectid = btrfs_ref_objectid(leaf, ref);
6767                 if (ref_objectid < BTRFS_FIRST_FREE_OBJECTID) {
6768                         if (first_time) {
6769                                 level = (int)ref_objectid;
6770                                 BUG_ON(level >= BTRFS_MAX_LEVEL);
6771                                 ref_path->lowest_level = level;
6772                                 ref_path->current_level = level;
6773                                 ref_path->nodes[level] = bytenr;
6774                         } else {
6775                                 WARN_ON(ref_objectid != level);
6776                         }
6777                 } else {
6778                         WARN_ON(level != -1);
6779                 }
6780                 first_time = 0;
6781
6782                 if (ref_path->lowest_level == level) {
6783                         ref_path->owner_objectid = ref_objectid;
6784                         ref_path->num_refs = btrfs_ref_num_refs(leaf, ref);
6785                 }
6786
6787                 /*
6788                  * the block is tree root or the block isn't in reference
6789                  * counted tree.
6790                  */
6791                 if (found_key.objectid == found_key.offset ||
6792                     is_cowonly_root(btrfs_ref_root(leaf, ref))) {
6793                         ref_path->root_objectid = btrfs_ref_root(leaf, ref);
6794                         ref_path->root_generation =
6795                                 btrfs_ref_generation(leaf, ref);
6796                         if (level < 0) {
6797                                 /* special reference from the tree log */
6798                                 ref_path->nodes[0] = found_key.offset;
6799                                 ref_path->current_level = 0;
6800                         }
6801                         ret = 0;
6802                         goto out;
6803                 }
6804
6805                 level++;
6806                 BUG_ON(ref_path->nodes[level] != 0);
6807                 ref_path->nodes[level] = found_key.offset;
6808                 ref_path->current_level = level;
6809
6810                 /*
6811                  * the reference was created in the running transaction,
6812                  * no need to continue walking up.
6813                  */
6814                 if (btrfs_ref_generation(leaf, ref) == trans->transid) {
6815                         ref_path->root_objectid = btrfs_ref_root(leaf, ref);
6816                         ref_path->root_generation =
6817                                 btrfs_ref_generation(leaf, ref);
6818                         ret = 0;
6819                         goto out;
6820                 }
6821
6822                 btrfs_release_path(extent_root, path);
6823                 cond_resched();
6824         }
6825         /* reached max tree level, but no tree root found. */
6826         BUG();
6827 out:
6828         btrfs_free_path(path);
6829         return ret;
6830 }
6831
6832 static int btrfs_first_ref_path(struct btrfs_trans_handle *trans,
6833                                 struct btrfs_root *extent_root,
6834                                 struct btrfs_ref_path *ref_path,
6835                                 u64 extent_start)
6836 {
6837         memset(ref_path, 0, sizeof(*ref_path));
6838         ref_path->extent_start = extent_start;
6839
6840         return __next_ref_path(trans, extent_root, ref_path, 1);
6841 }
6842
6843 static int btrfs_next_ref_path(struct btrfs_trans_handle *trans,
6844                                struct btrfs_root *extent_root,
6845                                struct btrfs_ref_path *ref_path)
6846 {
6847         return __next_ref_path(trans, extent_root, ref_path, 0);
6848 }
6849
6850 static noinline int get_new_locations(struct inode *reloc_inode,
6851                                       struct btrfs_key *extent_key,
6852                                       u64 offset, int no_fragment,
6853                                       struct disk_extent **extents,
6854                                       int *nr_extents)
6855 {
6856         struct btrfs_root *root = BTRFS_I(reloc_inode)->root;
6857         struct btrfs_path *path;
6858         struct btrfs_file_extent_item *fi;
6859         struct extent_buffer *leaf;
6860         struct disk_extent *exts = *extents;
6861         struct btrfs_key found_key;
6862         u64 cur_pos;
6863         u64 last_byte;
6864         u32 nritems;
6865         int nr = 0;
6866         int max = *nr_extents;
6867         int ret;
6868
6869         WARN_ON(!no_fragment && *extents);
6870         if (!exts) {
6871                 max = 1;
6872                 exts = kmalloc(sizeof(*exts) * max, GFP_NOFS);
6873                 if (!exts)
6874                         return -ENOMEM;
6875         }
6876
6877         path = btrfs_alloc_path();
6878         BUG_ON(!path);
6879
6880         cur_pos = extent_key->objectid - offset;
6881         last_byte = extent_key->objectid + extent_key->offset;
6882         ret = btrfs_lookup_file_extent(NULL, root, path, reloc_inode->i_ino,
6883                                        cur_pos, 0);
6884         if (ret < 0)
6885                 goto out;
6886         if (ret > 0) {
6887                 ret = -ENOENT;
6888                 goto out;
6889         }
6890
6891         while (1) {
6892                 leaf = path->nodes[0];
6893                 nritems = btrfs_header_nritems(leaf);
6894                 if (path->slots[0] >= nritems) {
6895                         ret = btrfs_next_leaf(root, path);
6896                         if (ret < 0)
6897                                 goto out;
6898                         if (ret > 0)
6899                                 break;
6900                         leaf = path->nodes[0];
6901                 }
6902
6903                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
6904                 if (found_key.offset != cur_pos ||
6905                     found_key.type != BTRFS_EXTENT_DATA_KEY ||
6906                     found_key.objectid != reloc_inode->i_ino)
6907                         break;
6908
6909                 fi = btrfs_item_ptr(leaf, path->slots[0],
6910                                     struct btrfs_file_extent_item);
6911                 if (btrfs_file_extent_type(leaf, fi) !=
6912                     BTRFS_FILE_EXTENT_REG ||
6913                     btrfs_file_extent_disk_bytenr(leaf, fi) == 0)
6914                         break;
6915
6916                 if (nr == max) {
6917                         struct disk_extent *old = exts;
6918                         max *= 2;
6919                         exts = kzalloc(sizeof(*exts) * max, GFP_NOFS);
6920                         memcpy(exts, old, sizeof(*exts) * nr);
6921                         if (old != *extents)
6922                                 kfree(old);
6923                 }
6924
6925                 exts[nr].disk_bytenr =
6926                         btrfs_file_extent_disk_bytenr(leaf, fi);
6927                 exts[nr].disk_num_bytes =
6928                         btrfs_file_extent_disk_num_bytes(leaf, fi);
6929                 exts[nr].offset = btrfs_file_extent_offset(leaf, fi);
6930                 exts[nr].num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
6931                 exts[nr].ram_bytes = btrfs_file_extent_ram_bytes(leaf, fi);
6932                 exts[nr].compression = btrfs_file_extent_compression(leaf, fi);
6933                 exts[nr].encryption = btrfs_file_extent_encryption(leaf, fi);
6934                 exts[nr].other_encoding = btrfs_file_extent_other_encoding(leaf,
6935                                                                            fi);
6936                 BUG_ON(exts[nr].offset > 0);
6937                 BUG_ON(exts[nr].compression || exts[nr].encryption);
6938                 BUG_ON(exts[nr].num_bytes != exts[nr].disk_num_bytes);
6939
6940                 cur_pos += exts[nr].num_bytes;
6941                 nr++;
6942
6943                 if (cur_pos + offset >= last_byte)
6944                         break;
6945
6946                 if (no_fragment) {
6947                         ret = 1;
6948                         goto out;
6949                 }
6950                 path->slots[0]++;
6951         }
6952
6953         BUG_ON(cur_pos + offset > last_byte);
6954         if (cur_pos + offset < last_byte) {
6955                 ret = -ENOENT;
6956                 goto out;
6957         }
6958         ret = 0;
6959 out:
6960         btrfs_free_path(path);
6961         if (ret) {
6962                 if (exts != *extents)
6963                         kfree(exts);
6964         } else {
6965                 *extents = exts;
6966                 *nr_extents = nr;
6967         }
6968         return ret;
6969 }
6970
6971 static noinline int replace_one_extent(struct btrfs_trans_handle *trans,
6972                                         struct btrfs_root *root,
6973                                         struct btrfs_path *path,
6974                                         struct btrfs_key *extent_key,
6975                                         struct btrfs_key *leaf_key,
6976                                         struct btrfs_ref_path *ref_path,
6977                                         struct disk_extent *new_extents,
6978                                         int nr_extents)
6979 {
6980         struct extent_buffer *leaf;
6981         struct btrfs_file_extent_item *fi;
6982         struct inode *inode = NULL;
6983         struct btrfs_key key;
6984         u64 lock_start = 0;
6985         u64 lock_end = 0;
6986         u64 num_bytes;
6987         u64 ext_offset;
6988         u64 search_end = (u64)-1;
6989         u32 nritems;
6990         int nr_scaned = 0;
6991         int extent_locked = 0;
6992         int extent_type;
6993         int ret;
6994
6995         memcpy(&key, leaf_key, sizeof(key));
6996         if (ref_path->owner_objectid != BTRFS_MULTIPLE_OBJECTIDS) {
6997                 if (key.objectid < ref_path->owner_objectid ||
6998                     (key.objectid == ref_path->owner_objectid &&
6999                      key.type < BTRFS_EXTENT_DATA_KEY)) {
7000                         key.objectid = ref_path->owner_objectid;
7001                         key.type = BTRFS_EXTENT_DATA_KEY;
7002                         key.offset = 0;
7003                 }
7004         }
7005
7006         while (1) {
7007                 ret = btrfs_search_slot(trans, root, &key, path, 0, 1);
7008                 if (ret < 0)
7009                         goto out;
7010
7011                 leaf = path->nodes[0];
7012                 nritems = btrfs_header_nritems(leaf);
7013 next:
7014                 if (extent_locked && ret > 0) {
7015                         /*
7016                          * the file extent item was modified by someone
7017                          * before the extent got locked.
