fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include "compat.h"
  27 #include "hash.h"
  28 #include "ctree.h"
  29 #include "disk-io.h"
  30 #include "print-tree.h"
  31 #include "transaction.h"
  32 #include "volumes.h"
  33 #include "locking.h"
  34 #include "free-space-cache.h"
  35
  36 /* control flags for do_chunk_alloc's force field
  37  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  38  * if we really need one.
  39  *
  40  * CHUNK_ALLOC_FORCE means it must try to allocate one
  41  *
  42  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  43  * if we have very few chunks already allocated.  This is
  44  * used as part of the clustering code to help make sure
  45  * we have a good pool of storage to cluster in, without
  46  * filling the FS with empty chunks
  47  *
  48  */
  49 enum {
  50         CHUNK_ALLOC_NO_FORCE = 0,
  51         CHUNK_ALLOC_FORCE = 1,
  52         CHUNK_ALLOC_LIMITED = 2,
  53 };
  54
  55 /*
  56  * Control how reservations are dealt with.
  57  *
  58  * RESERVE_FREE - freeing a reservation.
  59  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  60  *   ENOSPC accounting
  61  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  62  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  63  */
  64 enum {
  65         RESERVE_FREE = 0,
  66         RESERVE_ALLOC = 1,
  67         RESERVE_ALLOC_NO_ACCOUNT = 2,
  68 };
  69
  70 static int update_block_group(struct btrfs_trans_handle *trans,
  71                               struct btrfs_root *root,
  72                               u64 bytenr, u64 num_bytes, int alloc);
  73 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  74                                 struct btrfs_root *root,
  75                                 u64 bytenr, u64 num_bytes, u64 parent,
  76                                 u64 root_objectid, u64 owner_objectid,
  77                                 u64 owner_offset, int refs_to_drop,
  78                                 struct btrfs_delayed_extent_op *extra_op);
  79 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  80                                     struct extent_buffer *leaf,
  81                                     struct btrfs_extent_item *ei);
  82 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  83                                       struct btrfs_root *root,
  84                                       u64 parent, u64 root_objectid,
  85                                       u64 flags, u64 owner, u64 offset,
  86                                       struct btrfs_key *ins, int ref_mod);
  87 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  88                                      struct btrfs_root *root,
  89                                      u64 parent, u64 root_objectid,
  90                                      u64 flags, struct btrfs_disk_key *key,
  91                                      int level, struct btrfs_key *ins);
  92 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  93                           struct btrfs_root *extent_root, u64 alloc_bytes,
  94                           u64 flags, int force);
  95 static int find_next_key(struct btrfs_path *path, int level,
  96                          struct btrfs_key *key);
  97 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
  98                             int dump_block_groups);
  99 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 100                                        u64 num_bytes, int reserve);
 101
 102 static noinline int
 103 block_group_cache_done(struct btrfs_block_group_cache *cache)
 104 {
 105         smp_mb();
 106         return cache->cached == BTRFS_CACHE_FINISHED;
 107 }
 108
 109 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 110 {
 111         return (cache->flags & bits) == bits;
 112 }
 113
 114 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 115 {
 116         atomic_inc(&cache->count);
 117 }
 118
 119 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 120 {
 121         if (atomic_dec_and_test(&cache->count)) {
 122                 WARN_ON(cache->pinned > 0);
 123                 WARN_ON(cache->reserved > 0);
 124                 kfree(cache->free_space_ctl);
 125                 kfree(cache);
 126         }
 127 }
 128
 129 /*
 130  * this adds the block group to the fs_info rb tree for the block group
 131  * cache
 132  */
 133 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 134                                 struct btrfs_block_group_cache *block_group)
 135 {
 136         struct rb_node **p;
 137         struct rb_node *parent = NULL;
 138         struct btrfs_block_group_cache *cache;
 139
 140         spin_lock(&info->block_group_cache_lock);
 141         p = &info->block_group_cache_tree.rb_node;
 142
 143         while (*p) {
 144                 parent = *p;
 145                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 146                                  cache_node);
 147                 if (block_group->key.objectid < cache->key.objectid) {
 148                         p = &(*p)->rb_left;
 149                 } else if (block_group->key.objectid > cache->key.objectid) {
 150                         p = &(*p)->rb_right;
 151                 } else {
 152                         spin_unlock(&info->block_group_cache_lock);
 153                         return -EEXIST;
 154                 }
 155         }
 156
 157         rb_link_node(&block_group->cache_node, parent, p);
 158         rb_insert_color(&block_group->cache_node,
 159                         &info->block_group_cache_tree);
 160         spin_unlock(&info->block_group_cache_lock);
 161
 162         return 0;
 163 }
 164
 165 /*
 166  * This will return the block group at or after bytenr if contains is 0, else
 167  * it will return the block group that contains the bytenr
 168  */
 169 static struct btrfs_block_group_cache *
 170 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 171                               int contains)
 172 {
 173         struct btrfs_block_group_cache *cache, *ret = NULL;
 174         struct rb_node *n;
 175         u64 end, start;
 176
 177         spin_lock(&info->block_group_cache_lock);
 178         n = info->block_group_cache_tree.rb_node;
 179
 180         while (n) {
 181                 cache = rb_entry(n, struct btrfs_block_group_cache,
 182                                  cache_node);
 183                 end = cache->key.objectid + cache->key.offset - 1;
 184                 start = cache->key.objectid;
 185
 186                 if (bytenr < start) {
 187                         if (!contains && (!ret || start < ret->key.objectid))
 188                                 ret = cache;
 189                         n = n->rb_left;
 190                 } else if (bytenr > start) {
 191                         if (contains && bytenr <= end) {
 192                                 ret = cache;
 193                                 break;
 194                         }
 195                         n = n->rb_right;
 196                 } else {
 197                         ret = cache;
 198                         break;
 199                 }
 200         }
 201         if (ret)
 202                 btrfs_get_block_group(ret);
 203         spin_unlock(&info->block_group_cache_lock);
 204
 205         return ret;
 206 }
 207
 208 static int add_excluded_extent(struct btrfs_root *root,
 209                                u64 start, u64 num_bytes)
 210 {
 211         u64 end = start + num_bytes - 1;
 212         set_extent_bits(&root->fs_info->freed_extents[0],
 213                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 214         set_extent_bits(&root->fs_info->freed_extents[1],
 215                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 216         return 0;
 217 }
 218
 219 static void free_excluded_extents(struct btrfs_root *root,
 220                                   struct btrfs_block_group_cache *cache)
 221 {
 222         u64 start, end;
 223
 224         start = cache->key.objectid;
 225         end = start + cache->key.offset - 1;
 226
 227         clear_extent_bits(&root->fs_info->freed_extents[0],
 228                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 229         clear_extent_bits(&root->fs_info->freed_extents[1],
 230                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 231 }
 232
 233 static int exclude_super_stripes(struct btrfs_root *root,
 234                                  struct btrfs_block_group_cache *cache)
 235 {
 236         u64 bytenr;
 237         u64 *logical;
 238         int stripe_len;
 239         int i, nr, ret;
 240
 241         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 242                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 243                 cache->bytes_super += stripe_len;
 244                 ret = add_excluded_extent(root, cache->key.objectid,
 245                                           stripe_len);
 246                 BUG_ON(ret);
 247         }
 248
 249         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 250                 bytenr = btrfs_sb_offset(i);
 251                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 252                                        cache->key.objectid, bytenr,
 253                                        0, &logical, &nr, &stripe_len);
 254                 BUG_ON(ret);
 255
 256                 while (nr--) {
 257                         cache->bytes_super += stripe_len;
 258                         ret = add_excluded_extent(root, logical[nr],
 259                                                   stripe_len);
 260                         BUG_ON(ret);
 261                 }
 262
 263                 kfree(logical);
 264         }
 265         return 0;
 266 }
 267
 268 static struct btrfs_caching_control *
 269 get_caching_control(struct btrfs_block_group_cache *cache)
 270 {
 271         struct btrfs_caching_control *ctl;
 272
 273         spin_lock(&cache->lock);
 274         if (cache->cached != BTRFS_CACHE_STARTED) {
 275                 spin_unlock(&cache->lock);
 276                 return NULL;
 277         }
 278
 279         /* We're loading it the fast way, so we don't have a caching_ctl. */
 280         if (!cache->caching_ctl) {
 281                 spin_unlock(&cache->lock);
 282                 return NULL;
 283         }
 284
 285         ctl = cache->caching_ctl;
 286         atomic_inc(&ctl->count);
 287         spin_unlock(&cache->lock);
 288         return ctl;
 289 }
 290
 291 static void put_caching_control(struct btrfs_caching_control *ctl)
 292 {
 293         if (atomic_dec_and_test(&ctl->count))
 294                 kfree(ctl);
 295 }
 296
 297 /*
 298  * this is only called by cache_block_group, since we could have freed extents
 299  * we need to check the pinned_extents for any extents that can't be used yet
 300  * since their free space will be released as soon as the transaction commits.
 301  */
 302 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 303                               struct btrfs_fs_info *info, u64 start, u64 end)
 304 {
 305         u64 extent_start, extent_end, size, total_added = 0;
 306         int ret;
 307
 308         while (start < end) {
 309                 ret = find_first_extent_bit(info->pinned_extents, start,
 310                                             &extent_start, &extent_end,
 311                                             EXTENT_DIRTY | EXTENT_UPTODATE);
 312                 if (ret)
 313                         break;
 314
 315                 if (extent_start <= start) {
 316                         start = extent_end + 1;
 317                 } else if (extent_start > start && extent_start < end) {
 318                         size = extent_start - start;
 319                         total_added += size;
 320                         ret = btrfs_add_free_space(block_group, start,
 321                                                    size);
 322                         BUG_ON(ret);
 323                         start = extent_end + 1;
 324                 } else {
 325                         break;
 326                 }
 327         }
 328
 329         if (start < end) {
 330                 size = end - start;
 331                 total_added += size;
 332                 ret = btrfs_add_free_space(block_group, start, size);
 333                 BUG_ON(ret);
 334         }
 335
 336         return total_added;
 337 }
 338
 339 static noinline void caching_thread(struct btrfs_work *work)
 340 {
 341         struct btrfs_block_group_cache *block_group;
 342         struct btrfs_fs_info *fs_info;
 343         struct btrfs_caching_control *caching_ctl;
 344         struct btrfs_root *extent_root;
 345         struct btrfs_path *path;
 346         struct extent_buffer *leaf;
 347         struct btrfs_key key;
 348         u64 total_found = 0;
 349         u64 last = 0;
 350         u32 nritems;
 351         int ret = 0;
 352
 353         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 354         block_group = caching_ctl->block_group;
 355         fs_info = block_group->fs_info;
 356         extent_root = fs_info->extent_root;
 357
 358         path = btrfs_alloc_path();
 359         if (!path)
 360                 goto out;
 361
 362         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 363
 364         /*
 365          * We don't want to deadlock with somebody trying to allocate a new
 366          * extent for the extent root while also trying to search the extent
 367          * root to add free space.  So we skip locking and search the commit
 368          * root, since its read-only
 369          */
 370         path->skip_locking = 1;
 371         path->search_commit_root = 1;
 372         path->reada = 1;
 373
 374         key.objectid = last;
 375         key.offset = 0;
 376         key.type = BTRFS_EXTENT_ITEM_KEY;
 377 again:
 378         mutex_lock(&caching_ctl->mutex);
 379         /* need to make sure the commit_root doesn't disappear */
 380         down_read(&fs_info->extent_commit_sem);
 381
 382         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 383         if (ret < 0)
 384                 goto err;
 385
 386         leaf = path->nodes[0];
 387         nritems = btrfs_header_nritems(leaf);
 388
 389         while (1) {
 390                 if (btrfs_fs_closing(fs_info) > 1) {
 391                         last = (u64)-1;
 392                         break;
 393                 }
 394
 395                 if (path->slots[0] < nritems) {
 396                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 397                 } else {
 398                         ret = find_next_key(path, 0, &key);
 399                         if (ret)
 400                                 break;
 401
 402                         if (need_resched() ||
 403                             btrfs_next_leaf(extent_root, path)) {
 404                                 caching_ctl->progress = last;
 405                                 btrfs_release_path(path);
 406                                 up_read(&fs_info->extent_commit_sem);
 407                                 mutex_unlock(&caching_ctl->mutex);
 408                                 cond_resched();
 409                                 goto again;
 410                         }
 411                         leaf = path->nodes[0];
 412                         nritems = btrfs_header_nritems(leaf);
 413                         continue;
 414                 }
 415
 416                 if (key.objectid < block_group->key.objectid) {
 417                         path->slots[0]++;
 418                         continue;
 419                 }
 420
 421                 if (key.objectid >= block_group->key.objectid +
 422                     block_group->key.offset)
 423                         break;
 424
 425                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 426                         total_found += add_new_free_space(block_group,
 427                                                           fs_info, last,
 428                                                           key.objectid);
 429                         last = key.objectid + key.offset;
 430
 431                         if (total_found > (1024 * 1024 * 2)) {
 432                                 total_found = 0;
 433                                 wake_up(&caching_ctl->wait);
 434                         }
 435                 }
 436                 path->slots[0]++;
 437         }
 438         ret = 0;
 439
 440         total_found += add_new_free_space(block_group, fs_info, last,
 441                                           block_group->key.objectid +
 442                                           block_group->key.offset);
 443         caching_ctl->progress = (u64)-1;
 444
 445         spin_lock(&block_group->lock);
 446         block_group->caching_ctl = NULL;
 447         block_group->cached = BTRFS_CACHE_FINISHED;
 448         spin_unlock(&block_group->lock);
 449
 450 err:
 451         btrfs_free_path(path);
 452         up_read(&fs_info->extent_commit_sem);
 453
 454         free_excluded_extents(extent_root, block_group);
 455
 456         mutex_unlock(&caching_ctl->mutex);
 457 out:
 458         wake_up(&caching_ctl->wait);
 459
 460         put_caching_control(caching_ctl);
 461         btrfs_put_block_group(block_group);
 462 }
 463
 464 static int cache_block_group(struct btrfs_block_group_cache *cache,
 465                              struct btrfs_trans_handle *trans,
 466                              struct btrfs_root *root,
 467                              int load_cache_only)
 468 {
 469         struct btrfs_fs_info *fs_info = cache->fs_info;
 470         struct btrfs_caching_control *caching_ctl;
 471         int ret = 0;
 472
 473         smp_mb();
 474         if (cache->cached != BTRFS_CACHE_NO)
 475                 return 0;
 476
 477         /*
 478          * We can't do the read from on-disk cache during a commit since we need
 479          * to have the normal tree locking.  Also if we are currently trying to
 480          * allocate blocks for the tree root we can't do the fast caching since
 481          * we likely hold important locks.
 482          */
 483         if (trans && (!trans->transaction->in_commit) &&
 484             (root && root != root->fs_info->tree_root)) {
 485                 spin_lock(&cache->lock);
 486                 if (cache->cached != BTRFS_CACHE_NO) {
 487                         spin_unlock(&cache->lock);
 488                         return 0;
 489                 }
 490                 cache->cached = BTRFS_CACHE_STARTED;
 491                 spin_unlock(&cache->lock);
 492
 493                 ret = load_free_space_cache(fs_info, cache);
 494
 495                 spin_lock(&cache->lock);
 496                 if (ret == 1) {
 497                         cache->cached = BTRFS_CACHE_FINISHED;
 498                         cache->last_byte_to_unpin = (u64)-1;
 499                 } else {
 500                         cache->cached = BTRFS_CACHE_NO;
 501                 }
 502                 spin_unlock(&cache->lock);
 503                 if (ret == 1) {
 504                         free_excluded_extents(fs_info->extent_root, cache);
 505                         return 0;
 506                 }
 507         }
 508
 509         if (load_cache_only)
 510                 return 0;
 511
 512         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 513         BUG_ON(!caching_ctl);
 514
 515         INIT_LIST_HEAD(&caching_ctl->list);
 516         mutex_init(&caching_ctl->mutex);
 517         init_waitqueue_head(&caching_ctl->wait);
 518         caching_ctl->block_group = cache;
 519         caching_ctl->progress = cache->key.objectid;
 520         /* one for caching kthread, one for caching block group list */
 521         atomic_set(&caching_ctl->count, 2);
 522         caching_ctl->work.func = caching_thread;
 523
 524         spin_lock(&cache->lock);
 525         if (cache->cached != BTRFS_CACHE_NO) {
 526                 spin_unlock(&cache->lock);
 527                 kfree(caching_ctl);
 528                 return 0;
 529         }
 530         cache->caching_ctl = caching_ctl;
 531         cache->cached = BTRFS_CACHE_STARTED;
 532         spin_unlock(&cache->lock);
 533
 534         down_write(&fs_info->extent_commit_sem);
 535         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 536         up_write(&fs_info->extent_commit_sem);
 537
 538         btrfs_get_block_group(cache);
 539
 540         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 541
 542         return ret;
 543 }
 544
 545 /*
 546  * return the block group that starts at or after bytenr
 547  */
 548 static struct btrfs_block_group_cache *
 549 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 550 {
 551         struct btrfs_block_group_cache *cache;
 552
 553         cache = block_group_cache_tree_search(info, bytenr, 0);
 554
 555         return cache;
 556 }
 557
 558 /*
 559  * return the block group that contains the given bytenr
 560  */
 561 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 562                                                  struct btrfs_fs_info *info,
 563                                                  u64 bytenr)
 564 {
 565         struct btrfs_block_group_cache *cache;
 566
 567         cache = block_group_cache_tree_search(info, bytenr, 1);
 568
 569         return cache;
 570 }
 571
 572 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 573                                                   u64 flags)
 574 {
 575         struct list_head *head = &info->space_info;
 576         struct btrfs_space_info *found;
 577
 578         flags &= BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_SYSTEM |
 579                  BTRFS_BLOCK_GROUP_METADATA;
 580
 581         rcu_read_lock();
 582         list_for_each_entry_rcu(found, head, list) {
 583                 if (found->flags & flags) {
 584                         rcu_read_unlock();
 585                         return found;
 586                 }
 587         }
 588         rcu_read_unlock();
 589         return NULL;
 590 }
 591
 592 /*
 593  * after adding space to the filesystem, we need to clear the full flags
 594  * on all the space infos.
 595  */
 596 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 597 {
 598         struct list_head *head = &info->space_info;
 599         struct btrfs_space_info *found;
 600
 601         rcu_read_lock();
 602         list_for_each_entry_rcu(found, head, list)
 603                 found->full = 0;
 604         rcu_read_unlock();
 605 }
 606
 607 static u64 div_factor(u64 num, int factor)
 608 {
 609         if (factor == 10)
 610                 return num;
 611         num *= factor;
 612         do_div(num, 10);
 613         return num;
 614 }
 615
 616 static u64 div_factor_fine(u64 num, int factor)
 617 {
 618         if (factor == 100)
 619                 return num;
 620         num *= factor;
 621         do_div(num, 100);
 622         return num;
 623 }
 624
 625 u64 btrfs_find_block_group(struct btrfs_root *root,
 626                            u64 search_start, u64 search_hint, int owner)
 627 {
 628         struct btrfs_block_group_cache *cache;
 629         u64 used;
 630         u64 last = max(search_hint, search_start);
 631         u64 group_start = 0;
 632         int full_search = 0;
 633         int factor = 9;
 634         int wrapped = 0;
 635 again:
 636         while (1) {
 637                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
 638                 if (!cache)
 639                         break;
 640
 641                 spin_lock(&cache->lock);
 642                 last = cache->key.objectid + cache->key.offset;
 643                 used = btrfs_block_group_used(&cache->item);
 644
 645                 if ((full_search || !cache->ro) &&
 646                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
 647                         if (used + cache->pinned + cache->reserved <
 648                             div_factor(cache->key.offset, factor)) {
 649                                 group_start = cache->key.objectid;
 650                                 spin_unlock(&cache->lock);
 651                                 btrfs_put_block_group(cache);
 652                                 goto found;
 653                         }
 654                 }
 655                 spin_unlock(&cache->lock);
 656                 btrfs_put_block_group(cache);
 657                 cond_resched();
 658         }
 659         if (!wrapped) {
 660                 last = search_start;
 661                 wrapped = 1;
 662                 goto again;
 663         }
 664         if (!full_search && factor < 10) {
 665                 last = search_start;
 666                 full_search = 1;
 667                 factor = 10;
 668                 goto again;
 669         }
 670 found:
 671         return group_start;
 672 }
 673
 674 /* simple helper to search for an existing extent at a given offset */
 675 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 676 {
 677         int ret;
 678         struct btrfs_key key;
 679         struct btrfs_path *path;
 680
 681         path = btrfs_alloc_path();
 682         if (!path)
 683                 return -ENOMEM;
 684
 685         key.objectid = start;
 686         key.offset = len;
 687         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 688         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 689                                 0, 0);
 690         btrfs_free_path(path);
 691         return ret;
 692 }
 693
 694 /*
 695  * helper function to lookup reference count and flags of extent.
 696  *
 697  * the head node for delayed ref is used to store the sum of all the
 698  * reference count modifications queued up in the rbtree. the head
 699  * node may also store the extent flags to set. This way you can check
 700  * to see what the reference count and extent flags would be if all of
 701  * the delayed refs are not processed.
 702  */
 703 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 704                              struct btrfs_root *root, u64 bytenr,
 705                              u64 num_bytes, u64 *refs, u64 *flags)
 706 {
 707         struct btrfs_delayed_ref_head *head;
 708         struct btrfs_delayed_ref_root *delayed_refs;
 709         struct btrfs_path *path;
 710         struct btrfs_extent_item *ei;
 711         struct extent_buffer *leaf;
 712         struct btrfs_key key;
 713         u32 item_size;
 714         u64 num_refs;
 715         u64 extent_flags;
 716         int ret;
 717
 718         path = btrfs_alloc_path();
 719         if (!path)
 720                 return -ENOMEM;
 721
 722         key.objectid = bytenr;
 723         key.type = BTRFS_EXTENT_ITEM_KEY;
 724         key.offset = num_bytes;
 725         if (!trans) {
 726                 path->skip_locking = 1;
 727                 path->search_commit_root = 1;
 728         }
 729 again:
 730         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 731                                 &key, path, 0, 0);
 732         if (ret < 0)
 733                 goto out_free;
 734
 735         if (ret == 0) {
 736                 leaf = path->nodes[0];
 737                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 738                 if (item_size >= sizeof(*ei)) {
 739                         ei = btrfs_item_ptr(leaf, path->slots[0],
 740                                             struct btrfs_extent_item);
 741                         num_refs = btrfs_extent_refs(leaf, ei);
 742                         extent_flags = btrfs_extent_flags(leaf, ei);
 743                 } else {
 744 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 745                         struct btrfs_extent_item_v0 *ei0;
 746                         BUG_ON(item_size != sizeof(*ei0));
 747                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 748                                              struct btrfs_extent_item_v0);
 749                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 750                         /* FIXME: this isn't correct for data */
 751                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 752 #else
 753                         BUG();
 754 #endif
 755                 }
 756                 BUG_ON(num_refs == 0);
 757         } else {
 758                 num_refs = 0;
 759                 extent_flags = 0;
 760                 ret = 0;
 761         }
 762
 763         if (!trans)
 764                 goto out;
 765
 766         delayed_refs = &trans->transaction->delayed_refs;
 767         spin_lock(&delayed_refs->lock);
 768         head = btrfs_find_delayed_ref_head(trans, bytenr);
 769         if (head) {
 770                 if (!mutex_trylock(&head->mutex)) {
 771                         atomic_inc(&head->node.refs);
 772                         spin_unlock(&delayed_refs->lock);
 773
 774                         btrfs_release_path(path);
 775
 776                         /*
 777                          * Mutex was contended, block until it's released and try
 778                          * again
 779                          */
 780                         mutex_lock(&head->mutex);
 781                         mutex_unlock(&head->mutex);
 782                         btrfs_put_delayed_ref(&head->node);
 783                         goto again;
 784                 }
 785                 if (head->extent_op && head->extent_op->update_flags)
 786                         extent_flags |= head->extent_op->flags_to_set;
 787                 else
 788                         BUG_ON(num_refs == 0);
 789
 790                 num_refs += head->node.ref_mod;
 791                 mutex_unlock(&head->mutex);
 792         }
 793         spin_unlock(&delayed_refs->lock);
 794 out:
 795         WARN_ON(num_refs == 0);
 796         if (refs)
 797                 *refs = num_refs;
 798         if (flags)
 799                 *flags = extent_flags;
 800 out_free:
 801         btrfs_free_path(path);
 802         return ret;
 803 }
 804
 805 /*
 806  * Back reference rules.  Back refs have three main goals:
 807  *
 808  * 1) differentiate between all holders of references to an extent so that
 809  *    when a reference is dropped we can make sure it was a valid reference
 810  *    before freeing the extent.
 811  *
 812  * 2) Provide enough information to quickly find the holders of an extent
 813  *    if we notice a given block is corrupted or bad.
 814  *
 815  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 816  *    maintenance.  This is actually the same as #2, but with a slightly
 817  *    different use case.
 818  *
 819  * There are two kinds of back refs. The implicit back refs is optimized
 820  * for pointers in non-shared tree blocks. For a given pointer in a block,
 821  * back refs of this kind provide information about the block's owner tree
 822  * and the pointer's key. These information allow us to find the block by
 823  * b-tree searching. The full back refs is for pointers in tree blocks not
 824  * referenced by their owner trees. The location of tree block is recorded
 825  * in the back refs. Actually the full back refs is generic, and can be
 826  * used in all cases the implicit back refs is used. The major shortcoming
 827  * of the full back refs is its overhead. Every time a tree block gets
 828  * COWed, we have to update back refs entry for all pointers in it.
 829  *
 830  * For a newly allocated tree block, we use implicit back refs for
 831  * pointers in it. This means most tree related operations only involve
 832  * implicit back refs. For a tree block created in old transaction, the
 833  * only way to drop a reference to it is COW it. So we can detect the
 834  * event that tree block loses its owner tree's reference and do the
 835  * back refs conversion.
 836  *
 837  * When a tree block is COW'd through a tree, there are four cases:
 838  *
 839  * The reference count of the block is one and the tree is the block's
 840  * owner tree. Nothing to do in this case.
 841  *
 842  * The reference count of the block is one and the tree is not the
 843  * block's owner tree. In this case, full back refs is used for pointers
 844  * in the block. Remove these full back refs, add implicit back refs for
 845  * every pointers in the new block.
 846  *
 847  * The reference count of the block is greater than one and the tree is
 848  * the block's owner tree. In this case, implicit back refs is used for
 849  * pointers in the block. Add full back refs for every pointers in the
 850  * block, increase lower level extents' reference counts. The original
 851  * implicit back refs are entailed to the new block.
 852  *
 853  * The reference count of the block is greater than one and the tree is
 854  * not the block's owner tree. Add implicit back refs for every pointer in
 855  * the new block, increase lower level extents' reference count.
 856  *
 857  * Back Reference Key composing:
 858  *
 859  * The key objectid corresponds to the first byte in the extent,
 860  * The key type is used to differentiate between types of back refs.
 861  * There are different meanings of the key offset for different types
 862  * of back refs.
 863  *
 864  * File extents can be referenced by:
 865  *
 866  * - multiple snapshots, subvolumes, or different generations in one subvol
 867  * - different files inside a single subvolume
 868  * - different offsets inside a file (bookend extents in file.c)
 869  *
 870  * The extent ref structure for the implicit back refs has fields for:
 871  *
 872  * - Objectid of the subvolume root
 873  * - objectid of the file holding the reference
 874  * - original offset in the file
 875  * - how many bookend extents
 876  *
 877  * The key offset for the implicit back refs is hash of the first
 878  * three fields.
 879  *
 880  * The extent ref structure for the full back refs has field for:
 881  *
 882  * - number of pointers in the tree leaf
 883  *
 884  * The key offset for the implicit back refs is the first byte of
 885  * the tree leaf
 886  *
 887  * When a file extent is allocated, The implicit back refs is used.
 888  * the fields are filled in:
 889  *
 890  *     (root_key.objectid, inode objectid, offset in file, 1)
 891  *
 892  * When a file extent is removed file truncation, we find the
 893  * corresponding implicit back refs and check the following fields:
 894  *
 895  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 896  *
 897  * Btree extents can be referenced by:
 898  *
 899  * - Different subvolumes
 900  *
 901  * Both the implicit back refs and the full back refs for tree blocks
 902  * only consist of key. The key offset for the implicit back refs is
 903  * objectid of block's owner tree. The key offset for the full back refs
 904  * is the first byte of parent block.
 905  *
 906  * When implicit back refs is used, information about the lowest key and
 907  * level of the tree block are required. These information are stored in
 908  * tree block info structure.
