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 /*
  38  * control flags for do_chunk_alloc's force field
  39  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  40  * if we really need one.
  41  *
  42  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  43  * if we have very few chunks already allocated.  This is
  44  * used as part of the clustering code to help make sure
  45  * we have a good pool of storage to cluster in, without
  46  * filling the FS with empty chunks
  47  *
  48  * CHUNK_ALLOC_FORCE means it must try to allocate one
  49  *
  50  */
  51 enum {
  52         CHUNK_ALLOC_NO_FORCE = 0,
  53         CHUNK_ALLOC_LIMITED = 1,
  54         CHUNK_ALLOC_FORCE = 2,
  55 };
  56
  57 /*
  58  * Control how reservations are dealt with.
  59  *
  60  * RESERVE_FREE - freeing a reservation.
  61  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  62  *   ENOSPC accounting
  63  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  64  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  65  */
  66 enum {
  67         RESERVE_FREE = 0,
  68         RESERVE_ALLOC = 1,
  69         RESERVE_ALLOC_NO_ACCOUNT = 2,
  70 };
  71
  72 static int update_block_group(struct btrfs_trans_handle *trans,
  73                               struct btrfs_root *root,
  74                               u64 bytenr, u64 num_bytes, int alloc);
  75 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  76                                 struct btrfs_root *root,
  77                                 u64 bytenr, u64 num_bytes, u64 parent,
  78                                 u64 root_objectid, u64 owner_objectid,
  79                                 u64 owner_offset, int refs_to_drop,
  80                                 struct btrfs_delayed_extent_op *extra_op);
  81 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  82                                     struct extent_buffer *leaf,
  83                                     struct btrfs_extent_item *ei);
  84 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  85                                       struct btrfs_root *root,
  86                                       u64 parent, u64 root_objectid,
  87                                       u64 flags, u64 owner, u64 offset,
  88                                       struct btrfs_key *ins, int ref_mod);
  89 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  90                                      struct btrfs_root *root,
  91                                      u64 parent, u64 root_objectid,
  92                                      u64 flags, struct btrfs_disk_key *key,
  93                                      int level, struct btrfs_key *ins);
  94 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  95                           struct btrfs_root *extent_root, u64 alloc_bytes,
  96                           u64 flags, int force);
  97 static int find_next_key(struct btrfs_path *path, int level,
  98                          struct btrfs_key *key);
  99 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 100                             int dump_block_groups);
 101 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 102                                        u64 num_bytes, int reserve);
 103
 104 static noinline int
 105 block_group_cache_done(struct btrfs_block_group_cache *cache)
 106 {
 107         smp_mb();
 108         return cache->cached == BTRFS_CACHE_FINISHED;
 109 }
 110
 111 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 112 {
 113         return (cache->flags & bits) == bits;
 114 }
 115
 116 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 117 {
 118         atomic_inc(&cache->count);
 119 }
 120
 121 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 122 {
 123         if (atomic_dec_and_test(&cache->count)) {
 124                 WARN_ON(cache->pinned > 0);
 125                 WARN_ON(cache->reserved > 0);
 126                 kfree(cache->free_space_ctl);
 127                 kfree(cache);
 128         }
 129 }
 130
 131 /*
 132  * this adds the block group to the fs_info rb tree for the block group
 133  * cache
 134  */
 135 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 136                                 struct btrfs_block_group_cache *block_group)
 137 {
 138         struct rb_node **p;
 139         struct rb_node *parent = NULL;
 140         struct btrfs_block_group_cache *cache;
 141
 142         spin_lock(&info->block_group_cache_lock);
 143         p = &info->block_group_cache_tree.rb_node;
 144
 145         while (*p) {
 146                 parent = *p;
 147                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 148                                  cache_node);
 149                 if (block_group->key.objectid < cache->key.objectid) {
 150                         p = &(*p)->rb_left;
 151                 } else if (block_group->key.objectid > cache->key.objectid) {
 152                         p = &(*p)->rb_right;
 153                 } else {
 154                         spin_unlock(&info->block_group_cache_lock);
 155                         return -EEXIST;
 156                 }
 157         }
 158
 159         rb_link_node(&block_group->cache_node, parent, p);
 160         rb_insert_color(&block_group->cache_node,
 161                         &info->block_group_cache_tree);
 162         spin_unlock(&info->block_group_cache_lock);
 163
 164         return 0;
 165 }
 166
 167 /*
 168  * This will return the block group at or after bytenr if contains is 0, else
 169  * it will return the block group that contains the bytenr
 170  */
 171 static struct btrfs_block_group_cache *
 172 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 173                               int contains)
 174 {
 175         struct btrfs_block_group_cache *cache, *ret = NULL;
 176         struct rb_node *n;
 177         u64 end, start;
 178
 179         spin_lock(&info->block_group_cache_lock);
 180         n = info->block_group_cache_tree.rb_node;
 181
 182         while (n) {
 183                 cache = rb_entry(n, struct btrfs_block_group_cache,
 184                                  cache_node);
 185                 end = cache->key.objectid + cache->key.offset - 1;
 186                 start = cache->key.objectid;
 187
 188                 if (bytenr < start) {
 189                         if (!contains && (!ret || start < ret->key.objectid))
 190                                 ret = cache;
 191                         n = n->rb_left;
 192                 } else if (bytenr > start) {
 193                         if (contains && bytenr <= end) {
 194                                 ret = cache;
 195                                 break;
 196                         }
 197                         n = n->rb_right;
 198                 } else {
 199                         ret = cache;
 200                         break;
 201                 }
 202         }
 203         if (ret)
 204                 btrfs_get_block_group(ret);
 205         spin_unlock(&info->block_group_cache_lock);
 206
 207         return ret;
 208 }
 209
 210 static int add_excluded_extent(struct btrfs_root *root,
 211                                u64 start, u64 num_bytes)
 212 {
 213         u64 end = start + num_bytes - 1;
 214         set_extent_bits(&root->fs_info->freed_extents[0],
 215                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 216         set_extent_bits(&root->fs_info->freed_extents[1],
 217                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 218         return 0;
 219 }
 220
 221 static void free_excluded_extents(struct btrfs_root *root,
 222                                   struct btrfs_block_group_cache *cache)
 223 {
 224         u64 start, end;
 225
 226         start = cache->key.objectid;
 227         end = start + cache->key.offset - 1;
 228
 229         clear_extent_bits(&root->fs_info->freed_extents[0],
 230                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 231         clear_extent_bits(&root->fs_info->freed_extents[1],
 232                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 233 }
 234
 235 static int exclude_super_stripes(struct btrfs_root *root,
 236                                  struct btrfs_block_group_cache *cache)
 237 {
 238         u64 bytenr;
 239         u64 *logical;
 240         int stripe_len;
 241         int i, nr, ret;
 242
 243         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 244                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 245                 cache->bytes_super += stripe_len;
 246                 ret = add_excluded_extent(root, cache->key.objectid,
 247                                           stripe_len);
 248                 BUG_ON(ret);
 249         }
 250
 251         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 252                 bytenr = btrfs_sb_offset(i);
 253                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 254                                        cache->key.objectid, bytenr,
 255                                        0, &logical, &nr, &stripe_len);
 256                 BUG_ON(ret);
 257
 258                 while (nr--) {
 259                         cache->bytes_super += stripe_len;
 260                         ret = add_excluded_extent(root, logical[nr],
 261                                                   stripe_len);
 262                         BUG_ON(ret);
 263                 }
 264
 265                 kfree(logical);
 266         }
 267         return 0;
 268 }
 269
 270 static struct btrfs_caching_control *
 271 get_caching_control(struct btrfs_block_group_cache *cache)
 272 {
 273         struct btrfs_caching_control *ctl;
 274
 275         spin_lock(&cache->lock);
 276         if (cache->cached != BTRFS_CACHE_STARTED) {
 277                 spin_unlock(&cache->lock);
 278                 return NULL;
 279         }
 280
 281         /* We're loading it the fast way, so we don't have a caching_ctl. */
 282         if (!cache->caching_ctl) {
 283                 spin_unlock(&cache->lock);
 284                 return NULL;
 285         }
 286
 287         ctl = cache->caching_ctl;
 288         atomic_inc(&ctl->count);
 289         spin_unlock(&cache->lock);
 290         return ctl;
 291 }
 292
 293 static void put_caching_control(struct btrfs_caching_control *ctl)
 294 {
 295         if (atomic_dec_and_test(&ctl->count))
 296                 kfree(ctl);
 297 }
 298
 299 /*
 300  * this is only called by cache_block_group, since we could have freed extents
 301  * we need to check the pinned_extents for any extents that can't be used yet
 302  * since their free space will be released as soon as the transaction commits.
 303  */
 304 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 305                               struct btrfs_fs_info *info, u64 start, u64 end)
 306 {
 307         u64 extent_start, extent_end, size, total_added = 0;
 308         int ret;
 309
 310         while (start < end) {
 311                 ret = find_first_extent_bit(info->pinned_extents, start,
 312                                             &extent_start, &extent_end,
 313                                             EXTENT_DIRTY | EXTENT_UPTODATE);
 314                 if (ret)
 315                         break;
 316
 317                 if (extent_start <= start) {
 318                         start = extent_end + 1;
 319                 } else if (extent_start > start && extent_start < end) {
 320                         size = extent_start - start;
 321                         total_added += size;
 322                         ret = btrfs_add_free_space(block_group, start,
 323                                                    size);
 324                         BUG_ON(ret);
 325                         start = extent_end + 1;
 326                 } else {
 327                         break;
 328                 }
 329         }
 330
 331         if (start < end) {
 332                 size = end - start;
 333                 total_added += size;
 334                 ret = btrfs_add_free_space(block_group, start, size);
 335                 BUG_ON(ret);
 336         }
 337
 338         return total_added;
 339 }
 340
 341 static noinline void caching_thread(struct btrfs_work *work)
 342 {
 343         struct btrfs_block_group_cache *block_group;
 344         struct btrfs_fs_info *fs_info;
 345         struct btrfs_caching_control *caching_ctl;
 346         struct btrfs_root *extent_root;
 347         struct btrfs_path *path;
 348         struct extent_buffer *leaf;
 349         struct btrfs_key key;
 350         u64 total_found = 0;
 351         u64 last = 0;
 352         u32 nritems;
 353         int ret = 0;
 354
 355         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 356         block_group = caching_ctl->block_group;
 357         fs_info = block_group->fs_info;
 358         extent_root = fs_info->extent_root;
 359
 360         path = btrfs_alloc_path();
 361         if (!path)
 362                 goto out;
 363
 364         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 365
 366         /*
 367          * We don't want to deadlock with somebody trying to allocate a new
 368          * extent for the extent root while also trying to search the extent
 369          * root to add free space.  So we skip locking and search the commit
 370          * root, since its read-only
 371          */
 372         path->skip_locking = 1;
 373         path->search_commit_root = 1;
 374         path->reada = 1;
 375
 376         key.objectid = last;
 377         key.offset = 0;
 378         key.type = BTRFS_EXTENT_ITEM_KEY;
 379 again:
 380         mutex_lock(&caching_ctl->mutex);
 381         /* need to make sure the commit_root doesn't disappear */
 382         down_read(&fs_info->extent_commit_sem);
 383
 384         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 385         if (ret < 0)
 386                 goto err;
 387
 388         leaf = path->nodes[0];
 389         nritems = btrfs_header_nritems(leaf);
 390
 391         while (1) {
 392                 if (btrfs_fs_closing(fs_info) > 1) {
 393                         last = (u64)-1;
 394                         break;
 395                 }
 396
 397                 if (path->slots[0] < nritems) {
 398                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 399                 } else {
 400                         ret = find_next_key(path, 0, &key);
 401                         if (ret)
 402                                 break;
 403
 404                         if (need_resched() ||
 405                             btrfs_next_leaf(extent_root, path)) {
 406                                 caching_ctl->progress = last;
 407                                 btrfs_release_path(path);
 408                                 up_read(&fs_info->extent_commit_sem);
 409                                 mutex_unlock(&caching_ctl->mutex);
 410                                 cond_resched();
 411                                 goto again;
 412                         }
 413                         leaf = path->nodes[0];
 414                         nritems = btrfs_header_nritems(leaf);
 415                         continue;
 416                 }
 417
 418                 if (key.objectid < block_group->key.objectid) {
 419                         path->slots[0]++;
 420                         continue;
 421                 }
 422
 423                 if (key.objectid >= block_group->key.objectid +
 424                     block_group->key.offset)
 425                         break;
 426
 427                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 428                         total_found += add_new_free_space(block_group,
 429                                                           fs_info, last,
 430                                                           key.objectid);
 431                         last = key.objectid + key.offset;
 432
 433                         if (total_found > (1024 * 1024 * 2)) {
 434                                 total_found = 0;
 435                                 wake_up(&caching_ctl->wait);
 436                         }
 437                 }
 438                 path->slots[0]++;
 439         }
 440         ret = 0;
 441
 442         total_found += add_new_free_space(block_group, fs_info, last,
 443                                           block_group->key.objectid +
 444                                           block_group->key.offset);
 445         caching_ctl->progress = (u64)-1;
 446
 447         spin_lock(&block_group->lock);
 448         block_group->caching_ctl = NULL;
 449         block_group->cached = BTRFS_CACHE_FINISHED;
 450         spin_unlock(&block_group->lock);
 451
 452 err:
 453         btrfs_free_path(path);
 454         up_read(&fs_info->extent_commit_sem);
 455
 456         free_excluded_extents(extent_root, block_group);
 457
 458         mutex_unlock(&caching_ctl->mutex);
 459 out:
 460         wake_up(&caching_ctl->wait);
 461
 462         put_caching_control(caching_ctl);
 463         btrfs_put_block_group(block_group);
 464 }
 465
 466 static int cache_block_group(struct btrfs_block_group_cache *cache,
 467                              struct btrfs_trans_handle *trans,
 468                              struct btrfs_root *root,
 469                              int load_cache_only)
 470 {
 471         DEFINE_WAIT(wait);
 472         struct btrfs_fs_info *fs_info = cache->fs_info;
 473         struct btrfs_caching_control *caching_ctl;
 474         int ret = 0;
 475
 476         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 477         BUG_ON(!caching_ctl);
 478
 479         INIT_LIST_HEAD(&caching_ctl->list);
 480         mutex_init(&caching_ctl->mutex);
 481         init_waitqueue_head(&caching_ctl->wait);
 482         caching_ctl->block_group = cache;
 483         caching_ctl->progress = cache->key.objectid;
 484         atomic_set(&caching_ctl->count, 1);
 485         caching_ctl->work.func = caching_thread;
 486
 487         spin_lock(&cache->lock);
 488         /*
 489          * This should be a rare occasion, but this could happen I think in the
 490          * case where one thread starts to load the space cache info, and then
 491          * some other thread starts a transaction commit which tries to do an
 492          * allocation while the other thread is still loading the space cache
 493          * info.  The previous loop should have kept us from choosing this block
 494          * group, but if we've moved to the state where we will wait on caching
 495          * block groups we need to first check if we're doing a fast load here,
 496          * so we can wait for it to finish, otherwise we could end up allocating
 497          * from a block group who's cache gets evicted for one reason or
 498          * another.