7018                          */
7019                         unlock_extent(&BTRFS_I(inode)->io_tree, lock_start,
7020                                       lock_end, GFP_NOFS);
7021                         extent_locked = 0;
7022                 }
7023
7024                 if (path->slots[0] >= nritems) {
7025                         if (++nr_scaned > 2)
7026                                 break;
7027
7028                         BUG_ON(extent_locked);
7029                         ret = btrfs_next_leaf(root, path);
7030                         if (ret < 0)
7031                                 goto out;
7032                         if (ret > 0)
7033                                 break;
7034                         leaf = path->nodes[0];
7035                         nritems = btrfs_header_nritems(leaf);
7036                 }
7037
7038                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
7039
7040                 if (ref_path->owner_objectid != BTRFS_MULTIPLE_OBJECTIDS) {
7041                         if ((key.objectid > ref_path->owner_objectid) ||
7042                             (key.objectid == ref_path->owner_objectid &&
7043                              key.type > BTRFS_EXTENT_DATA_KEY) ||
7044                             key.offset >= search_end)
7045                                 break;
7046                 }
7047
7048                 if (inode && key.objectid != inode->i_ino) {
7049                         BUG_ON(extent_locked);
7050                         btrfs_release_path(root, path);
7051                         mutex_unlock(&inode->i_mutex);
7052                         iput(inode);
7053                         inode = NULL;
7054                         continue;
7055                 }
7056
7057                 if (key.type != BTRFS_EXTENT_DATA_KEY) {
7058                         path->slots[0]++;
7059                         ret = 1;
7060                         goto next;
7061                 }
7062                 fi = btrfs_item_ptr(leaf, path->slots[0],
7063                                     struct btrfs_file_extent_item);
7064                 extent_type = btrfs_file_extent_type(leaf, fi);
7065                 if ((extent_type != BTRFS_FILE_EXTENT_REG &&
7066                      extent_type != BTRFS_FILE_EXTENT_PREALLOC) ||
7067                     (btrfs_file_extent_disk_bytenr(leaf, fi) !=
7068                      extent_key->objectid)) {
7069                         path->slots[0]++;
7070                         ret = 1;
7071                         goto next;
7072                 }
7073
7074                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
7075                 ext_offset = btrfs_file_extent_offset(leaf, fi);
7076
7077                 if (search_end == (u64)-1) {
7078                         search_end = key.offset - ext_offset +
7079                                 btrfs_file_extent_ram_bytes(leaf, fi);
7080                 }
7081
7082                 if (!extent_locked) {
7083                         lock_start = key.offset;
7084                         lock_end = lock_start + num_bytes - 1;
7085                 } else {
7086                         if (lock_start > key.offset ||
7087                             lock_end + 1 < key.offset + num_bytes) {
7088                                 unlock_extent(&BTRFS_I(inode)->io_tree,
7089                                               lock_start, lock_end, GFP_NOFS);
7090                                 extent_locked = 0;
7091                         }
7092                 }
7093
7094                 if (!inode) {
7095                         btrfs_release_path(root, path);
7096
7097                         inode = btrfs_iget_locked(root->fs_info->sb,
7098                                                   key.objectid, root);
7099                         if (inode->i_state & I_NEW) {
7100                                 BTRFS_I(inode)->root = root;
7101                                 BTRFS_I(inode)->location.objectid =
7102                                         key.objectid;
7103                                 BTRFS_I(inode)->location.type =
7104                                         BTRFS_INODE_ITEM_KEY;
7105                                 BTRFS_I(inode)->location.offset = 0;
7106                                 btrfs_read_locked_inode(inode);
7107                                 unlock_new_inode(inode);
7108                         }
7109                         /*
7110                          * some code call btrfs_commit_transaction while
7111                          * holding the i_mutex, so we can't use mutex_lock
7112                          * here.
7113                          */
7114                         if (is_bad_inode(inode) ||
7115                             !mutex_trylock(&inode->i_mutex)) {
7116                                 iput(inode);
7117                                 inode = NULL;
7118                                 key.offset = (u64)-1;
7119                                 goto skip;
7120                         }
7121                 }
7122
7123                 if (!extent_locked) {
7124                         struct btrfs_ordered_extent *ordered;
7125
7126                         btrfs_release_path(root, path);
7127
7128                         lock_extent(&BTRFS_I(inode)->io_tree, lock_start,
7129                                     lock_end, GFP_NOFS);
7130                         ordered = btrfs_lookup_first_ordered_extent(inode,
7131                                                                     lock_end);
7132                         if (ordered &&
7133                             ordered->file_offset <= lock_end &&
7134                             ordered->file_offset + ordered->len > lock_start) {
7135                                 unlock_extent(&BTRFS_I(inode)->io_tree,
7136                                               lock_start, lock_end, GFP_NOFS);
7137                                 btrfs_start_ordered_extent(inode, ordered, 1);
7138                                 btrfs_put_ordered_extent(ordered);
7139                                 key.offset += num_bytes;
7140                                 goto skip;
7141                         }
7142                         if (ordered)
7143                                 btrfs_put_ordered_extent(ordered);
7144
7145                         extent_locked = 1;
7146                         continue;
7147                 }
7148
7149                 if (nr_extents == 1) {
7150                         /* update extent pointer in place */
7151                         btrfs_set_file_extent_disk_bytenr(leaf, fi,
7152                                                 new_extents[0].disk_bytenr);
7153                         btrfs_set_file_extent_disk_num_bytes(leaf, fi,
7154                                                 new_extents[0].disk_num_bytes);
7155                         btrfs_mark_buffer_dirty(leaf);
7156
7157                         btrfs_drop_extent_cache(inode, key.offset,
7158                                                 key.offset + num_bytes - 1, 0);
7159
7160                         ret = btrfs_inc_extent_ref(trans, root,
7161                                                 new_extents[0].disk_bytenr,
7162                                                 new_extents[0].disk_num_bytes,
7163                                                 leaf->start,
7164                                                 root->root_key.objectid,
7165                                                 trans->transid,
7166                                                 key.objectid);
7167                         BUG_ON(ret);
7168
7169                         ret = btrfs_free_extent(trans, root,
7170                                                 extent_key->objectid,
7171                                                 extent_key->offset,
7172                                                 leaf->start,
7173                                                 btrfs_header_owner(leaf),
7174                                                 btrfs_header_generation(leaf),
7175                                                 key.objectid, 0);
7176                         BUG_ON(ret);
7177
7178                         btrfs_release_path(root, path);
7179                         key.offset += num_bytes;
7180                 } else {
7181                         BUG_ON(1);
7182 #if 0
7183                         u64 alloc_hint;
7184                         u64 extent_len;
7185                         int i;
7186                         /*
7187                          * drop old extent pointer at first, then insert the
7188                          * new pointers one bye one
7189                          */
7190                         btrfs_release_path(root, path);
7191                         ret = btrfs_drop_extents(trans, root, inode, key.offset,
7192                                                  key.offset + num_bytes,
7193                                                  key.offset, &alloc_hint);
7194                         BUG_ON(ret);
7195
7196                         for (i = 0; i < nr_extents; i++) {
7197                                 if (ext_offset >= new_extents[i].num_bytes) {
7198                                         ext_offset -= new_extents[i].num_bytes;
7199                                         continue;
7200                                 }
7201                                 extent_len = min(new_extents[i].num_bytes -
7202                                                  ext_offset, num_bytes);
7203
7204                                 ret = btrfs_insert_empty_item(trans, root,
7205                                                               path, &key,
7206                                                               sizeof(*fi));
7207                                 BUG_ON(ret);
7208
7209                                 leaf = path->nodes[0];
7210                                 fi = btrfs_item_ptr(leaf, path->slots[0],