 909  */
 910
 911 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 912 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 913                                   struct btrfs_root *root,
 914                                   struct btrfs_path *path,
 915                                   u64 owner, u32 extra_size)
 916 {
 917         struct btrfs_extent_item *item;
 918         struct btrfs_extent_item_v0 *ei0;
 919         struct btrfs_extent_ref_v0 *ref0;
 920         struct btrfs_tree_block_info *bi;
 921         struct extent_buffer *leaf;
 922         struct btrfs_key key;
 923         struct btrfs_key found_key;
 924         u32 new_size = sizeof(*item);
 925         u64 refs;
 926         int ret;
 927
 928         leaf = path->nodes[0];
 929         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 930
 931         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 932         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 933                              struct btrfs_extent_item_v0);
 934         refs = btrfs_extent_refs_v0(leaf, ei0);
 935
 936         if (owner == (u64)-1) {
 937                 while (1) {
 938                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 939                                 ret = btrfs_next_leaf(root, path);
 940                                 if (ret < 0)
 941                                         return ret;
 942                                 BUG_ON(ret > 0);
 943                                 leaf = path->nodes[0];
 944                         }
 945                         btrfs_item_key_to_cpu(leaf, &found_key,
 946                                               path->slots[0]);
 947                         BUG_ON(key.objectid != found_key.objectid);
 948                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 949                                 path->slots[0]++;
 950                                 continue;
 951                         }
 952                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 953                                               struct btrfs_extent_ref_v0);
 954                         owner = btrfs_ref_objectid_v0(leaf, ref0);
 955                         break;
 956                 }
 957         }
 958         btrfs_release_path(path);
 959
 960         if (owner < BTRFS_FIRST_FREE_OBJECTID)
 961                 new_size += sizeof(*bi);
 962
 963         new_size -= sizeof(*ei0);
 964         ret = btrfs_search_slot(trans, root, &key, path,
 965                                 new_size + extra_size, 1);
 966         if (ret < 0)
 967                 return ret;
 968         BUG_ON(ret);
 969
 970         ret = btrfs_extend_item(trans, root, path, new_size);
 971
 972         leaf = path->nodes[0];
 973         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 974         btrfs_set_extent_refs(leaf, item, refs);
 975         /* FIXME: get real generation */
 976         btrfs_set_extent_generation(leaf, item, 0);
 977         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 978                 btrfs_set_extent_flags(leaf, item,
 979                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
 980                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
 981                 bi = (struct btrfs_tree_block_info *)(item + 1);
 982                 /* FIXME: get first key of the block */
 983                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
 984                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
 985         } else {
 986                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
 987         }
 988         btrfs_mark_buffer_dirty(leaf);
 989         return 0;
 990 }
 991 #endif
 992
 993 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
 994 {
 995         u32 high_crc = ~(u32)0;
 996         u32 low_crc = ~(u32)0;
 997         __le64 lenum;
 998
 999         lenum = cpu_to_le64(root_objectid);
1000         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1001         lenum = cpu_to_le64(owner);
1002         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1003         lenum = cpu_to_le64(offset);
1004         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1005
1006         return ((u64)high_crc << 31) ^ (u64)low_crc;
1007 }
1008
1009 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1010                                      struct btrfs_extent_data_ref *ref)
1011 {
1012         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1013                                     btrfs_extent_data_ref_objectid(leaf, ref),
1014                                     btrfs_extent_data_ref_offset(leaf, ref));
1015 }
1016
1017 static int match_extent_data_ref(struct extent_buffer *leaf,
1018                                  struct btrfs_extent_data_ref *ref,
1019                                  u64 root_objectid, u64 owner, u64 offset)
1020 {
1021         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1022             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1023             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1024                 return 0;
1025         return 1;
1026 }
1027
1028 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1029                                            struct btrfs_root *root,
1030                                            struct btrfs_path *path,
1031                                            u64 bytenr, u64 parent,
1032                                            u64 root_objectid,
1033                                            u64 owner, u64 offset)
1034 {
1035         struct btrfs_key key;
1036         struct btrfs_extent_data_ref *ref;
1037         struct extent_buffer *leaf;
1038         u32 nritems;
1039         int ret;
1040         int recow;
1041         int err = -ENOENT;
1042
1043         key.objectid = bytenr;
1044         if (parent) {
1045                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1046                 key.offset = parent;
1047         } else {
1048                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1049                 key.offset = hash_extent_data_ref(root_objectid,
1050                                                   owner, offset);
1051         }
1052 again:
1053         recow = 0;
1054         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1055         if (ret < 0) {
1056                 err = ret;
1057                 goto fail;
1058         }
1059
1060         if (parent) {
1061                 if (!ret)
1062                         return 0;
1063 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1064                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1065                 btrfs_release_path(path);
1066                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1067                 if (ret < 0) {
1068                         err = ret;
1069                         goto fail;
1070                 }
1071                 if (!ret)
1072                         return 0;
1073 #endif
1074                 goto fail;
1075         }
1076
1077         leaf = path->nodes[0];
1078         nritems = btrfs_header_nritems(leaf);
1079         while (1) {
1080                 if (path->slots[0] >= nritems) {
1081                         ret = btrfs_next_leaf(root, path);
1082                         if (ret < 0)
1083                                 err = ret;
1084                         if (ret)
1085                                 goto fail;
1086
1087                         leaf = path->nodes[0];
1088                         nritems = btrfs_header_nritems(leaf);
1089                         recow = 1;
1090                 }
1091
1092                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1093                 if (key.objectid != bytenr ||
1094                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1095                         goto fail;
1096
1097                 ref = btrfs_item_ptr(leaf, path->slots[0],
1098                                      struct btrfs_extent_data_ref);
1099
1100                 if (match_extent_data_ref(leaf, ref, root_objectid,
1101                                           owner, offset)) {
1102                         if (recow) {
1103                                 btrfs_release_path(path);
1104                                 goto again;
1105                         }
1106                         err = 0;
1107                         break;
1108                 }
1109                 path->slots[0]++;
1110         }
1111 fail:
1112         return err;
1113 }
1114
1115 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1116                                            struct btrfs_root *root,
1117                                            struct btrfs_path *path,
1118                                            u64 bytenr, u64 parent,
1119                                            u64 root_objectid, u64 owner,
1120                                            u64 offset, int refs_to_add)
1121 {
1122         struct btrfs_key key;
1123         struct extent_buffer *leaf;
1124         u32 size;
1125         u32 num_refs;
1126         int ret;
1127
1128         key.objectid = bytenr;
1129         if (parent) {
1130                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1131                 key.offset = parent;
1132                 size = sizeof(struct btrfs_shared_data_ref);
1133         } else {
1134                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1135                 key.offset = hash_extent_data_ref(root_objectid,
1136                                                   owner, offset);
1137                 size = sizeof(struct btrfs_extent_data_ref);
1138         }
1139
1140         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1141         if (ret && ret != -EEXIST)
1142                 goto fail;
1143
1144         leaf = path->nodes[0];
1145         if (parent) {
1146                 struct btrfs_shared_data_ref *ref;
1147                 ref = btrfs_item_ptr(leaf, path->slots[0],
1148                                      struct btrfs_shared_data_ref);
1149                 if (ret == 0) {
1150                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1151                 } else {
1152                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1153                         num_refs += refs_to_add;
1154                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1155                 }
1156         } else {
1157                 struct btrfs_extent_data_ref *ref;
1158                 while (ret == -EEXIST) {
1159                         ref = btrfs_item_ptr(leaf, path->slots[0],
1160                                              struct btrfs_extent_data_ref);
1161                         if (match_extent_data_ref(leaf, ref, root_objectid,
1162                                                   owner, offset))
1163                                 break;
1164                         btrfs_release_path(path);
1165                         key.offset++;
1166                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1167                                                       size);
1168                         if (ret && ret != -EEXIST)
1169                                 goto fail;
1170
1171                         leaf = path->nodes[0];
1172                 }
1173                 ref = btrfs_item_ptr(leaf, path->slots[0],
1174                                      struct btrfs_extent_data_ref);
1175                 if (ret == 0) {
1176                         btrfs_set_extent_data_ref_root(leaf, ref,
1177                                                        root_objectid);
1178                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1179                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1180                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1181                 } else {
1182                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1183                         num_refs += refs_to_add;
1184                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1185                 }
1186         }
1187         btrfs_mark_buffer_dirty(leaf);
1188         ret = 0;
1189 fail:
1190         btrfs_release_path(path);
1191         return ret;
1192 }
1193
1194 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1195                                            struct btrfs_root *root,
1196                                            struct btrfs_path *path,
1197                                            int refs_to_drop)
1198 {
1199         struct btrfs_key key;
1200         struct btrfs_extent_data_ref *ref1 = NULL;
1201         struct btrfs_shared_data_ref *ref2 = NULL;
1202         struct extent_buffer *leaf;
1203         u32 num_refs = 0;
1204         int ret = 0;
1205
1206         leaf = path->nodes[0];
1207         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1208
1209         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1210                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1211                                       struct btrfs_extent_data_ref);
1212                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1213         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1214                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1215                                       struct btrfs_shared_data_ref);
1216                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1217 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1218         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1219                 struct btrfs_extent_ref_v0 *ref0;
1220                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1221                                       struct btrfs_extent_ref_v0);
1222                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1223 #endif
1224         } else {
1225                 BUG();
1226         }
1227
1228         BUG_ON(num_refs < refs_to_drop);
1229         num_refs -= refs_to_drop;
1230
1231         if (num_refs == 0) {
1232                 ret = btrfs_del_item(trans, root, path);
1233         } else {
1234                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1235                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1236                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1237                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1238 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1239                 else {
1240                         struct btrfs_extent_ref_v0 *ref0;
1241                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1242                                         struct btrfs_extent_ref_v0);
1243                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1244                 }
1245 #endif
1246                 btrfs_mark_buffer_dirty(leaf);
1247         }
1248         return ret;
1249 }
1250
1251 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1252                                           struct btrfs_path *path,
1253                                           struct btrfs_extent_inline_ref *iref)
1254 {
1255         struct btrfs_key key;
1256         struct extent_buffer *leaf;
1257         struct btrfs_extent_data_ref *ref1;
1258         struct btrfs_shared_data_ref *ref2;
1259         u32 num_refs = 0;
1260
1261         leaf = path->nodes[0];
1262         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1263         if (iref) {
1264                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1265                     BTRFS_EXTENT_DATA_REF_KEY) {
1266                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1267                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1268                 } else {
1269                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1270                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1271                 }
1272         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1273                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1274                                       struct btrfs_extent_data_ref);
1275                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1276         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1277                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1278                                       struct btrfs_shared_data_ref);
1279                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1280 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1281         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1282                 struct btrfs_extent_ref_v0 *ref0;
1283                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1284                                       struct btrfs_extent_ref_v0);
1285                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1286 #endif
1287         } else {
1288                 WARN_ON(1);
1289         }
1290         return num_refs;
1291 }
1292
1293 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1294                                           struct btrfs_root *root,
1295                                           struct btrfs_path *path,
1296                                           u64 bytenr, u64 parent,
1297                                           u64 root_objectid)
1298 {
1299         struct btrfs_key key;
1300         int ret;
1301
1302         key.objectid = bytenr;
1303         if (parent) {
1304                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1305                 key.offset = parent;
1306         } else {
1307                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1308                 key.offset = root_objectid;
1309         }
1310
1311         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1312         if (ret > 0)
1313                 ret = -ENOENT;
1314 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1315         if (ret == -ENOENT && parent) {
1316                 btrfs_release_path(path);
1317                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1318                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1319                 if (ret > 0)
1320                         ret = -ENOENT;
1321         }
1322 #endif
1323         return ret;
1324 }
1325
1326 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1327                                           struct btrfs_root *root,
1328                                           struct btrfs_path *path,
1329                                           u64 bytenr, u64 parent,
1330                                           u64 root_objectid)
1331 {
1332         struct btrfs_key key;
1333         int ret;
1334
1335         key.objectid = bytenr;
1336         if (parent) {
1337                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1338                 key.offset = parent;
1339         } else {
1340                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1341                 key.offset = root_objectid;
1342         }
1343
1344         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1345         btrfs_release_path(path);
1346         return ret;
1347 }
1348
1349 static inline int extent_ref_type(u64 parent, u64 owner)
1350 {
1351         int type;
1352         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1353                 if (parent > 0)
1354                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1355                 else
1356                         type = BTRFS_TREE_BLOCK_REF_KEY;
1357         } else {
1358                 if (parent > 0)
1359                         type = BTRFS_SHARED_DATA_REF_KEY;
1360                 else
1361                         type = BTRFS_EXTENT_DATA_REF_KEY;
1362         }
1363         return type;
1364 }
1365
1366 static int find_next_key(struct btrfs_path *path, int level,
1367                          struct btrfs_key *key)
1368
1369 {
1370         for (; level < BTRFS_MAX_LEVEL; level++) {
1371                 if (!path->nodes[level])
1372                         break;
1373                 if (path->slots[level] + 1 >=
1374                     btrfs_header_nritems(path->nodes[level]))
1375                         continue;
1376                 if (level == 0)
1377                         btrfs_item_key_to_cpu(path->nodes[level], key,
1378                                               path->slots[level] + 1);
1379                 else
1380                         btrfs_node_key_to_cpu(path->nodes[level], key,
1381                                               path->slots[level] + 1);
1382                 return 0;
1383         }
1384         return 1;
1385 }
1386
1387 /*
1388  * look for inline back ref. if back ref is found, *ref_ret is set
1389  * to the address of inline back ref, and 0 is returned.
1390  *
1391  * if back ref isn't found, *ref_ret is set to the address where it
1392  * should be inserted, and -ENOENT is returned.
1393  *
1394  * if insert is true and there are too many inline back refs, the path
1395  * points to the extent item, and -EAGAIN is returned.
1396  *
1397  * NOTE: inline back refs are ordered in the same way that back ref
1398  *       items in the tree are ordered.
1399  */
1400 static noinline_for_stack
1401 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1402                                  struct btrfs_root *root,
1403                                  struct btrfs_path *path,
1404                                  struct btrfs_extent_inline_ref **ref_ret,
1405                                  u64 bytenr, u64 num_bytes,
1406                                  u64 parent, u64 root_objectid,
1407                                  u64 owner, u64 offset, int insert)
1408 {
1409         struct btrfs_key key;
1410         struct extent_buffer *leaf;
1411         struct btrfs_extent_item *ei;
1412         struct btrfs_extent_inline_ref *iref;
1413         u64 flags;
1414         u64 item_size;
1415         unsigned long ptr;
1416         unsigned long end;
1417         int extra_size;
1418         int type;
1419         int want;
1420         int ret;
1421         int err = 0;
1422
1423         key.objectid = bytenr;
1424         key.type = BTRFS_EXTENT_ITEM_KEY;
1425         key.offset = num_bytes;
1426
1427         want = extent_ref_type(parent, owner);
1428         if (insert) {
1429                 extra_size = btrfs_extent_inline_ref_size(want);
1430                 path->keep_locks = 1;
1431         } else
1432                 extra_size = -1;
1433         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1434         if (ret < 0) {
1435                 err = ret;
1436                 goto out;
1437         }
1438         BUG_ON(ret);
1439
1440         leaf = path->nodes[0];
1441         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1442 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1443         if (item_size < sizeof(*ei)) {
1444                 if (!insert) {
1445                         err = -ENOENT;
1446                         goto out;
1447                 }
1448                 ret = convert_extent_item_v0(trans, root, path, owner,
1449                                              extra_size);
1450                 if (ret < 0) {
1451                         err = ret;
1452                         goto out;
1453                 }
1454                 leaf = path->nodes[0];
1455                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1456         }
1457 #endif
1458         BUG_ON(item_size < sizeof(*ei));
1459
1460         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1461         flags = btrfs_extent_flags(leaf, ei);
1462
1463         ptr = (unsigned long)(ei + 1);
1464         end = (unsigned long)ei + item_size;
1465
1466         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1467                 ptr += sizeof(struct btrfs_tree_block_info);
1468                 BUG_ON(ptr > end);
1469         } else {
1470                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1471         }
1472
1473         err = -ENOENT;
1474         while (1) {
1475                 if (ptr >= end) {
1476                         WARN_ON(ptr > end);
1477                         break;
1478                 }
1479                 iref = (struct btrfs_extent_inline_ref *)ptr;
1480                 type = btrfs_extent_inline_ref_type(leaf, iref);
1481                 if (want < type)
1482                         break;
1483                 if (want > type) {
1484                         ptr += btrfs_extent_inline_ref_size(type);
1485                         continue;
1486                 }
1487
1488                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1489                         struct btrfs_extent_data_ref *dref;
1490                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1491                         if (match_extent_data_ref(leaf, dref, root_objectid,
1492                                                   owner, offset)) {
1493                                 err = 0;
1494                                 break;
1495                         }
1496                         if (hash_extent_data_ref_item(leaf, dref) <
1497                             hash_extent_data_ref(root_objectid, owner, offset))
1498                                 break;
1499                 } else {
1500                         u64 ref_offset;
1501                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1502                         if (parent > 0) {
1503                                 if (parent == ref_offset) {
1504                                         err = 0;
1505                                         break;
1506                                 }
1507                                 if (ref_offset < parent)
1508                                         break;
1509                         } else {
1510                                 if (root_objectid == ref_offset) {
1511                                         err = 0;
1512                                         break;
1513                                 }
1514                                 if (ref_offset < root_objectid)
1515                                         break;
1516                         }
1517                 }
1518                 ptr += btrfs_extent_inline_ref_size(type);
1519         }
1520         if (err == -ENOENT && insert) {
1521                 if (item_size + extra_size >=
1522                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1523                         err = -EAGAIN;
1524                         goto out;
1525                 }
1526                 /*
1527                  * To add new inline back ref, we have to make sure
1528                  * there is no corresponding back ref item.
1529                  * For simplicity, we just do not add new inline back
1530                  * ref if there is any kind of item for this block
1531                  */
1532                 if (find_next_key(path, 0, &key) == 0 &&
1533                     key.objectid == bytenr &&
1534                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1535                         err = -EAGAIN;
1536                         goto out;
1537                 }
1538         }
1539         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1540 out:
1541         if (insert) {
1542                 path->keep_locks = 0;
1543                 btrfs_unlock_up_safe(path, 1);
1544         }
1545         return err;
1546 }
1547
1548 /*
1549  * helper to add new inline back ref
1550  */
1551 static noinline_for_stack
1552 int setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1553                                 struct btrfs_root *root,
1554                                 struct btrfs_path *path,
1555                                 struct btrfs_extent_inline_ref *iref,
1556                                 u64 parent, u64 root_objectid,
1557                                 u64 owner, u64 offset, int refs_to_add,
1558                                 struct btrfs_delayed_extent_op *extent_op)
1559 {
1560         struct extent_buffer *leaf;
1561         struct btrfs_extent_item *ei;
1562         unsigned long ptr;
1563         unsigned long end;
1564         unsigned long item_offset;
1565         u64 refs;
1566         int size;
1567         int type;
1568         int ret;
1569
1570         leaf = path->nodes[0];
1571         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1572         item_offset = (unsigned long)iref - (unsigned long)ei;
1573
1574         type = extent_ref_type(parent, owner);
1575         size = btrfs_extent_inline_ref_size(type);
1576
1577         ret = btrfs_extend_item(trans, root, path, size);
1578
1579         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1580         refs = btrfs_extent_refs(leaf, ei);
1581         refs += refs_to_add;
1582         btrfs_set_extent_refs(leaf, ei, refs);
1583         if (extent_op)
1584                 __run_delayed_extent_op(extent_op, leaf, ei);
1585
1586         ptr = (unsigned long)ei + item_offset;
1587         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1588         if (ptr < end - size)
1589                 memmove_extent_buffer(leaf, ptr + size, ptr,
1590                                       end - size - ptr);
1591
1592         iref = (struct btrfs_extent_inline_ref *)ptr;
1593         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1594         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1595                 struct btrfs_extent_data_ref *dref;
1596                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1597                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1598                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1599                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1600                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1601         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1602                 struct btrfs_shared_data_ref *sref;
1603                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1604                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1605                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1606         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1607                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1608         } else {
1609                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1610         }
1611         btrfs_mark_buffer_dirty(leaf);
1612         return 0;
1613 }
1614
1615 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1616                                  struct btrfs_root *root,
1617                                  struct btrfs_path *path,
1618                                  struct btrfs_extent_inline_ref **ref_ret,
1619                                  u64 bytenr, u64 num_bytes, u64 parent,
1620                                  u64 root_objectid, u64 owner, u64 offset)
1621 {
1622         int ret;
1623
1624         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1625                                            bytenr, num_bytes, parent,
1626                                            root_objectid, owner, offset, 0);
1627         if (ret != -ENOENT)
1628                 return ret;
1629
1630         btrfs_release_path(path);
1631         *ref_ret = NULL;
1632
1633         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1634                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1635                                             root_objectid);
1636         } else {
1637                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1638                                              root_objectid, owner, offset);
1639         }
1640         return ret;
1641 }
1642
1643 /*
1644  * helper to update/remove inline back ref
1645  */
1646 static noinline_for_stack
1647 int update_inline_extent_backref(struct btrfs_trans_handle *trans,
1648                                  struct btrfs_root *root,
1649                                  struct btrfs_path *path,
1650                                  struct btrfs_extent_inline_ref *iref,
1651                                  int refs_to_mod,
1652                                  struct btrfs_delayed_extent_op *extent_op)
1653 {
1654         struct extent_buffer *leaf;
1655         struct btrfs_extent_item *ei;
1656         struct btrfs_extent_data_ref *dref = NULL;
1657         struct btrfs_shared_data_ref *sref = NULL;
1658         unsigned long ptr;
1659         unsigned long end;
1660         u32 item_size;
1661         int size;
1662         int type;
1663         int ret;
1664         u64 refs;
1665
1666         leaf = path->nodes[0];
1667         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1668         refs = btrfs_extent_refs(leaf, ei);
1669         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1670         refs += refs_to_mod;
1671         btrfs_set_extent_refs(leaf, ei, refs);
1672         if (extent_op)
1673                 __run_delayed_extent_op(extent_op, leaf, ei);
1674
1675         type = btrfs_extent_inline_ref_type(leaf, iref);
1676
1677         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1678                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1679                 refs = btrfs_extent_data_ref_count(leaf, dref);
1680         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1681                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1682                 refs = btrfs_shared_data_ref_count(leaf, sref);
1683         } else {
1684                 refs = 1;
1685                 BUG_ON(refs_to_mod != -1);
1686         }
1687
1688         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1689         refs += refs_to_mod;
1690
1691         if (refs > 0) {
1692                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1693                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1694                 else
1695                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1696         } else {
1697                 size =  btrfs_extent_inline_ref_size(type);
1698                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1699                 ptr = (unsigned long)iref;
1700                 end = (unsigned long)ei + item_size;
1701                 if (ptr + size < end)
1702                         memmove_extent_buffer(leaf, ptr, ptr + size,
1703                                               end - ptr - size);
1704                 item_size -= size;
1705                 ret = btrfs_truncate_item(trans, root, path, item_size, 1);
1706         }
1707         btrfs_mark_buffer_dirty(leaf);
1708         return 0;
1709 }
1710
1711 static noinline_for_stack
1712 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1713                                  struct btrfs_root *root,
1714                                  struct btrfs_path *path,
1715                                  u64 bytenr, u64 num_bytes, u64 parent,
1716                                  u64 root_objectid, u64 owner,
1717                                  u64 offset, int refs_to_add,
1718                                  struct btrfs_delayed_extent_op *extent_op)
1719 {
1720         struct btrfs_extent_inline_ref *iref;
1721         int ret;
1722
1723         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1724                                            bytenr, num_bytes, parent,
1725                                            root_objectid, owner, offset, 1);
1726         if (ret == 0) {
1727                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1728                 ret = update_inline_extent_backref(trans, root, path, iref,
1729                                                    refs_to_add, extent_op);
1730         } else if (ret == -ENOENT) {
1731                 ret = setup_inline_extent_backref(trans, root, path, iref,
1732                                                   parent, root_objectid,
1733                                                   owner, offset, refs_to_add,
1734                                                   extent_op);
1735         }
1736         return ret;
1737 }
1738
1739 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1740                                  struct btrfs_root *root,
1741                                  struct btrfs_path *path,
1742                                  u64 bytenr, u64 parent, u64 root_objectid,
1743                                  u64 owner, u64 offset, int refs_to_add)
1744 {
1745         int ret;
1746         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1747                 BUG_ON(refs_to_add != 1);
1748                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1749                                             parent, root_objectid);
1750         } else {
1751                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1752                                              parent, root_objectid,
1753                                              owner, offset, refs_to_add);
1754         }
1755         return ret;
1756 }
1757
1758 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1759                                  struct btrfs_root *root,
1760                                  struct btrfs_path *path,
1761                                  struct btrfs_extent_inline_ref *iref,
1762                                  int refs_to_drop, int is_data)
1763 {
1764         int ret;
1765
1766         BUG_ON(!is_data && refs_to_drop != 1);
1767         if (iref) {
1768                 ret = update_inline_extent_backref(trans, root, path, iref,
1769                                                    -refs_to_drop, NULL);
1770         } else if (is_data) {
1771                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1772         } else {
1773                 ret = btrfs_del_item(trans, root, path);
1774         }
1775         return ret;
1776 }
1777
1778 static int btrfs_issue_discard(struct block_device *bdev,
1779                                 u64 start, u64 len)
1780 {
1781         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1782 }
1783
1784 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1785                                 u64 num_bytes, u64 *actual_bytes)
1786 {
1787         int ret;
1788         u64 discarded_bytes = 0;
1789         struct btrfs_multi_bio *multi = NULL;
1790
1791
1792         /* Tell the block device(s) that the sectors can be discarded */
1793         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1794                               bytenr, &num_bytes, &multi, 0);
1795         if (!ret) {
1796                 struct btrfs_bio_stripe *stripe = multi->stripes;
1797                 int i;
1798
1799
1800                 for (i = 0; i < multi->num_stripes; i++, stripe++) {
1801                         if (!stripe->dev->can_discard)
1802                                 continue;
1803
1804                         ret = btrfs_issue_discard(stripe->dev->bdev,
1805                                                   stripe->physical,
1806                                                   stripe->length);
1807                         if (!ret)
1808                                 discarded_bytes += stripe->length;
1809                         else if (ret != -EOPNOTSUPP)
1810                                 break;
1811
1812                         /*
1813                          * Just in case we get back EOPNOTSUPP for some reason,
1814                          * just ignore the return value so we don't screw up
1815                          * people calling discard_extent.