 499          */
 500         while (cache->cached == BTRFS_CACHE_FAST) {
 501                 struct btrfs_caching_control *ctl;
 502
 503                 ctl = cache->caching_ctl;
 504                 atomic_inc(&ctl->count);
 505                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 506                 spin_unlock(&cache->lock);
 507
 508                 schedule();
 509
 510                 finish_wait(&ctl->wait, &wait);
 511                 put_caching_control(ctl);
 512                 spin_lock(&cache->lock);
 513         }
 514
 515         if (cache->cached != BTRFS_CACHE_NO) {
 516                 spin_unlock(&cache->lock);
 517                 kfree(caching_ctl);
 518                 return 0;
 519         }
 520         WARN_ON(cache->caching_ctl);
 521         cache->caching_ctl = caching_ctl;
 522         cache->cached = BTRFS_CACHE_FAST;
 523         spin_unlock(&cache->lock);
 524
 525         /*
 526          * We can't do the read from on-disk cache during a commit since we need
 527          * to have the normal tree locking.  Also if we are currently trying to
 528          * allocate blocks for the tree root we can't do the fast caching since
 529          * we likely hold important locks.
 530          */
 531         if (trans && (!trans->transaction->in_commit) &&
 532             (root && root != root->fs_info->tree_root) &&
 533             btrfs_test_opt(root, SPACE_CACHE)) {
 534                 ret = load_free_space_cache(fs_info, cache);
 535
 536                 spin_lock(&cache->lock);
 537                 if (ret == 1) {
 538                         cache->caching_ctl = NULL;
 539                         cache->cached = BTRFS_CACHE_FINISHED;
 540                         cache->last_byte_to_unpin = (u64)-1;
 541                 } else {
 542                         if (load_cache_only) {
 543                                 cache->caching_ctl = NULL;
 544                                 cache->cached = BTRFS_CACHE_NO;
 545                         } else {
 546                                 cache->cached = BTRFS_CACHE_STARTED;
 547                         }
 548                 }
 549                 spin_unlock(&cache->lock);
 550                 wake_up(&caching_ctl->wait);
 551                 if (ret == 1) {
 552                         put_caching_control(caching_ctl);
 553                         free_excluded_extents(fs_info->extent_root, cache);
 554                         return 0;
 555                 }
 556         } else {
 557                 /*
 558                  * We are not going to do the fast caching, set cached to the
 559                  * appropriate value and wakeup any waiters.
 560                  */
 561                 spin_lock(&cache->lock);
 562                 if (load_cache_only) {
 563                         cache->caching_ctl = NULL;
 564                         cache->cached = BTRFS_CACHE_NO;
 565                 } else {
 566                         cache->cached = BTRFS_CACHE_STARTED;
 567                 }
 568                 spin_unlock(&cache->lock);
 569                 wake_up(&caching_ctl->wait);
 570         }
 571
 572         if (load_cache_only) {
 573                 put_caching_control(caching_ctl);
 574                 return 0;
 575         }
 576
 577         down_write(&fs_info->extent_commit_sem);
 578         atomic_inc(&caching_ctl->count);
 579         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 580         up_write(&fs_info->extent_commit_sem);
 581
 582         btrfs_get_block_group(cache);
 583
 584         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 585
 586         return ret;
 587 }
 588
 589 /*
 590  * return the block group that starts at or after bytenr
 591  */
 592 static struct btrfs_block_group_cache *
 593 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 594 {
 595         struct btrfs_block_group_cache *cache;
 596
 597         cache = block_group_cache_tree_search(info, bytenr, 0);
 598
 599         return cache;
 600 }
 601
 602 /*
 603  * return the block group that contains the given bytenr
 604  */
 605 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 606                                                  struct btrfs_fs_info *info,
 607                                                  u64 bytenr)
 608 {
 609         struct btrfs_block_group_cache *cache;
 610
 611         cache = block_group_cache_tree_search(info, bytenr, 1);
 612
 613         return cache;
 614 }
 615
 616 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 617                                                   u64 flags)
 618 {
 619         struct list_head *head = &info->space_info;
 620         struct btrfs_space_info *found;
 621
 622         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 623
 624         rcu_read_lock();
 625         list_for_each_entry_rcu(found, head, list) {
 626                 if (found->flags & flags) {
 627                         rcu_read_unlock();
 628                         return found;
 629                 }
 630         }
 631         rcu_read_unlock();
 632         return NULL;
 633 }
 634
 635 /*
 636  * after adding space to the filesystem, we need to clear the full flags
 637  * on all the space infos.
 638  */
 639 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 640 {
 641         struct list_head *head = &info->space_info;
 642         struct btrfs_space_info *found;
 643
 644         rcu_read_lock();
 645         list_for_each_entry_rcu(found, head, list)
 646                 found->full = 0;
 647         rcu_read_unlock();
 648 }
 649
 650 static u64 div_factor(u64 num, int factor)
 651 {
 652         if (factor == 10)
 653                 return num;
 654         num *= factor;
 655         do_div(num, 10);
 656         return num;
 657 }
 658
 659 static u64 div_factor_fine(u64 num, int factor)
 660 {
 661         if (factor == 100)
 662                 return num;
 663         num *= factor;
 664         do_div(num, 100);
 665         return num;
 666 }
 667
 668 u64 btrfs_find_block_group(struct btrfs_root *root,
 669                            u64 search_start, u64 search_hint, int owner)
 670 {
 671         struct btrfs_block_group_cache *cache;
 672         u64 used;
 673         u64 last = max(search_hint, search_start);
 674         u64 group_start = 0;
 675         int full_search = 0;
 676         int factor = 9;
 677         int wrapped = 0;
 678 again:
 679         while (1) {
 680                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
 681                 if (!cache)
 682                         break;
 683
 684                 spin_lock(&cache->lock);
 685                 last = cache->key.objectid + cache->key.offset;
 686                 used = btrfs_block_group_used(&cache->item);
 687
 688                 if ((full_search || !cache->ro) &&
 689                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
 690                         if (used + cache->pinned + cache->reserved <
 691                             div_factor(cache->key.offset, factor)) {
 692                                 group_start = cache->key.objectid;
 693                                 spin_unlock(&cache->lock);
 694                                 btrfs_put_block_group(cache);
 695                                 goto found;
 696                         }
 697                 }
 698                 spin_unlock(&cache->lock);
 699                 btrfs_put_block_group(cache);
 700                 cond_resched();
 701         }
 702         if (!wrapped) {
 703                 last = search_start;
 704                 wrapped = 1;
 705                 goto again;
 706         }
 707         if (!full_search && factor < 10) {
 708                 last = search_start;
 709                 full_search = 1;
 710                 factor = 10;
 711                 goto again;
 712         }
 713 found:
 714         return group_start;
 715 }
 716
 717 /* simple helper to search for an existing extent at a given offset */
 718 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 719 {
 720         int ret;
 721         struct btrfs_key key;
 722         struct btrfs_path *path;
 723
 724         path = btrfs_alloc_path();
 725         if (!path)
 726                 return -ENOMEM;
 727
 728         key.objectid = start;
 729         key.offset = len;
 730         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 731         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 732                                 0, 0);
 733         btrfs_free_path(path);
 734         return ret;
 735 }
 736
 737 /*
 738  * helper function to lookup reference count and flags of extent.
 739  *
 740  * the head node for delayed ref is used to store the sum of all the
 741  * reference count modifications queued up in the rbtree. the head
 742  * node may also store the extent flags to set. This way you can check
 743  * to see what the reference count and extent flags would be if all of
 744  * the delayed refs are not processed.
 745  */
 746 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 747                              struct btrfs_root *root, u64 bytenr,
 748                              u64 num_bytes, u64 *refs, u64 *flags)
 749 {
 750         struct btrfs_delayed_ref_head *head;
 751         struct btrfs_delayed_ref_root *delayed_refs;
 752         struct btrfs_path *path;
 753         struct btrfs_extent_item *ei;
 754         struct extent_buffer *leaf;
 755         struct btrfs_key key;
 756         u32 item_size;
 757         u64 num_refs;
 758         u64 extent_flags;
 759         int ret;
 760
 761         path = btrfs_alloc_path();
 762         if (!path)
 763                 return -ENOMEM;
 764
 765         key.objectid = bytenr;
 766         key.type = BTRFS_EXTENT_ITEM_KEY;
 767         key.offset = num_bytes;
 768         if (!trans) {
 769                 path->skip_locking = 1;
 770                 path->search_commit_root = 1;
 771         }
 772 again:
 773         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 774                                 &key, path, 0, 0);
 775         if (ret < 0)
 776                 goto out_free;
 777
 778         if (ret == 0) {
 779                 leaf = path->nodes[0];
 780                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 781                 if (item_size >= sizeof(*ei)) {
 782                         ei = btrfs_item_ptr(leaf, path->slots[0],
 783                                             struct btrfs_extent_item);
 784                         num_refs = btrfs_extent_refs(leaf, ei);
 785                         extent_flags = btrfs_extent_flags(leaf, ei);
 786                 } else {
 787 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 788                         struct btrfs_extent_item_v0 *ei0;
 789                         BUG_ON(item_size != sizeof(*ei0));
 790                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 791                                              struct btrfs_extent_item_v0);
 792                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 793                         /* FIXME: this isn't correct for data */
 794                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 795 #else
 796                         BUG();
 797 #endif
 798                 }
 799                 BUG_ON(num_refs == 0);
 800         } else {
 801                 num_refs = 0;
 802                 extent_flags = 0;
 803                 ret = 0;
 804         }
 805
 806         if (!trans)
 807                 goto out;
 808
 809         delayed_refs = &trans->transaction->delayed_refs;
 810         spin_lock(&delayed_refs->lock);
 811         head = btrfs_find_delayed_ref_head(trans, bytenr);
 812         if (head) {
 813                 if (!mutex_trylock(&head->mutex)) {
 814                         atomic_inc(&head->node.refs);
 815                         spin_unlock(&delayed_refs->lock);
 816
 817                         btrfs_release_path(path);
 818
 819                         /*
 820                          * Mutex was contended, block until it's released and try
 821                          * again
 822                          */
 823                         mutex_lock(&head->mutex);
 824                         mutex_unlock(&head->mutex);
 825                         btrfs_put_delayed_ref(&head->node);
 826                         goto again;
 827                 }
 828                 if (head->extent_op && head->extent_op->update_flags)
 829                         extent_flags |= head->extent_op->flags_to_set;
 830                 else
 831                         BUG_ON(num_refs == 0);
 832
 833                 num_refs += head->node.ref_mod;
 834                 mutex_unlock(&head->mutex);
 835         }
 836         spin_unlock(&delayed_refs->lock);
 837 out:
 838         WARN_ON(num_refs == 0);
 839         if (refs)
 840                 *refs = num_refs;
 841         if (flags)
 842                 *flags = extent_flags;
 843 out_free:
 844         btrfs_free_path(path);
 845         return ret;
 846 }
 847
 848 /*
 849  * Back reference rules.  Back refs have three main goals:
 850  *
 851  * 1) differentiate between all holders of references to an extent so that
 852  *    when a reference is dropped we can make sure it was a valid reference
 853  *    before freeing the extent.
 854  *
 855  * 2) Provide enough information to quickly find the holders of an extent
 856  *    if we notice a given block is corrupted or bad.
 857  *
 858  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 859  *    maintenance.  This is actually the same as #2, but with a slightly
 860  *    different use case.
 861  *
 862  * There are two kinds of back refs. The implicit back refs is optimized
 863  * for pointers in non-shared tree blocks. For a given pointer in a block,
 864  * back refs of this kind provide information about the block's owner tree
 865  * and the pointer's key. These information allow us to find the block by
 866  * b-tree searching. The full back refs is for pointers in tree blocks not
 867  * referenced by their owner trees. The location of tree block is recorded
 868  * in the back refs. Actually the full back refs is generic, and can be
 869  * used in all cases the implicit back refs is used. The major shortcoming
 870  * of the full back refs is its overhead. Every time a tree block gets
 871  * COWed, we have to update back refs entry for all pointers in it.