7211                                                 struct btrfs_file_extent_item);
7212                                 btrfs_set_file_extent_generation(leaf, fi,
7213                                                         trans->transid);
7214                                 btrfs_set_file_extent_type(leaf, fi,
7215                                                         BTRFS_FILE_EXTENT_REG);
7216                                 btrfs_set_file_extent_disk_bytenr(leaf, fi,
7217                                                 new_extents[i].disk_bytenr);
7218                                 btrfs_set_file_extent_disk_num_bytes(leaf, fi,
7219                                                 new_extents[i].disk_num_bytes);
7220                                 btrfs_set_file_extent_ram_bytes(leaf, fi,
7221                                                 new_extents[i].ram_bytes);
7222
7223                                 btrfs_set_file_extent_compression(leaf, fi,
7224                                                 new_extents[i].compression);
7225                                 btrfs_set_file_extent_encryption(leaf, fi,
7226                                                 new_extents[i].encryption);
7227                                 btrfs_set_file_extent_other_encoding(leaf, fi,
7228                                                 new_extents[i].other_encoding);
7229
7230                                 btrfs_set_file_extent_num_bytes(leaf, fi,
7231                                                         extent_len);
7232                                 ext_offset += new_extents[i].offset;
7233                                 btrfs_set_file_extent_offset(leaf, fi,
7234                                                         ext_offset);
7235                                 btrfs_mark_buffer_dirty(leaf);
7236
7237                                 btrfs_drop_extent_cache(inode, key.offset,
7238                                                 key.offset + extent_len - 1, 0);
7239
7240                                 ret = btrfs_inc_extent_ref(trans, root,
7241                                                 new_extents[i].disk_bytenr,
7242                                                 new_extents[i].disk_num_bytes,
7243                                                 leaf->start,
7244                                                 root->root_key.objectid,
7245                                                 trans->transid, key.objectid);
7246                                 BUG_ON(ret);
7247                                 btrfs_release_path(root, path);
7248
7249                                 inode_add_bytes(inode, extent_len);
7250
7251                                 ext_offset = 0;
7252                                 num_bytes -= extent_len;
7253                                 key.offset += extent_len;
7254
7255                                 if (num_bytes == 0)
7256                                         break;
7257                         }
7258                         BUG_ON(i >= nr_extents);
7259 #endif
7260                 }
7261
7262                 if (extent_locked) {
7263                         unlock_extent(&BTRFS_I(inode)->io_tree, lock_start,
7264                                       lock_end, GFP_NOFS);
7265                         extent_locked = 0;
7266                 }
7267 skip:
7268                 if (ref_path->owner_objectid != BTRFS_MULTIPLE_OBJECTIDS &&
7269                     key.offset >= search_end)
7270                         break;
7271
7272                 cond_resched();
7273         }
7274         ret = 0;
7275 out:
7276         btrfs_release_path(root, path);
7277         if (inode) {
7278                 mutex_unlock(&inode->i_mutex);
7279                 if (extent_locked) {
7280                         unlock_extent(&BTRFS_I(inode)->io_tree, lock_start,
7281                                       lock_end, GFP_NOFS);
7282                 }
7283                 iput(inode);
7284         }
7285         return ret;
7286 }
7287
7288 int btrfs_reloc_tree_cache_ref(struct btrfs_trans_handle *trans,
7289                                struct btrfs_root *root,
7290                                struct extent_buffer *buf, u64 orig_start)
7291 {
7292         int level;
7293         int ret;
7294
7295         BUG_ON(btrfs_header_generation(buf) != trans->transid);
7296         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7297
7298         level = btrfs_header_level(buf);
7299         if (level == 0) {
7300                 struct btrfs_leaf_ref *ref;
7301                 struct btrfs_leaf_ref *orig_ref;
7302
7303                 orig_ref = btrfs_lookup_leaf_ref(root, orig_start);
7304                 if (!orig_ref)
7305                         return -ENOENT;
7306
7307                 ref = btrfs_alloc_leaf_ref(root, orig_ref->nritems);
7308                 if (!ref) {
7309                         btrfs_free_leaf_ref(root, orig_ref);
7310                         return -ENOMEM;
7311                 }
7312
7313                 ref->nritems = orig_ref->nritems;
7314                 memcpy(ref->extents, orig_ref->extents,
7315                         sizeof(ref->extents[0]) * ref->nritems);
7316
7317                 btrfs_free_leaf_ref(root, orig_ref);
7318
7319                 ref->root_gen = trans->transid;
7320                 ref->bytenr = buf->start;
7321                 ref->owner = btrfs_header_owner(buf);
7322                 ref->generation = btrfs_header_generation(buf);
7323
7324                 ret = btrfs_add_leaf_ref(root, ref, 0);
7325                 WARN_ON(ret);
7326                 btrfs_free_leaf_ref(root, ref);
7327         }
7328         return 0;
7329 }
7330
7331 static noinline int invalidate_extent_cache(struct btrfs_root *root,
7332                                         struct extent_buffer *leaf,
7333                                         struct btrfs_block_group_cache *group,
7334                                         struct btrfs_root *target_root)
7335 {
7336         struct btrfs_key key;
7337         struct inode *inode = NULL;
7338         struct btrfs_file_extent_item *fi;
7339         struct extent_state *cached_state = NULL;
7340         u64 num_bytes;
7341         u64 skip_objectid = 0;
7342         u32 nritems;
7343         u32 i;
7344
7345         nritems = btrfs_header_nritems(leaf);
7346         for (i = 0; i < nritems; i++) {
7347                 btrfs_item_key_to_cpu(leaf, &key, i);
7348                 if (key.objectid == skip_objectid ||
7349                     key.type != BTRFS_EXTENT_DATA_KEY)
7350                         continue;
7351                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
7352                 if (btrfs_file_extent_type(leaf, fi) ==
7353                     BTRFS_FILE_EXTENT_INLINE)
7354                         continue;
7355                 if (btrfs_file_extent_disk_bytenr(leaf, fi) == 0)
7356                         continue;
7357                 if (!inode || inode->i_ino != key.objectid) {
7358                         iput(inode);
7359                         inode = btrfs_ilookup(target_root->fs_info->sb,
7360                                               key.objectid, target_root, 1);
7361                 }
7362                 if (!inode) {
7363                         skip_objectid = key.objectid;
7364                         continue;
7365                 }
7366                 num_bytes = btrfs_file_extent_num_bytes(leaf, fi);
7367
7368                 lock_extent_bits(&BTRFS_I(inode)->io_tree, key.offset,
7369                                  key.offset + num_bytes - 1, 0, &cached_state,
7370                                  GFP_NOFS);
7371                 btrfs_drop_extent_cache(inode, key.offset,
7372                                         key.offset + num_bytes - 1, 1);
7373                 unlock_extent_cached(&BTRFS_I(inode)->io_tree, key.offset,
7374                                      key.offset + num_bytes - 1, &cached_state,
7375                                      GFP_NOFS);
7376                 cond_resched();
7377         }
7378         iput(inode);
7379         return 0;
7380 }
7381
7382 static noinline int replace_extents_in_leaf(struct btrfs_trans_handle *trans,
7383                                         struct btrfs_root *root,
7384                                         struct extent_buffer *leaf,
7385                                         struct btrfs_block_group_cache *group,
7386                                         struct inode *reloc_inode)
7387 {
7388         struct btrfs_key key;
7389         struct btrfs_key extent_key;
7390         struct btrfs_file_extent_item *fi;
7391         struct btrfs_leaf_ref *ref;
7392         struct disk_extent *new_extent;
7393         u64 bytenr;
7394         u64 num_bytes;
7395         u32 nritems;
7396         u32 i;
7397         int ext_index;
7398         int nr_extent;
7399         int ret;
7400
7401         new_extent = kmalloc(sizeof(*new_extent), GFP_NOFS);
7402         BUG_ON(!new_extent);
7403
7404         ref = btrfs_lookup_leaf_ref(root, leaf->start);
7405         BUG_ON(!ref);
7406
7407         ext_index = -1;
7408         nritems = btrfs_header_nritems(leaf);
7409         for (i = 0; i < nritems; i++) {
7410                 btrfs_item_key_to_cpu(leaf, &key, i);
7411                 if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
7412                         continue;
7413                 fi = btrfs_item_ptr(leaf, i, struct btrfs_file_extent_item);
7414                 if (btrfs_file_extent_type(leaf, fi) ==
7415                     BTRFS_FILE_EXTENT_INLINE)
7416                         continue;
7417                 bytenr = btrfs_file_extent_disk_bytenr(leaf, fi);
7418                 num_bytes = btrfs_file_extent_disk_num_bytes(leaf, fi);
7419                 if (bytenr == 0)
7420                         continue;
7421
7422                 ext_index++;
7423                 if (bytenr >= group->key.objectid + group->key.offset ||
7424                     bytenr + num_bytes <= group->key.objectid)