1816                          */
1817                         ret = 0;
1818                 }
1819                 kfree(multi);
1820         }
1821
1822         if (actual_bytes)
1823                 *actual_bytes = discarded_bytes;
1824
1825
1826         return ret;
1827 }
1828
1829 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1830                          struct btrfs_root *root,
1831                          u64 bytenr, u64 num_bytes, u64 parent,
1832                          u64 root_objectid, u64 owner, u64 offset)
1833 {
1834         int ret;
1835         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1836                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1837
1838         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1839                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
1840                                         parent, root_objectid, (int)owner,
1841                                         BTRFS_ADD_DELAYED_REF, NULL);
1842         } else {
1843                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
1844                                         parent, root_objectid, owner, offset,
1845                                         BTRFS_ADD_DELAYED_REF, NULL);
1846         }
1847         return ret;
1848 }
1849
1850 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1851                                   struct btrfs_root *root,
1852                                   u64 bytenr, u64 num_bytes,
1853                                   u64 parent, u64 root_objectid,
1854                                   u64 owner, u64 offset, int refs_to_add,
1855                                   struct btrfs_delayed_extent_op *extent_op)
1856 {
1857         struct btrfs_path *path;
1858         struct extent_buffer *leaf;
1859         struct btrfs_extent_item *item;
1860         u64 refs;
1861         int ret;
1862         int err = 0;
1863
1864         path = btrfs_alloc_path();
1865         if (!path)
1866                 return -ENOMEM;
1867
1868         path->reada = 1;
1869         path->leave_spinning = 1;
1870         /* this will setup the path even if it fails to insert the back ref */
1871         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1872                                            path, bytenr, num_bytes, parent,
1873                                            root_objectid, owner, offset,
1874                                            refs_to_add, extent_op);
1875         if (ret == 0)
1876                 goto out;
1877
1878         if (ret != -EAGAIN) {
1879                 err = ret;
1880                 goto out;
1881         }
1882
1883         leaf = path->nodes[0];
1884         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1885         refs = btrfs_extent_refs(leaf, item);
1886         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1887         if (extent_op)
1888                 __run_delayed_extent_op(extent_op, leaf, item);
1889
1890         btrfs_mark_buffer_dirty(leaf);
1891         btrfs_release_path(path);
1892
1893         path->reada = 1;
1894         path->leave_spinning = 1;
1895
1896         /* now insert the actual backref */
1897         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1898                                     path, bytenr, parent, root_objectid,
1899                                     owner, offset, refs_to_add);
1900         BUG_ON(ret);
1901 out:
1902         btrfs_free_path(path);
1903         return err;
1904 }
1905
1906 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1907                                 struct btrfs_root *root,
1908                                 struct btrfs_delayed_ref_node *node,
1909                                 struct btrfs_delayed_extent_op *extent_op,
1910                                 int insert_reserved)
1911 {
1912         int ret = 0;
1913         struct btrfs_delayed_data_ref *ref;
1914         struct btrfs_key ins;
1915         u64 parent = 0;
1916         u64 ref_root = 0;
1917         u64 flags = 0;
1918
1919         ins.objectid = node->bytenr;
1920         ins.offset = node->num_bytes;
1921         ins.type = BTRFS_EXTENT_ITEM_KEY;
1922
1923         ref = btrfs_delayed_node_to_data_ref(node);
1924         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1925                 parent = ref->parent;
1926         else
1927                 ref_root = ref->root;
1928
1929         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1930                 if (extent_op) {
1931                         BUG_ON(extent_op->update_key);
1932                         flags |= extent_op->flags_to_set;
1933                 }
1934                 ret = alloc_reserved_file_extent(trans, root,
1935                                                  parent, ref_root, flags,
1936                                                  ref->objectid, ref->offset,
1937                                                  &ins, node->ref_mod);
1938         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1939                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1940                                              node->num_bytes, parent,
1941                                              ref_root, ref->objectid,
1942                                              ref->offset, node->ref_mod,
1943                                              extent_op);
1944         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1945                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1946                                           node->num_bytes, parent,
1947                                           ref_root, ref->objectid,
1948                                           ref->offset, node->ref_mod,
1949                                           extent_op);
1950         } else {
1951                 BUG();
1952         }
1953         return ret;
1954 }
1955
1956 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1957                                     struct extent_buffer *leaf,
1958                                     struct btrfs_extent_item *ei)
1959 {
1960         u64 flags = btrfs_extent_flags(leaf, ei);
1961         if (extent_op->update_flags) {
1962                 flags |= extent_op->flags_to_set;
1963                 btrfs_set_extent_flags(leaf, ei, flags);
1964         }
1965
1966         if (extent_op->update_key) {
1967                 struct btrfs_tree_block_info *bi;
1968                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1969                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1970                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1971         }
1972 }
1973
1974 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1975                                  struct btrfs_root *root,
1976                                  struct btrfs_delayed_ref_node *node,
1977                                  struct btrfs_delayed_extent_op *extent_op)
1978 {
1979         struct btrfs_key key;
1980         struct btrfs_path *path;
1981         struct btrfs_extent_item *ei;
1982         struct extent_buffer *leaf;
1983         u32 item_size;
1984         int ret;
1985         int err = 0;
1986
1987         path = btrfs_alloc_path();
1988         if (!path)
1989                 return -ENOMEM;
1990
1991         key.objectid = node->bytenr;
1992         key.type = BTRFS_EXTENT_ITEM_KEY;
1993         key.offset = node->num_bytes;
1994
1995         path->reada = 1;
1996         path->leave_spinning = 1;
1997         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
1998                                 path, 0, 1);
1999         if (ret < 0) {
2000                 err = ret;
2001                 goto out;
2002         }
2003         if (ret > 0) {
2004                 err = -EIO;
2005                 goto out;
2006         }
2007
2008         leaf = path->nodes[0];
2009         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2010 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2011         if (item_size < sizeof(*ei)) {
2012                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2013                                              path, (u64)-1, 0);
2014                 if (ret < 0) {
2015                         err = ret;
2016                         goto out;
2017                 }
2018                 leaf = path->nodes[0];
2019                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2020         }
2021 #endif
2022         BUG_ON(item_size < sizeof(*ei));
2023         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2024         __run_delayed_extent_op(extent_op, leaf, ei);
2025
2026         btrfs_mark_buffer_dirty(leaf);
2027 out:
2028         btrfs_free_path(path);
2029         return err;
2030 }
2031
2032 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2033                                 struct btrfs_root *root,
2034                                 struct btrfs_delayed_ref_node *node,
2035                                 struct btrfs_delayed_extent_op *extent_op,
2036                                 int insert_reserved)
2037 {
2038         int ret = 0;
2039         struct btrfs_delayed_tree_ref *ref;
2040         struct btrfs_key ins;
2041         u64 parent = 0;
2042         u64 ref_root = 0;
2043
2044         ins.objectid = node->bytenr;
2045         ins.offset = node->num_bytes;
2046         ins.type = BTRFS_EXTENT_ITEM_KEY;
2047
2048         ref = btrfs_delayed_node_to_tree_ref(node);
2049         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2050                 parent = ref->parent;
2051         else
2052                 ref_root = ref->root;
2053
2054         BUG_ON(node->ref_mod != 1);
2055         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2056                 BUG_ON(!extent_op || !extent_op->update_flags ||
2057                        !extent_op->update_key);
2058                 ret = alloc_reserved_tree_block(trans, root,
2059                                                 parent, ref_root,
2060                                                 extent_op->flags_to_set,
2061                                                 &extent_op->key,
2062                                                 ref->level, &ins);
2063         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2064                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2065                                              node->num_bytes, parent, ref_root,
2066                                              ref->level, 0, 1, extent_op);
2067         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2068                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2069                                           node->num_bytes, parent, ref_root,
2070                                           ref->level, 0, 1, extent_op);
2071         } else {
2072                 BUG();
2073         }
2074         return ret;
2075 }
2076
2077 /* helper function to actually process a single delayed ref entry */
2078 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2079                                struct btrfs_root *root,
2080                                struct btrfs_delayed_ref_node *node,
2081                                struct btrfs_delayed_extent_op *extent_op,
2082                                int insert_reserved)
2083 {
2084         int ret;
2085         if (btrfs_delayed_ref_is_head(node)) {
2086                 struct btrfs_delayed_ref_head *head;
2087                 /*
2088                  * we've hit the end of the chain and we were supposed
2089                  * to insert this extent into the tree.  But, it got
2090                  * deleted before we ever needed to insert it, so all
2091                  * we have to do is clean up the accounting
2092                  */
2093                 BUG_ON(extent_op);
2094                 head = btrfs_delayed_node_to_head(node);
2095                 if (insert_reserved) {
2096                         btrfs_pin_extent(root, node->bytenr,
2097                                          node->num_bytes, 1);
2098                         if (head->is_data) {
2099                                 ret = btrfs_del_csums(trans, root,
2100                                                       node->bytenr,
2101                                                       node->num_bytes);
2102                                 BUG_ON(ret);
2103                         }
2104                 }
2105                 mutex_unlock(&head->mutex);
2106                 return 0;
2107         }
2108
2109         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2110             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2111                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2112                                            insert_reserved);
2113         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2114                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2115                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2116                                            insert_reserved);
2117         else
2118                 BUG();
2119         return ret;
2120 }
2121
2122 static noinline struct btrfs_delayed_ref_node *
2123 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2124 {
2125         struct rb_node *node;
2126         struct btrfs_delayed_ref_node *ref;
2127         int action = BTRFS_ADD_DELAYED_REF;
2128 again:
2129         /*
2130          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2131          * this prevents ref count from going down to zero when
2132          * there still are pending delayed ref.
2133          */
2134         node = rb_prev(&head->node.rb_node);
2135         while (1) {
2136                 if (!node)
2137                         break;
2138                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2139                                 rb_node);
2140                 if (ref->bytenr != head->node.bytenr)
2141                         break;
2142                 if (ref->action == action)
2143                         return ref;
2144                 node = rb_prev(node);
2145         }
2146         if (action == BTRFS_ADD_DELAYED_REF) {
2147                 action = BTRFS_DROP_DELAYED_REF;
2148                 goto again;
2149         }
2150         return NULL;
2151 }
2152
2153 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2154                                        struct btrfs_root *root,
2155                                        struct list_head *cluster)
2156 {
2157         struct btrfs_delayed_ref_root *delayed_refs;
2158         struct btrfs_delayed_ref_node *ref;
2159         struct btrfs_delayed_ref_head *locked_ref = NULL;
2160         struct btrfs_delayed_extent_op *extent_op;
2161         int ret;
2162         int count = 0;
2163         int must_insert_reserved = 0;
2164
2165         delayed_refs = &trans->transaction->delayed_refs;
2166         while (1) {
2167                 if (!locked_ref) {
2168                         /* pick a new head ref from the cluster list */
2169                         if (list_empty(cluster))
2170                                 break;
2171
2172                         locked_ref = list_entry(cluster->next,
2173                                      struct btrfs_delayed_ref_head, cluster);
2174
2175                         /* grab the lock that says we are going to process
2176                          * all the refs for this head */
2177                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2178
2179                         /*
2180                          * we may have dropped the spin lock to get the head
2181                          * mutex lock, and that might have given someone else
2182                          * time to free the head.  If that's true, it has been
2183                          * removed from our list and we can move on.
2184                          */
2185                         if (ret == -EAGAIN) {
2186                                 locked_ref = NULL;
2187                                 count++;
2188                                 continue;
2189                         }
2190                 }
2191
2192                 /*
2193                  * record the must insert reserved flag before we
2194                  * drop the spin lock.
2195                  */
2196                 must_insert_reserved = locked_ref->must_insert_reserved;
2197                 locked_ref->must_insert_reserved = 0;
2198
2199                 extent_op = locked_ref->extent_op;
2200                 locked_ref->extent_op = NULL;
2201
2202                 /*
2203                  * locked_ref is the head node, so we have to go one
2204                  * node back for any delayed ref updates
2205                  */
2206                 ref = select_delayed_ref(locked_ref);
2207                 if (!ref) {
2208                         /* All delayed refs have been processed, Go ahead
2209                          * and send the head node to run_one_delayed_ref,
2210                          * so that any accounting fixes can happen
2211                          */
2212                         ref = &locked_ref->node;
2213
2214                         if (extent_op && must_insert_reserved) {
2215                                 kfree(extent_op);
2216                                 extent_op = NULL;
2217                         }
2218
2219                         if (extent_op) {
2220                                 spin_unlock(&delayed_refs->lock);
2221
2222                                 ret = run_delayed_extent_op(trans, root,
2223                                                             ref, extent_op);
2224                                 BUG_ON(ret);
2225                                 kfree(extent_op);
2226
2227                                 cond_resched();
2228                                 spin_lock(&delayed_refs->lock);
2229                                 continue;
2230                         }
2231
2232                         list_del_init(&locked_ref->cluster);
2233                         locked_ref = NULL;
2234                 }
2235
2236                 ref->in_tree = 0;
2237                 rb_erase(&ref->rb_node, &delayed_refs->root);
2238                 delayed_refs->num_entries--;
2239
2240                 spin_unlock(&delayed_refs->lock);
2241
2242                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2243                                           must_insert_reserved);
2244                 BUG_ON(ret);
2245
2246                 btrfs_put_delayed_ref(ref);
2247                 kfree(extent_op);
2248                 count++;
2249
2250                 cond_resched();
2251                 spin_lock(&delayed_refs->lock);
2252         }
2253         return count;
2254 }
2255
2256 /*
2257  * this starts processing the delayed reference count updates and
2258  * extent insertions we have queued up so far.  count can be
2259  * 0, which means to process everything in the tree at the start
2260  * of the run (but not newly added entries), or it can be some target
2261  * number you'd like to process.
2262  */
2263 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2264                            struct btrfs_root *root, unsigned long count)
2265 {
2266         struct rb_node *node;
2267         struct btrfs_delayed_ref_root *delayed_refs;
2268         struct btrfs_delayed_ref_node *ref;
2269         struct list_head cluster;
2270         int ret;
2271         int run_all = count == (unsigned long)-1;
2272         int run_most = 0;
2273
2274         if (root == root->fs_info->extent_root)
2275                 root = root->fs_info->tree_root;
2276
2277         delayed_refs = &trans->transaction->delayed_refs;
2278         INIT_LIST_HEAD(&cluster);
2279 again:
2280         spin_lock(&delayed_refs->lock);
2281         if (count == 0) {
2282                 count = delayed_refs->num_entries * 2;
2283                 run_most = 1;
2284         }
2285         while (1) {
2286                 if (!(run_all || run_most) &&
2287                     delayed_refs->num_heads_ready < 64)
2288                         break;
2289
2290                 /*
2291                  * go find something we can process in the rbtree.  We start at
2292                  * the beginning of the tree, and then build a cluster
2293                  * of refs to process starting at the first one we are able to
2294                  * lock
2295                  */
2296                 ret = btrfs_find_ref_cluster(trans, &cluster,
2297                                              delayed_refs->run_delayed_start);
2298                 if (ret)
2299                         break;
2300
2301                 ret = run_clustered_refs(trans, root, &cluster);
2302                 BUG_ON(ret < 0);
2303
2304                 count -= min_t(unsigned long, ret, count);
2305
2306                 if (count == 0)
2307                         break;
2308         }
2309
2310         if (run_all) {
2311                 node = rb_first(&delayed_refs->root);
2312                 if (!node)
2313                         goto out;
2314                 count = (unsigned long)-1;
2315
2316                 while (node) {
2317                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2318                                        rb_node);
2319                         if (btrfs_delayed_ref_is_head(ref)) {
2320                                 struct btrfs_delayed_ref_head *head;
2321
2322                                 head = btrfs_delayed_node_to_head(ref);
2323                                 atomic_inc(&ref->refs);
2324
2325                                 spin_unlock(&delayed_refs->lock);
2326                                 /*
2327                                  * Mutex was contended, block until it's
2328                                  * released and try again
2329                                  */
2330                                 mutex_lock(&head->mutex);
2331                                 mutex_unlock(&head->mutex);
2332
2333                                 btrfs_put_delayed_ref(ref);
2334                                 cond_resched();
2335                                 goto again;
2336                         }
2337                         node = rb_next(node);
2338                 }
2339                 spin_unlock(&delayed_refs->lock);
2340                 schedule_timeout(1);
2341                 goto again;
2342         }
2343 out:
2344         spin_unlock(&delayed_refs->lock);
2345         return 0;
2346 }
2347
2348 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2349                                 struct btrfs_root *root,
2350                                 u64 bytenr, u64 num_bytes, u64 flags,
2351                                 int is_data)
2352 {
2353         struct btrfs_delayed_extent_op *extent_op;
2354         int ret;
2355
2356         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2357         if (!extent_op)
2358                 return -ENOMEM;
2359
2360         extent_op->flags_to_set = flags;
2361         extent_op->update_flags = 1;
2362         extent_op->update_key = 0;
2363         extent_op->is_data = is_data ? 1 : 0;
2364
2365         ret = btrfs_add_delayed_extent_op(trans, bytenr, num_bytes, extent_op);
2366         if (ret)
2367                 kfree(extent_op);
2368         return ret;
2369 }
2370
2371 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2372                                       struct btrfs_root *root,
2373                                       struct btrfs_path *path,
2374                                       u64 objectid, u64 offset, u64 bytenr)
2375 {
2376         struct btrfs_delayed_ref_head *head;
2377         struct btrfs_delayed_ref_node *ref;
2378         struct btrfs_delayed_data_ref *data_ref;
2379         struct btrfs_delayed_ref_root *delayed_refs;
2380         struct rb_node *node;
2381         int ret = 0;
2382
2383         ret = -ENOENT;
2384         delayed_refs = &trans->transaction->delayed_refs;
2385         spin_lock(&delayed_refs->lock);
2386         head = btrfs_find_delayed_ref_head(trans, bytenr);
2387         if (!head)
2388                 goto out;
2389
2390         if (!mutex_trylock(&head->mutex)) {
2391                 atomic_inc(&head->node.refs);
2392                 spin_unlock(&delayed_refs->lock);
2393
2394                 btrfs_release_path(path);
2395
2396                 /*
2397                  * Mutex was contended, block until it's released and let
2398                  * caller try again
2399                  */
2400                 mutex_lock(&head->mutex);
2401                 mutex_unlock(&head->mutex);
2402                 btrfs_put_delayed_ref(&head->node);
2403                 return -EAGAIN;
2404         }
2405
2406         node = rb_prev(&head->node.rb_node);
2407         if (!node)
2408                 goto out_unlock;
2409
2410         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2411
2412         if (ref->bytenr != bytenr)
2413                 goto out_unlock;
2414
2415         ret = 1;
2416         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2417                 goto out_unlock;
2418
2419         data_ref = btrfs_delayed_node_to_data_ref(ref);
2420
2421         node = rb_prev(node);
2422         if (node) {
2423                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2424                 if (ref->bytenr == bytenr)
2425                         goto out_unlock;
2426         }
2427
2428         if (data_ref->root != root->root_key.objectid ||
2429             data_ref->objectid != objectid || data_ref->offset != offset)
2430                 goto out_unlock;
2431
2432         ret = 0;
2433 out_unlock:
2434         mutex_unlock(&head->mutex);
2435 out:
2436         spin_unlock(&delayed_refs->lock);
2437         return ret;
2438 }
2439
2440 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2441                                         struct btrfs_root *root,
2442                                         struct btrfs_path *path,
2443                                         u64 objectid, u64 offset, u64 bytenr)
2444 {
2445         struct btrfs_root *extent_root = root->fs_info->extent_root;
2446         struct extent_buffer *leaf;
2447         struct btrfs_extent_data_ref *ref;
2448         struct btrfs_extent_inline_ref *iref;
2449         struct btrfs_extent_item *ei;
2450         struct btrfs_key key;
2451         u32 item_size;
2452         int ret;
2453
2454         key.objectid = bytenr;
2455         key.offset = (u64)-1;
2456         key.type = BTRFS_EXTENT_ITEM_KEY;
2457
2458         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2459         if (ret < 0)
2460                 goto out;
2461         BUG_ON(ret == 0);
2462
2463         ret = -ENOENT;
2464         if (path->slots[0] == 0)
2465                 goto out;
2466
2467         path->slots[0]--;
2468         leaf = path->nodes[0];
2469         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2470
2471         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2472                 goto out;
2473
2474         ret = 1;
2475         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2476 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2477         if (item_size < sizeof(*ei)) {
2478                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2479                 goto out;
2480         }
2481 #endif
2482         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2483
2484         if (item_size != sizeof(*ei) +
2485             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2486                 goto out;
2487
2488         if (btrfs_extent_generation(leaf, ei) <=
2489             btrfs_root_last_snapshot(&root->root_item))
2490                 goto out;
2491
2492         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2493         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2494             BTRFS_EXTENT_DATA_REF_KEY)
2495                 goto out;
2496
2497         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2498         if (btrfs_extent_refs(leaf, ei) !=
2499             btrfs_extent_data_ref_count(leaf, ref) ||
2500             btrfs_extent_data_ref_root(leaf, ref) !=
2501             root->root_key.objectid ||
2502             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2503             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2504                 goto out;
2505
2506         ret = 0;
2507 out:
2508         return ret;
2509 }
2510
2511 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2512                           struct btrfs_root *root,
2513                           u64 objectid, u64 offset, u64 bytenr)
2514 {
2515         struct btrfs_path *path;
2516         int ret;
2517         int ret2;
2518
2519         path = btrfs_alloc_path();
2520         if (!path)
2521                 return -ENOENT;
2522
2523         do {
2524                 ret = check_committed_ref(trans, root, path, objectid,
2525                                           offset, bytenr);
2526                 if (ret && ret != -ENOENT)
2527                         goto out;
2528
2529                 ret2 = check_delayed_ref(trans, root, path, objectid,
2530                                          offset, bytenr);
2531         } while (ret2 == -EAGAIN);
2532
2533         if (ret2 && ret2 != -ENOENT) {
2534                 ret = ret2;
2535                 goto out;
2536         }
2537
2538         if (ret != -ENOENT || ret2 != -ENOENT)
2539                 ret = 0;
2540 out:
2541         btrfs_free_path(path);
2542         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2543                 WARN_ON(ret > 0);
2544         return ret;
2545 }
2546
2547 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2548                            struct btrfs_root *root,
2549                            struct extent_buffer *buf,
2550                            int full_backref, int inc)
2551 {
2552         u64 bytenr;
2553         u64 num_bytes;
2554         u64 parent;
2555         u64 ref_root;
2556         u32 nritems;
2557         struct btrfs_key key;
2558         struct btrfs_file_extent_item *fi;
2559         int i;
2560         int level;
2561         int ret = 0;
2562         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2563                             u64, u64, u64, u64, u64, u64);
2564
2565         ref_root = btrfs_header_owner(buf);
2566         nritems = btrfs_header_nritems(buf);
2567         level = btrfs_header_level(buf);
2568
2569         if (!root->ref_cows && level == 0)
2570                 return 0;
2571
2572         if (inc)
2573                 process_func = btrfs_inc_extent_ref;
2574         else
2575                 process_func = btrfs_free_extent;
2576
2577         if (full_backref)
2578                 parent = buf->start;
2579         else
2580                 parent = 0;
2581
2582         for (i = 0; i < nritems; i++) {
2583                 if (level == 0) {
2584                         btrfs_item_key_to_cpu(buf, &key, i);
2585                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2586                                 continue;
2587                         fi = btrfs_item_ptr(buf, i,
2588                                             struct btrfs_file_extent_item);
2589                         if (btrfs_file_extent_type(buf, fi) ==
2590                             BTRFS_FILE_EXTENT_INLINE)
2591                                 continue;
2592                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2593                         if (bytenr == 0)
2594                                 continue;
2595
2596                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2597                         key.offset -= btrfs_file_extent_offset(buf, fi);
2598                         ret = process_func(trans, root, bytenr, num_bytes,
2599                                            parent, ref_root, key.objectid,
2600                                            key.offset);
2601                         if (ret)
2602                                 goto fail;
2603                 } else {
2604                         bytenr = btrfs_node_blockptr(buf, i);
2605                         num_bytes = btrfs_level_size(root, level - 1);
2606                         ret = process_func(trans, root, bytenr, num_bytes,
2607                                            parent, ref_root, level - 1, 0);
2608                         if (ret)
2609                                 goto fail;
2610                 }
2611         }
2612         return 0;
2613 fail:
2614         BUG();
2615         return ret;
2616 }
2617
2618 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2619                   struct extent_buffer *buf, int full_backref)
2620 {
2621         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2622 }
2623
2624 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2625                   struct extent_buffer *buf, int full_backref)
2626 {
2627         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2628 }
2629
2630 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2631                                  struct btrfs_root *root,
2632                                  struct btrfs_path *path,
2633                                  struct btrfs_block_group_cache *cache)
2634 {
2635         int ret;
2636         struct btrfs_root *extent_root = root->fs_info->extent_root;
2637         unsigned long bi;
2638         struct extent_buffer *leaf;
2639
2640         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2641         if (ret < 0)
2642                 goto fail;
2643         BUG_ON(ret);
2644
2645         leaf = path->nodes[0];
2646         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2647         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2648         btrfs_mark_buffer_dirty(leaf);
2649         btrfs_release_path(path);
2650 fail:
2651         if (ret)
2652                 return ret;
2653         return 0;
2654
2655 }
2656
2657 static struct btrfs_block_group_cache *
2658 next_block_group(struct btrfs_root *root,
2659                  struct btrfs_block_group_cache *cache)
2660 {
2661         struct rb_node *node;
2662         spin_lock(&root->fs_info->block_group_cache_lock);
2663         node = rb_next(&cache->cache_node);
2664         btrfs_put_block_group(cache);
2665         if (node) {
2666                 cache = rb_entry(node, struct btrfs_block_group_cache,
2667                                  cache_node);
2668                 btrfs_get_block_group(cache);
2669         } else
2670                 cache = NULL;
2671         spin_unlock(&root->fs_info->block_group_cache_lock);
2672         return cache;
2673 }
2674
2675 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2676                             struct btrfs_trans_handle *trans,
2677                             struct btrfs_path *path)
2678 {
2679         struct btrfs_root *root = block_group->fs_info->tree_root;
2680         struct inode *inode = NULL;
2681         u64 alloc_hint = 0;
2682         int dcs = BTRFS_DC_ERROR;
2683         int num_pages = 0;
2684         int retries = 0;
2685         int ret = 0;
2686
2687         /*
2688          * If this block group is smaller than 100 megs don't bother caching the
2689          * block group.
2690          */
2691         if (block_group->key.offset < (100 * 1024 * 1024)) {
2692                 spin_lock(&block_group->lock);
2693                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2694                 spin_unlock(&block_group->lock);
2695                 return 0;
2696         }
2697
2698 again:
2699         inode = lookup_free_space_inode(root, block_group, path);
2700         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2701                 ret = PTR_ERR(inode);
2702                 btrfs_release_path(path);
2703                 goto out;
2704         }
2705
2706         if (IS_ERR(inode)) {
2707                 BUG_ON(retries);
2708                 retries++;
2709
2710                 if (block_group->ro)
2711                         goto out_free;
2712
2713                 ret = create_free_space_inode(root, trans, block_group, path);
2714                 if (ret)
2715                         goto out_free;
2716                 goto again;
2717         }
2718
2719         /*
2720          * We want to set the generation to 0, that way if anything goes wrong
2721          * from here on out we know not to trust this cache when we load up next
2722          * time.
2723          */
2724         BTRFS_I(inode)->generation = 0;
2725         ret = btrfs_update_inode(trans, root, inode);
2726         WARN_ON(ret);
2727
2728         if (i_size_read(inode) > 0) {
2729                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2730                                                       inode);
2731                 if (ret)
2732                         goto out_put;
2733         }
2734
2735         spin_lock(&block_group->lock);
2736         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2737                 /* We're not cached, don't bother trying to write stuff out */
2738                 dcs = BTRFS_DC_WRITTEN;
2739                 spin_unlock(&block_group->lock);
2740                 goto out_put;
2741         }
2742         spin_unlock(&block_group->lock);
2743
2744         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2745         if (!num_pages)
2746                 num_pages = 1;
2747
2748         /*
2749          * Just to make absolutely sure we have enough space, we're going to
2750          * preallocate 12 pages worth of space for each block group.  In
2751          * practice we ought to use at most 8, but we need extra space so we can
2752          * add our header and have a terminator between the extents and the
2753          * bitmaps.
2754          */
2755         num_pages *= 16;
2756         num_pages *= PAGE_CACHE_SIZE;
2757
2758         ret = btrfs_delalloc_reserve_space(inode, num_pages);
2759         if (ret)
2760                 goto out_put;
2761
2762         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2763                                               num_pages, num_pages,
2764                                               &alloc_hint);
2765         if (!ret) {
2766                 dcs = BTRFS_DC_SETUP;
2767                 btrfs_free_reserved_data_space(inode, num_pages);
2768         } else {
2769                 btrfs_delalloc_release_space(inode, num_pages);
2770         }
2771
2772 out_put:
2773         iput(inode);
2774 out_free:
2775         btrfs_release_path(path);
2776 out:
2777         spin_lock(&block_group->lock);
2778         block_group->disk_cache_state = dcs;
2779         spin_unlock(&block_group->lock);
2780
2781         return ret;
2782 }
2783
2784 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2785                                    struct btrfs_root *root)
2786 {
2787         struct btrfs_block_group_cache *cache;
2788         int err = 0;
2789         struct btrfs_path *path;
2790         u64 last = 0;
2791
2792         path = btrfs_alloc_path();
2793         if (!path)
2794                 return -ENOMEM;
2795
2796 again:
2797         while (1) {
2798                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2799                 while (cache) {
2800                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2801                                 break;
2802                         cache = next_block_group(root, cache);
2803                 }
2804                 if (!cache) {
2805                         if (last == 0)
2806                                 break;
2807                         last = 0;
2808                         continue;
2809                 }
2810                 err = cache_save_setup(cache, trans, path);
2811                 last = cache->key.objectid + cache->key.offset;
2812                 btrfs_put_block_group(cache);
2813         }
2814
2815         while (1) {
2816                 if (last == 0) {
2817                         err = btrfs_run_delayed_refs(trans, root,
2818                                                      (unsigned long)-1);
2819                         BUG_ON(err);
2820                 }
2821
2822                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2823                 while (cache) {
2824                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2825                                 btrfs_put_block_group(cache);
2826                                 goto again;
2827                         }
2828
2829                         if (cache->dirty)
2830                                 break;
2831                         cache = next_block_group(root, cache);
2832                 }
2833                 if (!cache) {
2834                         if (last == 0)
2835                                 break;
2836                         last = 0;
2837                         continue;
2838                 }
2839
2840                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
2841                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
2842                 cache->dirty = 0;
2843                 last = cache->key.objectid + cache->key.offset;
2844
2845                 err = write_one_cache_group(trans, root, path, cache);
2846                 BUG_ON(err);
2847                 btrfs_put_block_group(cache);
2848         }
2849
2850         while (1) {
2851                 /*
2852                  * I don't think this is needed since we're just marking our
2853                  * preallocated extent as written, but just in case it can't
2854                  * hurt.
2855                  */
2856                 if (last == 0) {
2857                         err = btrfs_run_delayed_refs(trans, root,
2858                                                      (unsigned long)-1);
2859                         BUG_ON(err);
2860                 }
2861
2862                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2863                 while (cache) {
2864                         /*
2865                          * Really this shouldn't happen, but it could if we
2866                          * couldn't write the entire preallocated extent and
2867                          * splitting the extent resulted in a new block.
2868                          */
2869                         if (cache->dirty) {
2870                                 btrfs_put_block_group(cache);
2871                                 goto again;
2872                         }
2873                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2874                                 break;
2875                         cache = next_block_group(root, cache);
2876                 }
2877                 if (!cache) {
2878                         if (last == 0)
2879                                 break;
2880                         last = 0;
2881                         continue;
2882                 }
2883
2884                 btrfs_write_out_cache(root, trans, cache, path);
2885
2886                 /*
2887                  * If we didn't have an error then the cache state is still
2888                  * NEED_WRITE, so we can set it to WRITTEN.
