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
3380                 spin_lock(&space_info->lock);
3381                 if (reserved > space_info->bytes_may_use)
3382                         reclaimed += reserved - space_info->bytes_may_use;
3383                 reserved = space_info->bytes_may_use;
3384                 spin_unlock(&space_info->lock);
3385
3386                 loops++;
3387
3388                 if (reserved == 0 || reclaimed >= max_reclaim)
3389                         break;
3390
3391                 if (trans && trans->transaction->blocked)
3392                         return -EAGAIN;
3393
3394                 if (wait_ordered && !trans) {
3395                         btrfs_wait_ordered_extents(root, 0, 0);
3396                 } else {
3397                         time_left = schedule_timeout_interruptible(1);
3398
3399                         /* We were interrupted, exit */
3400                         if (time_left)
3401                                 break;
3402                 }
3403
3404                 /* we've kicked the IO a few times, if anything has been freed,
3405                  * exit.  There is no sense in looping here for a long time
3406                  * when we really need to commit the transaction, or there are
3407                  * just too many writers without enough free space
3408                  */
3409
3410                 if (loops > 3) {
3411                         smp_mb();
3412                         if (progress != space_info->reservation_progress)
3413                                 break;
3414                 }
3415
3416         }
3417
3418         return reclaimed >= to_reclaim;
3419 }
3420
3421 /**
3422  * maybe_commit_transaction - possibly commit the transaction if its ok to
3423  * @root - the root we're allocating for
3424  * @bytes - the number of bytes we want to reserve
3425  * @force - force the commit
3426  *
3427  * This will check to make sure that committing the transaction will actually
3428  * get us somewhere and then commit the transaction if it does.  Otherwise it
3429  * will return -ENOSPC.
3430  */
3431 static int may_commit_transaction(struct btrfs_root *root,
3432                                   struct btrfs_space_info *space_info,
3433                                   u64 bytes, int force)
3434 {
3435         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3436         struct btrfs_trans_handle *trans;
3437
3438         trans = (struct btrfs_trans_handle *)current->journal_info;
3439         if (trans)
3440                 return -EAGAIN;
3441
3442         if (force)
3443                 goto commit;
3444
3445         /* See if there is enough pinned space to make this reservation */
3446         spin_lock(&space_info->lock);
3447         if (space_info->bytes_pinned >= bytes) {
3448                 spin_unlock(&space_info->lock);
3449                 goto commit;
3450         }
3451         spin_unlock(&space_info->lock);
3452
3453         /*
3454          * See if there is some space in the delayed insertion reservation for
3455          * this reservation.
3456          */
3457         if (space_info != delayed_rsv->space_info)
3458                 return -ENOSPC;
3459
3460         spin_lock(&delayed_rsv->lock);
3461         if (delayed_rsv->size < bytes) {
3462                 spin_unlock(&delayed_rsv->lock);
3463                 return -ENOSPC;
3464         }
3465         spin_unlock(&delayed_rsv->lock);
3466
3467 commit:
3468         trans = btrfs_join_transaction(root);
3469         if (IS_ERR(trans))
3470                 return -ENOSPC;
3471
3472         return btrfs_commit_transaction(trans, root);
3473 }
3474
3475 /**
3476  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3477  * @root - the root we're allocating for
3478  * @block_rsv - the block_rsv we're allocating for
3479  * @orig_bytes - the number of bytes we want
3480  * @flush - wether or not we can flush to make our reservation
3481  *
3482  * This will reserve orgi_bytes number of bytes from the space info associated
3483  * with the block_rsv.  If there is not enough space it will make an attempt to
3484  * flush out space to make room.  It will do this by flushing delalloc if
3485  * possible or committing the transaction.  If flush is 0 then no attempts to
3486  * regain reservations will be made and this will fail if there is not enough
3487  * space already.
3488  */
3489 static int reserve_metadata_bytes(struct btrfs_root *root,
3490                                   struct btrfs_block_rsv *block_rsv,
3491                                   u64 orig_bytes, int flush)
3492 {
3493         struct btrfs_space_info *space_info = block_rsv->space_info;
3494         u64 used;
3495         u64 num_bytes = orig_bytes;
3496         int retries = 0;
3497         int ret = 0;
3498         bool committed = false;
3499         bool flushing = false;
3500         bool wait_ordered = false;
3501
3502 again:
3503         ret = 0;
3504         spin_lock(&space_info->lock);
3505         /*
3506          * We only want to wait if somebody other than us is flushing and we are
3507          * actually alloed to flush.
3508          */
3509         while (flush && !flushing && space_info->flush) {
3510                 spin_unlock(&space_info->lock);
3511                 /*
3512                  * If we have a trans handle we can't wait because the flusher
3513                  * may have to commit the transaction, which would mean we would
3514                  * deadlock since we are waiting for the flusher to finish, but
3515                  * hold the current transaction open.
3516                  */
3517                 if (current->journal_info)
3518                         return -EAGAIN;
3519                 ret = wait_event_interruptible(space_info->wait,
3520                                                !space_info->flush);
3521                 /* Must have been interrupted, return */
3522                 if (ret)
3523                         return -EINTR;
3524
3525                 spin_lock(&space_info->lock);
3526         }
3527
3528         ret = -ENOSPC;
3529         used = space_info->bytes_used + space_info->bytes_reserved +
3530                 space_info->bytes_pinned + space_info->bytes_readonly +
3531                 space_info->bytes_may_use;
3532
3533         /*
3534          * The idea here is that we've not already over-reserved the block group
3535          * then we can go ahead and save our reservation first and then start
3536          * flushing if we need to.  Otherwise if we've already overcommitted
3537          * lets start flushing stuff first and then come back and try to make
3538          * our reservation.
3539          */
3540         if (used <= space_info->total_bytes) {
3541                 if (used + orig_bytes <= space_info->total_bytes) {
3542                         space_info->bytes_may_use += orig_bytes;
3543                         ret = 0;
3544                 } else {
3545                         /*
3546                          * Ok set num_bytes to orig_bytes since we aren't
3547                          * overocmmitted, this way we only try and reclaim what
3548                          * we need.
3549                          */
3550                         num_bytes = orig_bytes;
3551                 }
3552         } else {
3553                 /*
3554                  * Ok we're over committed, set num_bytes to the overcommitted
3555                  * amount plus the amount of bytes that we need for this
3556                  * reservation.
3557                  */
3558                 wait_ordered = true;
3559                 num_bytes = used - space_info->total_bytes +
3560                         (orig_bytes * (retries + 1));
3561         }
3562
3563         if (ret) {
3564                 u64 profile = btrfs_get_alloc_profile(root, 0);
3565                 u64 avail;
3566
3567                 /*
3568                  * If we have a lot of space that's pinned, don't bother doing
3569                  * the overcommit dance yet and just commit the transaction.
3570                  */
3571                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3572                 do_div(avail, 10);
3573                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3574                         space_info->flush = 1;
3575                         flushing = true;
3576                         spin_unlock(&space_info->lock);
3577                         ret = may_commit_transaction(root, space_info,
3578                                                      orig_bytes, 1);
3579                         if (ret)
3580                                 goto out;
3581                         committed = true;
3582                         goto again;
3583                 }
3584
3585                 spin_lock(&root->fs_info->free_chunk_lock);
3586                 avail = root->fs_info->free_chunk_space;
3587
3588                 /*
3589                  * If we have dup, raid1 or raid10 then only half of the free
3590                  * space is actually useable.
3591                  */
3592                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3593                                BTRFS_BLOCK_GROUP_RAID1 |
3594                                BTRFS_BLOCK_GROUP_RAID10))
3595                         avail >>= 1;
3596
3597                 /*
3598                  * If we aren't flushing don't let us overcommit too much, say
3599                  * 1/8th of the space.  If we can flush, let it overcommit up to
3600                  * 1/2 of the space.
3601                  */
3602                 if (flush)
3603                         avail >>= 3;
3604                 else
3605                         avail >>= 1;
3606                  spin_unlock(&root->fs_info->free_chunk_lock);
3607
3608                 if (used + num_bytes < space_info->total_bytes + avail) {
3609                         space_info->bytes_may_use += orig_bytes;
3610                         ret = 0;
3611                 } else {
3612                         wait_ordered = true;
3613                 }
3614         }
3615
3616         /*
3617          * Couldn't make our reservation, save our place so while we're trying
3618          * to reclaim space we can actually use it instead of somebody else
3619          * stealing it from us.
3620          */
3621         if (ret && flush) {
3622                 flushing = true;
3623                 space_info->flush = 1;
3624         }
3625
3626         spin_unlock(&space_info->lock);
3627
3628         if (!ret || !flush)
3629                 goto out;
3630
3631         /*
3632          * We do synchronous shrinking since we don't actually unreserve
3633          * metadata until after the IO is completed.
3634          */
3635         ret = shrink_delalloc(root, num_bytes, wait_ordered);
3636         if (ret < 0)
3637                 goto out;
3638
3639         ret = 0;
3640
3641         /*
3642          * So if we were overcommitted it's possible that somebody else flushed
3643          * out enough space and we simply didn't have enough space to reclaim,
3644          * so go back around and try again.
3645          */
3646         if (retries < 2) {
3647                 wait_ordered = true;
3648                 retries++;
3649                 goto again;
3650         }
3651
3652         ret = -ENOSPC;
3653         if (committed)
3654                 goto out;
3655
3656         ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3657         if (!ret) {
3658                 committed = true;
3659                 goto again;
3660         }
3661
3662 out:
3663         if (flushing) {
3664                 spin_lock(&space_info->lock);
3665                 space_info->flush = 0;
3666                 wake_up_all(&space_info->wait);
3667                 spin_unlock(&space_info->lock);
3668         }
3669         return ret;
3670 }
3671
3672 static struct btrfs_block_rsv *get_block_rsv(struct btrfs_trans_handle *trans,
3673                                              struct btrfs_root *root)
3674 {
3675         struct btrfs_block_rsv *block_rsv = NULL;
3676
3677         if (root->ref_cows || root == root->fs_info->csum_root)
3678                 block_rsv = trans->block_rsv;
3679
3680         if (!block_rsv)
3681                 block_rsv = root->block_rsv;
3682
3683         if (!block_rsv)
3684                 block_rsv = &root->fs_info->empty_block_rsv;
3685
3686         return block_rsv;
3687 }
3688
3689 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3690                                u64 num_bytes)
3691 {
3692         int ret = -ENOSPC;
3693         spin_lock(&block_rsv->lock);
3694         if (block_rsv->reserved >= num_bytes) {
3695                 block_rsv->reserved -= num_bytes;
3696                 if (block_rsv->reserved < block_rsv->size)
3697                         block_rsv->full = 0;
3698                 ret = 0;
3699         }
3700         spin_unlock(&block_rsv->lock);
3701         return ret;
3702 }
3703
3704 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3705                                 u64 num_bytes, int update_size)
3706 {
3707         spin_lock(&block_rsv->lock);
3708         block_rsv->reserved += num_bytes;
3709         if (update_size)
3710                 block_rsv->size += num_bytes;
3711         else if (block_rsv->reserved >= block_rsv->size)
3712                 block_rsv->full = 1;
3713         spin_unlock(&block_rsv->lock);
3714 }
3715
3716 static void block_rsv_release_bytes(struct btrfs_block_rsv *block_rsv,
3717                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3718 {
3719         struct btrfs_space_info *space_info = block_rsv->space_info;
3720
3721         spin_lock(&block_rsv->lock);
3722         if (num_bytes == (u64)-1)
3723                 num_bytes = block_rsv->size;
3724         block_rsv->size -= num_bytes;
3725         if (block_rsv->reserved >= block_rsv->size) {
3726                 num_bytes = block_rsv->reserved - block_rsv->size;
3727                 block_rsv->reserved = block_rsv->size;
3728                 block_rsv->full = 1;
3729         } else {
3730                 num_bytes = 0;
3731         }
3732         spin_unlock(&block_rsv->lock);
3733
3734         if (num_bytes > 0) {
3735                 if (dest) {
3736                         spin_lock(&dest->lock);
3737                         if (!dest->full) {
3738                                 u64 bytes_to_add;
3739
3740                                 bytes_to_add = dest->size - dest->reserved;
3741                                 bytes_to_add = min(num_bytes, bytes_to_add);
3742                                 dest->reserved += bytes_to_add;
3743                                 if (dest->reserved >= dest->size)
3744                                         dest->full = 1;
3745                                 num_bytes -= bytes_to_add;
3746                         }
3747                         spin_unlock(&dest->lock);
3748                 }
3749                 if (num_bytes) {
3750                         spin_lock(&space_info->lock);
3751                         space_info->bytes_may_use -= num_bytes;
3752                         space_info->reservation_progress++;
3753                         spin_unlock(&space_info->lock);
3754                 }
3755         }
3756 }
3757
3758 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3759                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3760 {
3761         int ret;
3762
3763         ret = block_rsv_use_bytes(src, num_bytes);
3764         if (ret)
3765                 return ret;
3766
3767         block_rsv_add_bytes(dst, num_bytes, 1);
3768         return 0;
3769 }
3770
3771 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3772 {
3773         memset(rsv, 0, sizeof(*rsv));
3774         spin_lock_init(&rsv->lock);
3775 }
3776
3777 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3778 {
3779         struct btrfs_block_rsv *block_rsv;
3780         struct btrfs_fs_info *fs_info = root->fs_info;
3781
3782         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3783         if (!block_rsv)
3784                 return NULL;
3785
3786         btrfs_init_block_rsv(block_rsv);
3787         block_rsv->space_info = __find_space_info(fs_info,
3788                                                   BTRFS_BLOCK_GROUP_METADATA);
3789         return block_rsv;
3790 }
3791
3792 void btrfs_free_block_rsv(struct btrfs_root *root,
3793                           struct btrfs_block_rsv *rsv)
3794 {
3795         btrfs_block_rsv_release(root, rsv, (u64)-1);
3796         kfree(rsv);
3797 }
3798
3799 int btrfs_block_rsv_add(struct btrfs_root *root,
3800                         struct btrfs_block_rsv *block_rsv,
3801                         u64 num_bytes)
3802 {
3803         int ret;
3804
3805         if (num_bytes == 0)
3806                 return 0;
3807
3808         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, 1);
3809         if (!ret) {
3810                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3811                 return 0;
3812         }
3813
3814         return ret;
3815 }
3816
3817 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
3818                                 struct btrfs_block_rsv *block_rsv,
3819                                 u64 num_bytes)
3820 {
3821         int ret;
3822
3823         if (num_bytes == 0)
3824                 return 0;
3825
3826         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, 0);
3827         if (!ret) {
3828                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3829                 return 0;
3830         }
3831
3832         return ret;
3833 }
3834
3835 int btrfs_block_rsv_check(struct btrfs_root *root,
3836                           struct btrfs_block_rsv *block_rsv, int min_factor)
3837 {
3838         u64 num_bytes = 0;
3839         int ret = -ENOSPC;
3840
3841         if (!block_rsv)
3842                 return 0;
3843
3844         spin_lock(&block_rsv->lock);
3845         num_bytes = div_factor(block_rsv->size, min_factor);
3846         if (block_rsv->reserved >= num_bytes)
3847                 ret = 0;
3848         spin_unlock(&block_rsv->lock);
3849
3850         return ret;
3851 }
3852
3853 int btrfs_block_rsv_refill(struct btrfs_root *root,
3854                           struct btrfs_block_rsv *block_rsv,
3855                           u64 min_reserved)
3856 {
3857         u64 num_bytes = 0;
3858         int ret = -ENOSPC;
3859
3860         if (!block_rsv)
3861                 return 0;
3862
3863         spin_lock(&block_rsv->lock);
3864         num_bytes = min_reserved;
3865         if (block_rsv->reserved >= num_bytes)
3866                 ret = 0;
3867         else
3868                 num_bytes -= block_rsv->reserved;
3869         spin_unlock(&block_rsv->lock);
3870
3871         if (!ret)
3872                 return 0;
3873
3874         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, 1);
3875         if (!ret) {
3876                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
3877                 return 0;
3878         }
3879
3880         return ret;
3881 }
3882
3883 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
3884                             struct btrfs_block_rsv *dst_rsv,
3885                             u64 num_bytes)
3886 {
3887         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3888 }
3889
3890 void btrfs_block_rsv_release(struct btrfs_root *root,
3891                              struct btrfs_block_rsv *block_rsv,
3892                              u64 num_bytes)
3893 {
3894         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3895         if (global_rsv->full || global_rsv == block_rsv ||
3896             block_rsv->space_info != global_rsv->space_info)
3897                 global_rsv = NULL;
3898         block_rsv_release_bytes(block_rsv, global_rsv, num_bytes);
3899 }
3900
3901 /*
3902  * helper to calculate size of global block reservation.