7425                         continue;
7426
7427                 extent_key.objectid = bytenr;
7428                 extent_key.offset = num_bytes;
7429                 extent_key.type = BTRFS_EXTENT_ITEM_KEY;
7430                 nr_extent = 1;
7431                 ret = get_new_locations(reloc_inode, &extent_key,
7432                                         group->key.objectid, 1,
7433                                         &new_extent, &nr_extent);
7434                 if (ret > 0)
7435                         continue;
7436                 BUG_ON(ret < 0);
7437
7438                 BUG_ON(ref->extents[ext_index].bytenr != bytenr);
7439                 BUG_ON(ref->extents[ext_index].num_bytes != num_bytes);
7440                 ref->extents[ext_index].bytenr = new_extent->disk_bytenr;
7441                 ref->extents[ext_index].num_bytes = new_extent->disk_num_bytes;
7442
7443                 btrfs_set_file_extent_disk_bytenr(leaf, fi,
7444                                                 new_extent->disk_bytenr);
7445                 btrfs_set_file_extent_disk_num_bytes(leaf, fi,
7446                                                 new_extent->disk_num_bytes);
7447                 btrfs_mark_buffer_dirty(leaf);
7448
7449                 ret = btrfs_inc_extent_ref(trans, root,
7450                                         new_extent->disk_bytenr,
7451                                         new_extent->disk_num_bytes,
7452                                         leaf->start,
7453                                         root->root_key.objectid,
7454                                         trans->transid, key.objectid);
7455                 BUG_ON(ret);
7456
7457                 ret = btrfs_free_extent(trans, root,
7458                                         bytenr, num_bytes, leaf->start,
7459                                         btrfs_header_owner(leaf),
7460                                         btrfs_header_generation(leaf),
7461                                         key.objectid, 0);
7462                 BUG_ON(ret);
7463                 cond_resched();
7464         }
7465         kfree(new_extent);
7466         BUG_ON(ext_index + 1 != ref->nritems);
7467         btrfs_free_leaf_ref(root, ref);
7468         return 0;
7469 }
7470
7471 int btrfs_free_reloc_root(struct btrfs_trans_handle *trans,
7472                           struct btrfs_root *root)
7473 {
7474         struct btrfs_root *reloc_root;
7475         int ret;
7476
7477         if (root->reloc_root) {
7478                 reloc_root = root->reloc_root;
7479                 root->reloc_root = NULL;
7480                 list_add(&reloc_root->dead_list,
7481                          &root->fs_info->dead_reloc_roots);
7482
7483                 btrfs_set_root_bytenr(&reloc_root->root_item,
7484                                       reloc_root->node->start);
7485                 btrfs_set_root_level(&root->root_item,
7486                                      btrfs_header_level(reloc_root->node));
7487                 memset(&reloc_root->root_item.drop_progress, 0,
7488                         sizeof(struct btrfs_disk_key));
7489                 reloc_root->root_item.drop_level = 0;
7490
7491                 ret = btrfs_update_root(trans, root->fs_info->tree_root,
7492                                         &reloc_root->root_key,
7493                                         &reloc_root->root_item);
7494                 BUG_ON(ret);
7495         }
7496         return 0;
7497 }
7498
7499 int btrfs_drop_dead_reloc_roots(struct btrfs_root *root)
7500 {
7501         struct btrfs_trans_handle *trans;
7502         struct btrfs_root *reloc_root;
7503         struct btrfs_root *prev_root = NULL;
7504         struct list_head dead_roots;
7505         int ret;
7506         unsigned long nr;
7507
7508         INIT_LIST_HEAD(&dead_roots);
7509         list_splice_init(&root->fs_info->dead_reloc_roots, &dead_roots);
7510
7511         while (!list_empty(&dead_roots)) {
7512                 reloc_root = list_entry(dead_roots.prev,
7513                                         struct btrfs_root, dead_list);
7514                 list_del_init(&reloc_root->dead_list);
7515
7516                 BUG_ON(reloc_root->commit_root != NULL);
7517                 while (1) {
7518                         trans = btrfs_join_transaction(root, 1);
7519                         BUG_ON(IS_ERR(trans));
7520
7521                         mutex_lock(&root->fs_info->drop_mutex);
7522                         ret = btrfs_drop_snapshot(trans, reloc_root);
7523                         if (ret != -EAGAIN)
7524                                 break;
7525                         mutex_unlock(&root->fs_info->drop_mutex);
7526
7527                         nr = trans->blocks_used;
7528                         ret = btrfs_end_transaction(trans, root);
7529                         BUG_ON(ret);
7530                         btrfs_btree_balance_dirty(root, nr);
7531                 }
7532
7533                 free_extent_buffer(reloc_root->node);
7534
7535                 ret = btrfs_del_root(trans, root->fs_info->tree_root,
7536                                      &reloc_root->root_key);
7537                 BUG_ON(ret);
7538                 mutex_unlock(&root->fs_info->drop_mutex);
7539
7540                 nr = trans->blocks_used;
7541                 ret = btrfs_end_transaction(trans, root);
7542                 BUG_ON(ret);
7543                 btrfs_btree_balance_dirty(root, nr);
7544
7545                 kfree(prev_root);
7546                 prev_root = reloc_root;
7547         }
7548         if (prev_root) {
7549                 btrfs_remove_leaf_refs(prev_root, (u64)-1, 0);
7550                 kfree(prev_root);
7551         }
7552         return 0;
7553 }
7554
7555 int btrfs_add_dead_reloc_root(struct btrfs_root *root)
7556 {
7557         list_add(&root->dead_list, &root->fs_info->dead_reloc_roots);
7558         return 0;
7559 }
7560
7561 int btrfs_cleanup_reloc_trees(struct btrfs_root *root)
7562 {
7563         struct btrfs_root *reloc_root;
7564         struct btrfs_trans_handle *trans;
7565         struct btrfs_key location;
7566         int found;
7567         int ret;
7568
7569         mutex_lock(&root->fs_info->tree_reloc_mutex);
7570         ret = btrfs_find_dead_roots(root, BTRFS_TREE_RELOC_OBJECTID, NULL);
7571         BUG_ON(ret);
7572         found = !list_empty(&root->fs_info->dead_reloc_roots);
7573         mutex_unlock(&root->fs_info->tree_reloc_mutex);
7574
7575         if (found) {
7576                 trans = btrfs_start_transaction(root, 1);
7577                 BUG_ON(!trans);
7578                 ret = btrfs_commit_transaction(trans, root);
7579                 BUG_ON(ret);
7580         }
7581
7582         location.objectid = BTRFS_DATA_RELOC_TREE_OBJECTID;
7583         location.offset = (u64)-1;
7584         location.type = BTRFS_ROOT_ITEM_KEY;
7585
7586         reloc_root = btrfs_read_fs_root_no_name(root->fs_info, &location);
7587         BUG_ON(!reloc_root);
7588         btrfs_orphan_cleanup(reloc_root);
7589         return 0;
7590 }
7591
7592 static noinline int init_reloc_tree(struct btrfs_trans_handle *trans,
7593                                     struct btrfs_root *root)
7594 {
7595         struct btrfs_root *reloc_root;
7596         struct extent_buffer *eb;
7597         struct btrfs_root_item *root_item;
7598         struct btrfs_key root_key;
7599         int ret;
7600
7601         BUG_ON(!root->ref_cows);
7602         if (root->reloc_root)
7603                 return 0;
7604
7605         root_item = kmalloc(sizeof(*root_item), GFP_NOFS);
7606         BUG_ON(!root_item);
7607
7608         ret = btrfs_copy_root(trans, root, root->commit_root,
7609                               &eb, BTRFS_TREE_RELOC_OBJECTID);
7610         BUG_ON(ret);
7611
7612         root_key.objectid = BTRFS_TREE_RELOC_OBJECTID;
7613         root_key.offset = root->root_key.objectid;
7614         root_key.type = BTRFS_ROOT_ITEM_KEY;
7615
7616         memcpy(root_item, &root->root_item, sizeof(root_item));
7617         btrfs_set_root_refs(root_item, 0);
7618         btrfs_set_root_bytenr(root_item, eb->start);
7619         btrfs_set_root_level(root_item, btrfs_header_level(eb));
7620         btrfs_set_root_generation(root_item, trans->transid);
7621
7622         btrfs_tree_unlock(eb);
7623         free_extent_buffer(eb);
7624
7625         ret = btrfs_insert_root(trans, root->fs_info->tree_root,
7626                                 &root_key, root_item);
7627         BUG_ON(ret);
7628         kfree(root_item);
7629
7630         reloc_root = btrfs_read_fs_root_no_radix(root->fs_info->tree_root,
7631                                                  &root_key);
7632         BUG_ON(!reloc_root);
7633         reloc_root->last_trans = trans->transid;
7634         reloc_root->commit_root = NULL;
7635         reloc_root->ref_tree = &root->fs_info->reloc_ref_tree;
7636
7637         root->reloc_root = reloc_root;
7638         return 0;
7639 }
7640
7641 /*
7642  * Core function of space balance.
7643  *
7644  * The idea is using reloc trees to relocate tree blocks in reference
7645  * counted roots. There is one reloc tree for each subvol, and all
7646  * reloc trees share same root key objectid. Reloc trees are snapshots
7647  * of the latest committed roots of subvols (root->commit_root).
7648  *
7649  * To relocate a tree block referenced by a subvol, there are two steps.
7650  * COW the block through subvol's reloc tree, then update block pointer
7651  * in the subvol to point to the new block. Since all reloc trees share
7652  * same root key objectid, doing special handing for tree blocks owned
7653  * by them is easy. Once a tree block has been COWed in one reloc tree,
7654  * we can use the resulting new block directly when the same block is
7655  * required to COW again through other reloc trees. By this way, relocated
7656  * tree blocks are shared between reloc trees, so they are also shared
7657  * between subvols.