2889                  */
2890                 if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2891                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
2892                 last = cache->key.objectid + cache->key.offset;
2893                 btrfs_put_block_group(cache);
2894         }
2895
2896         btrfs_free_path(path);
2897         return 0;
2898 }
2899
2900 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
2901 {
2902         struct btrfs_block_group_cache *block_group;
2903         int readonly = 0;
2904
2905         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
2906         if (!block_group || block_group->ro)
2907                 readonly = 1;
2908         if (block_group)
2909                 btrfs_put_block_group(block_group);
2910         return readonly;
2911 }
2912
2913 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
2914                              u64 total_bytes, u64 bytes_used,
2915                              struct btrfs_space_info **space_info)
2916 {
2917         struct btrfs_space_info *found;
2918         int i;
2919         int factor;
2920
2921         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
2922                      BTRFS_BLOCK_GROUP_RAID10))
2923                 factor = 2;
2924         else
2925                 factor = 1;
2926
2927         found = __find_space_info(info, flags);
2928         if (found) {
2929                 spin_lock(&found->lock);
2930                 found->total_bytes += total_bytes;
2931                 found->disk_total += total_bytes * factor;
2932                 found->bytes_used += bytes_used;
2933                 found->disk_used += bytes_used * factor;
2934                 found->full = 0;
2935                 spin_unlock(&found->lock);
2936                 *space_info = found;
2937                 return 0;
2938         }
2939         found = kzalloc(sizeof(*found), GFP_NOFS);
2940         if (!found)
2941                 return -ENOMEM;
2942
2943         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
2944                 INIT_LIST_HEAD(&found->block_groups[i]);
2945         init_rwsem(&found->groups_sem);
2946         spin_lock_init(&found->lock);
2947         found->flags = flags & (BTRFS_BLOCK_GROUP_DATA |
2948                                 BTRFS_BLOCK_GROUP_SYSTEM |
2949                                 BTRFS_BLOCK_GROUP_METADATA);
2950         found->total_bytes = total_bytes;
2951         found->disk_total = total_bytes * factor;
2952         found->bytes_used = bytes_used;
2953         found->disk_used = bytes_used * factor;
2954         found->bytes_pinned = 0;
2955         found->bytes_reserved = 0;
2956         found->bytes_readonly = 0;
2957         found->bytes_may_use = 0;
2958         found->full = 0;
2959         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
2960         found->chunk_alloc = 0;
2961         found->flush = 0;
2962         init_waitqueue_head(&found->wait);
2963         *space_info = found;
2964         list_add_rcu(&found->list, &info->space_info);
2965         return 0;
2966 }
2967
2968 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
2969 {
2970         u64 extra_flags = flags & (BTRFS_BLOCK_GROUP_RAID0 |
2971                                    BTRFS_BLOCK_GROUP_RAID1 |
2972                                    BTRFS_BLOCK_GROUP_RAID10 |
2973                                    BTRFS_BLOCK_GROUP_DUP);
2974         if (extra_flags) {
2975                 if (flags & BTRFS_BLOCK_GROUP_DATA)
2976                         fs_info->avail_data_alloc_bits |= extra_flags;
2977                 if (flags & BTRFS_BLOCK_GROUP_METADATA)
2978                         fs_info->avail_metadata_alloc_bits |= extra_flags;
2979                 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
2980                         fs_info->avail_system_alloc_bits |= extra_flags;
2981         }
2982 }
2983
2984 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
2985 {
2986         /*
2987          * we add in the count of missing devices because we want
2988          * to make sure that any RAID levels on a degraded FS
2989          * continue to be honored.
2990          */
2991         u64 num_devices = root->fs_info->fs_devices->rw_devices +
2992                 root->fs_info->fs_devices->missing_devices;
2993
2994         if (num_devices == 1)
2995                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
2996         if (num_devices < 4)
2997                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
2998
2999         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3000             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3001                       BTRFS_BLOCK_GROUP_RAID10))) {
3002                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3003         }
3004
3005         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3006             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3007                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3008         }
3009
3010         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3011             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3012              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3013              (flags & BTRFS_BLOCK_GROUP_DUP)))
3014                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3015         return flags;
3016 }
3017
3018 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3019 {
3020         if (flags & BTRFS_BLOCK_GROUP_DATA)
3021                 flags |= root->fs_info->avail_data_alloc_bits &
3022                          root->fs_info->data_alloc_profile;
3023         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3024                 flags |= root->fs_info->avail_system_alloc_bits &
3025                          root->fs_info->system_alloc_profile;
3026         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3027                 flags |= root->fs_info->avail_metadata_alloc_bits &
3028                          root->fs_info->metadata_alloc_profile;
3029         return btrfs_reduce_alloc_profile(root, flags);
3030 }
3031
3032 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3033 {
3034         u64 flags;
3035
3036         if (data)
3037                 flags = BTRFS_BLOCK_GROUP_DATA;
3038         else if (root == root->fs_info->chunk_root)
3039                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3040         else
3041                 flags = BTRFS_BLOCK_GROUP_METADATA;
3042
3043         return get_alloc_profile(root, flags);
3044 }
3045
3046 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3047 {
3048         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3049                                                        BTRFS_BLOCK_GROUP_DATA);
3050 }
3051
3052 /*
3053  * This will check the space that the inode allocates from to make sure we have
3054  * enough space for bytes.
3055  */
3056 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3057 {
3058         struct btrfs_space_info *data_sinfo;
3059         struct btrfs_root *root = BTRFS_I(inode)->root;
3060         u64 used;
3061         int ret = 0, committed = 0, alloc_chunk = 1;
3062
3063         /* make sure bytes are sectorsize aligned */
3064         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3065
3066         if (root == root->fs_info->tree_root ||
3067             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3068                 alloc_chunk = 0;
3069                 committed = 1;
3070         }
3071
3072         data_sinfo = BTRFS_I(inode)->space_info;
3073         if (!data_sinfo)
3074                 goto alloc;
3075
3076 again:
3077         /* make sure we have enough space to handle the data first */
3078         spin_lock(&data_sinfo->lock);
3079         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3080                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3081                 data_sinfo->bytes_may_use;
3082
3083         if (used + bytes > data_sinfo->total_bytes) {
3084                 struct btrfs_trans_handle *trans;
3085
3086                 /*
3087                  * if we don't have enough free bytes in this space then we need
3088                  * to alloc a new chunk.
3089                  */
3090                 if (!data_sinfo->full && alloc_chunk) {
3091                         u64 alloc_target;
3092
3093                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3094                         spin_unlock(&data_sinfo->lock);
3095 alloc:
3096                         alloc_target = btrfs_get_alloc_profile(root, 1);
3097                         trans = btrfs_join_transaction(root);
3098                         if (IS_ERR(trans))
3099                                 return PTR_ERR(trans);
3100
3101                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3102                                              bytes + 2 * 1024 * 1024,
3103                                              alloc_target,
3104                                              CHUNK_ALLOC_NO_FORCE);
3105                         btrfs_end_transaction(trans, root);
3106                         if (ret < 0) {
3107                                 if (ret != -ENOSPC)
3108                                         return ret;
3109                                 else
3110                                         goto commit_trans;
3111                         }
3112
3113                         if (!data_sinfo) {
3114                                 btrfs_set_inode_space_info(root, inode);
3115                                 data_sinfo = BTRFS_I(inode)->space_info;
3116                         }
3117                         goto again;
3118                 }
3119
3120                 /*
3121                  * If we have less pinned bytes than we want to allocate then
3122                  * don't bother committing the transaction, it won't help us.
3123                  */
3124                 if (data_sinfo->bytes_pinned < bytes)
3125                         committed = 1;
3126                 spin_unlock(&data_sinfo->lock);
3127
3128                 /* commit the current transaction and try again */
3129 commit_trans:
3130                 if (!committed &&
3131                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3132                         committed = 1;
3133                         trans = btrfs_join_transaction(root);
3134                         if (IS_ERR(trans))
3135                                 return PTR_ERR(trans);
3136                         ret = btrfs_commit_transaction(trans, root);
3137                         if (ret)
3138                                 return ret;
3139                         goto again;
3140                 }
3141
3142                 return -ENOSPC;
3143         }
3144         data_sinfo->bytes_may_use += bytes;
3145         spin_unlock(&data_sinfo->lock);
3146
3147         return 0;
3148 }
3149
3150 /*
3151  * Called if we need to clear a data reservation for this inode.
3152  */
3153 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3154 {
3155         struct btrfs_root *root = BTRFS_I(inode)->root;
3156         struct btrfs_space_info *data_sinfo;
3157
3158         /* make sure bytes are sectorsize aligned */
3159         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3160
3161         data_sinfo = BTRFS_I(inode)->space_info;
3162         spin_lock(&data_sinfo->lock);
3163         data_sinfo->bytes_may_use -= bytes;
3164         spin_unlock(&data_sinfo->lock);
3165 }
3166
3167 static void force_metadata_allocation(struct btrfs_fs_info *info)
3168 {
3169         struct list_head *head = &info->space_info;
3170         struct btrfs_space_info *found;
3171
3172         rcu_read_lock();
3173         list_for_each_entry_rcu(found, head, list) {
3174                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3175                         found->force_alloc = CHUNK_ALLOC_FORCE;
3176         }
3177         rcu_read_unlock();
3178 }
3179
3180 static int should_alloc_chunk(struct btrfs_root *root,
3181                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3182                               int force)
3183 {
3184         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3185         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3186         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3187         u64 thresh;
3188
3189         if (force == CHUNK_ALLOC_FORCE)
3190                 return 1;
3191
3192         /*
3193          * We need to take into account the global rsv because for all intents
3194          * and purposes it's used space.  Don't worry about locking the
3195          * global_rsv, it doesn't change except when the transaction commits.
3196          */
3197         num_allocated += global_rsv->size;
3198
3199         /*
3200          * in limited mode, we want to have some free space up to
3201          * about 1% of the FS size.
3202          */
3203         if (force == CHUNK_ALLOC_LIMITED) {
3204                 thresh = btrfs_super_total_bytes(&root->fs_info->super_copy);
3205                 thresh = max_t(u64, 64 * 1024 * 1024,
3206                                div_factor_fine(thresh, 1));
3207
3208                 if (num_bytes - num_allocated < thresh)
3209                         return 1;
3210         }
3211
3212         /*
3213          * we have two similar checks here, one based on percentage
3214          * and once based on a hard number of 256MB.  The idea
3215          * is that if we have a good amount of free
3216          * room, don't allocate a chunk.  A good mount is
3217          * less than 80% utilized of the chunks we have allocated,
3218          * or more than 256MB free
3219          */
3220         if (num_allocated + alloc_bytes + 256 * 1024 * 1024 < num_bytes)
3221                 return 0;
3222
3223         if (num_allocated + alloc_bytes < div_factor(num_bytes, 8))
3224                 return 0;
3225
3226         thresh = btrfs_super_total_bytes(&root->fs_info->super_copy);
3227
3228         /* 256MB or 5% of the FS */
3229         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 5));
3230
3231         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 3))
3232                 return 0;
3233         return 1;
3234 }
3235
3236 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3237                           struct btrfs_root *extent_root, u64 alloc_bytes,
3238                           u64 flags, int force)
3239 {
3240         struct btrfs_space_info *space_info;
3241         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3242         int wait_for_alloc = 0;
3243         int ret = 0;
3244
3245         flags = btrfs_reduce_alloc_profile(extent_root, flags);
3246
3247         space_info = __find_space_info(extent_root->fs_info, flags);
3248         if (!space_info) {
3249                 ret = update_space_info(extent_root->fs_info, flags,
3250                                         0, 0, &space_info);
3251                 BUG_ON(ret);
3252         }
3253         BUG_ON(!space_info);
3254
3255 again:
3256         spin_lock(&space_info->lock);
3257         if (space_info->force_alloc)
3258                 force = space_info->force_alloc;
3259         if (space_info->full) {
3260                 spin_unlock(&space_info->lock);
3261                 return 0;
3262         }
3263
3264         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3265                 spin_unlock(&space_info->lock);
3266                 return 0;
3267         } else if (space_info->chunk_alloc) {
3268                 wait_for_alloc = 1;
3269         } else {
3270                 space_info->chunk_alloc = 1;
3271         }
3272
3273         spin_unlock(&space_info->lock);
3274
3275         mutex_lock(&fs_info->chunk_mutex);
3276
3277         /*
3278          * The chunk_mutex is held throughout the entirety of a chunk
3279          * allocation, so once we've acquired the chunk_mutex we know that the
3280          * other guy is done and we need to recheck and see if we should
3281          * allocate.
3282          */
3283         if (wait_for_alloc) {
3284                 mutex_unlock(&fs_info->chunk_mutex);
3285                 wait_for_alloc = 0;
3286                 goto again;
3287         }
3288
3289         /*
3290          * If we have mixed data/metadata chunks we want to make sure we keep
3291          * allocating mixed chunks instead of individual chunks.
3292          */
3293         if (btrfs_mixed_space_info(space_info))
3294                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3295
3296         /*
3297          * if we're doing a data chunk, go ahead and make sure that
3298          * we keep a reasonable number of metadata chunks allocated in the
3299          * FS as well.
3300          */
3301         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3302                 fs_info->data_chunk_allocations++;
3303                 if (!(fs_info->data_chunk_allocations %
3304                       fs_info->metadata_ratio))
3305                         force_metadata_allocation(fs_info);
3306         }
3307
3308         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3309         if (ret < 0 && ret != -ENOSPC)
3310                 goto out;
3311
3312         spin_lock(&space_info->lock);
3313         if (ret)
3314                 space_info->full = 1;
3315         else
3316                 ret = 1;
3317
3318         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3319         space_info->chunk_alloc = 0;
3320         spin_unlock(&space_info->lock);
3321 out:
3322         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3323         return ret;
3324 }
3325
3326 /*
3327  * shrink metadata reservation for delalloc
3328  */
3329 static int shrink_delalloc(struct btrfs_trans_handle *trans,
3330                            struct btrfs_root *root, u64 to_reclaim, int sync)
3331 {
3332         struct btrfs_block_rsv *block_rsv;
3333         struct btrfs_space_info *space_info;
3334         u64 reserved;
3335         u64 max_reclaim;
3336         u64 reclaimed = 0;
3337         long time_left;
3338         int nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3339         int loops = 0;
3340         unsigned long progress;
3341
3342         block_rsv = &root->fs_info->delalloc_block_rsv;
3343         space_info = block_rsv->space_info;
3344
3345         smp_mb();
3346         reserved = space_info->bytes_may_use;
3347         progress = space_info->reservation_progress;
3348
3349         if (reserved == 0)
3350                 return 0;
3351
3352         smp_mb();
3353         if (root->fs_info->delalloc_bytes == 0) {
3354                 if (trans)
3355                         return 0;
3356                 btrfs_wait_ordered_extents(root, 0, 0);
3357                 return 0;
3358         }
3359
3360         max_reclaim = min(reserved, to_reclaim);
3361
3362         while (loops < 1024) {
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_may_use)
3371                         reclaimed += reserved - space_info->bytes_may_use;
3372                 reserved = space_info->bytes_may_use;
3373                 spin_unlock(&space_info->lock);
3374
3375                 loops++;
3376
3377                 if (reserved == 0 || reclaimed >= max_reclaim)
3378                         break;
3379
3380                 if (trans && trans->transaction->blocked)
3381                         return -EAGAIN;
3382
3383                 time_left = schedule_timeout_interruptible(1);
3384
3385                 /* We were interrupted, exit */
3386                 if (time_left)
3387                         break;
3388
3389                 /* we've kicked the IO a few times, if anything has been freed,
3390                  * exit.  There is no sense in looping here for a long time
3391                  * when we really need to commit the transaction, or there are
3392                  * just too many writers without enough free space
3393                  */
3394
3395                 if (loops > 3) {
3396                         smp_mb();
3397                         if (progress != space_info->reservation_progress)
3398                                 break;
3399                 }
3400
3401         }
3402         if (reclaimed >= to_reclaim && !trans)
3403                 btrfs_wait_ordered_extents(root, 0, 0);
3404         return reclaimed >= to_reclaim;
3405 }
3406
3407 /*
3408  * Retries tells us how many times we've called reserve_metadata_bytes.  The
3409  * idea is if this is the first call (retries == 0) then we will add to our
3410  * reserved count if we can't make the allocation in order to hold our place
3411  * while we go and try and free up space.  That way for retries > 1 we don't try
3412  * and add space, we just check to see if the amount of unused space is >= the
3413  * total space, meaning that our reservation is valid.
3414  *
3415  * However if we don't intend to retry this reservation, pass -1 as retries so
3416  * that it short circuits this logic.
3417  */
3418 static int reserve_metadata_bytes(struct btrfs_trans_handle *trans,
3419                                   struct btrfs_root *root,
3420                                   struct btrfs_block_rsv *block_rsv,
3421                                   u64 orig_bytes, int flush)
3422 {
3423         struct btrfs_space_info *space_info = block_rsv->space_info;
3424         u64 unused;
3425         u64 num_bytes = orig_bytes;
3426         int retries = 0;
3427         int ret = 0;
3428         bool committed = false;
3429         bool flushing = false;
3430 again:
3431         ret = 0;
3432         spin_lock(&space_info->lock);
3433         /*
3434          * We only want to wait if somebody other than us is flushing and we are
3435          * actually alloed to flush.
3436          */
3437         while (flush && !flushing && space_info->flush) {
3438                 spin_unlock(&space_info->lock);
3439                 /*
3440                  * If we have a trans handle we can't wait because the flusher
3441                  * may have to commit the transaction, which would mean we would
3442                  * deadlock since we are waiting for the flusher to finish, but
3443                  * hold the current transaction open.
3444                  */
3445                 if (trans)
3446                         return -EAGAIN;
3447                 ret = wait_event_interruptible(space_info->wait,
3448                                                !space_info->flush);
3449                 /* Must have been interrupted, return */
3450                 if (ret)
3451                         return -EINTR;
3452
3453                 spin_lock(&space_info->lock);
3454         }
3455
3456         ret = -ENOSPC;
3457         unused = space_info->bytes_used + space_info->bytes_reserved +
3458                  space_info->bytes_pinned + space_info->bytes_readonly +
3459                  space_info->bytes_may_use;
3460
3461         /*
3462          * The idea here is that we've not already over-reserved the block group
3463          * then we can go ahead and save our reservation first and then start
3464          * flushing if we need to.  Otherwise if we've already overcommitted
3465          * lets start flushing stuff first and then come back and try to make
3466          * our reservation.
3467          */
3468         if (unused <= space_info->total_bytes) {
3469                 unused = space_info->total_bytes - unused;
3470                 if (unused >= num_bytes) {
3471                         space_info->bytes_may_use += orig_bytes;
3472                         ret = 0;
3473                 } else {
3474                         /*
3475                          * Ok set num_bytes to orig_bytes since we aren't
3476                          * overocmmitted, this way we only try and reclaim what
3477                          * we need.
3478                          */
3479                         num_bytes = orig_bytes;
3480                 }
3481         } else {
3482                 /*
3483                  * Ok we're over committed, set num_bytes to the overcommitted
3484                  * amount plus the amount of bytes that we need for this
3485                  * reservation.
3486                  */
3487                 num_bytes = unused - space_info->total_bytes +
3488                         (orig_bytes * (retries + 1));
3489         }
3490
3491         /*
3492          * Couldn't make our reservation, save our place so while we're trying
3493          * to reclaim space we can actually use it instead of somebody else
3494          * stealing it from us.
3495          */
3496         if (ret && flush) {
3497                 flushing = true;
3498                 space_info->flush = 1;
3499         }
3500
3501         spin_unlock(&space_info->lock);
3502
3503         if (!ret || !flush)
3504                 goto out;
3505
3506         /*
3507          * We do synchronous shrinking since we don't actually unreserve
3508          * metadata until after the IO is completed.
3509          */
3510         ret = shrink_delalloc(trans, root, num_bytes, 1);
3511         if (ret < 0)
3512                 goto out;
3513
3514         ret = 0;
3515
3516         /*
3517          * So if we were overcommitted it's possible that somebody else flushed
3518          * out enough space and we simply didn't have enough space to reclaim,
3519          * so go back around and try again.
3520          */
3521         if (retries < 2) {
3522                 retries++;
3523                 goto again;
3524         }
3525
3526         /*
3527          * Not enough space to be reclaimed, don't bother committing the
3528          * transaction.
3529          */
3530         spin_lock(&space_info->lock);
3531         if (space_info->bytes_pinned < orig_bytes)
3532                 ret = -ENOSPC;
3533         spin_unlock(&space_info->lock);
3534         if (ret)
3535                 goto out;
3536
3537         ret = -EAGAIN;
3538         if (trans)
3539                 goto out;
3540
3541         ret = -ENOSPC;
3542         if (committed)
3543                 goto out;
3544
3545         trans = btrfs_join_transaction(root);
3546         if (IS_ERR(trans))
3547                 goto out;
3548         ret = btrfs_commit_transaction(trans, root);
3549         if (!ret) {
3550                 trans = NULL;
3551                 committed = true;
3552                 goto again;
3553         }
3554
3555 out:
3556         if (flushing) {
3557                 spin_lock(&space_info->lock);
3558                 space_info->flush = 0;
3559                 wake_up_all(&space_info->wait);
3560                 spin_unlock(&space_info->lock);
3561         }
3562         return ret;
3563 }
3564
3565 static struct btrfs_block_rsv *get_block_rsv(struct btrfs_trans_handle *trans,
3566                                              struct btrfs_root *root)
3567 {
3568         struct btrfs_block_rsv *block_rsv = NULL;
3569
3570         if (root->ref_cows || root == root->fs_info->csum_root)
3571                 block_rsv = trans->block_rsv;
3572
3573         if (!block_rsv)
3574                 block_rsv = root->block_rsv;
3575
3576         if (!block_rsv)
3577                 block_rsv = &root->fs_info->empty_block_rsv;
3578
3579         return block_rsv;
3580 }
3581
3582 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3583                                u64 num_bytes)
3584 {
3585         int ret = -ENOSPC;
3586         spin_lock(&block_rsv->lock);
3587         if (block_rsv->reserved >= num_bytes) {
3588                 block_rsv->reserved -= num_bytes;
3589                 if (block_rsv->reserved < block_rsv->size)
3590                         block_rsv->full = 0;
3591                 ret = 0;
3592         }
3593         spin_unlock(&block_rsv->lock);
3594         return ret;
3595 }
3596
3597 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3598                                 u64 num_bytes, int update_size)
3599 {
3600         spin_lock(&block_rsv->lock);
3601         block_rsv->reserved += num_bytes;
3602         if (update_size)
3603                 block_rsv->size += num_bytes;
3604         else if (block_rsv->reserved >= block_rsv->size)
3605                 block_rsv->full = 1;
3606         spin_unlock(&block_rsv->lock);
3607 }
3608
3609 static void block_rsv_release_bytes(struct btrfs_block_rsv *block_rsv,
3610                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3611 {
3612         struct btrfs_space_info *space_info = block_rsv->space_info;
3613
3614         spin_lock(&block_rsv->lock);
3615         if (num_bytes == (u64)-1)
3616                 num_bytes = block_rsv->size;
3617         block_rsv->size -= num_bytes;
3618         if (block_rsv->reserved >= block_rsv->size) {
3619                 num_bytes = block_rsv->reserved - block_rsv->size;
3620                 block_rsv->reserved = block_rsv->size;
3621                 block_rsv->full = 1;
3622         } else {
3623                 num_bytes = 0;
3624         }
3625         spin_unlock(&block_rsv->lock);
3626
3627         if (num_bytes > 0) {
3628                 if (dest) {
3629                         spin_lock(&dest->lock);
3630                         if (!dest->full) {
3631                                 u64 bytes_to_add;
3632
3633                                 bytes_to_add = dest->size - dest->reserved;
3634                                 bytes_to_add = min(num_bytes, bytes_to_add);
3635                                 dest->reserved += bytes_to_add;
3636                                 if (dest->reserved >= dest->size)
3637                                         dest->full = 1;
3638                                 num_bytes -= bytes_to_add;
3639                         }
3640                         spin_unlock(&dest->lock);
3641                 }
3642                 if (num_bytes) {
3643                         spin_lock(&space_info->lock);
3644                         space_info->bytes_may_use -= num_bytes;
3645                         space_info->reservation_progress++;
3646                         spin_unlock(&space_info->lock);
3647                 }
3648         }
3649 }
3650
3651 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3652                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3653 {
3654         int ret;
3655
3656         ret = block_rsv_use_bytes(src, num_bytes);
3657         if (ret)
3658                 return ret;
3659
3660         block_rsv_add_bytes(dst, num_bytes, 1);
3661         return 0;
3662 }
3663
3664 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3665 {
3666         memset(rsv, 0, sizeof(*rsv));
3667         spin_lock_init(&rsv->lock);
3668 }
3669
3670 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3671 {
3672         struct btrfs_block_rsv *block_rsv;
3673         struct btrfs_fs_info *fs_info = root->fs_info;
3674
3675         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3676         if (!block_rsv)
3677                 return NULL;
3678
3679         btrfs_init_block_rsv(block_rsv);
3680         block_rsv->space_info = __find_space_info(fs_info,
3681                                                   BTRFS_BLOCK_GROUP_METADATA);
3682         return block_rsv;
3683 }
3684
3685 void btrfs_free_block_rsv(struct btrfs_root *root,
3686                           struct btrfs_block_rsv *rsv)
3687 {
3688         btrfs_block_rsv_release(root, rsv, (u64)-1);
3689         kfree(rsv);
3690 }
3691
3692 int btrfs_block_rsv_add(struct btrfs_trans_handle *trans,
3693                         struct btrfs_root *root,
3694                         struct btrfs_block_rsv *block_rsv,
3695                         u64 num_bytes)
3696 {
3697         int ret;
3698
3699         if (num_bytes == 0)
3700                 return 0;
3701
3702         ret = reserve_metadata_bytes(trans, root, block_rsv, num_bytes, 1);
3703         if (!ret) {
3704                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3705                 return 0;
3706         }
3707
3708         return ret;
3709 }
3710
3711 int btrfs_block_rsv_check(struct btrfs_trans_handle *trans,
3712                           struct btrfs_root *root,
3713                           struct btrfs_block_rsv *block_rsv,
3714                           u64 min_reserved, int min_factor, int flush)
3715 {
3716         u64 num_bytes = 0;
3717         int ret = -ENOSPC;
3718
3719         if (!block_rsv)
3720                 return 0;
3721
3722         spin_lock(&block_rsv->lock);
3723         if (min_factor > 0)
3724                 num_bytes = div_factor(block_rsv->size, min_factor);
3725         if (min_reserved > num_bytes)
3726                 num_bytes = min_reserved;
3727
3728         if (block_rsv->reserved >= num_bytes)
3729                 ret = 0;
3730         else
3731                 num_bytes -= block_rsv->reserved;
3732         spin_unlock(&block_rsv->lock);
3733
3734         if (!ret)
3735                 return 0;
3736
3737         ret = reserve_metadata_bytes(trans, root, block_rsv, num_bytes, flush);
3738         if (!ret) {
3739                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
3740                 return 0;
3741         }
3742
3743         return ret;
3744 }
3745
3746 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
3747                             struct btrfs_block_rsv *dst_rsv,
3748                             u64 num_bytes)
3749 {
3750         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3751 }
3752
3753 void btrfs_block_rsv_release(struct btrfs_root *root,
3754                              struct btrfs_block_rsv *block_rsv,
3755                              u64 num_bytes)
3756 {
3757         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3758         if (global_rsv->full || global_rsv == block_rsv ||
3759             block_rsv->space_info != global_rsv->space_info)
3760                 global_rsv = NULL;
3761         block_rsv_release_bytes(block_rsv, global_rsv, num_bytes);
3762 }
3763
3764 /*
3765  * helper to calculate size of global block reservation.
3766  * the desired value is sum of space used by extent tree,
3767  * checksum tree and root tree
3768  */
3769 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
3770 {
3771         struct btrfs_space_info *sinfo;
3772         u64 num_bytes;
3773         u64 meta_used;
3774         u64 data_used;
3775         int csum_size = btrfs_super_csum_size(&fs_info->super_copy);
3776
3777         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
3778         spin_lock(&sinfo->lock);
3779         data_used = sinfo->bytes_used;
3780         spin_unlock(&sinfo->lock);
3781
3782         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3783         spin_lock(&sinfo->lock);
3784         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
3785                 data_used = 0;
3786         meta_used = sinfo->bytes_used;
3787         spin_unlock(&sinfo->lock);
3788
3789         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
3790                     csum_size * 2;
3791         num_bytes += div64_u64(data_used + meta_used, 50);
3792
3793         if (num_bytes * 3 > meta_used)
3794                 num_bytes = div64_u64(meta_used, 3);
3795
3796         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
3797 }
3798
3799 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
3800 {
3801         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
3802         struct btrfs_space_info *sinfo = block_rsv->space_info;
3803         u64 num_bytes;
3804
3805         num_bytes = calc_global_metadata_size(fs_info);
3806
3807         spin_lock(&block_rsv->lock);
3808         spin_lock(&sinfo->lock);
3809
3810         block_rsv->size = num_bytes;
3811
3812         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
3813                     sinfo->bytes_reserved + sinfo->bytes_readonly +
3814                     sinfo->bytes_may_use;
3815
3816         if (sinfo->total_bytes > num_bytes) {
3817                 num_bytes = sinfo->total_bytes - num_bytes;
3818                 block_rsv->reserved += num_bytes;
3819                 sinfo->bytes_may_use += num_bytes;
3820         }
3821
3822         if (block_rsv->reserved >= block_rsv->size) {
3823                 num_bytes = block_rsv->reserved - block_rsv->size;
3824                 sinfo->bytes_may_use -= num_bytes;
3825                 sinfo->reservation_progress++;
3826                 block_rsv->reserved = block_rsv->size;
3827                 block_rsv->full = 1;
3828         }
3829
3830         spin_unlock(&sinfo->lock);
3831         spin_unlock(&block_rsv->lock);
3832 }
3833
3834 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
3835 {
3836         struct btrfs_space_info *space_info;
3837
3838         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3839         fs_info->chunk_block_rsv.space_info = space_info;
3840
3841         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3842         fs_info->global_block_rsv.space_info = space_info;
3843         fs_info->delalloc_block_rsv.space_info = space_info;
3844         fs_info->trans_block_rsv.space_info = space_info;
3845         fs_info->empty_block_rsv.space_info = space_info;
3846
3847         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
3848         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
3849         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
3850         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
3851         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
3852
3853         update_global_block_rsv(fs_info);
3854 }
3855
3856 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
3857 {
3858         block_rsv_release_bytes(&fs_info->global_block_rsv, NULL, (u64)-1);
3859         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
3860         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
3861         WARN_ON(fs_info->trans_block_rsv.size > 0);
3862         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
3863         WARN_ON(fs_info->chunk_block_rsv.size > 0);
3864         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
3865 }
3866
3867 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
3868                                   struct btrfs_root *root)
3869 {
3870         struct btrfs_block_rsv *block_rsv;
3871
3872         if (!trans->bytes_reserved)
3873                 return;
3874
3875         block_rsv = &root->fs_info->trans_block_rsv;
3876         btrfs_block_rsv_release(root, block_rsv, trans->bytes_reserved);
3877         trans->bytes_reserved = 0;
3878 }
3879
3880 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
3881                                   struct inode *inode)
3882 {
3883         struct btrfs_root *root = BTRFS_I(inode)->root;
3884         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
3885         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
3886
3887         /*
3888          * We need to hold space in order to delete our orphan item once we've
3889          * added it, so this takes the reservation so we can release it later
3890          * when we are truly done with the orphan item.
