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 <linux/ratelimit.h>
  27 #include "compat.h"
  28 #include "hash.h"
  29 #include "ctree.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "transaction.h"
  33 #include "volumes.h"
  34 #include "locking.h"
  35 #include "free-space-cache.h"
  36
  37 /* control flags for do_chunk_alloc's force field
  38  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  39  * if we really need one.
  40  *
  41  * CHUNK_ALLOC_FORCE means it must try to allocate one
  42  *
  43  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  44  * if we have very few chunks already allocated.  This is
  45  * used as part of the clustering code to help make sure
  46  * we have a good pool of storage to cluster in, without
  47  * filling the FS with empty chunks
  48  *
  49  */
  50 enum {
  51         CHUNK_ALLOC_NO_FORCE = 0,
  52         CHUNK_ALLOC_FORCE = 1,
  53         CHUNK_ALLOC_LIMITED = 2,
  54 };
  55
  56 /*
  57  * Control how reservations are dealt with.
  58  *
  59  * RESERVE_FREE - freeing a reservation.
  60  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  61  *   ENOSPC accounting
  62  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  63  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  64  */
  65 enum {
  66         RESERVE_FREE = 0,
  67         RESERVE_ALLOC = 1,
  68         RESERVE_ALLOC_NO_ACCOUNT = 2,
  69 };
  70
  71 static int update_block_group(struct btrfs_trans_handle *trans,
  72                               struct btrfs_root *root,
  73                               u64 bytenr, u64 num_bytes, int alloc);
  74 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  75                                 struct btrfs_root *root,
  76                                 u64 bytenr, u64 num_bytes, u64 parent,
  77                                 u64 root_objectid, u64 owner_objectid,
  78                                 u64 owner_offset, int refs_to_drop,
  79                                 struct btrfs_delayed_extent_op *extra_op);
  80 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  81                                     struct extent_buffer *leaf,
  82                                     struct btrfs_extent_item *ei);
  83 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  84                                       struct btrfs_root *root,
  85                                       u64 parent, u64 root_objectid,
  86                                       u64 flags, u64 owner, u64 offset,
  87                                       struct btrfs_key *ins, int ref_mod);
  88 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  89                                      struct btrfs_root *root,
  90                                      u64 parent, u64 root_objectid,
  91                                      u64 flags, struct btrfs_disk_key *key,
  92                                      int level, struct btrfs_key *ins);
  93 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  94                           struct btrfs_root *extent_root, u64 alloc_bytes,
  95                           u64 flags, int force);
  96 static int find_next_key(struct btrfs_path *path, int level,
  97                          struct btrfs_key *key);
  98 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
  99                             int dump_block_groups);
 100 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 101                                        u64 num_bytes, int reserve);
 102
 103 static noinline int
 104 block_group_cache_done(struct btrfs_block_group_cache *cache)
 105 {
 106         smp_mb();
 107         return cache->cached == BTRFS_CACHE_FINISHED;
 108 }
 109
 110 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 111 {
 112         return (cache->flags & bits) == bits;
 113 }
 114
 115 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 116 {
 117         atomic_inc(&cache->count);
 118 }
 119
 120 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 121 {
 122         if (atomic_dec_and_test(&cache->count)) {
 123                 WARN_ON(cache->pinned > 0);
 124                 WARN_ON(cache->reserved > 0);
 125                 kfree(cache->free_space_ctl);
 126                 kfree(cache);
 127         }
 128 }
 129
 130 /*
 131  * this adds the block group to the fs_info rb tree for the block group
 132  * cache
 133  */
 134 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 135                                 struct btrfs_block_group_cache *block_group)
 136 {
 137         struct rb_node **p;
 138         struct rb_node *parent = NULL;
 139         struct btrfs_block_group_cache *cache;
 140
 141         spin_lock(&info->block_group_cache_lock);
 142         p = &info->block_group_cache_tree.rb_node;
 143
 144         while (*p) {
 145                 parent = *p;
 146                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 147                                  cache_node);
 148                 if (block_group->key.objectid < cache->key.objectid) {
 149                         p = &(*p)->rb_left;
 150                 } else if (block_group->key.objectid > cache->key.objectid) {
 151                         p = &(*p)->rb_right;
 152                 } else {
 153                         spin_unlock(&info->block_group_cache_lock);
 154                         return -EEXIST;
 155                 }
 156         }
 157
 158         rb_link_node(&block_group->cache_node, parent, p);
 159         rb_insert_color(&block_group->cache_node,
 160                         &info->block_group_cache_tree);
 161         spin_unlock(&info->block_group_cache_lock);
 162
 163         return 0;
 164 }
 165
 166 /*
 167  * This will return the block group at or after bytenr if contains is 0, else
 168  * it will return the block group that contains the bytenr
 169  */
 170 static struct btrfs_block_group_cache *
 171 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 172                               int contains)
 173 {
 174         struct btrfs_block_group_cache *cache, *ret = NULL;
 175         struct rb_node *n;
 176         u64 end, start;
 177
 178         spin_lock(&info->block_group_cache_lock);
 179         n = info->block_group_cache_tree.rb_node;
 180
 181         while (n) {
 182                 cache = rb_entry(n, struct btrfs_block_group_cache,
 183                                  cache_node);
 184                 end = cache->key.objectid + cache->key.offset - 1;
 185                 start = cache->key.objectid;
 186
 187                 if (bytenr < start) {
 188                         if (!contains && (!ret || start < ret->key.objectid))
 189                                 ret = cache;
 190                         n = n->rb_left;
 191                 } else if (bytenr > start) {
 192                         if (contains && bytenr <= end) {
 193                                 ret = cache;
 194                                 break;
 195                         }
 196                         n = n->rb_right;
 197                 } else {
 198                         ret = cache;
 199                         break;
 200                 }
 201         }
 202         if (ret)
 203                 btrfs_get_block_group(ret);
 204         spin_unlock(&info->block_group_cache_lock);
 205
 206         return ret;
 207 }
 208
 209 static int add_excluded_extent(struct btrfs_root *root,
 210                                u64 start, u64 num_bytes)
 211 {
 212         u64 end = start + num_bytes - 1;
 213         set_extent_bits(&root->fs_info->freed_extents[0],
 214                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 215         set_extent_bits(&root->fs_info->freed_extents[1],
 216                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 217         return 0;
 218 }
 219
 220 static void free_excluded_extents(struct btrfs_root *root,
 221                                   struct btrfs_block_group_cache *cache)
 222 {
 223         u64 start, end;
 224
 225         start = cache->key.objectid;
 226         end = start + cache->key.offset - 1;
 227
 228         clear_extent_bits(&root->fs_info->freed_extents[0],
 229                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 230         clear_extent_bits(&root->fs_info->freed_extents[1],
 231                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 232 }
 233
 234 static int exclude_super_stripes(struct btrfs_root *root,
 235                                  struct btrfs_block_group_cache *cache)
 236 {
 237         u64 bytenr;
 238         u64 *logical;
 239         int stripe_len;
 240         int i, nr, ret;
 241
 242         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 243                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 244                 cache->bytes_super += stripe_len;
 245                 ret = add_excluded_extent(root, cache->key.objectid,
 246                                           stripe_len);
 247                 BUG_ON(ret);
 248         }
 249
 250         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 251                 bytenr = btrfs_sb_offset(i);
 252                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 253                                        cache->key.objectid, bytenr,
 254                                        0, &logical, &nr, &stripe_len);
 255                 BUG_ON(ret);
 256
 257                 while (nr--) {
 258                         cache->bytes_super += stripe_len;
 259                         ret = add_excluded_extent(root, logical[nr],
 260                                                   stripe_len);
 261                         BUG_ON(ret);
 262                 }
 263
 264                 kfree(logical);
 265         }
 266         return 0;
 267 }
 268
 269 static struct btrfs_caching_control *
 270 get_caching_control(struct btrfs_block_group_cache *cache)
 271 {
 272         struct btrfs_caching_control *ctl;
 273
 274         spin_lock(&cache->lock);
 275         if (cache->cached != BTRFS_CACHE_STARTED) {
 276                 spin_unlock(&cache->lock);
 277                 return NULL;
 278         }
 279
 280         /* We're loading it the fast way, so we don't have a caching_ctl. */
 281         if (!cache->caching_ctl) {
 282                 spin_unlock(&cache->lock);
 283                 return NULL;
 284         }
 285
 286         ctl = cache->caching_ctl;
 287         atomic_inc(&ctl->count);
 288         spin_unlock(&cache->lock);
 289         return ctl;
 290 }
 291
 292 static void put_caching_control(struct btrfs_caching_control *ctl)
 293 {
 294         if (atomic_dec_and_test(&ctl->count))
 295                 kfree(ctl);
 296 }
 297
 298 /*
 299  * this is only called by cache_block_group, since we could have freed extents
 300  * we need to check the pinned_extents for any extents that can't be used yet
 301  * since their free space will be released as soon as the transaction commits.
 302  */
 303 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 304                               struct btrfs_fs_info *info, u64 start, u64 end)
 305 {
 306         u64 extent_start, extent_end, size, total_added = 0;
 307         int ret;
 308
 309         while (start < end) {
 310                 ret = find_first_extent_bit(info->pinned_extents, start,
 311                                             &extent_start, &extent_end,
 312                                             EXTENT_DIRTY | EXTENT_UPTODATE);
 313                 if (ret)
 314                         break;
 315
 316                 if (extent_start <= start) {
 317                         start = extent_end + 1;
 318                 } else if (extent_start > start && extent_start < end) {
 319                         size = extent_start - start;
 320                         total_added += size;
 321                         ret = btrfs_add_free_space(block_group, start,
 322                                                    size);
 323                         BUG_ON(ret);
 324                         start = extent_end + 1;
 325                 } else {
 326                         break;
 327                 }
 328         }
 329
 330         if (start < end) {
 331                 size = end - start;
 332                 total_added += size;
 333                 ret = btrfs_add_free_space(block_group, start, size);
 334                 BUG_ON(ret);
 335         }
 336
 337         return total_added;
 338 }
 339
 340 static noinline void caching_thread(struct btrfs_work *work)
 341 {
 342         struct btrfs_block_group_cache *block_group;
 343         struct btrfs_fs_info *fs_info;
 344         struct btrfs_caching_control *caching_ctl;
 345         struct btrfs_root *extent_root;
 346         struct btrfs_path *path;
 347         struct extent_buffer *leaf;
 348         struct btrfs_key key;
 349         u64 total_found = 0;
 350         u64 last = 0;
 351         u32 nritems;
 352         int ret = 0;
 353
 354         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 355         block_group = caching_ctl->block_group;
 356         fs_info = block_group->fs_info;
 357         extent_root = fs_info->extent_root;
 358
 359         path = btrfs_alloc_path();
 360         if (!path)
 361                 goto out;
 362
 363         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 364
 365         /*
 366          * We don't want to deadlock with somebody trying to allocate a new
 367          * extent for the extent root while also trying to search the extent
 368          * root to add free space.  So we skip locking and search the commit
 369          * root, since its read-only
 370          */
 371         path->skip_locking = 1;
 372         path->search_commit_root = 1;
 373         path->reada = 1;
 374
 375         key.objectid = last;
 376         key.offset = 0;
 377         key.type = BTRFS_EXTENT_ITEM_KEY;
 378 again:
 379         mutex_lock(&caching_ctl->mutex);
 380         /* need to make sure the commit_root doesn't disappear */
 381         down_read(&fs_info->extent_commit_sem);
 382
 383         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 384         if (ret < 0)
 385                 goto err;
 386
 387         leaf = path->nodes[0];
 388         nritems = btrfs_header_nritems(leaf);
 389
 390         while (1) {
 391                 if (btrfs_fs_closing(fs_info) > 1) {
 392                         last = (u64)-1;
 393                         break;
 394                 }
 395
 396                 if (path->slots[0] < nritems) {
 397                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 398                 } else {
 399                         ret = find_next_key(path, 0, &key);
 400                         if (ret)
 401                                 break;
 402
 403                         if (need_resched() ||
 404                             btrfs_next_leaf(extent_root, path)) {
 405                                 caching_ctl->progress = last;
 406                                 btrfs_release_path(path);
 407                                 up_read(&fs_info->extent_commit_sem);
 408                                 mutex_unlock(&caching_ctl->mutex);
 409                                 cond_resched();
 410                                 goto again;
 411                         }
 412                         leaf = path->nodes[0];
 413                         nritems = btrfs_header_nritems(leaf);
 414                         continue;
 415                 }
 416
 417                 if (key.objectid < block_group->key.objectid) {
 418                         path->slots[0]++;
 419                         continue;
 420                 }
 421
 422                 if (key.objectid >= block_group->key.objectid +
 423                     block_group->key.offset)
 424                         break;
 425
 426                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 427                         total_found += add_new_free_space(block_group,
 428                                                           fs_info, last,
 429                                                           key.objectid);
 430                         last = key.objectid + key.offset;
 431
 432                         if (total_found > (1024 * 1024 * 2)) {
 433                                 total_found = 0;
 434                                 wake_up(&caching_ctl->wait);
 435                         }
 436                 }
 437                 path->slots[0]++;
 438         }
 439         ret = 0;
 440
 441         total_found += add_new_free_space(block_group, fs_info, last,
 442                                           block_group->key.objectid +
 443                                           block_group->key.offset);
 444         caching_ctl->progress = (u64)-1;
 445
 446         spin_lock(&block_group->lock);
 447         block_group->caching_ctl = NULL;
 448         block_group->cached = BTRFS_CACHE_FINISHED;
 449         spin_unlock(&block_group->lock);
 450
 451 err:
 452         btrfs_free_path(path);
 453         up_read(&fs_info->extent_commit_sem);
 454
 455         free_excluded_extents(extent_root, block_group);
 456
 457         mutex_unlock(&caching_ctl->mutex);
 458 out:
 459         wake_up(&caching_ctl->wait);
 460
 461         put_caching_control(caching_ctl);
 462         btrfs_put_block_group(block_group);
 463 }
 464
 465 static int cache_block_group(struct btrfs_block_group_cache *cache,
 466                              struct btrfs_trans_handle *trans,
 467                              struct btrfs_root *root,
 468                              int load_cache_only)
 469 {
 470         struct btrfs_fs_info *fs_info = cache->fs_info;
 471         struct btrfs_caching_control *caching_ctl;
 472         int ret = 0;
 473
 474         smp_mb();
 475         if (cache->cached != BTRFS_CACHE_NO)
 476                 return 0;
 477
 478         /*
 479          * We can't do the read from on-disk cache during a commit since we need
 480          * to have the normal tree locking.  Also if we are currently trying to
 481          * allocate blocks for the tree root we can't do the fast caching since
 482          * we likely hold important locks.
 483          */
 484         if (trans && (!trans->transaction->in_commit) &&
 485             (root && root != root->fs_info->tree_root) &&
 486             btrfs_test_opt(root, SPACE_CACHE)) {
 487                 spin_lock(&cache->lock);
 488                 if (cache->cached != BTRFS_CACHE_NO) {
 489                         spin_unlock(&cache->lock);
 490                         return 0;
 491                 }
 492                 cache->cached = BTRFS_CACHE_STARTED;
 493                 spin_unlock(&cache->lock);
 494
 495                 ret = load_free_space_cache(fs_info, cache);
 496
 497                 spin_lock(&cache->lock);
 498                 if (ret == 1) {
 499                         cache->cached = BTRFS_CACHE_FINISHED;
 500                         cache->last_byte_to_unpin = (u64)-1;
 501                 } else {
 502                         cache->cached = BTRFS_CACHE_NO;
 503                 }
 504                 spin_unlock(&cache->lock);
 505                 if (ret == 1) {
 506                         free_excluded_extents(fs_info->extent_root, cache);
 507                         return 0;
 508                 }
 509         }
 510
 511         if (load_cache_only)
 512                 return 0;
 513
 514         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 515         BUG_ON(!caching_ctl);
 516
 517         INIT_LIST_HEAD(&caching_ctl->list);
 518         mutex_init(&caching_ctl->mutex);
 519         init_waitqueue_head(&caching_ctl->wait);
 520         caching_ctl->block_group = cache;
 521         caching_ctl->progress = cache->key.objectid;
 522         /* one for caching kthread, one for caching block group list */
 523         atomic_set(&caching_ctl->count, 2);
 524         caching_ctl->work.func = caching_thread;
 525
 526         spin_lock(&cache->lock);
 527         if (cache->cached != BTRFS_CACHE_NO) {
 528                 spin_unlock(&cache->lock);
 529                 kfree(caching_ctl);
 530                 return 0;
 531         }
 532         cache->caching_ctl = caching_ctl;
 533         cache->cached = BTRFS_CACHE_STARTED;
 534         spin_unlock(&cache->lock);
 535
 536         down_write(&fs_info->extent_commit_sem);
 537         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 538         up_write(&fs_info->extent_commit_sem);
 539
 540         btrfs_get_block_group(cache);
 541
 542         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 543
 544         return ret;
 545 }
 546
 547 /*
 548  * return the block group that starts at or after bytenr
 549  */
 550 static struct btrfs_block_group_cache *
 551 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 552 {
 553         struct btrfs_block_group_cache *cache;
 554
 555         cache = block_group_cache_tree_search(info, bytenr, 0);
 556
 557         return cache;
 558 }
 559
 560 /*
 561  * return the block group that contains the given bytenr
 562  */
 563 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 564                                                  struct btrfs_fs_info *info,
 565                                                  u64 bytenr)
 566 {
 567         struct btrfs_block_group_cache *cache;
 568
 569         cache = block_group_cache_tree_search(info, bytenr, 1);
 570
 571         return cache;
 572 }
 573
 574 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 575                                                   u64 flags)
 576 {
 577         struct list_head *head = &info->space_info;
 578         struct btrfs_space_info *found;
 579
 580         flags &= BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_SYSTEM |
 581                  BTRFS_BLOCK_GROUP_METADATA;
 582
 583         rcu_read_lock();
 584         list_for_each_entry_rcu(found, head, list) {
 585                 if (found->flags & flags) {
 586                         rcu_read_unlock();
 587                         return found;
 588                 }
 589         }
 590         rcu_read_unlock();
 591         return NULL;
 592 }
 593
 594 /*
 595  * after adding space to the filesystem, we need to clear the full flags
 596  * on all the space infos.
 597  */
 598 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 599 {
 600         struct list_head *head = &info->space_info;
 601         struct btrfs_space_info *found;
 602
 603         rcu_read_lock();
 604         list_for_each_entry_rcu(found, head, list)
 605                 found->full = 0;
 606         rcu_read_unlock();
 607 }
 608
 609 static u64 div_factor(u64 num, int factor)
 610 {
 611         if (factor == 10)
 612                 return num;
 613         num *= factor;
 614         do_div(num, 10);
 615         return num;
 616 }
 617
 618 static u64 div_factor_fine(u64 num, int factor)
 619 {
 620         if (factor == 100)
 621                 return num;
 622         num *= factor;
 623         do_div(num, 100);
 624         return num;
 625 }
 626
 627 u64 btrfs_find_block_group(struct btrfs_root *root,
 628                            u64 search_start, u64 search_hint, int owner)
 629 {
 630         struct btrfs_block_group_cache *cache;
 631         u64 used;
 632         u64 last = max(search_hint, search_start);
 633         u64 group_start = 0;
 634         int full_search = 0;
 635         int factor = 9;
 636         int wrapped = 0;
 637 again:
 638         while (1) {
 639                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
 640                 if (!cache)
 641                         break;
 642
 643                 spin_lock(&cache->lock);
 644                 last = cache->key.objectid + cache->key.offset;
 645                 used = btrfs_block_group_used(&cache->item);
 646
 647                 if ((full_search || !cache->ro) &&
 648                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
 649                         if (used + cache->pinned + cache->reserved <
 650                             div_factor(cache->key.offset, factor)) {
 651                                 group_start = cache->key.objectid;
 652                                 spin_unlock(&cache->lock);
 653                                 btrfs_put_block_group(cache);
 654                                 goto found;
 655                         }
 656                 }
 657                 spin_unlock(&cache->lock);
 658                 btrfs_put_block_group(cache);
 659                 cond_resched();
 660         }
 661         if (!wrapped) {
 662                 last = search_start;
 663                 wrapped = 1;
 664                 goto again;
 665         }
 666         if (!full_search && factor < 10) {
 667                 last = search_start;
 668                 full_search = 1;
 669                 factor = 10;
 670                 goto again;
 671         }
 672 found:
 673         return group_start;
 674 }
 675
 676 /* simple helper to search for an existing extent at a given offset */
 677 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 678 {
 679         int ret;
 680         struct btrfs_key key;
 681         struct btrfs_path *path;
 682
 683         path = btrfs_alloc_path();
 684         if (!path)
 685                 return -ENOMEM;
 686
 687         key.objectid = start;
 688         key.offset = len;
 689         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 690         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 691                                 0, 0);
 692         btrfs_free_path(path);
 693         return ret;
 694 }
 695
 696 /*
 697  * helper function to lookup reference count and flags of extent.
 698  *
 699  * the head node for delayed ref is used to store the sum of all the
 700  * reference count modifications queued up in the rbtree. the head
 701  * node may also store the extent flags to set. This way you can check
 702  * to see what the reference count and extent flags would be if all of
 703  * the delayed refs are not processed.
 704  */
 705 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 706                              struct btrfs_root *root, u64 bytenr,
 707                              u64 num_bytes, u64 *refs, u64 *flags)
 708 {
 709         struct btrfs_delayed_ref_head *head;
 710         struct btrfs_delayed_ref_root *delayed_refs;
 711         struct btrfs_path *path;
 712         struct btrfs_extent_item *ei;
 713         struct extent_buffer *leaf;
 714         struct btrfs_key key;
 715         u32 item_size;
 716         u64 num_refs;
 717         u64 extent_flags;
 718         int ret;
 719
 720         path = btrfs_alloc_path();
 721         if (!path)
 722                 return -ENOMEM;
 723
 724         key.objectid = bytenr;
 725         key.type = BTRFS_EXTENT_ITEM_KEY;
 726         key.offset = num_bytes;
 727         if (!trans) {
 728                 path->skip_locking = 1;
 729                 path->search_commit_root = 1;
 730         }
 731 again:
 732         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 733                                 &key, path, 0, 0);
 734         if (ret < 0)
 735                 goto out_free;
 736
 737         if (ret == 0) {
 738                 leaf = path->nodes[0];
 739                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 740                 if (item_size >= sizeof(*ei)) {
 741                         ei = btrfs_item_ptr(leaf, path->slots[0],
 742                                             struct btrfs_extent_item);
 743                         num_refs = btrfs_extent_refs(leaf, ei);
 744                         extent_flags = btrfs_extent_flags(leaf, ei);
 745                 } else {
 746 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 747                         struct btrfs_extent_item_v0 *ei0;
 748                         BUG_ON(item_size != sizeof(*ei0));
 749                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 750                                              struct btrfs_extent_item_v0);
 751                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 752                         /* FIXME: this isn't correct for data */
 753                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 754 #else
 755                         BUG();
 756 #endif
 757                 }
 758                 BUG_ON(num_refs == 0);
 759         } else {
 760                 num_refs = 0;
 761                 extent_flags = 0;
 762                 ret = 0;
 763         }
 764
 765         if (!trans)
 766                 goto out;
 767
 768         delayed_refs = &trans->transaction->delayed_refs;
 769         spin_lock(&delayed_refs->lock);
 770         head = btrfs_find_delayed_ref_head(trans, bytenr);
 771         if (head) {
 772                 if (!mutex_trylock(&head->mutex)) {
 773                         atomic_inc(&head->node.refs);
 774                         spin_unlock(&delayed_refs->lock);
 775
 776                         btrfs_release_path(path);
 777
 778                         /*
 779                          * Mutex was contended, block until it's released and try
 780                          * again
 781                          */
 782                         mutex_lock(&head->mutex);
 783                         mutex_unlock(&head->mutex);
 784                         btrfs_put_delayed_ref(&head->node);
 785                         goto again;
 786                 }
 787                 if (head->extent_op && head->extent_op->update_flags)
 788                         extent_flags |= head->extent_op->flags_to_set;
 789                 else
 790                         BUG_ON(num_refs == 0);
 791
 792                 num_refs += head->node.ref_mod;
 793                 mutex_unlock(&head->mutex);
 794         }
 795         spin_unlock(&delayed_refs->lock);
 796 out:
 797         WARN_ON(num_refs == 0);
 798         if (refs)
 799                 *refs = num_refs;
 800         if (flags)
 801                 *flags = extent_flags;
 802 out_free:
 803         btrfs_free_path(path);
 804         return ret;
 805 }
 806
 807 /*
 808  * Back reference rules.  Back refs have three main goals:
 809  *
 810  * 1) differentiate between all holders of references to an extent so that
 811  *    when a reference is dropped we can make sure it was a valid reference
 812  *    before freeing the extent.
 813  *
 814  * 2) Provide enough information to quickly find the holders of an extent
 815  *    if we notice a given block is corrupted or bad.
 816  *
 817  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 818  *    maintenance.  This is actually the same as #2, but with a slightly
 819  *    different use case.
 820  *
 821  * There are two kinds of back refs. The implicit back refs is optimized
 822  * for pointers in non-shared tree blocks. For a given pointer in a block,
 823  * back refs of this kind provide information about the block's owner tree
 824  * and the pointer's key. These information allow us to find the block by
 825  * b-tree searching. The full back refs is for pointers in tree blocks not
 826  * referenced by their owner trees. The location of tree block is recorded
 827  * in the back refs. Actually the full back refs is generic, and can be
 828  * used in all cases the implicit back refs is used. The major shortcoming
 829  * of the full back refs is its overhead. Every time a tree block gets
 830  * COWed, we have to update back refs entry for all pointers in it.
 831  *
 832  * For a newly allocated tree block, we use implicit back refs for
 833  * pointers in it. This means most tree related operations only involve
 834  * implicit back refs. For a tree block created in old transaction, the
 835  * only way to drop a reference to it is COW it. So we can detect the
 836  * event that tree block loses its owner tree's reference and do the
 837  * back refs conversion.
 838  *
 839  * When a tree block is COW'd through a tree, there are four cases:
 840  *
 841  * The reference count of the block is one and the tree is the block's
 842  * owner tree. Nothing to do in this case.
 843  *
 844  * The reference count of the block is one and the tree is not the
 845  * block's owner tree. In this case, full back refs is used for pointers
 846  * in the block. Remove these full back refs, add implicit back refs for
 847  * every pointers in the new block.
 848  *
 849  * The reference count of the block is greater than one and the tree is
 850  * the block's owner tree. In this case, implicit back refs is used for
 851  * pointers in the block. Add full back refs for every pointers in the
 852  * block, increase lower level extents' reference counts. The original
 853  * implicit back refs are entailed to the new block.
 854  *
 855  * The reference count of the block is greater than one and the tree is
 856  * not the block's owner tree. Add implicit back refs for every pointer in
 857  * the new block, increase lower level extents' reference count.
 858  *
 859  * Back Reference Key composing:
 860  *
 861  * The key objectid corresponds to the first byte in the extent,
 862  * The key type is used to differentiate between types of back refs.
 863  * There are different meanings of the key offset for different types
 864  * of back refs.
 865  *
 866  * File extents can be referenced by:
 867  *
 868  * - multiple snapshots, subvolumes, or different generations in one subvol
 869  * - different files inside a single subvolume
 870  * - different offsets inside a file (bookend extents in file.c)
 871  *
 872  * The extent ref structure for the implicit back refs has fields for:
 873  *
 874  * - Objectid of the subvolume root
 875  * - objectid of the file holding the reference
 876  * - original offset in the file
 877  * - how many bookend extents
 878  *
 879  * The key offset for the implicit back refs is hash of the first
 880  * three fields.
 881  *
 882  * The extent ref structure for the full back refs has field for:
 883  *
 884  * - number of pointers in the tree leaf
 885  *
 886  * The key offset for the implicit back refs is the first byte of
 887  * the tree leaf
 888  *
 889  * When a file extent is allocated, The implicit back refs is used.
 890  * the fields are filled in:
 891  *
 892  *     (root_key.objectid, inode objectid, offset in file, 1)
 893  *
 894  * When a file extent is removed file truncation, we find the
 895  * corresponding implicit back refs and check the following fields:
 896  *
 897  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 898  *
 899  * Btree extents can be referenced by:
 900  *
 901  * - Different subvolumes
 902  *
 903  * Both the implicit back refs and the full back refs for tree blocks
 904  * only consist of key. The key offset for the implicit back refs is
 905  * objectid of block's owner tree. The key offset for the full back refs
 906  * is the first byte of parent block.
 907  *
 908  * When implicit back refs is used, information about the lowest key and
 909  * level of the tree block are required. These information are stored in
 910  * tree block info structure.