3903  * the desired value is sum of space used by extent tree,
3904  * checksum tree and root tree
3905  */
3906 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
3907 {
3908         struct btrfs_space_info *sinfo;
3909         u64 num_bytes;
3910         u64 meta_used;
3911         u64 data_used;
3912         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
3913
3914         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
3915         spin_lock(&sinfo->lock);
3916         data_used = sinfo->bytes_used;
3917         spin_unlock(&sinfo->lock);
3918
3919         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3920         spin_lock(&sinfo->lock);
3921         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
3922                 data_used = 0;
3923         meta_used = sinfo->bytes_used;
3924         spin_unlock(&sinfo->lock);
3925
3926         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
3927                     csum_size * 2;
3928         num_bytes += div64_u64(data_used + meta_used, 50);
3929
3930         if (num_bytes * 3 > meta_used)
3931                 num_bytes = div64_u64(meta_used, 3);
3932
3933         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
3934 }
3935
3936 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
3937 {
3938         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
3939         struct btrfs_space_info *sinfo = block_rsv->space_info;
3940         u64 num_bytes;
3941
3942         num_bytes = calc_global_metadata_size(fs_info);
3943
3944         spin_lock(&block_rsv->lock);
3945         spin_lock(&sinfo->lock);
3946
3947         block_rsv->size = num_bytes;
3948
3949         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
3950                     sinfo->bytes_reserved + sinfo->bytes_readonly +
3951                     sinfo->bytes_may_use;
3952
3953         if (sinfo->total_bytes > num_bytes) {
3954                 num_bytes = sinfo->total_bytes - num_bytes;
3955                 block_rsv->reserved += num_bytes;
3956                 sinfo->bytes_may_use += num_bytes;
3957         }
3958
3959         if (block_rsv->reserved >= block_rsv->size) {
3960                 num_bytes = block_rsv->reserved - block_rsv->size;
3961                 sinfo->bytes_may_use -= num_bytes;
3962                 sinfo->reservation_progress++;
3963                 block_rsv->reserved = block_rsv->size;
3964                 block_rsv->full = 1;
3965         }
3966
3967         spin_unlock(&sinfo->lock);
3968         spin_unlock(&block_rsv->lock);
3969 }
3970
3971 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
3972 {
3973         struct btrfs_space_info *space_info;
3974
3975         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3976         fs_info->chunk_block_rsv.space_info = space_info;
3977
3978         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3979         fs_info->global_block_rsv.space_info = space_info;
3980         fs_info->delalloc_block_rsv.space_info = space_info;
3981         fs_info->trans_block_rsv.space_info = space_info;
3982         fs_info->empty_block_rsv.space_info = space_info;
3983         fs_info->delayed_block_rsv.space_info = space_info;
3984
3985         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
3986         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
3987         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
3988         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
3989         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
3990
3991         update_global_block_rsv(fs_info);
3992 }
3993
3994 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
3995 {
3996         block_rsv_release_bytes(&fs_info->global_block_rsv, NULL, (u64)-1);
3997         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
3998         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
3999         WARN_ON(fs_info->trans_block_rsv.size > 0);
4000         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4001         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4002         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4003         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4004         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4005 }
4006
4007 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4008                                   struct btrfs_root *root)
4009 {
4010         if (!trans->bytes_reserved)
4011                 return;
4012
4013         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4014         trans->bytes_reserved = 0;
4015 }
4016
4017 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4018                                   struct inode *inode)
4019 {
4020         struct btrfs_root *root = BTRFS_I(inode)->root;
4021         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4022         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4023
4024         /*
4025          * We need to hold space in order to delete our orphan item once we've
4026          * added it, so this takes the reservation so we can release it later
4027          * when we are truly done with the orphan item.
4028          */
4029         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4030         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4031 }
4032
4033 void btrfs_orphan_release_metadata(struct inode *inode)
4034 {
4035         struct btrfs_root *root = BTRFS_I(inode)->root;
4036         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4037         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4038 }
4039
4040 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4041                                 struct btrfs_pending_snapshot *pending)
4042 {
4043         struct btrfs_root *root = pending->root;
4044         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4045         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4046         /*
4047          * two for root back/forward refs, two for directory entries
4048          * and one for root of the snapshot.
4049          */
4050         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4051         dst_rsv->space_info = src_rsv->space_info;
4052         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4053 }
4054
4055 /**
4056  * drop_outstanding_extent - drop an outstanding extent
4057  * @inode: the inode we're dropping the extent for
4058  *
4059  * This is called when we are freeing up an outstanding extent, either called
4060  * after an error or after an extent is written.  This will return the number of
4061  * reserved extents that need to be freed.  This must be called with
4062  * BTRFS_I(inode)->lock held.
4063  */
4064 static unsigned drop_outstanding_extent(struct inode *inode)
4065 {
4066         unsigned dropped_extents = 0;
4067
4068         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4069         BTRFS_I(inode)->outstanding_extents--;
4070
4071         /*
4072          * If we have more or the same amount of outsanding extents than we have
4073          * reserved then we need to leave the reserved extents count alone.
4074          */
4075         if (BTRFS_I(inode)->outstanding_extents >=
4076             BTRFS_I(inode)->reserved_extents)
4077                 return 0;
4078
4079         dropped_extents = BTRFS_I(inode)->reserved_extents -
4080                 BTRFS_I(inode)->outstanding_extents;
4081         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4082         return dropped_extents;
4083 }
4084
4085 /**
4086  * calc_csum_metadata_size - return the amount of metada space that must be
4087  *      reserved/free'd for the given bytes.
4088  * @inode: the inode we're manipulating
4089  * @num_bytes: the number of bytes in question
4090  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4091  *
4092  * This adjusts the number of csum_bytes in the inode and then returns the
4093  * correct amount of metadata that must either be reserved or freed.  We
4094  * calculate how many checksums we can fit into one leaf and then divide the
4095  * number of bytes that will need to be checksumed by this value to figure out
4096  * how many checksums will be required.  If we are adding bytes then the number
4097  * may go up and we will return the number of additional bytes that must be
4098  * reserved.  If it is going down we will return the number of bytes that must
4099  * be freed.
4100  *
4101  * This must be called with BTRFS_I(inode)->lock held.
4102  */
4103 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4104                                    int reserve)
4105 {
4106         struct btrfs_root *root = BTRFS_I(inode)->root;
4107         u64 csum_size;
4108         int num_csums_per_leaf;
4109         int num_csums;
4110         int old_csums;
4111
4112         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4113             BTRFS_I(inode)->csum_bytes == 0)
4114                 return 0;
4115
4116         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4117         if (reserve)
4118                 BTRFS_I(inode)->csum_bytes += num_bytes;
4119         else
4120                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4121         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4122         num_csums_per_leaf = (int)div64_u64(csum_size,
4123                                             sizeof(struct btrfs_csum_item) +
4124                                             sizeof(struct btrfs_disk_key));
4125         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4126         num_csums = num_csums + num_csums_per_leaf - 1;
4127         num_csums = num_csums / num_csums_per_leaf;
4128
4129         old_csums = old_csums + num_csums_per_leaf - 1;
4130         old_csums = old_csums / num_csums_per_leaf;
4131
4132         /* No change, no need to reserve more */
4133         if (old_csums == num_csums)
4134                 return 0;
4135
4136         if (reserve)
4137                 return btrfs_calc_trans_metadata_size(root,
4138                                                       num_csums - old_csums);
4139
4140         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4141 }
4142
4143 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4144 {
4145         struct btrfs_root *root = BTRFS_I(inode)->root;
4146         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4147         u64 to_reserve = 0;
4148         unsigned nr_extents = 0;
4149         int flush = 1;
4150         int ret;
4151
4152         if (btrfs_is_free_space_inode(root, inode))
4153                 flush = 0;
4154
4155         if (flush && btrfs_transaction_in_commit(root->fs_info))
4156                 schedule_timeout(1);
4157
4158         num_bytes = ALIGN(num_bytes, root->sectorsize);
4159
4160         spin_lock(&BTRFS_I(inode)->lock);
4161         BTRFS_I(inode)->outstanding_extents++;
4162
4163         if (BTRFS_I(inode)->outstanding_extents >
4164             BTRFS_I(inode)->reserved_extents) {
4165                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4166                         BTRFS_I(inode)->reserved_extents;
4167                 BTRFS_I(inode)->reserved_extents += nr_extents;
4168
4169                 to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4170         }
4171         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4172         spin_unlock(&BTRFS_I(inode)->lock);
4173
4174         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4175         if (ret) {
4176                 u64 to_free = 0;
4177                 unsigned dropped;
4178
4179                 spin_lock(&BTRFS_I(inode)->lock);
4180                 dropped = drop_outstanding_extent(inode);
4181                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4182                 spin_unlock(&BTRFS_I(inode)->lock);
4183                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4184
4185                 /*
4186                  * Somebody could have come in and twiddled with the
4187                  * reservation, so if we have to free more than we would have
4188                  * reserved from this reservation go ahead and release those
4189                  * bytes.
4190                  */
4191                 to_free -= to_reserve;
4192                 if (to_free)
4193                         btrfs_block_rsv_release(root, block_rsv, to_free);
4194                 return ret;
4195         }
4196
4197         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4198
4199         return 0;
4200 }
4201
4202 /**
4203  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4204  * @inode: the inode to release the reservation for
4205  * @num_bytes: the number of bytes we're releasing
4206  *
4207  * This will release the metadata reservation for an inode.  This can be called
4208  * once we complete IO for a given set of bytes to release their metadata
4209  * reservations.
4210  */
4211 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4212 {
4213         struct btrfs_root *root = BTRFS_I(inode)->root;
4214         u64 to_free = 0;
4215         unsigned dropped;
4216
4217         num_bytes = ALIGN(num_bytes, root->sectorsize);
4218         spin_lock(&BTRFS_I(inode)->lock);
4219         dropped = drop_outstanding_extent(inode);
4220
4221         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4222         spin_unlock(&BTRFS_I(inode)->lock);
4223         if (dropped > 0)
4224                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4225
4226         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4227                                 to_free);
4228 }
4229
4230 /**
4231  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4232  * @inode: inode we're writing to
4233  * @num_bytes: the number of bytes we want to allocate
4234  *
4235  * This will do the following things
4236  *
4237  * o reserve space in the data space info for num_bytes
4238  * o reserve space in the metadata space info based on number of outstanding
4239  *   extents and how much csums will be needed
4240  * o add to the inodes ->delalloc_bytes
4241  * o add it to the fs_info's delalloc inodes list.
4242  *
4243  * This will return 0 for success and -ENOSPC if there is no space left.
4244  */
4245 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4246 {
4247         int ret;
4248
4249         ret = btrfs_check_data_free_space(inode, num_bytes);
4250         if (ret)
4251                 return ret;
4252
4253         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4254         if (ret) {
4255                 btrfs_free_reserved_data_space(inode, num_bytes);
4256                 return ret;
4257         }
4258
4259         return 0;
4260 }
4261
4262 /**
4263  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4264  * @inode: inode we're releasing space for
4265  * @num_bytes: the number of bytes we want to free up
4266  *
4267  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4268  * called in the case that we don't need the metadata AND data reservations
4269  * anymore.  So if there is an error or we insert an inline extent.
4270  *
4271  * This function will release the metadata space that was not used and will
4272  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4273  * list if there are no delalloc bytes left.
4274  */
4275 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4276 {
4277         btrfs_delalloc_release_metadata(inode, num_bytes);
4278         btrfs_free_reserved_data_space(inode, num_bytes);
4279 }
4280
4281 static int update_block_group(struct btrfs_trans_handle *trans,
4282                               struct btrfs_root *root,
4283                               u64 bytenr, u64 num_bytes, int alloc)
4284 {
4285         struct btrfs_block_group_cache *cache = NULL;
4286         struct btrfs_fs_info *info = root->fs_info;
4287         u64 total = num_bytes;
4288         u64 old_val;
4289         u64 byte_in_group;
4290         int factor;
4291
4292         /* block accounting for super block */
4293         spin_lock(&info->delalloc_lock);
4294         old_val = btrfs_super_bytes_used(info->super_copy);
4295         if (alloc)
4296                 old_val += num_bytes;
4297         else
4298                 old_val -= num_bytes;
4299         btrfs_set_super_bytes_used(info->super_copy, old_val);
4300         spin_unlock(&info->delalloc_lock);
4301
4302         while (total) {
4303                 cache = btrfs_lookup_block_group(info, bytenr);
4304                 if (!cache)
4305                         return -1;
4306                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4307                                     BTRFS_BLOCK_GROUP_RAID1 |
4308                                     BTRFS_BLOCK_GROUP_RAID10))
4309                         factor = 2;
4310                 else
4311                         factor = 1;
4312                 /*
4313                  * If this block group has free space cache written out, we
4314                  * need to make sure to load it if we are removing space.  This
4315                  * is because we need the unpinning stage to actually add the
4316                  * space back to the block group, otherwise we will leak space.
4317                  */
4318                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4319                         cache_block_group(cache, trans, NULL, 1);
4320
4321                 byte_in_group = bytenr - cache->key.objectid;
4322                 WARN_ON(byte_in_group > cache->key.offset);
4323
4324                 spin_lock(&cache->space_info->lock);
4325                 spin_lock(&cache->lock);
4326
4327                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4328                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4329                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4330
4331                 cache->dirty = 1;
4332                 old_val = btrfs_block_group_used(&cache->item);
4333                 num_bytes = min(total, cache->key.offset - byte_in_group);
4334                 if (alloc) {
4335                         old_val += num_bytes;
4336                         btrfs_set_block_group_used(&cache->item, old_val);
4337                         cache->reserved -= num_bytes;
4338                         cache->space_info->bytes_reserved -= num_bytes;
4339                         cache->space_info->bytes_used += num_bytes;
4340                         cache->space_info->disk_used += num_bytes * factor;
4341                         spin_unlock(&cache->lock);
4342                         spin_unlock(&cache->space_info->lock);
4343                 } else {
4344                         old_val -= num_bytes;
4345                         btrfs_set_block_group_used(&cache->item, old_val);
4346                         cache->pinned += num_bytes;
4347                         cache->space_info->bytes_pinned += num_bytes;
4348                         cache->space_info->bytes_used -= num_bytes;
4349                         cache->space_info->disk_used -= num_bytes * factor;
4350                         spin_unlock(&cache->lock);
4351                         spin_unlock(&cache->space_info->lock);
4352
4353                         set_extent_dirty(info->pinned_extents,
4354                                          bytenr, bytenr + num_bytes - 1,
4355                                          GFP_NOFS | __GFP_NOFAIL);
4356                 }
4357                 btrfs_put_block_group(cache);
4358                 total -= num_bytes;
4359                 bytenr += num_bytes;
4360         }
4361         return 0;
4362 }
4363
4364 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4365 {
4366         struct btrfs_block_group_cache *cache;
4367         u64 bytenr;
4368
4369         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4370         if (!cache)
4371                 return 0;
4372
4373         bytenr = cache->key.objectid;
4374         btrfs_put_block_group(cache);
4375
4376         return bytenr;
4377 }
4378
4379 static int pin_down_extent(struct btrfs_root *root,
4380                            struct btrfs_block_group_cache *cache,
4381                            u64 bytenr, u64 num_bytes, int reserved)
4382 {
4383         spin_lock(&cache->space_info->lock);
4384         spin_lock(&cache->lock);
4385         cache->pinned += num_bytes;
4386         cache->space_info->bytes_pinned += num_bytes;
4387         if (reserved) {
4388                 cache->reserved -= num_bytes;
4389                 cache->space_info->bytes_reserved -= num_bytes;
4390         }
4391         spin_unlock(&cache->lock);
4392         spin_unlock(&cache->space_info->lock);
4393
4394         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4395                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4396         return 0;
4397 }
4398
4399 /*
4400  * this function must be called within transaction
4401  */
4402 int btrfs_pin_extent(struct btrfs_root *root,
4403                      u64 bytenr, u64 num_bytes, int reserved)
4404 {
4405         struct btrfs_block_group_cache *cache;
4406
4407         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4408         BUG_ON(!cache);
4409
4410         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4411
4412         btrfs_put_block_group(cache);
4413         return 0;
4414 }
4415
4416 /*
4417  * this function must be called within transaction
4418  */
4419 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4420                                     struct btrfs_root *root,
4421                                     u64 bytenr, u64 num_bytes)
4422 {
4423         struct btrfs_block_group_cache *cache;
4424
4425         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4426         BUG_ON(!cache);
4427
4428         /*
4429          * pull in the free space cache (if any) so that our pin
4430          * removes the free space from the cache.  We have load_only set
4431          * to one because the slow code to read in the free extents does check
4432          * the pinned extents.
4433          */
4434         cache_block_group(cache, trans, root, 1);
4435
4436         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4437
4438         /* remove us from the free space cache (if we're there at all) */
4439         btrfs_remove_free_space(cache, bytenr, num_bytes);
4440         btrfs_put_block_group(cache);
4441         return 0;
4442 }
4443
4444 /**
4445  * btrfs_update_reserved_bytes - update the block_group and space info counters
4446  * @cache:      The cache we are manipulating
4447  * @num_bytes:  The number of bytes in question
4448  * @reserve:    One of the reservation enums
4449  *
4450  * This is called by the allocator when it reserves space, or by somebody who is
4451  * freeing space that was never actually used on disk.  For example if you
4452  * reserve some space for a new leaf in transaction A and before transaction A
4453  * commits you free that leaf, you call this with reserve set to 0 in order to
4454  * clear the reservation.