 872  *
 873  * For a newly allocated tree block, we use implicit back refs for
 874  * pointers in it. This means most tree related operations only involve
 875  * implicit back refs. For a tree block created in old transaction, the
 876  * only way to drop a reference to it is COW it. So we can detect the
 877  * event that tree block loses its owner tree's reference and do the
 878  * back refs conversion.
 879  *
 880  * When a tree block is COW'd through a tree, there are four cases:
 881  *
 882  * The reference count of the block is one and the tree is the block's
 883  * owner tree. Nothing to do in this case.
 884  *
 885  * The reference count of the block is one and the tree is not the
 886  * block's owner tree. In this case, full back refs is used for pointers
 887  * in the block. Remove these full back refs, add implicit back refs for
 888  * every pointers in the new block.
 889  *
 890  * The reference count of the block is greater than one and the tree is
 891  * the block's owner tree. In this case, implicit back refs is used for
 892  * pointers in the block. Add full back refs for every pointers in the
 893  * block, increase lower level extents' reference counts. The original
 894  * implicit back refs are entailed to the new block.
 895  *
 896  * The reference count of the block is greater than one and the tree is
 897  * not the block's owner tree. Add implicit back refs for every pointer in
 898  * the new block, increase lower level extents' reference count.
 899  *
 900  * Back Reference Key composing:
 901  *
 902  * The key objectid corresponds to the first byte in the extent,
 903  * The key type is used to differentiate between types of back refs.
 904  * There are different meanings of the key offset for different types
 905  * of back refs.
 906  *
 907  * File extents can be referenced by:
 908  *
 909  * - multiple snapshots, subvolumes, or different generations in one subvol
 910  * - different files inside a single subvolume
 911  * - different offsets inside a file (bookend extents in file.c)
 912  *
 913  * The extent ref structure for the implicit back refs has fields for:
 914  *
 915  * - Objectid of the subvolume root
 916  * - objectid of the file holding the reference
 917  * - original offset in the file
 918  * - how many bookend extents
 919  *
 920  * The key offset for the implicit back refs is hash of the first
 921  * three fields.
 922  *
 923  * The extent ref structure for the full back refs has field for:
 924  *
 925  * - number of pointers in the tree leaf
 926  *
 927  * The key offset for the implicit back refs is the first byte of
 928  * the tree leaf
 929  *
 930  * When a file extent is allocated, The implicit back refs is used.
 931  * the fields are filled in:
 932  *
 933  *     (root_key.objectid, inode objectid, offset in file, 1)
 934  *
 935  * When a file extent is removed file truncation, we find the
 936  * corresponding implicit back refs and check the following fields:
 937  *
 938  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 939  *
 940  * Btree extents can be referenced by:
 941  *
 942  * - Different subvolumes
 943  *
 944  * Both the implicit back refs and the full back refs for tree blocks
 945  * only consist of key. The key offset for the implicit back refs is
 946  * objectid of block's owner tree. The key offset for the full back refs
 947  * is the first byte of parent block.
 948  *
 949  * When implicit back refs is used, information about the lowest key and
 950  * level of the tree block are required. These information are stored in
 951  * tree block info structure.
 952  */
 953
 954 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 955 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 956                                   struct btrfs_root *root,
 957                                   struct btrfs_path *path,
 958                                   u64 owner, u32 extra_size)
 959 {
 960         struct btrfs_extent_item *item;
 961         struct btrfs_extent_item_v0 *ei0;
 962         struct btrfs_extent_ref_v0 *ref0;
 963         struct btrfs_tree_block_info *bi;
 964         struct extent_buffer *leaf;
 965         struct btrfs_key key;
 966         struct btrfs_key found_key;
 967         u32 new_size = sizeof(*item);
 968         u64 refs;
 969         int ret;
 970
 971         leaf = path->nodes[0];
 972         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 973
 974         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 975         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 976                              struct btrfs_extent_item_v0);
 977         refs = btrfs_extent_refs_v0(leaf, ei0);
 978
 979         if (owner == (u64)-1) {
 980                 while (1) {
 981                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 982                                 ret = btrfs_next_leaf(root, path);
 983                                 if (ret < 0)
 984                                         return ret;
 985                                 BUG_ON(ret > 0);
 986                                 leaf = path->nodes[0];
 987                         }
 988                         btrfs_item_key_to_cpu(leaf, &found_key,
 989                                               path->slots[0]);
 990                         BUG_ON(key.objectid != found_key.objectid);
 991                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 992                                 path->slots[0]++;
 993                                 continue;
 994                         }
 995                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 996                                               struct btrfs_extent_ref_v0);
 997                         owner = btrfs_ref_objectid_v0(leaf, ref0);
 998                         break;
 999                 }
1000         }
1001         btrfs_release_path(path);
1002
1003         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1004                 new_size += sizeof(*bi);
1005
1006         new_size -= sizeof(*ei0);
1007         ret = btrfs_search_slot(trans, root, &key, path,
1008                                 new_size + extra_size, 1);
1009         if (ret < 0)
1010                 return ret;
1011         BUG_ON(ret);
1012
1013         btrfs_extend_item(trans, root, path, new_size);
1014
1015         leaf = path->nodes[0];
1016         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1017         btrfs_set_extent_refs(leaf, item, refs);
1018         /* FIXME: get real generation */
1019         btrfs_set_extent_generation(leaf, item, 0);
1020         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1021                 btrfs_set_extent_flags(leaf, item,
1022                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1023                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1024                 bi = (struct btrfs_tree_block_info *)(item + 1);
1025                 /* FIXME: get first key of the block */
1026                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1027                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1028         } else {
1029                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1030         }
1031         btrfs_mark_buffer_dirty(leaf);
1032         return 0;
1033 }
1034 #endif
1035
1036 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1037 {
1038         u32 high_crc = ~(u32)0;
1039         u32 low_crc = ~(u32)0;
1040         __le64 lenum;
1041
1042         lenum = cpu_to_le64(root_objectid);
1043         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1044         lenum = cpu_to_le64(owner);
1045         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1046         lenum = cpu_to_le64(offset);
1047         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1048
1049         return ((u64)high_crc << 31) ^ (u64)low_crc;
1050 }
1051
1052 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1053                                      struct btrfs_extent_data_ref *ref)
1054 {
1055         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1056                                     btrfs_extent_data_ref_objectid(leaf, ref),
1057                                     btrfs_extent_data_ref_offset(leaf, ref));
1058 }
1059
1060 static int match_extent_data_ref(struct extent_buffer *leaf,
1061                                  struct btrfs_extent_data_ref *ref,
1062                                  u64 root_objectid, u64 owner, u64 offset)
1063 {
1064         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1065             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1066             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1067                 return 0;
1068         return 1;
1069 }
1070
1071 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1072                                            struct btrfs_root *root,
1073                                            struct btrfs_path *path,
1074                                            u64 bytenr, u64 parent,
1075                                            u64 root_objectid,
1076                                            u64 owner, u64 offset)
1077 {
1078         struct btrfs_key key;
1079         struct btrfs_extent_data_ref *ref;
1080         struct extent_buffer *leaf;
1081         u32 nritems;
1082         int ret;
1083         int recow;
1084         int err = -ENOENT;
1085
1086         key.objectid = bytenr;
1087         if (parent) {
1088                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1089                 key.offset = parent;
1090         } else {
1091                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1092                 key.offset = hash_extent_data_ref(root_objectid,
1093                                                   owner, offset);
1094         }
1095 again:
1096         recow = 0;
1097         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1098         if (ret < 0) {
1099                 err = ret;
1100                 goto fail;
1101         }
1102
1103         if (parent) {
1104                 if (!ret)
1105                         return 0;
1106 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1107                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1108                 btrfs_release_path(path);
1109                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1110                 if (ret < 0) {
1111                         err = ret;
1112                         goto fail;
1113                 }
1114                 if (!ret)
1115                         return 0;
1116 #endif
1117                 goto fail;
1118         }
1119
1120         leaf = path->nodes[0];
1121         nritems = btrfs_header_nritems(leaf);
1122         while (1) {
1123                 if (path->slots[0] >= nritems) {
1124                         ret = btrfs_next_leaf(root, path);
1125                         if (ret < 0)
1126                                 err = ret;
1127                         if (ret)
1128                                 goto fail;
1129
1130                         leaf = path->nodes[0];
1131                         nritems = btrfs_header_nritems(leaf);
1132                         recow = 1;
1133                 }
1134
1135                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1136                 if (key.objectid != bytenr ||
1137                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1138                         goto fail;
1139
1140                 ref = btrfs_item_ptr(leaf, path->slots[0],
1141                                      struct btrfs_extent_data_ref);
1142
1143                 if (match_extent_data_ref(leaf, ref, root_objectid,
1144                                           owner, offset)) {
1145                         if (recow) {
1146                                 btrfs_release_path(path);
1147                                 goto again;
1148                         }
1149                         err = 0;
1150                         break;
1151                 }
1152                 path->slots[0]++;
1153         }
1154 fail:
1155         return err;
1156 }
1157
1158 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1159                                            struct btrfs_root *root,
1160                                            struct btrfs_path *path,
1161                                            u64 bytenr, u64 parent,
1162                                            u64 root_objectid, u64 owner,
1163                                            u64 offset, int refs_to_add)
1164 {
1165         struct btrfs_key key;
1166         struct extent_buffer *leaf;
1167         u32 size;
1168         u32 num_refs;
1169         int ret;
1170
1171         key.objectid = bytenr;
1172         if (parent) {
1173                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1174                 key.offset = parent;
1175                 size = sizeof(struct btrfs_shared_data_ref);
1176         } else {
1177                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1178                 key.offset = hash_extent_data_ref(root_objectid,
1179                                                   owner, offset);
1180                 size = sizeof(struct btrfs_extent_data_ref);
1181         }
1182
1183         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1184         if (ret && ret != -EEXIST)
1185                 goto fail;
1186
1187         leaf = path->nodes[0];
1188         if (parent) {
1189                 struct btrfs_shared_data_ref *ref;
1190                 ref = btrfs_item_ptr(leaf, path->slots[0],
1191                                      struct btrfs_shared_data_ref);
1192                 if (ret == 0) {
1193                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1194                 } else {
1195                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1196                         num_refs += refs_to_add;
1197                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1198                 }
1199         } else {
1200                 struct btrfs_extent_data_ref *ref;
1201                 while (ret == -EEXIST) {
1202                         ref = btrfs_item_ptr(leaf, path->slots[0],
1203                                              struct btrfs_extent_data_ref);
1204                         if (match_extent_data_ref(leaf, ref, root_objectid,
1205                                                   owner, offset))
1206                                 break;
1207                         btrfs_release_path(path);
1208                         key.offset++;
1209                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1210                                                       size);
1211                         if (ret && ret != -EEXIST)
1212                                 goto fail;
1213
1214                         leaf = path->nodes[0];
1215                 }
1216                 ref = btrfs_item_ptr(leaf, path->slots[0],
1217                                      struct btrfs_extent_data_ref);
1218                 if (ret == 0) {
1219                         btrfs_set_extent_data_ref_root(leaf, ref,
1220                                                        root_objectid);
1221                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1222                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1223                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1224                 } else {
1225                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1226                         num_refs += refs_to_add;
1227                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1228                 }
1229         }
1230         btrfs_mark_buffer_dirty(leaf);
1231         ret = 0;
1232 fail:
1233         btrfs_release_path(path);
1234         return ret;
1235 }
1236
1237 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1238                                            struct btrfs_root *root,
1239                                            struct btrfs_path *path,
1240                                            int refs_to_drop)
1241 {
1242         struct btrfs_key key;
1243         struct btrfs_extent_data_ref *ref1 = NULL;
1244         struct btrfs_shared_data_ref *ref2 = NULL;
1245         struct extent_buffer *leaf;
1246         u32 num_refs = 0;
1247         int ret = 0;
1248
1249         leaf = path->nodes[0];
1250         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1251
1252         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1253                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1254                                       struct btrfs_extent_data_ref);
1255                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1256         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1257                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1258                                       struct btrfs_shared_data_ref);
1259                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1260 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1261         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1262                 struct btrfs_extent_ref_v0 *ref0;
1263                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1264                                       struct btrfs_extent_ref_v0);
1265                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1266 #endif
1267         } else {
1268                 BUG();
1269         }
1270
1271         BUG_ON(num_refs < refs_to_drop);
1272         num_refs -= refs_to_drop;
1273
1274         if (num_refs == 0) {
1275                 ret = btrfs_del_item(trans, root, path);
1276         } else {
1277                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1278                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1279                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1280                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1281 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1282                 else {
1283                         struct btrfs_extent_ref_v0 *ref0;
1284                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1285                                         struct btrfs_extent_ref_v0);
1286                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1287                 }
1288 #endif
1289                 btrfs_mark_buffer_dirty(leaf);
1290         }
1291         return ret;
1292 }
1293
1294 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1295                                           struct btrfs_path *path,
1296                                           struct btrfs_extent_inline_ref *iref)
1297 {
1298         struct btrfs_key key;
1299         struct extent_buffer *leaf;
1300         struct btrfs_extent_data_ref *ref1;
1301         struct btrfs_shared_data_ref *ref2;
1302         u32 num_refs = 0;
1303
1304         leaf = path->nodes[0];
1305         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1306         if (iref) {
1307                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1308                     BTRFS_EXTENT_DATA_REF_KEY) {
1309                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1310                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1311                 } else {
1312                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1313                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1314                 }
1315         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1316                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1317                                       struct btrfs_extent_data_ref);
1318                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1319         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1320                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1321                                       struct btrfs_shared_data_ref);
1322                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1323 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1324         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1325                 struct btrfs_extent_ref_v0 *ref0;
1326                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1327                                       struct btrfs_extent_ref_v0);
1328                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1329 #endif
1330         } else {
1331                 WARN_ON(1);
1332         }
1333         return num_refs;
1334 }
1335
1336 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1337                                           struct btrfs_root *root,
1338                                           struct btrfs_path *path,
1339                                           u64 bytenr, u64 parent,
1340                                           u64 root_objectid)
1341 {
1342         struct btrfs_key key;
1343         int ret;
1344
1345         key.objectid = bytenr;
1346         if (parent) {
1347                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1348                 key.offset = parent;
1349         } else {
1350                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1351                 key.offset = root_objectid;
1352         }
1353
1354         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1355         if (ret > 0)
1356                 ret = -ENOENT;
1357 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1358         if (ret == -ENOENT && parent) {
1359                 btrfs_release_path(path);
1360                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1361                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1362                 if (ret > 0)
1363                         ret = -ENOENT;
1364         }
1365 #endif
1366         return ret;
1367 }
1368
1369 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1370                                           struct btrfs_root *root,
1371                                           struct btrfs_path *path,
1372                                           u64 bytenr, u64 parent,
1373                                           u64 root_objectid)
1374 {
1375         struct btrfs_key key;
1376         int ret;
1377
1378         key.objectid = bytenr;
1379         if (parent) {
1380                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1381                 key.offset = parent;
1382         } else {
1383                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1384                 key.offset = root_objectid;
1385         }
1386
1387         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1388         btrfs_release_path(path);
1389         return ret;
1390 }
1391
1392 static inline int extent_ref_type(u64 parent, u64 owner)
1393 {
1394         int type;
1395         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1396                 if (parent > 0)
1397                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1398                 else
1399                         type = BTRFS_TREE_BLOCK_REF_KEY;
1400         } else {
1401                 if (parent > 0)
1402                         type = BTRFS_SHARED_DATA_REF_KEY;
1403                 else
1404                         type = BTRFS_EXTENT_DATA_REF_KEY;
1405         }
1406         return type;
1407 }
1408
1409 static int find_next_key(struct btrfs_path *path, int level,
1410                          struct btrfs_key *key)
1411
1412 {
1413         for (; level < BTRFS_MAX_LEVEL; level++) {
1414                 if (!path->nodes[level])
1415                         break;
1416                 if (path->slots[level] + 1 >=
1417                     btrfs_header_nritems(path->nodes[level]))
1418                         continue;
1419                 if (level == 0)
1420                         btrfs_item_key_to_cpu(path->nodes[level], key,
1421                                               path->slots[level] + 1);
1422                 else
1423                         btrfs_node_key_to_cpu(path->nodes[level], key,
1424                                               path->slots[level] + 1);
1425                 return 0;
1426         }
1427         return 1;
1428 }
1429
1430 /*
1431  * look for inline back ref. if back ref is found, *ref_ret is set
1432  * to the address of inline back ref, and 0 is returned.