7658  */
7659 static noinline int relocate_one_path(struct btrfs_trans_handle *trans,
7660                                       struct btrfs_root *root,
7661                                       struct btrfs_path *path,
7662                                       struct btrfs_key *first_key,
7663                                       struct btrfs_ref_path *ref_path,
7664                                       struct btrfs_block_group_cache *group,
7665                                       struct inode *reloc_inode)
7666 {
7667         struct btrfs_root *reloc_root;
7668         struct extent_buffer *eb = NULL;
7669         struct btrfs_key *keys;
7670         u64 *nodes;
7671         int level;
7672         int shared_level;
7673         int lowest_level = 0;
7674         int ret;
7675
7676         if (ref_path->owner_objectid < BTRFS_FIRST_FREE_OBJECTID)
7677                 lowest_level = ref_path->owner_objectid;
7678
7679         if (!root->ref_cows) {
7680                 path->lowest_level = lowest_level;
7681                 ret = btrfs_search_slot(trans, root, first_key, path, 0, 1);
7682                 BUG_ON(ret < 0);
7683                 path->lowest_level = 0;
7684                 btrfs_release_path(root, path);
7685                 return 0;
7686         }
7687
7688         mutex_lock(&root->fs_info->tree_reloc_mutex);
7689         ret = init_reloc_tree(trans, root);
7690         BUG_ON(ret);
7691         reloc_root = root->reloc_root;
7692
7693         shared_level = ref_path->shared_level;
7694         ref_path->shared_level = BTRFS_MAX_LEVEL - 1;
7695
7696         keys = ref_path->node_keys;
7697         nodes = ref_path->new_nodes;
7698         memset(&keys[shared_level + 1], 0,
7699                sizeof(*keys) * (BTRFS_MAX_LEVEL - shared_level - 1));
7700         memset(&nodes[shared_level + 1], 0,
7701                sizeof(*nodes) * (BTRFS_MAX_LEVEL - shared_level - 1));
7702
7703         if (nodes[lowest_level] == 0) {
7704                 path->lowest_level = lowest_level;
7705                 ret = btrfs_search_slot(trans, reloc_root, first_key, path,
7706                                         0, 1);
7707                 BUG_ON(ret);
7708                 for (level = lowest_level; level < BTRFS_MAX_LEVEL; level++) {
7709                         eb = path->nodes[level];
7710                         if (!eb || eb == reloc_root->node)
7711                                 break;
7712                         nodes[level] = eb->start;
7713                         if (level == 0)
7714                                 btrfs_item_key_to_cpu(eb, &keys[level], 0);
7715                         else
7716                                 btrfs_node_key_to_cpu(eb, &keys[level], 0);
7717                 }
7718                 if (nodes[0] &&
7719                     ref_path->owner_objectid >= BTRFS_FIRST_FREE_OBJECTID) {
7720                         eb = path->nodes[0];
7721                         ret = replace_extents_in_leaf(trans, reloc_root, eb,
7722                                                       group, reloc_inode);
7723                         BUG_ON(ret);
7724                 }
7725                 btrfs_release_path(reloc_root, path);
7726         } else {
7727                 ret = btrfs_merge_path(trans, reloc_root, keys, nodes,
7728                                        lowest_level);
7729                 BUG_ON(ret);
7730         }
7731
7732         /*
7733          * replace tree blocks in the fs tree with tree blocks in
7734          * the reloc tree.
7735          */
7736         ret = btrfs_merge_path(trans, root, keys, nodes, lowest_level);
7737         BUG_ON(ret < 0);
7738
7739         if (ref_path->owner_objectid >= BTRFS_FIRST_FREE_OBJECTID) {
7740                 ret = btrfs_search_slot(trans, reloc_root, first_key, path,
7741                                         0, 0);
7742                 BUG_ON(ret);
7743                 extent_buffer_get(path->nodes[0]);
7744                 eb = path->nodes[0];
7745                 btrfs_release_path(reloc_root, path);
7746                 ret = invalidate_extent_cache(reloc_root, eb, group, root);
7747                 BUG_ON(ret);
7748                 free_extent_buffer(eb);
7749         }
7750
7751         mutex_unlock(&root->fs_info->tree_reloc_mutex);
7752         path->lowest_level = 0;
7753         return 0;
7754 }
7755
7756 static noinline int relocate_tree_block(struct btrfs_trans_handle *trans,
7757                                         struct btrfs_root *root,
7758                                         struct btrfs_path *path,
7759                                         struct btrfs_key *first_key,
7760                                         struct btrfs_ref_path *ref_path)
7761 {
7762         int ret;
7763
7764         ret = relocate_one_path(trans, root, path, first_key,
7765                                 ref_path, NULL, NULL);
7766         BUG_ON(ret);
7767
7768         return 0;
7769 }
7770
7771 static noinline int del_extent_zero(struct btrfs_trans_handle *trans,
7772                                     struct btrfs_root *extent_root,
7773                                     struct btrfs_path *path,
7774                                     struct btrfs_key *extent_key)
7775 {
7776         int ret;
7777
7778         ret = btrfs_search_slot(trans, extent_root, extent_key, path, -1, 1);
7779         if (ret)
7780                 goto out;
7781         ret = btrfs_del_item(trans, extent_root, path);
7782 out:
7783         btrfs_release_path(extent_root, path);
7784         return ret;
7785 }
7786
7787 static noinline struct btrfs_root *read_ref_root(struct btrfs_fs_info *fs_info,
7788                                                 struct btrfs_ref_path *ref_path)
7789 {
7790         struct btrfs_key root_key;
7791
7792         root_key.objectid = ref_path->root_objectid;
7793         root_key.type = BTRFS_ROOT_ITEM_KEY;
7794         if (is_cowonly_root(ref_path->root_objectid))
7795                 root_key.offset = 0;
7796         else
7797                 root_key.offset = (u64)-1;
7798
7799         return btrfs_read_fs_root_no_name(fs_info, &root_key);
7800 }
7801
7802 static noinline int relocate_one_extent(struct btrfs_root *extent_root,
7803                                         struct btrfs_path *path,
7804                                         struct btrfs_key *extent_key,
7805                                         struct btrfs_block_group_cache *group,
7806                                         struct inode *reloc_inode, int pass)
7807 {
7808         struct btrfs_trans_handle *trans;
7809         struct btrfs_root *found_root;
7810         struct btrfs_ref_path *ref_path = NULL;
7811         struct disk_extent *new_extents = NULL;
7812         int nr_extents = 0;
7813         int loops;
7814         int ret;
7815         int level;
7816         struct btrfs_key first_key;
7817         u64 prev_block = 0;
7818
7819
7820         trans = btrfs_start_transaction(extent_root, 1);
7821         BUG_ON(!trans);
7822
7823         if (extent_key->objectid == 0) {
7824                 ret = del_extent_zero(trans, extent_root, path, extent_key);
7825                 goto out;
7826         }
7827
7828         ref_path = kmalloc(sizeof(*ref_path), GFP_NOFS);
7829         if (!ref_path) {
7830                 ret = -ENOMEM;
7831                 goto out;
7832         }
7833
7834         for (loops = 0; ; loops++) {
7835                 if (loops == 0) {
7836                         ret = btrfs_first_ref_path(trans, extent_root, ref_path,
7837                                                    extent_key->objectid);
7838                 } else {
7839                         ret = btrfs_next_ref_path(trans, extent_root, ref_path);
7840                 }
7841                 if (ret < 0)
7842                         goto out;
7843                 if (ret > 0)
7844                         break;
7845
7846                 if (ref_path->root_objectid == BTRFS_TREE_LOG_OBJECTID ||
7847                     ref_path->root_objectid == BTRFS_TREE_RELOC_OBJECTID)
7848                         continue;
7849
7850                 found_root = read_ref_root(extent_root->fs_info, ref_path);
7851                 BUG_ON(!found_root);
7852                 /*
7853                  * for reference counted tree, only process reference paths
7854                  * rooted at the latest committed root.
7855                  */
7856                 if (found_root->ref_cows &&
7857                     ref_path->root_generation != found_root->root_key.offset)
7858                         continue;
7859
7860                 if (ref_path->owner_objectid >= BTRFS_FIRST_FREE_OBJECTID) {
7861                         if (pass == 0) {
7862                                 /*
7863                                  * copy data extents to new locations
7864                                  */
7865                                 u64 group_start = group->key.objectid;
7866                                 ret = relocate_data_extent(reloc_inode,
7867                                                            extent_key,
7868                                                            group_start);
7869                                 if (ret < 0)
7870                                         goto out;
7871                                 break;
7872                         }
7873                         level = 0;
7874                 } else {
7875                         level = ref_path->owner_objectid;
7876                 }
7877
7878                 if (prev_block != ref_path->nodes[level]) {
7879                         struct extent_buffer *eb;
7880                         u64 block_start = ref_path->nodes[level];
7881                         u64 block_size = btrfs_level_size(found_root, level);
7882
7883                         eb = read_tree_block(found_root, block_start,
7884                                              block_size, 0);
7885                         btrfs_tree_lock(eb);
7886                         BUG_ON(level != btrfs_header_level(eb));
7887
7888                         if (level == 0)
7889                                 btrfs_item_key_to_cpu(eb, &first_key, 0);
7890                         else
7891                                 btrfs_node_key_to_cpu(eb, &first_key, 0);
7892
7893                         btrfs_tree_unlock(eb);
7894                         free_extent_buffer(eb);
7895                         prev_block = block_start;
7896                 }
7897
7898                 mutex_lock(&extent_root->fs_info->trans_mutex);
7899                 btrfs_record_root_in_trans(found_root);
7900                 mutex_unlock(&extent_root->fs_info->trans_mutex);
7901                 if (ref_path->owner_objectid >= BTRFS_FIRST_FREE_OBJECTID) {
7902                         /*
7903                          * try to update data extent references while
7904                          * keeping metadata shared between snapshots.