4455  *
4456  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4457  * ENOSPC accounting.  For data we handle the reservation through clearing the
4458  * delalloc bits in the io_tree.  We have to do this since we could end up
4459  * allocating less disk space for the amount of data we have reserved in the
4460  * case of compression.
4461  *
4462  * If this is a reservation and the block group has become read only we cannot
4463  * make the reservation and return -EAGAIN, otherwise this function always
4464  * succeeds.
4465  */
4466 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4467                                        u64 num_bytes, int reserve)
4468 {
4469         struct btrfs_space_info *space_info = cache->space_info;
4470         int ret = 0;
4471         spin_lock(&space_info->lock);
4472         spin_lock(&cache->lock);
4473         if (reserve != RESERVE_FREE) {
4474                 if (cache->ro) {
4475                         ret = -EAGAIN;
4476                 } else {
4477                         cache->reserved += num_bytes;
4478                         space_info->bytes_reserved += num_bytes;
4479                         if (reserve == RESERVE_ALLOC) {
4480                                 BUG_ON(space_info->bytes_may_use < num_bytes);
4481                                 space_info->bytes_may_use -= num_bytes;
4482                         }
4483                 }
4484         } else {
4485                 if (cache->ro)
4486                         space_info->bytes_readonly += num_bytes;
4487                 cache->reserved -= num_bytes;
4488                 space_info->bytes_reserved -= num_bytes;
4489                 space_info->reservation_progress++;
4490         }
4491         spin_unlock(&cache->lock);
4492         spin_unlock(&space_info->lock);
4493         return ret;
4494 }
4495
4496 int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4497                                 struct btrfs_root *root)
4498 {
4499         struct btrfs_fs_info *fs_info = root->fs_info;
4500         struct btrfs_caching_control *next;
4501         struct btrfs_caching_control *caching_ctl;
4502         struct btrfs_block_group_cache *cache;
4503
4504         down_write(&fs_info->extent_commit_sem);
4505
4506         list_for_each_entry_safe(caching_ctl, next,
4507                                  &fs_info->caching_block_groups, list) {
4508                 cache = caching_ctl->block_group;
4509                 if (block_group_cache_done(cache)) {
4510                         cache->last_byte_to_unpin = (u64)-1;
4511                         list_del_init(&caching_ctl->list);
4512                         put_caching_control(caching_ctl);
4513                 } else {
4514                         cache->last_byte_to_unpin = caching_ctl->progress;
4515                 }
4516         }
4517
4518         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4519                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4520         else
4521                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4522
4523         up_write(&fs_info->extent_commit_sem);
4524
4525         update_global_block_rsv(fs_info);
4526         return 0;
4527 }
4528
4529 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4530 {
4531         struct btrfs_fs_info *fs_info = root->fs_info;
4532         struct btrfs_block_group_cache *cache = NULL;
4533         u64 len;
4534
4535         while (start <= end) {
4536                 if (!cache ||
4537                     start >= cache->key.objectid + cache->key.offset) {
4538                         if (cache)
4539                                 btrfs_put_block_group(cache);
4540                         cache = btrfs_lookup_block_group(fs_info, start);
4541                         BUG_ON(!cache);
4542                 }
4543
4544                 len = cache->key.objectid + cache->key.offset - start;
4545                 len = min(len, end + 1 - start);
4546
4547                 if (start < cache->last_byte_to_unpin) {
4548                         len = min(len, cache->last_byte_to_unpin - start);
4549                         btrfs_add_free_space(cache, start, len);
4550                 }
4551
4552                 start += len;
4553
4554                 spin_lock(&cache->space_info->lock);
4555                 spin_lock(&cache->lock);
4556                 cache->pinned -= len;
4557                 cache->space_info->bytes_pinned -= len;
4558                 if (cache->ro)
4559                         cache->space_info->bytes_readonly += len;
4560                 spin_unlock(&cache->lock);
4561                 spin_unlock(&cache->space_info->lock);
4562         }
4563
4564         if (cache)
4565                 btrfs_put_block_group(cache);
4566         return 0;
4567 }
4568
4569 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4570                                struct btrfs_root *root)
4571 {
4572         struct btrfs_fs_info *fs_info = root->fs_info;
4573         struct extent_io_tree *unpin;
4574         u64 start;
4575         u64 end;
4576         int ret;
4577
4578         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4579                 unpin = &fs_info->freed_extents[1];
4580         else
4581                 unpin = &fs_info->freed_extents[0];
4582
4583         while (1) {
4584                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4585                                             EXTENT_DIRTY);
4586                 if (ret)
4587                         break;
4588
4589                 if (btrfs_test_opt(root, DISCARD))
4590                         ret = btrfs_discard_extent(root, start,
4591                                                    end + 1 - start, NULL);
4592
4593                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4594                 unpin_extent_range(root, start, end);
4595                 cond_resched();
4596         }
4597
4598         return 0;
4599 }
4600
4601 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4602                                 struct btrfs_root *root,
4603                                 u64 bytenr, u64 num_bytes, u64 parent,
4604                                 u64 root_objectid, u64 owner_objectid,
4605                                 u64 owner_offset, int refs_to_drop,
4606                                 struct btrfs_delayed_extent_op *extent_op)
4607 {
4608         struct btrfs_key key;
4609         struct btrfs_path *path;
4610         struct btrfs_fs_info *info = root->fs_info;
4611         struct btrfs_root *extent_root = info->extent_root;
4612         struct extent_buffer *leaf;
4613         struct btrfs_extent_item *ei;
4614         struct btrfs_extent_inline_ref *iref;
4615         int ret;
4616         int is_data;
4617         int extent_slot = 0;
4618         int found_extent = 0;
4619         int num_to_del = 1;
4620         u32 item_size;
4621         u64 refs;
4622
4623         path = btrfs_alloc_path();
4624         if (!path)
4625                 return -ENOMEM;
4626
4627         path->reada = 1;
4628         path->leave_spinning = 1;
4629
4630         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4631         BUG_ON(!is_data && refs_to_drop != 1);
4632
4633         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4634                                     bytenr, num_bytes, parent,
4635                                     root_objectid, owner_objectid,
4636                                     owner_offset);
4637         if (ret == 0) {
4638                 extent_slot = path->slots[0];
4639                 while (extent_slot >= 0) {
4640                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4641                                               extent_slot);
4642                         if (key.objectid != bytenr)
4643                                 break;
4644                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4645                             key.offset == num_bytes) {
4646                                 found_extent = 1;
4647                                 break;
4648                         }
4649                         if (path->slots[0] - extent_slot > 5)
4650                                 break;
4651                         extent_slot--;
4652                 }
4653 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4654                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4655                 if (found_extent && item_size < sizeof(*ei))
4656                         found_extent = 0;
4657 #endif
4658                 if (!found_extent) {
4659                         BUG_ON(iref);
4660                         ret = remove_extent_backref(trans, extent_root, path,
4661                                                     NULL, refs_to_drop,
4662                                                     is_data);
4663                         BUG_ON(ret);
4664                         btrfs_release_path(path);
4665                         path->leave_spinning = 1;
4666
4667                         key.objectid = bytenr;
4668                         key.type = BTRFS_EXTENT_ITEM_KEY;
4669                         key.offset = num_bytes;
4670
4671                         ret = btrfs_search_slot(trans, extent_root,
4672                                                 &key, path, -1, 1);
4673                         if (ret) {
4674                                 printk(KERN_ERR "umm, got %d back from search"
4675                                        ", was looking for %llu\n", ret,
4676                                        (unsigned long long)bytenr);
4677                                 if (ret > 0)
4678                                         btrfs_print_leaf(extent_root,
4679                                                          path->nodes[0]);
4680                         }
4681                         BUG_ON(ret);
4682                         extent_slot = path->slots[0];
4683                 }
4684         } else {
4685                 btrfs_print_leaf(extent_root, path->nodes[0]);
4686                 WARN_ON(1);
4687                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4688                        "parent %llu root %llu  owner %llu offset %llu\n",
4689                        (unsigned long long)bytenr,
4690                        (unsigned long long)parent,
4691                        (unsigned long long)root_objectid,
4692                        (unsigned long long)owner_objectid,
4693                        (unsigned long long)owner_offset);
4694         }
4695
4696         leaf = path->nodes[0];
4697         item_size = btrfs_item_size_nr(leaf, extent_slot);
4698 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4699         if (item_size < sizeof(*ei)) {
4700                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4701                 ret = convert_extent_item_v0(trans, extent_root, path,
4702                                              owner_objectid, 0);
4703                 BUG_ON(ret < 0);
4704
4705                 btrfs_release_path(path);
4706                 path->leave_spinning = 1;
4707
4708                 key.objectid = bytenr;
4709                 key.type = BTRFS_EXTENT_ITEM_KEY;
4710                 key.offset = num_bytes;
4711
4712                 ret = btrfs_search_slot(trans, extent_root, &key, path,
4713                                         -1, 1);
4714                 if (ret) {
4715                         printk(KERN_ERR "umm, got %d back from search"
4716                                ", was looking for %llu\n", ret,
4717                                (unsigned long long)bytenr);
4718                         btrfs_print_leaf(extent_root, path->nodes[0]);
4719                 }
4720                 BUG_ON(ret);
4721                 extent_slot = path->slots[0];
4722                 leaf = path->nodes[0];
4723                 item_size = btrfs_item_size_nr(leaf, extent_slot);
4724         }
4725 #endif
4726         BUG_ON(item_size < sizeof(*ei));
4727         ei = btrfs_item_ptr(leaf, extent_slot,
4728                             struct btrfs_extent_item);
4729         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
4730                 struct btrfs_tree_block_info *bi;
4731                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4732                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4733                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4734         }
4735
4736         refs = btrfs_extent_refs(leaf, ei);
4737         BUG_ON(refs < refs_to_drop);
4738         refs -= refs_to_drop;
4739
4740         if (refs > 0) {
4741                 if (extent_op)
4742                         __run_delayed_extent_op(extent_op, leaf, ei);
4743                 /*
4744                  * In the case of inline back ref, reference count will
4745                  * be updated by remove_extent_backref
4746                  */
4747                 if (iref) {
4748                         BUG_ON(!found_extent);
4749                 } else {
4750                         btrfs_set_extent_refs(leaf, ei, refs);
4751                         btrfs_mark_buffer_dirty(leaf);
4752                 }
4753                 if (found_extent) {
4754                         ret = remove_extent_backref(trans, extent_root, path,
4755                                                     iref, refs_to_drop,
4756                                                     is_data);
4757                         BUG_ON(ret);
4758                 }
4759         } else {
4760                 if (found_extent) {
4761                         BUG_ON(is_data && refs_to_drop !=
4762                                extent_data_ref_count(root, path, iref));
4763                         if (iref) {
4764                                 BUG_ON(path->slots[0] != extent_slot);
4765                         } else {
4766                                 BUG_ON(path->slots[0] != extent_slot + 1);
4767                                 path->slots[0] = extent_slot;
4768                                 num_to_del = 2;
4769                         }
4770                 }
4771
4772                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
4773                                       num_to_del);
4774                 BUG_ON(ret);
4775                 btrfs_release_path(path);
4776
4777                 if (is_data) {
4778                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
4779                         BUG_ON(ret);
4780                 } else {
4781                         invalidate_mapping_pages(info->btree_inode->i_mapping,
4782                              bytenr >> PAGE_CACHE_SHIFT,
4783                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
4784                 }
4785
4786                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
4787                 BUG_ON(ret);
4788         }
4789         btrfs_free_path(path);
4790         return ret;
4791 }
4792
4793 /*
4794  * when we free an block, it is possible (and likely) that we free the last
4795  * delayed ref for that extent as well.  This searches the delayed ref tree for
4796  * a given extent, and if there are no other delayed refs to be processed, it
4797  * removes it from the tree.
4798  */
4799 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
4800                                       struct btrfs_root *root, u64 bytenr)
4801 {
4802         struct btrfs_delayed_ref_head *head;
4803         struct btrfs_delayed_ref_root *delayed_refs;
4804         struct btrfs_delayed_ref_node *ref;
4805         struct rb_node *node;
4806         int ret = 0;
4807
4808         delayed_refs = &trans->transaction->delayed_refs;
4809         spin_lock(&delayed_refs->lock);
4810         head = btrfs_find_delayed_ref_head(trans, bytenr);
4811         if (!head)
4812                 goto out;
4813
4814         node = rb_prev(&head->node.rb_node);
4815         if (!node)
4816                 goto out;
4817
4818         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
4819
4820         /* there are still entries for this ref, we can't drop it */
4821         if (ref->bytenr == bytenr)
4822                 goto out;
4823
4824         if (head->extent_op) {
4825                 if (!head->must_insert_reserved)
4826                         goto out;
4827                 kfree(head->extent_op);
4828                 head->extent_op = NULL;
4829         }
4830
4831         /*
4832          * waiting for the lock here would deadlock.  If someone else has it
4833          * locked they are already in the process of dropping it anyway
4834          */
4835         if (!mutex_trylock(&head->mutex))
4836                 goto out;
4837
4838         /*
4839          * at this point we have a head with no other entries.  Go
4840          * ahead and process it.
4841          */
4842         head->node.in_tree = 0;
4843         rb_erase(&head->node.rb_node, &delayed_refs->root);
4844
4845         delayed_refs->num_entries--;
4846
4847         /*
4848          * we don't take a ref on the node because we're removing it from the
4849          * tree, so we just steal the ref the tree was holding.
4850          */
4851         delayed_refs->num_heads--;
4852         if (list_empty(&head->cluster))
4853                 delayed_refs->num_heads_ready--;
4854
4855         list_del_init(&head->cluster);
4856         spin_unlock(&delayed_refs->lock);
4857
4858         BUG_ON(head->extent_op);
4859         if (head->must_insert_reserved)
4860                 ret = 1;
4861
4862         mutex_unlock(&head->mutex);
4863         btrfs_put_delayed_ref(&head->node);
4864         return ret;
4865 out:
4866         spin_unlock(&delayed_refs->lock);
4867         return 0;
4868 }
4869
4870 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
4871                            struct btrfs_root *root,
4872                            struct extent_buffer *buf,
4873                            u64 parent, int last_ref)
4874 {
4875         struct btrfs_block_group_cache *cache = NULL;
4876         int ret;
4877
4878         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4879                 ret = btrfs_add_delayed_tree_ref(trans, buf->start, buf->len,
4880                                                 parent, root->root_key.objectid,
4881                                                 btrfs_header_level(buf),
4882                                                 BTRFS_DROP_DELAYED_REF, NULL);
4883                 BUG_ON(ret);
4884         }
4885
4886         if (!last_ref)
4887                 return;
4888
4889         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
4890
4891         if (btrfs_header_generation(buf) == trans->transid) {
4892                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4893                         ret = check_ref_cleanup(trans, root, buf->start);
4894                         if (!ret)
4895                                 goto out;
4896                 }
4897
4898                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
4899                         pin_down_extent(root, cache, buf->start, buf->len, 1);
4900                         goto out;
4901                 }
4902
4903                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
4904
4905                 btrfs_add_free_space(cache, buf->start, buf->len);
4906                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
4907         }
4908 out:
4909         /*
4910          * Deleting the buffer, clear the corrupt flag since it doesn't matter
4911          * anymore.
4912          */
4913         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
4914         btrfs_put_block_group(cache);
4915 }
4916
4917 int btrfs_free_extent(struct btrfs_trans_handle *trans,
4918                       struct btrfs_root *root,
4919                       u64 bytenr, u64 num_bytes, u64 parent,
4920                       u64 root_objectid, u64 owner, u64 offset)
4921 {
4922         int ret;
4923
4924         /*
4925          * tree log blocks never actually go into the extent allocation
4926          * tree, just update pinning info and exit early.
4927          */
4928         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
4929                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
4930                 /* unlocks the pinned mutex */
4931                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
4932                 ret = 0;
4933         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
4934                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
4935                                         parent, root_objectid, (int)owner,
4936                                         BTRFS_DROP_DELAYED_REF, NULL);
4937                 BUG_ON(ret);
4938         } else {
4939                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
4940                                         parent, root_objectid, owner,
4941                                         offset, BTRFS_DROP_DELAYED_REF, NULL);
4942                 BUG_ON(ret);
4943         }
4944         return ret;
4945 }
4946
4947 static u64 stripe_align(struct btrfs_root *root, u64 val)
4948 {
4949         u64 mask = ((u64)root->stripesize - 1);
4950         u64 ret = (val + mask) & ~mask;
4951         return ret;
4952 }
4953
4954 /*
4955  * when we wait for progress in the block group caching, its because
4956  * our allocation attempt failed at least once.  So, we must sleep
4957  * and let some progress happen before we try again.
4958  *
4959  * This function will sleep at least once waiting for new free space to
4960  * show up, and then it will check the block group free space numbers
4961  * for our min num_bytes.  Another option is to have it go ahead
4962  * and look in the rbtree for a free extent of a given size, but this
4963  * is a good start.