1433  *
1434  * if back ref isn't found, *ref_ret is set to the address where it
1435  * should be inserted, and -ENOENT is returned.
1436  *
1437  * if insert is true and there are too many inline back refs, the path
1438  * points to the extent item, and -EAGAIN is returned.
1439  *
1440  * NOTE: inline back refs are ordered in the same way that back ref
1441  *       items in the tree are ordered.
1442  */
1443 static noinline_for_stack
1444 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1445                                  struct btrfs_root *root,
1446                                  struct btrfs_path *path,
1447                                  struct btrfs_extent_inline_ref **ref_ret,
1448                                  u64 bytenr, u64 num_bytes,
1449                                  u64 parent, u64 root_objectid,
1450                                  u64 owner, u64 offset, int insert)
1451 {
1452         struct btrfs_key key;
1453         struct extent_buffer *leaf;
1454         struct btrfs_extent_item *ei;
1455         struct btrfs_extent_inline_ref *iref;
1456         u64 flags;
1457         u64 item_size;
1458         unsigned long ptr;
1459         unsigned long end;
1460         int extra_size;
1461         int type;
1462         int want;
1463         int ret;
1464         int err = 0;
1465
1466         key.objectid = bytenr;
1467         key.type = BTRFS_EXTENT_ITEM_KEY;
1468         key.offset = num_bytes;
1469
1470         want = extent_ref_type(parent, owner);
1471         if (insert) {
1472                 extra_size = btrfs_extent_inline_ref_size(want);
1473                 path->keep_locks = 1;
1474         } else
1475                 extra_size = -1;
1476         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1477         if (ret < 0) {
1478                 err = ret;
1479                 goto out;
1480         }
1481         BUG_ON(ret);
1482
1483         leaf = path->nodes[0];
1484         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1485 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1486         if (item_size < sizeof(*ei)) {
1487                 if (!insert) {
1488                         err = -ENOENT;
1489                         goto out;
1490                 }
1491                 ret = convert_extent_item_v0(trans, root, path, owner,
1492                                              extra_size);
1493                 if (ret < 0) {
1494                         err = ret;
1495                         goto out;
1496                 }
1497                 leaf = path->nodes[0];
1498                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1499         }
1500 #endif
1501         BUG_ON(item_size < sizeof(*ei));
1502
1503         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1504         flags = btrfs_extent_flags(leaf, ei);
1505
1506         ptr = (unsigned long)(ei + 1);
1507         end = (unsigned long)ei + item_size;
1508
1509         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1510                 ptr += sizeof(struct btrfs_tree_block_info);
1511                 BUG_ON(ptr > end);
1512         } else {
1513                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1514         }
1515
1516         err = -ENOENT;
1517         while (1) {
1518                 if (ptr >= end) {
1519                         WARN_ON(ptr > end);
1520                         break;
1521                 }
1522                 iref = (struct btrfs_extent_inline_ref *)ptr;
1523                 type = btrfs_extent_inline_ref_type(leaf, iref);
1524                 if (want < type)
1525                         break;
1526                 if (want > type) {
1527                         ptr += btrfs_extent_inline_ref_size(type);
1528                         continue;
1529                 }
1530
1531                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1532                         struct btrfs_extent_data_ref *dref;
1533                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1534                         if (match_extent_data_ref(leaf, dref, root_objectid,
1535                                                   owner, offset)) {
1536                                 err = 0;
1537                                 break;
1538                         }
1539                         if (hash_extent_data_ref_item(leaf, dref) <
1540                             hash_extent_data_ref(root_objectid, owner, offset))
1541                                 break;
1542                 } else {
1543                         u64 ref_offset;
1544                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1545                         if (parent > 0) {
1546                                 if (parent == ref_offset) {
1547                                         err = 0;
1548                                         break;
1549                                 }
1550                                 if (ref_offset < parent)
1551                                         break;
1552                         } else {
1553                                 if (root_objectid == ref_offset) {
1554                                         err = 0;
1555                                         break;
1556                                 }
1557                                 if (ref_offset < root_objectid)
1558                                         break;
1559                         }
1560                 }
1561                 ptr += btrfs_extent_inline_ref_size(type);
1562         }
1563         if (err == -ENOENT && insert) {
1564                 if (item_size + extra_size >=
1565                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1566                         err = -EAGAIN;
1567                         goto out;
1568                 }
1569                 /*
1570                  * To add new inline back ref, we have to make sure
1571                  * there is no corresponding back ref item.
1572                  * For simplicity, we just do not add new inline back
1573                  * ref if there is any kind of item for this block
1574                  */
1575                 if (find_next_key(path, 0, &key) == 0 &&
1576                     key.objectid == bytenr &&
1577                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1578                         err = -EAGAIN;
1579                         goto out;
1580                 }
1581         }
1582         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1583 out:
1584         if (insert) {
1585                 path->keep_locks = 0;
1586                 btrfs_unlock_up_safe(path, 1);
1587         }
1588         return err;
1589 }
1590
1591 /*
1592  * helper to add new inline back ref
1593  */
1594 static noinline_for_stack
1595 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1596                                  struct btrfs_root *root,
1597                                  struct btrfs_path *path,
1598                                  struct btrfs_extent_inline_ref *iref,
1599                                  u64 parent, u64 root_objectid,
1600                                  u64 owner, u64 offset, int refs_to_add,
1601                                  struct btrfs_delayed_extent_op *extent_op)
1602 {
1603         struct extent_buffer *leaf;
1604         struct btrfs_extent_item *ei;
1605         unsigned long ptr;
1606         unsigned long end;
1607         unsigned long item_offset;
1608         u64 refs;
1609         int size;
1610         int type;
1611
1612         leaf = path->nodes[0];
1613         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1614         item_offset = (unsigned long)iref - (unsigned long)ei;
1615
1616         type = extent_ref_type(parent, owner);
1617         size = btrfs_extent_inline_ref_size(type);
1618
1619         btrfs_extend_item(trans, root, path, size);
1620
1621         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1622         refs = btrfs_extent_refs(leaf, ei);
1623         refs += refs_to_add;
1624         btrfs_set_extent_refs(leaf, ei, refs);
1625         if (extent_op)
1626                 __run_delayed_extent_op(extent_op, leaf, ei);
1627
1628         ptr = (unsigned long)ei + item_offset;
1629         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1630         if (ptr < end - size)
1631                 memmove_extent_buffer(leaf, ptr + size, ptr,
1632                                       end - size - ptr);
1633
1634         iref = (struct btrfs_extent_inline_ref *)ptr;
1635         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1636         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1637                 struct btrfs_extent_data_ref *dref;
1638                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1639                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1640                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1641                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1642                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1643         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1644                 struct btrfs_shared_data_ref *sref;
1645                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1646                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1647                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1648         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1649                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1650         } else {
1651                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1652         }
1653         btrfs_mark_buffer_dirty(leaf);
1654 }
1655
1656 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1657                                  struct btrfs_root *root,
1658                                  struct btrfs_path *path,
1659                                  struct btrfs_extent_inline_ref **ref_ret,
1660                                  u64 bytenr, u64 num_bytes, u64 parent,
1661                                  u64 root_objectid, u64 owner, u64 offset)
1662 {
1663         int ret;
1664
1665         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1666                                            bytenr, num_bytes, parent,
1667                                            root_objectid, owner, offset, 0);
1668         if (ret != -ENOENT)
1669                 return ret;
1670
1671         btrfs_release_path(path);
1672         *ref_ret = NULL;
1673
1674         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1675                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1676                                             root_objectid);
1677         } else {
1678                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1679                                              root_objectid, owner, offset);
1680         }
1681         return ret;
1682 }
1683
1684 /*
1685  * helper to update/remove inline back ref
1686  */
1687 static noinline_for_stack
1688 void update_inline_extent_backref(struct btrfs_trans_handle *trans,
1689                                   struct btrfs_root *root,
1690                                   struct btrfs_path *path,
1691                                   struct btrfs_extent_inline_ref *iref,
1692                                   int refs_to_mod,
1693                                   struct btrfs_delayed_extent_op *extent_op)
1694 {
1695         struct extent_buffer *leaf;
1696         struct btrfs_extent_item *ei;
1697         struct btrfs_extent_data_ref *dref = NULL;
1698         struct btrfs_shared_data_ref *sref = NULL;
1699         unsigned long ptr;
1700         unsigned long end;
1701         u32 item_size;
1702         int size;
1703         int type;
1704         u64 refs;
1705
1706         leaf = path->nodes[0];
1707         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1708         refs = btrfs_extent_refs(leaf, ei);
1709         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1710         refs += refs_to_mod;
1711         btrfs_set_extent_refs(leaf, ei, refs);
1712         if (extent_op)
1713                 __run_delayed_extent_op(extent_op, leaf, ei);
1714
1715         type = btrfs_extent_inline_ref_type(leaf, iref);
1716
1717         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1718                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1719                 refs = btrfs_extent_data_ref_count(leaf, dref);
1720         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1721                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1722                 refs = btrfs_shared_data_ref_count(leaf, sref);
1723         } else {
1724                 refs = 1;
1725                 BUG_ON(refs_to_mod != -1);
1726         }
1727
1728         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1729         refs += refs_to_mod;
1730
1731         if (refs > 0) {
1732                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1733                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1734                 else
1735                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1736         } else {
1737                 size =  btrfs_extent_inline_ref_size(type);
1738                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1739                 ptr = (unsigned long)iref;
1740                 end = (unsigned long)ei + item_size;
1741                 if (ptr + size < end)
1742                         memmove_extent_buffer(leaf, ptr, ptr + size,
1743                                               end - ptr - size);
1744                 item_size -= size;
1745                 btrfs_truncate_item(trans, root, path, item_size, 1);
1746         }
1747         btrfs_mark_buffer_dirty(leaf);
1748 }
1749
1750 static noinline_for_stack
1751 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1752                                  struct btrfs_root *root,
1753                                  struct btrfs_path *path,
1754                                  u64 bytenr, u64 num_bytes, u64 parent,
1755                                  u64 root_objectid, u64 owner,
1756                                  u64 offset, int refs_to_add,
1757                                  struct btrfs_delayed_extent_op *extent_op)
1758 {
1759         struct btrfs_extent_inline_ref *iref;
1760         int ret;
1761
1762         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1763                                            bytenr, num_bytes, parent,
1764                                            root_objectid, owner, offset, 1);
1765         if (ret == 0) {
1766                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1767                 update_inline_extent_backref(trans, root, path, iref,
1768                                              refs_to_add, extent_op);
1769         } else if (ret == -ENOENT) {
1770                 setup_inline_extent_backref(trans, root, path, iref, parent,
1771                                             root_objectid, owner, offset,
1772                                             refs_to_add, extent_op);
1773                 ret = 0;
1774         }
1775         return ret;
1776 }
1777
1778 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1779                                  struct btrfs_root *root,
1780                                  struct btrfs_path *path,
1781                                  u64 bytenr, u64 parent, u64 root_objectid,
1782                                  u64 owner, u64 offset, int refs_to_add)
1783 {
1784         int ret;
1785         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1786                 BUG_ON(refs_to_add != 1);
1787                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1788                                             parent, root_objectid);
1789         } else {
1790                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1791                                              parent, root_objectid,
1792                                              owner, offset, refs_to_add);
1793         }
1794         return ret;
1795 }
1796
1797 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1798                                  struct btrfs_root *root,
1799                                  struct btrfs_path *path,
1800                                  struct btrfs_extent_inline_ref *iref,
1801                                  int refs_to_drop, int is_data)
1802 {
1803         int ret = 0;
1804
1805         BUG_ON(!is_data && refs_to_drop != 1);
1806         if (iref) {
1807                 update_inline_extent_backref(trans, root, path, iref,
1808                                              -refs_to_drop, NULL);
1809         } else if (is_data) {
1810                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1811         } else {
1812                 ret = btrfs_del_item(trans, root, path);
1813         }
1814         return ret;
1815 }
1816
1817 static int btrfs_issue_discard(struct block_device *bdev,
1818                                 u64 start, u64 len)
1819 {
1820         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1821 }
1822
1823 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1824                                 u64 num_bytes, u64 *actual_bytes)
1825 {
1826         int ret;
1827         u64 discarded_bytes = 0;
1828         struct btrfs_bio *bbio = NULL;
1829
1830
1831         /* Tell the block device(s) that the sectors can be discarded */
1832         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1833                               bytenr, &num_bytes, &bbio, 0);
1834         if (!ret) {
1835                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1836                 int i;
1837
1838
1839                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1840                         if (!stripe->dev->can_discard)
1841                                 continue;
1842
1843                         ret = btrfs_issue_discard(stripe->dev->bdev,
1844                                                   stripe->physical,
1845                                                   stripe->length);
1846                         if (!ret)
1847                                 discarded_bytes += stripe->length;
1848                         else if (ret != -EOPNOTSUPP)
1849                                 break;
1850
1851                         /*
1852                          * Just in case we get back EOPNOTSUPP for some reason,
1853                          * just ignore the return value so we don't screw up
1854                          * people calling discard_extent.