 911  */
 912
 913 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 914 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 915                                   struct btrfs_root *root,
 916                                   struct btrfs_path *path,
 917                                   u64 owner, u32 extra_size)
 918 {
 919         struct btrfs_extent_item *item;
 920         struct btrfs_extent_item_v0 *ei0;
 921         struct btrfs_extent_ref_v0 *ref0;
 922         struct btrfs_tree_block_info *bi;
 923         struct extent_buffer *leaf;
 924         struct btrfs_key key;
 925         struct btrfs_key found_key;
 926         u32 new_size = sizeof(*item);
 927         u64 refs;
 928         int ret;
 929
 930         leaf = path->nodes[0];
 931         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 932
 933         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 934         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 935                              struct btrfs_extent_item_v0);
 936         refs = btrfs_extent_refs_v0(leaf, ei0);
 937
 938         if (owner == (u64)-1) {
 939                 while (1) {
 940                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 941                                 ret = btrfs_next_leaf(root, path);
 942                                 if (ret < 0)
 943                                         return ret;
 944                                 BUG_ON(ret > 0);
 945                                 leaf = path->nodes[0];
 946                         }
 947                         btrfs_item_key_to_cpu(leaf, &found_key,
 948                                               path->slots[0]);
 949                         BUG_ON(key.objectid != found_key.objectid);
 950                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 951                                 path->slots[0]++;
 952                                 continue;
 953                         }
 954                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 955                                               struct btrfs_extent_ref_v0);
 956                         owner = btrfs_ref_objectid_v0(leaf, ref0);
 957                         break;
 958                 }
 959         }
 960         btrfs_release_path(path);
 961
 962         if (owner < BTRFS_FIRST_FREE_OBJECTID)
 963                 new_size += sizeof(*bi);
 964
 965         new_size -= sizeof(*ei0);
 966         ret = btrfs_search_slot(trans, root, &key, path,
 967                                 new_size + extra_size, 1);
 968         if (ret < 0)
 969                 return ret;
 970         BUG_ON(ret);
 971
 972         ret = btrfs_extend_item(trans, root, path, new_size);
 973
 974         leaf = path->nodes[0];
 975         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 976         btrfs_set_extent_refs(leaf, item, refs);
 977         /* FIXME: get real generation */
 978         btrfs_set_extent_generation(leaf, item, 0);
 979         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 980                 btrfs_set_extent_flags(leaf, item,
 981                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
 982                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
 983                 bi = (struct btrfs_tree_block_info *)(item + 1);
 984                 /* FIXME: get first key of the block */
 985                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
 986                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
 987         } else {
 988                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
 989         }
 990         btrfs_mark_buffer_dirty(leaf);
 991         return 0;
 992 }
 993 #endif
 994
 995 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
 996 {
 997         u32 high_crc = ~(u32)0;
 998         u32 low_crc = ~(u32)0;
 999         __le64 lenum;
1000
1001         lenum = cpu_to_le64(root_objectid);
1002         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1003         lenum = cpu_to_le64(owner);
1004         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1005         lenum = cpu_to_le64(offset);
1006         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1007
1008         return ((u64)high_crc << 31) ^ (u64)low_crc;
1009 }
1010
1011 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1012                                      struct btrfs_extent_data_ref *ref)
1013 {
1014         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1015                                     btrfs_extent_data_ref_objectid(leaf, ref),
1016                                     btrfs_extent_data_ref_offset(leaf, ref));
1017 }
1018
1019 static int match_extent_data_ref(struct extent_buffer *leaf,
1020                                  struct btrfs_extent_data_ref *ref,
1021                                  u64 root_objectid, u64 owner, u64 offset)
1022 {
1023         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1024             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1025             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1026                 return 0;
1027         return 1;
1028 }
1029
1030 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1031                                            struct btrfs_root *root,
1032                                            struct btrfs_path *path,
1033                                            u64 bytenr, u64 parent,
1034                                            u64 root_objectid,
1035                                            u64 owner, u64 offset)
1036 {
1037         struct btrfs_key key;
1038         struct btrfs_extent_data_ref *ref;
1039         struct extent_buffer *leaf;
1040         u32 nritems;
1041         int ret;
1042         int recow;
1043         int err = -ENOENT;
1044
1045         key.objectid = bytenr;
1046         if (parent) {
1047                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1048                 key.offset = parent;
1049         } else {
1050                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1051                 key.offset = hash_extent_data_ref(root_objectid,
1052                                                   owner, offset);
1053         }
1054 again:
1055         recow = 0;
1056         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1057         if (ret < 0) {
1058                 err = ret;
1059                 goto fail;
1060         }
1061
1062         if (parent) {
1063                 if (!ret)
1064                         return 0;
1065 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1066                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1067                 btrfs_release_path(path);
1068                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1069                 if (ret < 0) {
1070                         err = ret;
1071                         goto fail;
1072                 }
1073                 if (!ret)
1074                         return 0;
1075 #endif
1076                 goto fail;
1077         }
1078
1079         leaf = path->nodes[0];
1080         nritems = btrfs_header_nritems(leaf);
1081         while (1) {
1082                 if (path->slots[0] >= nritems) {
1083                         ret = btrfs_next_leaf(root, path);
1084                         if (ret < 0)
1085                                 err = ret;
1086                         if (ret)
1087                                 goto fail;
1088
1089                         leaf = path->nodes[0];
1090                         nritems = btrfs_header_nritems(leaf);
1091                         recow = 1;
1092                 }
1093
1094                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1095                 if (key.objectid != bytenr ||
1096                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1097                         goto fail;
1098
1099                 ref = btrfs_item_ptr(leaf, path->slots[0],
1100                                      struct btrfs_extent_data_ref);
1101
1102                 if (match_extent_data_ref(leaf, ref, root_objectid,
1103                                           owner, offset)) {
1104                         if (recow) {
1105                                 btrfs_release_path(path);
1106                                 goto again;
1107                         }
1108                         err = 0;
1109                         break;
1110                 }
1111                 path->slots[0]++;
1112         }
1113 fail:
1114         return err;
1115 }
1116
1117 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1118                                            struct btrfs_root *root,
1119                                            struct btrfs_path *path,
1120                                            u64 bytenr, u64 parent,
1121                                            u64 root_objectid, u64 owner,
1122                                            u64 offset, int refs_to_add)
1123 {
1124         struct btrfs_key key;
1125         struct extent_buffer *leaf;
1126         u32 size;
1127         u32 num_refs;
1128         int ret;
1129
1130         key.objectid = bytenr;
1131         if (parent) {
1132                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1133                 key.offset = parent;
1134                 size = sizeof(struct btrfs_shared_data_ref);
1135         } else {
1136                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1137                 key.offset = hash_extent_data_ref(root_objectid,
1138                                                   owner, offset);
1139                 size = sizeof(struct btrfs_extent_data_ref);
1140         }
1141
1142         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1143         if (ret && ret != -EEXIST)
1144                 goto fail;
1145
1146         leaf = path->nodes[0];
1147         if (parent) {
1148                 struct btrfs_shared_data_ref *ref;
1149                 ref = btrfs_item_ptr(leaf, path->slots[0],
1150                                      struct btrfs_shared_data_ref);
1151                 if (ret == 0) {
1152                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1153                 } else {
1154                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1155                         num_refs += refs_to_add;
1156                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1157                 }
1158         } else {
1159                 struct btrfs_extent_data_ref *ref;
1160                 while (ret == -EEXIST) {
1161                         ref = btrfs_item_ptr(leaf, path->slots[0],
1162                                              struct btrfs_extent_data_ref);
1163                         if (match_extent_data_ref(leaf, ref, root_objectid,
1164                                                   owner, offset))
1165                                 break;
1166                         btrfs_release_path(path);
1167                         key.offset++;
1168                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1169                                                       size);
1170                         if (ret && ret != -EEXIST)
1171                                 goto fail;
1172
1173                         leaf = path->nodes[0];
1174                 }
1175                 ref = btrfs_item_ptr(leaf, path->slots[0],
1176                                      struct btrfs_extent_data_ref);
1177                 if (ret == 0) {
1178                         btrfs_set_extent_data_ref_root(leaf, ref,
1179                                                        root_objectid);
1180                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1181                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1182                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1183                 } else {
1184                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1185                         num_refs += refs_to_add;
1186                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1187                 }
1188         }
1189         btrfs_mark_buffer_dirty(leaf);
1190         ret = 0;
1191 fail:
1192         btrfs_release_path(path);
1193         return ret;
1194 }
1195
1196 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1197                                            struct btrfs_root *root,
1198                                            struct btrfs_path *path,
1199                                            int refs_to_drop)
1200 {
1201         struct btrfs_key key;
1202         struct btrfs_extent_data_ref *ref1 = NULL;
1203         struct btrfs_shared_data_ref *ref2 = NULL;
1204         struct extent_buffer *leaf;
1205         u32 num_refs = 0;
1206         int ret = 0;
1207
1208         leaf = path->nodes[0];
1209         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1210
1211         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1212                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1213                                       struct btrfs_extent_data_ref);
1214                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1215         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1216                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1217                                       struct btrfs_shared_data_ref);
1218                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1219 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1220         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1221                 struct btrfs_extent_ref_v0 *ref0;
1222                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1223                                       struct btrfs_extent_ref_v0);
1224                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1225 #endif
1226         } else {
1227                 BUG();
1228         }
1229
1230         BUG_ON(num_refs < refs_to_drop);
1231         num_refs -= refs_to_drop;
1232
1233         if (num_refs == 0) {
1234                 ret = btrfs_del_item(trans, root, path);
1235         } else {
1236                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1237                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1238                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1239                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1240 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1241                 else {
1242                         struct btrfs_extent_ref_v0 *ref0;
1243                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1244                                         struct btrfs_extent_ref_v0);
1245                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1246                 }
1247 #endif
1248                 btrfs_mark_buffer_dirty(leaf);
1249         }
1250         return ret;
1251 }
1252
1253 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1254                                           struct btrfs_path *path,
1255                                           struct btrfs_extent_inline_ref *iref)
1256 {
1257         struct btrfs_key key;
1258         struct extent_buffer *leaf;
1259         struct btrfs_extent_data_ref *ref1;
1260         struct btrfs_shared_data_ref *ref2;
1261         u32 num_refs = 0;
1262
1263         leaf = path->nodes[0];
1264         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1265         if (iref) {
1266                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1267                     BTRFS_EXTENT_DATA_REF_KEY) {
1268                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1269                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1270                 } else {
1271                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1272                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1273                 }
1274         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1275                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1276                                       struct btrfs_extent_data_ref);
1277                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1278         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1279                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1280                                       struct btrfs_shared_data_ref);
1281                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1282 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1283         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1284                 struct btrfs_extent_ref_v0 *ref0;
1285                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1286                                       struct btrfs_extent_ref_v0);
1287                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1288 #endif
1289         } else {
1290                 WARN_ON(1);
1291         }
1292         return num_refs;
1293 }
1294
1295 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1296                                           struct btrfs_root *root,
1297                                           struct btrfs_path *path,
1298                                           u64 bytenr, u64 parent,
1299                                           u64 root_objectid)
1300 {
1301         struct btrfs_key key;
1302         int ret;
1303
1304         key.objectid = bytenr;
1305         if (parent) {
1306                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1307                 key.offset = parent;
1308         } else {
1309                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1310                 key.offset = root_objectid;
1311         }
1312
1313         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1314         if (ret > 0)
1315                 ret = -ENOENT;
1316 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1317         if (ret == -ENOENT && parent) {
1318                 btrfs_release_path(path);
1319                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1320                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1321                 if (ret > 0)
1322                         ret = -ENOENT;
1323         }
1324 #endif
1325         return ret;
1326 }
1327
1328 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1329                                           struct btrfs_root *root,
1330                                           struct btrfs_path *path,
1331                                           u64 bytenr, u64 parent,
1332                                           u64 root_objectid)
1333 {
1334         struct btrfs_key key;
1335         int ret;
1336
1337         key.objectid = bytenr;
1338         if (parent) {
1339                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1340                 key.offset = parent;
1341         } else {
1342                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1343                 key.offset = root_objectid;
1344         }
1345
1346         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1347         btrfs_release_path(path);
1348         return ret;
1349 }
1350
1351 static inline int extent_ref_type(u64 parent, u64 owner)
1352 {
1353         int type;
1354         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1355                 if (parent > 0)
1356                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1357                 else
1358                         type = BTRFS_TREE_BLOCK_REF_KEY;
1359         } else {
1360                 if (parent > 0)
1361                         type = BTRFS_SHARED_DATA_REF_KEY;
1362                 else
1363                         type = BTRFS_EXTENT_DATA_REF_KEY;
1364         }
1365         return type;
1366 }
1367
1368 static int find_next_key(struct btrfs_path *path, int level,
1369                          struct btrfs_key *key)
1370
1371 {
1372         for (; level < BTRFS_MAX_LEVEL; level++) {
1373                 if (!path->nodes[level])
1374                         break;
1375                 if (path->slots[level] + 1 >=
1376                     btrfs_header_nritems(path->nodes[level]))
1377                         continue;
1378                 if (level == 0)
1379                         btrfs_item_key_to_cpu(path->nodes[level], key,
1380                                               path->slots[level] + 1);
1381                 else
1382                         btrfs_node_key_to_cpu(path->nodes[level], key,
1383                                               path->slots[level] + 1);
1384                 return 0;
1385         }
1386         return 1;
1387 }
1388
1389 /*
1390  * look for inline back ref. if back ref is found, *ref_ret is set
1391  * to the address of inline back ref, and 0 is returned.
1392  *
1393  * if back ref isn't found, *ref_ret is set to the address where it
1394  * should be inserted, and -ENOENT is returned.
1395  *
1396  * if insert is true and there are too many inline back refs, the path
1397  * points to the extent item, and -EAGAIN is returned.
1398  *
1399  * NOTE: inline back refs are ordered in the same way that back ref
1400  *       items in the tree are ordered.
1401  */
1402 static noinline_for_stack
1403 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1404                                  struct btrfs_root *root,
1405                                  struct btrfs_path *path,
1406                                  struct btrfs_extent_inline_ref **ref_ret,
1407                                  u64 bytenr, u64 num_bytes,
1408                                  u64 parent, u64 root_objectid,
1409                                  u64 owner, u64 offset, int insert)
1410 {
1411         struct btrfs_key key;
1412         struct extent_buffer *leaf;
1413         struct btrfs_extent_item *ei;
1414         struct btrfs_extent_inline_ref *iref;
1415         u64 flags;
1416         u64 item_size;
1417         unsigned long ptr;
1418         unsigned long end;
1419         int extra_size;
1420         int type;
1421         int want;
1422         int ret;
1423         int err = 0;
1424
1425         key.objectid = bytenr;
1426         key.type = BTRFS_EXTENT_ITEM_KEY;
1427         key.offset = num_bytes;
1428
1429         want = extent_ref_type(parent, owner);
1430         if (insert) {
1431                 extra_size = btrfs_extent_inline_ref_size(want);
1432                 path->keep_locks = 1;
1433         } else
1434                 extra_size = -1;
1435         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1436         if (ret < 0) {
1437                 err = ret;
1438                 goto out;
1439         }
1440         BUG_ON(ret);
1441
1442         leaf = path->nodes[0];
1443         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1444 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1445         if (item_size < sizeof(*ei)) {
1446                 if (!insert) {
1447                         err = -ENOENT;
1448                         goto out;
1449                 }
1450                 ret = convert_extent_item_v0(trans, root, path, owner,
1451                                              extra_size);
1452                 if (ret < 0) {
1453                         err = ret;
1454                         goto out;
1455                 }
1456                 leaf = path->nodes[0];
1457                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1458         }
1459 #endif
1460         BUG_ON(item_size < sizeof(*ei));
1461
1462         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1463         flags = btrfs_extent_flags(leaf, ei);
1464
1465         ptr = (unsigned long)(ei + 1);
1466         end = (unsigned long)ei + item_size;
1467
1468         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1469                 ptr += sizeof(struct btrfs_tree_block_info);
1470                 BUG_ON(ptr > end);
1471         } else {
1472                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1473         }
1474
1475         err = -ENOENT;
1476         while (1) {
1477                 if (ptr >= end) {
1478                         WARN_ON(ptr > end);
1479                         break;
1480                 }
1481                 iref = (struct btrfs_extent_inline_ref *)ptr;
1482                 type = btrfs_extent_inline_ref_type(leaf, iref);
1483                 if (want < type)
1484                         break;
1485                 if (want > type) {
1486                         ptr += btrfs_extent_inline_ref_size(type);
1487                         continue;
1488                 }
1489
1490                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1491                         struct btrfs_extent_data_ref *dref;
1492                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1493                         if (match_extent_data_ref(leaf, dref, root_objectid,
1494                                                   owner, offset)) {
1495                                 err = 0;
1496                                 break;
1497                         }
1498                         if (hash_extent_data_ref_item(leaf, dref) <
1499                             hash_extent_data_ref(root_objectid, owner, offset))
1500                                 break;
1501                 } else {
1502                         u64 ref_offset;
1503                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1504                         if (parent > 0) {
1505                                 if (parent == ref_offset) {
1506                                         err = 0;
1507                                         break;
1508                                 }
1509                                 if (ref_offset < parent)
1510                                         break;
1511                         } else {
1512                                 if (root_objectid == ref_offset) {
1513                                         err = 0;
1514                                         break;
1515                                 }
1516                                 if (ref_offset < root_objectid)
1517                                         break;
1518                         }
1519                 }
1520                 ptr += btrfs_extent_inline_ref_size(type);
1521         }
1522         if (err == -ENOENT && insert) {
1523                 if (item_size + extra_size >=
1524                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1525                         err = -EAGAIN;
1526                         goto out;
1527                 }
1528                 /*
1529                  * To add new inline back ref, we have to make sure
1530                  * there is no corresponding back ref item.
1531                  * For simplicity, we just do not add new inline back
1532                  * ref if there is any kind of item for this block
1533                  */
1534                 if (find_next_key(path, 0, &key) == 0 &&
1535                     key.objectid == bytenr &&
1536                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1537                         err = -EAGAIN;
1538                         goto out;
1539                 }
1540         }
1541         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1542 out:
1543         if (insert) {
1544                 path->keep_locks = 0;
1545                 btrfs_unlock_up_safe(path, 1);
1546         }
1547         return err;
1548 }
1549
1550 /*
1551  * helper to add new inline back ref
1552  */
1553 static noinline_for_stack
1554 int setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1555                                 struct btrfs_root *root,
1556                                 struct btrfs_path *path,
1557                                 struct btrfs_extent_inline_ref *iref,
1558                                 u64 parent, u64 root_objectid,
1559                                 u64 owner, u64 offset, int refs_to_add,
1560                                 struct btrfs_delayed_extent_op *extent_op)
1561 {
1562         struct extent_buffer *leaf;
1563         struct btrfs_extent_item *ei;
1564         unsigned long ptr;
1565         unsigned long end;
1566         unsigned long item_offset;
1567         u64 refs;
1568         int size;
1569         int type;
1570         int ret;
1571
1572         leaf = path->nodes[0];
1573         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1574         item_offset = (unsigned long)iref - (unsigned long)ei;
1575
1576         type = extent_ref_type(parent, owner);
1577         size = btrfs_extent_inline_ref_size(type);
1578
1579         ret = btrfs_extend_item(trans, root, path, size);
1580
1581         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1582         refs = btrfs_extent_refs(leaf, ei);
1583         refs += refs_to_add;
1584         btrfs_set_extent_refs(leaf, ei, refs);
1585         if (extent_op)
1586                 __run_delayed_extent_op(extent_op, leaf, ei);
1587
1588         ptr = (unsigned long)ei + item_offset;
1589         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1590         if (ptr < end - size)
1591                 memmove_extent_buffer(leaf, ptr + size, ptr,
1592                                       end - size - ptr);
1593
1594         iref = (struct btrfs_extent_inline_ref *)ptr;
1595         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1596         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1597                 struct btrfs_extent_data_ref *dref;
1598                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1599                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1600                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1601                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1602                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1603         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1604                 struct btrfs_shared_data_ref *sref;
1605                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1606                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1607                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1608         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1609                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1610         } else {
1611                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1612         }
1613         btrfs_mark_buffer_dirty(leaf);
1614         return 0;
1615 }
1616
1617 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1618                                  struct btrfs_root *root,
1619                                  struct btrfs_path *path,
1620                                  struct btrfs_extent_inline_ref **ref_ret,
1621                                  u64 bytenr, u64 num_bytes, u64 parent,
1622                                  u64 root_objectid, u64 owner, u64 offset)
1623 {
1624         int ret;
1625
1626         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1627                                            bytenr, num_bytes, parent,
1628                                            root_objectid, owner, offset, 0);
1629         if (ret != -ENOENT)
1630                 return ret;
1631
1632         btrfs_release_path(path);
1633         *ref_ret = NULL;
1634
1635         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1636                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1637                                             root_objectid);
1638         } else {
1639                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1640                                              root_objectid, owner, offset);
1641         }
1642         return ret;
1643 }
1644
1645 /*
1646  * helper to update/remove inline back ref
1647  */
1648 static noinline_for_stack
1649 int update_inline_extent_backref(struct btrfs_trans_handle *trans,
1650                                  struct btrfs_root *root,
1651                                  struct btrfs_path *path,
1652                                  struct btrfs_extent_inline_ref *iref,
1653                                  int refs_to_mod,
1654                                  struct btrfs_delayed_extent_op *extent_op)
1655 {
1656         struct extent_buffer *leaf;
1657         struct btrfs_extent_item *ei;
1658         struct btrfs_extent_data_ref *dref = NULL;
1659         struct btrfs_shared_data_ref *sref = NULL;
1660         unsigned long ptr;
1661         unsigned long end;
1662         u32 item_size;
1663         int size;
1664         int type;
1665         int ret;
1666         u64 refs;
1667
1668         leaf = path->nodes[0];
1669         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1670         refs = btrfs_extent_refs(leaf, ei);
1671         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1672         refs += refs_to_mod;
1673         btrfs_set_extent_refs(leaf, ei, refs);
1674         if (extent_op)
1675                 __run_delayed_extent_op(extent_op, leaf, ei);
1676
1677         type = btrfs_extent_inline_ref_type(leaf, iref);
1678
1679         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1680                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1681                 refs = btrfs_extent_data_ref_count(leaf, dref);
1682         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1683                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1684                 refs = btrfs_shared_data_ref_count(leaf, sref);
1685         } else {
1686                 refs = 1;
1687                 BUG_ON(refs_to_mod != -1);
1688         }
1689
1690         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1691         refs += refs_to_mod;
1692
1693         if (refs > 0) {
1694                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1695                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1696                 else
1697                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1698         } else {
1699                 size =  btrfs_extent_inline_ref_size(type);
1700                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1701                 ptr = (unsigned long)iref;
1702                 end = (unsigned long)ei + item_size;
1703                 if (ptr + size < end)
1704                         memmove_extent_buffer(leaf, ptr, ptr + size,
1705                                               end - ptr - size);
1706                 item_size -= size;
1707                 ret = btrfs_truncate_item(trans, root, path, item_size, 1);
1708         }
1709         btrfs_mark_buffer_dirty(leaf);
1710         return 0;
1711 }
1712
1713 static noinline_for_stack
1714 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1715                                  struct btrfs_root *root,
1716                                  struct btrfs_path *path,
1717                                  u64 bytenr, u64 num_bytes, u64 parent,
1718                                  u64 root_objectid, u64 owner,
1719                                  u64 offset, int refs_to_add,
1720                                  struct btrfs_delayed_extent_op *extent_op)
1721 {
1722         struct btrfs_extent_inline_ref *iref;
1723         int ret;
1724
1725         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1726                                            bytenr, num_bytes, parent,
1727                                            root_objectid, owner, offset, 1);
1728         if (ret == 0) {
1729                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1730                 ret = update_inline_extent_backref(trans, root, path, iref,
1731                                                    refs_to_add, extent_op);
1732         } else if (ret == -ENOENT) {
1733                 ret = setup_inline_extent_backref(trans, root, path, iref,
1734                                                   parent, root_objectid,
1735                                                   owner, offset, refs_to_add,
1736                                                   extent_op);
1737         }
1738         return ret;
1739 }
1740
1741 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1742                                  struct btrfs_root *root,
1743                                  struct btrfs_path *path,
1744                                  u64 bytenr, u64 parent, u64 root_objectid,
1745                                  u64 owner, u64 offset, int refs_to_add)
1746 {
1747         int ret;
1748         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1749                 BUG_ON(refs_to_add != 1);
1750                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1751                                             parent, root_objectid);
1752         } else {
1753                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1754                                              parent, root_objectid,
1755                                              owner, offset, refs_to_add);
1756         }
1757         return ret;
1758 }
1759
1760 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1761                                  struct btrfs_root *root,
1762                                  struct btrfs_path *path,
1763                                  struct btrfs_extent_inline_ref *iref,
1764                                  int refs_to_drop, int is_data)
1765 {
1766         int ret;
1767
1768         BUG_ON(!is_data && refs_to_drop != 1);
1769         if (iref) {
1770                 ret = update_inline_extent_backref(trans, root, path, iref,
1771                                                    -refs_to_drop, NULL);
1772         } else if (is_data) {
1773                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1774         } else {
1775                 ret = btrfs_del_item(trans, root, path);
1776         }
1777         return ret;
1778 }
1779
1780 static int btrfs_issue_discard(struct block_device *bdev,
1781                                 u64 start, u64 len)
1782 {
1783         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1784 }
1785
1786 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1787                                 u64 num_bytes, u64 *actual_bytes)
1788 {
1789         int ret;
1790         u64 discarded_bytes = 0;
1791         struct btrfs_bio *bbio = NULL;
1792
1793
1794         /* Tell the block device(s) that the sectors can be discarded */
1795         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1796                               bytenr, &num_bytes, &bbio, 0);
1797         if (!ret) {
1798                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1799                 int i;
1800
1801
1802                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1803                         if (!stripe->dev->can_discard)
1804                                 continue;
1805
1806                         ret = btrfs_issue_discard(stripe->dev->bdev,
1807                                                   stripe->physical,
1808                                                   stripe->length);
1809                         if (!ret)
1810                                 discarded_bytes += stripe->length;
1811                         else if (ret != -EOPNOTSUPP)
1812                                 break;
1813
1814                         /*
1815                          * Just in case we get back EOPNOTSUPP for some reason,
1816                          * just ignore the return value so we don't screw up
1817                          * people calling discard_extent.
1818                          */
1819                         ret = 0;
1820                 }
1821                 kfree(bbio);
1822         }
1823
1824         if (actual_bytes)
1825                 *actual_bytes = discarded_bytes;
1826
1827
1828         return ret;
1829 }
1830
1831 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1832                          struct btrfs_root *root,
1833                          u64 bytenr, u64 num_bytes, u64 parent,
1834                          u64 root_objectid, u64 owner, u64 offset)
1835 {
1836         int ret;
1837         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1838                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1839
1840         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1841                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
1842                                         parent, root_objectid, (int)owner,
1843                                         BTRFS_ADD_DELAYED_REF, NULL);
1844         } else {
1845                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
1846                                         parent, root_objectid, owner, offset,
1847                                         BTRFS_ADD_DELAYED_REF, NULL);
1848         }
1849         return ret;
1850 }
1851
1852 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1853                                   struct btrfs_root *root,
1854                                   u64 bytenr, u64 num_bytes,
1855                                   u64 parent, u64 root_objectid,
1856                                   u64 owner, u64 offset, int refs_to_add,
1857                                   struct btrfs_delayed_extent_op *extent_op)
1858 {
1859         struct btrfs_path *path;
1860         struct extent_buffer *leaf;
1861         struct btrfs_extent_item *item;
1862         u64 refs;
1863         int ret;
1864         int err = 0;
1865
1866         path = btrfs_alloc_path();
1867         if (!path)
1868                 return -ENOMEM;
1869
1870         path->reada = 1;
1871         path->leave_spinning = 1;
1872         /* this will setup the path even if it fails to insert the back ref */
1873         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1874                                            path, bytenr, num_bytes, parent,
1875                                            root_objectid, owner, offset,
1876                                            refs_to_add, extent_op);
1877         if (ret == 0)
1878                 goto out;
1879
1880         if (ret != -EAGAIN) {
1881                 err = ret;
1882                 goto out;
1883         }
1884
1885         leaf = path->nodes[0];
1886         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1887         refs = btrfs_extent_refs(leaf, item);
1888         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1889         if (extent_op)
1890                 __run_delayed_extent_op(extent_op, leaf, item);
1891
1892         btrfs_mark_buffer_dirty(leaf);
1893         btrfs_release_path(path);
1894
1895         path->reada = 1;
1896         path->leave_spinning = 1;
1897
1898         /* now insert the actual backref */
1899         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1900                                     path, bytenr, parent, root_objectid,
1901                                     owner, offset, refs_to_add);
1902         BUG_ON(ret);
1903 out:
1904         btrfs_free_path(path);
1905         return err;
1906 }
1907
1908 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1909                                 struct btrfs_root *root,
1910                                 struct btrfs_delayed_ref_node *node,
1911                                 struct btrfs_delayed_extent_op *extent_op,
1912                                 int insert_reserved)
1913 {
1914         int ret = 0;
1915         struct btrfs_delayed_data_ref *ref;
1916         struct btrfs_key ins;
1917         u64 parent = 0;
1918         u64 ref_root = 0;
1919         u64 flags = 0;
1920
1921         ins.objectid = node->bytenr;
1922         ins.offset = node->num_bytes;
1923         ins.type = BTRFS_EXTENT_ITEM_KEY;
1924
1925         ref = btrfs_delayed_node_to_data_ref(node);
1926         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1927                 parent = ref->parent;
1928         else
1929                 ref_root = ref->root;
1930
1931         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1932                 if (extent_op) {
1933                         BUG_ON(extent_op->update_key);
1934                         flags |= extent_op->flags_to_set;
1935                 }
1936                 ret = alloc_reserved_file_extent(trans, root,
1937                                                  parent, ref_root, flags,
1938                                                  ref->objectid, ref->offset,
1939                                                  &ins, node->ref_mod);
1940         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1941                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1942                                              node->num_bytes, parent,
1943                                              ref_root, ref->objectid,
1944                                              ref->offset, node->ref_mod,
1945                                              extent_op);
1946         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1947                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1948                                           node->num_bytes, parent,
1949                                           ref_root, ref->objectid,
1950                                           ref->offset, node->ref_mod,
1951                                           extent_op);
1952         } else {
1953                 BUG();
1954         }
1955         return ret;
1956 }
1957
1958 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1959                                     struct extent_buffer *leaf,
1960                                     struct btrfs_extent_item *ei)
1961 {
1962         u64 flags = btrfs_extent_flags(leaf, ei);
1963         if (extent_op->update_flags) {
1964                 flags |= extent_op->flags_to_set;
1965                 btrfs_set_extent_flags(leaf, ei, flags);
1966         }
1967
1968         if (extent_op->update_key) {
1969                 struct btrfs_tree_block_info *bi;
1970                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1971                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1972                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1973         }
1974 }
1975
1976 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1977                                  struct btrfs_root *root,
1978                                  struct btrfs_delayed_ref_node *node,
1979                                  struct btrfs_delayed_extent_op *extent_op)
1980 {
1981         struct btrfs_key key;
1982         struct btrfs_path *path;
1983         struct btrfs_extent_item *ei;
1984         struct extent_buffer *leaf;
1985         u32 item_size;
1986         int ret;
1987         int err = 0;
1988
1989         path = btrfs_alloc_path();
1990         if (!path)
1991                 return -ENOMEM;
1992
1993         key.objectid = node->bytenr;
1994         key.type = BTRFS_EXTENT_ITEM_KEY;
1995         key.offset = node->num_bytes;
1996
1997         path->reada = 1;
1998         path->leave_spinning = 1;
1999         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2000                                 path, 0, 1);
2001         if (ret < 0) {
2002                 err = ret;
2003                 goto out;
2004         }
2005         if (ret > 0) {
2006                 err = -EIO;
2007                 goto out;
2008         }
2009
2010         leaf = path->nodes[0];
2011         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2012 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2013         if (item_size < sizeof(*ei)) {
2014                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2015                                              path, (u64)-1, 0);
2016                 if (ret < 0) {
2017                         err = ret;
2018                         goto out;
2019                 }
2020                 leaf = path->nodes[0];
2021                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2022         }
2023 #endif
2024         BUG_ON(item_size < sizeof(*ei));
2025         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2026         __run_delayed_extent_op(extent_op, leaf, ei);
2027
2028         btrfs_mark_buffer_dirty(leaf);
2029 out:
2030         btrfs_free_path(path);
2031         return err;
2032 }
2033
2034 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2035                                 struct btrfs_root *root,
2036                                 struct btrfs_delayed_ref_node *node,
2037                                 struct btrfs_delayed_extent_op *extent_op,
2038                                 int insert_reserved)
2039 {
2040         int ret = 0;
2041         struct btrfs_delayed_tree_ref *ref;
2042         struct btrfs_key ins;
2043         u64 parent = 0;
2044         u64 ref_root = 0;
2045
2046         ins.objectid = node->bytenr;
2047         ins.offset = node->num_bytes;
2048         ins.type = BTRFS_EXTENT_ITEM_KEY;
2049
2050         ref = btrfs_delayed_node_to_tree_ref(node);
2051         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2052                 parent = ref->parent;
2053         else
2054                 ref_root = ref->root;
2055
2056         BUG_ON(node->ref_mod != 1);
2057         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2058                 BUG_ON(!extent_op || !extent_op->update_flags ||
2059                        !extent_op->update_key);
2060                 ret = alloc_reserved_tree_block(trans, root,
2061                                                 parent, ref_root,
2062                                                 extent_op->flags_to_set,
2063                                                 &extent_op->key,
2064                                                 ref->level, &ins);
2065         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2066                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2067                                              node->num_bytes, parent, ref_root,
2068                                              ref->level, 0, 1, extent_op);
2069         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2070                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2071                                           node->num_bytes, parent, ref_root,
2072                                           ref->level, 0, 1, extent_op);
2073         } else {
2074                 BUG();
2075         }
2076         return ret;
2077 }
2078
2079 /* helper function to actually process a single delayed ref entry */
2080 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2081                                struct btrfs_root *root,
2082                                struct btrfs_delayed_ref_node *node,
2083                                struct btrfs_delayed_extent_op *extent_op,
2084                                int insert_reserved)
2085 {
2086         int ret;
2087         if (btrfs_delayed_ref_is_head(node)) {
2088                 struct btrfs_delayed_ref_head *head;
2089                 /*
2090                  * we've hit the end of the chain and we were supposed
2091                  * to insert this extent into the tree.  But, it got
2092                  * deleted before we ever needed to insert it, so all
2093                  * we have to do is clean up the accounting
2094                  */
2095                 BUG_ON(extent_op);
2096                 head = btrfs_delayed_node_to_head(node);
2097                 if (insert_reserved) {
2098                         btrfs_pin_extent(root, node->bytenr,
2099                                          node->num_bytes, 1);
2100                         if (head->is_data) {
2101                                 ret = btrfs_del_csums(trans, root,
2102                                                       node->bytenr,
2103                                                       node->num_bytes);
2104                                 BUG_ON(ret);
2105                         }
2106                 }
2107                 mutex_unlock(&head->mutex);
2108                 return 0;
2109         }
2110
2111         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2112             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2113                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2114                                            insert_reserved);
2115         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2116                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2117                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2118                                            insert_reserved);
2119         else
2120                 BUG();
2121         return ret;
2122 }
2123
2124 static noinline struct btrfs_delayed_ref_node *
2125 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2126 {
2127         struct rb_node *node;
2128         struct btrfs_delayed_ref_node *ref;
2129         int action = BTRFS_ADD_DELAYED_REF;
2130 again:
2131         /*
2132          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2133          * this prevents ref count from going down to zero when
2134          * there still are pending delayed ref.