4964  */
4965 static noinline int
4966 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
4967                                 u64 num_bytes)
4968 {
4969         struct btrfs_caching_control *caching_ctl;
4970         DEFINE_WAIT(wait);
4971
4972         caching_ctl = get_caching_control(cache);
4973         if (!caching_ctl)
4974                 return 0;
4975
4976         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
4977                    (cache->free_space_ctl->free_space >= num_bytes));
4978
4979         put_caching_control(caching_ctl);
4980         return 0;
4981 }
4982
4983 static noinline int
4984 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
4985 {
4986         struct btrfs_caching_control *caching_ctl;
4987         DEFINE_WAIT(wait);
4988
4989         caching_ctl = get_caching_control(cache);
4990         if (!caching_ctl)
4991                 return 0;
4992
4993         wait_event(caching_ctl->wait, block_group_cache_done(cache));
4994
4995         put_caching_control(caching_ctl);
4996         return 0;
4997 }
4998
4999 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5000 {
5001         int index;
5002         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
5003                 index = 0;
5004         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
5005                 index = 1;
5006         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
5007                 index = 2;
5008         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
5009                 index = 3;
5010         else
5011                 index = 4;
5012         return index;
5013 }
5014
5015 enum btrfs_loop_type {
5016         LOOP_FIND_IDEAL = 0,
5017         LOOP_CACHING_NOWAIT = 1,
5018         LOOP_CACHING_WAIT = 2,
5019         LOOP_ALLOC_CHUNK = 3,
5020         LOOP_NO_EMPTY_SIZE = 4,
5021 };
5022
5023 /*
5024  * walks the btree of allocated extents and find a hole of a given size.
5025  * The key ins is changed to record the hole:
5026  * ins->objectid == block start
5027  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5028  * ins->offset == number of blocks
5029  * Any available blocks before search_start are skipped.
5030  */
5031 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5032                                      struct btrfs_root *orig_root,
5033                                      u64 num_bytes, u64 empty_size,
5034                                      u64 search_start, u64 search_end,
5035                                      u64 hint_byte, struct btrfs_key *ins,
5036                                      u64 data)
5037 {
5038         int ret = 0;
5039         struct btrfs_root *root = orig_root->fs_info->extent_root;
5040         struct btrfs_free_cluster *last_ptr = NULL;
5041         struct btrfs_block_group_cache *block_group = NULL;
5042         int empty_cluster = 2 * 1024 * 1024;
5043         int allowed_chunk_alloc = 0;
5044         int done_chunk_alloc = 0;
5045         struct btrfs_space_info *space_info;
5046         int last_ptr_loop = 0;
5047         int loop = 0;
5048         int index = 0;
5049         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5050                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5051         bool found_uncached_bg = false;
5052         bool failed_cluster_refill = false;
5053         bool failed_alloc = false;
5054         bool use_cluster = true;
5055         bool have_caching_bg = false;
5056         u64 ideal_cache_percent = 0;
5057         u64 ideal_cache_offset = 0;
5058
5059         WARN_ON(num_bytes < root->sectorsize);
5060         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5061         ins->objectid = 0;
5062         ins->offset = 0;
5063
5064         space_info = __find_space_info(root->fs_info, data);
5065         if (!space_info) {
5066                 printk(KERN_ERR "No space info for %llu\n", data);
5067                 return -ENOSPC;
5068         }
5069
5070         /*
5071          * If the space info is for both data and metadata it means we have a
5072          * small filesystem and we can't use the clustering stuff.
5073          */
5074         if (btrfs_mixed_space_info(space_info))
5075                 use_cluster = false;
5076
5077         if (orig_root->ref_cows || empty_size)
5078                 allowed_chunk_alloc = 1;
5079
5080         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5081                 last_ptr = &root->fs_info->meta_alloc_cluster;
5082                 if (!btrfs_test_opt(root, SSD))
5083                         empty_cluster = 64 * 1024;
5084         }
5085
5086         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5087             btrfs_test_opt(root, SSD)) {
5088                 last_ptr = &root->fs_info->data_alloc_cluster;
5089         }
5090
5091         if (last_ptr) {
5092                 spin_lock(&last_ptr->lock);
5093                 if (last_ptr->block_group)
5094                         hint_byte = last_ptr->window_start;
5095                 spin_unlock(&last_ptr->lock);
5096         }
5097
5098         search_start = max(search_start, first_logical_byte(root, 0));
5099         search_start = max(search_start, hint_byte);
5100
5101         if (!last_ptr)
5102                 empty_cluster = 0;
5103
5104         if (search_start == hint_byte) {
5105 ideal_cache:
5106                 block_group = btrfs_lookup_block_group(root->fs_info,
5107                                                        search_start);
5108                 /*
5109                  * we don't want to use the block group if it doesn't match our
5110                  * allocation bits, or if its not cached.
5111                  *
5112                  * However if we are re-searching with an ideal block group
5113                  * picked out then we don't care that the block group is cached.
5114                  */
5115                 if (block_group && block_group_bits(block_group, data) &&
5116                     (block_group->cached != BTRFS_CACHE_NO ||
5117                      search_start == ideal_cache_offset)) {
5118                         down_read(&space_info->groups_sem);
5119                         if (list_empty(&block_group->list) ||
5120                             block_group->ro) {
5121                                 /*
5122                                  * someone is removing this block group,
5123                                  * we can't jump into the have_block_group
5124                                  * target because our list pointers are not
5125                                  * valid
5126                                  */
5127                                 btrfs_put_block_group(block_group);
5128                                 up_read(&space_info->groups_sem);
5129                         } else {
5130                                 index = get_block_group_index(block_group);
5131                                 goto have_block_group;
5132                         }
5133                 } else if (block_group) {
5134                         btrfs_put_block_group(block_group);
5135                 }
5136         }
5137 search:
5138         have_caching_bg = false;
5139         down_read(&space_info->groups_sem);
5140         list_for_each_entry(block_group, &space_info->block_groups[index],
5141                             list) {
5142                 u64 offset;
5143                 int cached;
5144
5145                 btrfs_get_block_group(block_group);
5146                 search_start = block_group->key.objectid;
5147
5148                 /*
5149                  * this can happen if we end up cycling through all the
5150                  * raid types, but we want to make sure we only allocate
5151                  * for the proper type.
5152                  */
5153                 if (!block_group_bits(block_group, data)) {
5154                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5155                                 BTRFS_BLOCK_GROUP_RAID1 |
5156                                 BTRFS_BLOCK_GROUP_RAID10;
5157
5158                         /*
5159                          * if they asked for extra copies and this block group
5160                          * doesn't provide them, bail.  This does allow us to
5161                          * fill raid0 from raid1.
5162                          */
5163                         if ((data & extra) && !(block_group->flags & extra))
5164                                 goto loop;
5165                 }
5166
5167 have_block_group:
5168                 if (unlikely(block_group->cached == BTRFS_CACHE_NO)) {
5169                         u64 free_percent;
5170
5171                         ret = cache_block_group(block_group, trans,
5172                                                 orig_root, 1);
5173                         if (block_group->cached == BTRFS_CACHE_FINISHED)
5174                                 goto have_block_group;
5175
5176                         free_percent = btrfs_block_group_used(&block_group->item);
5177                         free_percent *= 100;
5178                         free_percent = div64_u64(free_percent,
5179                                                  block_group->key.offset);
5180                         free_percent = 100 - free_percent;
5181                         if (free_percent > ideal_cache_percent &&
5182                             likely(!block_group->ro)) {
5183                                 ideal_cache_offset = block_group->key.objectid;
5184                                 ideal_cache_percent = free_percent;
5185                         }
5186
5187                         /*
5188                          * The caching workers are limited to 2 threads, so we
5189                          * can queue as much work as we care to.
5190                          */
5191                         if (loop > LOOP_FIND_IDEAL) {
5192                                 ret = cache_block_group(block_group, trans,
5193                                                         orig_root, 0);
5194                                 BUG_ON(ret);
5195                         }
5196                         found_uncached_bg = true;
5197
5198                         /*
5199                          * If loop is set for cached only, try the next block
5200                          * group.
5201                          */
5202                         if (loop == LOOP_FIND_IDEAL)
5203                                 goto loop;
5204                 }
5205
5206                 cached = block_group_cache_done(block_group);
5207                 if (unlikely(!cached))
5208                         found_uncached_bg = true;
5209
5210                 if (unlikely(block_group->ro))
5211                         goto loop;
5212
5213                 spin_lock(&block_group->free_space_ctl->tree_lock);
5214                 if (cached &&
5215                     block_group->free_space_ctl->free_space <
5216                     num_bytes + empty_size) {
5217                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5218                         goto loop;
5219                 }
5220                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5221
5222                 /*
5223                  * Ok we want to try and use the cluster allocator, so lets look
5224                  * there, unless we are on LOOP_NO_EMPTY_SIZE, since we will
5225                  * have tried the cluster allocator plenty of times at this
5226                  * point and not have found anything, so we are likely way too
5227                  * fragmented for the clustering stuff to find anything, so lets
5228                  * just skip it and let the allocator find whatever block it can
5229                  * find
5230                  */
5231                 if (last_ptr && loop < LOOP_NO_EMPTY_SIZE) {
5232                         /*
5233                          * the refill lock keeps out other
5234                          * people trying to start a new cluster
5235                          */
5236                         spin_lock(&last_ptr->refill_lock);
5237                         if (last_ptr->block_group &&
5238                             (last_ptr->block_group->ro ||
5239                             !block_group_bits(last_ptr->block_group, data))) {
5240                                 offset = 0;
5241                                 goto refill_cluster;
5242                         }
5243
5244                         offset = btrfs_alloc_from_cluster(block_group, last_ptr,
5245                                                  num_bytes, search_start);
5246                         if (offset) {
5247                                 /* we have a block, we're done */
5248                                 spin_unlock(&last_ptr->refill_lock);
5249                                 goto checks;
5250                         }
5251
5252                         spin_lock(&last_ptr->lock);
5253                         /*
5254                          * whoops, this cluster doesn't actually point to
5255                          * this block group.  Get a ref on the block
5256                          * group is does point to and try again
5257                          */
5258                         if (!last_ptr_loop && last_ptr->block_group &&
5259                             last_ptr->block_group != block_group &&
5260                             index <=
5261                                  get_block_group_index(last_ptr->block_group)) {
5262
5263                                 btrfs_put_block_group(block_group);
5264                                 block_group = last_ptr->block_group;
5265                                 btrfs_get_block_group(block_group);
5266                                 spin_unlock(&last_ptr->lock);
5267                                 spin_unlock(&last_ptr->refill_lock);
5268
5269                                 last_ptr_loop = 1;
5270                                 search_start = block_group->key.objectid;
5271                                 /*
5272                                  * we know this block group is properly
5273                                  * in the list because
5274                                  * btrfs_remove_block_group, drops the
5275                                  * cluster before it removes the block
5276                                  * group from the list
5277                                  */
5278                                 goto have_block_group;
5279                         }
5280                         spin_unlock(&last_ptr->lock);
5281 refill_cluster:
5282                         /*
5283                          * this cluster didn't work out, free it and
5284                          * start over
5285                          */
5286                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5287
5288                         last_ptr_loop = 0;
5289
5290                         /* allocate a cluster in this block group */
5291                         ret = btrfs_find_space_cluster(trans, root,
5292                                                block_group, last_ptr,
5293                                                offset, num_bytes,
5294                                                empty_cluster + empty_size);
5295                         if (ret == 0) {
5296                                 /*
5297                                  * now pull our allocation out of this
5298                                  * cluster
5299                                  */
5300                                 offset = btrfs_alloc_from_cluster(block_group,
5301                                                   last_ptr, num_bytes,
5302                                                   search_start);
5303                                 if (offset) {
5304                                         /* we found one, proceed */
5305                                         spin_unlock(&last_ptr->refill_lock);
5306                                         goto checks;
5307                                 }
5308                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5309                                    && !failed_cluster_refill) {
5310                                 spin_unlock(&last_ptr->refill_lock);
5311
5312                                 failed_cluster_refill = true;
5313                                 wait_block_group_cache_progress(block_group,
5314                                        num_bytes + empty_cluster + empty_size);
5315                                 goto have_block_group;
5316                         }
5317
5318                         /*
5319                          * at this point we either didn't find a cluster
5320                          * or we weren't able to allocate a block from our
5321                          * cluster.  Free the cluster we've been trying
5322                          * to use, and go to the next block group
5323                          */
5324                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5325                         spin_unlock(&last_ptr->refill_lock);
5326                         goto loop;
5327                 }
5328
5329                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5330                                                     num_bytes, empty_size);
5331                 /*
5332                  * If we didn't find a chunk, and we haven't failed on this
5333                  * block group before, and this block group is in the middle of
5334                  * caching and we are ok with waiting, then go ahead and wait
5335                  * for progress to be made, and set failed_alloc to true.
5336                  *
5337                  * If failed_alloc is true then we've already waited on this
5338                  * block group once and should move on to the next block group.
5339                  */
5340                 if (!offset && !failed_alloc && !cached &&
5341                     loop > LOOP_CACHING_NOWAIT) {
5342                         wait_block_group_cache_progress(block_group,
5343                                                 num_bytes + empty_size);
5344                         failed_alloc = true;
5345                         goto have_block_group;
5346                 } else if (!offset) {
5347                         if (!cached)
5348                                 have_caching_bg = true;
5349                         goto loop;
5350                 }
5351 checks:
5352                 search_start = stripe_align(root, offset);
5353                 /* move on to the next group */
5354                 if (search_start + num_bytes >= search_end) {
5355                         btrfs_add_free_space(block_group, offset, num_bytes);
5356                         goto loop;
5357                 }
5358
5359                 /* move on to the next group */
5360                 if (search_start + num_bytes >
5361                     block_group->key.objectid + block_group->key.offset) {
5362                         btrfs_add_free_space(block_group, offset, num_bytes);
5363                         goto loop;
5364                 }
5365
5366                 ins->objectid = search_start;
5367                 ins->offset = num_bytes;
5368
5369                 if (offset < search_start)
5370                         btrfs_add_free_space(block_group, offset,
5371                                              search_start - offset);
5372                 BUG_ON(offset > search_start);
5373
5374                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
5375                                                   alloc_type);
5376                 if (ret == -EAGAIN) {
5377                         btrfs_add_free_space(block_group, offset, num_bytes);
5378                         goto loop;
5379                 }
5380
5381                 /* we are all good, lets return */
5382                 ins->objectid = search_start;
5383                 ins->offset = num_bytes;
5384
5385                 if (offset < search_start)
5386                         btrfs_add_free_space(block_group, offset,
5387                                              search_start - offset);
5388                 BUG_ON(offset > search_start);
5389                 btrfs_put_block_group(block_group);
5390                 break;
5391 loop:
5392                 failed_cluster_refill = false;
5393                 failed_alloc = false;
5394                 BUG_ON(index != get_block_group_index(block_group));
5395                 btrfs_put_block_group(block_group);
5396         }
5397         up_read(&space_info->groups_sem);
5398
5399         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
5400                 goto search;
5401
5402         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5403                 goto search;
5404
5405         /* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
5406          *                      for them to make caching progress.  Also
5407          *                      determine the best possible bg to cache
5408          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5409          *                      caching kthreads as we move along
5410          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5411          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5412          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5413          *                      again
5414          */
5415         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5416                 index = 0;
5417                 if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
5418                         found_uncached_bg = false;
5419                         loop++;
5420                         if (!ideal_cache_percent)
5421                                 goto search;
5422
5423                         /*
5424                          * 1 of the following 2 things have happened so far
5425                          *
5426                          * 1) We found an ideal block group for caching that
5427                          * is mostly full and will cache quickly, so we might
5428                          * as well wait for it.
5429                          *
5430                          * 2) We searched for cached only and we didn't find
5431                          * anything, and we didn't start any caching kthreads
5432                          * either, so chances are we will loop through and
5433                          * start a couple caching kthreads, and then come back
5434                          * around and just wait for them.  This will be slower
5435                          * because we will have 2 caching kthreads reading at
5436                          * the same time when we could have just started one
5437                          * and waited for it to get far enough to give us an
5438                          * allocation, so go ahead and go to the wait caching
5439                          * loop.
5440                          */
5441                         loop = LOOP_CACHING_WAIT;
5442                         search_start = ideal_cache_offset;
5443                         ideal_cache_percent = 0;
5444                         goto ideal_cache;
5445                 } else if (loop == LOOP_FIND_IDEAL) {
5446                         /*
5447                          * Didn't find a uncached bg, wait on anything we find
5448                          * next.
5449                          */
5450                         loop = LOOP_CACHING_WAIT;
5451                         goto search;
5452                 }
5453
5454                 loop++;
5455
5456                 if (loop == LOOP_ALLOC_CHUNK) {
5457                        if (allowed_chunk_alloc) {
5458                                 ret = do_chunk_alloc(trans, root, num_bytes +
5459                                                      2 * 1024 * 1024, data,
5460                                                      CHUNK_ALLOC_LIMITED);
5461                                 allowed_chunk_alloc = 0;
5462                                 if (ret == 1)
5463                                         done_chunk_alloc = 1;
5464                         } else if (!done_chunk_alloc &&
5465                                    space_info->force_alloc ==
5466                                    CHUNK_ALLOC_NO_FORCE) {
5467                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5468                         }
5469
5470                        /*
5471                         * We didn't allocate a chunk, go ahead and drop the
5472                         * empty size and loop again.
