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