7905                          */
7906                         if (pass == 1) {
7907                                 ret = relocate_one_path(trans, found_root,
7908                                                 path, &first_key, ref_path,
7909                                                 group, reloc_inode);
7910                                 if (ret < 0)
7911                                         goto out;
7912                                 continue;
7913                         }
7914                         /*
7915                          * use fallback method to process the remaining
7916                          * references.
7917                          */
7918                         if (!new_extents) {
7919                                 u64 group_start = group->key.objectid;
7920                                 new_extents = kmalloc(sizeof(*new_extents),
7921                                                       GFP_NOFS);
7922                                 nr_extents = 1;
7923                                 ret = get_new_locations(reloc_inode,
7924                                                         extent_key,
7925                                                         group_start, 1,
7926                                                         &new_extents,
7927                                                         &nr_extents);
7928                                 if (ret)
7929                                         goto out;
7930                         }
7931                         ret = replace_one_extent(trans, found_root,
7932                                                 path, extent_key,
7933                                                 &first_key, ref_path,
7934                                                 new_extents, nr_extents);
7935                 } else {
7936                         ret = relocate_tree_block(trans, found_root, path,
7937                                                   &first_key, ref_path);
7938                 }
7939                 if (ret < 0)
7940                         goto out;
7941         }
7942         ret = 0;
7943 out:
7944         btrfs_end_transaction(trans, extent_root);
7945         kfree(new_extents);
7946         kfree(ref_path);
7947         return ret;
7948 }
7949 #endif
7950
7951 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7952 {
7953         u64 num_devices;
7954         u64 stripped = BTRFS_BLOCK_GROUP_RAID0 |
7955                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7956
7957         /*
7958          * we add in the count of missing devices because we want
7959          * to make sure that any RAID levels on a degraded FS
7960          * continue to be honored.
7961          */
7962         num_devices = root->fs_info->fs_devices->rw_devices +
7963                 root->fs_info->fs_devices->missing_devices;
7964
7965         if (num_devices == 1) {
7966                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7967                 stripped = flags & ~stripped;
7968
7969                 /* turn raid0 into single device chunks */
7970                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7971                         return stripped;
7972
7973                 /* turn mirroring into duplication */
7974                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7975                              BTRFS_BLOCK_GROUP_RAID10))
7976                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7977                 return flags;
7978         } else {
7979                 /* they already had raid on here, just return */
7980                 if (flags & stripped)
7981                         return flags;
7982
7983                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7984                 stripped = flags & ~stripped;
7985
7986                 /* switch duplicated blocks with raid1 */
7987                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7988                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7989
7990                 /* turn single device chunks into raid0 */
7991                 return stripped | BTRFS_BLOCK_GROUP_RAID0;
7992         }
7993         return flags;
7994 }
7995
7996 static int set_block_group_ro(struct btrfs_block_group_cache *cache)
7997 {
7998         struct btrfs_space_info *sinfo = cache->space_info;
7999         u64 num_bytes;
8000         int ret = -ENOSPC;
8001
8002         if (cache->ro)
8003                 return 0;
8004
8005         spin_lock(&sinfo->lock);
8006         spin_lock(&cache->lock);
8007         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8008                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8009
8010         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8011             sinfo->bytes_may_use + sinfo->bytes_readonly +
8012             cache->reserved_pinned + num_bytes <= sinfo->total_bytes) {
8013                 sinfo->bytes_readonly += num_bytes;
8014                 sinfo->bytes_reserved += cache->reserved_pinned;
8015                 cache->reserved_pinned = 0;
8016                 cache->ro = 1;
8017                 ret = 0;
8018         }
8019
8020         spin_unlock(&cache->lock);
8021         spin_unlock(&sinfo->lock);
8022         return ret;
8023 }
8024
8025 int btrfs_set_block_group_ro(struct btrfs_root *root,
8026                              struct btrfs_block_group_cache *cache)
8027
8028 {
8029         struct btrfs_trans_handle *trans;
8030         u64 alloc_flags;
8031         int ret;
8032
8033         BUG_ON(cache->ro);
8034
8035         trans = btrfs_join_transaction(root, 1);
8036         BUG_ON(IS_ERR(trans));
8037
8038         alloc_flags = update_block_group_flags(root, cache->flags);
8039         if (alloc_flags != cache->flags)
8040                 do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags, 1);
8041
8042         ret = set_block_group_ro(cache);
8043         if (!ret)
8044                 goto out;
8045         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
8046         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags, 1);
8047         if (ret < 0)
8048                 goto out;
8049         ret = set_block_group_ro(cache);
8050 out:
8051         btrfs_end_transaction(trans, root);
8052         return ret;
8053 }
8054
8055 /*
8056  * helper to account the unused space of all the readonly block group in the
8057  * list. takes mirrors into account.
8058  */
8059 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
8060 {
8061         struct btrfs_block_group_cache *block_group;
8062         u64 free_bytes = 0;
8063         int factor;
8064
8065         list_for_each_entry(block_group, groups_list, list) {
8066                 spin_lock(&block_group->lock);
8067
8068                 if (!block_group->ro) {
8069                         spin_unlock(&block_group->lock);
8070                         continue;
8071                 }
8072
8073                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
8074                                           BTRFS_BLOCK_GROUP_RAID10 |
8075                                           BTRFS_BLOCK_GROUP_DUP))
8076                         factor = 2;
8077                 else
8078                         factor = 1;
8079
8080                 free_bytes += (block_group->key.offset -
8081                                btrfs_block_group_used(&block_group->item)) *
8082                                factor;
8083
8084                 spin_unlock(&block_group->lock);
8085         }
8086
8087         return free_bytes;
8088 }
8089
8090 /*
8091  * helper to account the unused space of all the readonly block group in the
8092  * space_info. takes mirrors into account.
8093  */
8094 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
8095 {
8096         int i;
8097         u64 free_bytes = 0;
8098
8099         spin_lock(&sinfo->lock);
8100
8101         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
8102                 if (!list_empty(&sinfo->block_groups[i]))
8103                         free_bytes += __btrfs_get_ro_block_group_free_space(
8104                                                 &sinfo->block_groups[i]);
8105
8106         spin_unlock(&sinfo->lock);
8107
8108         return free_bytes;
8109 }
8110
8111 int btrfs_set_block_group_rw(struct btrfs_root *root,
8112                               struct btrfs_block_group_cache *cache)
8113 {
8114         struct btrfs_space_info *sinfo = cache->space_info;
8115         u64 num_bytes;
8116
8117         BUG_ON(!cache->ro);
8118
8119         spin_lock(&sinfo->lock);
8120         spin_lock(&cache->lock);
8121         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8122                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8123         sinfo->bytes_readonly -= num_bytes;
8124         cache->ro = 0;
8125         spin_unlock(&cache->lock);
8126         spin_unlock(&sinfo->lock);
8127         return 0;
8128 }
8129
8130 /*
8131  * checks to see if its even possible to relocate this block group.
8132  *
8133  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
8134  * ok to go ahead and try.
8135  */
8136 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
8137 {
8138         struct btrfs_block_group_cache *block_group;
8139         struct btrfs_space_info *space_info;
8140         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
8141         struct btrfs_device *device;
8142         int full = 0;
8143         int ret = 0;
8144
8145         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
8146
8147         /* odd, couldn't find the block group, leave it alone */
8148         if (!block_group)
8149                 return -1;
8150
8151         /* no bytes used, we're good */
8152         if (!btrfs_block_group_used(&block_group->item))
8153                 goto out;
8154
8155         space_info = block_group->space_info;
8156         spin_lock(&space_info->lock);
8157
8158         full = space_info->full;
8159
8160         /*
8161          * if this is the last block group we have in this space, we can't
8162          * relocate it unless we're able to allocate a new chunk below.
8163          *
8164          * Otherwise, we need to make sure we have room in the space to handle
8165          * all of the extents from this block group.  If we can, we're good
8166          */
8167         if ((space_info->total_bytes != block_group->key.offset) &&
8168            (space_info->bytes_used + space_info->bytes_reserved +
8169             space_info->bytes_pinned + space_info->bytes_readonly +
8170             btrfs_block_group_used(&block_group->item) <
8171             space_info->total_bytes)) {
8172                 spin_unlock(&space_info->lock);
8173                 goto out;
8174         }
8175         spin_unlock(&space_info->lock);
8176
8177         /*
8178          * ok we don't have enough space, but maybe we have free space on our
8179          * devices to allocate new chunks for relocation, so loop through our
8180          * alloc devices and guess if we have enough space.  However, if we
8181          * were marked as full, then we know there aren't enough chunks, and we
8182          * can just return.