5473                         */
5474                        if (!done_chunk_alloc)
5475                                loop = LOOP_NO_EMPTY_SIZE;
5476                 }
5477
5478                 if (loop == LOOP_NO_EMPTY_SIZE) {
5479                         empty_size = 0;
5480                         empty_cluster = 0;
5481                 }
5482
5483                 goto search;
5484         } else if (!ins->objectid) {
5485                 ret = -ENOSPC;
5486         } else if (ins->objectid) {
5487                 ret = 0;
5488         }
5489
5490         return ret;
5491 }
5492
5493 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5494                             int dump_block_groups)
5495 {
5496         struct btrfs_block_group_cache *cache;
5497         int index = 0;
5498
5499         spin_lock(&info->lock);
5500         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5501                (unsigned long long)info->flags,
5502                (unsigned long long)(info->total_bytes - info->bytes_used -
5503                                     info->bytes_pinned - info->bytes_reserved -
5504                                     info->bytes_readonly),
5505                (info->full) ? "" : "not ");
5506         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5507                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5508                (unsigned long long)info->total_bytes,
5509                (unsigned long long)info->bytes_used,
5510                (unsigned long long)info->bytes_pinned,
5511                (unsigned long long)info->bytes_reserved,
5512                (unsigned long long)info->bytes_may_use,
5513                (unsigned long long)info->bytes_readonly);
5514         spin_unlock(&info->lock);
5515
5516         if (!dump_block_groups)
5517                 return;
5518
5519         down_read(&info->groups_sem);
5520 again:
5521         list_for_each_entry(cache, &info->block_groups[index], list) {
5522                 spin_lock(&cache->lock);
5523                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5524                        "%llu pinned %llu reserved\n",
5525                        (unsigned long long)cache->key.objectid,
5526                        (unsigned long long)cache->key.offset,
5527                        (unsigned long long)btrfs_block_group_used(&cache->item),
5528                        (unsigned long long)cache->pinned,
5529                        (unsigned long long)cache->reserved);
5530                 btrfs_dump_free_space(cache, bytes);
5531                 spin_unlock(&cache->lock);
5532         }
5533         if (++index < BTRFS_NR_RAID_TYPES)
5534                 goto again;
5535         up_read(&info->groups_sem);
5536 }
5537
5538 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5539                          struct btrfs_root *root,
5540                          u64 num_bytes, u64 min_alloc_size,
5541                          u64 empty_size, u64 hint_byte,
5542                          u64 search_end, struct btrfs_key *ins,
5543                          u64 data)
5544 {
5545         int ret;
5546         u64 search_start = 0;
5547
5548         data = btrfs_get_alloc_profile(root, data);
5549 again:
5550         /*
5551          * the only place that sets empty_size is btrfs_realloc_node, which
5552          * is not called recursively on allocations
5553          */
5554         if (empty_size || root->ref_cows)
5555                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5556                                      num_bytes + 2 * 1024 * 1024, data,
5557                                      CHUNK_ALLOC_NO_FORCE);
5558
5559         WARN_ON(num_bytes < root->sectorsize);
5560         ret = find_free_extent(trans, root, num_bytes, empty_size,
5561                                search_start, search_end, hint_byte,
5562                                ins, data);
5563
5564         if (ret == -ENOSPC && num_bytes > min_alloc_size) {
5565                 num_bytes = num_bytes >> 1;
5566                 num_bytes = num_bytes & ~(root->sectorsize - 1);
5567                 num_bytes = max(num_bytes, min_alloc_size);
5568                 do_chunk_alloc(trans, root->fs_info->extent_root,
5569                                num_bytes, data, CHUNK_ALLOC_FORCE);
5570                 goto again;
5571         }
5572         if (ret == -ENOSPC && btrfs_test_opt(root, ENOSPC_DEBUG)) {
5573                 struct btrfs_space_info *sinfo;
5574
5575                 sinfo = __find_space_info(root->fs_info, data);
5576                 printk(KERN_ERR "btrfs allocation failed flags %llu, "
5577                        "wanted %llu\n", (unsigned long long)data,
5578                        (unsigned long long)num_bytes);
5579                 dump_space_info(sinfo, num_bytes, 1);
5580         }
5581
5582         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
5583
5584         return ret;
5585 }
5586
5587 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
5588                                         u64 start, u64 len, int pin)
5589 {
5590         struct btrfs_block_group_cache *cache;
5591         int ret = 0;
5592
5593         cache = btrfs_lookup_block_group(root->fs_info, start);
5594         if (!cache) {
5595                 printk(KERN_ERR "Unable to find block group for %llu\n",
5596                        (unsigned long long)start);
5597                 return -ENOSPC;
5598         }
5599
5600         if (btrfs_test_opt(root, DISCARD))
5601                 ret = btrfs_discard_extent(root, start, len, NULL);
5602
5603         if (pin)
5604                 pin_down_extent(root, cache, start, len, 1);
5605         else {
5606                 btrfs_add_free_space(cache, start, len);
5607                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
5608         }
5609         btrfs_put_block_group(cache);
5610
5611         trace_btrfs_reserved_extent_free(root, start, len);
5612
5613         return ret;
5614 }
5615
5616 int btrfs_free_reserved_extent(struct btrfs_root *root,
5617                                         u64 start, u64 len)
5618 {
5619         return __btrfs_free_reserved_extent(root, start, len, 0);
5620 }
5621
5622 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
5623                                        u64 start, u64 len)
5624 {
5625         return __btrfs_free_reserved_extent(root, start, len, 1);
5626 }
5627
5628 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5629                                       struct btrfs_root *root,
5630                                       u64 parent, u64 root_objectid,
5631                                       u64 flags, u64 owner, u64 offset,
5632                                       struct btrfs_key *ins, int ref_mod)
5633 {
5634         int ret;
5635         struct btrfs_fs_info *fs_info = root->fs_info;
5636         struct btrfs_extent_item *extent_item;
5637         struct btrfs_extent_inline_ref *iref;
5638         struct btrfs_path *path;
5639         struct extent_buffer *leaf;
5640         int type;
5641         u32 size;
5642
5643         if (parent > 0)
5644                 type = BTRFS_SHARED_DATA_REF_KEY;
5645         else
5646                 type = BTRFS_EXTENT_DATA_REF_KEY;
5647
5648         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5649
5650         path = btrfs_alloc_path();
5651         if (!path)
5652                 return -ENOMEM;
5653
5654         path->leave_spinning = 1;
5655         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5656                                       ins, size);
5657         BUG_ON(ret);
5658
5659         leaf = path->nodes[0];
5660         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5661                                      struct btrfs_extent_item);
5662         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
5663         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5664         btrfs_set_extent_flags(leaf, extent_item,
5665                                flags | BTRFS_EXTENT_FLAG_DATA);
5666
5667         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
5668         btrfs_set_extent_inline_ref_type(leaf, iref, type);
5669         if (parent > 0) {
5670                 struct btrfs_shared_data_ref *ref;
5671                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
5672                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5673                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
5674         } else {
5675                 struct btrfs_extent_data_ref *ref;
5676                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
5677                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
5678                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
5679                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
5680                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
5681         }
5682
5683         btrfs_mark_buffer_dirty(path->nodes[0]);
5684         btrfs_free_path(path);
5685
5686         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5687         if (ret) {
5688                 printk(KERN_ERR "btrfs update block group failed for %llu "
5689                        "%llu\n", (unsigned long long)ins->objectid,
5690                        (unsigned long long)ins->offset);
5691                 BUG();
5692         }
5693         return ret;
5694 }
5695
5696 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
5697                                      struct btrfs_root *root,
5698                                      u64 parent, u64 root_objectid,
5699                                      u64 flags, struct btrfs_disk_key *key,
5700                                      int level, struct btrfs_key *ins)
5701 {
5702         int ret;
5703         struct btrfs_fs_info *fs_info = root->fs_info;
5704         struct btrfs_extent_item *extent_item;
5705         struct btrfs_tree_block_info *block_info;
5706         struct btrfs_extent_inline_ref *iref;
5707         struct btrfs_path *path;
5708         struct extent_buffer *leaf;
5709         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
5710
5711         path = btrfs_alloc_path();
5712         if (!path)
5713                 return -ENOMEM;
5714
5715         path->leave_spinning = 1;
5716         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5717                                       ins, size);
5718         BUG_ON(ret);
5719
5720         leaf = path->nodes[0];
5721         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5722                                      struct btrfs_extent_item);
5723         btrfs_set_extent_refs(leaf, extent_item, 1);
5724         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5725         btrfs_set_extent_flags(leaf, extent_item,
5726                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
5727         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
5728
5729         btrfs_set_tree_block_key(leaf, block_info, key);
5730         btrfs_set_tree_block_level(leaf, block_info, level);
5731
5732         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
5733         if (parent > 0) {
5734                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
5735                 btrfs_set_extent_inline_ref_type(leaf, iref,
5736                                                  BTRFS_SHARED_BLOCK_REF_KEY);
5737                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5738         } else {
5739                 btrfs_set_extent_inline_ref_type(leaf, iref,
5740                                                  BTRFS_TREE_BLOCK_REF_KEY);
5741                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
5742         }
5743
5744         btrfs_mark_buffer_dirty(leaf);
5745         btrfs_free_path(path);
5746
5747         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5748         if (ret) {
5749                 printk(KERN_ERR "btrfs update block group failed for %llu "
5750                        "%llu\n", (unsigned long long)ins->objectid,
5751                        (unsigned long long)ins->offset);
5752                 BUG();
5753         }
5754         return ret;
5755 }
5756
5757 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5758                                      struct btrfs_root *root,
5759                                      u64 root_objectid, u64 owner,
5760                                      u64 offset, struct btrfs_key *ins)
5761 {
5762         int ret;
5763
5764         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
5765
5766         ret = btrfs_add_delayed_data_ref(trans, ins->objectid, ins->offset,
5767                                          0, root_objectid, owner, offset,
5768                                          BTRFS_ADD_DELAYED_EXTENT, NULL);
5769         return ret;
5770 }
5771
5772 /*
5773  * this is used by the tree logging recovery code.  It records that
5774  * an extent has been allocated and makes sure to clear the free
5775  * space cache bits as well
5776  */
5777 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
5778                                    struct btrfs_root *root,
5779                                    u64 root_objectid, u64 owner, u64 offset,
5780                                    struct btrfs_key *ins)
5781 {
5782         int ret;
5783         struct btrfs_block_group_cache *block_group;
5784         struct btrfs_caching_control *caching_ctl;
5785         u64 start = ins->objectid;
5786         u64 num_bytes = ins->offset;
5787
5788         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
5789         cache_block_group(block_group, trans, NULL, 0);
5790         caching_ctl = get_caching_control(block_group);
5791
5792         if (!caching_ctl) {
5793                 BUG_ON(!block_group_cache_done(block_group));
5794                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5795                 BUG_ON(ret);
5796         } else {
5797                 mutex_lock(&caching_ctl->mutex);
5798
5799                 if (start >= caching_ctl->progress) {
5800                         ret = add_excluded_extent(root, start, num_bytes);
5801                         BUG_ON(ret);
5802                 } else if (start + num_bytes <= caching_ctl->progress) {
5803                         ret = btrfs_remove_free_space(block_group,
5804                                                       start, num_bytes);
5805                         BUG_ON(ret);
5806                 } else {
5807                         num_bytes = caching_ctl->progress - start;
5808                         ret = btrfs_remove_free_space(block_group,
5809                                                       start, num_bytes);
5810                         BUG_ON(ret);
5811
5812                         start = caching_ctl->progress;
5813                         num_bytes = ins->objectid + ins->offset -
5814                                     caching_ctl->progress;
5815                         ret = add_excluded_extent(root, start, num_bytes);
5816                         BUG_ON(ret);
5817                 }
5818
5819                 mutex_unlock(&caching_ctl->mutex);
5820                 put_caching_control(caching_ctl);
5821         }
5822
5823         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
5824                                           RESERVE_ALLOC_NO_ACCOUNT);
5825         BUG_ON(ret);
5826         btrfs_put_block_group(block_group);
5827         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
5828                                          0, owner, offset, ins, 1);
5829         return ret;
5830 }
5831
5832 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
5833                                             struct btrfs_root *root,
5834                                             u64 bytenr, u32 blocksize,
5835                                             int level)
5836 {
5837         struct extent_buffer *buf;
5838
5839         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
5840         if (!buf)
5841                 return ERR_PTR(-ENOMEM);
5842         btrfs_set_header_generation(buf, trans->transid);
5843         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
5844         btrfs_tree_lock(buf);
5845         clean_tree_block(trans, root, buf);
5846
5847         btrfs_set_lock_blocking(buf);
5848         btrfs_set_buffer_uptodate(buf);
5849
5850         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
5851                 /*
5852                  * we allow two log transactions at a time, use different
5853                  * EXENT bit to differentiate dirty pages.
5854                  */
5855                 if (root->log_transid % 2 == 0)
5856                         set_extent_dirty(&root->dirty_log_pages, buf->start,
5857                                         buf->start + buf->len - 1, GFP_NOFS);
5858                 else
5859                         set_extent_new(&root->dirty_log_pages, buf->start,
5860                                         buf->start + buf->len - 1, GFP_NOFS);
5861         } else {
5862                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
5863                          buf->start + buf->len - 1, GFP_NOFS);
5864         }
5865         trans->blocks_used++;
5866         /* this returns a buffer locked for blocking */
5867         return buf;
5868 }
5869
5870 static struct btrfs_block_rsv *
5871 use_block_rsv(struct btrfs_trans_handle *trans,
5872               struct btrfs_root *root, u32 blocksize)
5873 {
5874         struct btrfs_block_rsv *block_rsv;
5875         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5876         int ret;
5877
5878         block_rsv = get_block_rsv(trans, root);
5879
5880         if (block_rsv->size == 0) {
5881                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
5882                 /*
5883                  * If we couldn't reserve metadata bytes try and use some from
5884                  * the global reserve.
5885                  */
5886                 if (ret && block_rsv != global_rsv) {
5887                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5888                         if (!ret)
5889                                 return global_rsv;
5890                         return ERR_PTR(ret);
5891                 } else if (ret) {
5892                         return ERR_PTR(ret);
5893                 }
5894                 return block_rsv;
5895         }
5896
5897         ret = block_rsv_use_bytes(block_rsv, blocksize);
5898         if (!ret)
5899                 return block_rsv;
5900         if (ret) {
5901                 static DEFINE_RATELIMIT_STATE(_rs,
5902                                 DEFAULT_RATELIMIT_INTERVAL,
5903                                 /*DEFAULT_RATELIMIT_BURST*/ 2);
5904                 if (__ratelimit(&_rs)) {
5905                         printk(KERN_DEBUG "btrfs: block rsv returned %d\n", ret);
5906                         WARN_ON(1);
5907                 }
5908                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
5909                 if (!ret) {
5910                         return block_rsv;
5911                 } else if (ret && block_rsv != global_rsv) {
5912                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5913                         if (!ret)
5914                                 return global_rsv;
5915                 }
5916         }
5917
5918         return ERR_PTR(-ENOSPC);
5919 }
5920
5921 static void unuse_block_rsv(struct btrfs_block_rsv *block_rsv, u32 blocksize)
5922 {
5923         block_rsv_add_bytes(block_rsv, blocksize, 0);
5924         block_rsv_release_bytes(block_rsv, NULL, 0);
5925 }
5926
5927 /*
5928  * finds a free extent and does all the dirty work required for allocation
5929  * returns the key for the extent through ins, and a tree buffer for
5930  * the first block of the extent through buf.
5931  *
5932  * returns the tree buffer or NULL.