2135          */
2136         node = rb_prev(&head->node.rb_node);
2137         while (1) {
2138                 if (!node)
2139                         break;
2140                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2141                                 rb_node);
2142                 if (ref->bytenr != head->node.bytenr)
2143                         break;
2144                 if (ref->action == action)
2145                         return ref;
2146                 node = rb_prev(node);
2147         }
2148         if (action == BTRFS_ADD_DELAYED_REF) {
2149                 action = BTRFS_DROP_DELAYED_REF;
2150                 goto again;
2151         }
2152         return NULL;
2153 }
2154
2155 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2156                                        struct btrfs_root *root,
2157                                        struct list_head *cluster)
2158 {
2159         struct btrfs_delayed_ref_root *delayed_refs;
2160         struct btrfs_delayed_ref_node *ref;
2161         struct btrfs_delayed_ref_head *locked_ref = NULL;
2162         struct btrfs_delayed_extent_op *extent_op;
2163         int ret;
2164         int count = 0;
2165         int must_insert_reserved = 0;
2166
2167         delayed_refs = &trans->transaction->delayed_refs;
2168         while (1) {
2169                 if (!locked_ref) {
2170                         /* pick a new head ref from the cluster list */
2171                         if (list_empty(cluster))
2172                                 break;
2173
2174                         locked_ref = list_entry(cluster->next,
2175                                      struct btrfs_delayed_ref_head, cluster);
2176
2177                         /* grab the lock that says we are going to process
2178                          * all the refs for this head */
2179                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2180
2181                         /*
2182                          * we may have dropped the spin lock to get the head
2183                          * mutex lock, and that might have given someone else
2184                          * time to free the head.  If that's true, it has been
2185                          * removed from our list and we can move on.
2186                          */
2187                         if (ret == -EAGAIN) {
2188                                 locked_ref = NULL;
2189                                 count++;
2190                                 continue;
2191                         }
2192                 }
2193
2194                 /*
2195                  * record the must insert reserved flag before we
2196                  * drop the spin lock.
2197                  */
2198                 must_insert_reserved = locked_ref->must_insert_reserved;
2199                 locked_ref->must_insert_reserved = 0;
2200
2201                 extent_op = locked_ref->extent_op;
2202                 locked_ref->extent_op = NULL;
2203
2204                 /*
2205                  * locked_ref is the head node, so we have to go one
2206                  * node back for any delayed ref updates
2207                  */
2208                 ref = select_delayed_ref(locked_ref);
2209                 if (!ref) {
2210                         /* All delayed refs have been processed, Go ahead
2211                          * and send the head node to run_one_delayed_ref,
2212                          * so that any accounting fixes can happen
2213                          */
2214                         ref = &locked_ref->node;
2215
2216                         if (extent_op && must_insert_reserved) {
2217                                 kfree(extent_op);
2218                                 extent_op = NULL;
2219                         }
2220
2221                         if (extent_op) {
2222                                 spin_unlock(&delayed_refs->lock);
2223
2224                                 ret = run_delayed_extent_op(trans, root,
2225                                                             ref, extent_op);
2226                                 BUG_ON(ret);
2227                                 kfree(extent_op);
2228
2229                                 cond_resched();
2230                                 spin_lock(&delayed_refs->lock);
2231                                 continue;
2232                         }
2233
2234                         list_del_init(&locked_ref->cluster);
2235                         locked_ref = NULL;
2236                 }
2237
2238                 ref->in_tree = 0;
2239                 rb_erase(&ref->rb_node, &delayed_refs->root);
2240                 delayed_refs->num_entries--;
2241
2242                 spin_unlock(&delayed_refs->lock);
2243
2244                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2245                                           must_insert_reserved);
2246                 BUG_ON(ret);
2247
2248                 btrfs_put_delayed_ref(ref);
2249                 kfree(extent_op);
2250                 count++;
2251
2252                 cond_resched();
2253                 spin_lock(&delayed_refs->lock);
2254         }
2255         return count;
2256 }
2257
2258 /*
2259  * this starts processing the delayed reference count updates and
2260  * extent insertions we have queued up so far.  count can be
2261  * 0, which means to process everything in the tree at the start
2262  * of the run (but not newly added entries), or it can be some target
2263  * number you'd like to process.
2264  */
2265 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2266                            struct btrfs_root *root, unsigned long count)
2267 {
2268         struct rb_node *node;
2269         struct btrfs_delayed_ref_root *delayed_refs;
2270         struct btrfs_delayed_ref_node *ref;
2271         struct list_head cluster;
2272         int ret;
2273         int run_all = count == (unsigned long)-1;
2274         int run_most = 0;
2275
2276         if (root == root->fs_info->extent_root)
2277                 root = root->fs_info->tree_root;
2278
2279         delayed_refs = &trans->transaction->delayed_refs;
2280         INIT_LIST_HEAD(&cluster);
2281 again:
2282         spin_lock(&delayed_refs->lock);
2283         if (count == 0) {
2284                 count = delayed_refs->num_entries * 2;
2285                 run_most = 1;
2286         }
2287         while (1) {
2288                 if (!(run_all || run_most) &&
2289                     delayed_refs->num_heads_ready < 64)
2290                         break;
2291
2292                 /*
2293                  * go find something we can process in the rbtree.  We start at
2294                  * the beginning of the tree, and then build a cluster
2295                  * of refs to process starting at the first one we are able to
2296                  * lock
2297                  */
2298                 ret = btrfs_find_ref_cluster(trans, &cluster,
2299                                              delayed_refs->run_delayed_start);
2300                 if (ret)
2301                         break;
2302
2303                 ret = run_clustered_refs(trans, root, &cluster);
2304                 BUG_ON(ret < 0);
2305
2306                 count -= min_t(unsigned long, ret, count);
2307
2308                 if (count == 0)
2309                         break;
2310         }
2311
2312         if (run_all) {
2313                 node = rb_first(&delayed_refs->root);
2314                 if (!node)
2315                         goto out;
2316                 count = (unsigned long)-1;
2317
2318                 while (node) {
2319                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2320                                        rb_node);
2321                         if (btrfs_delayed_ref_is_head(ref)) {
2322                                 struct btrfs_delayed_ref_head *head;
2323
2324                                 head = btrfs_delayed_node_to_head(ref);
2325                                 atomic_inc(&ref->refs);
2326
2327                                 spin_unlock(&delayed_refs->lock);
2328                                 /*
2329                                  * Mutex was contended, block until it's
2330                                  * released and try again
2331                                  */
2332                                 mutex_lock(&head->mutex);
2333                                 mutex_unlock(&head->mutex);
2334
2335                                 btrfs_put_delayed_ref(ref);
2336                                 cond_resched();
2337                                 goto again;
2338                         }
2339                         node = rb_next(node);
2340                 }
2341                 spin_unlock(&delayed_refs->lock);
2342                 schedule_timeout(1);
2343                 goto again;
2344         }
2345 out:
2346         spin_unlock(&delayed_refs->lock);
2347         return 0;
2348 }
2349
2350 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2351                                 struct btrfs_root *root,
2352                                 u64 bytenr, u64 num_bytes, u64 flags,
2353                                 int is_data)
2354 {
2355         struct btrfs_delayed_extent_op *extent_op;
2356         int ret;
2357
2358         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2359         if (!extent_op)
2360                 return -ENOMEM;
2361
2362         extent_op->flags_to_set = flags;
2363         extent_op->update_flags = 1;
2364         extent_op->update_key = 0;
2365         extent_op->is_data = is_data ? 1 : 0;
2366
2367         ret = btrfs_add_delayed_extent_op(trans, bytenr, num_bytes, extent_op);
2368         if (ret)
2369                 kfree(extent_op);
2370         return ret;
2371 }
2372
2373 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2374                                       struct btrfs_root *root,
2375                                       struct btrfs_path *path,
2376                                       u64 objectid, u64 offset, u64 bytenr)
2377 {
2378         struct btrfs_delayed_ref_head *head;
2379         struct btrfs_delayed_ref_node *ref;
2380         struct btrfs_delayed_data_ref *data_ref;
2381         struct btrfs_delayed_ref_root *delayed_refs;
2382         struct rb_node *node;
2383         int ret = 0;
2384
2385         ret = -ENOENT;
2386         delayed_refs = &trans->transaction->delayed_refs;
2387         spin_lock(&delayed_refs->lock);
2388         head = btrfs_find_delayed_ref_head(trans, bytenr);
2389         if (!head)
2390                 goto out;
2391
2392         if (!mutex_trylock(&head->mutex)) {
2393                 atomic_inc(&head->node.refs);
2394                 spin_unlock(&delayed_refs->lock);
2395
2396                 btrfs_release_path(path);
2397
2398                 /*
2399                  * Mutex was contended, block until it's released and let
2400                  * caller try again
2401                  */
2402                 mutex_lock(&head->mutex);
2403                 mutex_unlock(&head->mutex);
2404                 btrfs_put_delayed_ref(&head->node);
2405                 return -EAGAIN;
2406         }
2407
2408         node = rb_prev(&head->node.rb_node);
2409         if (!node)
2410                 goto out_unlock;
2411
2412         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2413
2414         if (ref->bytenr != bytenr)
2415                 goto out_unlock;
2416
2417         ret = 1;
2418         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2419                 goto out_unlock;
2420
2421         data_ref = btrfs_delayed_node_to_data_ref(ref);
2422
2423         node = rb_prev(node);
2424         if (node) {
2425                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2426                 if (ref->bytenr == bytenr)
2427                         goto out_unlock;
2428         }
2429
2430         if (data_ref->root != root->root_key.objectid ||
2431             data_ref->objectid != objectid || data_ref->offset != offset)
2432                 goto out_unlock;
2433
2434         ret = 0;
2435 out_unlock:
2436         mutex_unlock(&head->mutex);
2437 out:
2438         spin_unlock(&delayed_refs->lock);
2439         return ret;
2440 }
2441
2442 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2443                                         struct btrfs_root *root,
2444                                         struct btrfs_path *path,
2445                                         u64 objectid, u64 offset, u64 bytenr)
2446 {
2447         struct btrfs_root *extent_root = root->fs_info->extent_root;
2448         struct extent_buffer *leaf;
2449         struct btrfs_extent_data_ref *ref;
2450         struct btrfs_extent_inline_ref *iref;
2451         struct btrfs_extent_item *ei;
2452         struct btrfs_key key;
2453         u32 item_size;
2454         int ret;
2455
2456         key.objectid = bytenr;
2457         key.offset = (u64)-1;
2458         key.type = BTRFS_EXTENT_ITEM_KEY;
2459
2460         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2461         if (ret < 0)
2462                 goto out;
2463         BUG_ON(ret == 0);
2464
2465         ret = -ENOENT;
2466         if (path->slots[0] == 0)
2467                 goto out;
2468
2469         path->slots[0]--;
2470         leaf = path->nodes[0];
2471         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2472
2473         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2474                 goto out;
2475
2476         ret = 1;
2477         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2478 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2479         if (item_size < sizeof(*ei)) {
2480                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2481                 goto out;
2482         }
2483 #endif
2484         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2485
2486         if (item_size != sizeof(*ei) +
2487             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2488                 goto out;
2489
2490         if (btrfs_extent_generation(leaf, ei) <=
2491             btrfs_root_last_snapshot(&root->root_item))
2492                 goto out;
2493
2494         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2495         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2496             BTRFS_EXTENT_DATA_REF_KEY)
2497                 goto out;
2498
2499         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2500         if (btrfs_extent_refs(leaf, ei) !=
2501             btrfs_extent_data_ref_count(leaf, ref) ||
2502             btrfs_extent_data_ref_root(leaf, ref) !=
2503             root->root_key.objectid ||
2504             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2505             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2506                 goto out;
2507
2508         ret = 0;
2509 out:
2510         return ret;
2511 }
2512
2513 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2514                           struct btrfs_root *root,
2515                           u64 objectid, u64 offset, u64 bytenr)
2516 {
2517         struct btrfs_path *path;
2518         int ret;
2519         int ret2;
2520
2521         path = btrfs_alloc_path();
2522         if (!path)
2523                 return -ENOENT;
2524
2525         do {
2526                 ret = check_committed_ref(trans, root, path, objectid,
2527                                           offset, bytenr);
2528                 if (ret && ret != -ENOENT)
2529                         goto out;
2530
2531                 ret2 = check_delayed_ref(trans, root, path, objectid,
2532                                          offset, bytenr);
2533         } while (ret2 == -EAGAIN);
2534
2535         if (ret2 && ret2 != -ENOENT) {
2536                 ret = ret2;
2537                 goto out;
2538         }
2539
2540         if (ret != -ENOENT || ret2 != -ENOENT)
2541                 ret = 0;
2542 out:
2543         btrfs_free_path(path);
2544         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2545                 WARN_ON(ret > 0);
2546         return ret;
2547 }
2548
2549 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2550                            struct btrfs_root *root,
2551                            struct extent_buffer *buf,
2552                            int full_backref, int inc)
2553 {
2554         u64 bytenr;
2555         u64 num_bytes;
2556         u64 parent;
2557         u64 ref_root;
2558         u32 nritems;
2559         struct btrfs_key key;
2560         struct btrfs_file_extent_item *fi;
2561         int i;
2562         int level;
2563         int ret = 0;
2564         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2565                             u64, u64, u64, u64, u64, u64);
2566
2567         ref_root = btrfs_header_owner(buf);
2568         nritems = btrfs_header_nritems(buf);
2569         level = btrfs_header_level(buf);
2570
2571         if (!root->ref_cows && level == 0)
2572                 return 0;
2573
2574         if (inc)
2575                 process_func = btrfs_inc_extent_ref;
2576         else
2577                 process_func = btrfs_free_extent;
2578
2579         if (full_backref)
2580                 parent = buf->start;
2581         else
2582                 parent = 0;
2583
2584         for (i = 0; i < nritems; i++) {
2585                 if (level == 0) {
2586                         btrfs_item_key_to_cpu(buf, &key, i);
2587                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2588                                 continue;
2589                         fi = btrfs_item_ptr(buf, i,
2590                                             struct btrfs_file_extent_item);
2591                         if (btrfs_file_extent_type(buf, fi) ==
2592                             BTRFS_FILE_EXTENT_INLINE)
2593                                 continue;
2594                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2595                         if (bytenr == 0)
2596                                 continue;
2597
2598                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2599                         key.offset -= btrfs_file_extent_offset(buf, fi);
2600                         ret = process_func(trans, root, bytenr, num_bytes,
2601                                            parent, ref_root, key.objectid,
2602                                            key.offset);
2603                         if (ret)
2604                                 goto fail;
2605                 } else {
2606                         bytenr = btrfs_node_blockptr(buf, i);
2607                         num_bytes = btrfs_level_size(root, level - 1);
2608                         ret = process_func(trans, root, bytenr, num_bytes,
2609                                            parent, ref_root, level - 1, 0);
2610                         if (ret)
2611                                 goto fail;
2612                 }
2613         }
2614         return 0;
2615 fail:
2616         BUG();
2617         return ret;
2618 }
2619
2620 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2621                   struct extent_buffer *buf, int full_backref)
2622 {
2623         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2624 }
2625
2626 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2627                   struct extent_buffer *buf, int full_backref)
2628 {
2629         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2630 }
2631
2632 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2633                                  struct btrfs_root *root,
2634                                  struct btrfs_path *path,
2635                                  struct btrfs_block_group_cache *cache)
2636 {
2637         int ret;
2638         struct btrfs_root *extent_root = root->fs_info->extent_root;
2639         unsigned long bi;
2640         struct extent_buffer *leaf;
2641
2642         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2643         if (ret < 0)
2644                 goto fail;
2645         BUG_ON(ret);
2646
2647         leaf = path->nodes[0];
2648         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2649         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2650         btrfs_mark_buffer_dirty(leaf);
2651         btrfs_release_path(path);
2652 fail:
2653         if (ret)
2654                 return ret;
2655         return 0;
2656
2657 }
2658
2659 static struct btrfs_block_group_cache *
2660 next_block_group(struct btrfs_root *root,
2661                  struct btrfs_block_group_cache *cache)
2662 {
2663         struct rb_node *node;
2664         spin_lock(&root->fs_info->block_group_cache_lock);
2665         node = rb_next(&cache->cache_node);
2666         btrfs_put_block_group(cache);
2667         if (node) {
2668                 cache = rb_entry(node, struct btrfs_block_group_cache,
2669                                  cache_node);
2670                 btrfs_get_block_group(cache);
2671         } else
2672                 cache = NULL;
2673         spin_unlock(&root->fs_info->block_group_cache_lock);
2674         return cache;
2675 }
2676
2677 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2678                             struct btrfs_trans_handle *trans,
2679                             struct btrfs_path *path)
2680 {
2681         struct btrfs_root *root = block_group->fs_info->tree_root;
2682         struct inode *inode = NULL;
2683         u64 alloc_hint = 0;
2684         int dcs = BTRFS_DC_ERROR;
2685         int num_pages = 0;
2686         int retries = 0;
2687         int ret = 0;
2688
2689         /*
2690          * If this block group is smaller than 100 megs don't bother caching the
2691          * block group.
2692          */
2693         if (block_group->key.offset < (100 * 1024 * 1024)) {
2694                 spin_lock(&block_group->lock);
2695                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2696                 spin_unlock(&block_group->lock);
2697                 return 0;
2698         }
2699
2700 again:
2701         inode = lookup_free_space_inode(root, block_group, path);
2702         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2703                 ret = PTR_ERR(inode);
2704                 btrfs_release_path(path);
2705                 goto out;
2706         }
2707
2708         if (IS_ERR(inode)) {
2709                 BUG_ON(retries);
2710                 retries++;
2711
2712                 if (block_group->ro)
2713                         goto out_free;
2714
2715                 ret = create_free_space_inode(root, trans, block_group, path);
2716                 if (ret)
2717                         goto out_free;
2718                 goto again;
2719         }
2720
2721         /* We've already setup this transaction, go ahead and exit */
2722         if (block_group->cache_generation == trans->transid &&
2723             i_size_read(inode)) {
2724                 dcs = BTRFS_DC_SETUP;
2725                 goto out_put;
2726         }
2727
2728         /*
2729          * We want to set the generation to 0, that way if anything goes wrong
2730          * from here on out we know not to trust this cache when we load up next
2731          * time.
2732          */
2733         BTRFS_I(inode)->generation = 0;
2734         ret = btrfs_update_inode(trans, root, inode);
2735         WARN_ON(ret);
2736
2737         if (i_size_read(inode) > 0) {
2738                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2739                                                       inode);
2740                 if (ret)
2741                         goto out_put;
2742         }
2743
2744         spin_lock(&block_group->lock);
2745         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2746                 /* We're not cached, don't bother trying to write stuff out */
2747                 dcs = BTRFS_DC_WRITTEN;
2748                 spin_unlock(&block_group->lock);
2749                 goto out_put;
2750         }
2751         spin_unlock(&block_group->lock);
2752
2753         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2754         if (!num_pages)
2755                 num_pages = 1;
2756
2757         /*
2758          * Just to make absolutely sure we have enough space, we're going to
2759          * preallocate 12 pages worth of space for each block group.  In
2760          * practice we ought to use at most 8, but we need extra space so we can
2761          * add our header and have a terminator between the extents and the
2762          * bitmaps.
2763          */
2764         num_pages *= 16;
2765         num_pages *= PAGE_CACHE_SIZE;
2766
2767         ret = btrfs_check_data_free_space(inode, num_pages);
2768         if (ret)
2769                 goto out_put;
2770
2771         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2772                                               num_pages, num_pages,
2773                                               &alloc_hint);
2774         if (!ret)
2775                 dcs = BTRFS_DC_SETUP;
2776         btrfs_free_reserved_data_space(inode, num_pages);
2777
2778 out_put:
2779         iput(inode);
2780 out_free:
2781         btrfs_release_path(path);
2782 out:
2783         spin_lock(&block_group->lock);
2784         if (!ret)
2785                 block_group->cache_generation = trans->transid;
2786         block_group->disk_cache_state = dcs;
2787         spin_unlock(&block_group->lock);
2788
2789         return ret;
2790 }
2791
2792 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2793                                    struct btrfs_root *root)
2794 {
2795         struct btrfs_block_group_cache *cache;
2796         int err = 0;
2797         struct btrfs_path *path;
2798         u64 last = 0;
2799
2800         path = btrfs_alloc_path();
2801         if (!path)
2802                 return -ENOMEM;
2803
2804 again:
2805         while (1) {
2806                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2807                 while (cache) {
2808                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2809                                 break;
2810                         cache = next_block_group(root, cache);
2811                 }
2812                 if (!cache) {
2813                         if (last == 0)
2814                                 break;
2815                         last = 0;
2816                         continue;
2817                 }
2818                 err = cache_save_setup(cache, trans, path);
2819                 last = cache->key.objectid + cache->key.offset;
2820                 btrfs_put_block_group(cache);
2821         }
2822
2823         while (1) {
2824                 if (last == 0) {
2825                         err = btrfs_run_delayed_refs(trans, root,
2826                                                      (unsigned long)-1);
2827                         BUG_ON(err);
2828                 }
2829
2830                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2831                 while (cache) {
2832                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2833                                 btrfs_put_block_group(cache);
2834                                 goto again;
2835                         }
2836
2837                         if (cache->dirty)
2838                                 break;
2839                         cache = next_block_group(root, cache);
2840                 }
2841                 if (!cache) {
2842                         if (last == 0)
2843                                 break;
2844                         last = 0;
2845                         continue;
2846                 }
2847
2848                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
2849                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
2850                 cache->dirty = 0;
2851                 last = cache->key.objectid + cache->key.offset;
2852
2853                 err = write_one_cache_group(trans, root, path, cache);
2854                 BUG_ON(err);
2855                 btrfs_put_block_group(cache);
2856         }
2857
2858         while (1) {
2859                 /*
2860                  * I don't think this is needed since we're just marking our
2861                  * preallocated extent as written, but just in case it can't
2862                  * hurt.
2863                  */
2864                 if (last == 0) {
2865                         err = btrfs_run_delayed_refs(trans, root,
2866                                                      (unsigned long)-1);
2867                         BUG_ON(err);
2868                 }
2869
2870                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2871                 while (cache) {
2872                         /*
2873                          * Really this shouldn't happen, but it could if we
2874                          * couldn't write the entire preallocated extent and
2875                          * splitting the extent resulted in a new block.
2876                          */
2877                         if (cache->dirty) {
2878                                 btrfs_put_block_group(cache);
2879                                 goto again;
2880                         }
2881                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2882                                 break;
2883                         cache = next_block_group(root, cache);
2884                 }
2885                 if (!cache) {
2886                         if (last == 0)
2887                                 break;
2888                         last = 0;
2889                         continue;
2890                 }
2891
2892                 btrfs_write_out_cache(root, trans, cache, path);
2893
2894                 /*
2895                  * If we didn't have an error then the cache state is still
2896                  * NEED_WRITE, so we can set it to WRITTEN.
2897                  */
2898                 if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2899                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
2900                 last = cache->key.objectid + cache->key.offset;
2901                 btrfs_put_block_group(cache);
2902         }
2903
2904         btrfs_free_path(path);
2905         return 0;
2906 }
2907
2908 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
2909 {
2910         struct btrfs_block_group_cache *block_group;
2911         int readonly = 0;
2912
2913         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
2914         if (!block_group || block_group->ro)
2915                 readonly = 1;
2916         if (block_group)
2917                 btrfs_put_block_group(block_group);
2918         return readonly;
2919 }
2920
2921 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
2922                              u64 total_bytes, u64 bytes_used,
2923                              struct btrfs_space_info **space_info)
2924 {
2925         struct btrfs_space_info *found;
2926         int i;
2927         int factor;
2928
2929         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
2930                      BTRFS_BLOCK_GROUP_RAID10))
2931                 factor = 2;
2932         else
2933                 factor = 1;
2934
2935         found = __find_space_info(info, flags);
2936         if (found) {
2937                 spin_lock(&found->lock);
2938                 found->total_bytes += total_bytes;
2939                 found->disk_total += total_bytes * factor;
2940                 found->bytes_used += bytes_used;
2941                 found->disk_used += bytes_used * factor;
2942                 found->full = 0;
2943                 spin_unlock(&found->lock);
2944                 *space_info = found;
2945                 return 0;
2946         }
2947         found = kzalloc(sizeof(*found), GFP_NOFS);
2948         if (!found)
2949                 return -ENOMEM;
2950
2951         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
2952                 INIT_LIST_HEAD(&found->block_groups[i]);
2953         init_rwsem(&found->groups_sem);
2954         spin_lock_init(&found->lock);
2955         found->flags = flags & (BTRFS_BLOCK_GROUP_DATA |
2956                                 BTRFS_BLOCK_GROUP_SYSTEM |
2957                                 BTRFS_BLOCK_GROUP_METADATA);
2958         found->total_bytes = total_bytes;
2959         found->disk_total = total_bytes * factor;
2960         found->bytes_used = bytes_used;
2961         found->disk_used = bytes_used * factor;
2962         found->bytes_pinned = 0;
2963         found->bytes_reserved = 0;
2964         found->bytes_readonly = 0;
2965         found->bytes_may_use = 0;
2966         found->full = 0;
2967         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
2968         found->chunk_alloc = 0;
2969         found->flush = 0;
2970         init_waitqueue_head(&found->wait);
2971         *space_info = found;
2972         list_add_rcu(&found->list, &info->space_info);
2973         return 0;
2974 }
2975
2976 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
2977 {
2978         u64 extra_flags = flags & (BTRFS_BLOCK_GROUP_RAID0 |
2979                                    BTRFS_BLOCK_GROUP_RAID1 |
2980                                    BTRFS_BLOCK_GROUP_RAID10 |
2981                                    BTRFS_BLOCK_GROUP_DUP);
2982         if (extra_flags) {
2983                 if (flags & BTRFS_BLOCK_GROUP_DATA)
2984                         fs_info->avail_data_alloc_bits |= extra_flags;
2985                 if (flags & BTRFS_BLOCK_GROUP_METADATA)
2986                         fs_info->avail_metadata_alloc_bits |= extra_flags;
2987                 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
2988                         fs_info->avail_system_alloc_bits |= extra_flags;
2989         }
2990 }
2991
2992 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
2993 {
2994         /*
2995          * we add in the count of missing devices because we want
2996          * to make sure that any RAID levels on a degraded FS
2997          * continue to be honored.