3891          */
3892         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
3893         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3894 }
3895
3896 void btrfs_orphan_release_metadata(struct inode *inode)
3897 {
3898         struct btrfs_root *root = BTRFS_I(inode)->root;
3899         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
3900         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
3901 }
3902
3903 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
3904                                 struct btrfs_pending_snapshot *pending)
3905 {
3906         struct btrfs_root *root = pending->root;
3907         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
3908         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
3909         /*
3910          * two for root back/forward refs, two for directory entries
3911          * and one for root of the snapshot.
3912          */
3913         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
3914         dst_rsv->space_info = src_rsv->space_info;
3915         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3916 }
3917
3918 /**
3919  * drop_outstanding_extent - drop an outstanding extent
3920  * @inode: the inode we're dropping the extent for
3921  *
3922  * This is called when we are freeing up an outstanding extent, either called
3923  * after an error or after an extent is written.  This will return the number of
3924  * reserved extents that need to be freed.  This must be called with
3925  * BTRFS_I(inode)->lock held.
3926  */
3927 static unsigned drop_outstanding_extent(struct inode *inode)
3928 {
3929         unsigned dropped_extents = 0;
3930
3931         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
3932         BTRFS_I(inode)->outstanding_extents--;
3933
3934         /*
3935          * If we have more or the same amount of outsanding extents than we have
3936          * reserved then we need to leave the reserved extents count alone.
3937          */
3938         if (BTRFS_I(inode)->outstanding_extents >=
3939             BTRFS_I(inode)->reserved_extents)
3940                 return 0;
3941
3942         dropped_extents = BTRFS_I(inode)->reserved_extents -
3943                 BTRFS_I(inode)->outstanding_extents;
3944         BTRFS_I(inode)->reserved_extents -= dropped_extents;
3945         return dropped_extents;
3946 }
3947
3948 /**
3949  * calc_csum_metadata_size - return the amount of metada space that must be
3950  *      reserved/free'd for the given bytes.
3951  * @inode: the inode we're manipulating
3952  * @num_bytes: the number of bytes in question
3953  * @reserve: 1 if we are reserving space, 0 if we are freeing space
3954  *
3955  * This adjusts the number of csum_bytes in the inode and then returns the
3956  * correct amount of metadata that must either be reserved or freed.  We
3957  * calculate how many checksums we can fit into one leaf and then divide the
3958  * number of bytes that will need to be checksumed by this value to figure out
3959  * how many checksums will be required.  If we are adding bytes then the number
3960  * may go up and we will return the number of additional bytes that must be
3961  * reserved.  If it is going down we will return the number of bytes that must
3962  * be freed.
3963  *
3964  * This must be called with BTRFS_I(inode)->lock held.
3965  */
3966 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
3967                                    int reserve)
3968 {
3969         struct btrfs_root *root = BTRFS_I(inode)->root;
3970         u64 csum_size;
3971         int num_csums_per_leaf;
3972         int num_csums;
3973         int old_csums;
3974
3975         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
3976             BTRFS_I(inode)->csum_bytes == 0)
3977                 return 0;
3978
3979         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
3980         if (reserve)
3981                 BTRFS_I(inode)->csum_bytes += num_bytes;
3982         else
3983                 BTRFS_I(inode)->csum_bytes -= num_bytes;
3984         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
3985         num_csums_per_leaf = (int)div64_u64(csum_size,
3986                                             sizeof(struct btrfs_csum_item) +
3987                                             sizeof(struct btrfs_disk_key));
3988         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
3989         num_csums = num_csums + num_csums_per_leaf - 1;
3990         num_csums = num_csums / num_csums_per_leaf;
3991
3992         old_csums = old_csums + num_csums_per_leaf - 1;
3993         old_csums = old_csums / num_csums_per_leaf;
3994
3995         /* No change, no need to reserve more */
3996         if (old_csums == num_csums)
3997                 return 0;
3998
3999         if (reserve)
4000                 return btrfs_calc_trans_metadata_size(root,
4001                                                       num_csums - old_csums);
4002
4003         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4004 }
4005
4006 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4007 {
4008         struct btrfs_root *root = BTRFS_I(inode)->root;
4009         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4010         u64 to_reserve = 0;
4011         unsigned nr_extents = 0;
4012         int flush = 1;
4013         int ret;
4014
4015         if (btrfs_is_free_space_inode(root, inode))
4016                 flush = 0;
4017
4018         if (flush && btrfs_transaction_in_commit(root->fs_info))
4019                 schedule_timeout(1);
4020
4021         num_bytes = ALIGN(num_bytes, root->sectorsize);
4022
4023         spin_lock(&BTRFS_I(inode)->lock);
4024         BTRFS_I(inode)->outstanding_extents++;
4025
4026         if (BTRFS_I(inode)->outstanding_extents >
4027             BTRFS_I(inode)->reserved_extents) {
4028                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4029                         BTRFS_I(inode)->reserved_extents;
4030                 BTRFS_I(inode)->reserved_extents += nr_extents;
4031
4032                 to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4033         }
4034         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4035         spin_unlock(&BTRFS_I(inode)->lock);
4036
4037         ret = reserve_metadata_bytes(NULL, root, block_rsv, to_reserve, flush);
4038         if (ret) {
4039                 u64 to_free = 0;
4040                 unsigned dropped;
4041
4042                 spin_lock(&BTRFS_I(inode)->lock);
4043                 dropped = drop_outstanding_extent(inode);
4044                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4045                 spin_unlock(&BTRFS_I(inode)->lock);
4046                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4047
4048                 /*
4049                  * Somebody could have come in and twiddled with the
4050                  * reservation, so if we have to free more than we would have
4051                  * reserved from this reservation go ahead and release those
4052                  * bytes.
4053                  */
4054                 to_free -= to_reserve;
4055                 if (to_free)
4056                         btrfs_block_rsv_release(root, block_rsv, to_free);
4057                 return ret;
4058         }
4059
4060         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4061
4062         return 0;
4063 }
4064
4065 /**
4066  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4067  * @inode: the inode to release the reservation for
4068  * @num_bytes: the number of bytes we're releasing
4069  *
4070  * This will release the metadata reservation for an inode.  This can be called
4071  * once we complete IO for a given set of bytes to release their metadata
4072  * reservations.
4073  */
4074 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4075 {
4076         struct btrfs_root *root = BTRFS_I(inode)->root;
4077         u64 to_free = 0;
4078         unsigned dropped;
4079
4080         num_bytes = ALIGN(num_bytes, root->sectorsize);
4081         spin_lock(&BTRFS_I(inode)->lock);
4082         dropped = drop_outstanding_extent(inode);
4083
4084         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4085         spin_unlock(&BTRFS_I(inode)->lock);
4086         if (dropped > 0)
4087                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4088
4089         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4090                                 to_free);
4091 }
4092
4093 /**
4094  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4095  * @inode: inode we're writing to
4096  * @num_bytes: the number of bytes we want to allocate
4097  *
4098  * This will do the following things
4099  *
4100  * o reserve space in the data space info for num_bytes
4101  * o reserve space in the metadata space info based on number of outstanding
4102  *   extents and how much csums will be needed
4103  * o add to the inodes ->delalloc_bytes
4104  * o add it to the fs_info's delalloc inodes list.
4105  *
4106  * This will return 0 for success and -ENOSPC if there is no space left.
4107  */
4108 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4109 {
4110         int ret;
4111
4112         ret = btrfs_check_data_free_space(inode, num_bytes);
4113         if (ret)
4114                 return ret;
4115
4116         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4117         if (ret) {
4118                 btrfs_free_reserved_data_space(inode, num_bytes);
4119                 return ret;
4120         }
4121
4122         return 0;
4123 }
4124
4125 /**
4126  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4127  * @inode: inode we're releasing space for
4128  * @num_bytes: the number of bytes we want to free up
4129  *
4130  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4131  * called in the case that we don't need the metadata AND data reservations
4132  * anymore.  So if there is an error or we insert an inline extent.
4133  *
4134  * This function will release the metadata space that was not used and will
4135  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4136  * list if there are no delalloc bytes left.
4137  */
4138 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4139 {
4140         btrfs_delalloc_release_metadata(inode, num_bytes);
4141         btrfs_free_reserved_data_space(inode, num_bytes);
4142 }
4143
4144 static int update_block_group(struct btrfs_trans_handle *trans,
4145                               struct btrfs_root *root,
4146                               u64 bytenr, u64 num_bytes, int alloc)
4147 {
4148         struct btrfs_block_group_cache *cache = NULL;
4149         struct btrfs_fs_info *info = root->fs_info;
4150         u64 total = num_bytes;
4151         u64 old_val;
4152         u64 byte_in_group;
4153         int factor;
4154
4155         /* block accounting for super block */
4156         spin_lock(&info->delalloc_lock);
4157         old_val = btrfs_super_bytes_used(&info->super_copy);
4158         if (alloc)
4159                 old_val += num_bytes;
4160         else
4161                 old_val -= num_bytes;
4162         btrfs_set_super_bytes_used(&info->super_copy, old_val);
4163         spin_unlock(&info->delalloc_lock);
4164
4165         while (total) {
4166                 cache = btrfs_lookup_block_group(info, bytenr);
4167                 if (!cache)
4168                         return -1;
4169                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4170                                     BTRFS_BLOCK_GROUP_RAID1 |
4171                                     BTRFS_BLOCK_GROUP_RAID10))
4172                         factor = 2;
4173                 else
4174                         factor = 1;
4175                 /*
4176                  * If this block group has free space cache written out, we
4177                  * need to make sure to load it if we are removing space.  This
4178                  * is because we need the unpinning stage to actually add the
4179                  * space back to the block group, otherwise we will leak space.
4180                  */
4181                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4182                         cache_block_group(cache, trans, NULL, 1);
4183
4184                 byte_in_group = bytenr - cache->key.objectid;
4185                 WARN_ON(byte_in_group > cache->key.offset);
4186
4187                 spin_lock(&cache->space_info->lock);
4188                 spin_lock(&cache->lock);
4189
4190                 if (btrfs_super_cache_generation(&info->super_copy) != 0 &&
4191                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4192                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4193
4194                 cache->dirty = 1;
4195                 old_val = btrfs_block_group_used(&cache->item);
4196                 num_bytes = min(total, cache->key.offset - byte_in_group);
4197                 if (alloc) {
4198                         old_val += num_bytes;
4199                         btrfs_set_block_group_used(&cache->item, old_val);
4200                         cache->reserved -= num_bytes;
4201                         cache->space_info->bytes_reserved -= num_bytes;
4202                         cache->space_info->bytes_used += num_bytes;
4203                         cache->space_info->disk_used += num_bytes * factor;
4204                         spin_unlock(&cache->lock);
4205                         spin_unlock(&cache->space_info->lock);
4206                 } else {
4207                         old_val -= num_bytes;
4208                         btrfs_set_block_group_used(&cache->item, old_val);
4209                         cache->pinned += num_bytes;
4210                         cache->space_info->bytes_pinned += num_bytes;
4211                         cache->space_info->bytes_used -= num_bytes;
4212                         cache->space_info->disk_used -= num_bytes * factor;
4213                         spin_unlock(&cache->lock);
4214                         spin_unlock(&cache->space_info->lock);
4215
4216                         set_extent_dirty(info->pinned_extents,
4217                                          bytenr, bytenr + num_bytes - 1,
4218                                          GFP_NOFS | __GFP_NOFAIL);
4219                 }
4220                 btrfs_put_block_group(cache);
4221                 total -= num_bytes;
4222                 bytenr += num_bytes;
4223         }
4224         return 0;
4225 }
4226
4227 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4228 {
4229         struct btrfs_block_group_cache *cache;
4230         u64 bytenr;
4231
4232         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4233         if (!cache)
4234                 return 0;
4235
4236         bytenr = cache->key.objectid;
4237         btrfs_put_block_group(cache);
4238
4239         return bytenr;
4240 }
4241
4242 static int pin_down_extent(struct btrfs_root *root,
4243                            struct btrfs_block_group_cache *cache,
4244                            u64 bytenr, u64 num_bytes, int reserved)
4245 {
4246         spin_lock(&cache->space_info->lock);
4247         spin_lock(&cache->lock);
4248         cache->pinned += num_bytes;
4249         cache->space_info->bytes_pinned += num_bytes;
4250         if (reserved) {
4251                 cache->reserved -= num_bytes;
4252                 cache->space_info->bytes_reserved -= num_bytes;
4253         }
4254         spin_unlock(&cache->lock);
4255         spin_unlock(&cache->space_info->lock);
4256
4257         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4258                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4259         return 0;
4260 }
4261
4262 /*
4263  * this function must be called within transaction
4264  */
4265 int btrfs_pin_extent(struct btrfs_root *root,
4266                      u64 bytenr, u64 num_bytes, int reserved)
4267 {
4268         struct btrfs_block_group_cache *cache;
4269
4270         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4271         BUG_ON(!cache);
4272
4273         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4274
4275         btrfs_put_block_group(cache);
4276         return 0;
4277 }
4278
4279 /**
4280  * btrfs_update_reserved_bytes - update the block_group and space info counters
4281  * @cache:      The cache we are manipulating
4282  * @num_bytes:  The number of bytes in question
4283  * @reserve:    One of the reservation enums
4284  *
4285  * This is called by the allocator when it reserves space, or by somebody who is
4286  * freeing space that was never actually used on disk.  For example if you
4287  * reserve some space for a new leaf in transaction A and before transaction A
4288  * commits you free that leaf, you call this with reserve set to 0 in order to
4289  * clear the reservation.
4290  *
4291  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4292  * ENOSPC accounting.  For data we handle the reservation through clearing the
4293  * delalloc bits in the io_tree.  We have to do this since we could end up
4294  * allocating less disk space for the amount of data we have reserved in the
4295  * case of compression.
4296  *
4297  * If this is a reservation and the block group has become read only we cannot
4298  * make the reservation and return -EAGAIN, otherwise this function always
4299  * succeeds.
4300  */
4301 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4302                                        u64 num_bytes, int reserve)
4303 {
4304         struct btrfs_space_info *space_info = cache->space_info;
4305         int ret = 0;
4306         spin_lock(&space_info->lock);
4307         spin_lock(&cache->lock);
4308         if (reserve != RESERVE_FREE) {
4309                 if (cache->ro) {
4310                         ret = -EAGAIN;
4311                 } else {
4312                         cache->reserved += num_bytes;
4313                         space_info->bytes_reserved += num_bytes;
4314                         if (reserve == RESERVE_ALLOC) {
4315                                 BUG_ON(space_info->bytes_may_use < num_bytes);
4316                                 space_info->bytes_may_use -= num_bytes;
4317                         }
4318                 }
4319         } else {
4320                 if (cache->ro)
4321                         space_info->bytes_readonly += num_bytes;
4322                 cache->reserved -= num_bytes;
4323                 space_info->bytes_reserved -= num_bytes;
4324                 space_info->reservation_progress++;
4325         }
4326         spin_unlock(&cache->lock);
4327         spin_unlock(&space_info->lock);
4328         return ret;
4329 }
4330
4331 int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4332                                 struct btrfs_root *root)
4333 {
4334         struct btrfs_fs_info *fs_info = root->fs_info;
4335         struct btrfs_caching_control *next;
4336         struct btrfs_caching_control *caching_ctl;
4337         struct btrfs_block_group_cache *cache;
4338
4339         down_write(&fs_info->extent_commit_sem);
4340
4341         list_for_each_entry_safe(caching_ctl, next,
4342                                  &fs_info->caching_block_groups, list) {
4343                 cache = caching_ctl->block_group;
4344                 if (block_group_cache_done(cache)) {
4345                         cache->last_byte_to_unpin = (u64)-1;
4346                         list_del_init(&caching_ctl->list);
4347                         put_caching_control(caching_ctl);
4348                 } else {
4349                         cache->last_byte_to_unpin = caching_ctl->progress;
4350                 }
4351         }
4352
4353         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4354                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4355         else
4356                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4357
4358         up_write(&fs_info->extent_commit_sem);
4359
4360         update_global_block_rsv(fs_info);
4361         return 0;
4362 }
4363
4364 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4365 {
4366         struct btrfs_fs_info *fs_info = root->fs_info;
4367         struct btrfs_block_group_cache *cache = NULL;
4368         u64 len;
4369
4370         while (start <= end) {
4371                 if (!cache ||
4372                     start >= cache->key.objectid + cache->key.offset) {
4373                         if (cache)
4374                                 btrfs_put_block_group(cache);
4375                         cache = btrfs_lookup_block_group(fs_info, start);
4376                         BUG_ON(!cache);
4377                 }
4378
4379                 len = cache->key.objectid + cache->key.offset - start;
4380                 len = min(len, end + 1 - start);
4381
4382                 if (start < cache->last_byte_to_unpin) {
4383                         len = min(len, cache->last_byte_to_unpin - start);
4384                         btrfs_add_free_space(cache, start, len);
4385                 }
4386
4387                 start += len;
4388
4389                 spin_lock(&cache->space_info->lock);
4390                 spin_lock(&cache->lock);
4391                 cache->pinned -= len;
4392                 cache->space_info->bytes_pinned -= len;
4393                 if (cache->ro)
4394                         cache->space_info->bytes_readonly += len;
4395                 spin_unlock(&cache->lock);
4396                 spin_unlock(&cache->space_info->lock);
4397         }
4398
4399         if (cache)
4400                 btrfs_put_block_group(cache);
4401         return 0;
4402 }
4403
4404 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4405                                struct btrfs_root *root)
4406 {
4407         struct btrfs_fs_info *fs_info = root->fs_info;
4408         struct extent_io_tree *unpin;
4409         u64 start;
4410         u64 end;
4411         int ret;
4412
4413         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4414                 unpin = &fs_info->freed_extents[1];
4415         else
4416                 unpin = &fs_info->freed_extents[0];
4417
4418         while (1) {
4419                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4420                                             EXTENT_DIRTY);
4421                 if (ret)
4422                         break;
4423
4424                 if (btrfs_test_opt(root, DISCARD))
4425                         ret = btrfs_discard_extent(root, start,
4426                                                    end + 1 - start, NULL);
4427
4428                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4429                 unpin_extent_range(root, start, end);
4430                 cond_resched();
4431         }
4432
4433         return 0;
4434 }
4435
4436 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4437                                 struct btrfs_root *root,
4438                                 u64 bytenr, u64 num_bytes, u64 parent,
4439                                 u64 root_objectid, u64 owner_objectid,
4440                                 u64 owner_offset, int refs_to_drop,
4441                                 struct btrfs_delayed_extent_op *extent_op)
4442 {
4443         struct btrfs_key key;
4444         struct btrfs_path *path;
4445         struct btrfs_fs_info *info = root->fs_info;
4446         struct btrfs_root *extent_root = info->extent_root;
4447         struct extent_buffer *leaf;
4448         struct btrfs_extent_item *ei;
4449         struct btrfs_extent_inline_ref *iref;
4450         int ret;
4451         int is_data;
4452         int extent_slot = 0;
4453         int found_extent = 0;
4454         int num_to_del = 1;
4455         u32 item_size;
4456         u64 refs;
4457
4458         path = btrfs_alloc_path();
4459         if (!path)
4460                 return -ENOMEM;
4461
4462         path->reada = 1;
4463         path->leave_spinning = 1;
4464
4465         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4466         BUG_ON(!is_data && refs_to_drop != 1);
4467
4468         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4469                                     bytenr, num_bytes, parent,
4470                                     root_objectid, owner_objectid,
4471                                     owner_offset);
4472         if (ret == 0) {
4473                 extent_slot = path->slots[0];
4474                 while (extent_slot >= 0) {
4475                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4476                                               extent_slot);
4477                         if (key.objectid != bytenr)
4478                                 break;
4479                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4480                             key.offset == num_bytes) {
4481                                 found_extent = 1;
4482                                 break;
4483                         }
4484                         if (path->slots[0] - extent_slot > 5)
4485                                 break;
4486                         extent_slot--;
4487                 }
4488 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4489                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4490                 if (found_extent && item_size < sizeof(*ei))
4491                         found_extent = 0;
4492 #endif
4493                 if (!found_extent) {
4494                         BUG_ON(iref);
4495                         ret = remove_extent_backref(trans, extent_root, path,
4496                                                     NULL, refs_to_drop,
4497                                                     is_data);
4498                         BUG_ON(ret);
4499                         btrfs_release_path(path);
4500                         path->leave_spinning = 1;
4501
4502                         key.objectid = bytenr;
4503                         key.type = BTRFS_EXTENT_ITEM_KEY;
4504                         key.offset = num_bytes;
4505
4506                         ret = btrfs_search_slot(trans, extent_root,
4507                                                 &key, path, -1, 1);
4508                         if (ret) {
4509                                 printk(KERN_ERR "umm, got %d back from search"
4510                                        ", was looking for %llu\n", ret,
4511                                        (unsigned long long)bytenr);
4512                                 if (ret > 0)
4513                                         btrfs_print_leaf(extent_root,
4514                                                          path->nodes[0]);
4515                         }
4516                         BUG_ON(ret);
4517                         extent_slot = path->slots[0];
4518                 }
4519         } else {
4520                 btrfs_print_leaf(extent_root, path->nodes[0]);
4521                 WARN_ON(1);
4522                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4523                        "parent %llu root %llu  owner %llu offset %llu\n",
4524                        (unsigned long long)bytenr,
4525                        (unsigned long long)parent,
4526                        (unsigned long long)root_objectid,
4527                        (unsigned long long)owner_objectid,
4528                        (unsigned long long)owner_offset);
4529         }
4530
4531         leaf = path->nodes[0];
4532         item_size = btrfs_item_size_nr(leaf, extent_slot);
4533 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4534         if (item_size < sizeof(*ei)) {
4535                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4536                 ret = convert_extent_item_v0(trans, extent_root, path,
4537                                              owner_objectid, 0);
4538                 BUG_ON(ret < 0);
4539
4540                 btrfs_release_path(path);
4541                 path->leave_spinning = 1;
4542
4543                 key.objectid = bytenr;
4544                 key.type = BTRFS_EXTENT_ITEM_KEY;
4545                 key.offset = num_bytes;
4546
4547                 ret = btrfs_search_slot(trans, extent_root, &key, path,
4548                                         -1, 1);
4549                 if (ret) {
4550                         printk(KERN_ERR "umm, got %d back from search"
4551                                ", was looking for %llu\n", ret,
4552                                (unsigned long long)bytenr);
4553                         btrfs_print_leaf(extent_root, path->nodes[0]);
4554                 }
4555                 BUG_ON(ret);
4556                 extent_slot = path->slots[0];
4557                 leaf = path->nodes[0];
4558                 item_size = btrfs_item_size_nr(leaf, extent_slot);
4559         }
4560 #endif
4561         BUG_ON(item_size < sizeof(*ei));
4562         ei = btrfs_item_ptr(leaf, extent_slot,
4563                             struct btrfs_extent_item);
4564         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
4565                 struct btrfs_tree_block_info *bi;
4566                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4567                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4568                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4569         }
4570
4571         refs = btrfs_extent_refs(leaf, ei);
4572         BUG_ON(refs < refs_to_drop);
4573         refs -= refs_to_drop;
4574
4575         if (refs > 0) {
4576                 if (extent_op)
4577                         __run_delayed_extent_op(extent_op, leaf, ei);
4578                 /*
4579                  * In the case of inline back ref, reference count will
4580                  * be updated by remove_extent_backref
4581                  */
4582                 if (iref) {
4583                         BUG_ON(!found_extent);
4584                 } else {
4585                         btrfs_set_extent_refs(leaf, ei, refs);
4586                         btrfs_mark_buffer_dirty(leaf);
4587                 }
4588                 if (found_extent) {
4589                         ret = remove_extent_backref(trans, extent_root, path,
4590                                                     iref, refs_to_drop,
4591                                                     is_data);
4592                         BUG_ON(ret);
4593                 }
4594         } else {
4595                 if (found_extent) {
4596                         BUG_ON(is_data && refs_to_drop !=
4597                                extent_data_ref_count(root, path, iref));
4598                         if (iref) {
4599                                 BUG_ON(path->slots[0] != extent_slot);
4600                         } else {
4601                                 BUG_ON(path->slots[0] != extent_slot + 1);
4602                                 path->slots[0] = extent_slot;
4603                                 num_to_del = 2;
4604                         }
4605                 }
4606
4607                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
4608                                       num_to_del);
4609                 BUG_ON(ret);
4610                 btrfs_release_path(path);
4611
4612                 if (is_data) {
4613                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
4614                         BUG_ON(ret);
4615                 } else {
4616                         invalidate_mapping_pages(info->btree_inode->i_mapping,
4617                              bytenr >> PAGE_CACHE_SHIFT,
4618                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
4619                 }
4620
4621                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
4622                 BUG_ON(ret);
4623         }
4624         btrfs_free_path(path);
4625         return ret;
4626 }
4627
4628 /*
4629  * when we free an block, it is possible (and likely) that we free the last
4630  * delayed ref for that extent as well.  This searches the delayed ref tree for
4631  * a given extent, and if there are no other delayed refs to be processed, it
4632  * removes it from the tree.
4633  */
4634 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
4635                                       struct btrfs_root *root, u64 bytenr)
4636 {
4637         struct btrfs_delayed_ref_head *head;
4638         struct btrfs_delayed_ref_root *delayed_refs;
4639         struct btrfs_delayed_ref_node *ref;
4640         struct rb_node *node;
4641         int ret = 0;
4642
4643         delayed_refs = &trans->transaction->delayed_refs;
4644         spin_lock(&delayed_refs->lock);
4645         head = btrfs_find_delayed_ref_head(trans, bytenr);
4646         if (!head)
4647                 goto out;
4648
4649         node = rb_prev(&head->node.rb_node);
4650         if (!node)
4651                 goto out;
4652
4653         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
4654
4655         /* there are still entries for this ref, we can't drop it */
4656         if (ref->bytenr == bytenr)
4657                 goto out;
4658
4659         if (head->extent_op) {
4660                 if (!head->must_insert_reserved)
4661                         goto out;
4662                 kfree(head->extent_op);
4663                 head->extent_op = NULL;
4664         }
4665
4666         /*
4667          * waiting for the lock here would deadlock.  If someone else has it
4668          * locked they are already in the process of dropping it anyway
4669          */
4670         if (!mutex_trylock(&head->mutex))
4671                 goto out;
4672
4673         /*
4674          * at this point we have a head with no other entries.  Go
4675          * ahead and process it.
4676          */
4677         head->node.in_tree = 0;
4678         rb_erase(&head->node.rb_node, &delayed_refs->root);
4679
4680         delayed_refs->num_entries--;
4681
4682         /*
4683          * we don't take a ref on the node because we're removing it from the
4684          * tree, so we just steal the ref the tree was holding.
4685          */
4686         delayed_refs->num_heads--;
4687         if (list_empty(&head->cluster))
4688                 delayed_refs->num_heads_ready--;
4689
4690         list_del_init(&head->cluster);
4691         spin_unlock(&delayed_refs->lock);
4692
4693         BUG_ON(head->extent_op);
4694         if (head->must_insert_reserved)
4695                 ret = 1;
4696
4697         mutex_unlock(&head->mutex);
4698         btrfs_put_delayed_ref(&head->node);
4699         return ret;
4700 out:
4701         spin_unlock(&delayed_refs->lock);
4702         return 0;
4703 }
4704
4705 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
4706                            struct btrfs_root *root,
4707                            struct extent_buffer *buf,
4708                            u64 parent, int last_ref)
4709 {
4710         struct btrfs_block_rsv *block_rsv;
4711         struct btrfs_block_group_cache *cache = NULL;
4712         int ret;
4713
4714         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4715                 ret = btrfs_add_delayed_tree_ref(trans, buf->start, buf->len,
4716                                                 parent, root->root_key.objectid,
4717                                                 btrfs_header_level(buf),
4718                                                 BTRFS_DROP_DELAYED_REF, NULL);
4719                 BUG_ON(ret);
4720         }
4721
4722         if (!last_ref)
4723                 return;
4724
4725         block_rsv = get_block_rsv(trans, root);
4726         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
4727         if (block_rsv->space_info != cache->space_info)
4728                 goto out;
4729
4730         if (btrfs_header_generation(buf) == trans->transid) {
4731                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4732                         ret = check_ref_cleanup(trans, root, buf->start);
4733                         if (!ret)
4734                                 goto out;
4735                 }
4736
4737                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
4738                         pin_down_extent(root, cache, buf->start, buf->len, 1);
4739                         goto out;
4740                 }
4741
4742                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
4743
4744                 btrfs_add_free_space(cache, buf->start, buf->len);
4745                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
4746         }
4747 out:
4748         /*
4749          * Deleting the buffer, clear the corrupt flag since it doesn't matter
4750          * anymore.