5933  */
5934 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
5935                                         struct btrfs_root *root, u32 blocksize,
5936                                         u64 parent, u64 root_objectid,
5937                                         struct btrfs_disk_key *key, int level,
5938                                         u64 hint, u64 empty_size)
5939 {
5940         struct btrfs_key ins;
5941         struct btrfs_block_rsv *block_rsv;
5942         struct extent_buffer *buf;
5943         u64 flags = 0;
5944         int ret;
5945
5946
5947         block_rsv = use_block_rsv(trans, root, blocksize);
5948         if (IS_ERR(block_rsv))
5949                 return ERR_CAST(block_rsv);
5950
5951         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
5952                                    empty_size, hint, (u64)-1, &ins, 0);
5953         if (ret) {
5954                 unuse_block_rsv(block_rsv, blocksize);
5955                 return ERR_PTR(ret);
5956         }
5957
5958         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
5959                                     blocksize, level);
5960         BUG_ON(IS_ERR(buf));
5961
5962         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
5963                 if (parent == 0)
5964                         parent = ins.objectid;
5965                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
5966         } else
5967                 BUG_ON(parent > 0);
5968
5969         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
5970                 struct btrfs_delayed_extent_op *extent_op;
5971                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
5972                 BUG_ON(!extent_op);
5973                 if (key)
5974                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
5975                 else
5976                         memset(&extent_op->key, 0, sizeof(extent_op->key));
5977                 extent_op->flags_to_set = flags;
5978                 extent_op->update_key = 1;
5979                 extent_op->update_flags = 1;
5980                 extent_op->is_data = 0;
5981
5982                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
5983                                         ins.offset, parent, root_objectid,
5984                                         level, BTRFS_ADD_DELAYED_EXTENT,
5985                                         extent_op);
5986                 BUG_ON(ret);
5987         }
5988         return buf;
5989 }
5990
5991 struct walk_control {
5992         u64 refs[BTRFS_MAX_LEVEL];
5993         u64 flags[BTRFS_MAX_LEVEL];
5994         struct btrfs_key update_progress;
5995         int stage;
5996         int level;
5997         int shared_level;
5998         int update_ref;
5999         int keep_locks;
6000         int reada_slot;
6001         int reada_count;
6002 };
6003
6004 #define DROP_REFERENCE  1
6005 #define UPDATE_BACKREF  2
6006
6007 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6008                                      struct btrfs_root *root,
6009                                      struct walk_control *wc,
6010                                      struct btrfs_path *path)
6011 {
6012         u64 bytenr;
6013         u64 generation;
6014         u64 refs;
6015         u64 flags;
6016         u32 nritems;
6017         u32 blocksize;
6018         struct btrfs_key key;
6019         struct extent_buffer *eb;
6020         int ret;
6021         int slot;
6022         int nread = 0;
6023
6024         if (path->slots[wc->level] < wc->reada_slot) {
6025                 wc->reada_count = wc->reada_count * 2 / 3;
6026                 wc->reada_count = max(wc->reada_count, 2);
6027         } else {
6028                 wc->reada_count = wc->reada_count * 3 / 2;
6029                 wc->reada_count = min_t(int, wc->reada_count,
6030                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6031         }
6032
6033         eb = path->nodes[wc->level];
6034         nritems = btrfs_header_nritems(eb);
6035         blocksize = btrfs_level_size(root, wc->level - 1);
6036
6037         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6038                 if (nread >= wc->reada_count)
6039                         break;
6040
6041                 cond_resched();
6042                 bytenr = btrfs_node_blockptr(eb, slot);
6043                 generation = btrfs_node_ptr_generation(eb, slot);
6044
6045                 if (slot == path->slots[wc->level])
6046                         goto reada;
6047
6048                 if (wc->stage == UPDATE_BACKREF &&
6049                     generation <= root->root_key.offset)
6050                         continue;
6051
6052                 /* We don't lock the tree block, it's OK to be racy here */
6053                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6054                                                &refs, &flags);
6055                 BUG_ON(ret);
6056                 BUG_ON(refs == 0);
6057
6058                 if (wc->stage == DROP_REFERENCE) {
6059                         if (refs == 1)
6060                                 goto reada;
6061
6062                         if (wc->level == 1 &&
6063                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6064                                 continue;
6065                         if (!wc->update_ref ||
6066                             generation <= root->root_key.offset)
6067                                 continue;
6068                         btrfs_node_key_to_cpu(eb, &key, slot);
6069                         ret = btrfs_comp_cpu_keys(&key,
6070                                                   &wc->update_progress);
6071                         if (ret < 0)
6072                                 continue;
6073                 } else {
6074                         if (wc->level == 1 &&
6075                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6076                                 continue;
6077                 }
6078 reada:
6079                 ret = readahead_tree_block(root, bytenr, blocksize,
6080                                            generation);
6081                 if (ret)
6082                         break;
6083                 nread++;
6084         }
6085         wc->reada_slot = slot;
6086 }
6087
6088 /*
6089  * hepler to process tree block while walking down the tree.
6090  *
6091  * when wc->stage == UPDATE_BACKREF, this function updates
6092  * back refs for pointers in the block.
6093  *
6094  * NOTE: return value 1 means we should stop walking down.
6095  */
6096 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6097                                    struct btrfs_root *root,
6098                                    struct btrfs_path *path,
6099                                    struct walk_control *wc, int lookup_info)
6100 {
6101         int level = wc->level;
6102         struct extent_buffer *eb = path->nodes[level];
6103         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6104         int ret;
6105
6106         if (wc->stage == UPDATE_BACKREF &&
6107             btrfs_header_owner(eb) != root->root_key.objectid)
6108                 return 1;
6109
6110         /*
6111          * when reference count of tree block is 1, it won't increase
6112          * again. once full backref flag is set, we never clear it.
6113          */
6114         if (lookup_info &&
6115             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6116              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6117                 BUG_ON(!path->locks[level]);
6118                 ret = btrfs_lookup_extent_info(trans, root,
6119                                                eb->start, eb->len,
6120                                                &wc->refs[level],
6121                                                &wc->flags[level]);
6122                 BUG_ON(ret);
6123                 BUG_ON(wc->refs[level] == 0);
6124         }
6125
6126         if (wc->stage == DROP_REFERENCE) {
6127                 if (wc->refs[level] > 1)
6128                         return 1;
6129
6130                 if (path->locks[level] && !wc->keep_locks) {
6131                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6132                         path->locks[level] = 0;
6133                 }
6134                 return 0;
6135         }
6136
6137         /* wc->stage == UPDATE_BACKREF */
6138         if (!(wc->flags[level] & flag)) {
6139                 BUG_ON(!path->locks[level]);
6140                 ret = btrfs_inc_ref(trans, root, eb, 1);
6141                 BUG_ON(ret);
6142                 ret = btrfs_dec_ref(trans, root, eb, 0);
6143                 BUG_ON(ret);
6144                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6145                                                   eb->len, flag, 0);
6146                 BUG_ON(ret);
6147                 wc->flags[level] |= flag;
6148         }
6149
6150         /*
6151          * the block is shared by multiple trees, so it's not good to
6152          * keep the tree lock
6153          */
6154         if (path->locks[level] && level > 0) {
6155                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6156                 path->locks[level] = 0;
6157         }
6158         return 0;
6159 }
6160
6161 /*
6162  * hepler to process tree block pointer.
6163  *
6164  * when wc->stage == DROP_REFERENCE, this function checks
6165  * reference count of the block pointed to. if the block
6166  * is shared and we need update back refs for the subtree
6167  * rooted at the block, this function changes wc->stage to
6168  * UPDATE_BACKREF. if the block is shared and there is no
6169  * need to update back, this function drops the reference
6170  * to the block.
6171  *
6172  * NOTE: return value 1 means we should stop walking down.
6173  */
6174 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6175                                  struct btrfs_root *root,
6176                                  struct btrfs_path *path,
6177                                  struct walk_control *wc, int *lookup_info)
6178 {
6179         u64 bytenr;
6180         u64 generation;
6181         u64 parent;
6182         u32 blocksize;
6183         struct btrfs_key key;
6184         struct extent_buffer *next;
6185         int level = wc->level;
6186         int reada = 0;
6187         int ret = 0;
6188
6189         generation = btrfs_node_ptr_generation(path->nodes[level],
6190                                                path->slots[level]);
6191         /*
6192          * if the lower level block was created before the snapshot
6193          * was created, we know there is no need to update back refs
6194          * for the subtree
6195          */
6196         if (wc->stage == UPDATE_BACKREF &&
6197             generation <= root->root_key.offset) {
6198                 *lookup_info = 1;
6199                 return 1;
6200         }
6201
6202         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6203         blocksize = btrfs_level_size(root, level - 1);
6204
6205         next = btrfs_find_tree_block(root, bytenr, blocksize);
6206         if (!next) {
6207                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6208                 if (!next)
6209                         return -ENOMEM;
6210                 reada = 1;
6211         }
6212         btrfs_tree_lock(next);
6213         btrfs_set_lock_blocking(next);
6214
6215         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6216                                        &wc->refs[level - 1],
6217                                        &wc->flags[level - 1]);
6218         BUG_ON(ret);
6219         BUG_ON(wc->refs[level - 1] == 0);
6220         *lookup_info = 0;
6221
6222         if (wc->stage == DROP_REFERENCE) {
6223                 if (wc->refs[level - 1] > 1) {
6224                         if (level == 1 &&
6225                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6226                                 goto skip;
6227
6228                         if (!wc->update_ref ||
6229                             generation <= root->root_key.offset)
6230                                 goto skip;
6231
6232                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6233                                               path->slots[level]);
6234                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6235                         if (ret < 0)
6236                                 goto skip;
6237
6238                         wc->stage = UPDATE_BACKREF;
6239                         wc->shared_level = level - 1;
6240                 }
6241         } else {
6242                 if (level == 1 &&
6243                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6244                         goto skip;
6245         }
6246
6247         if (!btrfs_buffer_uptodate(next, generation)) {
6248                 btrfs_tree_unlock(next);
6249                 free_extent_buffer(next);
6250                 next = NULL;
6251                 *lookup_info = 1;
6252         }
6253
6254         if (!next) {
6255                 if (reada && level == 1)
6256                         reada_walk_down(trans, root, wc, path);
6257                 next = read_tree_block(root, bytenr, blocksize, generation);
6258                 if (!next)
6259                         return -EIO;
6260                 btrfs_tree_lock(next);
6261                 btrfs_set_lock_blocking(next);
6262         }
6263
6264         level--;
6265         BUG_ON(level != btrfs_header_level(next));
6266         path->nodes[level] = next;
6267         path->slots[level] = 0;
6268         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6269         wc->level = level;
6270         if (wc->level == 1)
6271                 wc->reada_slot = 0;
6272         return 0;
6273 skip:
6274         wc->refs[level - 1] = 0;
6275         wc->flags[level - 1] = 0;
6276         if (wc->stage == DROP_REFERENCE) {
6277                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6278                         parent = path->nodes[level]->start;
6279                 } else {
6280                         BUG_ON(root->root_key.objectid !=
6281                                btrfs_header_owner(path->nodes[level]));
6282                         parent = 0;
6283                 }
6284
6285                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6286                                         root->root_key.objectid, level - 1, 0);
6287                 BUG_ON(ret);
6288         }
6289         btrfs_tree_unlock(next);
6290         free_extent_buffer(next);
6291         *lookup_info = 1;
6292         return 1;
6293 }
6294
6295 /*
6296  * hepler to process tree block while walking up the tree.
6297  *
6298  * when wc->stage == DROP_REFERENCE, this function drops
6299  * reference count on the block.
6300  *
6301  * when wc->stage == UPDATE_BACKREF, this function changes
6302  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6303  * to UPDATE_BACKREF previously while processing the block.
6304  *
6305  * NOTE: return value 1 means we should stop walking up.
6306  */
6307 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6308                                  struct btrfs_root *root,
6309                                  struct btrfs_path *path,
6310                                  struct walk_control *wc)
6311 {
6312         int ret;
6313         int level = wc->level;
6314         struct extent_buffer *eb = path->nodes[level];
6315         u64 parent = 0;
6316
6317         if (wc->stage == UPDATE_BACKREF) {
6318                 BUG_ON(wc->shared_level < level);
6319                 if (level < wc->shared_level)
6320                         goto out;
6321
6322                 ret = find_next_key(path, level + 1, &wc->update_progress);
6323                 if (ret > 0)
6324                         wc->update_ref = 0;
6325
6326                 wc->stage = DROP_REFERENCE;
6327                 wc->shared_level = -1;
6328                 path->slots[level] = 0;
6329
6330                 /*
6331                  * check reference count again if the block isn't locked.
6332                  * we should start walking down the tree again if reference
6333                  * count is one.
6334                  */
6335                 if (!path->locks[level]) {
6336                         BUG_ON(level == 0);
6337                         btrfs_tree_lock(eb);
6338                         btrfs_set_lock_blocking(eb);
6339                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6340
6341                         ret = btrfs_lookup_extent_info(trans, root,
6342                                                        eb->start, eb->len,
6343                                                        &wc->refs[level],
6344                                                        &wc->flags[level]);
6345                         BUG_ON(ret);
6346                         BUG_ON(wc->refs[level] == 0);
6347                         if (wc->refs[level] == 1) {
6348                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6349                                 return 1;
6350                         }
6351                 }
6352         }
6353
6354         /* wc->stage == DROP_REFERENCE */
6355         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6356
6357         if (wc->refs[level] == 1) {
6358                 if (level == 0) {
6359                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6360                                 ret = btrfs_dec_ref(trans, root, eb, 1);
6361                         else
6362                                 ret = btrfs_dec_ref(trans, root, eb, 0);
6363                         BUG_ON(ret);
6364                 }
6365                 /* make block locked assertion in clean_tree_block happy */
6366                 if (!path->locks[level] &&
6367                     btrfs_header_generation(eb) == trans->transid) {
6368                         btrfs_tree_lock(eb);
6369                         btrfs_set_lock_blocking(eb);
6370                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6371                 }
6372                 clean_tree_block(trans, root, eb);
6373         }
6374
6375         if (eb == root->node) {
6376                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6377                         parent = eb->start;
6378                 else
6379                         BUG_ON(root->root_key.objectid !=
6380                                btrfs_header_owner(eb));
6381         } else {
6382                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6383                         parent = path->nodes[level + 1]->start;
6384                 else
6385                         BUG_ON(root->root_key.objectid !=
6386                                btrfs_header_owner(path->nodes[level + 1]));
6387         }
6388
6389         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6390 out:
6391         wc->refs[level] = 0;
6392         wc->flags[level] = 0;
6393         return 0;
6394 }
6395
6396 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6397                                    struct btrfs_root *root,
6398                                    struct btrfs_path *path,
6399                                    struct walk_control *wc)
6400 {
6401         int level = wc->level;
6402         int lookup_info = 1;
6403         int ret;
6404
6405         while (level >= 0) {
6406                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6407                 if (ret > 0)
6408                         break;
6409
6410                 if (level == 0)
6411                         break;
6412
6413                 if (path->slots[level] >=
6414                     btrfs_header_nritems(path->nodes[level]))
6415                         break;
6416
6417                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6418                 if (ret > 0) {
6419                         path->slots[level]++;
6420                         continue;
6421                 } else if (ret < 0)
6422                         return ret;
6423                 level = wc->level;
6424         }
6425         return 0;
6426 }
6427
6428 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6429                                  struct btrfs_root *root,
6430                                  struct btrfs_path *path,
6431                                  struct walk_control *wc, int max_level)
6432 {
6433         int level = wc->level;
6434         int ret;
6435
6436         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6437         while (level < max_level && path->nodes[level]) {
6438                 wc->level = level;
6439                 if (path->slots[level] + 1 <
6440                     btrfs_header_nritems(path->nodes[level])) {
6441                         path->slots[level]++;
6442                         return 0;
6443                 } else {
6444                         ret = walk_up_proc(trans, root, path, wc);
6445                         if (ret > 0)
6446                                 return 0;
6447
6448                         if (path->locks[level]) {
6449                                 btrfs_tree_unlock_rw(path->nodes[level],
6450                                                      path->locks[level]);
6451                                 path->locks[level] = 0;
6452                         }
6453                         free_extent_buffer(path->nodes[level]);
6454                         path->nodes[level] = NULL;
6455                         level++;
6456                 }
6457         }
6458         return 1;
6459 }
6460
6461 /*
6462  * drop a subvolume tree.
6463  *
6464  * this function traverses the tree freeing any blocks that only
6465  * referenced by the tree.
6466  *
6467  * when a shared tree block is found. this function decreases its
6468  * reference count by one. if update_ref is true, this function
6469  * also make sure backrefs for the shared block and all lower level
6470  * blocks are properly updated.