2998          */
2999         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3000                 root->fs_info->fs_devices->missing_devices;
3001
3002         if (num_devices == 1)
3003                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3004         if (num_devices < 4)
3005                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3006
3007         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3008             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3009                       BTRFS_BLOCK_GROUP_RAID10))) {
3010                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3011         }
3012
3013         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3014             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3015                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3016         }
3017
3018         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3019             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3020              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3021              (flags & BTRFS_BLOCK_GROUP_DUP)))
3022                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3023         return flags;
3024 }
3025
3026 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3027 {
3028         if (flags & BTRFS_BLOCK_GROUP_DATA)
3029                 flags |= root->fs_info->avail_data_alloc_bits &
3030                          root->fs_info->data_alloc_profile;
3031         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3032                 flags |= root->fs_info->avail_system_alloc_bits &
3033                          root->fs_info->system_alloc_profile;
3034         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3035                 flags |= root->fs_info->avail_metadata_alloc_bits &
3036                          root->fs_info->metadata_alloc_profile;
3037         return btrfs_reduce_alloc_profile(root, flags);
3038 }
3039
3040 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3041 {
3042         u64 flags;
3043
3044         if (data)
3045                 flags = BTRFS_BLOCK_GROUP_DATA;
3046         else if (root == root->fs_info->chunk_root)
3047                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3048         else
3049                 flags = BTRFS_BLOCK_GROUP_METADATA;
3050
3051         return get_alloc_profile(root, flags);
3052 }
3053
3054 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3055 {
3056         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3057                                                        BTRFS_BLOCK_GROUP_DATA);
3058 }
3059
3060 /*
3061  * This will check the space that the inode allocates from to make sure we have
3062  * enough space for bytes.
3063  */
3064 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3065 {
3066         struct btrfs_space_info *data_sinfo;
3067         struct btrfs_root *root = BTRFS_I(inode)->root;
3068         u64 used;
3069         int ret = 0, committed = 0, alloc_chunk = 1;
3070
3071         /* make sure bytes are sectorsize aligned */
3072         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3073
3074         if (root == root->fs_info->tree_root ||
3075             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3076                 alloc_chunk = 0;
3077                 committed = 1;
3078         }
3079
3080         data_sinfo = BTRFS_I(inode)->space_info;
3081         if (!data_sinfo)
3082                 goto alloc;
3083
3084 again:
3085         /* make sure we have enough space to handle the data first */
3086         spin_lock(&data_sinfo->lock);
3087         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3088                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3089                 data_sinfo->bytes_may_use;
3090
3091         if (used + bytes > data_sinfo->total_bytes) {
3092                 struct btrfs_trans_handle *trans;
3093
3094                 /*
3095                  * if we don't have enough free bytes in this space then we need
3096                  * to alloc a new chunk.
3097                  */
3098                 if (!data_sinfo->full && alloc_chunk) {
3099                         u64 alloc_target;
3100
3101                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3102                         spin_unlock(&data_sinfo->lock);
3103 alloc:
3104                         alloc_target = btrfs_get_alloc_profile(root, 1);
3105                         trans = btrfs_join_transaction(root);
3106                         if (IS_ERR(trans))
3107                                 return PTR_ERR(trans);
3108
3109                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3110                                              bytes + 2 * 1024 * 1024,
3111                                              alloc_target,
3112                                              CHUNK_ALLOC_NO_FORCE);
3113                         btrfs_end_transaction(trans, root);
3114                         if (ret < 0) {
3115                                 if (ret != -ENOSPC)
3116                                         return ret;
3117                                 else
3118                                         goto commit_trans;
3119                         }
3120
3121                         if (!data_sinfo) {
3122                                 btrfs_set_inode_space_info(root, inode);
3123                                 data_sinfo = BTRFS_I(inode)->space_info;
3124                         }
3125                         goto again;
3126                 }
3127
3128                 /*
3129                  * If we have less pinned bytes than we want to allocate then
3130                  * don't bother committing the transaction, it won't help us.
3131                  */
3132                 if (data_sinfo->bytes_pinned < bytes)
3133                         committed = 1;
3134                 spin_unlock(&data_sinfo->lock);
3135
3136                 /* commit the current transaction and try again */
3137 commit_trans:
3138                 if (!committed &&
3139                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3140                         committed = 1;
3141                         trans = btrfs_join_transaction(root);
3142                         if (IS_ERR(trans))
3143                                 return PTR_ERR(trans);
3144                         ret = btrfs_commit_transaction(trans, root);
3145                         if (ret)
3146                                 return ret;
3147                         goto again;
3148                 }
3149
3150                 return -ENOSPC;
3151         }
3152         data_sinfo->bytes_may_use += bytes;
3153         spin_unlock(&data_sinfo->lock);
3154
3155         return 0;
3156 }
3157
3158 /*
3159  * Called if we need to clear a data reservation for this inode.
3160  */
3161 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3162 {
3163         struct btrfs_root *root = BTRFS_I(inode)->root;
3164         struct btrfs_space_info *data_sinfo;
3165
3166         /* make sure bytes are sectorsize aligned */
3167         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3168
3169         data_sinfo = BTRFS_I(inode)->space_info;
3170         spin_lock(&data_sinfo->lock);
3171         data_sinfo->bytes_may_use -= bytes;
3172         spin_unlock(&data_sinfo->lock);
3173 }
3174
3175 static void force_metadata_allocation(struct btrfs_fs_info *info)
3176 {
3177         struct list_head *head = &info->space_info;
3178         struct btrfs_space_info *found;
3179
3180         rcu_read_lock();
3181         list_for_each_entry_rcu(found, head, list) {
3182                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3183                         found->force_alloc = CHUNK_ALLOC_FORCE;
3184         }
3185         rcu_read_unlock();
3186 }
3187
3188 static int should_alloc_chunk(struct btrfs_root *root,
3189                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3190                               int force)
3191 {
3192         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3193         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3194         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3195         u64 thresh;
3196
3197         if (force == CHUNK_ALLOC_FORCE)
3198                 return 1;
3199
3200         /*
3201          * We need to take into account the global rsv because for all intents
3202          * and purposes it's used space.  Don't worry about locking the
3203          * global_rsv, it doesn't change except when the transaction commits.
3204          */
3205         num_allocated += global_rsv->size;
3206
3207         /*
3208          * in limited mode, we want to have some free space up to
3209          * about 1% of the FS size.
3210          */
3211         if (force == CHUNK_ALLOC_LIMITED) {
3212                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3213                 thresh = max_t(u64, 64 * 1024 * 1024,
3214                                div_factor_fine(thresh, 1));
3215
3216                 if (num_bytes - num_allocated < thresh)
3217                         return 1;
3218         }
3219
3220         /*
3221          * we have two similar checks here, one based on percentage
3222          * and once based on a hard number of 256MB.  The idea
3223          * is that if we have a good amount of free
3224          * room, don't allocate a chunk.  A good mount is
3225          * less than 80% utilized of the chunks we have allocated,
3226          * or more than 256MB free
3227          */
3228         if (num_allocated + alloc_bytes + 256 * 1024 * 1024 < num_bytes)
3229                 return 0;
3230
3231         if (num_allocated + alloc_bytes < div_factor(num_bytes, 8))
3232                 return 0;
3233
3234         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3235
3236         /* 256MB or 5% of the FS */
3237         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 5));
3238
3239         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 3))
3240                 return 0;
3241         return 1;
3242 }
3243
3244 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3245                           struct btrfs_root *extent_root, u64 alloc_bytes,
3246                           u64 flags, int force)
3247 {
3248         struct btrfs_space_info *space_info;
3249         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3250         int wait_for_alloc = 0;
3251         int ret = 0;
3252
3253         flags = btrfs_reduce_alloc_profile(extent_root, flags);
3254
3255         space_info = __find_space_info(extent_root->fs_info, flags);
3256         if (!space_info) {
3257                 ret = update_space_info(extent_root->fs_info, flags,
3258                                         0, 0, &space_info);
3259                 BUG_ON(ret);
3260         }
3261         BUG_ON(!space_info);
3262
3263 again:
3264         spin_lock(&space_info->lock);
3265         if (space_info->force_alloc)
3266                 force = space_info->force_alloc;
3267         if (space_info->full) {
3268                 spin_unlock(&space_info->lock);
3269                 return 0;
3270         }
3271
3272         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3273                 spin_unlock(&space_info->lock);
3274                 return 0;
3275         } else if (space_info->chunk_alloc) {
3276                 wait_for_alloc = 1;
3277         } else {
3278                 space_info->chunk_alloc = 1;
3279         }
3280
3281         spin_unlock(&space_info->lock);
3282
3283         mutex_lock(&fs_info->chunk_mutex);
3284
3285         /*
3286          * The chunk_mutex is held throughout the entirety of a chunk
3287          * allocation, so once we've acquired the chunk_mutex we know that the
3288          * other guy is done and we need to recheck and see if we should
3289          * allocate.
3290          */
3291         if (wait_for_alloc) {
3292                 mutex_unlock(&fs_info->chunk_mutex);
3293                 wait_for_alloc = 0;
3294                 goto again;
3295         }
3296
3297         /*
3298          * If we have mixed data/metadata chunks we want to make sure we keep
3299          * allocating mixed chunks instead of individual chunks.
3300          */
3301         if (btrfs_mixed_space_info(space_info))
3302                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3303
3304         /*
3305          * if we're doing a data chunk, go ahead and make sure that
3306          * we keep a reasonable number of metadata chunks allocated in the
3307          * FS as well.
3308          */
3309         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3310                 fs_info->data_chunk_allocations++;
3311                 if (!(fs_info->data_chunk_allocations %
3312                       fs_info->metadata_ratio))
3313                         force_metadata_allocation(fs_info);
3314         }
3315
3316         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3317         if (ret < 0 && ret != -ENOSPC)
3318                 goto out;
3319
3320         spin_lock(&space_info->lock);
3321         if (ret)
3322                 space_info->full = 1;
3323         else
3324                 ret = 1;
3325
3326         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3327         space_info->chunk_alloc = 0;
3328         spin_unlock(&space_info->lock);
3329 out:
3330         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3331         return ret;
3332 }
3333
3334 /*
3335  * shrink metadata reservation for delalloc
3336  */
3337 static int shrink_delalloc(struct btrfs_root *root, u64 to_reclaim,
3338                            bool wait_ordered)
3339 {
3340         struct btrfs_block_rsv *block_rsv;
3341         struct btrfs_space_info *space_info;
3342         struct btrfs_trans_handle *trans;
3343         u64 reserved;
3344         u64 max_reclaim;
3345         u64 reclaimed = 0;
3346         long time_left;
3347         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3348         int loops = 0;
3349         unsigned long progress;
3350
3351         trans = (struct btrfs_trans_handle *)current->journal_info;
3352         block_rsv = &root->fs_info->delalloc_block_rsv;
3353         space_info = block_rsv->space_info;
3354
3355         smp_mb();
3356         reserved = space_info->bytes_may_use;
3357         progress = space_info->reservation_progress;
3358
3359         if (reserved == 0)
3360                 return 0;
3361
3362         smp_mb();
3363         if (root->fs_info->delalloc_bytes == 0) {
3364                 if (trans)
3365                         return 0;
3366                 btrfs_wait_ordered_extents(root, 0, 0);
3367                 return 0;
3368         }
3369
3370         max_reclaim = min(reserved, to_reclaim);
3371         nr_pages = max_t(unsigned long, nr_pages,
3372                          max_reclaim >> PAGE_CACHE_SHIFT);
3373         while (loops < 1024) {
3374                 /* have the flusher threads jump in and do some IO */
3375                 smp_mb();
3376                 nr_pages = min_t(unsigned long, nr_pages,
3377                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3378                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3379                                                 WB_REASON_FS_FREE_SPACE);
3380
3381                 spin_lock(&space_info->lock);
3382                 if (reserved > space_info->bytes_may_use)
3383                         reclaimed += reserved - space_info->bytes_may_use;
3384                 reserved = space_info->bytes_may_use;
3385                 spin_unlock(&space_info->lock);
3386
3387                 loops++;
3388
3389                 if (reserved == 0 || reclaimed >= max_reclaim)
3390                         break;
3391
3392                 if (trans && trans->transaction->blocked)
3393                         return -EAGAIN;
3394
3395                 if (wait_ordered && !trans) {
3396                         btrfs_wait_ordered_extents(root, 0, 0);
3397                 } else {
3398                         time_left = schedule_timeout_interruptible(1);
3399
3400                         /* We were interrupted, exit */
3401                         if (time_left)
3402                                 break;
3403                 }
3404
3405                 /* we've kicked the IO a few times, if anything has been freed,
3406                  * exit.  There is no sense in looping here for a long time
3407                  * when we really need to commit the transaction, or there are
3408                  * just too many writers without enough free space
3409                  */
3410
3411                 if (loops > 3) {
3412                         smp_mb();
3413                         if (progress != space_info->reservation_progress)
3414                                 break;
3415                 }
3416
3417         }
3418
3419         return reclaimed >= to_reclaim;
3420 }
3421
3422 /**
3423  * maybe_commit_transaction - possibly commit the transaction if its ok to
3424  * @root - the root we're allocating for
3425  * @bytes - the number of bytes we want to reserve
3426  * @force - force the commit
3427  *
3428  * This will check to make sure that committing the transaction will actually
3429  * get us somewhere and then commit the transaction if it does.  Otherwise it
3430  * will return -ENOSPC.
3431  */
3432 static int may_commit_transaction(struct btrfs_root *root,
3433                                   struct btrfs_space_info *space_info,
3434                                   u64 bytes, int force)
3435 {
3436         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3437         struct btrfs_trans_handle *trans;
3438
3439         trans = (struct btrfs_trans_handle *)current->journal_info;
3440         if (trans)
3441                 return -EAGAIN;
3442
3443         if (force)
3444                 goto commit;
3445
3446         /* See if there is enough pinned space to make this reservation */
3447         spin_lock(&space_info->lock);
3448         if (space_info->bytes_pinned >= bytes) {
3449                 spin_unlock(&space_info->lock);
3450                 goto commit;
3451         }
3452         spin_unlock(&space_info->lock);
3453
3454         /*
3455          * See if there is some space in the delayed insertion reservation for
3456          * this reservation.
3457          */
3458         if (space_info != delayed_rsv->space_info)
3459                 return -ENOSPC;
3460
3461         spin_lock(&delayed_rsv->lock);
3462         if (delayed_rsv->size < bytes) {
3463                 spin_unlock(&delayed_rsv->lock);
3464                 return -ENOSPC;
3465         }
3466         spin_unlock(&delayed_rsv->lock);
3467
3468 commit:
3469         trans = btrfs_join_transaction(root);
3470         if (IS_ERR(trans))
3471                 return -ENOSPC;
3472
3473         return btrfs_commit_transaction(trans, root);
3474 }
3475
3476 /**
3477  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3478  * @root - the root we're allocating for
3479  * @block_rsv - the block_rsv we're allocating for
3480  * @orig_bytes - the number of bytes we want
3481  * @flush - wether or not we can flush to make our reservation
3482  *
3483  * This will reserve orgi_bytes number of bytes from the space info associated
3484  * with the block_rsv.  If there is not enough space it will make an attempt to
3485  * flush out space to make room.  It will do this by flushing delalloc if
3486  * possible or committing the transaction.  If flush is 0 then no attempts to
3487  * regain reservations will be made and this will fail if there is not enough
3488  * space already.
3489  */
3490 static int reserve_metadata_bytes(struct btrfs_root *root,
3491                                   struct btrfs_block_rsv *block_rsv,
3492                                   u64 orig_bytes, int flush)
3493 {
3494         struct btrfs_space_info *space_info = block_rsv->space_info;
3495         u64 used;
3496         u64 num_bytes = orig_bytes;
3497         int retries = 0;
3498         int ret = 0;
3499         bool committed = false;
3500         bool flushing = false;
3501         bool wait_ordered = false;
3502
3503 again:
3504         ret = 0;
3505         spin_lock(&space_info->lock);
3506         /*
3507          * We only want to wait if somebody other than us is flushing and we are
3508          * actually alloed to flush.
3509          */
3510         while (flush && !flushing && space_info->flush) {
3511                 spin_unlock(&space_info->lock);
3512                 /*
3513                  * If we have a trans handle we can't wait because the flusher
3514                  * may have to commit the transaction, which would mean we would
3515                  * deadlock since we are waiting for the flusher to finish, but
3516                  * hold the current transaction open.
3517                  */
3518                 if (current->journal_info)
3519                         return -EAGAIN;
3520                 ret = wait_event_interruptible(space_info->wait,
3521                                                !space_info->flush);
3522                 /* Must have been interrupted, return */
3523                 if (ret)
3524                         return -EINTR;
3525
3526                 spin_lock(&space_info->lock);
3527         }
3528
3529         ret = -ENOSPC;
3530         used = space_info->bytes_used + space_info->bytes_reserved +
3531                 space_info->bytes_pinned + space_info->bytes_readonly +
3532                 space_info->bytes_may_use;
3533
3534         /*
3535          * The idea here is that we've not already over-reserved the block group
3536          * then we can go ahead and save our reservation first and then start
3537          * flushing if we need to.  Otherwise if we've already overcommitted
3538          * lets start flushing stuff first and then come back and try to make
3539          * our reservation.
3540          */
3541         if (used <= space_info->total_bytes) {
3542                 if (used + orig_bytes <= space_info->total_bytes) {
3543                         space_info->bytes_may_use += orig_bytes;
3544                         ret = 0;
3545                 } else {
3546                         /*
3547                          * Ok set num_bytes to orig_bytes since we aren't
3548                          * overocmmitted, this way we only try and reclaim what
3549                          * we need.
3550                          */
3551                         num_bytes = orig_bytes;
3552                 }
3553         } else {
3554                 /*
3555                  * Ok we're over committed, set num_bytes to the overcommitted
3556                  * amount plus the amount of bytes that we need for this
3557                  * reservation.
3558                  */
3559                 wait_ordered = true;
3560                 num_bytes = used - space_info->total_bytes +
3561                         (orig_bytes * (retries + 1));
3562         }
3563
3564         if (ret) {
3565                 u64 profile = btrfs_get_alloc_profile(root, 0);
3566                 u64 avail;
3567
3568                 /*
3569                  * If we have a lot of space that's pinned, don't bother doing
3570                  * the overcommit dance yet and just commit the transaction.
3571                  */
3572                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3573                 do_div(avail, 10);
3574                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3575                         space_info->flush = 1;
3576                         flushing = true;
3577                         spin_unlock(&space_info->lock);
3578                         ret = may_commit_transaction(root, space_info,
3579                                                      orig_bytes, 1);
3580                         if (ret)
3581                                 goto out;
3582                         committed = true;
3583                         goto again;
3584                 }
3585
3586                 spin_lock(&root->fs_info->free_chunk_lock);
3587                 avail = root->fs_info->free_chunk_space;
3588
3589                 /*
3590                  * If we have dup, raid1 or raid10 then only half of the free
3591                  * space is actually useable.
3592                  */
3593                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3594                                BTRFS_BLOCK_GROUP_RAID1 |
3595                                BTRFS_BLOCK_GROUP_RAID10))
3596                         avail >>= 1;
3597
3598                 /*
3599                  * If we aren't flushing don't let us overcommit too much, say
3600                  * 1/8th of the space.  If we can flush, let it overcommit up to
3601                  * 1/2 of the space.
3602                  */
3603                 if (flush)
3604                         avail >>= 3;
3605                 else
3606                         avail >>= 1;
3607                  spin_unlock(&root->fs_info->free_chunk_lock);
3608
3609                 if (used + num_bytes < space_info->total_bytes + avail) {
3610                         space_info->bytes_may_use += orig_bytes;
3611                         ret = 0;
3612                 } else {
3613                         wait_ordered = true;
3614                 }
3615         }
3616
3617         /*
3618          * Couldn't make our reservation, save our place so while we're trying
3619          * to reclaim space we can actually use it instead of somebody else
3620          * stealing it from us.
3621          */
3622         if (ret && flush) {
3623                 flushing = true;
3624                 space_info->flush = 1;
3625         }
3626
3627         spin_unlock(&space_info->lock);
3628
3629         if (!ret || !flush)
3630                 goto out;
3631
3632         /*
3633          * We do synchronous shrinking since we don't actually unreserve
3634          * metadata until after the IO is completed.
3635          */
3636         ret = shrink_delalloc(root, num_bytes, wait_ordered);
3637         if (ret < 0)
3638                 goto out;
3639
3640         ret = 0;
3641
3642         /*
3643          * So if we were overcommitted it's possible that somebody else flushed
3644          * out enough space and we simply didn't have enough space to reclaim,
3645          * so go back around and try again.
3646          */
3647         if (retries < 2) {
3648                 wait_ordered = true;
3649                 retries++;
3650                 goto again;
3651         }
3652
3653         ret = -ENOSPC;
3654         if (committed)
3655                 goto out;
3656
3657         ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3658         if (!ret) {
3659                 committed = true;
3660                 goto again;
3661         }
3662
3663 out:
3664         if (flushing) {
3665                 spin_lock(&space_info->lock);
3666                 space_info->flush = 0;
3667                 wake_up_all(&space_info->wait);
3668                 spin_unlock(&space_info->lock);
3669         }
3670         return ret;
3671 }
3672
3673 static struct btrfs_block_rsv *get_block_rsv(struct btrfs_trans_handle *trans,
3674                                              struct btrfs_root *root)
3675 {
3676         struct btrfs_block_rsv *block_rsv = NULL;
3677
3678         if (root->ref_cows || root == root->fs_info->csum_root)
3679                 block_rsv = trans->block_rsv;
3680
3681         if (!block_rsv)
3682                 block_rsv = root->block_rsv;
3683
3684         if (!block_rsv)
3685                 block_rsv = &root->fs_info->empty_block_rsv;
3686
3687         return block_rsv;
3688 }
3689
3690 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3691                                u64 num_bytes)
3692 {
3693         int ret = -ENOSPC;
3694         spin_lock(&block_rsv->lock);
3695         if (block_rsv->reserved >= num_bytes) {
3696                 block_rsv->reserved -= num_bytes;
3697                 if (block_rsv->reserved < block_rsv->size)
3698                         block_rsv->full = 0;
3699                 ret = 0;
3700         }
3701         spin_unlock(&block_rsv->lock);
3702         return ret;
3703 }
3704
3705 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3706                                 u64 num_bytes, int update_size)
3707 {
3708         spin_lock(&block_rsv->lock);
3709         block_rsv->reserved += num_bytes;
3710         if (update_size)
3711                 block_rsv->size += num_bytes;
3712         else if (block_rsv->reserved >= block_rsv->size)
3713                 block_rsv->full = 1;
3714         spin_unlock(&block_rsv->lock);
3715 }
3716
3717 static void block_rsv_release_bytes(struct btrfs_block_rsv *block_rsv,
3718                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3719 {
3720         struct btrfs_space_info *space_info = block_rsv->space_info;
3721
3722         spin_lock(&block_rsv->lock);
3723         if (num_bytes == (u64)-1)
3724                 num_bytes = block_rsv->size;
3725         block_rsv->size -= num_bytes;
3726         if (block_rsv->reserved >= block_rsv->size) {
3727                 num_bytes = block_rsv->reserved - block_rsv->size;
3728                 block_rsv->reserved = block_rsv->size;
3729                 block_rsv->full = 1;
3730         } else {
3731                 num_bytes = 0;
3732         }
3733         spin_unlock(&block_rsv->lock);
3734
3735         if (num_bytes > 0) {
3736                 if (dest) {
3737                         spin_lock(&dest->lock);
3738                         if (!dest->full) {
3739                                 u64 bytes_to_add;
3740
3741                                 bytes_to_add = dest->size - dest->reserved;
3742                                 bytes_to_add = min(num_bytes, bytes_to_add);
3743                                 dest->reserved += bytes_to_add;
3744                                 if (dest->reserved >= dest->size)
3745                                         dest->full = 1;
3746                                 num_bytes -= bytes_to_add;
3747                         }
3748                         spin_unlock(&dest->lock);
3749                 }
3750                 if (num_bytes) {
3751                         spin_lock(&space_info->lock);
3752                         space_info->bytes_may_use -= num_bytes;
3753                         space_info->reservation_progress++;
3754                         spin_unlock(&space_info->lock);
3755                 }
3756         }
3757 }
3758
3759 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3760                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3761 {
3762         int ret;
3763
3764         ret = block_rsv_use_bytes(src, num_bytes);
3765         if (ret)
3766                 return ret;
3767
3768         block_rsv_add_bytes(dst, num_bytes, 1);
3769         return 0;
3770 }
3771
3772 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3773 {
3774         memset(rsv, 0, sizeof(*rsv));
3775         spin_lock_init(&rsv->lock);
3776 }
3777
3778 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3779 {
3780         struct btrfs_block_rsv *block_rsv;
3781         struct btrfs_fs_info *fs_info = root->fs_info;
3782
3783         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3784         if (!block_rsv)
3785                 return NULL;
3786
3787         btrfs_init_block_rsv(block_rsv);
3788         block_rsv->space_info = __find_space_info(fs_info,
3789                                                   BTRFS_BLOCK_GROUP_METADATA);
3790         return block_rsv;
3791 }
3792
3793 void btrfs_free_block_rsv(struct btrfs_root *root,
3794                           struct btrfs_block_rsv *rsv)
3795 {
3796         btrfs_block_rsv_release(root, rsv, (u64)-1);
3797         kfree(rsv);
3798 }
3799
3800 static inline int __block_rsv_add(struct btrfs_root *root,
3801                                   struct btrfs_block_rsv *block_rsv,
3802                                   u64 num_bytes, int flush)
3803 {
3804         int ret;
3805
3806         if (num_bytes == 0)
3807                 return 0;
3808
3809         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
3810         if (!ret) {
3811                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3812                 return 0;
3813         }
3814
3815         return ret;
3816 }
3817
3818 int btrfs_block_rsv_add(struct btrfs_root *root,
3819                         struct btrfs_block_rsv *block_rsv,
3820                         u64 num_bytes)
3821 {
3822         return __block_rsv_add(root, block_rsv, num_bytes, 1);
3823 }
3824
3825 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
3826                                 struct btrfs_block_rsv *block_rsv,
3827                                 u64 num_bytes)
3828 {
3829         return __block_rsv_add(root, block_rsv, num_bytes, 0);
3830 }
3831
3832 int btrfs_block_rsv_check(struct btrfs_root *root,
3833                           struct btrfs_block_rsv *block_rsv, int min_factor)
3834 {
3835         u64 num_bytes = 0;
3836         int ret = -ENOSPC;
3837
3838         if (!block_rsv)
3839                 return 0;
3840
3841         spin_lock(&block_rsv->lock);
3842         num_bytes = div_factor(block_rsv->size, min_factor);
3843         if (block_rsv->reserved >= num_bytes)
3844                 ret = 0;
3845         spin_unlock(&block_rsv->lock);
3846
3847         return ret;
3848 }
3849
3850 int btrfs_block_rsv_refill(struct btrfs_root *root,
3851                           struct btrfs_block_rsv *block_rsv,
3852                           u64 min_reserved)
3853 {
3854         u64 num_bytes = 0;
3855         int ret = -ENOSPC;
3856
3857         if (!block_rsv)
3858                 return 0;
3859
3860         spin_lock(&block_rsv->lock);
3861         num_bytes = min_reserved;
3862         if (block_rsv->reserved >= num_bytes)
3863                 ret = 0;
3864         else
3865                 num_bytes -= block_rsv->reserved;
3866         spin_unlock(&block_rsv->lock);
3867
3868         if (!ret)
3869                 return 0;
3870
3871         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, 1);
3872         if (!ret) {
3873                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
3874                 return 0;
3875         }
3876
3877         return ret;
3878 }
3879
3880 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
3881                             struct btrfs_block_rsv *dst_rsv,
3882                             u64 num_bytes)
3883 {
3884         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3885 }
3886
3887 void btrfs_block_rsv_release(struct btrfs_root *root,
3888                              struct btrfs_block_rsv *block_rsv,
3889                              u64 num_bytes)
3890 {
3891         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3892         if (global_rsv->full || global_rsv == block_rsv ||
3893             block_rsv->space_info != global_rsv->space_info)
3894                 global_rsv = NULL;
3895         block_rsv_release_bytes(block_rsv, global_rsv, num_bytes);
3896 }
3897
3898 /*
3899  * helper to calculate size of global block reservation.