1855                          */
1856                         ret = 0;
1857                 }
1858                 kfree(bbio);
1859         }
1860
1861         if (actual_bytes)
1862                 *actual_bytes = discarded_bytes;
1863
1864
1865         return ret;
1866 }
1867
1868 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1869                          struct btrfs_root *root,
1870                          u64 bytenr, u64 num_bytes, u64 parent,
1871                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1872 {
1873         int ret;
1874         struct btrfs_fs_info *fs_info = root->fs_info;
1875
1876         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1877                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1878
1879         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1880                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1881                                         num_bytes,
1882                                         parent, root_objectid, (int)owner,
1883                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1884         } else {
1885                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1886                                         num_bytes,
1887                                         parent, root_objectid, owner, offset,
1888                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1889         }
1890         return ret;
1891 }
1892
1893 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1894                                   struct btrfs_root *root,
1895                                   u64 bytenr, u64 num_bytes,
1896                                   u64 parent, u64 root_objectid,
1897                                   u64 owner, u64 offset, int refs_to_add,
1898                                   struct btrfs_delayed_extent_op *extent_op)
1899 {
1900         struct btrfs_path *path;
1901         struct extent_buffer *leaf;
1902         struct btrfs_extent_item *item;
1903         u64 refs;
1904         int ret;
1905         int err = 0;
1906
1907         path = btrfs_alloc_path();
1908         if (!path)
1909                 return -ENOMEM;
1910
1911         path->reada = 1;
1912         path->leave_spinning = 1;
1913         /* this will setup the path even if it fails to insert the back ref */
1914         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1915                                            path, bytenr, num_bytes, parent,
1916                                            root_objectid, owner, offset,
1917                                            refs_to_add, extent_op);
1918         if (ret == 0)
1919                 goto out;
1920
1921         if (ret != -EAGAIN) {
1922                 err = ret;
1923                 goto out;
1924         }
1925
1926         leaf = path->nodes[0];
1927         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1928         refs = btrfs_extent_refs(leaf, item);
1929         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1930         if (extent_op)
1931                 __run_delayed_extent_op(extent_op, leaf, item);
1932
1933         btrfs_mark_buffer_dirty(leaf);
1934         btrfs_release_path(path);
1935
1936         path->reada = 1;
1937         path->leave_spinning = 1;
1938
1939         /* now insert the actual backref */
1940         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1941                                     path, bytenr, parent, root_objectid,
1942                                     owner, offset, refs_to_add);
1943         BUG_ON(ret);
1944 out:
1945         btrfs_free_path(path);
1946         return err;
1947 }
1948
1949 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1950                                 struct btrfs_root *root,
1951                                 struct btrfs_delayed_ref_node *node,
1952                                 struct btrfs_delayed_extent_op *extent_op,
1953                                 int insert_reserved)
1954 {
1955         int ret = 0;
1956         struct btrfs_delayed_data_ref *ref;
1957         struct btrfs_key ins;
1958         u64 parent = 0;
1959         u64 ref_root = 0;
1960         u64 flags = 0;
1961
1962         ins.objectid = node->bytenr;
1963         ins.offset = node->num_bytes;
1964         ins.type = BTRFS_EXTENT_ITEM_KEY;
1965
1966         ref = btrfs_delayed_node_to_data_ref(node);
1967         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1968                 parent = ref->parent;
1969         else
1970                 ref_root = ref->root;
1971
1972         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1973                 if (extent_op) {
1974                         BUG_ON(extent_op->update_key);
1975                         flags |= extent_op->flags_to_set;
1976                 }
1977                 ret = alloc_reserved_file_extent(trans, root,
1978                                                  parent, ref_root, flags,
1979                                                  ref->objectid, ref->offset,
1980                                                  &ins, node->ref_mod);
1981         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1982                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1983                                              node->num_bytes, parent,
1984                                              ref_root, ref->objectid,
1985                                              ref->offset, node->ref_mod,
1986                                              extent_op);
1987         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1988                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1989                                           node->num_bytes, parent,
1990                                           ref_root, ref->objectid,
1991                                           ref->offset, node->ref_mod,
1992                                           extent_op);
1993         } else {
1994                 BUG();
1995         }
1996         return ret;
1997 }
1998
1999 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2000                                     struct extent_buffer *leaf,
2001                                     struct btrfs_extent_item *ei)
2002 {
2003         u64 flags = btrfs_extent_flags(leaf, ei);
2004         if (extent_op->update_flags) {
2005                 flags |= extent_op->flags_to_set;
2006                 btrfs_set_extent_flags(leaf, ei, flags);
2007         }
2008
2009         if (extent_op->update_key) {
2010                 struct btrfs_tree_block_info *bi;
2011                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2012                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2013                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2014         }
2015 }
2016
2017 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2018                                  struct btrfs_root *root,
2019                                  struct btrfs_delayed_ref_node *node,
2020                                  struct btrfs_delayed_extent_op *extent_op)
2021 {
2022         struct btrfs_key key;
2023         struct btrfs_path *path;
2024         struct btrfs_extent_item *ei;
2025         struct extent_buffer *leaf;
2026         u32 item_size;
2027         int ret;
2028         int err = 0;
2029
2030         path = btrfs_alloc_path();
2031         if (!path)
2032                 return -ENOMEM;
2033
2034         key.objectid = node->bytenr;
2035         key.type = BTRFS_EXTENT_ITEM_KEY;
2036         key.offset = node->num_bytes;
2037
2038         path->reada = 1;
2039         path->leave_spinning = 1;
2040         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2041                                 path, 0, 1);
2042         if (ret < 0) {
2043                 err = ret;
2044                 goto out;
2045         }
2046         if (ret > 0) {
2047                 err = -EIO;
2048                 goto out;
2049         }
2050
2051         leaf = path->nodes[0];
2052         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2053 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2054         if (item_size < sizeof(*ei)) {
2055                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2056                                              path, (u64)-1, 0);
2057                 if (ret < 0) {
2058                         err = ret;
2059                         goto out;
2060                 }
2061                 leaf = path->nodes[0];
2062                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2063         }
2064 #endif
2065         BUG_ON(item_size < sizeof(*ei));
2066         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2067         __run_delayed_extent_op(extent_op, leaf, ei);
2068
2069         btrfs_mark_buffer_dirty(leaf);
2070 out:
2071         btrfs_free_path(path);
2072         return err;
2073 }
2074
2075 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2076                                 struct btrfs_root *root,
2077                                 struct btrfs_delayed_ref_node *node,
2078                                 struct btrfs_delayed_extent_op *extent_op,
2079                                 int insert_reserved)
2080 {
2081         int ret = 0;
2082         struct btrfs_delayed_tree_ref *ref;
2083         struct btrfs_key ins;
2084         u64 parent = 0;
2085         u64 ref_root = 0;
2086
2087         ins.objectid = node->bytenr;
2088         ins.offset = node->num_bytes;
2089         ins.type = BTRFS_EXTENT_ITEM_KEY;
2090
2091         ref = btrfs_delayed_node_to_tree_ref(node);
2092         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2093                 parent = ref->parent;
2094         else
2095                 ref_root = ref->root;
2096
2097         BUG_ON(node->ref_mod != 1);
2098         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2099                 BUG_ON(!extent_op || !extent_op->update_flags ||
2100                        !extent_op->update_key);
2101                 ret = alloc_reserved_tree_block(trans, root,
2102                                                 parent, ref_root,
2103                                                 extent_op->flags_to_set,
2104                                                 &extent_op->key,
2105                                                 ref->level, &ins);
2106         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2107                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2108                                              node->num_bytes, parent, ref_root,
2109                                              ref->level, 0, 1, extent_op);
2110         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2111                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2112                                           node->num_bytes, parent, ref_root,
2113                                           ref->level, 0, 1, extent_op);
2114         } else {
2115                 BUG();
2116         }
2117         return ret;
2118 }
2119
2120 /* helper function to actually process a single delayed ref entry */
2121 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2122                                struct btrfs_root *root,
2123                                struct btrfs_delayed_ref_node *node,
2124                                struct btrfs_delayed_extent_op *extent_op,
2125                                int insert_reserved)
2126 {
2127         int ret;
2128         if (btrfs_delayed_ref_is_head(node)) {
2129                 struct btrfs_delayed_ref_head *head;
2130                 /*
2131                  * we've hit the end of the chain and we were supposed
2132                  * to insert this extent into the tree.  But, it got
2133                  * deleted before we ever needed to insert it, so all
2134                  * we have to do is clean up the accounting
2135                  */
2136                 BUG_ON(extent_op);
2137                 head = btrfs_delayed_node_to_head(node);
2138                 if (insert_reserved) {
2139                         btrfs_pin_extent(root, node->bytenr,
2140                                          node->num_bytes, 1);
2141                         if (head->is_data) {
2142                                 ret = btrfs_del_csums(trans, root,
2143                                                       node->bytenr,
2144                                                       node->num_bytes);
2145                                 BUG_ON(ret);
2146                         }
2147                 }
2148                 mutex_unlock(&head->mutex);
2149                 return 0;
2150         }
2151
2152         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2153             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2154                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2155                                            insert_reserved);
2156         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2157                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2158                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2159                                            insert_reserved);
2160         else
2161                 BUG();
2162         return ret;
2163 }
2164
2165 static noinline struct btrfs_delayed_ref_node *
2166 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2167 {
2168         struct rb_node *node;
2169         struct btrfs_delayed_ref_node *ref;
2170         int action = BTRFS_ADD_DELAYED_REF;
2171 again:
2172         /*
2173          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2174          * this prevents ref count from going down to zero when
2175          * there still are pending delayed ref.
2176          */
2177         node = rb_prev(&head->node.rb_node);
2178         while (1) {
2179                 if (!node)
2180                         break;
2181                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2182                                 rb_node);
2183                 if (ref->bytenr != head->node.bytenr)
2184                         break;
2185                 if (ref->action == action)
2186                         return ref;
2187                 node = rb_prev(node);
2188         }
2189         if (action == BTRFS_ADD_DELAYED_REF) {
2190                 action = BTRFS_DROP_DELAYED_REF;
2191                 goto again;
2192         }
2193         return NULL;
2194 }
2195
2196 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2197                                        struct btrfs_root *root,
2198                                        struct list_head *cluster)
2199 {
2200         struct btrfs_delayed_ref_root *delayed_refs;
2201         struct btrfs_delayed_ref_node *ref;
2202         struct btrfs_delayed_ref_head *locked_ref = NULL;
2203         struct btrfs_delayed_extent_op *extent_op;
2204         int ret;
2205         int count = 0;
2206         int must_insert_reserved = 0;
2207
2208         delayed_refs = &trans->transaction->delayed_refs;
2209         while (1) {
2210                 if (!locked_ref) {
2211                         /* pick a new head ref from the cluster list */
2212                         if (list_empty(cluster))
2213                                 break;
2214
2215                         locked_ref = list_entry(cluster->next,
2216                                      struct btrfs_delayed_ref_head, cluster);
2217
2218                         /* grab the lock that says we are going to process
2219                          * all the refs for this head */
2220                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2221
2222                         /*
2223                          * we may have dropped the spin lock to get the head
2224                          * mutex lock, and that might have given someone else
2225                          * time to free the head.  If that's true, it has been
2226                          * removed from our list and we can move on.