6471  */
6472 void btrfs_drop_snapshot(struct btrfs_root *root,
6473                          struct btrfs_block_rsv *block_rsv, int update_ref)
6474 {
6475         struct btrfs_path *path;
6476         struct btrfs_trans_handle *trans;
6477         struct btrfs_root *tree_root = root->fs_info->tree_root;
6478         struct btrfs_root_item *root_item = &root->root_item;
6479         struct walk_control *wc;
6480         struct btrfs_key key;
6481         int err = 0;
6482         int ret;
6483         int level;
6484
6485         path = btrfs_alloc_path();
6486         if (!path) {
6487                 err = -ENOMEM;
6488                 goto out;
6489         }
6490
6491         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6492         if (!wc) {
6493                 btrfs_free_path(path);
6494                 err = -ENOMEM;
6495                 goto out;
6496         }
6497
6498         trans = btrfs_start_transaction(tree_root, 0);
6499         BUG_ON(IS_ERR(trans));
6500
6501         if (block_rsv)
6502                 trans->block_rsv = block_rsv;
6503
6504         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6505                 level = btrfs_header_level(root->node);
6506                 path->nodes[level] = btrfs_lock_root_node(root);
6507                 btrfs_set_lock_blocking(path->nodes[level]);
6508                 path->slots[level] = 0;
6509                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6510                 memset(&wc->update_progress, 0,
6511                        sizeof(wc->update_progress));
6512         } else {
6513                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6514                 memcpy(&wc->update_progress, &key,
6515                        sizeof(wc->update_progress));
6516
6517                 level = root_item->drop_level;
6518                 BUG_ON(level == 0);
6519                 path->lowest_level = level;
6520                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6521                 path->lowest_level = 0;
6522                 if (ret < 0) {
6523                         err = ret;
6524                         goto out_free;
6525                 }
6526                 WARN_ON(ret > 0);
6527
6528                 /*
6529                  * unlock our path, this is safe because only this
6530                  * function is allowed to delete this snapshot
6531                  */
6532                 btrfs_unlock_up_safe(path, 0);
6533
6534                 level = btrfs_header_level(root->node);
6535                 while (1) {
6536                         btrfs_tree_lock(path->nodes[level]);
6537                         btrfs_set_lock_blocking(path->nodes[level]);
6538
6539                         ret = btrfs_lookup_extent_info(trans, root,
6540                                                 path->nodes[level]->start,
6541                                                 path->nodes[level]->len,
6542                                                 &wc->refs[level],
6543                                                 &wc->flags[level]);
6544                         BUG_ON(ret);
6545                         BUG_ON(wc->refs[level] == 0);
6546
6547                         if (level == root_item->drop_level)
6548                                 break;
6549
6550                         btrfs_tree_unlock(path->nodes[level]);
6551                         WARN_ON(wc->refs[level] != 1);
6552                         level--;
6553                 }
6554         }
6555
6556         wc->level = level;
6557         wc->shared_level = -1;
6558         wc->stage = DROP_REFERENCE;
6559         wc->update_ref = update_ref;
6560         wc->keep_locks = 0;
6561         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6562
6563         while (1) {
6564                 ret = walk_down_tree(trans, root, path, wc);
6565                 if (ret < 0) {
6566                         err = ret;
6567                         break;
6568                 }
6569
6570                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
6571                 if (ret < 0) {
6572                         err = ret;
6573                         break;
6574                 }
6575
6576                 if (ret > 0) {
6577                         BUG_ON(wc->stage != DROP_REFERENCE);
6578                         break;
6579                 }
6580
6581                 if (wc->stage == DROP_REFERENCE) {
6582                         level = wc->level;
6583                         btrfs_node_key(path->nodes[level],
6584                                        &root_item->drop_progress,
6585                                        path->slots[level]);
6586                         root_item->drop_level = level;
6587                 }
6588
6589                 BUG_ON(wc->level == 0);
6590                 if (btrfs_should_end_transaction(trans, tree_root)) {
6591                         ret = btrfs_update_root(trans, tree_root,
6592                                                 &root->root_key,
6593                                                 root_item);
6594                         BUG_ON(ret);
6595
6596                         btrfs_end_transaction_throttle(trans, tree_root);
6597                         trans = btrfs_start_transaction(tree_root, 0);
6598                         BUG_ON(IS_ERR(trans));
6599                         if (block_rsv)
6600                                 trans->block_rsv = block_rsv;
6601                 }
6602         }
6603         btrfs_release_path(path);
6604         BUG_ON(err);
6605
6606         ret = btrfs_del_root(trans, tree_root, &root->root_key);
6607         BUG_ON(ret);
6608
6609         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
6610                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
6611                                            NULL, NULL);
6612                 BUG_ON(ret < 0);
6613                 if (ret > 0) {
6614                         /* if we fail to delete the orphan item this time
6615                          * around, it'll get picked up the next time.
6616                          *
6617                          * The most common failure here is just -ENOENT.
6618                          */
6619                         btrfs_del_orphan_item(trans, tree_root,
6620                                               root->root_key.objectid);
6621                 }
6622         }
6623
6624         if (root->in_radix) {
6625                 btrfs_free_fs_root(tree_root->fs_info, root);
6626         } else {
6627                 free_extent_buffer(root->node);
6628                 free_extent_buffer(root->commit_root);
6629                 kfree(root);
6630         }
6631 out_free:
6632         btrfs_end_transaction_throttle(trans, tree_root);
6633         kfree(wc);
6634         btrfs_free_path(path);
6635 out:
6636         if (err)
6637                 btrfs_std_error(root->fs_info, err);
6638         return;
6639 }
6640
6641 /*
6642  * drop subtree rooted at tree block 'node'.
6643  *
6644  * NOTE: this function will unlock and release tree block 'node'
6645  */
6646 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
6647                         struct btrfs_root *root,
6648                         struct extent_buffer *node,
6649                         struct extent_buffer *parent)
6650 {
6651         struct btrfs_path *path;
6652         struct walk_control *wc;
6653         int level;
6654         int parent_level;
6655         int ret = 0;
6656         int wret;
6657
6658         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
6659
6660         path = btrfs_alloc_path();
6661         if (!path)
6662                 return -ENOMEM;
6663
6664         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6665         if (!wc) {
6666                 btrfs_free_path(path);
6667                 return -ENOMEM;
6668         }
6669
6670         btrfs_assert_tree_locked(parent);
6671         parent_level = btrfs_header_level(parent);
6672         extent_buffer_get(parent);
6673         path->nodes[parent_level] = parent;
6674         path->slots[parent_level] = btrfs_header_nritems(parent);
6675
6676         btrfs_assert_tree_locked(node);
6677         level = btrfs_header_level(node);
6678         path->nodes[level] = node;
6679         path->slots[level] = 0;
6680         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6681
6682         wc->refs[parent_level] = 1;
6683         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6684         wc->level = level;
6685         wc->shared_level = -1;
6686         wc->stage = DROP_REFERENCE;
6687         wc->update_ref = 0;
6688         wc->keep_locks = 1;
6689         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6690
6691         while (1) {
6692                 wret = walk_down_tree(trans, root, path, wc);
6693                 if (wret < 0) {
6694                         ret = wret;
6695                         break;
6696                 }
6697
6698                 wret = walk_up_tree(trans, root, path, wc, parent_level);
6699                 if (wret < 0)
6700                         ret = wret;
6701                 if (wret != 0)
6702                         break;
6703         }
6704
6705         kfree(wc);
6706         btrfs_free_path(path);
6707         return ret;
6708 }
6709
6710 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
6711 {
6712         u64 num_devices;
6713         u64 stripped = BTRFS_BLOCK_GROUP_RAID0 |
6714                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
6715
6716         /*
6717          * we add in the count of missing devices because we want
6718          * to make sure that any RAID levels on a degraded FS
6719          * continue to be honored.
6720          */
6721         num_devices = root->fs_info->fs_devices->rw_devices +
6722                 root->fs_info->fs_devices->missing_devices;
6723
6724         if (num_devices == 1) {
6725                 stripped |= BTRFS_BLOCK_GROUP_DUP;
6726                 stripped = flags & ~stripped;
6727
6728                 /* turn raid0 into single device chunks */
6729                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
6730                         return stripped;
6731
6732                 /* turn mirroring into duplication */
6733                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
6734                              BTRFS_BLOCK_GROUP_RAID10))
6735                         return stripped | BTRFS_BLOCK_GROUP_DUP;
6736                 return flags;
6737         } else {
6738                 /* they already had raid on here, just return */
6739                 if (flags & stripped)
6740                         return flags;
6741
6742                 stripped |= BTRFS_BLOCK_GROUP_DUP;
6743                 stripped = flags & ~stripped;
6744
6745                 /* switch duplicated blocks with raid1 */
6746                 if (flags & BTRFS_BLOCK_GROUP_DUP)
6747                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
6748
6749                 /* turn single device chunks into raid0 */
6750                 return stripped | BTRFS_BLOCK_GROUP_RAID0;
6751         }
6752         return flags;
6753 }
6754
6755 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
6756 {
6757         struct btrfs_space_info *sinfo = cache->space_info;
6758         u64 num_bytes;
6759         u64 min_allocable_bytes;
6760         int ret = -ENOSPC;
6761
6762
6763         /*
6764          * We need some metadata space and system metadata space for
6765          * allocating chunks in some corner cases until we force to set
6766          * it to be readonly.
6767          */
6768         if ((sinfo->flags &
6769              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
6770             !force)
6771                 min_allocable_bytes = 1 * 1024 * 1024;
6772         else
6773                 min_allocable_bytes = 0;
6774
6775         spin_lock(&sinfo->lock);
6776         spin_lock(&cache->lock);
6777
6778         if (cache->ro) {
6779                 ret = 0;
6780                 goto out;
6781         }
6782
6783         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
6784                     cache->bytes_super - btrfs_block_group_used(&cache->item);
6785
6786         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
6787             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
6788             min_allocable_bytes <= sinfo->total_bytes) {
6789                 sinfo->bytes_readonly += num_bytes;
6790                 cache->ro = 1;
6791                 ret = 0;
6792         }
6793 out:
6794         spin_unlock(&cache->lock);
6795         spin_unlock(&sinfo->lock);
6796         return ret;
6797 }
6798
6799 int btrfs_set_block_group_ro(struct btrfs_root *root,
6800                              struct btrfs_block_group_cache *cache)
6801
6802 {
6803         struct btrfs_trans_handle *trans;
6804         u64 alloc_flags;
6805         int ret;
6806
6807         BUG_ON(cache->ro);
6808
6809         trans = btrfs_join_transaction(root);
6810         BUG_ON(IS_ERR(trans));
6811
6812         alloc_flags = update_block_group_flags(root, cache->flags);
6813         if (alloc_flags != cache->flags)
6814                 do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6815                                CHUNK_ALLOC_FORCE);
6816
6817         ret = set_block_group_ro(cache, 0);
6818         if (!ret)
6819                 goto out;
6820         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
6821         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6822                              CHUNK_ALLOC_FORCE);
6823         if (ret < 0)
6824                 goto out;
6825         ret = set_block_group_ro(cache, 0);
6826 out:
6827         btrfs_end_transaction(trans, root);
6828         return ret;
6829 }
6830
6831 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
6832                             struct btrfs_root *root, u64 type)
6833 {
6834         u64 alloc_flags = get_alloc_profile(root, type);
6835         return do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6836                               CHUNK_ALLOC_FORCE);
6837 }
6838
6839 /*
6840  * helper to account the unused space of all the readonly block group in the
6841  * list. takes mirrors into account.
6842  */
6843 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
6844 {
6845         struct btrfs_block_group_cache *block_group;
6846         u64 free_bytes = 0;
6847         int factor;
6848
6849         list_for_each_entry(block_group, groups_list, list) {
6850                 spin_lock(&block_group->lock);
6851
6852                 if (!block_group->ro) {
6853                         spin_unlock(&block_group->lock);
6854                         continue;
6855                 }
6856
6857                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
6858                                           BTRFS_BLOCK_GROUP_RAID10 |
6859                                           BTRFS_BLOCK_GROUP_DUP))
6860                         factor = 2;
6861                 else
6862                         factor = 1;
6863
6864                 free_bytes += (block_group->key.offset -
6865                                btrfs_block_group_used(&block_group->item)) *
6866                                factor;
6867
6868                 spin_unlock(&block_group->lock);
6869         }
6870
6871         return free_bytes;
6872 }
6873
6874 /*
6875  * helper to account the unused space of all the readonly block group in the
6876  * space_info. takes mirrors into account.
6877  */
6878 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
6879 {
6880         int i;
6881         u64 free_bytes = 0;
6882
6883         spin_lock(&sinfo->lock);
6884
6885         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
6886                 if (!list_empty(&sinfo->block_groups[i]))
6887                         free_bytes += __btrfs_get_ro_block_group_free_space(
6888                                                 &sinfo->block_groups[i]);
6889
6890         spin_unlock(&sinfo->lock);
6891
6892         return free_bytes;
6893 }
6894
6895 int btrfs_set_block_group_rw(struct btrfs_root *root,
6896                               struct btrfs_block_group_cache *cache)
6897 {
6898         struct btrfs_space_info *sinfo = cache->space_info;
6899         u64 num_bytes;
6900
6901         BUG_ON(!cache->ro);
6902
6903         spin_lock(&sinfo->lock);
6904         spin_lock(&cache->lock);
6905         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
6906                     cache->bytes_super - btrfs_block_group_used(&cache->item);
6907         sinfo->bytes_readonly -= num_bytes;
6908         cache->ro = 0;
6909         spin_unlock(&cache->lock);
6910         spin_unlock(&sinfo->lock);
6911         return 0;
6912 }
6913
6914 /*
6915  * checks to see if its even possible to relocate this block group.
6916  *
6917  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
6918  * ok to go ahead and try.
6919  */
6920 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
6921 {
6922         struct btrfs_block_group_cache *block_group;
6923         struct btrfs_space_info *space_info;
6924         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
6925         struct btrfs_device *device;
6926         u64 min_free;
6927         u64 dev_min = 1;
6928         u64 dev_nr = 0;
6929         int index;
6930         int full = 0;
6931         int ret = 0;
6932
6933         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
6934
6935         /* odd, couldn't find the block group, leave it alone */
6936         if (!block_group)
6937                 return -1;
6938
6939         min_free = btrfs_block_group_used(&block_group->item);
6940
6941         /* no bytes used, we're good */
6942         if (!min_free)
6943                 goto out;
6944
6945         space_info = block_group->space_info;
6946         spin_lock(&space_info->lock);
6947
6948         full = space_info->full;
6949
6950         /*
6951          * if this is the last block group we have in this space, we can't
6952          * relocate it unless we're able to allocate a new chunk below.
6953          *
6954          * Otherwise, we need to make sure we have room in the space to handle
6955          * all of the extents from this block group.  If we can, we're good
6956          */
6957         if ((space_info->total_bytes != block_group->key.offset) &&
6958             (space_info->bytes_used + space_info->bytes_reserved +
6959              space_info->bytes_pinned + space_info->bytes_readonly +
6960              min_free < space_info->total_bytes)) {
6961                 spin_unlock(&space_info->lock);
6962                 goto out;
6963         }
6964         spin_unlock(&space_info->lock);
6965
6966         /*
6967          * ok we don't have enough space, but maybe we have free space on our
6968          * devices to allocate new chunks for relocation, so loop through our
6969          * alloc devices and guess if we have enough space.  However, if we
6970          * were marked as full, then we know there aren't enough chunks, and we
6971          * can just return.
6972          */
6973         ret = -1;
6974         if (full)
6975                 goto out;
6976
6977         /*
6978          * index:
6979          *      0: raid10
6980          *      1: raid1
6981          *      2: dup
6982          *      3: raid0
6983          *      4: single
6984          */
6985         index = get_block_group_index(block_group);
6986         if (index == 0) {
6987                 dev_min = 4;
6988                 /* Divide by 2 */
6989                 min_free >>= 1;
6990         } else if (index == 1) {
6991                 dev_min = 2;
6992         } else if (index == 2) {
6993                 /* Multiply by 2 */
6994                 min_free <<= 1;
6995         } else if (index == 3) {
6996                 dev_min = fs_devices->rw_devices;
6997                 do_div(min_free, dev_min);
6998         }
6999
7000         mutex_lock(&root->fs_info->chunk_mutex);
7001         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7002                 u64 dev_offset;
7003
7004                 /*
7005                  * check to make sure we can actually find a chunk with enough
7006                  * space to fit our block group in.
7007                  */
7008                 if (device->total_bytes > device->bytes_used + min_free) {
7009                         ret = find_free_dev_extent(NULL, device, min_free,
7010                                                    &dev_offset, NULL);
7011                         if (!ret)
7012                                 dev_nr++;
7013
7014                         if (dev_nr >= dev_min)
7015                                 break;
7016
7017                         ret = -1;
7018                 }
7019         }
7020         mutex_unlock(&root->fs_info->chunk_mutex);
7021 out:
7022         btrfs_put_block_group(block_group);
7023         return ret;
7024 }
7025
7026 static int find_first_block_group(struct btrfs_root *root,
7027                 struct btrfs_path *path, struct btrfs_key *key)
7028 {
7029         int ret = 0;
7030         struct btrfs_key found_key;
7031         struct extent_buffer *leaf;
7032         int slot;
7033
7034         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7035         if (ret < 0)
7036                 goto out;
7037
7038         while (1) {
7039                 slot = path->slots[0];
7040                 leaf = path->nodes[0];
7041                 if (slot >= btrfs_header_nritems(leaf)) {
7042                         ret = btrfs_next_leaf(root, path);
7043                         if (ret == 0)
7044                                 continue;
7045                         if (ret < 0)
7046                                 goto out;
7047                         break;
7048                 }
7049                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7050
7051                 if (found_key.objectid >= key->objectid &&
7052                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7053                         ret = 0;
7054                         goto out;
7055                 }
7056                 path->slots[0]++;
7057         }
7058 out:
7059         return ret;
7060 }
7061
7062 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7063 {
7064         struct btrfs_block_group_cache *block_group;
7065         u64 last = 0;
7066
7067         while (1) {
7068                 struct inode *inode;
7069
7070                 block_group = btrfs_lookup_first_block_group(info, last);
7071                 while (block_group) {
7072                         spin_lock(&block_group->lock);
7073                         if (block_group->iref)
7074                                 break;
7075                         spin_unlock(&block_group->lock);
7076                         block_group = next_block_group(info->tree_root,
7077                                                        block_group);
7078                 }
7079                 if (!block_group) {
7080                         if (last == 0)
7081                                 break;
7082                         last = 0;
7083                         continue;
7084                 }
7085
7086                 inode = block_group->inode;
7087                 block_group->iref = 0;
7088                 block_group->inode = NULL;
7089                 spin_unlock(&block_group->lock);
7090                 iput(inode);
7091                 last = block_group->key.objectid + block_group->key.offset;
7092                 btrfs_put_block_group(block_group);
7093         }
7094 }
7095
7096 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7097 {
7098         struct btrfs_block_group_cache *block_group;
7099         struct btrfs_space_info *space_info;
7100         struct btrfs_caching_control *caching_ctl;
7101         struct rb_node *n;
7102
7103         down_write(&info->extent_commit_sem);
7104         while (!list_empty(&info->caching_block_groups)) {
7105                 caching_ctl = list_entry(info->caching_block_groups.next,
7106                                          struct btrfs_caching_control, list);
7107                 list_del(&caching_ctl->list);
7108                 put_caching_control(caching_ctl);
7109         }
7110         up_write(&info->extent_commit_sem);
7111
7112         spin_lock(&info->block_group_cache_lock);
7113         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7114                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7115                                        cache_node);
7116                 rb_erase(&block_group->cache_node,
7117                          &info->block_group_cache_tree);
7118                 spin_unlock(&info->block_group_cache_lock);
7119
7120                 down_write(&block_group->space_info->groups_sem);
7121                 list_del(&block_group->list);
7122                 up_write(&block_group->space_info->groups_sem);
7123
7124                 if (block_group->cached == BTRFS_CACHE_STARTED)
7125                         wait_block_group_cache_done(block_group);
7126
7127                 /*
7128                  * We haven't cached this block group, which means we could
7129                  * possibly have excluded extents on this block group.