3900  * the desired value is sum of space used by extent tree,
3901  * checksum tree and root tree
3902  */
3903 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
3904 {
3905         struct btrfs_space_info *sinfo;
3906         u64 num_bytes;
3907         u64 meta_used;
3908         u64 data_used;
3909         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
3910
3911         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
3912         spin_lock(&sinfo->lock);
3913         data_used = sinfo->bytes_used;
3914         spin_unlock(&sinfo->lock);
3915
3916         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3917         spin_lock(&sinfo->lock);
3918         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
3919                 data_used = 0;
3920         meta_used = sinfo->bytes_used;
3921         spin_unlock(&sinfo->lock);
3922
3923         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
3924                     csum_size * 2;
3925         num_bytes += div64_u64(data_used + meta_used, 50);
3926
3927         if (num_bytes * 3 > meta_used)
3928                 num_bytes = div64_u64(meta_used, 3);
3929
3930         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
3931 }
3932
3933 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
3934 {
3935         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
3936         struct btrfs_space_info *sinfo = block_rsv->space_info;
3937         u64 num_bytes;
3938
3939         num_bytes = calc_global_metadata_size(fs_info);
3940
3941         spin_lock(&block_rsv->lock);
3942         spin_lock(&sinfo->lock);
3943
3944         block_rsv->size = num_bytes;
3945
3946         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
3947                     sinfo->bytes_reserved + sinfo->bytes_readonly +
3948                     sinfo->bytes_may_use;
3949
3950         if (sinfo->total_bytes > num_bytes) {
3951                 num_bytes = sinfo->total_bytes - num_bytes;
3952                 block_rsv->reserved += num_bytes;
3953                 sinfo->bytes_may_use += num_bytes;
3954         }
3955
3956         if (block_rsv->reserved >= block_rsv->size) {
3957                 num_bytes = block_rsv->reserved - block_rsv->size;
3958                 sinfo->bytes_may_use -= num_bytes;
3959                 sinfo->reservation_progress++;
3960                 block_rsv->reserved = block_rsv->size;
3961                 block_rsv->full = 1;
3962         }
3963
3964         spin_unlock(&sinfo->lock);
3965         spin_unlock(&block_rsv->lock);
3966 }
3967
3968 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
3969 {
3970         struct btrfs_space_info *space_info;
3971
3972         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3973         fs_info->chunk_block_rsv.space_info = space_info;
3974
3975         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3976         fs_info->global_block_rsv.space_info = space_info;
3977         fs_info->delalloc_block_rsv.space_info = space_info;
3978         fs_info->trans_block_rsv.space_info = space_info;
3979         fs_info->empty_block_rsv.space_info = space_info;
3980         fs_info->delayed_block_rsv.space_info = space_info;
3981
3982         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
3983         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
3984         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
3985         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
3986         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
3987
3988         update_global_block_rsv(fs_info);
3989 }
3990
3991 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
3992 {
3993         block_rsv_release_bytes(&fs_info->global_block_rsv, NULL, (u64)-1);
3994         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
3995         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
3996         WARN_ON(fs_info->trans_block_rsv.size > 0);
3997         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
3998         WARN_ON(fs_info->chunk_block_rsv.size > 0);
3999         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4000         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4001         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4002 }
4003
4004 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4005                                   struct btrfs_root *root)
4006 {
4007         if (!trans->bytes_reserved)
4008                 return;
4009
4010         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4011         trans->bytes_reserved = 0;
4012 }
4013
4014 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4015                                   struct inode *inode)
4016 {
4017         struct btrfs_root *root = BTRFS_I(inode)->root;
4018         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4019         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4020
4021         /*
4022          * We need to hold space in order to delete our orphan item once we've
4023          * added it, so this takes the reservation so we can release it later
4024          * when we are truly done with the orphan item.
4025          */
4026         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4027         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4028 }
4029
4030 void btrfs_orphan_release_metadata(struct inode *inode)
4031 {
4032         struct btrfs_root *root = BTRFS_I(inode)->root;
4033         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4034         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4035 }
4036
4037 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4038                                 struct btrfs_pending_snapshot *pending)
4039 {
4040         struct btrfs_root *root = pending->root;
4041         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4042         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4043         /*
4044          * two for root back/forward refs, two for directory entries
4045          * and one for root of the snapshot.
4046          */
4047         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4048         dst_rsv->space_info = src_rsv->space_info;
4049         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4050 }
4051
4052 /**
4053  * drop_outstanding_extent - drop an outstanding extent
4054  * @inode: the inode we're dropping the extent for
4055  *
4056  * This is called when we are freeing up an outstanding extent, either called
4057  * after an error or after an extent is written.  This will return the number of
4058  * reserved extents that need to be freed.  This must be called with
4059  * BTRFS_I(inode)->lock held.
4060  */
4061 static unsigned drop_outstanding_extent(struct inode *inode)
4062 {
4063         unsigned drop_inode_space = 0;
4064         unsigned dropped_extents = 0;
4065
4066         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4067         BTRFS_I(inode)->outstanding_extents--;
4068
4069         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4070             BTRFS_I(inode)->delalloc_meta_reserved) {
4071                 drop_inode_space = 1;
4072                 BTRFS_I(inode)->delalloc_meta_reserved = 0;
4073         }
4074
4075         /*
4076          * If we have more or the same amount of outsanding extents than we have
4077          * reserved then we need to leave the reserved extents count alone.
4078          */
4079         if (BTRFS_I(inode)->outstanding_extents >=
4080             BTRFS_I(inode)->reserved_extents)
4081                 return drop_inode_space;
4082
4083         dropped_extents = BTRFS_I(inode)->reserved_extents -
4084                 BTRFS_I(inode)->outstanding_extents;
4085         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4086         return dropped_extents + drop_inode_space;
4087 }
4088
4089 /**
4090  * calc_csum_metadata_size - return the amount of metada space that must be
4091  *      reserved/free'd for the given bytes.
4092  * @inode: the inode we're manipulating
4093  * @num_bytes: the number of bytes in question
4094  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4095  *
4096  * This adjusts the number of csum_bytes in the inode and then returns the
4097  * correct amount of metadata that must either be reserved or freed.  We
4098  * calculate how many checksums we can fit into one leaf and then divide the
4099  * number of bytes that will need to be checksumed by this value to figure out
4100  * how many checksums will be required.  If we are adding bytes then the number
4101  * may go up and we will return the number of additional bytes that must be
4102  * reserved.  If it is going down we will return the number of bytes that must
4103  * be freed.
4104  *
4105  * This must be called with BTRFS_I(inode)->lock held.
4106  */
4107 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4108                                    int reserve)
4109 {
4110         struct btrfs_root *root = BTRFS_I(inode)->root;
4111         u64 csum_size;
4112         int num_csums_per_leaf;
4113         int num_csums;
4114         int old_csums;
4115
4116         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4117             BTRFS_I(inode)->csum_bytes == 0)
4118                 return 0;
4119
4120         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4121         if (reserve)
4122                 BTRFS_I(inode)->csum_bytes += num_bytes;
4123         else
4124                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4125         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4126         num_csums_per_leaf = (int)div64_u64(csum_size,
4127                                             sizeof(struct btrfs_csum_item) +
4128                                             sizeof(struct btrfs_disk_key));
4129         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4130         num_csums = num_csums + num_csums_per_leaf - 1;
4131         num_csums = num_csums / num_csums_per_leaf;
4132
4133         old_csums = old_csums + num_csums_per_leaf - 1;
4134         old_csums = old_csums / num_csums_per_leaf;
4135
4136         /* No change, no need to reserve more */
4137         if (old_csums == num_csums)
4138                 return 0;
4139
4140         if (reserve)
4141                 return btrfs_calc_trans_metadata_size(root,
4142                                                       num_csums - old_csums);
4143
4144         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4145 }
4146
4147 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4148 {
4149         struct btrfs_root *root = BTRFS_I(inode)->root;
4150         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4151         u64 to_reserve = 0;
4152         unsigned nr_extents = 0;
4153         int flush = 1;
4154         int ret;
4155
4156         if (btrfs_is_free_space_inode(root, inode))
4157                 flush = 0;
4158
4159         if (flush && btrfs_transaction_in_commit(root->fs_info))
4160                 schedule_timeout(1);
4161
4162         num_bytes = ALIGN(num_bytes, root->sectorsize);
4163
4164         spin_lock(&BTRFS_I(inode)->lock);
4165         BTRFS_I(inode)->outstanding_extents++;
4166
4167         if (BTRFS_I(inode)->outstanding_extents >
4168             BTRFS_I(inode)->reserved_extents) {
4169                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4170                         BTRFS_I(inode)->reserved_extents;
4171                 BTRFS_I(inode)->reserved_extents += nr_extents;
4172         }
4173
4174         /*
4175          * Add an item to reserve for updating the inode when we complete the
4176          * delalloc io.
4177          */
4178         if (!BTRFS_I(inode)->delalloc_meta_reserved) {
4179                 nr_extents++;
4180                 BTRFS_I(inode)->delalloc_meta_reserved = 1;
4181         }
4182
4183         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4184         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4185         spin_unlock(&BTRFS_I(inode)->lock);
4186
4187         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4188         if (ret) {
4189                 u64 to_free = 0;
4190                 unsigned dropped;
4191
4192                 spin_lock(&BTRFS_I(inode)->lock);
4193                 dropped = drop_outstanding_extent(inode);
4194                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4195                 spin_unlock(&BTRFS_I(inode)->lock);
4196                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4197
4198                 /*
4199                  * Somebody could have come in and twiddled with the
4200                  * reservation, so if we have to free more than we would have
4201                  * reserved from this reservation go ahead and release those
4202                  * bytes.
4203                  */
4204                 to_free -= to_reserve;
4205                 if (to_free)
4206                         btrfs_block_rsv_release(root, block_rsv, to_free);
4207                 return ret;
4208         }
4209
4210         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4211
4212         return 0;
4213 }
4214
4215 /**
4216  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4217  * @inode: the inode to release the reservation for
4218  * @num_bytes: the number of bytes we're releasing
4219  *
4220  * This will release the metadata reservation for an inode.  This can be called
4221  * once we complete IO for a given set of bytes to release their metadata
4222  * reservations.
4223  */
4224 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4225 {
4226         struct btrfs_root *root = BTRFS_I(inode)->root;
4227         u64 to_free = 0;
4228         unsigned dropped;
4229
4230         num_bytes = ALIGN(num_bytes, root->sectorsize);
4231         spin_lock(&BTRFS_I(inode)->lock);
4232         dropped = drop_outstanding_extent(inode);
4233
4234         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4235         spin_unlock(&BTRFS_I(inode)->lock);
4236         if (dropped > 0)
4237                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4238
4239         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4240                                 to_free);
4241 }
4242
4243 /**
4244  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4245  * @inode: inode we're writing to
4246  * @num_bytes: the number of bytes we want to allocate
4247  *
4248  * This will do the following things
4249  *
4250  * o reserve space in the data space info for num_bytes
4251  * o reserve space in the metadata space info based on number of outstanding
4252  *   extents and how much csums will be needed
4253  * o add to the inodes ->delalloc_bytes
4254  * o add it to the fs_info's delalloc inodes list.
4255  *
4256  * This will return 0 for success and -ENOSPC if there is no space left.
4257  */
4258 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4259 {
4260         int ret;
4261
4262         ret = btrfs_check_data_free_space(inode, num_bytes);
4263         if (ret)
4264                 return ret;
4265
4266         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4267         if (ret) {
4268                 btrfs_free_reserved_data_space(inode, num_bytes);
4269                 return ret;
4270         }
4271
4272         return 0;
4273 }
4274
4275 /**
4276  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4277  * @inode: inode we're releasing space for
4278  * @num_bytes: the number of bytes we want to free up
4279  *
4280  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4281  * called in the case that we don't need the metadata AND data reservations
4282  * anymore.  So if there is an error or we insert an inline extent.
4283  *
4284  * This function will release the metadata space that was not used and will
4285  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4286  * list if there are no delalloc bytes left.
4287  */
4288 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4289 {
4290         btrfs_delalloc_release_metadata(inode, num_bytes);
4291         btrfs_free_reserved_data_space(inode, num_bytes);
4292 }
4293
4294 static int update_block_group(struct btrfs_trans_handle *trans,
4295                               struct btrfs_root *root,
4296                               u64 bytenr, u64 num_bytes, int alloc)
4297 {
4298         struct btrfs_block_group_cache *cache = NULL;
4299         struct btrfs_fs_info *info = root->fs_info;
4300         u64 total = num_bytes;
4301         u64 old_val;
4302         u64 byte_in_group;
4303         int factor;
4304
4305         /* block accounting for super block */
4306         spin_lock(&info->delalloc_lock);
4307         old_val = btrfs_super_bytes_used(info->super_copy);
4308         if (alloc)
4309                 old_val += num_bytes;
4310         else
4311                 old_val -= num_bytes;
4312         btrfs_set_super_bytes_used(info->super_copy, old_val);
4313         spin_unlock(&info->delalloc_lock);
4314
4315         while (total) {
4316                 cache = btrfs_lookup_block_group(info, bytenr);
4317                 if (!cache)
4318                         return -1;
4319                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4320                                     BTRFS_BLOCK_GROUP_RAID1 |
4321                                     BTRFS_BLOCK_GROUP_RAID10))
4322                         factor = 2;
4323                 else
4324                         factor = 1;
4325                 /*
4326                  * If this block group has free space cache written out, we
4327                  * need to make sure to load it if we are removing space.  This
4328                  * is because we need the unpinning stage to actually add the
4329                  * space back to the block group, otherwise we will leak space.
4330                  */
4331                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4332                         cache_block_group(cache, trans, NULL, 1);
4333
4334                 byte_in_group = bytenr - cache->key.objectid;
4335                 WARN_ON(byte_in_group > cache->key.offset);
4336
4337                 spin_lock(&cache->space_info->lock);
4338                 spin_lock(&cache->lock);
4339
4340                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4341                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4342                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4343
4344                 cache->dirty = 1;
4345                 old_val = btrfs_block_group_used(&cache->item);
4346                 num_bytes = min(total, cache->key.offset - byte_in_group);
4347                 if (alloc) {
4348                         old_val += num_bytes;
4349                         btrfs_set_block_group_used(&cache->item, old_val);
4350                         cache->reserved -= num_bytes;
4351                         cache->space_info->bytes_reserved -= num_bytes;
4352                         cache->space_info->bytes_used += num_bytes;
4353                         cache->space_info->disk_used += num_bytes * factor;
4354                         spin_unlock(&cache->lock);
4355                         spin_unlock(&cache->space_info->lock);
4356                 } else {
4357                         old_val -= num_bytes;
4358                         btrfs_set_block_group_used(&cache->item, old_val);
4359                         cache->pinned += num_bytes;
4360                         cache->space_info->bytes_pinned += num_bytes;
4361                         cache->space_info->bytes_used -= num_bytes;
4362                         cache->space_info->disk_used -= num_bytes * factor;
4363                         spin_unlock(&cache->lock);
4364                         spin_unlock(&cache->space_info->lock);
4365
4366                         set_extent_dirty(info->pinned_extents,
4367                                          bytenr, bytenr + num_bytes - 1,
4368                                          GFP_NOFS | __GFP_NOFAIL);
4369                 }
4370                 btrfs_put_block_group(cache);
4371                 total -= num_bytes;
4372                 bytenr += num_bytes;
4373         }
4374         return 0;
4375 }
4376
4377 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4378 {
4379         struct btrfs_block_group_cache *cache;
4380         u64 bytenr;
4381
4382         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4383         if (!cache)
4384                 return 0;
4385
4386         bytenr = cache->key.objectid;
4387         btrfs_put_block_group(cache);
4388
4389         return bytenr;
4390 }
4391
4392 static int pin_down_extent(struct btrfs_root *root,
4393                            struct btrfs_block_group_cache *cache,
4394                            u64 bytenr, u64 num_bytes, int reserved)
4395 {
4396         spin_lock(&cache->space_info->lock);
4397         spin_lock(&cache->lock);
4398         cache->pinned += num_bytes;
4399         cache->space_info->bytes_pinned += num_bytes;
4400         if (reserved) {
4401                 cache->reserved -= num_bytes;
4402                 cache->space_info->bytes_reserved -= num_bytes;
4403         }
4404         spin_unlock(&cache->lock);
4405         spin_unlock(&cache->space_info->lock);
4406
4407         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4408                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4409         return 0;
4410 }
4411
4412 /*
4413  * this function must be called within transaction
4414  */
4415 int btrfs_pin_extent(struct btrfs_root *root,
4416                      u64 bytenr, u64 num_bytes, int reserved)
4417 {
4418         struct btrfs_block_group_cache *cache;
4419
4420         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4421         BUG_ON(!cache);
4422
4423         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4424
4425         btrfs_put_block_group(cache);
4426         return 0;
4427 }
4428
4429 /*
4430  * this function must be called within transaction
4431  */
4432 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4433                                     struct btrfs_root *root,
4434                                     u64 bytenr, u64 num_bytes)
4435 {
4436         struct btrfs_block_group_cache *cache;
4437
4438         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4439         BUG_ON(!cache);
4440
4441         /*
4442          * pull in the free space cache (if any) so that our pin
4443          * removes the free space from the cache.  We have load_only set
4444          * to one because the slow code to read in the free extents does check
4445          * the pinned extents.
4446          */
4447         cache_block_group(cache, trans, root, 1);
4448
4449         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4450
4451         /* remove us from the free space cache (if we're there at all) */
4452         btrfs_remove_free_space(cache, bytenr, num_bytes);
4453         btrfs_put_block_group(cache);
4454         return 0;
4455 }
4456
4457 /**
4458  * btrfs_update_reserved_bytes - update the block_group and space info counters
4459  * @cache:      The cache we are manipulating
4460  * @num_bytes:  The number of bytes in question
4461  * @reserve:    One of the reservation enums
4462  *
4463  * This is called by the allocator when it reserves space, or by somebody who is
4464  * freeing space that was never actually used on disk.  For example if you
4465  * reserve some space for a new leaf in transaction A and before transaction A
4466  * commits you free that leaf, you call this with reserve set to 0 in order to
4467  * clear the reservation.
4468  *
4469  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4470  * ENOSPC accounting.  For data we handle the reservation through clearing the
4471  * delalloc bits in the io_tree.  We have to do this since we could end up
4472  * allocating less disk space for the amount of data we have reserved in the
4473  * case of compression.
4474  *
4475  * If this is a reservation and the block group has become read only we cannot
4476  * make the reservation and return -EAGAIN, otherwise this function always
4477  * succeeds.
4478  */
4479 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4480                                        u64 num_bytes, int reserve)
4481 {
4482         struct btrfs_space_info *space_info = cache->space_info;
4483         int ret = 0;
4484         spin_lock(&space_info->lock);
4485         spin_lock(&cache->lock);
4486         if (reserve != RESERVE_FREE) {
4487                 if (cache->ro) {
4488                         ret = -EAGAIN;
4489                 } else {
4490                         cache->reserved += num_bytes;
4491                         space_info->bytes_reserved += num_bytes;
4492                         if (reserve == RESERVE_ALLOC) {
4493                                 BUG_ON(space_info->bytes_may_use < num_bytes);
4494                                 space_info->bytes_may_use -= num_bytes;
4495                         }
4496                 }
4497         } else {
4498                 if (cache->ro)
4499                         space_info->bytes_readonly += num_bytes;
4500                 cache->reserved -= num_bytes;
4501                 space_info->bytes_reserved -= num_bytes;
4502                 space_info->reservation_progress++;
4503         }
4504         spin_unlock(&cache->lock);
4505         spin_unlock(&space_info->lock);
4506         return ret;
4507 }
4508
4509 int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4510                                 struct btrfs_root *root)
4511 {
4512         struct btrfs_fs_info *fs_info = root->fs_info;
4513         struct btrfs_caching_control *next;
4514         struct btrfs_caching_control *caching_ctl;
4515         struct btrfs_block_group_cache *cache;
4516
4517         down_write(&fs_info->extent_commit_sem);
4518
4519         list_for_each_entry_safe(caching_ctl, next,
4520                                  &fs_info->caching_block_groups, list) {
4521                 cache = caching_ctl->block_group;
4522                 if (block_group_cache_done(cache)) {
4523                         cache->last_byte_to_unpin = (u64)-1;
4524                         list_del_init(&caching_ctl->list);
4525                         put_caching_control(caching_ctl);
4526                 } else {
4527                         cache->last_byte_to_unpin = caching_ctl->progress;
4528                 }
4529         }
4530
4531         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4532                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4533         else
4534                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4535
4536         up_write(&fs_info->extent_commit_sem);
4537
4538         update_global_block_rsv(fs_info);
4539         return 0;
4540 }
4541
4542 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4543 {
4544         struct btrfs_fs_info *fs_info = root->fs_info;
4545         struct btrfs_block_group_cache *cache = NULL;
4546         u64 len;
4547
4548         while (start <= end) {
4549                 if (!cache ||
4550                     start >= cache->key.objectid + cache->key.offset) {
4551                         if (cache)
4552                                 btrfs_put_block_group(cache);
4553                         cache = btrfs_lookup_block_group(fs_info, start);
4554                         BUG_ON(!cache);
4555                 }
4556
4557                 len = cache->key.objectid + cache->key.offset - start;
4558                 len = min(len, end + 1 - start);
4559
4560                 if (start < cache->last_byte_to_unpin) {
4561                         len = min(len, cache->last_byte_to_unpin - start);
4562                         btrfs_add_free_space(cache, start, len);
4563                 }
4564
4565                 start += len;
4566
4567                 spin_lock(&cache->space_info->lock);
4568                 spin_lock(&cache->lock);
4569                 cache->pinned -= len;
4570                 cache->space_info->bytes_pinned -= len;
4571                 if (cache->ro)
4572                         cache->space_info->bytes_readonly += len;
4573                 spin_unlock(&cache->lock);
4574                 spin_unlock(&cache->space_info->lock);
4575         }
4576
4577         if (cache)
4578                 btrfs_put_block_group(cache);
4579         return 0;
4580 }
4581
4582 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4583                                struct btrfs_root *root)
4584 {
4585         struct btrfs_fs_info *fs_info = root->fs_info;
4586         struct extent_io_tree *unpin;
4587         u64 start;
4588         u64 end;
4589         int ret;
4590
4591         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4592                 unpin = &fs_info->freed_extents[1];
4593         else
4594                 unpin = &fs_info->freed_extents[0];
4595
4596         while (1) {
4597                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4598                                             EXTENT_DIRTY);
4599                 if (ret)
4600                         break;
4601
4602                 if (btrfs_test_opt(root, DISCARD))
4603                         ret = btrfs_discard_extent(root, start,
4604                                                    end + 1 - start, NULL);
4605
4606                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4607                 unpin_extent_range(root, start, end);
4608                 cond_resched();
4609         }
4610
4611         return 0;
4612 }
4613
4614 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4615                                 struct btrfs_root *root,
4616                                 u64 bytenr, u64 num_bytes, u64 parent,
4617                                 u64 root_objectid, u64 owner_objectid,
4618                                 u64 owner_offset, int refs_to_drop,
4619                                 struct btrfs_delayed_extent_op *extent_op)
4620 {
4621         struct btrfs_key key;
4622         struct btrfs_path *path;
4623         struct btrfs_fs_info *info = root->fs_info;
4624         struct btrfs_root *extent_root = info->extent_root;
4625         struct extent_buffer *leaf;
4626         struct btrfs_extent_item *ei;
4627         struct btrfs_extent_inline_ref *iref;
4628         int ret;
4629         int is_data;
4630         int extent_slot = 0;
4631         int found_extent = 0;
4632         int num_to_del = 1;
4633         u32 item_size;
4634         u64 refs;
4635
4636         path = btrfs_alloc_path();
4637         if (!path)
4638                 return -ENOMEM;
4639
4640         path->reada = 1;
4641         path->leave_spinning = 1;
4642
4643         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4644         BUG_ON(!is_data && refs_to_drop != 1);
4645
4646         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4647                                     bytenr, num_bytes, parent,
4648                                     root_objectid, owner_objectid,
4649                                     owner_offset);
4650         if (ret == 0) {
4651                 extent_slot = path->slots[0];
4652                 while (extent_slot >= 0) {
4653                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4654                                               extent_slot);
4655                         if (key.objectid != bytenr)
4656                                 break;
4657                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4658                             key.offset == num_bytes) {
4659                                 found_extent = 1;
4660                                 break;
4661                         }
4662                         if (path->slots[0] - extent_slot > 5)
4663                                 break;
4664                         extent_slot--;
4665                 }
4666 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4667                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4668                 if (found_extent && item_size < sizeof(*ei))
4669                         found_extent = 0;
4670 #endif
4671                 if (!found_extent) {
4672                         BUG_ON(iref);
4673                         ret = remove_extent_backref(trans, extent_root, path,
4674                                                     NULL, refs_to_drop,
4675                                                     is_data);
4676                         BUG_ON(ret);
4677                         btrfs_release_path(path);
4678                         path->leave_spinning = 1;
4679
4680                         key.objectid = bytenr;
4681                         key.type = BTRFS_EXTENT_ITEM_KEY;
4682                         key.offset = num_bytes;
4683
4684                         ret = btrfs_search_slot(trans, extent_root,
4685                                                 &key, path, -1, 1);
4686                         if (ret) {
4687                                 printk(KERN_ERR "umm, got %d back from search"
4688                                        ", was looking for %llu\n", ret,
4689                                        (unsigned long long)bytenr);
4690                                 if (ret > 0)
4691                                         btrfs_print_leaf(extent_root,
4692                                                          path->nodes[0]);
4693                         }
4694                         BUG_ON(ret);
4695                         extent_slot = path->slots[0];
4696                 }
4697         } else {
4698                 btrfs_print_leaf(extent_root, path->nodes[0]);
4699                 WARN_ON(1);
4700                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4701                        "parent %llu root %llu  owner %llu offset %llu\n",
4702                        (unsigned long long)bytenr,
4703                        (unsigned long long)parent,
4704                        (unsigned long long)root_objectid,
4705                        (unsigned long long)owner_objectid,
4706                        (unsigned long long)owner_offset);
4707         }
4708
4709         leaf = path->nodes[0];
4710         item_size = btrfs_item_size_nr(leaf, extent_slot);
4711 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4712         if (item_size < sizeof(*ei)) {
4713                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4714                 ret = convert_extent_item_v0(trans, extent_root, path,
4715                                              owner_objectid, 0);
4716                 BUG_ON(ret < 0);
4717
4718                 btrfs_release_path(path);
4719                 path->leave_spinning = 1;
4720
4721                 key.objectid = bytenr;
4722                 key.type = BTRFS_EXTENT_ITEM_KEY;
4723                 key.offset = num_bytes;
4724
4725                 ret = btrfs_search_slot(trans, extent_root, &key, path,
4726                                         -1, 1);
4727                 if (ret) {
4728                         printk(KERN_ERR "umm, got %d back from search"
4729                                ", was looking for %llu\n", ret,
4730                                (unsigned long long)bytenr);
4731                         btrfs_print_leaf(extent_root, path->nodes[0]);
4732                 }
4733                 BUG_ON(ret);
4734                 extent_slot = path->slots[0];
4735                 leaf = path->nodes[0];
4736                 item_size = btrfs_item_size_nr(leaf, extent_slot);
4737         }
4738 #endif
4739         BUG_ON(item_size < sizeof(*ei));
4740         ei = btrfs_item_ptr(leaf, extent_slot,
4741                             struct btrfs_extent_item);
4742         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
4743                 struct btrfs_tree_block_info *bi;
4744                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4745                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4746                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4747         }
4748
4749         refs = btrfs_extent_refs(leaf, ei);
4750         BUG_ON(refs < refs_to_drop);
4751         refs -= refs_to_drop;
4752
4753         if (refs > 0) {
4754                 if (extent_op)
4755                         __run_delayed_extent_op(extent_op, leaf, ei);
4756                 /*
4757                  * In the case of inline back ref, reference count will
4758                  * be updated by remove_extent_backref
4759                  */
4760                 if (iref) {
4761                         BUG_ON(!found_extent);
4762                 } else {
4763                         btrfs_set_extent_refs(leaf, ei, refs);
4764                         btrfs_mark_buffer_dirty(leaf);
4765                 }
4766                 if (found_extent) {
4767                         ret = remove_extent_backref(trans, extent_root, path,
4768                                                     iref, refs_to_drop,
4769                                                     is_data);
4770                         BUG_ON(ret);
4771                 }
4772         } else {
4773                 if (found_extent) {
4774                         BUG_ON(is_data && refs_to_drop !=
4775                                extent_data_ref_count(root, path, iref));
4776                         if (iref) {
4777                                 BUG_ON(path->slots[0] != extent_slot);
4778                         } else {
4779                                 BUG_ON(path->slots[0] != extent_slot + 1);
4780                                 path->slots[0] = extent_slot;
4781                                 num_to_del = 2;
4782                         }
4783                 }
4784
4785                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
4786                                       num_to_del);
4787                 BUG_ON(ret);
4788                 btrfs_release_path(path);
4789
4790                 if (is_data) {
4791                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
4792                         BUG_ON(ret);
4793                 } else {
4794                         invalidate_mapping_pages(info->btree_inode->i_mapping,
4795                              bytenr >> PAGE_CACHE_SHIFT,
4796                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
4797                 }
4798
4799                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
4800                 BUG_ON(ret);
4801         }
4802         btrfs_free_path(path);
4803         return ret;
4804 }
4805
4806 /*
4807  * when we free an block, it is possible (and likely) that we free the last
4808  * delayed ref for that extent as well.  This searches the delayed ref tree for
4809  * a given extent, and if there are no other delayed refs to be processed, it
4810  * removes it from the tree.
4811  */
4812 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
4813                                       struct btrfs_root *root, u64 bytenr)
4814 {
4815         struct btrfs_delayed_ref_head *head;
4816         struct btrfs_delayed_ref_root *delayed_refs;
4817         struct btrfs_delayed_ref_node *ref;
4818         struct rb_node *node;
4819         int ret = 0;
4820
4821         delayed_refs = &trans->transaction->delayed_refs;
4822         spin_lock(&delayed_refs->lock);
4823         head = btrfs_find_delayed_ref_head(trans, bytenr);
4824         if (!head)
4825                 goto out;
4826
4827         node = rb_prev(&head->node.rb_node);
4828         if (!node)
4829                 goto out;
4830
4831         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
4832
4833         /* there are still entries for this ref, we can't drop it */
4834         if (ref->bytenr == bytenr)
4835                 goto out;
4836
4837         if (head->extent_op) {
4838                 if (!head->must_insert_reserved)
4839                         goto out;
4840                 kfree(head->extent_op);
4841                 head->extent_op = NULL;
4842         }
4843
4844         /*
4845          * waiting for the lock here would deadlock.  If someone else has it
4846          * locked they are already in the process of dropping it anyway
4847          */
4848         if (!mutex_trylock(&head->mutex))
4849                 goto out;
4850
4851         /*
4852          * at this point we have a head with no other entries.  Go
4853          * ahead and process it.