8183          */
8184         ret = -1;
8185         if (full)
8186                 goto out;
8187
8188         mutex_lock(&root->fs_info->chunk_mutex);
8189         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
8190                 u64 min_free = btrfs_block_group_used(&block_group->item);
8191                 u64 dev_offset;
8192
8193                 /*
8194                  * check to make sure we can actually find a chunk with enough
8195                  * space to fit our block group in.
8196                  */
8197                 if (device->total_bytes > device->bytes_used + min_free) {
8198                         ret = find_free_dev_extent(NULL, device, min_free,
8199                                                    &dev_offset, NULL);
8200                         if (!ret)
8201                                 break;
8202                         ret = -1;
8203                 }
8204         }
8205         mutex_unlock(&root->fs_info->chunk_mutex);
8206 out:
8207         btrfs_put_block_group(block_group);
8208         return ret;
8209 }
8210
8211 static int find_first_block_group(struct btrfs_root *root,
8212                 struct btrfs_path *path, struct btrfs_key *key)
8213 {
8214         int ret = 0;
8215         struct btrfs_key found_key;
8216         struct extent_buffer *leaf;
8217         int slot;
8218
8219         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
8220         if (ret < 0)
8221                 goto out;
8222
8223         while (1) {
8224                 slot = path->slots[0];
8225                 leaf = path->nodes[0];
8226                 if (slot >= btrfs_header_nritems(leaf)) {
8227                         ret = btrfs_next_leaf(root, path);
8228                         if (ret == 0)
8229                                 continue;
8230                         if (ret < 0)
8231                                 goto out;
8232                         break;
8233                 }
8234                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
8235
8236                 if (found_key.objectid >= key->objectid &&
8237                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
8238                         ret = 0;
8239                         goto out;
8240                 }
8241                 path->slots[0]++;
8242         }
8243 out:
8244         return ret;
8245 }
8246
8247 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
8248 {
8249         struct btrfs_block_group_cache *block_group;
8250         u64 last = 0;
8251
8252         while (1) {
8253                 struct inode *inode;
8254
8255                 block_group = btrfs_lookup_first_block_group(info, last);
8256                 while (block_group) {
8257                         spin_lock(&block_group->lock);
8258                         if (block_group->iref)
8259                                 break;
8260                         spin_unlock(&block_group->lock);
8261                         block_group = next_block_group(info->tree_root,
8262                                                        block_group);
8263                 }
8264                 if (!block_group) {
8265                         if (last == 0)
8266                                 break;
8267                         last = 0;
8268                         continue;
8269                 }
8270
8271                 inode = block_group->inode;
8272                 block_group->iref = 0;
8273                 block_group->inode = NULL;
8274                 spin_unlock(&block_group->lock);
8275                 iput(inode);
8276                 last = block_group->key.objectid + block_group->key.offset;
8277                 btrfs_put_block_group(block_group);
8278         }
8279 }
8280
8281 int btrfs_free_block_groups(struct btrfs_fs_info *info)
8282 {
8283         struct btrfs_block_group_cache *block_group;
8284         struct btrfs_space_info *space_info;
8285         struct btrfs_caching_control *caching_ctl;
8286         struct rb_node *n;
8287
8288         down_write(&info->extent_commit_sem);
8289         while (!list_empty(&info->caching_block_groups)) {
8290                 caching_ctl = list_entry(info->caching_block_groups.next,
8291                                          struct btrfs_caching_control, list);
8292                 list_del(&caching_ctl->list);
8293                 put_caching_control(caching_ctl);
8294         }
8295         up_write(&info->extent_commit_sem);
8296
8297         spin_lock(&info->block_group_cache_lock);
8298         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
8299                 block_group = rb_entry(n, struct btrfs_block_group_cache,
8300                                        cache_node);
8301                 rb_erase(&block_group->cache_node,
8302                          &info->block_group_cache_tree);
8303                 spin_unlock(&info->block_group_cache_lock);
8304
8305                 down_write(&block_group->space_info->groups_sem);
8306                 list_del(&block_group->list);
8307                 up_write(&block_group->space_info->groups_sem);
8308
8309                 if (block_group->cached == BTRFS_CACHE_STARTED)
8310                         wait_block_group_cache_done(block_group);
8311
8312                 btrfs_remove_free_space_cache(block_group);
8313                 btrfs_put_block_group(block_group);
8314
8315                 spin_lock(&info->block_group_cache_lock);
8316         }
8317         spin_unlock(&info->block_group_cache_lock);
8318
8319         /* now that all the block groups are freed, go through and
8320          * free all the space_info structs.  This is only called during
8321          * the final stages of unmount, and so we know nobody is
8322          * using them.  We call synchronize_rcu() once before we start,
8323          * just to be on the safe side.
8324          */
8325         synchronize_rcu();
8326
8327         release_global_block_rsv(info);
8328
8329         while(!list_empty(&info->space_info)) {
8330                 space_info = list_entry(info->space_info.next,
8331                                         struct btrfs_space_info,
8332                                         list);
8333                 if (space_info->bytes_pinned > 0 ||
8334                     space_info->bytes_reserved > 0) {
8335                         WARN_ON(1);
8336                         dump_space_info(space_info, 0, 0);
8337                 }
8338                 list_del(&space_info->list);
8339                 kfree(space_info);
8340         }
8341         return 0;
8342 }
8343
8344 static void __link_block_group(struct btrfs_space_info *space_info,
8345                                struct btrfs_block_group_cache *cache)
8346 {
8347         int index = get_block_group_index(cache);
8348
8349         down_write(&space_info->groups_sem);
8350         list_add_tail(&cache->list, &space_info->block_groups[index]);
8351         up_write(&space_info->groups_sem);
8352 }
8353
8354 int btrfs_read_block_groups(struct btrfs_root *root)
8355 {
8356         struct btrfs_path *path;
8357         int ret;
8358         struct btrfs_block_group_cache *cache;
8359         struct btrfs_fs_info *info = root->fs_info;
8360         struct btrfs_space_info *space_info;
8361         struct btrfs_key key;
8362         struct btrfs_key found_key;
8363         struct extent_buffer *leaf;
8364         int need_clear = 0;
8365         u64 cache_gen;
8366
8367         root = info->extent_root;
8368         key.objectid = 0;
8369         key.offset = 0;
8370         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
8371         path = btrfs_alloc_path();
8372         if (!path)
8373                 return -ENOMEM;
8374
8375         cache_gen = btrfs_super_cache_generation(&root->fs_info->super_copy);
8376         if (cache_gen != 0 &&
8377             btrfs_super_generation(&root->fs_info->super_copy) != cache_gen)
8378                 need_clear = 1;
8379         if (btrfs_test_opt(root, CLEAR_CACHE))
8380                 need_clear = 1;
8381         if (!btrfs_test_opt(root, SPACE_CACHE) && cache_gen)
8382                 printk(KERN_INFO "btrfs: disk space caching is enabled\n");
8383
8384         while (1) {
8385                 ret = find_first_block_group(root, path, &key);
8386                 if (ret > 0)
8387                         break;
8388                 if (ret != 0)
8389                         goto error;
8390                 leaf = path->nodes[0];
8391                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
8392                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
8393                 if (!cache) {
8394                         ret = -ENOMEM;
8395                         goto error;
8396                 }
8397
8398                 atomic_set(&cache->count, 1);
8399                 spin_lock_init(&cache->lock);
8400                 spin_lock_init(&cache->tree_lock);
8401                 cache->fs_info = info;
8402                 INIT_LIST_HEAD(&cache->list);
8403                 INIT_LIST_HEAD(&cache->cluster_list);
8404
8405                 if (need_clear)
8406                         cache->disk_cache_state = BTRFS_DC_CLEAR;
8407
8408                 /*
8409                  * we only want to have 32k of ram per block group for keeping
8410                  * track of free space, and if we pass 1/2 of that we want to
8411                  * start converting things over to using bitmaps
8412                  */
8413                 cache->extents_thresh = ((1024 * 32) / 2) /
8414                         sizeof(struct btrfs_free_space);
8415
8416                 read_extent_buffer(leaf, &cache->item,
8417                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
8418                                    sizeof(cache->item));
8419                 memcpy(&cache->key, &found_key, sizeof(found_key));
8420
8421                 key.objectid = found_key.objectid + found_key.offset;
8422                 btrfs_release_path(root, path);
8423                 cache->flags = btrfs_block_group_flags(&cache->item);
8424                 cache->sectorsize = root->sectorsize;
8425
8426                 /*
8427                  * check for two cases, either we are full, and therefore
8428                  * don't need to bother with the caching work since we won't
8429                  * find any space, or we are empty, and we can just add all
8430                  * the space in and be done with it.  This saves us _alot_ of
8431                  * time, particularly in the full case.