7130                  */
7131                 if (block_group->cached == BTRFS_CACHE_NO)
7132                         free_excluded_extents(info->extent_root, block_group);
7133
7134                 btrfs_remove_free_space_cache(block_group);
7135                 btrfs_put_block_group(block_group);
7136
7137                 spin_lock(&info->block_group_cache_lock);
7138         }
7139         spin_unlock(&info->block_group_cache_lock);
7140
7141         /* now that all the block groups are freed, go through and
7142          * free all the space_info structs.  This is only called during
7143          * the final stages of unmount, and so we know nobody is
7144          * using them.  We call synchronize_rcu() once before we start,
7145          * just to be on the safe side.
7146          */
7147         synchronize_rcu();
7148
7149         release_global_block_rsv(info);
7150
7151         while(!list_empty(&info->space_info)) {
7152                 space_info = list_entry(info->space_info.next,
7153                                         struct btrfs_space_info,
7154                                         list);
7155                 if (space_info->bytes_pinned > 0 ||
7156                     space_info->bytes_reserved > 0 ||
7157                     space_info->bytes_may_use > 0) {
7158                         WARN_ON(1);
7159                         dump_space_info(space_info, 0, 0);
7160                 }
7161                 list_del(&space_info->list);
7162                 kfree(space_info);
7163         }
7164         return 0;
7165 }
7166
7167 static void __link_block_group(struct btrfs_space_info *space_info,
7168                                struct btrfs_block_group_cache *cache)
7169 {
7170         int index = get_block_group_index(cache);
7171
7172         down_write(&space_info->groups_sem);
7173         list_add_tail(&cache->list, &space_info->block_groups[index]);
7174         up_write(&space_info->groups_sem);
7175 }
7176
7177 int btrfs_read_block_groups(struct btrfs_root *root)
7178 {
7179         struct btrfs_path *path;
7180         int ret;
7181         struct btrfs_block_group_cache *cache;
7182         struct btrfs_fs_info *info = root->fs_info;
7183         struct btrfs_space_info *space_info;
7184         struct btrfs_key key;
7185         struct btrfs_key found_key;
7186         struct extent_buffer *leaf;
7187         int need_clear = 0;
7188         u64 cache_gen;
7189
7190         root = info->extent_root;
7191         key.objectid = 0;
7192         key.offset = 0;
7193         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7194         path = btrfs_alloc_path();
7195         if (!path)
7196                 return -ENOMEM;
7197         path->reada = 1;
7198
7199         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7200         if (btrfs_test_opt(root, SPACE_CACHE) &&
7201             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7202                 need_clear = 1;
7203         if (btrfs_test_opt(root, CLEAR_CACHE))
7204                 need_clear = 1;
7205
7206         while (1) {
7207                 ret = find_first_block_group(root, path, &key);
7208                 if (ret > 0)
7209                         break;
7210                 if (ret != 0)
7211                         goto error;
7212                 leaf = path->nodes[0];
7213                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7214                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7215                 if (!cache) {
7216                         ret = -ENOMEM;
7217                         goto error;
7218                 }
7219                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7220                                                 GFP_NOFS);
7221                 if (!cache->free_space_ctl) {
7222                         kfree(cache);
7223                         ret = -ENOMEM;
7224                         goto error;
7225                 }
7226
7227                 atomic_set(&cache->count, 1);
7228                 spin_lock_init(&cache->lock);
7229                 cache->fs_info = info;
7230                 INIT_LIST_HEAD(&cache->list);
7231                 INIT_LIST_HEAD(&cache->cluster_list);
7232
7233                 if (need_clear)
7234                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7235
7236                 read_extent_buffer(leaf, &cache->item,
7237                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7238                                    sizeof(cache->item));
7239                 memcpy(&cache->key, &found_key, sizeof(found_key));
7240
7241                 key.objectid = found_key.objectid + found_key.offset;
7242                 btrfs_release_path(path);
7243                 cache->flags = btrfs_block_group_flags(&cache->item);
7244                 cache->sectorsize = root->sectorsize;
7245
7246                 btrfs_init_free_space_ctl(cache);
7247
7248                 /*
7249                  * We need to exclude the super stripes now so that the space
7250                  * info has super bytes accounted for, otherwise we'll think
7251                  * we have more space than we actually do.
7252                  */
7253                 exclude_super_stripes(root, cache);
7254
7255                 /*
7256                  * check for two cases, either we are full, and therefore
7257                  * don't need to bother with the caching work since we won't
7258                  * find any space, or we are empty, and we can just add all
7259                  * the space in and be done with it.  This saves us _alot_ of
7260                  * time, particularly in the full case.
7261                  */
7262                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7263                         cache->last_byte_to_unpin = (u64)-1;
7264                         cache->cached = BTRFS_CACHE_FINISHED;
7265                         free_excluded_extents(root, cache);
7266                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7267                         cache->last_byte_to_unpin = (u64)-1;
7268                         cache->cached = BTRFS_CACHE_FINISHED;
7269                         add_new_free_space(cache, root->fs_info,
7270                                            found_key.objectid,
7271                                            found_key.objectid +
7272                                            found_key.offset);
7273                         free_excluded_extents(root, cache);
7274                 }
7275
7276                 ret = update_space_info(info, cache->flags, found_key.offset,
7277                                         btrfs_block_group_used(&cache->item),
7278                                         &space_info);
7279                 BUG_ON(ret);
7280                 cache->space_info = space_info;
7281                 spin_lock(&cache->space_info->lock);
7282                 cache->space_info->bytes_readonly += cache->bytes_super;
7283                 spin_unlock(&cache->space_info->lock);
7284
7285                 __link_block_group(space_info, cache);
7286
7287                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7288                 BUG_ON(ret);
7289
7290                 set_avail_alloc_bits(root->fs_info, cache->flags);
7291                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7292                         set_block_group_ro(cache, 1);
7293         }
7294
7295         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7296                 if (!(get_alloc_profile(root, space_info->flags) &
7297                       (BTRFS_BLOCK_GROUP_RAID10 |
7298                        BTRFS_BLOCK_GROUP_RAID1 |
7299                        BTRFS_BLOCK_GROUP_DUP)))
7300                         continue;
7301                 /*
7302                  * avoid allocating from un-mirrored block group if there are
7303                  * mirrored block groups.
7304                  */
7305                 list_for_each_entry(cache, &space_info->block_groups[3], list)
7306                         set_block_group_ro(cache, 1);
7307                 list_for_each_entry(cache, &space_info->block_groups[4], list)
7308                         set_block_group_ro(cache, 1);
7309         }
7310
7311         init_global_block_rsv(info);
7312         ret = 0;
7313 error:
7314         btrfs_free_path(path);
7315         return ret;
7316 }
7317
7318 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7319                            struct btrfs_root *root, u64 bytes_used,
7320                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7321                            u64 size)
7322 {
7323         int ret;
7324         struct btrfs_root *extent_root;
7325         struct btrfs_block_group_cache *cache;
7326
7327         extent_root = root->fs_info->extent_root;
7328
7329         root->fs_info->last_trans_log_full_commit = trans->transid;
7330
7331         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7332         if (!cache)
7333                 return -ENOMEM;
7334         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7335                                         GFP_NOFS);
7336         if (!cache->free_space_ctl) {
7337                 kfree(cache);
7338                 return -ENOMEM;
7339         }
7340
7341         cache->key.objectid = chunk_offset;
7342         cache->key.offset = size;
7343         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7344         cache->sectorsize = root->sectorsize;
7345         cache->fs_info = root->fs_info;
7346
7347         atomic_set(&cache->count, 1);
7348         spin_lock_init(&cache->lock);
7349         INIT_LIST_HEAD(&cache->list);
7350         INIT_LIST_HEAD(&cache->cluster_list);
7351
7352         btrfs_init_free_space_ctl(cache);
7353
7354         btrfs_set_block_group_used(&cache->item, bytes_used);
7355         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7356         cache->flags = type;
7357         btrfs_set_block_group_flags(&cache->item, type);
7358
7359         cache->last_byte_to_unpin = (u64)-1;
7360         cache->cached = BTRFS_CACHE_FINISHED;
7361         exclude_super_stripes(root, cache);
7362
7363         add_new_free_space(cache, root->fs_info, chunk_offset,
7364                            chunk_offset + size);
7365
7366         free_excluded_extents(root, cache);
7367
7368         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7369                                 &cache->space_info);
7370         BUG_ON(ret);
7371
7372         spin_lock(&cache->space_info->lock);
7373         cache->space_info->bytes_readonly += cache->bytes_super;
7374         spin_unlock(&cache->space_info->lock);
7375
7376         __link_block_group(cache->space_info, cache);
7377
7378         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7379         BUG_ON(ret);
7380
7381         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
7382                                 sizeof(cache->item));
7383         BUG_ON(ret);
7384
7385         set_avail_alloc_bits(extent_root->fs_info, type);
7386
7387         return 0;
7388 }
7389
7390 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7391                              struct btrfs_root *root, u64 group_start)
7392 {
7393         struct btrfs_path *path;
7394         struct btrfs_block_group_cache *block_group;
7395         struct btrfs_free_cluster *cluster;
7396         struct btrfs_root *tree_root = root->fs_info->tree_root;
7397         struct btrfs_key key;
7398         struct inode *inode;
7399         int ret;
7400         int factor;
7401
7402         root = root->fs_info->extent_root;
7403
7404         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7405         BUG_ON(!block_group);
7406         BUG_ON(!block_group->ro);
7407
7408         /*
7409          * Free the reserved super bytes from this block group before
7410          * remove it.
7411          */
7412         free_excluded_extents(root, block_group);
7413
7414         memcpy(&key, &block_group->key, sizeof(key));
7415         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7416                                   BTRFS_BLOCK_GROUP_RAID1 |
7417                                   BTRFS_BLOCK_GROUP_RAID10))
7418                 factor = 2;
7419         else
7420                 factor = 1;
7421
7422         /* make sure this block group isn't part of an allocation cluster */
7423         cluster = &root->fs_info->data_alloc_cluster;
7424         spin_lock(&cluster->refill_lock);
7425         btrfs_return_cluster_to_free_space(block_group, cluster);
7426         spin_unlock(&cluster->refill_lock);
7427
7428         /*
7429          * make sure this block group isn't part of a metadata
7430          * allocation cluster
7431          */
7432         cluster = &root->fs_info->meta_alloc_cluster;
7433         spin_lock(&cluster->refill_lock);
7434         btrfs_return_cluster_to_free_space(block_group, cluster);
7435         spin_unlock(&cluster->refill_lock);
7436
7437         path = btrfs_alloc_path();
7438         if (!path) {
7439                 ret = -ENOMEM;
7440                 goto out;
7441         }
7442
7443         inode = lookup_free_space_inode(tree_root, block_group, path);
7444         if (!IS_ERR(inode)) {
7445                 ret = btrfs_orphan_add(trans, inode);
7446                 BUG_ON(ret);
7447                 clear_nlink(inode);
7448                 /* One for the block groups ref */
7449                 spin_lock(&block_group->lock);
7450                 if (block_group->iref) {
7451                         block_group->iref = 0;
7452                         block_group->inode = NULL;
7453                         spin_unlock(&block_group->lock);
7454                         iput(inode);
7455                 } else {
7456                         spin_unlock(&block_group->lock);
7457                 }
7458                 /* One for our lookup ref */
7459                 btrfs_add_delayed_iput(inode);
7460         }
7461
7462         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
7463         key.offset = block_group->key.objectid;
7464         key.type = 0;
7465
7466         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
7467         if (ret < 0)
7468                 goto out;
7469         if (ret > 0)
7470                 btrfs_release_path(path);
7471         if (ret == 0) {
7472                 ret = btrfs_del_item(trans, tree_root, path);
7473                 if (ret)
7474                         goto out;
7475                 btrfs_release_path(path);
7476         }
7477
7478         spin_lock(&root->fs_info->block_group_cache_lock);
7479         rb_erase(&block_group->cache_node,
7480                  &root->fs_info->block_group_cache_tree);
7481         spin_unlock(&root->fs_info->block_group_cache_lock);
7482
7483         down_write(&block_group->space_info->groups_sem);
7484         /*
7485          * we must use list_del_init so people can check to see if they
7486          * are still on the list after taking the semaphore
7487          */
7488         list_del_init(&block_group->list);
7489         up_write(&block_group->space_info->groups_sem);
7490
7491         if (block_group->cached == BTRFS_CACHE_STARTED)
7492                 wait_block_group_cache_done(block_group);
7493
7494         btrfs_remove_free_space_cache(block_group);
7495
7496         spin_lock(&block_group->space_info->lock);
7497         block_group->space_info->total_bytes -= block_group->key.offset;
7498         block_group->space_info->bytes_readonly -= block_group->key.offset;
7499         block_group->space_info->disk_total -= block_group->key.offset * factor;
7500         spin_unlock(&block_group->space_info->lock);
7501
7502         memcpy(&key, &block_group->key, sizeof(key));
7503
7504         btrfs_clear_space_info_full(root->fs_info);
7505
7506         btrfs_put_block_group(block_group);
7507         btrfs_put_block_group(block_group);
7508
7509         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
7510         if (ret > 0)
7511                 ret = -EIO;
7512         if (ret < 0)
7513                 goto out;
7514
7515         ret = btrfs_del_item(trans, root, path);
7516 out:
7517         btrfs_free_path(path);
7518         return ret;
7519 }
7520
7521 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
7522 {
7523         struct btrfs_space_info *space_info;
7524         struct btrfs_super_block *disk_super;
7525         u64 features;
7526         u64 flags;
7527         int mixed = 0;
7528         int ret;
7529
7530         disk_super = fs_info->super_copy;
7531         if (!btrfs_super_root(disk_super))
7532                 return 1;
7533
7534         features = btrfs_super_incompat_flags(disk_super);
7535         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
7536                 mixed = 1;
7537
7538         flags = BTRFS_BLOCK_GROUP_SYSTEM;
7539         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7540         if (ret)
7541                 goto out;
7542
7543         if (mixed) {
7544                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
7545                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7546         } else {
7547                 flags = BTRFS_BLOCK_GROUP_METADATA;
7548                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7549                 if (ret)
7550                         goto out;
7551
7552                 flags = BTRFS_BLOCK_GROUP_DATA;
7553                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7554         }
7555 out:
7556         return ret;
7557 }
7558
7559 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
7560 {
7561         return unpin_extent_range(root, start, end);
7562 }
7563
7564 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
7565                                u64 num_bytes, u64 *actual_bytes)
7566 {
7567         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
7568 }
7569
7570 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
7571 {
7572         struct btrfs_fs_info *fs_info = root->fs_info;
7573         struct btrfs_block_group_cache *cache = NULL;
7574         u64 group_trimmed;
7575         u64 start;
7576         u64 end;
7577         u64 trimmed = 0;
7578         int ret = 0;
7579
7580         cache = btrfs_lookup_block_group(fs_info, range->start);
7581
7582         while (cache) {
7583                 if (cache->key.objectid >= (range->start + range->len)) {
7584                         btrfs_put_block_group(cache);
7585                         break;
7586                 }
7587
7588                 start = max(range->start, cache->key.objectid);
7589                 end = min(range->start + range->len,
7590                                 cache->key.objectid + cache->key.offset);
7591
7592                 if (end - start >= range->minlen) {
7593                         if (!block_group_cache_done(cache)) {
7594                                 ret = cache_block_group(cache, NULL, root, 0);
7595                                 if (!ret)
7596                                         wait_block_group_cache_done(cache);
7597                         }
7598                         ret = btrfs_trim_block_group(cache,
7599                                                      &group_trimmed,
7600                                                      start,
7601                                                      end,
7602                                                      range->minlen);
7603
7604                         trimmed += group_trimmed;
7605                         if (ret) {
7606                                 btrfs_put_block_group(cache);
7607                                 break;
7608                         }
7609                 }
7610
7611                 cache = next_block_group(fs_info->tree_root, cache);
7612         }
7613
7614         range->len = trimmed;
7615         return ret;
7616 }