4751          */
4752         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
4753         btrfs_put_block_group(cache);
4754 }
4755
4756 int btrfs_free_extent(struct btrfs_trans_handle *trans,
4757                       struct btrfs_root *root,
4758                       u64 bytenr, u64 num_bytes, u64 parent,
4759                       u64 root_objectid, u64 owner, u64 offset)
4760 {
4761         int ret;
4762
4763         /*
4764          * tree log blocks never actually go into the extent allocation
4765          * tree, just update pinning info and exit early.
4766          */
4767         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
4768                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
4769                 /* unlocks the pinned mutex */
4770                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
4771                 ret = 0;
4772         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
4773                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
4774                                         parent, root_objectid, (int)owner,
4775                                         BTRFS_DROP_DELAYED_REF, NULL);
4776                 BUG_ON(ret);
4777         } else {
4778                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
4779                                         parent, root_objectid, owner,
4780                                         offset, BTRFS_DROP_DELAYED_REF, NULL);
4781                 BUG_ON(ret);
4782         }
4783         return ret;
4784 }
4785
4786 static u64 stripe_align(struct btrfs_root *root, u64 val)
4787 {
4788         u64 mask = ((u64)root->stripesize - 1);
4789         u64 ret = (val + mask) & ~mask;
4790         return ret;
4791 }
4792
4793 /*
4794  * when we wait for progress in the block group caching, its because
4795  * our allocation attempt failed at least once.  So, we must sleep
4796  * and let some progress happen before we try again.
4797  *
4798  * This function will sleep at least once waiting for new free space to
4799  * show up, and then it will check the block group free space numbers
4800  * for our min num_bytes.  Another option is to have it go ahead
4801  * and look in the rbtree for a free extent of a given size, but this
4802  * is a good start.
4803  */
4804 static noinline int
4805 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
4806                                 u64 num_bytes)
4807 {
4808         struct btrfs_caching_control *caching_ctl;
4809         DEFINE_WAIT(wait);
4810
4811         caching_ctl = get_caching_control(cache);
4812         if (!caching_ctl)
4813                 return 0;
4814
4815         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
4816                    (cache->free_space_ctl->free_space >= num_bytes));
4817
4818         put_caching_control(caching_ctl);
4819         return 0;
4820 }
4821
4822 static noinline int
4823 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
4824 {
4825         struct btrfs_caching_control *caching_ctl;
4826         DEFINE_WAIT(wait);
4827
4828         caching_ctl = get_caching_control(cache);
4829         if (!caching_ctl)
4830                 return 0;
4831
4832         wait_event(caching_ctl->wait, block_group_cache_done(cache));
4833
4834         put_caching_control(caching_ctl);
4835         return 0;
4836 }
4837
4838 static int get_block_group_index(struct btrfs_block_group_cache *cache)
4839 {
4840         int index;
4841         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
4842                 index = 0;
4843         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
4844                 index = 1;
4845         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
4846                 index = 2;
4847         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
4848                 index = 3;
4849         else
4850                 index = 4;
4851         return index;
4852 }
4853
4854 enum btrfs_loop_type {
4855         LOOP_FIND_IDEAL = 0,
4856         LOOP_CACHING_NOWAIT = 1,
4857         LOOP_CACHING_WAIT = 2,
4858         LOOP_ALLOC_CHUNK = 3,
4859         LOOP_NO_EMPTY_SIZE = 4,
4860 };
4861
4862 /*
4863  * walks the btree of allocated extents and find a hole of a given size.
4864  * The key ins is changed to record the hole:
4865  * ins->objectid == block start
4866  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4867  * ins->offset == number of blocks
4868  * Any available blocks before search_start are skipped.
4869  */
4870 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
4871                                      struct btrfs_root *orig_root,
4872                                      u64 num_bytes, u64 empty_size,
4873                                      u64 search_start, u64 search_end,
4874                                      u64 hint_byte, struct btrfs_key *ins,
4875                                      u64 data)
4876 {
4877         int ret = 0;
4878         struct btrfs_root *root = orig_root->fs_info->extent_root;
4879         struct btrfs_free_cluster *last_ptr = NULL;
4880         struct btrfs_block_group_cache *block_group = NULL;
4881         int empty_cluster = 2 * 1024 * 1024;
4882         int allowed_chunk_alloc = 0;
4883         int done_chunk_alloc = 0;
4884         struct btrfs_space_info *space_info;
4885         int last_ptr_loop = 0;
4886         int loop = 0;
4887         int index = 0;
4888         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
4889                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
4890         bool found_uncached_bg = false;
4891         bool failed_cluster_refill = false;
4892         bool failed_alloc = false;
4893         bool use_cluster = true;
4894         u64 ideal_cache_percent = 0;
4895         u64 ideal_cache_offset = 0;
4896
4897         WARN_ON(num_bytes < root->sectorsize);
4898         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
4899         ins->objectid = 0;
4900         ins->offset = 0;
4901
4902         space_info = __find_space_info(root->fs_info, data);
4903         if (!space_info) {
4904                 printk(KERN_ERR "No space info for %llu\n", data);
4905                 return -ENOSPC;
4906         }
4907
4908         /*
4909          * If the space info is for both data and metadata it means we have a
4910          * small filesystem and we can't use the clustering stuff.
4911          */
4912         if (btrfs_mixed_space_info(space_info))
4913                 use_cluster = false;
4914
4915         if (orig_root->ref_cows || empty_size)
4916                 allowed_chunk_alloc = 1;
4917
4918         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
4919                 last_ptr = &root->fs_info->meta_alloc_cluster;
4920                 if (!btrfs_test_opt(root, SSD))
4921                         empty_cluster = 64 * 1024;
4922         }
4923
4924         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
4925             btrfs_test_opt(root, SSD)) {
4926                 last_ptr = &root->fs_info->data_alloc_cluster;
4927         }
4928
4929         if (last_ptr) {
4930                 spin_lock(&last_ptr->lock);
4931                 if (last_ptr->block_group)
4932                         hint_byte = last_ptr->window_start;
4933                 spin_unlock(&last_ptr->lock);
4934         }
4935
4936         search_start = max(search_start, first_logical_byte(root, 0));
4937         search_start = max(search_start, hint_byte);
4938
4939         if (!last_ptr)
4940                 empty_cluster = 0;
4941
4942         if (search_start == hint_byte) {
4943 ideal_cache:
4944                 block_group = btrfs_lookup_block_group(root->fs_info,
4945                                                        search_start);
4946                 /*
4947                  * we don't want to use the block group if it doesn't match our
4948                  * allocation bits, or if its not cached.
4949                  *
4950                  * However if we are re-searching with an ideal block group
4951                  * picked out then we don't care that the block group is cached.
4952                  */
4953                 if (block_group && block_group_bits(block_group, data) &&
4954                     (block_group->cached != BTRFS_CACHE_NO ||
4955                      search_start == ideal_cache_offset)) {
4956                         down_read(&space_info->groups_sem);
4957                         if (list_empty(&block_group->list) ||
4958                             block_group->ro) {
4959                                 /*
4960                                  * someone is removing this block group,
4961                                  * we can't jump into the have_block_group
4962                                  * target because our list pointers are not
4963                                  * valid
4964                                  */
4965                                 btrfs_put_block_group(block_group);
4966                                 up_read(&space_info->groups_sem);
4967                         } else {
4968                                 index = get_block_group_index(block_group);
4969                                 goto have_block_group;
4970                         }
4971                 } else if (block_group) {
4972                         btrfs_put_block_group(block_group);
4973                 }
4974         }
4975 search:
4976         down_read(&space_info->groups_sem);
4977         list_for_each_entry(block_group, &space_info->block_groups[index],
4978                             list) {
4979                 u64 offset;
4980                 int cached;
4981
4982                 btrfs_get_block_group(block_group);
4983                 search_start = block_group->key.objectid;
4984
4985                 /*
4986                  * this can happen if we end up cycling through all the
4987                  * raid types, but we want to make sure we only allocate
4988                  * for the proper type.
4989                  */
4990                 if (!block_group_bits(block_group, data)) {
4991                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
4992                                 BTRFS_BLOCK_GROUP_RAID1 |
4993                                 BTRFS_BLOCK_GROUP_RAID10;
4994
4995                         /*
4996                          * if they asked for extra copies and this block group
4997                          * doesn't provide them, bail.  This does allow us to
4998                          * fill raid0 from raid1.
4999                          */
5000                         if ((data & extra) && !(block_group->flags & extra))
5001                                 goto loop;
5002                 }
5003
5004 have_block_group:
5005                 if (unlikely(block_group->cached == BTRFS_CACHE_NO)) {
5006                         u64 free_percent;
5007
5008                         ret = cache_block_group(block_group, trans,
5009                                                 orig_root, 1);
5010                         if (block_group->cached == BTRFS_CACHE_FINISHED)
5011                                 goto have_block_group;
5012
5013                         free_percent = btrfs_block_group_used(&block_group->item);
5014                         free_percent *= 100;
5015                         free_percent = div64_u64(free_percent,
5016                                                  block_group->key.offset);
5017                         free_percent = 100 - free_percent;
5018                         if (free_percent > ideal_cache_percent &&
5019                             likely(!block_group->ro)) {
5020                                 ideal_cache_offset = block_group->key.objectid;
5021                                 ideal_cache_percent = free_percent;
5022                         }
5023
5024                         /*
5025                          * The caching workers are limited to 2 threads, so we
5026                          * can queue as much work as we care to.
5027                          */
5028                         if (loop > LOOP_FIND_IDEAL) {
5029                                 ret = cache_block_group(block_group, trans,
5030                                                         orig_root, 0);
5031                                 BUG_ON(ret);
5032                         }
5033                         found_uncached_bg = true;
5034
5035                         /*
5036                          * If loop is set for cached only, try the next block
5037                          * group.
5038                          */
5039                         if (loop == LOOP_FIND_IDEAL)
5040                                 goto loop;
5041                 }
5042
5043                 cached = block_group_cache_done(block_group);
5044                 if (unlikely(!cached))
5045                         found_uncached_bg = true;
5046
5047                 if (unlikely(block_group->ro))
5048                         goto loop;
5049
5050                 spin_lock(&block_group->free_space_ctl->tree_lock);
5051                 if (cached &&
5052                     block_group->free_space_ctl->free_space <
5053                     num_bytes + empty_size) {
5054                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5055                         goto loop;
5056                 }
5057                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5058
5059                 /*
5060                  * Ok we want to try and use the cluster allocator, so lets look
5061                  * there, unless we are on LOOP_NO_EMPTY_SIZE, since we will
5062                  * have tried the cluster allocator plenty of times at this
5063                  * point and not have found anything, so we are likely way too
5064                  * fragmented for the clustering stuff to find anything, so lets
5065                  * just skip it and let the allocator find whatever block it can
5066                  * find
5067                  */
5068                 if (last_ptr && loop < LOOP_NO_EMPTY_SIZE) {
5069                         /*
5070                          * the refill lock keeps out other
5071                          * people trying to start a new cluster
5072                          */
5073                         spin_lock(&last_ptr->refill_lock);
5074                         if (last_ptr->block_group &&
5075                             (last_ptr->block_group->ro ||
5076                             !block_group_bits(last_ptr->block_group, data))) {
5077                                 offset = 0;
5078                                 goto refill_cluster;
5079                         }
5080
5081                         offset = btrfs_alloc_from_cluster(block_group, last_ptr,
5082                                                  num_bytes, search_start);
5083                         if (offset) {
5084                                 /* we have a block, we're done */
5085                                 spin_unlock(&last_ptr->refill_lock);
5086                                 goto checks;
5087                         }
5088
5089                         spin_lock(&last_ptr->lock);
5090                         /*
5091                          * whoops, this cluster doesn't actually point to
5092                          * this block group.  Get a ref on the block
5093                          * group is does point to and try again
5094                          */
5095                         if (!last_ptr_loop && last_ptr->block_group &&
5096                             last_ptr->block_group != block_group &&
5097                             index <=
5098                                  get_block_group_index(last_ptr->block_group)) {
5099
5100                                 btrfs_put_block_group(block_group);
5101                                 block_group = last_ptr->block_group;
5102                                 btrfs_get_block_group(block_group);
5103                                 spin_unlock(&last_ptr->lock);
5104                                 spin_unlock(&last_ptr->refill_lock);
5105
5106                                 last_ptr_loop = 1;
5107                                 search_start = block_group->key.objectid;
5108                                 /*
5109                                  * we know this block group is properly
5110                                  * in the list because
5111                                  * btrfs_remove_block_group, drops the
5112                                  * cluster before it removes the block
5113                                  * group from the list
5114                                  */
5115                                 goto have_block_group;
5116                         }
5117                         spin_unlock(&last_ptr->lock);
5118 refill_cluster:
5119                         /*
5120                          * this cluster didn't work out, free it and
5121                          * start over
5122                          */
5123                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5124
5125                         last_ptr_loop = 0;
5126
5127                         /* allocate a cluster in this block group */
5128                         ret = btrfs_find_space_cluster(trans, root,
5129                                                block_group, last_ptr,
5130                                                offset, num_bytes,
5131                                                empty_cluster + empty_size);
5132                         if (ret == 0) {
5133                                 /*
5134                                  * now pull our allocation out of this
5135                                  * cluster
5136                                  */
5137                                 offset = btrfs_alloc_from_cluster(block_group,
5138                                                   last_ptr, num_bytes,
5139                                                   search_start);
5140                                 if (offset) {
5141                                         /* we found one, proceed */
5142                                         spin_unlock(&last_ptr->refill_lock);
5143                                         goto checks;
5144                                 }
5145                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5146                                    && !failed_cluster_refill) {
5147                                 spin_unlock(&last_ptr->refill_lock);
5148
5149                                 failed_cluster_refill = true;
5150                                 wait_block_group_cache_progress(block_group,
5151                                        num_bytes + empty_cluster + empty_size);
5152                                 goto have_block_group;
5153                         }
5154
5155                         /*
5156                          * at this point we either didn't find a cluster
5157                          * or we weren't able to allocate a block from our
5158                          * cluster.  Free the cluster we've been trying
5159                          * to use, and go to the next block group
5160                          */
5161                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5162                         spin_unlock(&last_ptr->refill_lock);
5163                         goto loop;
5164                 }
5165
5166                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5167                                                     num_bytes, empty_size);
5168                 /*
5169                  * If we didn't find a chunk, and we haven't failed on this
5170                  * block group before, and this block group is in the middle of
5171                  * caching and we are ok with waiting, then go ahead and wait
5172                  * for progress to be made, and set failed_alloc to true.
5173                  *
5174                  * If failed_alloc is true then we've already waited on this
5175                  * block group once and should move on to the next block group.
5176                  */
5177                 if (!offset && !failed_alloc && !cached &&
5178                     loop > LOOP_CACHING_NOWAIT) {
5179                         wait_block_group_cache_progress(block_group,
5180                                                 num_bytes + empty_size);
5181                         failed_alloc = true;
5182                         goto have_block_group;
5183                 } else if (!offset) {
5184                         goto loop;
5185                 }
5186 checks:
5187                 search_start = stripe_align(root, offset);
5188                 /* move on to the next group */
5189                 if (search_start + num_bytes >= search_end) {
5190                         btrfs_add_free_space(block_group, offset, num_bytes);
5191                         goto loop;
5192                 }
5193
5194                 /* move on to the next group */
5195                 if (search_start + num_bytes >
5196                     block_group->key.objectid + block_group->key.offset) {
5197                         btrfs_add_free_space(block_group, offset, num_bytes);
5198                         goto loop;
5199                 }
5200
5201                 ins->objectid = search_start;
5202                 ins->offset = num_bytes;
5203
5204                 if (offset < search_start)
5205                         btrfs_add_free_space(block_group, offset,
5206                                              search_start - offset);
5207                 BUG_ON(offset > search_start);
5208
5209                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
5210                                                   alloc_type);
5211                 if (ret == -EAGAIN) {
5212                         btrfs_add_free_space(block_group, offset, num_bytes);
5213                         goto loop;
5214                 }
5215
5216                 /* we are all good, lets return */
5217                 ins->objectid = search_start;
5218                 ins->offset = num_bytes;
5219
5220                 if (offset < search_start)
5221                         btrfs_add_free_space(block_group, offset,
5222                                              search_start - offset);
5223                 BUG_ON(offset > search_start);
5224                 btrfs_put_block_group(block_group);
5225                 break;
5226 loop:
5227                 failed_cluster_refill = false;
5228                 failed_alloc = false;
5229                 BUG_ON(index != get_block_group_index(block_group));
5230                 btrfs_put_block_group(block_group);
5231         }
5232         up_read(&space_info->groups_sem);
5233
5234         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5235                 goto search;
5236
5237         /* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
5238          *                      for them to make caching progress.  Also
5239          *                      determine the best possible bg to cache
5240          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5241          *                      caching kthreads as we move along
5242          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5243          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5244          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5245          *                      again
5246          */
5247         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5248                 index = 0;
5249                 if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
5250                         found_uncached_bg = false;
5251                         loop++;
5252                         if (!ideal_cache_percent)
5253                                 goto search;
5254
5255                         /*
5256                          * 1 of the following 2 things have happened so far
5257                          *
5258                          * 1) We found an ideal block group for caching that
5259                          * is mostly full and will cache quickly, so we might
5260                          * as well wait for it.
5261                          *
5262                          * 2) We searched for cached only and we didn't find
5263                          * anything, and we didn't start any caching kthreads
5264                          * either, so chances are we will loop through and
5265                          * start a couple caching kthreads, and then come back
5266                          * around and just wait for them.  This will be slower
5267                          * because we will have 2 caching kthreads reading at
5268                          * the same time when we could have just started one
5269                          * and waited for it to get far enough to give us an
5270                          * allocation, so go ahead and go to the wait caching
5271                          * loop.
5272                          */
5273                         loop = LOOP_CACHING_WAIT;
5274                         search_start = ideal_cache_offset;
5275                         ideal_cache_percent = 0;
5276                         goto ideal_cache;
5277                 } else if (loop == LOOP_FIND_IDEAL) {
5278                         /*
5279                          * Didn't find a uncached bg, wait on anything we find
5280                          * next.
5281                          */
5282                         loop = LOOP_CACHING_WAIT;
5283                         goto search;
5284                 }
5285
5286                 loop++;
5287
5288                 if (loop == LOOP_ALLOC_CHUNK) {
5289                        if (allowed_chunk_alloc) {
5290                                 ret = do_chunk_alloc(trans, root, num_bytes +
5291                                                      2 * 1024 * 1024, data,
5292                                                      CHUNK_ALLOC_LIMITED);
5293                                 allowed_chunk_alloc = 0;
5294                                 if (ret == 1)
5295                                         done_chunk_alloc = 1;
5296                         } else if (!done_chunk_alloc &&
5297                                    space_info->force_alloc ==
5298                                    CHUNK_ALLOC_NO_FORCE) {
5299                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5300                         }
5301
5302                        /*
5303                         * We didn't allocate a chunk, go ahead and drop the
5304                         * empty size and loop again.
5305                         */
5306                        if (!done_chunk_alloc)
5307                                loop = LOOP_NO_EMPTY_SIZE;
5308                 }
5309
5310                 if (loop == LOOP_NO_EMPTY_SIZE) {
5311                         empty_size = 0;
5312                         empty_cluster = 0;
5313                 }
5314
5315                 goto search;
5316         } else if (!ins->objectid) {
5317                 ret = -ENOSPC;
5318         } else if (ins->objectid) {
5319                 ret = 0;
5320         }
5321
5322         return ret;
5323 }
5324
5325 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5326                             int dump_block_groups)
5327 {
5328         struct btrfs_block_group_cache *cache;
5329         int index = 0;
5330
5331         spin_lock(&info->lock);
5332         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5333                (unsigned long long)info->flags,
5334                (unsigned long long)(info->total_bytes - info->bytes_used -
5335                                     info->bytes_pinned - info->bytes_reserved -
5336                                     info->bytes_readonly),
5337                (info->full) ? "" : "not ");
5338         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5339                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5340                (unsigned long long)info->total_bytes,
5341                (unsigned long long)info->bytes_used,
5342                (unsigned long long)info->bytes_pinned,
5343                (unsigned long long)info->bytes_reserved,
5344                (unsigned long long)info->bytes_may_use,
5345                (unsigned long long)info->bytes_readonly);
5346         spin_unlock(&info->lock);
5347
5348         if (!dump_block_groups)
5349                 return;
5350
5351         down_read(&info->groups_sem);
5352 again:
5353         list_for_each_entry(cache, &info->block_groups[index], list) {
5354                 spin_lock(&cache->lock);
5355                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5356                        "%llu pinned %llu reserved\n",
5357                        (unsigned long long)cache->key.objectid,
5358                        (unsigned long long)cache->key.offset,
5359                        (unsigned long long)btrfs_block_group_used(&cache->item),
5360                        (unsigned long long)cache->pinned,
5361                        (unsigned long long)cache->reserved);
5362                 btrfs_dump_free_space(cache, bytes);
5363                 spin_unlock(&cache->lock);
5364         }
5365         if (++index < BTRFS_NR_RAID_TYPES)
5366                 goto again;
5367         up_read(&info->groups_sem);
5368 }
5369
5370 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5371                          struct btrfs_root *root,
5372                          u64 num_bytes, u64 min_alloc_size,
5373                          u64 empty_size, u64 hint_byte,
5374                          u64 search_end, struct btrfs_key *ins,
5375                          u64 data)
5376 {
5377         int ret;
5378         u64 search_start = 0;
5379
5380         data = btrfs_get_alloc_profile(root, data);
5381 again:
5382         /*
5383          * the only place that sets empty_size is btrfs_realloc_node, which
5384          * is not called recursively on allocations
5385          */
5386         if (empty_size || root->ref_cows)
5387                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5388                                      num_bytes + 2 * 1024 * 1024, data,
5389                                      CHUNK_ALLOC_NO_FORCE);
5390
5391         WARN_ON(num_bytes < root->sectorsize);
5392         ret = find_free_extent(trans, root, num_bytes, empty_size,
5393                                search_start, search_end, hint_byte,
5394                                ins, data);
5395
5396         if (ret == -ENOSPC && num_bytes > min_alloc_size) {
5397                 num_bytes = num_bytes >> 1;
5398                 num_bytes = num_bytes & ~(root->sectorsize - 1);
5399                 num_bytes = max(num_bytes, min_alloc_size);
5400                 do_chunk_alloc(trans, root->fs_info->extent_root,
5401                                num_bytes, data, CHUNK_ALLOC_FORCE);
5402                 goto again;
5403         }
5404         if (ret == -ENOSPC && btrfs_test_opt(root, ENOSPC_DEBUG)) {
5405                 struct btrfs_space_info *sinfo;
5406
5407                 sinfo = __find_space_info(root->fs_info, data);
5408                 printk(KERN_ERR "btrfs allocation failed flags %llu, "
5409                        "wanted %llu\n", (unsigned long long)data,
5410                        (unsigned long long)num_bytes);
5411                 dump_space_info(sinfo, num_bytes, 1);
5412         }
5413
5414         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
5415
5416         return ret;
5417 }
5418
5419 int btrfs_free_reserved_extent(struct btrfs_root *root, u64 start, u64 len)
5420 {
5421         struct btrfs_block_group_cache *cache;
5422         int ret = 0;
5423
5424         cache = btrfs_lookup_block_group(root->fs_info, start);
5425         if (!cache) {
5426                 printk(KERN_ERR "Unable to find block group for %llu\n",
5427                        (unsigned long long)start);
5428                 return -ENOSPC;
5429         }
5430
5431         if (btrfs_test_opt(root, DISCARD))
5432                 ret = btrfs_discard_extent(root, start, len, NULL);
5433
5434         btrfs_add_free_space(cache, start, len);
5435         btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
5436         btrfs_put_block_group(cache);
5437
5438         trace_btrfs_reserved_extent_free(root, start, len);
5439
5440         return ret;
5441 }
5442
5443 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5444                                       struct btrfs_root *root,
5445                                       u64 parent, u64 root_objectid,
5446                                       u64 flags, u64 owner, u64 offset,
5447                                       struct btrfs_key *ins, int ref_mod)
5448 {
5449         int ret;
5450         struct btrfs_fs_info *fs_info = root->fs_info;
5451         struct btrfs_extent_item *extent_item;
5452         struct btrfs_extent_inline_ref *iref;
5453         struct btrfs_path *path;
5454         struct extent_buffer *leaf;
5455         int type;
5456         u32 size;
5457
5458         if (parent > 0)
5459                 type = BTRFS_SHARED_DATA_REF_KEY;
5460         else
5461                 type = BTRFS_EXTENT_DATA_REF_KEY;
5462
5463         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5464
5465         path = btrfs_alloc_path();
5466         if (!path)
5467                 return -ENOMEM;
5468
5469         path->leave_spinning = 1;
5470         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5471                                       ins, size);
5472         BUG_ON(ret);
5473
5474         leaf = path->nodes[0];
5475         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5476                                      struct btrfs_extent_item);
5477         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
5478         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5479         btrfs_set_extent_flags(leaf, extent_item,
5480                                flags | BTRFS_EXTENT_FLAG_DATA);
5481
5482         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
5483         btrfs_set_extent_inline_ref_type(leaf, iref, type);
5484         if (parent > 0) {
5485                 struct btrfs_shared_data_ref *ref;
5486                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
5487                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5488                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
5489         } else {
5490                 struct btrfs_extent_data_ref *ref;
5491                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
5492                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
5493                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
5494                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
5495                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
5496         }
5497
5498         btrfs_mark_buffer_dirty(path->nodes[0]);
5499         btrfs_free_path(path);
5500
5501         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5502         if (ret) {
5503                 printk(KERN_ERR "btrfs update block group failed for %llu "
5504                        "%llu\n", (unsigned long long)ins->objectid,
5505                        (unsigned long long)ins->offset);
5506                 BUG();
5507         }
5508         return ret;
5509 }
5510
5511 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
5512                                      struct btrfs_root *root,
5513                                      u64 parent, u64 root_objectid,
5514                                      u64 flags, struct btrfs_disk_key *key,
5515                                      int level, struct btrfs_key *ins)
5516 {
5517         int ret;
5518         struct btrfs_fs_info *fs_info = root->fs_info;
5519         struct btrfs_extent_item *extent_item;
5520         struct btrfs_tree_block_info *block_info;
5521         struct btrfs_extent_inline_ref *iref;
5522         struct btrfs_path *path;
5523         struct extent_buffer *leaf;
5524         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
5525
5526         path = btrfs_alloc_path();
5527         if (!path)
5528                 return -ENOMEM;
5529
5530         path->leave_spinning = 1;
5531         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5532                                       ins, size);
5533         BUG_ON(ret);
5534
5535         leaf = path->nodes[0];
5536         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5537                                      struct btrfs_extent_item);
5538         btrfs_set_extent_refs(leaf, extent_item, 1);
5539         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5540         btrfs_set_extent_flags(leaf, extent_item,
5541                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
5542         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
5543
5544         btrfs_set_tree_block_key(leaf, block_info, key);
5545         btrfs_set_tree_block_level(leaf, block_info, level);
5546
5547         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
5548         if (parent > 0) {
5549                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
5550                 btrfs_set_extent_inline_ref_type(leaf, iref,
5551                                                  BTRFS_SHARED_BLOCK_REF_KEY);
5552                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5553         } else {
5554                 btrfs_set_extent_inline_ref_type(leaf, iref,
5555                                                  BTRFS_TREE_BLOCK_REF_KEY);
5556                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
5557         }
5558
5559         btrfs_mark_buffer_dirty(leaf);
5560         btrfs_free_path(path);
5561
5562         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5563         if (ret) {
5564                 printk(KERN_ERR "btrfs update block group failed for %llu "
5565                        "%llu\n", (unsigned long long)ins->objectid,
5566                        (unsigned long long)ins->offset);
5567                 BUG();
5568         }
5569         return ret;
5570 }
5571
5572 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5573                                      struct btrfs_root *root,
5574                                      u64 root_objectid, u64 owner,
5575                                      u64 offset, struct btrfs_key *ins)
5576 {
5577         int ret;
5578
5579         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
5580
5581         ret = btrfs_add_delayed_data_ref(trans, ins->objectid, ins->offset,
5582                                          0, root_objectid, owner, offset,
5583                                          BTRFS_ADD_DELAYED_EXTENT, NULL);
5584         return ret;
5585 }
5586
5587 /*
5588  * this is used by the tree logging recovery code.  It records that
5589  * an extent has been allocated and makes sure to clear the free
5590  * space cache bits as well
5591  */
5592 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
5593                                    struct btrfs_root *root,
5594                                    u64 root_objectid, u64 owner, u64 offset,
5595                                    struct btrfs_key *ins)
5596 {
5597         int ret;
5598         struct btrfs_block_group_cache *block_group;
5599         struct btrfs_caching_control *caching_ctl;
5600         u64 start = ins->objectid;
5601         u64 num_bytes = ins->offset;
5602
5603         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
5604         cache_block_group(block_group, trans, NULL, 0);
5605         caching_ctl = get_caching_control(block_group);
5606
5607         if (!caching_ctl) {
5608                 BUG_ON(!block_group_cache_done(block_group));
5609                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5610                 BUG_ON(ret);
5611         } else {
5612                 mutex_lock(&caching_ctl->mutex);
5613
5614                 if (start >= caching_ctl->progress) {
5615                         ret = add_excluded_extent(root, start, num_bytes);
5616                         BUG_ON(ret);
5617                 } else if (start + num_bytes <= caching_ctl->progress) {
5618                         ret = btrfs_remove_free_space(block_group,
5619                                                       start, num_bytes);
5620                         BUG_ON(ret);
5621                 } else {
5622                         num_bytes = caching_ctl->progress - start;
5623                         ret = btrfs_remove_free_space(block_group,
5624                                                       start, num_bytes);
5625                         BUG_ON(ret);
5626
5627                         start = caching_ctl->progress;
5628                         num_bytes = ins->objectid + ins->offset -
5629                                     caching_ctl->progress;
5630                         ret = add_excluded_extent(root, start, num_bytes);
5631                         BUG_ON(ret);
5632                 }
5633
5634                 mutex_unlock(&caching_ctl->mutex);
5635                 put_caching_control(caching_ctl);
5636         }
5637
5638         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
5639                                           RESERVE_ALLOC_NO_ACCOUNT);
5640         BUG_ON(ret);
5641         btrfs_put_block_group(block_group);
5642         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
5643                                          0, owner, offset, ins, 1);
5644         return ret;
5645 }
5646
5647 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
5648                                             struct btrfs_root *root,
5649                                             u64 bytenr, u32 blocksize,
5650                                             int level)
5651 {
5652         struct extent_buffer *buf;
5653
5654         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
5655         if (!buf)
5656                 return ERR_PTR(-ENOMEM);
5657         btrfs_set_header_generation(buf, trans->transid);
5658         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
5659         btrfs_tree_lock(buf);
5660         clean_tree_block(trans, root, buf);
5661
5662         btrfs_set_lock_blocking(buf);
5663         btrfs_set_buffer_uptodate(buf);
5664
5665         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
5666                 /*
5667                  * we allow two log transactions at a time, use different
5668                  * EXENT bit to differentiate dirty pages.