2227                          */
2228                         if (ret == -EAGAIN) {
2229                                 locked_ref = NULL;
2230                                 count++;
2231                                 continue;
2232                         }
2233                 }
2234
2235                 /*
2236                  * locked_ref is the head node, so we have to go one
2237                  * node back for any delayed ref updates
2238                  */
2239                 ref = select_delayed_ref(locked_ref);
2240
2241                 if (ref && ref->seq &&
2242                     btrfs_check_delayed_seq(delayed_refs, ref->seq)) {
2243                         /*
2244                          * there are still refs with lower seq numbers in the
2245                          * process of being added. Don't run this ref yet.
2246                          */
2247                         list_del_init(&locked_ref->cluster);
2248                         mutex_unlock(&locked_ref->mutex);
2249                         locked_ref = NULL;
2250                         delayed_refs->num_heads_ready++;
2251                         spin_unlock(&delayed_refs->lock);
2252                         cond_resched();
2253                         spin_lock(&delayed_refs->lock);
2254                         continue;
2255                 }
2256
2257                 /*
2258                  * record the must insert reserved flag before we
2259                  * drop the spin lock.
2260                  */
2261                 must_insert_reserved = locked_ref->must_insert_reserved;
2262                 locked_ref->must_insert_reserved = 0;
2263
2264                 extent_op = locked_ref->extent_op;
2265                 locked_ref->extent_op = NULL;
2266
2267                 if (!ref) {
2268                         /* All delayed refs have been processed, Go ahead
2269                          * and send the head node to run_one_delayed_ref,
2270                          * so that any accounting fixes can happen
2271                          */
2272                         ref = &locked_ref->node;
2273
2274                         if (extent_op && must_insert_reserved) {
2275                                 kfree(extent_op);
2276                                 extent_op = NULL;
2277                         }
2278
2279                         if (extent_op) {
2280                                 spin_unlock(&delayed_refs->lock);
2281
2282                                 ret = run_delayed_extent_op(trans, root,
2283                                                             ref, extent_op);
2284                                 BUG_ON(ret);
2285                                 kfree(extent_op);
2286
2287                                 goto next;
2288                         }
2289
2290                         list_del_init(&locked_ref->cluster);
2291                         locked_ref = NULL;
2292                 }
2293
2294                 ref->in_tree = 0;
2295                 rb_erase(&ref->rb_node, &delayed_refs->root);
2296                 delayed_refs->num_entries--;
2297                 /*
2298                  * we modified num_entries, but as we're currently running
2299                  * delayed refs, skip
2300                  *     wake_up(&delayed_refs->seq_wait);
2301                  * here.
2302                  */
2303                 spin_unlock(&delayed_refs->lock);
2304
2305                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2306                                           must_insert_reserved);
2307                 BUG_ON(ret);
2308
2309                 btrfs_put_delayed_ref(ref);
2310                 kfree(extent_op);
2311                 count++;
2312 next:
2313                 do_chunk_alloc(trans, root->fs_info->extent_root,
2314                                2 * 1024 * 1024,
2315                                btrfs_get_alloc_profile(root, 0),
2316                                CHUNK_ALLOC_NO_FORCE);
2317                 cond_resched();
2318                 spin_lock(&delayed_refs->lock);
2319         }
2320         return count;
2321 }
2322
2323
2324 static void wait_for_more_refs(struct btrfs_delayed_ref_root *delayed_refs,
2325                         unsigned long num_refs)
2326 {
2327         struct list_head *first_seq = delayed_refs->seq_head.next;
2328
2329         spin_unlock(&delayed_refs->lock);
2330         pr_debug("waiting for more refs (num %ld, first %p)\n",
2331                  num_refs, first_seq);
2332         wait_event(delayed_refs->seq_wait,
2333                    num_refs != delayed_refs->num_entries ||
2334                    delayed_refs->seq_head.next != first_seq);
2335         pr_debug("done waiting for more refs (num %ld, first %p)\n",
2336                  delayed_refs->num_entries, delayed_refs->seq_head.next);
2337         spin_lock(&delayed_refs->lock);
2338 }
2339
2340 /*
2341  * this starts processing the delayed reference count updates and
2342  * extent insertions we have queued up so far.  count can be
2343  * 0, which means to process everything in the tree at the start
2344  * of the run (but not newly added entries), or it can be some target
2345  * number you'd like to process.
2346  */
2347 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2348                            struct btrfs_root *root, unsigned long count)
2349 {
2350         struct rb_node *node;
2351         struct btrfs_delayed_ref_root *delayed_refs;
2352         struct btrfs_delayed_ref_node *ref;
2353         struct list_head cluster;
2354         int ret;
2355         u64 delayed_start;
2356         int run_all = count == (unsigned long)-1;
2357         int run_most = 0;
2358         unsigned long num_refs = 0;
2359         int consider_waiting;
2360
2361         if (root == root->fs_info->extent_root)
2362                 root = root->fs_info->tree_root;
2363
2364         do_chunk_alloc(trans, root->fs_info->extent_root,
2365                        2 * 1024 * 1024, btrfs_get_alloc_profile(root, 0),
2366                        CHUNK_ALLOC_NO_FORCE);
2367
2368         delayed_refs = &trans->transaction->delayed_refs;
2369         INIT_LIST_HEAD(&cluster);
2370 again:
2371         consider_waiting = 0;
2372         spin_lock(&delayed_refs->lock);
2373         if (count == 0) {
2374                 count = delayed_refs->num_entries * 2;
2375                 run_most = 1;
2376         }
2377         while (1) {
2378                 if (!(run_all || run_most) &&
2379                     delayed_refs->num_heads_ready < 64)
2380                         break;
2381
2382                 /*
2383                  * go find something we can process in the rbtree.  We start at
2384                  * the beginning of the tree, and then build a cluster
2385                  * of refs to process starting at the first one we are able to
2386                  * lock
2387                  */
2388                 delayed_start = delayed_refs->run_delayed_start;
2389                 ret = btrfs_find_ref_cluster(trans, &cluster,
2390                                              delayed_refs->run_delayed_start);
2391                 if (ret)
2392                         break;
2393
2394                 if (delayed_start >= delayed_refs->run_delayed_start) {
2395                         if (consider_waiting == 0) {
2396                                 /*
2397                                  * btrfs_find_ref_cluster looped. let's do one
2398                                  * more cycle. if we don't run any delayed ref
2399                                  * during that cycle (because we can't because
2400                                  * all of them are blocked) and if the number of
2401                                  * refs doesn't change, we avoid busy waiting.
2402                                  */
2403                                 consider_waiting = 1;
2404                                 num_refs = delayed_refs->num_entries;
2405                         } else {
2406                                 wait_for_more_refs(delayed_refs, num_refs);
2407                                 /*
2408                                  * after waiting, things have changed. we
2409                                  * dropped the lock and someone else might have
2410                                  * run some refs, built new clusters and so on.
2411                                  * therefore, we restart staleness detection.
2412                                  */
2413                                 consider_waiting = 0;
2414                         }
2415                 }
2416
2417                 ret = run_clustered_refs(trans, root, &cluster);
2418                 BUG_ON(ret < 0);
2419
2420                 count -= min_t(unsigned long, ret, count);
2421
2422                 if (count == 0)
2423                         break;
2424
2425                 if (ret || delayed_refs->run_delayed_start == 0) {
2426                         /* refs were run, let's reset staleness detection */
2427                         consider_waiting = 0;
2428                 }
2429         }
2430
2431         if (run_all) {
2432                 node = rb_first(&delayed_refs->root);
2433                 if (!node)
2434                         goto out;
2435                 count = (unsigned long)-1;
2436
2437                 while (node) {
2438                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2439                                        rb_node);
2440                         if (btrfs_delayed_ref_is_head(ref)) {
2441                                 struct btrfs_delayed_ref_head *head;
2442
2443                                 head = btrfs_delayed_node_to_head(ref);
2444                                 atomic_inc(&ref->refs);
2445
2446                                 spin_unlock(&delayed_refs->lock);
2447                                 /*
2448                                  * Mutex was contended, block until it's
2449                                  * released and try again
2450                                  */
2451                                 mutex_lock(&head->mutex);
2452                                 mutex_unlock(&head->mutex);
2453
2454                                 btrfs_put_delayed_ref(ref);
2455                                 cond_resched();
2456                                 goto again;
2457                         }
2458                         node = rb_next(node);
2459                 }
2460                 spin_unlock(&delayed_refs->lock);
2461                 schedule_timeout(1);
2462                 goto again;
2463         }
2464 out:
2465         spin_unlock(&delayed_refs->lock);
2466         return 0;
2467 }
2468
2469 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2470                                 struct btrfs_root *root,
2471                                 u64 bytenr, u64 num_bytes, u64 flags,
2472                                 int is_data)
2473 {
2474         struct btrfs_delayed_extent_op *extent_op;
2475         int ret;
2476
2477         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2478         if (!extent_op)
2479                 return -ENOMEM;
2480
2481         extent_op->flags_to_set = flags;
2482         extent_op->update_flags = 1;
2483         extent_op->update_key = 0;
2484         extent_op->is_data = is_data ? 1 : 0;
2485
2486         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2487                                           num_bytes, extent_op);
2488         if (ret)
2489                 kfree(extent_op);
2490         return ret;
2491 }
2492
2493 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2494                                       struct btrfs_root *root,
2495                                       struct btrfs_path *path,
2496                                       u64 objectid, u64 offset, u64 bytenr)
2497 {
2498         struct btrfs_delayed_ref_head *head;
2499         struct btrfs_delayed_ref_node *ref;
2500         struct btrfs_delayed_data_ref *data_ref;
2501         struct btrfs_delayed_ref_root *delayed_refs;
2502         struct rb_node *node;
2503         int ret = 0;
2504
2505         ret = -ENOENT;
2506         delayed_refs = &trans->transaction->delayed_refs;
2507         spin_lock(&delayed_refs->lock);
2508         head = btrfs_find_delayed_ref_head(trans, bytenr);
2509         if (!head)
2510                 goto out;
2511
2512         if (!mutex_trylock(&head->mutex)) {
2513                 atomic_inc(&head->node.refs);
2514                 spin_unlock(&delayed_refs->lock);
2515
2516                 btrfs_release_path(path);
2517
2518                 /*
2519                  * Mutex was contended, block until it's released and let
2520                  * caller try again
2521                  */
2522                 mutex_lock(&head->mutex);
2523                 mutex_unlock(&head->mutex);
2524                 btrfs_put_delayed_ref(&head->node);
2525                 return -EAGAIN;
2526         }
2527
2528         node = rb_prev(&head->node.rb_node);
2529         if (!node)
2530                 goto out_unlock;
2531
2532         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2533
2534         if (ref->bytenr != bytenr)
2535                 goto out_unlock;
2536
2537         ret = 1;
2538         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2539                 goto out_unlock;
2540
2541         data_ref = btrfs_delayed_node_to_data_ref(ref);
2542
2543         node = rb_prev(node);
2544         if (node) {
2545                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2546                 if (ref->bytenr == bytenr)
2547                         goto out_unlock;
2548         }
2549
2550         if (data_ref->root != root->root_key.objectid ||
2551             data_ref->objectid != objectid || data_ref->offset != offset)
2552                 goto out_unlock;
2553
2554         ret = 0;
2555 out_unlock:
2556         mutex_unlock(&head->mutex);
2557 out:
2558         spin_unlock(&delayed_refs->lock);
2559         return ret;
2560 }
2561
2562 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2563                                         struct btrfs_root *root,
2564                                         struct btrfs_path *path,
2565                                         u64 objectid, u64 offset, u64 bytenr)
2566 {
2567         struct btrfs_root *extent_root = root->fs_info->extent_root;
2568         struct extent_buffer *leaf;
2569         struct btrfs_extent_data_ref *ref;
2570         struct btrfs_extent_inline_ref *iref;
2571         struct btrfs_extent_item *ei;
2572         struct btrfs_key key;
2573         u32 item_size;
2574         int ret;
2575
2576         key.objectid = bytenr;
2577         key.offset = (u64)-1;
2578         key.type = BTRFS_EXTENT_ITEM_KEY;
2579
2580         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2581         if (ret < 0)
2582                 goto out;
2583         BUG_ON(ret == 0);
2584
2585         ret = -ENOENT;
2586         if (path->slots[0] == 0)
2587                 goto out;
2588
2589         path->slots[0]--;
2590         leaf = path->nodes[0];
2591         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2592
2593         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2594                 goto out;
2595
2596         ret = 1;
2597         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2598 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2599         if (item_size < sizeof(*ei)) {
2600                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2601                 goto out;
2602         }
2603 #endif
2604         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2605
2606         if (item_size != sizeof(*ei) +
2607             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2608                 goto out;
2609
2610         if (btrfs_extent_generation(leaf, ei) <=
2611             btrfs_root_last_snapshot(&root->root_item))
2612                 goto out;
2613
2614         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2615         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2616             BTRFS_EXTENT_DATA_REF_KEY)
2617                 goto out;
2618
2619         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2620         if (btrfs_extent_refs(leaf, ei) !=
2621             btrfs_extent_data_ref_count(leaf, ref) ||
2622             btrfs_extent_data_ref_root(leaf, ref) !=
2623             root->root_key.objectid ||
2624             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2625             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2626                 goto out;
2627
2628         ret = 0;
2629 out:
2630         return ret;
2631 }
2632
2633 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2634                           struct btrfs_root *root,
2635                           u64 objectid, u64 offset, u64 bytenr)
2636 {
2637         struct btrfs_path *path;
2638         int ret;
2639         int ret2;
2640
2641         path = btrfs_alloc_path();
2642         if (!path)
2643                 return -ENOENT;
2644
2645         do {
2646                 ret = check_committed_ref(trans, root, path, objectid,
2647                                           offset, bytenr);
2648                 if (ret && ret != -ENOENT)
2649                         goto out;
2650
2651                 ret2 = check_delayed_ref(trans, root, path, objectid,
2652                                          offset, bytenr);
2653         } while (ret2 == -EAGAIN);
2654
2655         if (ret2 && ret2 != -ENOENT) {
2656                 ret = ret2;
2657                 goto out;
2658         }
2659
2660         if (ret != -ENOENT || ret2 != -ENOENT)
2661                 ret = 0;
2662 out:
2663         btrfs_free_path(path);
2664         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2665                 WARN_ON(ret > 0);