4854          */
4855         head->node.in_tree = 0;
4856         rb_erase(&head->node.rb_node, &delayed_refs->root);
4857
4858         delayed_refs->num_entries--;
4859
4860         /*
4861          * we don't take a ref on the node because we're removing it from the
4862          * tree, so we just steal the ref the tree was holding.
4863          */
4864         delayed_refs->num_heads--;
4865         if (list_empty(&head->cluster))
4866                 delayed_refs->num_heads_ready--;
4867
4868         list_del_init(&head->cluster);
4869         spin_unlock(&delayed_refs->lock);
4870
4871         BUG_ON(head->extent_op);
4872         if (head->must_insert_reserved)
4873                 ret = 1;
4874
4875         mutex_unlock(&head->mutex);
4876         btrfs_put_delayed_ref(&head->node);
4877         return ret;
4878 out:
4879         spin_unlock(&delayed_refs->lock);
4880         return 0;
4881 }
4882
4883 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
4884                            struct btrfs_root *root,
4885                            struct extent_buffer *buf,
4886                            u64 parent, int last_ref)
4887 {
4888         struct btrfs_block_group_cache *cache = NULL;
4889         int ret;
4890
4891         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4892                 ret = btrfs_add_delayed_tree_ref(trans, buf->start, buf->len,
4893                                                 parent, root->root_key.objectid,
4894                                                 btrfs_header_level(buf),
4895                                                 BTRFS_DROP_DELAYED_REF, NULL);
4896                 BUG_ON(ret);
4897         }
4898
4899         if (!last_ref)
4900                 return;
4901
4902         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
4903
4904         if (btrfs_header_generation(buf) == trans->transid) {
4905                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4906                         ret = check_ref_cleanup(trans, root, buf->start);
4907                         if (!ret)
4908                                 goto out;
4909                 }
4910
4911                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
4912                         pin_down_extent(root, cache, buf->start, buf->len, 1);
4913                         goto out;
4914                 }
4915
4916                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
4917
4918                 btrfs_add_free_space(cache, buf->start, buf->len);
4919                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
4920         }
4921 out:
4922         /*
4923          * Deleting the buffer, clear the corrupt flag since it doesn't matter
4924          * anymore.
4925          */
4926         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
4927         btrfs_put_block_group(cache);
4928 }
4929
4930 int btrfs_free_extent(struct btrfs_trans_handle *trans,
4931                       struct btrfs_root *root,
4932                       u64 bytenr, u64 num_bytes, u64 parent,
4933                       u64 root_objectid, u64 owner, u64 offset)
4934 {
4935         int ret;
4936
4937         /*
4938          * tree log blocks never actually go into the extent allocation
4939          * tree, just update pinning info and exit early.
4940          */
4941         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
4942                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
4943                 /* unlocks the pinned mutex */
4944                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
4945                 ret = 0;
4946         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
4947                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
4948                                         parent, root_objectid, (int)owner,
4949                                         BTRFS_DROP_DELAYED_REF, NULL);
4950                 BUG_ON(ret);
4951         } else {
4952                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
4953                                         parent, root_objectid, owner,
4954                                         offset, BTRFS_DROP_DELAYED_REF, NULL);
4955                 BUG_ON(ret);
4956         }
4957         return ret;
4958 }
4959
4960 static u64 stripe_align(struct btrfs_root *root, u64 val)
4961 {
4962         u64 mask = ((u64)root->stripesize - 1);
4963         u64 ret = (val + mask) & ~mask;
4964         return ret;
4965 }
4966
4967 /*
4968  * when we wait for progress in the block group caching, its because
4969  * our allocation attempt failed at least once.  So, we must sleep
4970  * and let some progress happen before we try again.
4971  *
4972  * This function will sleep at least once waiting for new free space to
4973  * show up, and then it will check the block group free space numbers
4974  * for our min num_bytes.  Another option is to have it go ahead
4975  * and look in the rbtree for a free extent of a given size, but this
4976  * is a good start.
4977  */
4978 static noinline int
4979 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
4980                                 u64 num_bytes)
4981 {
4982         struct btrfs_caching_control *caching_ctl;
4983         DEFINE_WAIT(wait);
4984
4985         caching_ctl = get_caching_control(cache);
4986         if (!caching_ctl)
4987                 return 0;
4988
4989         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
4990                    (cache->free_space_ctl->free_space >= num_bytes));
4991
4992         put_caching_control(caching_ctl);
4993         return 0;
4994 }
4995
4996 static noinline int
4997 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
4998 {
4999         struct btrfs_caching_control *caching_ctl;
5000         DEFINE_WAIT(wait);
5001
5002         caching_ctl = get_caching_control(cache);
5003         if (!caching_ctl)
5004                 return 0;
5005
5006         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5007
5008         put_caching_control(caching_ctl);
5009         return 0;
5010 }
5011
5012 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5013 {
5014         int index;
5015         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
5016                 index = 0;
5017         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
5018                 index = 1;
5019         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
5020                 index = 2;
5021         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
5022                 index = 3;
5023         else
5024                 index = 4;
5025         return index;
5026 }
5027
5028 enum btrfs_loop_type {
5029         LOOP_FIND_IDEAL = 0,
5030         LOOP_CACHING_NOWAIT = 1,
5031         LOOP_CACHING_WAIT = 2,
5032         LOOP_ALLOC_CHUNK = 3,
5033         LOOP_NO_EMPTY_SIZE = 4,
5034 };
5035
5036 /*
5037  * walks the btree of allocated extents and find a hole of a given size.
5038  * The key ins is changed to record the hole:
5039  * ins->objectid == block start
5040  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5041  * ins->offset == number of blocks
5042  * Any available blocks before search_start are skipped.
5043  */
5044 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5045                                      struct btrfs_root *orig_root,
5046                                      u64 num_bytes, u64 empty_size,
5047                                      u64 search_start, u64 search_end,
5048                                      u64 hint_byte, struct btrfs_key *ins,
5049                                      u64 data)
5050 {
5051         int ret = 0;
5052         struct btrfs_root *root = orig_root->fs_info->extent_root;
5053         struct btrfs_free_cluster *last_ptr = NULL;
5054         struct btrfs_block_group_cache *block_group = NULL;
5055         int empty_cluster = 2 * 1024 * 1024;
5056         int allowed_chunk_alloc = 0;
5057         int done_chunk_alloc = 0;
5058         struct btrfs_space_info *space_info;
5059         int last_ptr_loop = 0;
5060         int loop = 0;
5061         int index = 0;
5062         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5063                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5064         bool found_uncached_bg = false;
5065         bool failed_cluster_refill = false;
5066         bool failed_alloc = false;
5067         bool use_cluster = true;
5068         bool have_caching_bg = false;
5069         u64 ideal_cache_percent = 0;
5070         u64 ideal_cache_offset = 0;
5071
5072         WARN_ON(num_bytes < root->sectorsize);
5073         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5074         ins->objectid = 0;
5075         ins->offset = 0;
5076
5077         space_info = __find_space_info(root->fs_info, data);
5078         if (!space_info) {
5079                 printk(KERN_ERR "No space info for %llu\n", data);
5080                 return -ENOSPC;
5081         }
5082
5083         /*
5084          * If the space info is for both data and metadata it means we have a
5085          * small filesystem and we can't use the clustering stuff.
5086          */
5087         if (btrfs_mixed_space_info(space_info))
5088                 use_cluster = false;
5089
5090         if (orig_root->ref_cows || empty_size)
5091                 allowed_chunk_alloc = 1;
5092
5093         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5094                 last_ptr = &root->fs_info->meta_alloc_cluster;
5095                 if (!btrfs_test_opt(root, SSD))
5096                         empty_cluster = 64 * 1024;
5097         }
5098
5099         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5100             btrfs_test_opt(root, SSD)) {
5101                 last_ptr = &root->fs_info->data_alloc_cluster;
5102         }
5103
5104         if (last_ptr) {
5105                 spin_lock(&last_ptr->lock);
5106                 if (last_ptr->block_group)
5107                         hint_byte = last_ptr->window_start;
5108                 spin_unlock(&last_ptr->lock);
5109         }
5110
5111         search_start = max(search_start, first_logical_byte(root, 0));
5112         search_start = max(search_start, hint_byte);
5113
5114         if (!last_ptr)
5115                 empty_cluster = 0;
5116
5117         if (search_start == hint_byte) {
5118 ideal_cache:
5119                 block_group = btrfs_lookup_block_group(root->fs_info,
5120                                                        search_start);
5121                 /*
5122                  * we don't want to use the block group if it doesn't match our
5123                  * allocation bits, or if its not cached.
5124                  *
5125                  * However if we are re-searching with an ideal block group
5126                  * picked out then we don't care that the block group is cached.
5127                  */
5128                 if (block_group && block_group_bits(block_group, data) &&
5129                     (block_group->cached != BTRFS_CACHE_NO ||
5130                      search_start == ideal_cache_offset)) {
5131                         down_read(&space_info->groups_sem);
5132                         if (list_empty(&block_group->list) ||
5133                             block_group->ro) {
5134                                 /*
5135                                  * someone is removing this block group,
5136                                  * we can't jump into the have_block_group
5137                                  * target because our list pointers are not
5138                                  * valid
5139                                  */
5140                                 btrfs_put_block_group(block_group);
5141                                 up_read(&space_info->groups_sem);
5142                         } else {
5143                                 index = get_block_group_index(block_group);
5144                                 goto have_block_group;
5145                         }
5146                 } else if (block_group) {
5147                         btrfs_put_block_group(block_group);
5148                 }
5149         }
5150 search:
5151         have_caching_bg = false;
5152         down_read(&space_info->groups_sem);
5153         list_for_each_entry(block_group, &space_info->block_groups[index],
5154                             list) {
5155                 u64 offset;
5156                 int cached;
5157
5158                 btrfs_get_block_group(block_group);
5159                 search_start = block_group->key.objectid;
5160
5161                 /*
5162                  * this can happen if we end up cycling through all the
5163                  * raid types, but we want to make sure we only allocate
5164                  * for the proper type.
5165                  */
5166                 if (!block_group_bits(block_group, data)) {
5167                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5168                                 BTRFS_BLOCK_GROUP_RAID1 |
5169                                 BTRFS_BLOCK_GROUP_RAID10;
5170
5171                         /*
5172                          * if they asked for extra copies and this block group
5173                          * doesn't provide them, bail.  This does allow us to
5174                          * fill raid0 from raid1.
5175                          */
5176                         if ((data & extra) && !(block_group->flags & extra))
5177                                 goto loop;
5178                 }
5179
5180 have_block_group:
5181                 if (unlikely(block_group->cached == BTRFS_CACHE_NO)) {
5182                         u64 free_percent;
5183
5184                         ret = cache_block_group(block_group, trans,
5185                                                 orig_root, 1);
5186                         if (block_group->cached == BTRFS_CACHE_FINISHED)
5187                                 goto have_block_group;
5188
5189                         free_percent = btrfs_block_group_used(&block_group->item);
5190                         free_percent *= 100;
5191                         free_percent = div64_u64(free_percent,
5192                                                  block_group->key.offset);
5193                         free_percent = 100 - free_percent;
5194                         if (free_percent > ideal_cache_percent &&
5195                             likely(!block_group->ro)) {
5196                                 ideal_cache_offset = block_group->key.objectid;
5197                                 ideal_cache_percent = free_percent;
5198                         }
5199
5200                         /*
5201                          * The caching workers are limited to 2 threads, so we
5202                          * can queue as much work as we care to.
5203                          */
5204                         if (loop > LOOP_FIND_IDEAL) {
5205                                 ret = cache_block_group(block_group, trans,
5206                                                         orig_root, 0);
5207                                 BUG_ON(ret);
5208                         }
5209                         found_uncached_bg = true;
5210
5211                         /*
5212                          * If loop is set for cached only, try the next block
5213                          * group.
5214                          */
5215                         if (loop == LOOP_FIND_IDEAL)
5216                                 goto loop;
5217                 }
5218
5219                 cached = block_group_cache_done(block_group);
5220                 if (unlikely(!cached))
5221                         found_uncached_bg = true;
5222
5223                 if (unlikely(block_group->ro))
5224                         goto loop;
5225
5226                 spin_lock(&block_group->free_space_ctl->tree_lock);
5227                 if (cached &&
5228                     block_group->free_space_ctl->free_space <
5229                     num_bytes + empty_size) {
5230                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5231                         goto loop;
5232                 }
5233                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5234
5235                 /*
5236                  * Ok we want to try and use the cluster allocator, so lets look
5237                  * there, unless we are on LOOP_NO_EMPTY_SIZE, since we will
5238                  * have tried the cluster allocator plenty of times at this
5239                  * point and not have found anything, so we are likely way too
5240                  * fragmented for the clustering stuff to find anything, so lets
5241                  * just skip it and let the allocator find whatever block it can
5242                  * find
5243                  */
5244                 if (last_ptr && loop < LOOP_NO_EMPTY_SIZE) {
5245                         /*
5246                          * the refill lock keeps out other
5247                          * people trying to start a new cluster
5248                          */
5249                         spin_lock(&last_ptr->refill_lock);
5250                         if (last_ptr->block_group &&
5251                             (last_ptr->block_group->ro ||
5252                             !block_group_bits(last_ptr->block_group, data))) {
5253                                 offset = 0;
5254                                 goto refill_cluster;
5255                         }
5256
5257                         offset = btrfs_alloc_from_cluster(block_group, last_ptr,
5258                                                  num_bytes, search_start);
5259                         if (offset) {
5260                                 /* we have a block, we're done */
5261                                 spin_unlock(&last_ptr->refill_lock);
5262                                 goto checks;
5263                         }
5264
5265                         spin_lock(&last_ptr->lock);
5266                         /*
5267                          * whoops, this cluster doesn't actually point to
5268                          * this block group.  Get a ref on the block
5269                          * group is does point to and try again
5270                          */
5271                         if (!last_ptr_loop && last_ptr->block_group &&
5272                             last_ptr->block_group != block_group &&
5273                             index <=
5274                                  get_block_group_index(last_ptr->block_group)) {
5275
5276                                 btrfs_put_block_group(block_group);
5277                                 block_group = last_ptr->block_group;
5278                                 btrfs_get_block_group(block_group);
5279                                 spin_unlock(&last_ptr->lock);
5280                                 spin_unlock(&last_ptr->refill_lock);
5281
5282                                 last_ptr_loop = 1;
5283                                 search_start = block_group->key.objectid;
5284                                 /*
5285                                  * we know this block group is properly
5286                                  * in the list because
5287                                  * btrfs_remove_block_group, drops the
5288                                  * cluster before it removes the block
5289                                  * group from the list
5290                                  */
5291                                 goto have_block_group;
5292                         }
5293                         spin_unlock(&last_ptr->lock);
5294 refill_cluster:
5295                         /*
5296                          * this cluster didn't work out, free it and
5297                          * start over
5298                          */
5299                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5300
5301                         last_ptr_loop = 0;
5302
5303                         /* allocate a cluster in this block group */
5304                         ret = btrfs_find_space_cluster(trans, root,
5305                                                block_group, last_ptr,
5306                                                offset, num_bytes,
5307                                                empty_cluster + empty_size);
5308                         if (ret == 0) {
5309                                 /*
5310                                  * now pull our allocation out of this
5311                                  * cluster
5312                                  */
5313                                 offset = btrfs_alloc_from_cluster(block_group,
5314                                                   last_ptr, num_bytes,
5315                                                   search_start);
5316                                 if (offset) {
5317                                         /* we found one, proceed */
5318                                         spin_unlock(&last_ptr->refill_lock);
5319                                         goto checks;
5320                                 }
5321                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5322                                    && !failed_cluster_refill) {
5323                                 spin_unlock(&last_ptr->refill_lock);
5324
5325                                 failed_cluster_refill = true;
5326                                 wait_block_group_cache_progress(block_group,
5327                                        num_bytes + empty_cluster + empty_size);
5328                                 goto have_block_group;
5329                         }
5330
5331                         /*
5332                          * at this point we either didn't find a cluster
5333                          * or we weren't able to allocate a block from our
5334                          * cluster.  Free the cluster we've been trying
5335                          * to use, and go to the next block group
5336                          */
5337                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5338                         spin_unlock(&last_ptr->refill_lock);
5339                         goto loop;
5340                 }
5341
5342                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5343                                                     num_bytes, empty_size);
5344                 /*
5345                  * If we didn't find a chunk, and we haven't failed on this
5346                  * block group before, and this block group is in the middle of
5347                  * caching and we are ok with waiting, then go ahead and wait
5348                  * for progress to be made, and set failed_alloc to true.
5349                  *
5350                  * If failed_alloc is true then we've already waited on this
5351                  * block group once and should move on to the next block group.
5352                  */
5353                 if (!offset && !failed_alloc && !cached &&
5354                     loop > LOOP_CACHING_NOWAIT) {
5355                         wait_block_group_cache_progress(block_group,
5356                                                 num_bytes + empty_size);
5357                         failed_alloc = true;
5358                         goto have_block_group;
5359                 } else if (!offset) {
5360                         if (!cached)
5361                                 have_caching_bg = true;
5362                         goto loop;
5363                 }
5364 checks:
5365                 search_start = stripe_align(root, offset);
5366                 /* move on to the next group */
5367                 if (search_start + num_bytes >= search_end) {
5368                         btrfs_add_free_space(block_group, offset, num_bytes);
5369                         goto loop;
5370                 }
5371
5372                 /* move on to the next group */
5373                 if (search_start + num_bytes >
5374                     block_group->key.objectid + block_group->key.offset) {
5375                         btrfs_add_free_space(block_group, offset, num_bytes);
5376                         goto loop;
5377                 }
5378
5379                 ins->objectid = search_start;
5380                 ins->offset = num_bytes;
5381
5382                 if (offset < search_start)
5383                         btrfs_add_free_space(block_group, offset,
5384                                              search_start - offset);
5385                 BUG_ON(offset > search_start);
5386
5387                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
5388                                                   alloc_type);
5389                 if (ret == -EAGAIN) {
5390                         btrfs_add_free_space(block_group, offset, num_bytes);
5391                         goto loop;
5392                 }
5393
5394                 /* we are all good, lets return */
5395                 ins->objectid = search_start;
5396                 ins->offset = num_bytes;
5397
5398                 if (offset < search_start)
5399                         btrfs_add_free_space(block_group, offset,
5400                                              search_start - offset);
5401                 BUG_ON(offset > search_start);
5402                 btrfs_put_block_group(block_group);
5403                 break;
5404 loop:
5405                 failed_cluster_refill = false;
5406                 failed_alloc = false;
5407                 BUG_ON(index != get_block_group_index(block_group));
5408                 btrfs_put_block_group(block_group);
5409         }
5410         up_read(&space_info->groups_sem);
5411
5412         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
5413                 goto search;
5414
5415         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5416                 goto search;
5417
5418         /* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
5419          *                      for them to make caching progress.  Also
5420          *                      determine the best possible bg to cache
5421          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5422          *                      caching kthreads as we move along
5423          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5424          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5425          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5426          *                      again
5427          */
5428         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5429                 index = 0;
5430                 if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
5431                         found_uncached_bg = false;
5432                         loop++;
5433                         if (!ideal_cache_percent)
5434                                 goto search;
5435
5436                         /*
5437                          * 1 of the following 2 things have happened so far
5438                          *
5439                          * 1) We found an ideal block group for caching that
5440                          * is mostly full and will cache quickly, so we might
5441                          * as well wait for it.
5442                          *
5443                          * 2) We searched for cached only and we didn't find
5444                          * anything, and we didn't start any caching kthreads
5445                          * either, so chances are we will loop through and
5446                          * start a couple caching kthreads, and then come back
5447                          * around and just wait for them.  This will be slower
5448                          * because we will have 2 caching kthreads reading at
5449                          * the same time when we could have just started one
5450                          * and waited for it to get far enough to give us an
5451                          * allocation, so go ahead and go to the wait caching
5452                          * loop.
5453                          */
5454                         loop = LOOP_CACHING_WAIT;
5455                         search_start = ideal_cache_offset;
5456                         ideal_cache_percent = 0;
5457                         goto ideal_cache;
5458                 } else if (loop == LOOP_FIND_IDEAL) {
5459                         /*
5460                          * Didn't find a uncached bg, wait on anything we find
5461                          * next.
5462                          */
5463                         loop = LOOP_CACHING_WAIT;
5464                         goto search;
5465                 }
5466
5467                 loop++;
5468
5469                 if (loop == LOOP_ALLOC_CHUNK) {
5470                        if (allowed_chunk_alloc) {
5471                                 ret = do_chunk_alloc(trans, root, num_bytes +
5472                                                      2 * 1024 * 1024, data,
5473                                                      CHUNK_ALLOC_LIMITED);
5474                                 allowed_chunk_alloc = 0;
5475                                 if (ret == 1)
5476                                         done_chunk_alloc = 1;
5477                         } else if (!done_chunk_alloc &&
5478                                    space_info->force_alloc ==
5479                                    CHUNK_ALLOC_NO_FORCE) {
5480                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5481                         }
5482
5483                        /*
5484                         * We didn't allocate a chunk, go ahead and drop the
5485                         * empty size and loop again.
5486                         */
5487                        if (!done_chunk_alloc)
5488                                loop = LOOP_NO_EMPTY_SIZE;
5489                 }
5490
5491                 if (loop == LOOP_NO_EMPTY_SIZE) {
5492                         empty_size = 0;
5493                         empty_cluster = 0;
5494                 }
5495
5496                 goto search;
5497         } else if (!ins->objectid) {
5498                 ret = -ENOSPC;
5499         } else if (ins->objectid) {
5500                 ret = 0;
5501         }
5502
5503         return ret;
5504 }
5505
5506 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5507                             int dump_block_groups)
5508 {
5509         struct btrfs_block_group_cache *cache;
5510         int index = 0;
5511
5512         spin_lock(&info->lock);
5513         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5514                (unsigned long long)info->flags,
5515                (unsigned long long)(info->total_bytes - info->bytes_used -
5516                                     info->bytes_pinned - info->bytes_reserved -
5517                                     info->bytes_readonly),
5518                (info->full) ? "" : "not ");
5519         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5520                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5521                (unsigned long long)info->total_bytes,
5522                (unsigned long long)info->bytes_used,
5523                (unsigned long long)info->bytes_pinned,
5524                (unsigned long long)info->bytes_reserved,
5525                (unsigned long long)info->bytes_may_use,
5526                (unsigned long long)info->bytes_readonly);
5527         spin_unlock(&info->lock);
5528
5529         if (!dump_block_groups)
5530                 return;
5531
5532         down_read(&info->groups_sem);
5533 again:
5534         list_for_each_entry(cache, &info->block_groups[index], list) {
5535                 spin_lock(&cache->lock);
5536                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5537                        "%llu pinned %llu reserved\n",
5538                        (unsigned long long)cache->key.objectid,
5539                        (unsigned long long)cache->key.offset,
5540                        (unsigned long long)btrfs_block_group_used(&cache->item),
5541                        (unsigned long long)cache->pinned,
5542                        (unsigned long long)cache->reserved);
5543                 btrfs_dump_free_space(cache, bytes);
5544                 spin_unlock(&cache->lock);
5545         }
5546         if (++index < BTRFS_NR_RAID_TYPES)
5547                 goto again;
5548         up_read(&info->groups_sem);
5549 }
5550
5551 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5552                          struct btrfs_root *root,
5553                          u64 num_bytes, u64 min_alloc_size,
5554                          u64 empty_size, u64 hint_byte,
5555                          u64 search_end, struct btrfs_key *ins,
5556                          u64 data)
5557 {
5558         int ret;
5559         u64 search_start = 0;
5560
5561         data = btrfs_get_alloc_profile(root, data);
5562 again:
5563         /*
5564          * the only place that sets empty_size is btrfs_realloc_node, which
5565          * is not called recursively on allocations
5566          */
5567         if (empty_size || root->ref_cows)
5568                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5569                                      num_bytes + 2 * 1024 * 1024, data,
5570                                      CHUNK_ALLOC_NO_FORCE);
5571
5572         WARN_ON(num_bytes < root->sectorsize);
5573         ret = find_free_extent(trans, root, num_bytes, empty_size,
5574                                search_start, search_end, hint_byte,
5575                                ins, data);
5576
5577         if (ret == -ENOSPC && num_bytes > min_alloc_size) {
5578                 num_bytes = num_bytes >> 1;
5579                 num_bytes = num_bytes & ~(root->sectorsize - 1);
5580                 num_bytes = max(num_bytes, min_alloc_size);
5581                 do_chunk_alloc(trans, root->fs_info->extent_root,
5582                                num_bytes, data, CHUNK_ALLOC_FORCE);
5583                 goto again;
5584         }
5585         if (ret == -ENOSPC && btrfs_test_opt(root, ENOSPC_DEBUG)) {
5586                 struct btrfs_space_info *sinfo;
5587
5588                 sinfo = __find_space_info(root->fs_info, data);
5589                 printk(KERN_ERR "btrfs allocation failed flags %llu, "
5590                        "wanted %llu\n", (unsigned long long)data,
5591                        (unsigned long long)num_bytes);
5592                 dump_space_info(sinfo, num_bytes, 1);
5593         }
5594
5595         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
5596
5597         return ret;
5598 }
5599
5600 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
5601                                         u64 start, u64 len, int pin)
5602 {
5603         struct btrfs_block_group_cache *cache;
5604         int ret = 0;
5605
5606         cache = btrfs_lookup_block_group(root->fs_info, start);
5607         if (!cache) {
5608                 printk(KERN_ERR "Unable to find block group for %llu\n",
5609                        (unsigned long long)start);
5610                 return -ENOSPC;
5611         }
5612
5613         if (btrfs_test_opt(root, DISCARD))
5614                 ret = btrfs_discard_extent(root, start, len, NULL);
5615
5616         if (pin)
5617                 pin_down_extent(root, cache, start, len, 1);
5618         else {
5619                 btrfs_add_free_space(cache, start, len);
5620                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
5621         }
5622         btrfs_put_block_group(cache);
5623
5624         trace_btrfs_reserved_extent_free(root, start, len);
5625
5626         return ret;
5627 }
5628
5629 int btrfs_free_reserved_extent(struct btrfs_root *root,
5630                                         u64 start, u64 len)
5631 {
5632         return __btrfs_free_reserved_extent(root, start, len, 0);
5633 }
5634
5635 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
5636                                        u64 start, u64 len)
5637 {
5638         return __btrfs_free_reserved_extent(root, start, len, 1);
5639 }
5640
5641 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5642                                       struct btrfs_root *root,
5643                                       u64 parent, u64 root_objectid,
5644                                       u64 flags, u64 owner, u64 offset,
5645                                       struct btrfs_key *ins, int ref_mod)
5646 {
5647         int ret;
5648         struct btrfs_fs_info *fs_info = root->fs_info;
5649         struct btrfs_extent_item *extent_item;
5650         struct btrfs_extent_inline_ref *iref;
5651         struct btrfs_path *path;
5652         struct extent_buffer *leaf;
5653         int type;
5654         u32 size;
5655
5656         if (parent > 0)
5657                 type = BTRFS_SHARED_DATA_REF_KEY;
5658         else
5659                 type = BTRFS_EXTENT_DATA_REF_KEY;
5660
5661         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5662
5663         path = btrfs_alloc_path();
5664         if (!path)
5665                 return -ENOMEM;
5666
5667         path->leave_spinning = 1;
5668         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5669                                       ins, size);
5670         BUG_ON(ret);
5671
5672         leaf = path->nodes[0];
5673         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5674                                      struct btrfs_extent_item);
5675         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
5676         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5677         btrfs_set_extent_flags(leaf, extent_item,
5678                                flags | BTRFS_EXTENT_FLAG_DATA);
5679
5680         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
5681         btrfs_set_extent_inline_ref_type(leaf, iref, type);
5682         if (parent > 0) {
5683                 struct btrfs_shared_data_ref *ref;
5684                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
5685                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5686                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
5687         } else {
5688                 struct btrfs_extent_data_ref *ref;
5689                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
5690                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
5691                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
5692                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
5693                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
5694         }
5695
5696         btrfs_mark_buffer_dirty(path->nodes[0]);
5697         btrfs_free_path(path);
5698
5699         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5700         if (ret) {
5701                 printk(KERN_ERR "btrfs update block group failed for %llu "
5702                        "%llu\n", (unsigned long long)ins->objectid,
5703                        (unsigned long long)ins->offset);
5704                 BUG();
5705         }
5706         return ret;
5707 }
5708
5709 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
5710                                      struct btrfs_root *root,
5711                                      u64 parent, u64 root_objectid,
5712                                      u64 flags, struct btrfs_disk_key *key,
5713                                      int level, struct btrfs_key *ins)
5714 {
5715         int ret;
5716         struct btrfs_fs_info *fs_info = root->fs_info;
5717         struct btrfs_extent_item *extent_item;
5718         struct btrfs_tree_block_info *block_info;
5719         struct btrfs_extent_inline_ref *iref;
5720         struct btrfs_path *path;
5721         struct extent_buffer *leaf;
5722         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
5723
5724         path = btrfs_alloc_path();
5725         if (!path)
5726                 return -ENOMEM;
5727
5728         path->leave_spinning = 1;
5729         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5730                                       ins, size);
5731         BUG_ON(ret);
5732
5733         leaf = path->nodes[0];
5734         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5735                                      struct btrfs_extent_item);
5736         btrfs_set_extent_refs(leaf, extent_item, 1);
5737         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5738         btrfs_set_extent_flags(leaf, extent_item,
5739                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
5740         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
5741
5742         btrfs_set_tree_block_key(leaf, block_info, key);
5743         btrfs_set_tree_block_level(leaf, block_info, level);
5744
5745         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
5746         if (parent > 0) {
5747                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
5748                 btrfs_set_extent_inline_ref_type(leaf, iref,
5749                                                  BTRFS_SHARED_BLOCK_REF_KEY);
5750                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5751         } else {
5752                 btrfs_set_extent_inline_ref_type(leaf, iref,
5753                                                  BTRFS_TREE_BLOCK_REF_KEY);
5754                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
5755         }
5756
5757         btrfs_mark_buffer_dirty(leaf);
5758         btrfs_free_path(path);
5759
5760         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5761         if (ret) {
5762                 printk(KERN_ERR "btrfs update block group failed for %llu "
5763                        "%llu\n", (unsigned long long)ins->objectid,
5764                        (unsigned long long)ins->offset);
5765                 BUG();
5766         }
5767         return ret;
5768 }
5769
5770 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5771                                      struct btrfs_root *root,
5772                                      u64 root_objectid, u64 owner,
5773                                      u64 offset, struct btrfs_key *ins)
5774 {
5775         int ret;
5776
5777         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
5778
5779         ret = btrfs_add_delayed_data_ref(trans, ins->objectid, ins->offset,
5780                                          0, root_objectid, owner, offset,
5781                                          BTRFS_ADD_DELAYED_EXTENT, NULL);
5782         return ret;
5783 }
5784
5785 /*
5786  * this is used by the tree logging recovery code.  It records that
5787  * an extent has been allocated and makes sure to clear the free
5788  * space cache bits as well
5789  */
5790 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
5791                                    struct btrfs_root *root,
5792                                    u64 root_objectid, u64 owner, u64 offset,
5793                                    struct btrfs_key *ins)
5794 {
5795         int ret;
5796         struct btrfs_block_group_cache *block_group;
5797         struct btrfs_caching_control *caching_ctl;
5798         u64 start = ins->objectid;
5799         u64 num_bytes = ins->offset;
5800
5801         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
5802         cache_block_group(block_group, trans, NULL, 0);
5803         caching_ctl = get_caching_control(block_group);
5804
5805         if (!caching_ctl) {
5806                 BUG_ON(!block_group_cache_done(block_group));
5807                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5808                 BUG_ON(ret);
5809         } else {
5810                 mutex_lock(&caching_ctl->mutex);
5811
5812                 if (start >= caching_ctl->progress) {
5813                         ret = add_excluded_extent(root, start, num_bytes);
5814                         BUG_ON(ret);
5815                 } else if (start + num_bytes <= caching_ctl->progress) {
5816                         ret = btrfs_remove_free_space(block_group,
5817                                                       start, num_bytes);
5818                         BUG_ON(ret);
5819                 } else {
5820                         num_bytes = caching_ctl->progress - start;
5821                         ret = btrfs_remove_free_space(block_group,
5822                                                       start, num_bytes);
5823                         BUG_ON(ret);
5824
5825                         start = caching_ctl->progress;
5826                         num_bytes = ins->objectid + ins->offset -
5827                                     caching_ctl->progress;
5828                         ret = add_excluded_extent(root, start, num_bytes);
5829                         BUG_ON(ret);
5830                 }
5831
5832                 mutex_unlock(&caching_ctl->mutex);
5833                 put_caching_control(caching_ctl);
5834         }
5835
5836         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
5837                                           RESERVE_ALLOC_NO_ACCOUNT);
5838         BUG_ON(ret);
5839         btrfs_put_block_group(block_group);
5840         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
5841                                          0, owner, offset, ins, 1);
5842         return ret;
5843 }
5844
5845 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
5846                                             struct btrfs_root *root,
5847                                             u64 bytenr, u32 blocksize,
5848                                             int level)
5849 {
5850         struct extent_buffer *buf;
5851
5852         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
5853         if (!buf)
5854                 return ERR_PTR(-ENOMEM);
5855         btrfs_set_header_generation(buf, trans->transid);
5856         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
5857         btrfs_tree_lock(buf);
5858         clean_tree_block(trans, root, buf);
5859
5860         btrfs_set_lock_blocking(buf);
5861         btrfs_set_buffer_uptodate(buf);
5862
5863         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
5864                 /*
5865                  * we allow two log transactions at a time, use different
5866                  * EXENT bit to differentiate dirty pages.