5669                  */
5670                 if (root->log_transid % 2 == 0)
5671                         set_extent_dirty(&root->dirty_log_pages, buf->start,
5672                                         buf->start + buf->len - 1, GFP_NOFS);
5673                 else
5674                         set_extent_new(&root->dirty_log_pages, buf->start,
5675                                         buf->start + buf->len - 1, GFP_NOFS);
5676         } else {
5677                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
5678                          buf->start + buf->len - 1, GFP_NOFS);
5679         }
5680         trans->blocks_used++;
5681         /* this returns a buffer locked for blocking */
5682         return buf;
5683 }
5684
5685 static struct btrfs_block_rsv *
5686 use_block_rsv(struct btrfs_trans_handle *trans,
5687               struct btrfs_root *root, u32 blocksize)
5688 {
5689         struct btrfs_block_rsv *block_rsv;
5690         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5691         int ret;
5692
5693         block_rsv = get_block_rsv(trans, root);
5694
5695         if (block_rsv->size == 0) {
5696                 ret = reserve_metadata_bytes(trans, root, block_rsv,
5697                                              blocksize, 0);
5698                 /*
5699                  * If we couldn't reserve metadata bytes try and use some from
5700                  * the global reserve.
5701                  */
5702                 if (ret && block_rsv != global_rsv) {
5703                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5704                         if (!ret)
5705                                 return global_rsv;
5706                         return ERR_PTR(ret);
5707                 } else if (ret) {
5708                         return ERR_PTR(ret);
5709                 }
5710                 return block_rsv;
5711         }
5712
5713         ret = block_rsv_use_bytes(block_rsv, blocksize);
5714         if (!ret)
5715                 return block_rsv;
5716         if (ret) {
5717                 WARN_ON(1);
5718                 ret = reserve_metadata_bytes(trans, root, block_rsv, blocksize,
5719                                              0);
5720                 if (!ret) {
5721                         return block_rsv;
5722                 } else if (ret && block_rsv != global_rsv) {
5723                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5724                         if (!ret)
5725                                 return global_rsv;
5726                 }
5727         }
5728
5729         return ERR_PTR(-ENOSPC);
5730 }
5731
5732 static void unuse_block_rsv(struct btrfs_block_rsv *block_rsv, u32 blocksize)
5733 {
5734         block_rsv_add_bytes(block_rsv, blocksize, 0);
5735         block_rsv_release_bytes(block_rsv, NULL, 0);
5736 }
5737
5738 /*
5739  * finds a free extent and does all the dirty work required for allocation
5740  * returns the key for the extent through ins, and a tree buffer for
5741  * the first block of the extent through buf.
5742  *
5743  * returns the tree buffer or NULL.
5744  */
5745 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
5746                                         struct btrfs_root *root, u32 blocksize,
5747                                         u64 parent, u64 root_objectid,
5748                                         struct btrfs_disk_key *key, int level,
5749                                         u64 hint, u64 empty_size)
5750 {
5751         struct btrfs_key ins;
5752         struct btrfs_block_rsv *block_rsv;
5753         struct extent_buffer *buf;
5754         u64 flags = 0;
5755         int ret;
5756
5757
5758         block_rsv = use_block_rsv(trans, root, blocksize);
5759         if (IS_ERR(block_rsv))
5760                 return ERR_CAST(block_rsv);
5761
5762         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
5763                                    empty_size, hint, (u64)-1, &ins, 0);
5764         if (ret) {
5765                 unuse_block_rsv(block_rsv, blocksize);
5766                 return ERR_PTR(ret);
5767         }
5768
5769         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
5770                                     blocksize, level);
5771         BUG_ON(IS_ERR(buf));
5772
5773         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
5774                 if (parent == 0)
5775                         parent = ins.objectid;
5776                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
5777         } else
5778                 BUG_ON(parent > 0);
5779
5780         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
5781                 struct btrfs_delayed_extent_op *extent_op;
5782                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
5783                 BUG_ON(!extent_op);
5784                 if (key)
5785                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
5786                 else
5787                         memset(&extent_op->key, 0, sizeof(extent_op->key));
5788                 extent_op->flags_to_set = flags;
5789                 extent_op->update_key = 1;
5790                 extent_op->update_flags = 1;
5791                 extent_op->is_data = 0;
5792
5793                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
5794                                         ins.offset, parent, root_objectid,
5795                                         level, BTRFS_ADD_DELAYED_EXTENT,
5796                                         extent_op);
5797                 BUG_ON(ret);
5798         }
5799         return buf;
5800 }
5801
5802 struct walk_control {
5803         u64 refs[BTRFS_MAX_LEVEL];
5804         u64 flags[BTRFS_MAX_LEVEL];
5805         struct btrfs_key update_progress;
5806         int stage;
5807         int level;
5808         int shared_level;
5809         int update_ref;
5810         int keep_locks;
5811         int reada_slot;
5812         int reada_count;
5813 };
5814
5815 #define DROP_REFERENCE  1
5816 #define UPDATE_BACKREF  2
5817
5818 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
5819                                      struct btrfs_root *root,
5820                                      struct walk_control *wc,
5821                                      struct btrfs_path *path)
5822 {
5823         u64 bytenr;
5824         u64 generation;
5825         u64 refs;
5826         u64 flags;
5827         u32 nritems;
5828         u32 blocksize;
5829         struct btrfs_key key;
5830         struct extent_buffer *eb;
5831         int ret;
5832         int slot;
5833         int nread = 0;
5834
5835         if (path->slots[wc->level] < wc->reada_slot) {
5836                 wc->reada_count = wc->reada_count * 2 / 3;
5837                 wc->reada_count = max(wc->reada_count, 2);
5838         } else {
5839                 wc->reada_count = wc->reada_count * 3 / 2;
5840                 wc->reada_count = min_t(int, wc->reada_count,
5841                                         BTRFS_NODEPTRS_PER_BLOCK(root));
5842         }
5843
5844         eb = path->nodes[wc->level];
5845         nritems = btrfs_header_nritems(eb);
5846         blocksize = btrfs_level_size(root, wc->level - 1);
5847
5848         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
5849                 if (nread >= wc->reada_count)
5850                         break;
5851
5852                 cond_resched();
5853                 bytenr = btrfs_node_blockptr(eb, slot);
5854                 generation = btrfs_node_ptr_generation(eb, slot);
5855
5856                 if (slot == path->slots[wc->level])
5857                         goto reada;
5858
5859                 if (wc->stage == UPDATE_BACKREF &&
5860                     generation <= root->root_key.offset)
5861                         continue;
5862
5863                 /* We don't lock the tree block, it's OK to be racy here */
5864                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
5865                                                &refs, &flags);
5866                 BUG_ON(ret);
5867                 BUG_ON(refs == 0);
5868
5869                 if (wc->stage == DROP_REFERENCE) {
5870                         if (refs == 1)
5871                                 goto reada;
5872
5873                         if (wc->level == 1 &&
5874                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5875                                 continue;
5876                         if (!wc->update_ref ||
5877                             generation <= root->root_key.offset)
5878                                 continue;
5879                         btrfs_node_key_to_cpu(eb, &key, slot);
5880                         ret = btrfs_comp_cpu_keys(&key,
5881                                                   &wc->update_progress);
5882                         if (ret < 0)
5883                                 continue;
5884                 } else {
5885                         if (wc->level == 1 &&
5886                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5887                                 continue;
5888                 }
5889 reada:
5890                 ret = readahead_tree_block(root, bytenr, blocksize,
5891                                            generation);
5892                 if (ret)
5893                         break;
5894                 nread++;
5895         }
5896         wc->reada_slot = slot;
5897 }
5898
5899 /*
5900  * hepler to process tree block while walking down the tree.
5901  *
5902  * when wc->stage == UPDATE_BACKREF, this function updates
5903  * back refs for pointers in the block.
5904  *
5905  * NOTE: return value 1 means we should stop walking down.
5906  */
5907 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
5908                                    struct btrfs_root *root,
5909                                    struct btrfs_path *path,
5910                                    struct walk_control *wc, int lookup_info)
5911 {
5912         int level = wc->level;
5913         struct extent_buffer *eb = path->nodes[level];
5914         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5915         int ret;
5916
5917         if (wc->stage == UPDATE_BACKREF &&
5918             btrfs_header_owner(eb) != root->root_key.objectid)
5919                 return 1;
5920
5921         /*
5922          * when reference count of tree block is 1, it won't increase
5923          * again. once full backref flag is set, we never clear it.
5924          */
5925         if (lookup_info &&
5926             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
5927              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
5928                 BUG_ON(!path->locks[level]);
5929                 ret = btrfs_lookup_extent_info(trans, root,
5930                                                eb->start, eb->len,
5931                                                &wc->refs[level],
5932                                                &wc->flags[level]);
5933                 BUG_ON(ret);
5934                 BUG_ON(wc->refs[level] == 0);
5935         }
5936
5937         if (wc->stage == DROP_REFERENCE) {
5938                 if (wc->refs[level] > 1)
5939                         return 1;
5940
5941                 if (path->locks[level] && !wc->keep_locks) {
5942                         btrfs_tree_unlock_rw(eb, path->locks[level]);
5943                         path->locks[level] = 0;
5944                 }
5945                 return 0;
5946         }
5947
5948         /* wc->stage == UPDATE_BACKREF */
5949         if (!(wc->flags[level] & flag)) {
5950                 BUG_ON(!path->locks[level]);
5951                 ret = btrfs_inc_ref(trans, root, eb, 1);
5952                 BUG_ON(ret);
5953                 ret = btrfs_dec_ref(trans, root, eb, 0);
5954                 BUG_ON(ret);
5955                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
5956                                                   eb->len, flag, 0);
5957                 BUG_ON(ret);
5958                 wc->flags[level] |= flag;
5959         }
5960
5961         /*
5962          * the block is shared by multiple trees, so it's not good to
5963          * keep the tree lock
5964          */
5965         if (path->locks[level] && level > 0) {
5966                 btrfs_tree_unlock_rw(eb, path->locks[level]);
5967                 path->locks[level] = 0;
5968         }
5969         return 0;
5970 }
5971
5972 /*
5973  * hepler to process tree block pointer.
5974  *
5975  * when wc->stage == DROP_REFERENCE, this function checks
5976  * reference count of the block pointed to. if the block
5977  * is shared and we need update back refs for the subtree
5978  * rooted at the block, this function changes wc->stage to
5979  * UPDATE_BACKREF. if the block is shared and there is no
5980  * need to update back, this function drops the reference
5981  * to the block.
5982  *
5983  * NOTE: return value 1 means we should stop walking down.
5984  */
5985 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
5986                                  struct btrfs_root *root,
5987                                  struct btrfs_path *path,
5988                                  struct walk_control *wc, int *lookup_info)
5989 {
5990         u64 bytenr;
5991         u64 generation;
5992         u64 parent;
5993         u32 blocksize;
5994         struct btrfs_key key;
5995         struct extent_buffer *next;
5996         int level = wc->level;
5997         int reada = 0;
5998         int ret = 0;
5999
6000         generation = btrfs_node_ptr_generation(path->nodes[level],
6001                                                path->slots[level]);
6002         /*
6003          * if the lower level block was created before the snapshot
6004          * was created, we know there is no need to update back refs
6005          * for the subtree
6006          */
6007         if (wc->stage == UPDATE_BACKREF &&
6008             generation <= root->root_key.offset) {
6009                 *lookup_info = 1;
6010                 return 1;
6011         }
6012
6013         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6014         blocksize = btrfs_level_size(root, level - 1);
6015
6016         next = btrfs_find_tree_block(root, bytenr, blocksize);
6017         if (!next) {
6018                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6019                 if (!next)
6020                         return -ENOMEM;
6021                 reada = 1;
6022         }
6023         btrfs_tree_lock(next);
6024         btrfs_set_lock_blocking(next);
6025
6026         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6027                                        &wc->refs[level - 1],
6028                                        &wc->flags[level - 1]);
6029         BUG_ON(ret);
6030         BUG_ON(wc->refs[level - 1] == 0);
6031         *lookup_info = 0;
6032
6033         if (wc->stage == DROP_REFERENCE) {
6034                 if (wc->refs[level - 1] > 1) {
6035                         if (level == 1 &&
6036                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6037                                 goto skip;
6038
6039                         if (!wc->update_ref ||
6040                             generation <= root->root_key.offset)
6041                                 goto skip;
6042
6043                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6044                                               path->slots[level]);
6045                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6046                         if (ret < 0)
6047                                 goto skip;
6048
6049                         wc->stage = UPDATE_BACKREF;
6050                         wc->shared_level = level - 1;
6051                 }
6052         } else {
6053                 if (level == 1 &&
6054                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6055                         goto skip;
6056         }
6057
6058         if (!btrfs_buffer_uptodate(next, generation)) {
6059                 btrfs_tree_unlock(next);
6060                 free_extent_buffer(next);
6061                 next = NULL;
6062                 *lookup_info = 1;
6063         }
6064
6065         if (!next) {
6066                 if (reada && level == 1)
6067                         reada_walk_down(trans, root, wc, path);
6068                 next = read_tree_block(root, bytenr, blocksize, generation);
6069                 if (!next)
6070                         return -EIO;
6071                 btrfs_tree_lock(next);
6072                 btrfs_set_lock_blocking(next);
6073         }
6074
6075         level--;
6076         BUG_ON(level != btrfs_header_level(next));
6077         path->nodes[level] = next;
6078         path->slots[level] = 0;
6079         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6080         wc->level = level;
6081         if (wc->level == 1)
6082                 wc->reada_slot = 0;
6083         return 0;
6084 skip:
6085         wc->refs[level - 1] = 0;
6086         wc->flags[level - 1] = 0;
6087         if (wc->stage == DROP_REFERENCE) {
6088                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6089                         parent = path->nodes[level]->start;
6090                 } else {
6091                         BUG_ON(root->root_key.objectid !=
6092                                btrfs_header_owner(path->nodes[level]));
6093                         parent = 0;
6094                 }
6095
6096                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6097                                         root->root_key.objectid, level - 1, 0);
6098                 BUG_ON(ret);
6099         }
6100         btrfs_tree_unlock(next);
6101         free_extent_buffer(next);
6102         *lookup_info = 1;
6103         return 1;
6104 }
6105
6106 /*
6107  * hepler to process tree block while walking up the tree.
6108  *
6109  * when wc->stage == DROP_REFERENCE, this function drops
6110  * reference count on the block.
6111  *
6112  * when wc->stage == UPDATE_BACKREF, this function changes
6113  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6114  * to UPDATE_BACKREF previously while processing the block.
6115  *
6116  * NOTE: return value 1 means we should stop walking up.
6117  */
6118 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6119                                  struct btrfs_root *root,
6120                                  struct btrfs_path *path,
6121                                  struct walk_control *wc)
6122 {
6123         int ret;
6124         int level = wc->level;
6125         struct extent_buffer *eb = path->nodes[level];
6126         u64 parent = 0;
6127
6128         if (wc->stage == UPDATE_BACKREF) {
6129                 BUG_ON(wc->shared_level < level);
6130                 if (level < wc->shared_level)
6131                         goto out;
6132
6133                 ret = find_next_key(path, level + 1, &wc->update_progress);
6134                 if (ret > 0)
6135                         wc->update_ref = 0;
6136
6137                 wc->stage = DROP_REFERENCE;
6138                 wc->shared_level = -1;
6139                 path->slots[level] = 0;
6140
6141                 /*
6142                  * check reference count again if the block isn't locked.
6143                  * we should start walking down the tree again if reference
6144                  * count is one.
6145                  */
6146                 if (!path->locks[level]) {
6147                         BUG_ON(level == 0);
6148                         btrfs_tree_lock(eb);
6149                         btrfs_set_lock_blocking(eb);
6150                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6151
6152                         ret = btrfs_lookup_extent_info(trans, root,
6153                                                        eb->start, eb->len,
6154                                                        &wc->refs[level],
6155                                                        &wc->flags[level]);
6156                         BUG_ON(ret);
6157                         BUG_ON(wc->refs[level] == 0);
6158                         if (wc->refs[level] == 1) {
6159                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6160                                 return 1;
6161                         }
6162                 }
6163         }
6164
6165         /* wc->stage == DROP_REFERENCE */
6166         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6167
6168         if (wc->refs[level] == 1) {
6169                 if (level == 0) {
6170                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6171                                 ret = btrfs_dec_ref(trans, root, eb, 1);
6172                         else
6173                                 ret = btrfs_dec_ref(trans, root, eb, 0);
6174                         BUG_ON(ret);
6175                 }
6176                 /* make block locked assertion in clean_tree_block happy */
6177                 if (!path->locks[level] &&
6178                     btrfs_header_generation(eb) == trans->transid) {
6179                         btrfs_tree_lock(eb);
6180                         btrfs_set_lock_blocking(eb);
6181                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6182                 }
6183                 clean_tree_block(trans, root, eb);
6184         }
6185
6186         if (eb == root->node) {
6187                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6188                         parent = eb->start;
6189                 else
6190                         BUG_ON(root->root_key.objectid !=
6191                                btrfs_header_owner(eb));
6192         } else {
6193                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6194                         parent = path->nodes[level + 1]->start;
6195                 else
6196                         BUG_ON(root->root_key.objectid !=
6197                                btrfs_header_owner(path->nodes[level + 1]));
6198         }
6199
6200         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6201 out:
6202         wc->refs[level] = 0;
6203         wc->flags[level] = 0;
6204         return 0;
6205 }
6206
6207 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6208                                    struct btrfs_root *root,
6209                                    struct btrfs_path *path,
6210                                    struct walk_control *wc)
6211 {
6212         int level = wc->level;
6213         int lookup_info = 1;
6214         int ret;
6215
6216         while (level >= 0) {
6217                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6218                 if (ret > 0)
6219                         break;
6220
6221                 if (level == 0)
6222                         break;
6223
6224                 if (path->slots[level] >=
6225                     btrfs_header_nritems(path->nodes[level]))
6226                         break;
6227
6228                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6229                 if (ret > 0) {
6230                         path->slots[level]++;
6231                         continue;
6232                 } else if (ret < 0)
6233                         return ret;
6234                 level = wc->level;
6235         }
6236         return 0;
6237 }
6238
6239 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6240                                  struct btrfs_root *root,
6241                                  struct btrfs_path *path,
6242                                  struct walk_control *wc, int max_level)
6243 {
6244         int level = wc->level;
6245         int ret;
6246
6247         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6248         while (level < max_level && path->nodes[level]) {
6249                 wc->level = level;
6250                 if (path->slots[level] + 1 <
6251                     btrfs_header_nritems(path->nodes[level])) {
6252                         path->slots[level]++;
6253                         return 0;
6254                 } else {
6255                         ret = walk_up_proc(trans, root, path, wc);
6256                         if (ret > 0)
6257                                 return 0;
6258
6259                         if (path->locks[level]) {
6260                                 btrfs_tree_unlock_rw(path->nodes[level],
6261                                                      path->locks[level]);
6262                                 path->locks[level] = 0;
6263                         }
6264                         free_extent_buffer(path->nodes[level]);
6265                         path->nodes[level] = NULL;
6266                         level++;
6267                 }
6268         }
6269         return 1;
6270 }
6271
6272 /*
6273  * drop a subvolume tree.
6274  *
6275  * this function traverses the tree freeing any blocks that only
6276  * referenced by the tree.
6277  *
6278  * when a shared tree block is found. this function decreases its
6279  * reference count by one. if update_ref is true, this function
6280  * also make sure backrefs for the shared block and all lower level
6281  * blocks are properly updated.
6282  */
6283 void btrfs_drop_snapshot(struct btrfs_root *root,
6284                          struct btrfs_block_rsv *block_rsv, int update_ref)
6285 {
6286         struct btrfs_path *path;
6287         struct btrfs_trans_handle *trans;
6288         struct btrfs_root *tree_root = root->fs_info->tree_root;
6289         struct btrfs_root_item *root_item = &root->root_item;
6290         struct walk_control *wc;
6291         struct btrfs_key key;
6292         int err = 0;
6293         int ret;
6294         int level;
6295
6296         path = btrfs_alloc_path();
6297         if (!path) {
6298                 err = -ENOMEM;
6299                 goto out;
6300         }
6301
6302         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6303         if (!wc) {
6304                 btrfs_free_path(path);
6305                 err = -ENOMEM;
6306                 goto out;
6307         }
6308
6309         trans = btrfs_start_transaction(tree_root, 0);
6310         BUG_ON(IS_ERR(trans));
6311
6312         if (block_rsv)
6313                 trans->block_rsv = block_rsv;
6314
6315         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6316                 level = btrfs_header_level(root->node);
6317                 path->nodes[level] = btrfs_lock_root_node(root);
6318                 btrfs_set_lock_blocking(path->nodes[level]);
6319                 path->slots[level] = 0;
6320                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6321                 memset(&wc->update_progress, 0,
6322                        sizeof(wc->update_progress));
6323         } else {
6324                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6325                 memcpy(&wc->update_progress, &key,
6326                        sizeof(wc->update_progress));
6327
6328                 level = root_item->drop_level;
6329                 BUG_ON(level == 0);
6330                 path->lowest_level = level;
6331                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6332                 path->lowest_level = 0;
6333                 if (ret < 0) {
6334                         err = ret;
6335                         goto out_free;
6336                 }
6337                 WARN_ON(ret > 0);
6338
6339                 /*
6340                  * unlock our path, this is safe because only this
6341                  * function is allowed to delete this snapshot
6342                  */
6343                 btrfs_unlock_up_safe(path, 0);
6344
6345                 level = btrfs_header_level(root->node);
6346                 while (1) {
6347                         btrfs_tree_lock(path->nodes[level]);
6348                         btrfs_set_lock_blocking(path->nodes[level]);
6349
6350                         ret = btrfs_lookup_extent_info(trans, root,
6351                                                 path->nodes[level]->start,
6352                                                 path->nodes[level]->len,
6353                                                 &wc->refs[level],
6354                                                 &wc->flags[level]);
6355                         BUG_ON(ret);
6356                         BUG_ON(wc->refs[level] == 0);
6357
6358                         if (level == root_item->drop_level)
6359                                 break;
6360
6361                         btrfs_tree_unlock(path->nodes[level]);
6362                         WARN_ON(wc->refs[level] != 1);
6363                         level--;
6364                 }
6365         }
6366
6367         wc->level = level;
6368         wc->shared_level = -1;
6369         wc->stage = DROP_REFERENCE;
6370         wc->update_ref = update_ref;
6371         wc->keep_locks = 0;
6372         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6373
6374         while (1) {
6375                 ret = walk_down_tree(trans, root, path, wc);
6376                 if (ret < 0) {
6377                         err = ret;
6378                         break;
6379                 }
6380
6381                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
6382                 if (ret < 0) {
6383                         err = ret;
6384                         break;
6385                 }
6386
6387                 if (ret > 0) {
6388                         BUG_ON(wc->stage != DROP_REFERENCE);
6389                         break;
6390                 }
6391
6392                 if (wc->stage == DROP_REFERENCE) {
6393                         level = wc->level;
6394                         btrfs_node_key(path->nodes[level],
6395                                        &root_item->drop_progress,
6396                                        path->slots[level]);
6397                         root_item->drop_level = level;
6398                 }
6399
6400                 BUG_ON(wc->level == 0);
6401                 if (btrfs_should_end_transaction(trans, tree_root)) {
6402                         ret = btrfs_update_root(trans, tree_root,
6403                                                 &root->root_key,
6404                                                 root_item);
6405                         BUG_ON(ret);
6406
6407                         btrfs_end_transaction_throttle(trans, tree_root);
6408                         trans = btrfs_start_transaction(tree_root, 0);
6409                         BUG_ON(IS_ERR(trans));
6410                         if (block_rsv)
6411                                 trans->block_rsv = block_rsv;
6412                 }
6413         }
6414         btrfs_release_path(path);
6415         BUG_ON(err);
6416
6417         ret = btrfs_del_root(trans, tree_root, &root->root_key);
6418         BUG_ON(ret);
6419
6420         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
6421                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
6422                                            NULL, NULL);
6423                 BUG_ON(ret < 0);
6424                 if (ret > 0) {
6425                         /* if we fail to delete the orphan item this time
6426                          * around, it'll get picked up the next time.
6427                          *
6428                          * The most common failure here is just -ENOENT.
6429                          */
6430                         btrfs_del_orphan_item(trans, tree_root,
6431                                               root->root_key.objectid);
6432                 }
6433         }
6434
6435         if (root->in_radix) {
6436                 btrfs_free_fs_root(tree_root->fs_info, root);
6437         } else {
6438                 free_extent_buffer(root->node);
6439                 free_extent_buffer(root->commit_root);
6440                 kfree(root);
6441         }
6442 out_free:
6443         btrfs_end_transaction_throttle(trans, tree_root);
6444         kfree(wc);
6445         btrfs_free_path(path);
6446 out:
6447         if (err)
6448                 btrfs_std_error(root->fs_info, err);
6449         return;
6450 }
6451
6452 /*
6453  * drop subtree rooted at tree block 'node'.
6454  *
6455  * NOTE: this function will unlock and release tree block 'node'
6456  */
6457 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
6458                         struct btrfs_root *root,
6459                         struct extent_buffer *node,
6460                         struct extent_buffer *parent)
6461 {
6462         struct btrfs_path *path;
6463         struct walk_control *wc;
6464         int level;
6465         int parent_level;
6466         int ret = 0;
6467         int wret;
6468
6469         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
6470
6471         path = btrfs_alloc_path();
6472         if (!path)
6473                 return -ENOMEM;
6474
6475         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6476         if (!wc) {
6477                 btrfs_free_path(path);
6478                 return -ENOMEM;
6479         }
6480
6481         btrfs_assert_tree_locked(parent);
6482         parent_level = btrfs_header_level(parent);
6483         extent_buffer_get(parent);
6484         path->nodes[parent_level] = parent;
6485         path->slots[parent_level] = btrfs_header_nritems(parent);
6486
6487         btrfs_assert_tree_locked(node);
6488         level = btrfs_header_level(node);
6489         path->nodes[level] = node;
6490         path->slots[level] = 0;
6491         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6492
6493         wc->refs[parent_level] = 1;
6494         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6495         wc->level = level;
6496         wc->shared_level = -1;
6497         wc->stage = DROP_REFERENCE;
6498         wc->update_ref = 0;
6499         wc->keep_locks = 1;
6500         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6501
6502         while (1) {
6503                 wret = walk_down_tree(trans, root, path, wc);
6504                 if (wret < 0) {
6505                         ret = wret;
6506                         break;
6507                 }
6508
6509                 wret = walk_up_tree(trans, root, path, wc, parent_level);
6510                 if (wret < 0)
6511                         ret = wret;
6512                 if (wret != 0)
6513                         break;
6514         }
6515
6516         kfree(wc);
6517         btrfs_free_path(path);
6518         return ret;
6519 }
6520
6521 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
6522 {
6523         u64 num_devices;
6524         u64 stripped = BTRFS_BLOCK_GROUP_RAID0 |
6525                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
6526
6527         /*
6528          * we add in the count of missing devices because we want
6529          * to make sure that any RAID levels on a degraded FS
6530          * continue to be honored.