8432                  */
8433                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
8434                         exclude_super_stripes(root, cache);
8435                         cache->last_byte_to_unpin = (u64)-1;
8436                         cache->cached = BTRFS_CACHE_FINISHED;
8437                         free_excluded_extents(root, cache);
8438                 } else if (btrfs_block_group_used(&cache->item) == 0) {
8439                         exclude_super_stripes(root, cache);
8440                         cache->last_byte_to_unpin = (u64)-1;
8441                         cache->cached = BTRFS_CACHE_FINISHED;
8442                         add_new_free_space(cache, root->fs_info,
8443                                            found_key.objectid,
8444                                            found_key.objectid +
8445                                            found_key.offset);
8446                         free_excluded_extents(root, cache);
8447                 }
8448
8449                 ret = update_space_info(info, cache->flags, found_key.offset,
8450                                         btrfs_block_group_used(&cache->item),
8451                                         &space_info);
8452                 BUG_ON(ret);
8453                 cache->space_info = space_info;
8454                 spin_lock(&cache->space_info->lock);
8455                 cache->space_info->bytes_readonly += cache->bytes_super;
8456                 spin_unlock(&cache->space_info->lock);
8457
8458                 __link_block_group(space_info, cache);
8459
8460                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
8461                 BUG_ON(ret);
8462
8463                 set_avail_alloc_bits(root->fs_info, cache->flags);
8464                 if (btrfs_chunk_readonly(root, cache->key.objectid))
8465                         set_block_group_ro(cache);
8466         }
8467
8468         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
8469                 if (!(get_alloc_profile(root, space_info->flags) &
8470                       (BTRFS_BLOCK_GROUP_RAID10 |
8471                        BTRFS_BLOCK_GROUP_RAID1 |
8472                        BTRFS_BLOCK_GROUP_DUP)))
8473                         continue;
8474                 /*
8475                  * avoid allocating from un-mirrored block group if there are
8476                  * mirrored block groups.
8477                  */
8478                 list_for_each_entry(cache, &space_info->block_groups[3], list)
8479                         set_block_group_ro(cache);
8480                 list_for_each_entry(cache, &space_info->block_groups[4], list)
8481                         set_block_group_ro(cache);
8482         }
8483
8484         init_global_block_rsv(info);
8485         ret = 0;
8486 error:
8487         btrfs_free_path(path);
8488         return ret;
8489 }
8490
8491 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
8492                            struct btrfs_root *root, u64 bytes_used,
8493                            u64 type, u64 chunk_objectid, u64 chunk_offset,
8494                            u64 size)
8495 {
8496         int ret;
8497         struct btrfs_root *extent_root;
8498         struct btrfs_block_group_cache *cache;
8499
8500         extent_root = root->fs_info->extent_root;
8501
8502         root->fs_info->last_trans_log_full_commit = trans->transid;
8503
8504         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8505         if (!cache)
8506                 return -ENOMEM;
8507
8508         cache->key.objectid = chunk_offset;
8509         cache->key.offset = size;
8510         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8511         cache->sectorsize = root->sectorsize;
8512         cache->fs_info = root->fs_info;
8513
8514         /*
8515          * we only want to have 32k of ram per block group for keeping track
8516          * of free space, and if we pass 1/2 of that we want to start
8517          * converting things over to using bitmaps
8518          */
8519         cache->extents_thresh = ((1024 * 32) / 2) /
8520                 sizeof(struct btrfs_free_space);
8521         atomic_set(&cache->count, 1);
8522         spin_lock_init(&cache->lock);
8523         spin_lock_init(&cache->tree_lock);
8524         INIT_LIST_HEAD(&cache->list);
8525         INIT_LIST_HEAD(&cache->cluster_list);
8526
8527         btrfs_set_block_group_used(&cache->item, bytes_used);
8528         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
8529         cache->flags = type;
8530         btrfs_set_block_group_flags(&cache->item, type);
8531
8532         cache->last_byte_to_unpin = (u64)-1;
8533         cache->cached = BTRFS_CACHE_FINISHED;
8534         exclude_super_stripes(root, cache);
8535
8536         add_new_free_space(cache, root->fs_info, chunk_offset,
8537                            chunk_offset + size);
8538
8539         free_excluded_extents(root, cache);
8540
8541         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
8542                                 &cache->space_info);
8543         BUG_ON(ret);
8544
8545         spin_lock(&cache->space_info->lock);
8546         cache->space_info->bytes_readonly += cache->bytes_super;
8547         spin_unlock(&cache->space_info->lock);
8548
8549         __link_block_group(cache->space_info, cache);
8550
8551         ret = btrfs_add_block_group_cache(root->fs_info, cache);
8552         BUG_ON(ret);
8553
8554         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
8555                                 sizeof(cache->item));
8556         BUG_ON(ret);
8557
8558         set_avail_alloc_bits(extent_root->fs_info, type);
8559
8560         return 0;
8561 }
8562
8563 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
8564                              struct btrfs_root *root, u64 group_start)
8565 {
8566         struct btrfs_path *path;
8567         struct btrfs_block_group_cache *block_group;
8568         struct btrfs_free_cluster *cluster;
8569         struct btrfs_root *tree_root = root->fs_info->tree_root;
8570         struct btrfs_key key;
8571         struct inode *inode;
8572         int ret;
8573         int factor;
8574
8575         root = root->fs_info->extent_root;
8576
8577         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
8578         BUG_ON(!block_group);
8579         BUG_ON(!block_group->ro);
8580
8581         memcpy(&key, &block_group->key, sizeof(key));
8582         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
8583                                   BTRFS_BLOCK_GROUP_RAID1 |
8584                                   BTRFS_BLOCK_GROUP_RAID10))
8585                 factor = 2;
8586         else
8587                 factor = 1;
8588
8589         /* make sure this block group isn't part of an allocation cluster */
8590         cluster = &root->fs_info->data_alloc_cluster;
8591         spin_lock(&cluster->refill_lock);
8592         btrfs_return_cluster_to_free_space(block_group, cluster);
8593         spin_unlock(&cluster->refill_lock);
8594
8595         /*
8596          * make sure this block group isn't part of a metadata
8597          * allocation cluster
8598          */
8599         cluster = &root->fs_info->meta_alloc_cluster;
8600         spin_lock(&cluster->refill_lock);
8601         btrfs_return_cluster_to_free_space(block_group, cluster);
8602         spin_unlock(&cluster->refill_lock);
8603
8604         path = btrfs_alloc_path();
8605         BUG_ON(!path);
8606
8607         inode = lookup_free_space_inode(root, block_group, path);
8608         if (!IS_ERR(inode)) {
8609                 btrfs_orphan_add(trans, inode);
8610                 clear_nlink(inode);
8611                 /* One for the block groups ref */
8612                 spin_lock(&block_group->lock);
8613                 if (block_group->iref) {
8614                         block_group->iref = 0;
8615                         block_group->inode = NULL;
8616                         spin_unlock(&block_group->lock);
8617                         iput(inode);
8618                 } else {
8619                         spin_unlock(&block_group->lock);
8620                 }
8621                 /* One for our lookup ref */
8622                 iput(inode);
8623         }
8624
8625         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
8626         key.offset = block_group->key.objectid;
8627         key.type = 0;
8628
8629         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
8630         if (ret < 0)
8631                 goto out;
8632         if (ret > 0)
8633                 btrfs_release_path(tree_root, path);
8634         if (ret == 0) {
8635                 ret = btrfs_del_item(trans, tree_root, path);
8636                 if (ret)
8637                         goto out;
8638                 btrfs_release_path(tree_root, path);
8639         }
8640
8641         spin_lock(&root->fs_info->block_group_cache_lock);
8642         rb_erase(&block_group->cache_node,
8643                  &root->fs_info->block_group_cache_tree);
8644         spin_unlock(&root->fs_info->block_group_cache_lock);
8645
8646         down_write(&block_group->space_info->groups_sem);
8647         /*
8648          * we must use list_del_init so people can check to see if they
8649          * are still on the list after taking the semaphore
8650          */
8651         list_del_init(&block_group->list);
8652         up_write(&block_group->space_info->groups_sem);
8653
8654         if (block_group->cached == BTRFS_CACHE_STARTED)
8655                 wait_block_group_cache_done(block_group);
8656
8657         btrfs_remove_free_space_cache(block_group);
8658
8659         spin_lock(&block_group->space_info->lock);
8660         block_group->space_info->total_bytes -= block_group->key.offset;
8661         block_group->space_info->bytes_readonly -= block_group->key.offset;
8662         block_group->space_info->disk_total -= block_group->key.offset * factor;
8663         spin_unlock(&block_group->space_info->lock);
8664
8665         memcpy(&key, &block_group->key, sizeof(key));
8666
8667         btrfs_clear_space_info_full(root->fs_info);
8668
8669         btrfs_put_block_group(block_group);
8670         btrfs_put_block_group(block_group);
8671
8672         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8673         if (ret > 0)
8674                 ret = -EIO;
8675         if (ret < 0)
8676                 goto out;
8677
8678         ret = btrfs_del_item(trans, root, path);
8679 out:
8680         btrfs_free_path(path);
8681         return ret;
8682 }
8683
8684 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8685 {
8686         return unpin_extent_range(root, start, end);
8687 }
8688
8689 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8690                                u64 num_bytes)
8691 {
8692         return btrfs_discard_extent(root, bytenr, num_bytes);
8693 }