2666         return ret;
2667 }
2668
2669 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2670                            struct btrfs_root *root,
2671                            struct extent_buffer *buf,
2672                            int full_backref, int inc, int for_cow)
2673 {
2674         u64 bytenr;
2675         u64 num_bytes;
2676         u64 parent;
2677         u64 ref_root;
2678         u32 nritems;
2679         struct btrfs_key key;
2680         struct btrfs_file_extent_item *fi;
2681         int i;
2682         int level;
2683         int ret = 0;
2684         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2685                             u64, u64, u64, u64, u64, u64, int);
2686
2687         ref_root = btrfs_header_owner(buf);
2688         nritems = btrfs_header_nritems(buf);
2689         level = btrfs_header_level(buf);
2690
2691         if (!root->ref_cows && level == 0)
2692                 return 0;
2693
2694         if (inc)
2695                 process_func = btrfs_inc_extent_ref;
2696         else
2697                 process_func = btrfs_free_extent;
2698
2699         if (full_backref)
2700                 parent = buf->start;
2701         else
2702                 parent = 0;
2703
2704         for (i = 0; i < nritems; i++) {
2705                 if (level == 0) {
2706                         btrfs_item_key_to_cpu(buf, &key, i);
2707                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2708                                 continue;
2709                         fi = btrfs_item_ptr(buf, i,
2710                                             struct btrfs_file_extent_item);
2711                         if (btrfs_file_extent_type(buf, fi) ==
2712                             BTRFS_FILE_EXTENT_INLINE)
2713                                 continue;
2714                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2715                         if (bytenr == 0)
2716                                 continue;
2717
2718                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2719                         key.offset -= btrfs_file_extent_offset(buf, fi);
2720                         ret = process_func(trans, root, bytenr, num_bytes,
2721                                            parent, ref_root, key.objectid,
2722                                            key.offset, for_cow);
2723                         if (ret)
2724                                 goto fail;
2725                 } else {
2726                         bytenr = btrfs_node_blockptr(buf, i);
2727                         num_bytes = btrfs_level_size(root, level - 1);
2728                         ret = process_func(trans, root, bytenr, num_bytes,
2729                                            parent, ref_root, level - 1, 0,
2730                                            for_cow);
2731                         if (ret)
2732                                 goto fail;
2733                 }
2734         }
2735         return 0;
2736 fail:
2737         BUG();
2738         return ret;
2739 }
2740
2741 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2742                   struct extent_buffer *buf, int full_backref, int for_cow)
2743 {
2744         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2745 }
2746
2747 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2748                   struct extent_buffer *buf, int full_backref, int for_cow)
2749 {
2750         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2751 }
2752
2753 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2754                                  struct btrfs_root *root,
2755                                  struct btrfs_path *path,
2756                                  struct btrfs_block_group_cache *cache)
2757 {
2758         int ret;
2759         struct btrfs_root *extent_root = root->fs_info->extent_root;
2760         unsigned long bi;
2761         struct extent_buffer *leaf;
2762
2763         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2764         if (ret < 0)
2765                 goto fail;
2766         BUG_ON(ret);
2767
2768         leaf = path->nodes[0];
2769         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2770         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2771         btrfs_mark_buffer_dirty(leaf);
2772         btrfs_release_path(path);
2773 fail:
2774         if (ret)
2775                 return ret;
2776         return 0;
2777
2778 }
2779
2780 static struct btrfs_block_group_cache *
2781 next_block_group(struct btrfs_root *root,
2782                  struct btrfs_block_group_cache *cache)
2783 {
2784         struct rb_node *node;
2785         spin_lock(&root->fs_info->block_group_cache_lock);
2786         node = rb_next(&cache->cache_node);
2787         btrfs_put_block_group(cache);
2788         if (node) {
2789                 cache = rb_entry(node, struct btrfs_block_group_cache,
2790                                  cache_node);
2791                 btrfs_get_block_group(cache);
2792         } else
2793                 cache = NULL;
2794         spin_unlock(&root->fs_info->block_group_cache_lock);
2795         return cache;
2796 }
2797
2798 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2799                             struct btrfs_trans_handle *trans,
2800                             struct btrfs_path *path)
2801 {
2802         struct btrfs_root *root = block_group->fs_info->tree_root;
2803         struct inode *inode = NULL;
2804         u64 alloc_hint = 0;
2805         int dcs = BTRFS_DC_ERROR;
2806         int num_pages = 0;
2807         int retries = 0;
2808         int ret = 0;
2809
2810         /*
2811          * If this block group is smaller than 100 megs don't bother caching the
2812          * block group.
2813          */
2814         if (block_group->key.offset < (100 * 1024 * 1024)) {
2815                 spin_lock(&block_group->lock);
2816                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2817                 spin_unlock(&block_group->lock);
2818                 return 0;
2819         }
2820
2821 again:
2822         inode = lookup_free_space_inode(root, block_group, path);
2823         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2824                 ret = PTR_ERR(inode);
2825                 btrfs_release_path(path);
2826                 goto out;
2827         }
2828
2829         if (IS_ERR(inode)) {
2830                 BUG_ON(retries);
2831                 retries++;
2832
2833                 if (block_group->ro)
2834                         goto out_free;
2835
2836                 ret = create_free_space_inode(root, trans, block_group, path);
2837                 if (ret)
2838                         goto out_free;
2839                 goto again;
2840         }
2841
2842         /* We've already setup this transaction, go ahead and exit */
2843         if (block_group->cache_generation == trans->transid &&
2844             i_size_read(inode)) {
2845                 dcs = BTRFS_DC_SETUP;
2846                 goto out_put;
2847         }
2848
2849         /*
2850          * We want to set the generation to 0, that way if anything goes wrong
2851          * from here on out we know not to trust this cache when we load up next
2852          * time.
2853          */
2854         BTRFS_I(inode)->generation = 0;
2855         ret = btrfs_update_inode(trans, root, inode);
2856         WARN_ON(ret);
2857
2858         if (i_size_read(inode) > 0) {
2859                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2860                                                       inode);
2861                 if (ret)
2862                         goto out_put;
2863         }
2864
2865         spin_lock(&block_group->lock);
2866         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2867                 /* We're not cached, don't bother trying to write stuff out */
2868                 dcs = BTRFS_DC_WRITTEN;
2869                 spin_unlock(&block_group->lock);
2870                 goto out_put;
2871         }
2872         spin_unlock(&block_group->lock);
2873
2874         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2875         if (!num_pages)
2876                 num_pages = 1;
2877
2878         /*
2879          * Just to make absolutely sure we have enough space, we're going to
2880          * preallocate 12 pages worth of space for each block group.  In
2881          * practice we ought to use at most 8, but we need extra space so we can
2882          * add our header and have a terminator between the extents and the
2883          * bitmaps.
2884          */
2885         num_pages *= 16;
2886         num_pages *= PAGE_CACHE_SIZE;
2887
2888         ret = btrfs_check_data_free_space(inode, num_pages);
2889         if (ret)
2890                 goto out_put;
2891
2892         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2893                                               num_pages, num_pages,
2894                                               &alloc_hint);
2895         if (!ret)
2896                 dcs = BTRFS_DC_SETUP;
2897         btrfs_free_reserved_data_space(inode, num_pages);
2898
2899 out_put:
2900         iput(inode);
2901 out_free:
2902         btrfs_release_path(path);
2903 out:
2904         spin_lock(&block_group->lock);
2905         if (!ret && dcs == BTRFS_DC_SETUP)
2906                 block_group->cache_generation = trans->transid;
2907         block_group->disk_cache_state = dcs;
2908         spin_unlock(&block_group->lock);
2909
2910         return ret;
2911 }
2912
2913 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2914                                    struct btrfs_root *root)
2915 {
2916         struct btrfs_block_group_cache *cache;
2917         int err = 0;
2918         struct btrfs_path *path;
2919         u64 last = 0;
2920
2921         path = btrfs_alloc_path();
2922         if (!path)
2923                 return -ENOMEM;
2924
2925 again:
2926         while (1) {
2927                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2928                 while (cache) {
2929                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2930                                 break;
2931                         cache = next_block_group(root, cache);
2932                 }
2933                 if (!cache) {
2934                         if (last == 0)
2935                                 break;
2936                         last = 0;
2937                         continue;
2938                 }
2939                 err = cache_save_setup(cache, trans, path);
2940                 last = cache->key.objectid + cache->key.offset;
2941                 btrfs_put_block_group(cache);
2942         }
2943
2944         while (1) {
2945                 if (last == 0) {
2946                         err = btrfs_run_delayed_refs(trans, root,
2947                                                      (unsigned long)-1);
2948                         BUG_ON(err);
2949                 }
2950
2951                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2952                 while (cache) {
2953                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2954                                 btrfs_put_block_group(cache);
2955                                 goto again;
2956                         }
2957
2958                         if (cache->dirty)
2959                                 break;
2960                         cache = next_block_group(root, cache);
2961                 }
2962                 if (!cache) {
2963                         if (last == 0)
2964                                 break;
2965                         last = 0;
2966                         continue;
2967                 }
2968
2969                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
2970                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
2971                 cache->dirty = 0;
2972                 last = cache->key.objectid + cache->key.offset;
2973
2974                 err = write_one_cache_group(trans, root, path, cache);
2975                 BUG_ON(err);
2976                 btrfs_put_block_group(cache);
2977         }
2978
2979         while (1) {
2980                 /*
2981                  * I don't think this is needed since we're just marking our
2982                  * preallocated extent as written, but just in case it can't
2983                  * hurt.
2984                  */
2985                 if (last == 0) {
2986                         err = btrfs_run_delayed_refs(trans, root,
2987                                                      (unsigned long)-1);
2988                         BUG_ON(err);
2989                 }
2990
2991                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2992                 while (cache) {
2993                         /*
2994                          * Really this shouldn't happen, but it could if we
2995                          * couldn't write the entire preallocated extent and
2996                          * splitting the extent resulted in a new block.
2997                          */
2998                         if (cache->dirty) {
2999                                 btrfs_put_block_group(cache);
3000                                 goto again;
3001                         }
3002                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3003                                 break;
3004                         cache = next_block_group(root, cache);
3005                 }
3006                 if (!cache) {
3007                         if (last == 0)
3008                                 break;
3009                         last = 0;
3010                         continue;
3011                 }
3012
3013                 btrfs_write_out_cache(root, trans, cache, path);
3014
3015                 /*
3016                  * If we didn't have an error then the cache state is still
3017                  * NEED_WRITE, so we can set it to WRITTEN.
3018                  */
3019                 if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3020                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3021                 last = cache->key.objectid + cache->key.offset;
3022                 btrfs_put_block_group(cache);
3023         }
3024
3025         btrfs_free_path(path);
3026         return 0;
3027 }
3028
3029 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3030 {
3031         struct btrfs_block_group_cache *block_group;
3032         int readonly = 0;
3033
3034         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3035         if (!block_group || block_group->ro)
3036                 readonly = 1;
3037         if (block_group)
3038                 btrfs_put_block_group(block_group);
3039         return readonly;
3040 }
3041
3042 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3043                              u64 total_bytes, u64 bytes_used,
3044                              struct btrfs_space_info **space_info)
3045 {
3046         struct btrfs_space_info *found;
3047         int i;
3048         int factor;
3049
3050         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3051                      BTRFS_BLOCK_GROUP_RAID10))
3052                 factor = 2;
3053         else
3054                 factor = 1;
3055
3056         found = __find_space_info(info, flags);
3057         if (found) {
3058                 spin_lock(&found->lock);
3059                 found->total_bytes += total_bytes;
3060                 found->disk_total += total_bytes * factor;
3061                 found->bytes_used += bytes_used;
3062                 found->disk_used += bytes_used * factor;
3063                 found->full = 0;
3064                 spin_unlock(&found->lock);
3065                 *space_info = found;
3066                 return 0;
3067         }
3068         found = kzalloc(sizeof(*found), GFP_NOFS);
3069         if (!found)
3070                 return -ENOMEM;
3071
3072         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3073                 INIT_LIST_HEAD(&found->block_groups[i]);
3074         init_rwsem(&found->groups_sem);
3075         spin_lock_init(&found->lock);
3076         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3077         found->total_bytes = total_bytes;
3078         found->disk_total = total_bytes * factor;
3079         found->bytes_used = bytes_used;
3080         found->disk_used = bytes_used * factor;
3081         found->bytes_pinned = 0;
3082         found->bytes_reserved = 0;
3083         found->bytes_readonly = 0;
3084         found->bytes_may_use = 0;
3085         found->full = 0;
3086         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3087         found->chunk_alloc = 0;
3088         found->flush = 0;
3089         init_waitqueue_head(&found->wait);
3090         *space_info = found;
3091         list_add_rcu(&found->list, &info->space_info);
3092         return 0;
3093 }
3094
3095 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3096 {
3097         u64 extra_flags = flags & BTRFS_BLOCK_GROUP_PROFILE_MASK;
3098
3099         /* chunk -> extended profile */
3100         if (extra_flags == 0)
3101                 extra_flags = BTRFS_AVAIL_ALLOC_BIT_SINGLE;
3102
3103         if (flags & BTRFS_BLOCK_GROUP_DATA)
3104                 fs_info->avail_data_alloc_bits |= extra_flags;
3105         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3106                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3107         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3108                 fs_info->avail_system_alloc_bits |= extra_flags;
3109 }
3110
3111 /*
3112  * @flags: available profiles in extended format (see ctree.h)
3113  *
3114  * Returns reduced profile in chunk format.  If profile changing is in
3115  * progress (either running or paused) picks the target profile (if it's
3116  * already available), otherwise falls back to plain reducing.