6531          */
6532         num_devices = root->fs_info->fs_devices->rw_devices +
6533                 root->fs_info->fs_devices->missing_devices;
6534
6535         if (num_devices == 1) {
6536                 stripped |= BTRFS_BLOCK_GROUP_DUP;
6537                 stripped = flags & ~stripped;
6538
6539                 /* turn raid0 into single device chunks */
6540                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
6541                         return stripped;
6542
6543                 /* turn mirroring into duplication */
6544                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
6545                              BTRFS_BLOCK_GROUP_RAID10))
6546                         return stripped | BTRFS_BLOCK_GROUP_DUP;
6547                 return flags;
6548         } else {
6549                 /* they already had raid on here, just return */
6550                 if (flags & stripped)
6551                         return flags;
6552
6553                 stripped |= BTRFS_BLOCK_GROUP_DUP;
6554                 stripped = flags & ~stripped;
6555
6556                 /* switch duplicated blocks with raid1 */
6557                 if (flags & BTRFS_BLOCK_GROUP_DUP)
6558                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
6559
6560                 /* turn single device chunks into raid0 */
6561                 return stripped | BTRFS_BLOCK_GROUP_RAID0;
6562         }
6563         return flags;
6564 }
6565
6566 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
6567 {
6568         struct btrfs_space_info *sinfo = cache->space_info;
6569         u64 num_bytes;
6570         u64 min_allocable_bytes;
6571         int ret = -ENOSPC;
6572
6573
6574         /*
6575          * We need some metadata space and system metadata space for
6576          * allocating chunks in some corner cases until we force to set
6577          * it to be readonly.
6578          */
6579         if ((sinfo->flags &
6580              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
6581             !force)
6582                 min_allocable_bytes = 1 * 1024 * 1024;
6583         else
6584                 min_allocable_bytes = 0;
6585
6586         spin_lock(&sinfo->lock);
6587         spin_lock(&cache->lock);
6588
6589         if (cache->ro) {
6590                 ret = 0;
6591                 goto out;
6592         }
6593
6594         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
6595                     cache->bytes_super - btrfs_block_group_used(&cache->item);
6596
6597         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
6598             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
6599             min_allocable_bytes <= sinfo->total_bytes) {
6600                 sinfo->bytes_readonly += num_bytes;
6601                 cache->ro = 1;
6602                 ret = 0;
6603         }
6604 out:
6605         spin_unlock(&cache->lock);
6606         spin_unlock(&sinfo->lock);
6607         return ret;
6608 }
6609
6610 int btrfs_set_block_group_ro(struct btrfs_root *root,
6611                              struct btrfs_block_group_cache *cache)
6612
6613 {
6614         struct btrfs_trans_handle *trans;
6615         u64 alloc_flags;
6616         int ret;
6617
6618         BUG_ON(cache->ro);
6619
6620         trans = btrfs_join_transaction(root);
6621         BUG_ON(IS_ERR(trans));
6622
6623         alloc_flags = update_block_group_flags(root, cache->flags);
6624         if (alloc_flags != cache->flags)
6625                 do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6626                                CHUNK_ALLOC_FORCE);
6627
6628         ret = set_block_group_ro(cache, 0);
6629         if (!ret)
6630                 goto out;
6631         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
6632         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6633                              CHUNK_ALLOC_FORCE);
6634         if (ret < 0)
6635                 goto out;
6636         ret = set_block_group_ro(cache, 0);
6637 out:
6638         btrfs_end_transaction(trans, root);
6639         return ret;
6640 }
6641
6642 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
6643                             struct btrfs_root *root, u64 type)
6644 {
6645         u64 alloc_flags = get_alloc_profile(root, type);
6646         return do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6647                               CHUNK_ALLOC_FORCE);
6648 }
6649
6650 /*
6651  * helper to account the unused space of all the readonly block group in the
6652  * list. takes mirrors into account.
6653  */
6654 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
6655 {
6656         struct btrfs_block_group_cache *block_group;
6657         u64 free_bytes = 0;
6658         int factor;
6659
6660         list_for_each_entry(block_group, groups_list, list) {
6661                 spin_lock(&block_group->lock);
6662
6663                 if (!block_group->ro) {
6664                         spin_unlock(&block_group->lock);
6665                         continue;
6666                 }
6667
6668                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
6669                                           BTRFS_BLOCK_GROUP_RAID10 |
6670                                           BTRFS_BLOCK_GROUP_DUP))
6671                         factor = 2;
6672                 else
6673                         factor = 1;
6674
6675                 free_bytes += (block_group->key.offset -
6676                                btrfs_block_group_used(&block_group->item)) *
6677                                factor;
6678
6679                 spin_unlock(&block_group->lock);
6680         }
6681
6682         return free_bytes;
6683 }
6684
6685 /*
6686  * helper to account the unused space of all the readonly block group in the
6687  * space_info. takes mirrors into account.
6688  */
6689 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
6690 {
6691         int i;
6692         u64 free_bytes = 0;
6693
6694         spin_lock(&sinfo->lock);
6695
6696         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
6697                 if (!list_empty(&sinfo->block_groups[i]))
6698                         free_bytes += __btrfs_get_ro_block_group_free_space(
6699                                                 &sinfo->block_groups[i]);
6700
6701         spin_unlock(&sinfo->lock);
6702
6703         return free_bytes;
6704 }
6705
6706 int btrfs_set_block_group_rw(struct btrfs_root *root,
6707                               struct btrfs_block_group_cache *cache)
6708 {
6709         struct btrfs_space_info *sinfo = cache->space_info;
6710         u64 num_bytes;
6711
6712         BUG_ON(!cache->ro);
6713
6714         spin_lock(&sinfo->lock);
6715         spin_lock(&cache->lock);
6716         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
6717                     cache->bytes_super - btrfs_block_group_used(&cache->item);
6718         sinfo->bytes_readonly -= num_bytes;
6719         cache->ro = 0;
6720         spin_unlock(&cache->lock);
6721         spin_unlock(&sinfo->lock);
6722         return 0;
6723 }
6724
6725 /*
6726  * checks to see if its even possible to relocate this block group.
6727  *
6728  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
6729  * ok to go ahead and try.
6730  */
6731 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
6732 {
6733         struct btrfs_block_group_cache *block_group;
6734         struct btrfs_space_info *space_info;
6735         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
6736         struct btrfs_device *device;
6737         u64 min_free;
6738         u64 dev_min = 1;
6739         u64 dev_nr = 0;
6740         int index;
6741         int full = 0;
6742         int ret = 0;
6743
6744         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
6745
6746         /* odd, couldn't find the block group, leave it alone */
6747         if (!block_group)
6748                 return -1;
6749
6750         min_free = btrfs_block_group_used(&block_group->item);
6751
6752         /* no bytes used, we're good */
6753         if (!min_free)
6754                 goto out;
6755
6756         space_info = block_group->space_info;
6757         spin_lock(&space_info->lock);
6758
6759         full = space_info->full;
6760
6761         /*
6762          * if this is the last block group we have in this space, we can't
6763          * relocate it unless we're able to allocate a new chunk below.
6764          *
6765          * Otherwise, we need to make sure we have room in the space to handle
6766          * all of the extents from this block group.  If we can, we're good
6767          */
6768         if ((space_info->total_bytes != block_group->key.offset) &&
6769             (space_info->bytes_used + space_info->bytes_reserved +
6770              space_info->bytes_pinned + space_info->bytes_readonly +
6771              min_free < space_info->total_bytes)) {
6772                 spin_unlock(&space_info->lock);
6773                 goto out;
6774         }
6775         spin_unlock(&space_info->lock);
6776
6777         /*
6778          * ok we don't have enough space, but maybe we have free space on our
6779          * devices to allocate new chunks for relocation, so loop through our
6780          * alloc devices and guess if we have enough space.  However, if we
6781          * were marked as full, then we know there aren't enough chunks, and we
6782          * can just return.
6783          */
6784         ret = -1;
6785         if (full)
6786                 goto out;
6787
6788         /*
6789          * index:
6790          *      0: raid10
6791          *      1: raid1
6792          *      2: dup
6793          *      3: raid0
6794          *      4: single
6795          */
6796         index = get_block_group_index(block_group);
6797         if (index == 0) {
6798                 dev_min = 4;
6799                 /* Divide by 2 */
6800                 min_free >>= 1;
6801         } else if (index == 1) {
6802                 dev_min = 2;
6803         } else if (index == 2) {
6804                 /* Multiply by 2 */
6805                 min_free <<= 1;
6806         } else if (index == 3) {
6807                 dev_min = fs_devices->rw_devices;
6808                 do_div(min_free, dev_min);
6809         }
6810
6811         mutex_lock(&root->fs_info->chunk_mutex);
6812         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
6813                 u64 dev_offset;
6814
6815                 /*
6816                  * check to make sure we can actually find a chunk with enough
6817                  * space to fit our block group in.
6818                  */
6819                 if (device->total_bytes > device->bytes_used + min_free) {
6820                         ret = find_free_dev_extent(NULL, device, min_free,
6821                                                    &dev_offset, NULL);
6822                         if (!ret)
6823                                 dev_nr++;
6824
6825                         if (dev_nr >= dev_min)
6826                                 break;
6827
6828                         ret = -1;
6829                 }
6830         }
6831         mutex_unlock(&root->fs_info->chunk_mutex);
6832 out:
6833         btrfs_put_block_group(block_group);
6834         return ret;
6835 }
6836
6837 static int find_first_block_group(struct btrfs_root *root,
6838                 struct btrfs_path *path, struct btrfs_key *key)
6839 {
6840         int ret = 0;
6841         struct btrfs_key found_key;
6842         struct extent_buffer *leaf;
6843         int slot;
6844
6845         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
6846         if (ret < 0)
6847                 goto out;
6848
6849         while (1) {
6850                 slot = path->slots[0];
6851                 leaf = path->nodes[0];
6852                 if (slot >= btrfs_header_nritems(leaf)) {
6853                         ret = btrfs_next_leaf(root, path);
6854                         if (ret == 0)
6855                                 continue;
6856                         if (ret < 0)
6857                                 goto out;
6858                         break;
6859                 }
6860                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
6861
6862                 if (found_key.objectid >= key->objectid &&
6863                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
6864                         ret = 0;
6865                         goto out;
6866                 }
6867                 path->slots[0]++;
6868         }
6869 out:
6870         return ret;
6871 }
6872
6873 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
6874 {
6875         struct btrfs_block_group_cache *block_group;
6876         u64 last = 0;
6877
6878         while (1) {
6879                 struct inode *inode;
6880
6881                 block_group = btrfs_lookup_first_block_group(info, last);
6882                 while (block_group) {
6883                         spin_lock(&block_group->lock);
6884                         if (block_group->iref)
6885                                 break;
6886                         spin_unlock(&block_group->lock);
6887                         block_group = next_block_group(info->tree_root,
6888                                                        block_group);
6889                 }
6890                 if (!block_group) {
6891                         if (last == 0)
6892                                 break;
6893                         last = 0;
6894                         continue;
6895                 }
6896
6897                 inode = block_group->inode;
6898                 block_group->iref = 0;
6899                 block_group->inode = NULL;
6900                 spin_unlock(&block_group->lock);
6901                 iput(inode);
6902                 last = block_group->key.objectid + block_group->key.offset;
6903                 btrfs_put_block_group(block_group);
6904         }
6905 }
6906
6907 int btrfs_free_block_groups(struct btrfs_fs_info *info)
6908 {
6909         struct btrfs_block_group_cache *block_group;
6910         struct btrfs_space_info *space_info;
6911         struct btrfs_caching_control *caching_ctl;
6912         struct rb_node *n;
6913
6914         down_write(&info->extent_commit_sem);
6915         while (!list_empty(&info->caching_block_groups)) {
6916                 caching_ctl = list_entry(info->caching_block_groups.next,
6917                                          struct btrfs_caching_control, list);
6918                 list_del(&caching_ctl->list);
6919                 put_caching_control(caching_ctl);
6920         }
6921         up_write(&info->extent_commit_sem);
6922
6923         spin_lock(&info->block_group_cache_lock);
6924         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
6925                 block_group = rb_entry(n, struct btrfs_block_group_cache,
6926                                        cache_node);
6927                 rb_erase(&block_group->cache_node,
6928                          &info->block_group_cache_tree);
6929                 spin_unlock(&info->block_group_cache_lock);
6930
6931                 down_write(&block_group->space_info->groups_sem);
6932                 list_del(&block_group->list);
6933                 up_write(&block_group->space_info->groups_sem);
6934
6935                 if (block_group->cached == BTRFS_CACHE_STARTED)
6936                         wait_block_group_cache_done(block_group);
6937
6938                 /*
6939                  * We haven't cached this block group, which means we could
6940                  * possibly have excluded extents on this block group.
6941                  */
6942                 if (block_group->cached == BTRFS_CACHE_NO)
6943                         free_excluded_extents(info->extent_root, block_group);
6944
6945                 btrfs_remove_free_space_cache(block_group);
6946                 btrfs_put_block_group(block_group);
6947
6948                 spin_lock(&info->block_group_cache_lock);
6949         }
6950         spin_unlock(&info->block_group_cache_lock);
6951
6952         /* now that all the block groups are freed, go through and
6953          * free all the space_info structs.  This is only called during
6954          * the final stages of unmount, and so we know nobody is
6955          * using them.  We call synchronize_rcu() once before we start,
6956          * just to be on the safe side.
6957          */
6958         synchronize_rcu();
6959
6960         release_global_block_rsv(info);
6961
6962         while(!list_empty(&info->space_info)) {
6963                 space_info = list_entry(info->space_info.next,
6964                                         struct btrfs_space_info,
6965                                         list);
6966                 if (space_info->bytes_pinned > 0 ||
6967                     space_info->bytes_reserved > 0 ||
6968                     space_info->bytes_may_use > 0) {
6969                         WARN_ON(1);
6970                         dump_space_info(space_info, 0, 0);
6971                 }
6972                 list_del(&space_info->list);
6973                 kfree(space_info);
6974         }
6975         return 0;
6976 }
6977
6978 static void __link_block_group(struct btrfs_space_info *space_info,
6979                                struct btrfs_block_group_cache *cache)
6980 {
6981         int index = get_block_group_index(cache);
6982
6983         down_write(&space_info->groups_sem);
6984         list_add_tail(&cache->list, &space_info->block_groups[index]);
6985         up_write(&space_info->groups_sem);
6986 }
6987
6988 int btrfs_read_block_groups(struct btrfs_root *root)
6989 {
6990         struct btrfs_path *path;
6991         int ret;
6992         struct btrfs_block_group_cache *cache;
6993         struct btrfs_fs_info *info = root->fs_info;
6994         struct btrfs_space_info *space_info;
6995         struct btrfs_key key;
6996         struct btrfs_key found_key;
6997         struct extent_buffer *leaf;
6998         int need_clear = 0;
6999         u64 cache_gen;
7000
7001         root = info->extent_root;
7002         key.objectid = 0;
7003         key.offset = 0;
7004         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7005         path = btrfs_alloc_path();
7006         if (!path)
7007                 return -ENOMEM;
7008         path->reada = 1;
7009
7010         cache_gen = btrfs_super_cache_generation(&root->fs_info->super_copy);
7011         if (cache_gen != 0 &&
7012             btrfs_super_generation(&root->fs_info->super_copy) != cache_gen)
7013                 need_clear = 1;
7014         if (btrfs_test_opt(root, CLEAR_CACHE))
7015                 need_clear = 1;
7016         if (!btrfs_test_opt(root, SPACE_CACHE) && cache_gen)
7017                 printk(KERN_INFO "btrfs: disk space caching is enabled\n");
7018
7019         while (1) {
7020                 ret = find_first_block_group(root, path, &key);
7021                 if (ret > 0)
7022                         break;
7023                 if (ret != 0)
7024                         goto error;
7025                 leaf = path->nodes[0];
7026                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7027                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7028                 if (!cache) {
7029                         ret = -ENOMEM;
7030                         goto error;
7031                 }
7032                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7033                                                 GFP_NOFS);
7034                 if (!cache->free_space_ctl) {
7035                         kfree(cache);
7036                         ret = -ENOMEM;
7037                         goto error;
7038                 }
7039
7040                 atomic_set(&cache->count, 1);
7041                 spin_lock_init(&cache->lock);
7042                 cache->fs_info = info;
7043                 INIT_LIST_HEAD(&cache->list);
7044                 INIT_LIST_HEAD(&cache->cluster_list);
7045
7046                 if (need_clear)
7047                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7048
7049                 read_extent_buffer(leaf, &cache->item,
7050                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7051                                    sizeof(cache->item));
7052                 memcpy(&cache->key, &found_key, sizeof(found_key));
7053
7054                 key.objectid = found_key.objectid + found_key.offset;
7055                 btrfs_release_path(path);
7056                 cache->flags = btrfs_block_group_flags(&cache->item);
7057                 cache->sectorsize = root->sectorsize;
7058
7059                 btrfs_init_free_space_ctl(cache);
7060
7061                 /*
7062                  * We need to exclude the super stripes now so that the space
7063                  * info has super bytes accounted for, otherwise we'll think
7064                  * we have more space than we actually do.
7065                  */
7066                 exclude_super_stripes(root, cache);
7067
7068                 /*
7069                  * check for two cases, either we are full, and therefore
7070                  * don't need to bother with the caching work since we won't
7071                  * find any space, or we are empty, and we can just add all
7072                  * the space in and be done with it.  This saves us _alot_ of
7073                  * time, particularly in the full case.
7074                  */
7075                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7076                         cache->last_byte_to_unpin = (u64)-1;
7077                         cache->cached = BTRFS_CACHE_FINISHED;
7078                         free_excluded_extents(root, cache);
7079                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7080                         cache->last_byte_to_unpin = (u64)-1;
7081                         cache->cached = BTRFS_CACHE_FINISHED;
7082                         add_new_free_space(cache, root->fs_info,
7083                                            found_key.objectid,
7084                                            found_key.objectid +
7085                                            found_key.offset);
7086                         free_excluded_extents(root, cache);
7087                 }
7088
7089                 ret = update_space_info(info, cache->flags, found_key.offset,
7090                                         btrfs_block_group_used(&cache->item),
7091                                         &space_info);
7092                 BUG_ON(ret);
7093                 cache->space_info = space_info;
7094                 spin_lock(&cache->space_info->lock);
7095                 cache->space_info->bytes_readonly += cache->bytes_super;
7096                 spin_unlock(&cache->space_info->lock);
7097
7098                 __link_block_group(space_info, cache);
7099
7100                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7101                 BUG_ON(ret);
7102
7103                 set_avail_alloc_bits(root->fs_info, cache->flags);
7104                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7105                         set_block_group_ro(cache, 1);
7106         }
7107
7108         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7109                 if (!(get_alloc_profile(root, space_info->flags) &
7110                       (BTRFS_BLOCK_GROUP_RAID10 |
7111                        BTRFS_BLOCK_GROUP_RAID1 |
7112                        BTRFS_BLOCK_GROUP_DUP)))
7113                         continue;
7114                 /*
7115                  * avoid allocating from un-mirrored block group if there are
7116                  * mirrored block groups.
7117                  */
7118                 list_for_each_entry(cache, &space_info->block_groups[3], list)
7119                         set_block_group_ro(cache, 1);
7120                 list_for_each_entry(cache, &space_info->block_groups[4], list)
7121                         set_block_group_ro(cache, 1);
7122         }
7123
7124         init_global_block_rsv(info);
7125         ret = 0;
7126 error:
7127         btrfs_free_path(path);
7128         return ret;
7129 }
7130
7131 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7132                            struct btrfs_root *root, u64 bytes_used,
7133                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7134                            u64 size)
7135 {
7136         int ret;
7137         struct btrfs_root *extent_root;
7138         struct btrfs_block_group_cache *cache;
7139
7140         extent_root = root->fs_info->extent_root;
7141
7142         root->fs_info->last_trans_log_full_commit = trans->transid;
7143
7144         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7145         if (!cache)
7146                 return -ENOMEM;
7147         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7148                                         GFP_NOFS);
7149         if (!cache->free_space_ctl) {
7150                 kfree(cache);
7151                 return -ENOMEM;
7152         }
7153
7154         cache->key.objectid = chunk_offset;
7155         cache->key.offset = size;
7156         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7157         cache->sectorsize = root->sectorsize;
7158         cache->fs_info = root->fs_info;
7159
7160         atomic_set(&cache->count, 1);
7161         spin_lock_init(&cache->lock);
7162         INIT_LIST_HEAD(&cache->list);
7163         INIT_LIST_HEAD(&cache->cluster_list);
7164
7165         btrfs_init_free_space_ctl(cache);
7166
7167         btrfs_set_block_group_used(&cache->item, bytes_used);
7168         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7169         cache->flags = type;
7170         btrfs_set_block_group_flags(&cache->item, type);
7171
7172         cache->last_byte_to_unpin = (u64)-1;
7173         cache->cached = BTRFS_CACHE_FINISHED;
7174         exclude_super_stripes(root, cache);
7175
7176         add_new_free_space(cache, root->fs_info, chunk_offset,
7177                            chunk_offset + size);
7178
7179         free_excluded_extents(root, cache);
7180
7181         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7182                                 &cache->space_info);
7183         BUG_ON(ret);
7184
7185         spin_lock(&cache->space_info->lock);
7186         cache->space_info->bytes_readonly += cache->bytes_super;
7187         spin_unlock(&cache->space_info->lock);
7188
7189         __link_block_group(cache->space_info, cache);
7190
7191         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7192         BUG_ON(ret);
7193
7194         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
7195                                 sizeof(cache->item));
7196         BUG_ON(ret);
7197
7198         set_avail_alloc_bits(extent_root->fs_info, type);
7199
7200         return 0;
7201 }
7202
7203 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7204                              struct btrfs_root *root, u64 group_start)
7205 {
7206         struct btrfs_path *path;
7207         struct btrfs_block_group_cache *block_group;
7208         struct btrfs_free_cluster *cluster;
7209         struct btrfs_root *tree_root = root->fs_info->tree_root;
7210         struct btrfs_key key;
7211         struct inode *inode;
7212         int ret;
7213         int factor;
7214
7215         root = root->fs_info->extent_root;
7216
7217         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7218         BUG_ON(!block_group);
7219         BUG_ON(!block_group->ro);
7220
7221         /*
7222          * Free the reserved super bytes from this block group before
7223          * remove it.
7224          */
7225         free_excluded_extents(root, block_group);
7226
7227         memcpy(&key, &block_group->key, sizeof(key));
7228         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7229                                   BTRFS_BLOCK_GROUP_RAID1 |
7230                                   BTRFS_BLOCK_GROUP_RAID10))
7231                 factor = 2;
7232         else
7233                 factor = 1;
7234
7235         /* make sure this block group isn't part of an allocation cluster */
7236         cluster = &root->fs_info->data_alloc_cluster;
7237         spin_lock(&cluster->refill_lock);
7238         btrfs_return_cluster_to_free_space(block_group, cluster);
7239         spin_unlock(&cluster->refill_lock);
7240
7241         /*
7242          * make sure this block group isn't part of a metadata
7243          * allocation cluster
7244          */
7245         cluster = &root->fs_info->meta_alloc_cluster;
7246         spin_lock(&cluster->refill_lock);
7247         btrfs_return_cluster_to_free_space(block_group, cluster);
7248         spin_unlock(&cluster->refill_lock);
7249
7250         path = btrfs_alloc_path();
7251         if (!path) {
7252                 ret = -ENOMEM;
7253                 goto out;
7254         }
7255
7256         inode = lookup_free_space_inode(root, block_group, path);
7257         if (!IS_ERR(inode)) {
7258                 ret = btrfs_orphan_add(trans, inode);
7259                 BUG_ON(ret);
7260                 clear_nlink(inode);
7261                 /* One for the block groups ref */
7262                 spin_lock(&block_group->lock);
7263                 if (block_group->iref) {
7264                         block_group->iref = 0;
7265                         block_group->inode = NULL;
7266                         spin_unlock(&block_group->lock);
7267                         iput(inode);
7268                 } else {
7269                         spin_unlock(&block_group->lock);
7270                 }
7271                 /* One for our lookup ref */
7272                 iput(inode);
7273         }
7274
7275         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
7276         key.offset = block_group->key.objectid;
7277         key.type = 0;
7278
7279         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
7280         if (ret < 0)
7281                 goto out;
7282         if (ret > 0)
7283                 btrfs_release_path(path);
7284         if (ret == 0) {
7285                 ret = btrfs_del_item(trans, tree_root, path);
7286                 if (ret)
7287                         goto out;
7288                 btrfs_release_path(path);
7289         }
7290
7291         spin_lock(&root->fs_info->block_group_cache_lock);
7292         rb_erase(&block_group->cache_node,
7293                  &root->fs_info->block_group_cache_tree);
7294         spin_unlock(&root->fs_info->block_group_cache_lock);
7295
7296         down_write(&block_group->space_info->groups_sem);
7297         /*
7298          * we must use list_del_init so people can check to see if they
7299          * are still on the list after taking the semaphore
7300          */
7301         list_del_init(&block_group->list);
7302         up_write(&block_group->space_info->groups_sem);
7303
7304         if (block_group->cached == BTRFS_CACHE_STARTED)
7305                 wait_block_group_cache_done(block_group);
7306
7307         btrfs_remove_free_space_cache(block_group);
7308
7309         spin_lock(&block_group->space_info->lock);
7310         block_group->space_info->total_bytes -= block_group->key.offset;
7311         block_group->space_info->bytes_readonly -= block_group->key.offset;
7312         block_group->space_info->disk_total -= block_group->key.offset * factor;
7313         spin_unlock(&block_group->space_info->lock);
7314
7315         memcpy(&key, &block_group->key, sizeof(key));
7316
7317         btrfs_clear_space_info_full(root->fs_info);
7318
7319         btrfs_put_block_group(block_group);
7320         btrfs_put_block_group(block_group);
7321
7322         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
7323         if (ret > 0)
7324                 ret = -EIO;
7325         if (ret < 0)
7326                 goto out;
7327
7328         ret = btrfs_del_item(trans, root, path);
7329 out:
7330         btrfs_free_path(path);
7331         return ret;
7332 }
7333
7334 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
7335 {
7336         struct btrfs_space_info *space_info;
7337         struct btrfs_super_block *disk_super;
7338         u64 features;
7339         u64 flags;
7340         int mixed = 0;
7341         int ret;
7342
7343         disk_super = &fs_info->super_copy;
7344         if (!btrfs_super_root(disk_super))
7345                 return 1;
7346
7347         features = btrfs_super_incompat_flags(disk_super);
7348         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
7349                 mixed = 1;
7350
7351         flags = BTRFS_BLOCK_GROUP_SYSTEM;
7352         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7353         if (ret)
7354                 goto out;
7355
7356         if (mixed) {
7357                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
7358                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7359         } else {
7360                 flags = BTRFS_BLOCK_GROUP_METADATA;
7361                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7362                 if (ret)
7363                         goto out;
7364
7365                 flags = BTRFS_BLOCK_GROUP_DATA;
7366                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7367         }
7368 out:
7369         return ret;
7370 }
7371
7372 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
7373 {
7374         return unpin_extent_range(root, start, end);
7375 }
7376
7377 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
7378                                u64 num_bytes, u64 *actual_bytes)
7379 {
7380         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
7381 }
7382
7383 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
7384 {
7385         struct btrfs_fs_info *fs_info = root->fs_info;
7386         struct btrfs_block_group_cache *cache = NULL;
7387         u64 group_trimmed;
7388         u64 start;
7389         u64 end;
7390         u64 trimmed = 0;
7391         int ret = 0;
7392
7393         cache = btrfs_lookup_block_group(fs_info, range->start);
7394
7395         while (cache) {
7396                 if (cache->key.objectid >= (range->start + range->len)) {
7397                         btrfs_put_block_group(cache);
7398                         break;
7399                 }
7400
7401                 start = max(range->start, cache->key.objectid);
7402                 end = min(range->start + range->len,
7403                                 cache->key.objectid + cache->key.offset);
7404
7405                 if (end - start >= range->minlen) {
7406                         if (!block_group_cache_done(cache)) {
7407                                 ret = cache_block_group(cache, NULL, root, 0);
7408                                 if (!ret)
7409                                         wait_block_group_cache_done(cache);
7410                         }
7411                         ret = btrfs_trim_block_group(cache,
7412                                                      &group_trimmed,
7413                                                      start,
7414                                                      end,
7415                                                      range->minlen);
7416
7417                         trimmed += group_trimmed;
7418                         if (ret) {
7419                                 btrfs_put_block_group(cache);
7420                                 break;
7421                         }
7422                 }
7423
7424                 cache = next_block_group(fs_info->tree_root, cache);
7425         }
7426
7427         range->len = trimmed;
7428         return ret;
7429 }