5867                  */
5868                 if (root->log_transid % 2 == 0)
5869                         set_extent_dirty(&root->dirty_log_pages, buf->start,
5870                                         buf->start + buf->len - 1, GFP_NOFS);
5871                 else
5872                         set_extent_new(&root->dirty_log_pages, buf->start,
5873                                         buf->start + buf->len - 1, GFP_NOFS);
5874         } else {
5875                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
5876                          buf->start + buf->len - 1, GFP_NOFS);
5877         }
5878         trans->blocks_used++;
5879         /* this returns a buffer locked for blocking */
5880         return buf;
5881 }
5882
5883 static struct btrfs_block_rsv *
5884 use_block_rsv(struct btrfs_trans_handle *trans,
5885               struct btrfs_root *root, u32 blocksize)
5886 {
5887         struct btrfs_block_rsv *block_rsv;
5888         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5889         int ret;
5890
5891         block_rsv = get_block_rsv(trans, root);
5892
5893         if (block_rsv->size == 0) {
5894                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
5895                 /*
5896                  * If we couldn't reserve metadata bytes try and use some from
5897                  * the global reserve.
5898                  */
5899                 if (ret && block_rsv != global_rsv) {
5900                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5901                         if (!ret)
5902                                 return global_rsv;
5903                         return ERR_PTR(ret);
5904                 } else if (ret) {
5905                         return ERR_PTR(ret);
5906                 }
5907                 return block_rsv;
5908         }
5909
5910         ret = block_rsv_use_bytes(block_rsv, blocksize);
5911         if (!ret)
5912                 return block_rsv;
5913         if (ret) {
5914                 static DEFINE_RATELIMIT_STATE(_rs,
5915                                 DEFAULT_RATELIMIT_INTERVAL,
5916                                 /*DEFAULT_RATELIMIT_BURST*/ 2);
5917                 if (__ratelimit(&_rs)) {
5918                         printk(KERN_DEBUG "btrfs: block rsv returned %d\n", ret);
5919                         WARN_ON(1);
5920                 }
5921                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
5922                 if (!ret) {
5923                         return block_rsv;
5924                 } else if (ret && block_rsv != global_rsv) {
5925                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5926                         if (!ret)
5927                                 return global_rsv;
5928                 }
5929         }
5930
5931         return ERR_PTR(-ENOSPC);
5932 }
5933
5934 static void unuse_block_rsv(struct btrfs_block_rsv *block_rsv, u32 blocksize)
5935 {
5936         block_rsv_add_bytes(block_rsv, blocksize, 0);
5937         block_rsv_release_bytes(block_rsv, NULL, 0);
5938 }
5939
5940 /*
5941  * finds a free extent and does all the dirty work required for allocation
5942  * returns the key for the extent through ins, and a tree buffer for
5943  * the first block of the extent through buf.
5944  *
5945  * returns the tree buffer or NULL.
5946  */
5947 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
5948                                         struct btrfs_root *root, u32 blocksize,
5949                                         u64 parent, u64 root_objectid,
5950                                         struct btrfs_disk_key *key, int level,
5951                                         u64 hint, u64 empty_size)
5952 {
5953         struct btrfs_key ins;
5954         struct btrfs_block_rsv *block_rsv;
5955         struct extent_buffer *buf;
5956         u64 flags = 0;
5957         int ret;
5958
5959
5960         block_rsv = use_block_rsv(trans, root, blocksize);
5961         if (IS_ERR(block_rsv))
5962                 return ERR_CAST(block_rsv);
5963
5964         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
5965                                    empty_size, hint, (u64)-1, &ins, 0);
5966         if (ret) {
5967                 unuse_block_rsv(block_rsv, blocksize);
5968                 return ERR_PTR(ret);
5969         }
5970
5971         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
5972                                     blocksize, level);
5973         BUG_ON(IS_ERR(buf));
5974
5975         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
5976                 if (parent == 0)
5977                         parent = ins.objectid;
5978                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
5979         } else
5980                 BUG_ON(parent > 0);
5981
5982         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
5983                 struct btrfs_delayed_extent_op *extent_op;
5984                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
5985                 BUG_ON(!extent_op);
5986                 if (key)
5987                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
5988                 else
5989                         memset(&extent_op->key, 0, sizeof(extent_op->key));
5990                 extent_op->flags_to_set = flags;
5991                 extent_op->update_key = 1;
5992                 extent_op->update_flags = 1;
5993                 extent_op->is_data = 0;
5994
5995                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
5996                                         ins.offset, parent, root_objectid,
5997                                         level, BTRFS_ADD_DELAYED_EXTENT,
5998                                         extent_op);
5999                 BUG_ON(ret);
6000         }
6001         return buf;
6002 }
6003
6004 struct walk_control {
6005         u64 refs[BTRFS_MAX_LEVEL];
6006         u64 flags[BTRFS_MAX_LEVEL];
6007         struct btrfs_key update_progress;
6008         int stage;
6009         int level;
6010         int shared_level;
6011         int update_ref;
6012         int keep_locks;
6013         int reada_slot;
6014         int reada_count;
6015 };
6016
6017 #define DROP_REFERENCE  1
6018 #define UPDATE_BACKREF  2
6019
6020 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6021                                      struct btrfs_root *root,
6022                                      struct walk_control *wc,
6023                                      struct btrfs_path *path)
6024 {
6025         u64 bytenr;
6026         u64 generation;
6027         u64 refs;
6028         u64 flags;
6029         u32 nritems;
6030         u32 blocksize;
6031         struct btrfs_key key;
6032         struct extent_buffer *eb;
6033         int ret;
6034         int slot;
6035         int nread = 0;
6036
6037         if (path->slots[wc->level] < wc->reada_slot) {
6038                 wc->reada_count = wc->reada_count * 2 / 3;
6039                 wc->reada_count = max(wc->reada_count, 2);
6040         } else {
6041                 wc->reada_count = wc->reada_count * 3 / 2;
6042                 wc->reada_count = min_t(int, wc->reada_count,
6043                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6044         }
6045
6046         eb = path->nodes[wc->level];
6047         nritems = btrfs_header_nritems(eb);
6048         blocksize = btrfs_level_size(root, wc->level - 1);
6049
6050         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6051                 if (nread >= wc->reada_count)
6052                         break;
6053
6054                 cond_resched();
6055                 bytenr = btrfs_node_blockptr(eb, slot);
6056                 generation = btrfs_node_ptr_generation(eb, slot);
6057
6058                 if (slot == path->slots[wc->level])
6059                         goto reada;
6060
6061                 if (wc->stage == UPDATE_BACKREF &&
6062                     generation <= root->root_key.offset)
6063                         continue;
6064
6065                 /* We don't lock the tree block, it's OK to be racy here */
6066                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6067                                                &refs, &flags);
6068                 BUG_ON(ret);
6069                 BUG_ON(refs == 0);
6070
6071                 if (wc->stage == DROP_REFERENCE) {
6072                         if (refs == 1)
6073                                 goto reada;
6074
6075                         if (wc->level == 1 &&
6076                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6077                                 continue;
6078                         if (!wc->update_ref ||
6079                             generation <= root->root_key.offset)
6080                                 continue;
6081                         btrfs_node_key_to_cpu(eb, &key, slot);
6082                         ret = btrfs_comp_cpu_keys(&key,
6083                                                   &wc->update_progress);
6084                         if (ret < 0)
6085                                 continue;
6086                 } else {
6087                         if (wc->level == 1 &&
6088                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6089                                 continue;
6090                 }
6091 reada:
6092                 ret = readahead_tree_block(root, bytenr, blocksize,
6093                                            generation);
6094                 if (ret)
6095                         break;
6096                 nread++;
6097         }
6098         wc->reada_slot = slot;
6099 }
6100
6101 /*
6102  * hepler to process tree block while walking down the tree.
6103  *
6104  * when wc->stage == UPDATE_BACKREF, this function updates
6105  * back refs for pointers in the block.
6106  *
6107  * NOTE: return value 1 means we should stop walking down.
6108  */
6109 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6110                                    struct btrfs_root *root,
6111                                    struct btrfs_path *path,
6112                                    struct walk_control *wc, int lookup_info)
6113 {
6114         int level = wc->level;
6115         struct extent_buffer *eb = path->nodes[level];
6116         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6117         int ret;
6118
6119         if (wc->stage == UPDATE_BACKREF &&
6120             btrfs_header_owner(eb) != root->root_key.objectid)
6121                 return 1;
6122
6123         /*
6124          * when reference count of tree block is 1, it won't increase
6125          * again. once full backref flag is set, we never clear it.
6126          */
6127         if (lookup_info &&
6128             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6129              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6130                 BUG_ON(!path->locks[level]);
6131                 ret = btrfs_lookup_extent_info(trans, root,
6132                                                eb->start, eb->len,
6133                                                &wc->refs[level],
6134                                                &wc->flags[level]);
6135                 BUG_ON(ret);
6136                 BUG_ON(wc->refs[level] == 0);
6137         }
6138
6139         if (wc->stage == DROP_REFERENCE) {
6140                 if (wc->refs[level] > 1)
6141                         return 1;
6142
6143                 if (path->locks[level] && !wc->keep_locks) {
6144                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6145                         path->locks[level] = 0;
6146                 }
6147                 return 0;
6148         }
6149
6150         /* wc->stage == UPDATE_BACKREF */
6151         if (!(wc->flags[level] & flag)) {
6152                 BUG_ON(!path->locks[level]);
6153                 ret = btrfs_inc_ref(trans, root, eb, 1);
6154                 BUG_ON(ret);
6155                 ret = btrfs_dec_ref(trans, root, eb, 0);
6156                 BUG_ON(ret);
6157                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6158                                                   eb->len, flag, 0);
6159                 BUG_ON(ret);
6160                 wc->flags[level] |= flag;
6161         }
6162
6163         /*
6164          * the block is shared by multiple trees, so it's not good to
6165          * keep the tree lock
6166          */
6167         if (path->locks[level] && level > 0) {
6168                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6169                 path->locks[level] = 0;
6170         }
6171         return 0;
6172 }
6173
6174 /*
6175  * hepler to process tree block pointer.
6176  *
6177  * when wc->stage == DROP_REFERENCE, this function checks
6178  * reference count of the block pointed to. if the block
6179  * is shared and we need update back refs for the subtree
6180  * rooted at the block, this function changes wc->stage to
6181  * UPDATE_BACKREF. if the block is shared and there is no
6182  * need to update back, this function drops the reference
6183  * to the block.
6184  *
6185  * NOTE: return value 1 means we should stop walking down.
6186  */
6187 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6188                                  struct btrfs_root *root,
6189                                  struct btrfs_path *path,
6190                                  struct walk_control *wc, int *lookup_info)
6191 {
6192         u64 bytenr;
6193         u64 generation;
6194         u64 parent;
6195         u32 blocksize;
6196         struct btrfs_key key;
6197         struct extent_buffer *next;
6198         int level = wc->level;
6199         int reada = 0;
6200         int ret = 0;
6201
6202         generation = btrfs_node_ptr_generation(path->nodes[level],
6203                                                path->slots[level]);
6204         /*
6205          * if the lower level block was created before the snapshot
6206          * was created, we know there is no need to update back refs
6207          * for the subtree
6208          */
6209         if (wc->stage == UPDATE_BACKREF &&
6210             generation <= root->root_key.offset) {
6211                 *lookup_info = 1;
6212                 return 1;
6213         }
6214
6215         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6216         blocksize = btrfs_level_size(root, level - 1);
6217
6218         next = btrfs_find_tree_block(root, bytenr, blocksize);
6219         if (!next) {
6220                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6221                 if (!next)
6222                         return -ENOMEM;
6223                 reada = 1;
6224         }
6225         btrfs_tree_lock(next);
6226         btrfs_set_lock_blocking(next);
6227
6228         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6229                                        &wc->refs[level - 1],
6230                                        &wc->flags[level - 1]);
6231         BUG_ON(ret);
6232         BUG_ON(wc->refs[level - 1] == 0);
6233         *lookup_info = 0;
6234
6235         if (wc->stage == DROP_REFERENCE) {
6236                 if (wc->refs[level - 1] > 1) {
6237                         if (level == 1 &&
6238                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6239                                 goto skip;
6240
6241                         if (!wc->update_ref ||
6242                             generation <= root->root_key.offset)
6243                                 goto skip;
6244
6245                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6246                                               path->slots[level]);
6247                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6248                         if (ret < 0)
6249                                 goto skip;
6250
6251                         wc->stage = UPDATE_BACKREF;
6252                         wc->shared_level = level - 1;
6253                 }
6254         } else {
6255                 if (level == 1 &&
6256                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6257                         goto skip;
6258         }
6259
6260         if (!btrfs_buffer_uptodate(next, generation)) {
6261                 btrfs_tree_unlock(next);
6262                 free_extent_buffer(next);
6263                 next = NULL;
6264                 *lookup_info = 1;
6265         }
6266
6267         if (!next) {
6268                 if (reada && level == 1)
6269                         reada_walk_down(trans, root, wc, path);
6270                 next = read_tree_block(root, bytenr, blocksize, generation);
6271                 if (!next)
6272                         return -EIO;
6273                 btrfs_tree_lock(next);
6274                 btrfs_set_lock_blocking(next);
6275         }
6276
6277         level--;
6278         BUG_ON(level != btrfs_header_level(next));
6279         path->nodes[level] = next;
6280         path->slots[level] = 0;
6281         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6282         wc->level = level;
6283         if (wc->level == 1)
6284                 wc->reada_slot = 0;
6285         return 0;
6286 skip:
6287         wc->refs[level - 1] = 0;
6288         wc->flags[level - 1] = 0;
6289         if (wc->stage == DROP_REFERENCE) {
6290                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6291                         parent = path->nodes[level]->start;
6292                 } else {
6293                         BUG_ON(root->root_key.objectid !=
6294                                btrfs_header_owner(path->nodes[level]));
6295                         parent = 0;
6296                 }
6297
6298                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6299                                         root->root_key.objectid, level - 1, 0);
6300                 BUG_ON(ret);
6301         }
6302         btrfs_tree_unlock(next);
6303         free_extent_buffer(next);
6304         *lookup_info = 1;
6305         return 1;
6306 }
6307
6308 /*
6309  * hepler to process tree block while walking up the tree.
6310  *
6311  * when wc->stage == DROP_REFERENCE, this function drops
6312  * reference count on the block.
6313  *
6314  * when wc->stage == UPDATE_BACKREF, this function changes
6315  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6316  * to UPDATE_BACKREF previously while processing the block.
6317  *
6318  * NOTE: return value 1 means we should stop walking up.
6319  */
6320 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6321                                  struct btrfs_root *root,
6322                                  struct btrfs_path *path,
6323                                  struct walk_control *wc)
6324 {
6325         int ret;
6326         int level = wc->level;
6327         struct extent_buffer *eb = path->nodes[level];
6328         u64 parent = 0;
6329
6330         if (wc->stage == UPDATE_BACKREF) {
6331                 BUG_ON(wc->shared_level < level);
6332                 if (level < wc->shared_level)
6333                         goto out;
6334
6335                 ret = find_next_key(path, level + 1, &wc->update_progress);
6336                 if (ret > 0)
6337                         wc->update_ref = 0;
6338
6339                 wc->stage = DROP_REFERENCE;
6340                 wc->shared_level = -1;
6341                 path->slots[level] = 0;
6342
6343                 /*
6344                  * check reference count again if the block isn't locked.
6345                  * we should start walking down the tree again if reference
6346                  * count is one.
6347                  */
6348                 if (!path->locks[level]) {
6349                         BUG_ON(level == 0);
6350                         btrfs_tree_lock(eb);
6351                         btrfs_set_lock_blocking(eb);
6352                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6353
6354                         ret = btrfs_lookup_extent_info(trans, root,
6355                                                        eb->start, eb->len,
6356                                                        &wc->refs[level],
6357                                                        &wc->flags[level]);
6358                         BUG_ON(ret);
6359                         BUG_ON(wc->refs[level] == 0);
6360                         if (wc->refs[level] == 1) {
6361                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6362                                 return 1;
6363                         }
6364                 }
6365         }
6366
6367         /* wc->stage == DROP_REFERENCE */
6368         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6369
6370         if (wc->refs[level] == 1) {
6371                 if (level == 0) {
6372                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6373                                 ret = btrfs_dec_ref(trans, root, eb, 1);
6374                         else
6375                                 ret = btrfs_dec_ref(trans, root, eb, 0);
6376                         BUG_ON(ret);
6377                 }
6378                 /* make block locked assertion in clean_tree_block happy */
6379                 if (!path->locks[level] &&
6380                     btrfs_header_generation(eb) == trans->transid) {
6381                         btrfs_tree_lock(eb);
6382                         btrfs_set_lock_blocking(eb);
6383                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6384                 }
6385                 clean_tree_block(trans, root, eb);
6386         }
6387
6388         if (eb == root->node) {
6389                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6390                         parent = eb->start;
6391                 else
6392                         BUG_ON(root->root_key.objectid !=
6393                                btrfs_header_owner(eb));
6394         } else {
6395                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6396                         parent = path->nodes[level + 1]->start;
6397                 else
6398                         BUG_ON(root->root_key.objectid !=
6399                                btrfs_header_owner(path->nodes[level + 1]));
6400         }
6401
6402         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6403 out:
6404         wc->refs[level] = 0;
6405         wc->flags[level] = 0;
6406         return 0;
6407 }
6408
6409 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6410                                    struct btrfs_root *root,
6411                                    struct btrfs_path *path,
6412                                    struct walk_control *wc)
6413 {
6414         int level = wc->level;
6415         int lookup_info = 1;
6416         int ret;
6417
6418         while (level >= 0) {
6419                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6420                 if (ret > 0)
6421                         break;
6422
6423                 if (level == 0)
6424                         break;
6425
6426                 if (path->slots[level] >=
6427                     btrfs_header_nritems(path->nodes[level]))
6428                         break;
6429
6430                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6431                 if (ret > 0) {
6432                         path->slots[level]++;
6433                         continue;
6434                 } else if (ret < 0)
6435                         return ret;
6436                 level = wc->level;
6437         }
6438         return 0;
6439 }
6440
6441 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6442                                  struct btrfs_root *root,
6443                                  struct btrfs_path *path,
6444                                  struct walk_control *wc, int max_level)
6445 {
6446         int level = wc->level;
6447         int ret;
6448
6449         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6450         while (level < max_level && path->nodes[level]) {
6451                 wc->level = level;
6452                 if (path->slots[level] + 1 <
6453                     btrfs_header_nritems(path->nodes[level])) {
6454                         path->slots[level]++;
6455                         return 0;
6456                 } else {
6457                         ret = walk_up_proc(trans, root, path, wc);
6458                         if (ret > 0)
6459                                 return 0;
6460
6461                         if (path->locks[level]) {
6462                                 btrfs_tree_unlock_rw(path->nodes[level],
6463                                                      path->locks[level]);
6464                                 path->locks[level] = 0;
6465                         }
6466                         free_extent_buffer(path->nodes[level]);
6467                         path->nodes[level] = NULL;
6468                         level++;
6469                 }
6470         }
6471         return 1;
6472 }
6473
6474 /*
6475  * drop a subvolume tree.
6476  *
6477  * this function traverses the tree freeing any blocks that only
6478  * referenced by the tree.
6479  *
6480  * when a shared tree block is found. this function decreases its
6481  * reference count by one. if update_ref is true, this function
6482  * also make sure backrefs for the shared block and all lower level
6483  * blocks are properly updated.
6484  */
6485 void btrfs_drop_snapshot(struct btrfs_root *root,
6486                          struct btrfs_block_rsv *block_rsv, int update_ref)
6487 {
6488         struct btrfs_path *path;
6489         struct btrfs_trans_handle *trans;
6490         struct btrfs_root *tree_root = root->fs_info->tree_root;
6491         struct btrfs_root_item *root_item = &root->root_item;
6492         struct walk_control *wc;
6493         struct btrfs_key key;
6494         int err = 0;
6495         int ret;
6496         int level;
6497
6498         path = btrfs_alloc_path();
6499         if (!path) {
6500                 err = -ENOMEM;
6501                 goto out;
6502         }
6503
6504         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6505         if (!wc) {
6506                 btrfs_free_path(path);
6507                 err = -ENOMEM;
6508                 goto out;
6509         }
6510
6511         trans = btrfs_start_transaction(tree_root, 0);
6512         BUG_ON(IS_ERR(trans));
6513
6514         if (block_rsv)
6515                 trans->block_rsv = block_rsv;
6516
6517         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6518                 level = btrfs_header_level(root->node);
6519                 path->nodes[level] = btrfs_lock_root_node(root);
6520                 btrfs_set_lock_blocking(path->nodes[level]);
6521                 path->slots[level] = 0;
6522                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6523                 memset(&wc->update_progress, 0,
6524                        sizeof(wc->update_progress));
6525         } else {
6526                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6527                 memcpy(&wc->update_progress, &key,
6528                        sizeof(wc->update_progress));
6529
6530                 level = root_item->drop_level;
6531                 BUG_ON(level == 0);
6532                 path->lowest_level = level;
6533                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6534                 path->lowest_level = 0;
6535                 if (ret < 0) {
6536                         err = ret;
6537                         goto out_free;
6538                 }
6539                 WARN_ON(ret > 0);
6540
6541                 /*
6542                  * unlock our path, this is safe because only this
6543                  * function is allowed to delete this snapshot
6544                  */
6545                 btrfs_unlock_up_safe(path, 0);
6546
6547                 level = btrfs_header_level(root->node);
6548                 while (1) {
6549                         btrfs_tree_lock(path->nodes[level]);
6550                         btrfs_set_lock_blocking(path->nodes[level]);
6551
6552                         ret = btrfs_lookup_extent_info(trans, root,
6553                                                 path->nodes[level]->start,
6554                                                 path->nodes[level]->len,
6555                                                 &wc->refs[level],
6556                                                 &wc->flags[level]);
6557                         BUG_ON(ret);
6558                         BUG_ON(wc->refs[level] == 0);
6559
6560                         if (level == root_item->drop_level)
6561                                 break;
6562
6563                         btrfs_tree_unlock(path->nodes[level]);
6564                         WARN_ON(wc->refs[level] != 1);
6565                         level--;
6566                 }
6567         }
6568
6569         wc->level = level;
6570         wc->shared_level = -1;
6571         wc->stage = DROP_REFERENCE;
6572         wc->update_ref = update_ref;
6573         wc->keep_locks = 0;
6574         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6575
6576         while (1) {
6577                 ret = walk_down_tree(trans, root, path, wc);
6578                 if (ret < 0) {
6579                         err = ret;
6580                         break;
6581                 }
6582
6583                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
6584                 if (ret < 0) {
6585                         err = ret;
6586                         break;
6587                 }
6588
6589                 if (ret > 0) {
6590                         BUG_ON(wc->stage != DROP_REFERENCE);
6591                         break;
6592                 }
6593
6594                 if (wc->stage == DROP_REFERENCE) {
6595                         level = wc->level;
6596                         btrfs_node_key(path->nodes[level],
6597                                        &root_item->drop_progress,
6598                                        path->slots[level]);
6599                         root_item->drop_level = level;
6600                 }
6601
6602                 BUG_ON(wc->level == 0);
6603                 if (btrfs_should_end_transaction(trans, tree_root)) {
6604                         ret = btrfs_update_root(trans, tree_root,
6605                                                 &root->root_key,
6606                                                 root_item);
6607                         BUG_ON(ret);
6608
6609                         btrfs_end_transaction_throttle(trans, tree_root);
6610                         trans = btrfs_start_transaction(tree_root, 0);
6611                         BUG_ON(IS_ERR(trans));
6612                         if (block_rsv)
6613                                 trans->block_rsv = block_rsv;
6614                 }
6615         }
6616         btrfs_release_path(path);
6617         BUG_ON(err);
6618
6619         ret = btrfs_del_root(trans, tree_root, &root->root_key);
6620         BUG_ON(ret);
6621
6622         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
6623                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
6624                                            NULL, NULL);
6625                 BUG_ON(ret < 0);
6626                 if (ret > 0) {
6627                         /* if we fail to delete the orphan item this time
6628                          * around, it'll get picked up the next time.
6629                          *
6630                          * The most common failure here is just -ENOENT.
6631                          */
6632                         btrfs_del_orphan_item(trans, tree_root,
6633                                               root->root_key.objectid);
6634                 }
6635         }
6636
6637         if (root->in_radix) {
6638                 btrfs_free_fs_root(tree_root->fs_info, root);
6639         } else {
6640                 free_extent_buffer(root->node);
6641                 free_extent_buffer(root->commit_root);
6642                 kfree(root);
6643         }
6644 out_free:
6645         btrfs_end_transaction_throttle(trans, tree_root);
6646         kfree(wc);
6647         btrfs_free_path(path);
6648 out:
6649         if (err)
6650                 btrfs_std_error(root->fs_info, err);
6651         return;
6652 }
6653
6654 /*
6655  * drop subtree rooted at tree block 'node'.
6656  *
6657  * NOTE: this function will unlock and release tree block 'node'
6658  */
6659 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
6660                         struct btrfs_root *root,
6661                         struct extent_buffer *node,
6662                         struct extent_buffer *parent)
6663 {
6664         struct btrfs_path *path;
6665         struct walk_control *wc;
6666         int level;
6667         int parent_level;
6668         int ret = 0;
6669         int wret;
6670
6671         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
6672
6673         path = btrfs_alloc_path();
6674         if (!path)
6675                 return -ENOMEM;
6676
6677         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6678         if (!wc) {
6679                 btrfs_free_path(path);
6680                 return -ENOMEM;
6681         }
6682
6683         btrfs_assert_tree_locked(parent);
6684         parent_level = btrfs_header_level(parent);
6685         extent_buffer_get(parent);
6686         path->nodes[parent_level] = parent;
6687         path->slots[parent_level] = btrfs_header_nritems(parent);
6688
6689         btrfs_assert_tree_locked(node);
6690         level = btrfs_header_level(node);
6691         path->nodes[level] = node;
6692         path->slots[level] = 0;
6693         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6694
6695         wc->refs[parent_level] = 1;
6696         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6697         wc->level = level;
6698         wc->shared_level = -1;
6699         wc->stage = DROP_REFERENCE;
6700         wc->update_ref = 0;
6701         wc->keep_locks = 1;
6702         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6703
6704         while (1) {
6705                 wret = walk_down_tree(trans, root, path, wc);
6706                 if (wret < 0) {
6707                         ret = wret;
6708                         break;
6709                 }
6710
6711                 wret = walk_up_tree(trans, root, path, wc, parent_level);
6712                 if (wret < 0)
6713                         ret = wret;
6714                 if (wret != 0)
6715                         break;
6716         }
6717
6718         kfree(wc);
6719         btrfs_free_path(path);
6720         return ret;
6721 }
6722
6723 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
6724 {
6725         u64 num_devices;
6726         u64 stripped = BTRFS_BLOCK_GROUP_RAID0 |
6727                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
6728
6729         /*
6730          * we add in the count of missing devices because we want
6731          * to make sure that any RAID levels on a degraded FS
6732          * continue to be honored.
6733          */
6734         num_devices = root->fs_info->fs_devices->rw_devices +
6735                 root->fs_info->fs_devices->missing_devices;
6736
6737         if (num_devices == 1) {
6738                 stripped |= BTRFS_BLOCK_GROUP_DUP;
6739                 stripped = flags & ~stripped;
6740
6741                 /* turn raid0 into single device chunks */
6742                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
6743                         return stripped;
6744
6745                 /* turn mirroring into duplication */
6746                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
6747                              BTRFS_BLOCK_GROUP_RAID10))
6748                         return stripped | BTRFS_BLOCK_GROUP_DUP;
6749                 return flags;
6750         } else {
6751                 /* they already had raid on here, just return */
6752                 if (flags & stripped)
6753                         return flags;
6754
6755                 stripped |= BTRFS_BLOCK_GROUP_DUP;
6756                 stripped = flags & ~stripped;
6757
6758                 /* switch duplicated blocks with raid1 */
6759                 if (flags & BTRFS_BLOCK_GROUP_DUP)
6760                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
6761
6762                 /* turn single device chunks into raid0 */
6763                 return stripped | BTRFS_BLOCK_GROUP_RAID0;
6764         }
6765         return flags;
6766 }
6767
6768 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
6769 {
6770         struct btrfs_space_info *sinfo = cache->space_info;
6771         u64 num_bytes;
6772         u64 min_allocable_bytes;
6773         int ret = -ENOSPC;
6774
6775
6776         /*
6777          * We need some metadata space and system metadata space for
6778          * allocating chunks in some corner cases until we force to set
6779          * it to be readonly.
6780          */
6781         if ((sinfo->flags &
6782              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
6783             !force)
6784                 min_allocable_bytes = 1 * 1024 * 1024;
6785         else
6786                 min_allocable_bytes = 0;
6787
6788         spin_lock(&sinfo->lock);
6789         spin_lock(&cache->lock);
6790
6791         if (cache->ro) {
6792                 ret = 0;
6793                 goto out;
6794         }
6795
6796         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
6797                     cache->bytes_super - btrfs_block_group_used(&cache->item);
6798
6799         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
6800             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
6801             min_allocable_bytes <= sinfo->total_bytes) {
6802                 sinfo->bytes_readonly += num_bytes;
6803                 cache->ro = 1;
6804                 ret = 0;
6805         }
6806 out:
6807         spin_unlock(&cache->lock);
6808         spin_unlock(&sinfo->lock);
6809         return ret;
6810 }
6811
6812 int btrfs_set_block_group_ro(struct btrfs_root *root,
6813                              struct btrfs_block_group_cache *cache)
6814
6815 {
6816         struct btrfs_trans_handle *trans;
6817         u64 alloc_flags;
6818         int ret;
6819
6820         BUG_ON(cache->ro);
6821
6822         trans = btrfs_join_transaction(root);
6823         BUG_ON(IS_ERR(trans));
6824
6825         alloc_flags = update_block_group_flags(root, cache->flags);
6826         if (alloc_flags != cache->flags)
6827                 do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6828                                CHUNK_ALLOC_FORCE);
6829
6830         ret = set_block_group_ro(cache, 0);
6831         if (!ret)
6832                 goto out;
6833         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
6834         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6835                              CHUNK_ALLOC_FORCE);
6836         if (ret < 0)
6837                 goto out;
6838         ret = set_block_group_ro(cache, 0);
6839 out:
6840         btrfs_end_transaction(trans, root);
6841         return ret;
6842 }
6843
6844 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
6845                             struct btrfs_root *root, u64 type)
6846 {
6847         u64 alloc_flags = get_alloc_profile(root, type);
6848         return do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6849                               CHUNK_ALLOC_FORCE);
6850 }
6851
6852 /*
6853  * helper to account the unused space of all the readonly block group in the
6854  * list. takes mirrors into account.
6855  */
6856 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
6857 {
6858         struct btrfs_block_group_cache *block_group;
6859         u64 free_bytes = 0;
6860         int factor;
6861
6862         list_for_each_entry(block_group, groups_list, list) {
6863                 spin_lock(&block_group->lock);
6864
6865                 if (!block_group->ro) {
6866                         spin_unlock(&block_group->lock);
6867                         continue;
6868                 }
6869
6870                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
6871                                           BTRFS_BLOCK_GROUP_RAID10 |
6872                                           BTRFS_BLOCK_GROUP_DUP))
6873                         factor = 2;
6874                 else
6875                         factor = 1;
6876
6877                 free_bytes += (block_group->key.offset -
6878                                btrfs_block_group_used(&block_group->item)) *
6879                                factor;
6880
6881                 spin_unlock(&block_group->lock);
6882         }
6883
6884         return free_bytes;
6885 }
6886
6887 /*
6888  * helper to account the unused space of all the readonly block group in the
6889  * space_info. takes mirrors into account.
6890  */
6891 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
6892 {
6893         int i;
6894         u64 free_bytes = 0;
6895
6896         spin_lock(&sinfo->lock);
6897
6898         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
6899                 if (!list_empty(&sinfo->block_groups[i]))
6900                         free_bytes += __btrfs_get_ro_block_group_free_space(
6901                                                 &sinfo->block_groups[i]);
6902
6903         spin_unlock(&sinfo->lock);
6904
6905         return free_bytes;
6906 }
6907
6908 int btrfs_set_block_group_rw(struct btrfs_root *root,
6909                               struct btrfs_block_group_cache *cache)
6910 {
6911         struct btrfs_space_info *sinfo = cache->space_info;
6912         u64 num_bytes;
6913
6914         BUG_ON(!cache->ro);
6915
6916         spin_lock(&sinfo->lock);
6917         spin_lock(&cache->lock);
6918         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
6919                     cache->bytes_super - btrfs_block_group_used(&cache->item);
6920         sinfo->bytes_readonly -= num_bytes;
6921         cache->ro = 0;
6922         spin_unlock(&cache->lock);
6923         spin_unlock(&sinfo->lock);
6924         return 0;
6925 }
6926
6927 /*
6928  * checks to see if its even possible to relocate this block group.
6929  *
6930  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
6931  * ok to go ahead and try.
6932  */
6933 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
6934 {
6935         struct btrfs_block_group_cache *block_group;
6936         struct btrfs_space_info *space_info;
6937         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
6938         struct btrfs_device *device;
6939         u64 min_free;
6940         u64 dev_min = 1;
6941         u64 dev_nr = 0;
6942         int index;
6943         int full = 0;
6944         int ret = 0;
6945
6946         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
6947
6948         /* odd, couldn't find the block group, leave it alone */
6949         if (!block_group)
6950                 return -1;
6951
6952         min_free = btrfs_block_group_used(&block_group->item);
6953
6954         /* no bytes used, we're good */
6955         if (!min_free)
6956                 goto out;
6957
6958         space_info = block_group->space_info;
6959         spin_lock(&space_info->lock);
6960
6961         full = space_info->full;
6962
6963         /*
6964          * if this is the last block group we have in this space, we can't
6965          * relocate it unless we're able to allocate a new chunk below.
6966          *
6967          * Otherwise, we need to make sure we have room in the space to handle
6968          * all of the extents from this block group.  If we can, we're good
6969          */
6970         if ((space_info->total_bytes != block_group->key.offset) &&
6971             (space_info->bytes_used + space_info->bytes_reserved +
6972              space_info->bytes_pinned + space_info->bytes_readonly +
6973              min_free < space_info->total_bytes)) {
6974                 spin_unlock(&space_info->lock);
6975                 goto out;
6976         }
6977         spin_unlock(&space_info->lock);
6978
6979         /*
6980          * ok we don't have enough space, but maybe we have free space on our
6981          * devices to allocate new chunks for relocation, so loop through our
6982          * alloc devices and guess if we have enough space.  However, if we
6983          * were marked as full, then we know there aren't enough chunks, and we
6984          * can just return.
6985          */
6986         ret = -1;
6987         if (full)
6988                 goto out;
6989
6990         /*
6991          * index:
6992          *      0: raid10
6993          *      1: raid1
6994          *      2: dup
6995          *      3: raid0
6996          *      4: single
6997          */
6998         index = get_block_group_index(block_group);
6999         if (index == 0) {
7000                 dev_min = 4;
7001                 /* Divide by 2 */
7002                 min_free >>= 1;
7003         } else if (index == 1) {
7004                 dev_min = 2;
7005         } else if (index == 2) {
7006                 /* Multiply by 2 */
7007                 min_free <<= 1;
7008         } else if (index == 3) {
7009                 dev_min = fs_devices->rw_devices;
7010                 do_div(min_free, dev_min);
7011         }
7012
7013         mutex_lock(&root->fs_info->chunk_mutex);
7014         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7015                 u64 dev_offset;
7016
7017                 /*
7018                  * check to make sure we can actually find a chunk with enough
7019                  * space to fit our block group in.
7020                  */
7021                 if (device->total_bytes > device->bytes_used + min_free) {
7022                         ret = find_free_dev_extent(NULL, device, min_free,
7023                                                    &dev_offset, NULL);
7024                         if (!ret)
7025                                 dev_nr++;
7026
7027                         if (dev_nr >= dev_min)
7028                                 break;
7029
7030                         ret = -1;
7031                 }
7032         }
7033         mutex_unlock(&root->fs_info->chunk_mutex);
7034 out:
7035         btrfs_put_block_group(block_group);
7036         return ret;
7037 }
7038
7039 static int find_first_block_group(struct btrfs_root *root,
7040                 struct btrfs_path *path, struct btrfs_key *key)
7041 {
7042         int ret = 0;
7043         struct btrfs_key found_key;
7044         struct extent_buffer *leaf;
7045         int slot;
7046
7047         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7048         if (ret < 0)
7049                 goto out;
7050
7051         while (1) {
7052                 slot = path->slots[0];
7053                 leaf = path->nodes[0];
7054                 if (slot >= btrfs_header_nritems(leaf)) {
7055                         ret = btrfs_next_leaf(root, path);
7056                         if (ret == 0)
7057                                 continue;
7058                         if (ret < 0)
7059                                 goto out;
7060                         break;
7061                 }
7062                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7063
7064                 if (found_key.objectid >= key->objectid &&
7065                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7066                         ret = 0;
7067                         goto out;
7068                 }
7069                 path->slots[0]++;
7070         }
7071 out:
7072         return ret;
7073 }
7074
7075 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7076 {
7077         struct btrfs_block_group_cache *block_group;
7078         u64 last = 0;
7079
7080         while (1) {
7081                 struct inode *inode;
7082
7083                 block_group = btrfs_lookup_first_block_group(info, last);
7084                 while (block_group) {
7085                         spin_lock(&block_group->lock);
7086                         if (block_group->iref)
7087                                 break;
7088                         spin_unlock(&block_group->lock);
7089                         block_group = next_block_group(info->tree_root,
7090                                                        block_group);
7091                 }
7092                 if (!block_group) {
7093                         if (last == 0)
7094                                 break;
7095                         last = 0;
7096                         continue;
7097                 }
7098
7099                 inode = block_group->inode;
7100                 block_group->iref = 0;
7101                 block_group->inode = NULL;
7102                 spin_unlock(&block_group->lock);
7103                 iput(inode);
7104                 last = block_group->key.objectid + block_group->key.offset;
7105                 btrfs_put_block_group(block_group);
7106         }
7107 }
7108
7109 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7110 {
7111         struct btrfs_block_group_cache *block_group;
7112         struct btrfs_space_info *space_info;
7113         struct btrfs_caching_control *caching_ctl;
7114         struct rb_node *n;
7115
7116         down_write(&info->extent_commit_sem);
7117         while (!list_empty(&info->caching_block_groups)) {
7118                 caching_ctl = list_entry(info->caching_block_groups.next,
7119                                          struct btrfs_caching_control, list);
7120                 list_del(&caching_ctl->list);
7121                 put_caching_control(caching_ctl);
7122         }
7123         up_write(&info->extent_commit_sem);
7124
7125         spin_lock(&info->block_group_cache_lock);
7126         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7127                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7128                                        cache_node);
7129                 rb_erase(&block_group->cache_node,
7130                          &info->block_group_cache_tree);
7131                 spin_unlock(&info->block_group_cache_lock);
7132
7133                 down_write(&block_group->space_info->groups_sem);
7134                 list_del(&block_group->list);
7135                 up_write(&block_group->space_info->groups_sem);
7136
7137                 if (block_group->cached == BTRFS_CACHE_STARTED)
7138                         wait_block_group_cache_done(block_group);
7139
7140                 /*
7141                  * We haven't cached this block group, which means we could
7142                  * possibly have excluded extents on this block group.
7143                  */
7144                 if (block_group->cached == BTRFS_CACHE_NO)
7145                         free_excluded_extents(info->extent_root, block_group);
7146
7147                 btrfs_remove_free_space_cache(block_group);
7148                 btrfs_put_block_group(block_group);
7149
7150                 spin_lock(&info->block_group_cache_lock);
7151         }
7152         spin_unlock(&info->block_group_cache_lock);
7153
7154         /* now that all the block groups are freed, go through and
7155          * free all the space_info structs.  This is only called during
7156          * the final stages of unmount, and so we know nobody is
7157          * using them.  We call synchronize_rcu() once before we start,
7158          * just to be on the safe side.
7159          */
7160         synchronize_rcu();
7161
7162         release_global_block_rsv(info);
7163
7164         while(!list_empty(&info->space_info)) {
7165                 space_info = list_entry(info->space_info.next,
7166                                         struct btrfs_space_info,
7167                                         list);
7168                 if (space_info->bytes_pinned > 0 ||
7169                     space_info->bytes_reserved > 0 ||
7170                     space_info->bytes_may_use > 0) {
7171                         WARN_ON(1);
7172                         dump_space_info(space_info, 0, 0);
7173                 }
7174                 list_del(&space_info->list);
7175                 kfree(space_info);
7176         }
7177         return 0;
7178 }
7179
7180 static void __link_block_group(struct btrfs_space_info *space_info,
7181                                struct btrfs_block_group_cache *cache)
7182 {
7183         int index = get_block_group_index(cache);
7184
7185         down_write(&space_info->groups_sem);
7186         list_add_tail(&cache->list, &space_info->block_groups[index]);
7187         up_write(&space_info->groups_sem);
7188 }
7189
7190 int btrfs_read_block_groups(struct btrfs_root *root)
7191 {
7192         struct btrfs_path *path;
7193         int ret;
7194         struct btrfs_block_group_cache *cache;
7195         struct btrfs_fs_info *info = root->fs_info;
7196         struct btrfs_space_info *space_info;
7197         struct btrfs_key key;
7198         struct btrfs_key found_key;
7199         struct extent_buffer *leaf;
7200         int need_clear = 0;
7201         u64 cache_gen;
7202
7203         root = info->extent_root;
7204         key.objectid = 0;
7205         key.offset = 0;
7206         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7207         path = btrfs_alloc_path();
7208         if (!path)
7209                 return -ENOMEM;
7210         path->reada = 1;
7211
7212         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7213         if (btrfs_test_opt(root, SPACE_CACHE) &&
7214             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7215                 need_clear = 1;
7216         if (btrfs_test_opt(root, CLEAR_CACHE))
7217                 need_clear = 1;
7218
7219         while (1) {
7220                 ret = find_first_block_group(root, path, &key);
7221                 if (ret > 0)
7222                         break;
7223                 if (ret != 0)
7224                         goto error;
7225                 leaf = path->nodes[0];
7226                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7227                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7228                 if (!cache) {
7229                         ret = -ENOMEM;
7230                         goto error;
7231                 }
7232                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7233                                                 GFP_NOFS);
7234                 if (!cache->free_space_ctl) {
7235                         kfree(cache);
7236                         ret = -ENOMEM;
7237                         goto error;
7238                 }
7239
7240                 atomic_set(&cache->count, 1);
7241                 spin_lock_init(&cache->lock);
7242                 cache->fs_info = info;
7243                 INIT_LIST_HEAD(&cache->list);
7244                 INIT_LIST_HEAD(&cache->cluster_list);
7245
7246                 if (need_clear)
7247                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7248
7249                 read_extent_buffer(leaf, &cache->item,
7250                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7251                                    sizeof(cache->item));
7252                 memcpy(&cache->key, &found_key, sizeof(found_key));
7253
7254                 key.objectid = found_key.objectid + found_key.offset;
7255                 btrfs_release_path(path);
7256                 cache->flags = btrfs_block_group_flags(&cache->item);
7257                 cache->sectorsize = root->sectorsize;
7258
7259                 btrfs_init_free_space_ctl(cache);
7260
7261                 /*
7262                  * We need to exclude the super stripes now so that the space
7263                  * info has super bytes accounted for, otherwise we'll think
7264                  * we have more space than we actually do.
7265                  */
7266                 exclude_super_stripes(root, cache);
7267
7268                 /*
7269                  * check for two cases, either we are full, and therefore
7270                  * don't need to bother with the caching work since we won't
7271                  * find any space, or we are empty, and we can just add all
7272                  * the space in and be done with it.  This saves us _alot_ of
7273                  * time, particularly in the full case.
7274                  */
7275                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7276                         cache->last_byte_to_unpin = (u64)-1;
7277                         cache->cached = BTRFS_CACHE_FINISHED;
7278                         free_excluded_extents(root, cache);
7279                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7280                         cache->last_byte_to_unpin = (u64)-1;
7281                         cache->cached = BTRFS_CACHE_FINISHED;
7282                         add_new_free_space(cache, root->fs_info,
7283                                            found_key.objectid,
7284                                            found_key.objectid +
7285                                            found_key.offset);
7286                         free_excluded_extents(root, cache);
7287                 }
7288
7289                 ret = update_space_info(info, cache->flags, found_key.offset,
7290                                         btrfs_block_group_used(&cache->item),
7291                                         &space_info);
7292                 BUG_ON(ret);
7293                 cache->space_info = space_info;
7294                 spin_lock(&cache->space_info->lock);
7295                 cache->space_info->bytes_readonly += cache->bytes_super;
7296                 spin_unlock(&cache->space_info->lock);
7297
7298                 __link_block_group(space_info, cache);
7299
7300                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7301                 BUG_ON(ret);
7302
7303                 set_avail_alloc_bits(root->fs_info, cache->flags);
7304                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7305                         set_block_group_ro(cache, 1);
7306         }
7307
7308         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7309                 if (!(get_alloc_profile(root, space_info->flags) &
7310                       (BTRFS_BLOCK_GROUP_RAID10 |
7311                        BTRFS_BLOCK_GROUP_RAID1 |
7312                        BTRFS_BLOCK_GROUP_DUP)))
7313                         continue;
7314                 /*
7315                  * avoid allocating from un-mirrored block group if there are
7316                  * mirrored block groups.
7317                  */
7318                 list_for_each_entry(cache, &space_info->block_groups[3], list)
7319                         set_block_group_ro(cache, 1);
7320                 list_for_each_entry(cache, &space_info->block_groups[4], list)
7321                         set_block_group_ro(cache, 1);
7322         }
7323
7324         init_global_block_rsv(info);
7325         ret = 0;
7326 error:
7327         btrfs_free_path(path);
7328         return ret;
7329 }
7330
7331 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7332                            struct btrfs_root *root, u64 bytes_used,
7333                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7334                            u64 size)
7335 {
7336         int ret;
7337         struct btrfs_root *extent_root;
7338         struct btrfs_block_group_cache *cache;
7339
7340         extent_root = root->fs_info->extent_root;
7341
7342         root->fs_info->last_trans_log_full_commit = trans->transid;
7343
7344         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7345         if (!cache)
7346                 return -ENOMEM;
7347         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7348                                         GFP_NOFS);
7349         if (!cache->free_space_ctl) {
7350                 kfree(cache);
7351                 return -ENOMEM;
7352         }
7353
7354         cache->key.objectid = chunk_offset;
7355         cache->key.offset = size;
7356         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7357         cache->sectorsize = root->sectorsize;
7358         cache->fs_info = root->fs_info;
7359
7360         atomic_set(&cache->count, 1);
7361         spin_lock_init(&cache->lock);
7362         INIT_LIST_HEAD(&cache->list);
7363         INIT_LIST_HEAD(&cache->cluster_list);
7364
7365         btrfs_init_free_space_ctl(cache);
7366
7367         btrfs_set_block_group_used(&cache->item, bytes_used);
7368         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7369         cache->flags = type;
7370         btrfs_set_block_group_flags(&cache->item, type);
7371
7372         cache->last_byte_to_unpin = (u64)-1;
7373         cache->cached = BTRFS_CACHE_FINISHED;
7374         exclude_super_stripes(root, cache);
7375
7376         add_new_free_space(cache, root->fs_info, chunk_offset,
7377                            chunk_offset + size);
7378
7379         free_excluded_extents(root, cache);
7380
7381         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7382                                 &cache->space_info);
7383         BUG_ON(ret);
7384
7385         spin_lock(&cache->space_info->lock);
7386         cache->space_info->bytes_readonly += cache->bytes_super;
7387         spin_unlock(&cache->space_info->lock);
7388
7389         __link_block_group(cache->space_info, cache);
7390
7391         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7392         BUG_ON(ret);
7393
7394         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
7395                                 sizeof(cache->item));
7396         BUG_ON(ret);
7397
7398         set_avail_alloc_bits(extent_root->fs_info, type);
7399
7400         return 0;
7401 }
7402
7403 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7404                              struct btrfs_root *root, u64 group_start)
7405 {
7406         struct btrfs_path *path;
7407         struct btrfs_block_group_cache *block_group;
7408         struct btrfs_free_cluster *cluster;
7409         struct btrfs_root *tree_root = root->fs_info->tree_root;
7410         struct btrfs_key key;
7411         struct inode *inode;
7412         int ret;
7413         int factor;
7414
7415         root = root->fs_info->extent_root;
7416
7417         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7418         BUG_ON(!block_group);
7419         BUG_ON(!block_group->ro);
7420
7421         /*
7422          * Free the reserved super bytes from this block group before
7423          * remove it.
7424          */
7425         free_excluded_extents(root, block_group);
7426
7427         memcpy(&key, &block_group->key, sizeof(key));
7428         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7429                                   BTRFS_BLOCK_GROUP_RAID1 |
7430                                   BTRFS_BLOCK_GROUP_RAID10))
7431                 factor = 2;
7432         else
7433                 factor = 1;
7434
7435         /* make sure this block group isn't part of an allocation cluster */
7436         cluster = &root->fs_info->data_alloc_cluster;
7437         spin_lock(&cluster->refill_lock);
7438         btrfs_return_cluster_to_free_space(block_group, cluster);
7439         spin_unlock(&cluster->refill_lock);
7440
7441         /*
7442          * make sure this block group isn't part of a metadata
7443          * allocation cluster
7444          */
7445         cluster = &root->fs_info->meta_alloc_cluster;
7446         spin_lock(&cluster->refill_lock);
7447         btrfs_return_cluster_to_free_space(block_group, cluster);
7448         spin_unlock(&cluster->refill_lock);
7449
7450         path = btrfs_alloc_path();
7451         if (!path) {
7452                 ret = -ENOMEM;
7453                 goto out;
7454         }
7455
7456         inode = lookup_free_space_inode(tree_root, block_group, path);
7457         if (!IS_ERR(inode)) {
7458                 ret = btrfs_orphan_add(trans, inode);
7459                 BUG_ON(ret);
7460                 clear_nlink(inode);
7461                 /* One for the block groups ref */
7462                 spin_lock(&block_group->lock);
7463                 if (block_group->iref) {
7464                         block_group->iref = 0;
7465                         block_group->inode = NULL;
7466                         spin_unlock(&block_group->lock);
7467                         iput(inode);
7468                 } else {
7469                         spin_unlock(&block_group->lock);
7470                 }
7471                 /* One for our lookup ref */
7472                 btrfs_add_delayed_iput(inode);
7473         }
7474
7475         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
7476         key.offset = block_group->key.objectid;
7477         key.type = 0;
7478
7479         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
7480         if (ret < 0)
7481                 goto out;
7482         if (ret > 0)
7483                 btrfs_release_path(path);
7484         if (ret == 0) {
7485                 ret = btrfs_del_item(trans, tree_root, path);
7486                 if (ret)
7487                         goto out;
7488                 btrfs_release_path(path);
7489         }
7490
7491         spin_lock(&root->fs_info->block_group_cache_lock);
7492         rb_erase(&block_group->cache_node,
7493                  &root->fs_info->block_group_cache_tree);
7494         spin_unlock(&root->fs_info->block_group_cache_lock);
7495
7496         down_write(&block_group->space_info->groups_sem);
7497         /*
7498          * we must use list_del_init so people can check to see if they
7499          * are still on the list after taking the semaphore
7500          */
7501         list_del_init(&block_group->list);
7502         up_write(&block_group->space_info->groups_sem);
7503
7504         if (block_group->cached == BTRFS_CACHE_STARTED)
7505                 wait_block_group_cache_done(block_group);
7506
7507         btrfs_remove_free_space_cache(block_group);
7508
7509         spin_lock(&block_group->space_info->lock);
7510         block_group->space_info->total_bytes -= block_group->key.offset;
7511         block_group->space_info->bytes_readonly -= block_group->key.offset;
7512         block_group->space_info->disk_total -= block_group->key.offset * factor;
7513         spin_unlock(&block_group->space_info->lock);
7514
7515         memcpy(&key, &block_group->key, sizeof(key));
7516
7517         btrfs_clear_space_info_full(root->fs_info);
7518
7519         btrfs_put_block_group(block_group);
7520         btrfs_put_block_group(block_group);
7521
7522         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
7523         if (ret > 0)
7524                 ret = -EIO;
7525         if (ret < 0)
7526                 goto out;
7527
7528         ret = btrfs_del_item(trans, root, path);
7529 out:
7530         btrfs_free_path(path);
7531         return ret;
7532 }
7533
7534 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
7535 {
7536         struct btrfs_space_info *space_info;
7537         struct btrfs_super_block *disk_super;
7538         u64 features;
7539         u64 flags;
7540         int mixed = 0;
7541         int ret;
7542
7543         disk_super = fs_info->super_copy;
7544         if (!btrfs_super_root(disk_super))
7545                 return 1;
7546
7547         features = btrfs_super_incompat_flags(disk_super);
7548         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
7549                 mixed = 1;
7550
7551         flags = BTRFS_BLOCK_GROUP_SYSTEM;
7552         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7553         if (ret)
7554                 goto out;
7555
7556         if (mixed) {
7557                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
7558                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7559         } else {
7560                 flags = BTRFS_BLOCK_GROUP_METADATA;
7561                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7562                 if (ret)
7563                         goto out;
7564
7565                 flags = BTRFS_BLOCK_GROUP_DATA;
7566                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7567         }
7568 out:
7569         return ret;
7570 }
7571
7572 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
7573 {
7574         return unpin_extent_range(root, start, end);
7575 }
7576
7577 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
7578                                u64 num_bytes, u64 *actual_bytes)
7579 {
7580         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
7581 }
7582
7583 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
7584 {
7585         struct btrfs_fs_info *fs_info = root->fs_info;
7586         struct btrfs_block_group_cache *cache = NULL;
7587         u64 group_trimmed;
7588         u64 start;
7589         u64 end;
7590         u64 trimmed = 0;
7591         int ret = 0;
7592
7593         cache = btrfs_lookup_block_group(fs_info, range->start);
7594
7595         while (cache) {
7596                 if (cache->key.objectid >= (range->start + range->len)) {
7597                         btrfs_put_block_group(cache);
7598                         break;
7599                 }
7600
7601                 start = max(range->start, cache->key.objectid);
7602                 end = min(range->start + range->len,
7603                                 cache->key.objectid + cache->key.offset);
7604
7605                 if (end - start >= range->minlen) {
7606                         if (!block_group_cache_done(cache)) {
7607                                 ret = cache_block_group(cache, NULL, root, 0);
7608                                 if (!ret)
7609                                         wait_block_group_cache_done(cache);
7610                         }
7611                         ret = btrfs_trim_block_group(cache,
7612                                                      &group_trimmed,
7613                                                      start,
7614                                                      end,
7615                                                      range->minlen);
7616
7617                         trimmed += group_trimmed;
7618                         if (ret) {
7619                                 btrfs_put_block_group(cache);
7620                                 break;
7621                         }
7622                 }
7623
7624                 cache = next_block_group(fs_info->tree_root, cache);
7625         }
7626
7627         range->len = trimmed;
7628         return ret;
7629 }