3117  */
3118 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3119 {
3120         /*
3121          * we add in the count of missing devices because we want
3122          * to make sure that any RAID levels on a degraded FS
3123          * continue to be honored.
3124          */
3125         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3126                 root->fs_info->fs_devices->missing_devices;
3127
3128         /* pick restriper's target profile if it's available */
3129         spin_lock(&root->fs_info->balance_lock);
3130         if (root->fs_info->balance_ctl) {
3131                 struct btrfs_balance_control *bctl = root->fs_info->balance_ctl;
3132                 u64 tgt = 0;
3133
3134                 if ((flags & BTRFS_BLOCK_GROUP_DATA) &&
3135                     (bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) &&
3136                     (flags & bctl->data.target)) {
3137                         tgt = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3138                 } else if ((flags & BTRFS_BLOCK_GROUP_SYSTEM) &&
3139                            (bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) &&
3140                            (flags & bctl->sys.target)) {
3141                         tgt = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3142                 } else if ((flags & BTRFS_BLOCK_GROUP_METADATA) &&
3143                            (bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) &&
3144                            (flags & bctl->meta.target)) {
3145                         tgt = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3146                 }
3147
3148                 if (tgt) {
3149                         spin_unlock(&root->fs_info->balance_lock);
3150                         flags = tgt;
3151                         goto out;
3152                 }
3153         }
3154         spin_unlock(&root->fs_info->balance_lock);
3155
3156         if (num_devices == 1)
3157                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3158         if (num_devices < 4)
3159                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3160
3161         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3162             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3163                       BTRFS_BLOCK_GROUP_RAID10))) {
3164                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3165         }
3166
3167         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3168             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3169                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3170         }
3171
3172         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3173             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3174              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3175              (flags & BTRFS_BLOCK_GROUP_DUP))) {
3176                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3177         }
3178
3179 out:
3180         /* extended -> chunk profile */
3181         flags &= ~BTRFS_AVAIL_ALLOC_BIT_SINGLE;
3182         return flags;
3183 }
3184
3185 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3186 {
3187         if (flags & BTRFS_BLOCK_GROUP_DATA)
3188                 flags |= root->fs_info->avail_data_alloc_bits;
3189         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3190                 flags |= root->fs_info->avail_system_alloc_bits;
3191         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3192                 flags |= root->fs_info->avail_metadata_alloc_bits;
3193
3194         return btrfs_reduce_alloc_profile(root, flags);
3195 }
3196
3197 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3198 {
3199         u64 flags;
3200
3201         if (data)
3202                 flags = BTRFS_BLOCK_GROUP_DATA;
3203         else if (root == root->fs_info->chunk_root)
3204                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3205         else
3206                 flags = BTRFS_BLOCK_GROUP_METADATA;
3207
3208         return get_alloc_profile(root, flags);
3209 }
3210
3211 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3212 {
3213         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3214                                                        BTRFS_BLOCK_GROUP_DATA);
3215 }
3216
3217 /*
3218  * This will check the space that the inode allocates from to make sure we have
3219  * enough space for bytes.
3220  */
3221 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3222 {
3223         struct btrfs_space_info *data_sinfo;
3224         struct btrfs_root *root = BTRFS_I(inode)->root;
3225         u64 used;
3226         int ret = 0, committed = 0, alloc_chunk = 1;
3227
3228         /* make sure bytes are sectorsize aligned */
3229         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3230
3231         if (root == root->fs_info->tree_root ||
3232             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3233                 alloc_chunk = 0;
3234                 committed = 1;
3235         }
3236
3237         data_sinfo = BTRFS_I(inode)->space_info;
3238         if (!data_sinfo)
3239                 goto alloc;
3240
3241 again:
3242         /* make sure we have enough space to handle the data first */
3243         spin_lock(&data_sinfo->lock);
3244         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3245                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3246                 data_sinfo->bytes_may_use;
3247
3248         if (used + bytes > data_sinfo->total_bytes) {
3249                 struct btrfs_trans_handle *trans;
3250
3251                 /*
3252                  * if we don't have enough free bytes in this space then we need
3253                  * to alloc a new chunk.
3254                  */
3255                 if (!data_sinfo->full && alloc_chunk) {
3256                         u64 alloc_target;
3257
3258                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3259                         spin_unlock(&data_sinfo->lock);
3260 alloc:
3261                         alloc_target = btrfs_get_alloc_profile(root, 1);
3262                         trans = btrfs_join_transaction(root);
3263                         if (IS_ERR(trans))
3264                                 return PTR_ERR(trans);
3265
3266                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3267                                              bytes + 2 * 1024 * 1024,
3268                                              alloc_target,
3269                                              CHUNK_ALLOC_NO_FORCE);
3270                         btrfs_end_transaction(trans, root);
3271                         if (ret < 0) {
3272                                 if (ret != -ENOSPC)
3273                                         return ret;
3274                                 else
3275                                         goto commit_trans;
3276                         }
3277
3278                         if (!data_sinfo) {
3279                                 btrfs_set_inode_space_info(root, inode);
3280                                 data_sinfo = BTRFS_I(inode)->space_info;
3281                         }
3282                         goto again;
3283                 }
3284
3285                 /*
3286                  * If we have less pinned bytes than we want to allocate then
3287                  * don't bother committing the transaction, it won't help us.
3288                  */
3289                 if (data_sinfo->bytes_pinned < bytes)
3290                         committed = 1;
3291                 spin_unlock(&data_sinfo->lock);
3292
3293                 /* commit the current transaction and try again */
3294 commit_trans:
3295                 if (!committed &&
3296                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3297                         committed = 1;
3298                         trans = btrfs_join_transaction(root);
3299                         if (IS_ERR(trans))
3300                                 return PTR_ERR(trans);
3301                         ret = btrfs_commit_transaction(trans, root);
3302                         if (ret)
3303                                 return ret;
3304                         goto again;
3305                 }
3306
3307                 return -ENOSPC;
3308         }
3309         data_sinfo->bytes_may_use += bytes;
3310         trace_btrfs_space_reservation(root->fs_info, "space_info",
3311                                       (u64)(unsigned long)data_sinfo,
3312                                       bytes, 1);
3313         spin_unlock(&data_sinfo->lock);
3314
3315         return 0;
3316 }
3317
3318 /*
3319  * Called if we need to clear a data reservation for this inode.
3320  */
3321 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3322 {
3323         struct btrfs_root *root = BTRFS_I(inode)->root;
3324         struct btrfs_space_info *data_sinfo;
3325
3326         /* make sure bytes are sectorsize aligned */
3327         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3328
3329         data_sinfo = BTRFS_I(inode)->space_info;
3330         spin_lock(&data_sinfo->lock);
3331         data_sinfo->bytes_may_use -= bytes;
3332         trace_btrfs_space_reservation(root->fs_info, "space_info",
3333                                       (u64)(unsigned long)data_sinfo,
3334                                       bytes, 0);
3335         spin_unlock(&data_sinfo->lock);
3336 }
3337
3338 static void force_metadata_allocation(struct btrfs_fs_info *info)
3339 {
3340         struct list_head *head = &info->space_info;
3341         struct btrfs_space_info *found;
3342
3343         rcu_read_lock();
3344         list_for_each_entry_rcu(found, head, list) {
3345                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3346                         found->force_alloc = CHUNK_ALLOC_FORCE;
3347         }
3348         rcu_read_unlock();
3349 }
3350
3351 static int should_alloc_chunk(struct btrfs_root *root,
3352                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3353                               int force)
3354 {
3355         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3356         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3357         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3358         u64 thresh;
3359
3360         if (force == CHUNK_ALLOC_FORCE)
3361                 return 1;
3362
3363         /*
3364          * We need to take into account the global rsv because for all intents
3365          * and purposes it's used space.  Don't worry about locking the
3366          * global_rsv, it doesn't change except when the transaction commits.
3367          */
3368         num_allocated += global_rsv->size;
3369
3370         /*
3371          * in limited mode, we want to have some free space up to
3372          * about 1% of the FS size.
3373          */
3374         if (force == CHUNK_ALLOC_LIMITED) {
3375                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3376                 thresh = max_t(u64, 64 * 1024 * 1024,
3377                                div_factor_fine(thresh, 1));
3378
3379                 if (num_bytes - num_allocated < thresh)
3380                         return 1;
3381         }
3382         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3383
3384         /* 256MB or 2% of the FS */
3385         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 2));
3386         /* system chunks need a much small threshold */
3387         if (sinfo->flags & BTRFS_BLOCK_GROUP_SYSTEM)
3388                 thresh = 32 * 1024 * 1024;
3389
3390         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 8))
3391                 return 0;
3392         return 1;
3393 }
3394
3395 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3396                           struct btrfs_root *extent_root, u64 alloc_bytes,
3397                           u64 flags, int force)
3398 {
3399         struct btrfs_space_info *space_info;
3400         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3401         int wait_for_alloc = 0;
3402         int ret = 0;
3403
3404         BUG_ON(!profile_is_valid(flags, 0));
3405
3406         space_info = __find_space_info(extent_root->fs_info, flags);
3407         if (!space_info) {
3408                 ret = update_space_info(extent_root->fs_info, flags,
3409                                         0, 0, &space_info);
3410                 BUG_ON(ret);
3411         }
3412         BUG_ON(!space_info);
3413
3414 again:
3415         spin_lock(&space_info->lock);
3416         if (force < space_info->force_alloc)
3417                 force = space_info->force_alloc;
3418         if (space_info->full) {
3419                 spin_unlock(&space_info->lock);
3420                 return 0;
3421         }
3422
3423         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3424                 spin_unlock(&space_info->lock);
3425                 return 0;
3426         } else if (space_info->chunk_alloc) {
3427                 wait_for_alloc = 1;
3428         } else {
3429                 space_info->chunk_alloc = 1;
3430         }
3431
3432         spin_unlock(&space_info->lock);
3433
3434         mutex_lock(&fs_info->chunk_mutex);
3435
3436         /*
3437          * The chunk_mutex is held throughout the entirety of a chunk
3438          * allocation, so once we've acquired the chunk_mutex we know that the
3439          * other guy is done and we need to recheck and see if we should
3440          * allocate.
3441          */
3442         if (wait_for_alloc) {
3443                 mutex_unlock(&fs_info->chunk_mutex);
3444                 wait_for_alloc = 0;
3445                 goto again;
3446         }
3447
3448         /*
3449          * If we have mixed data/metadata chunks we want to make sure we keep
3450          * allocating mixed chunks instead of individual chunks.
3451          */
3452         if (btrfs_mixed_space_info(space_info))
3453                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3454
3455         /*
3456          * if we're doing a data chunk, go ahead and make sure that
3457          * we keep a reasonable number of metadata chunks allocated in the
3458          * FS as well.
3459          */
3460         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3461                 fs_info->data_chunk_allocations++;
3462                 if (!(fs_info->data_chunk_allocations %
3463                       fs_info->metadata_ratio))
3464                         force_metadata_allocation(fs_info);
3465         }
3466
3467         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3468         if (ret < 0 && ret != -ENOSPC)
3469                 goto out;
3470
3471         spin_lock(&space_info->lock);
3472         if (ret)
3473                 space_info->full = 1;
3474         else
3475                 ret = 1;
3476
3477         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3478         space_info->chunk_alloc = 0;
3479         spin_unlock(&space_info->lock);
3480 out:
3481         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3482         return ret;
3483 }
3484
3485 /*
3486  * shrink metadata reservation for delalloc
3487  */
3488 static int shrink_delalloc(struct btrfs_root *root, u64 to_reclaim,
3489                            bool wait_ordered)
3490 {
3491         struct btrfs_block_rsv *block_rsv;
3492         struct btrfs_space_info *space_info;
3493         struct btrfs_trans_handle *trans;
3494         u64 reserved;
3495         u64 max_reclaim;
3496         u64 reclaimed = 0;
3497         long time_left;
3498         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3499         int loops = 0;
3500         unsigned long progress;
3501
3502         trans = (struct btrfs_trans_handle *)current->journal_info;
3503         block_rsv = &root->fs_info->delalloc_block_rsv;
3504         space_info = block_rsv->space_info;
3505
3506         smp_mb();
3507         reserved = space_info->bytes_may_use;
3508         progress = space_info->reservation_progress;
3509
3510         if (reserved == 0)
3511                 return 0;
3512
3513         smp_mb();
3514         if (root->fs_info->delalloc_bytes == 0) {
3515                 if (trans)
3516                         return 0;
3517                 btrfs_wait_ordered_extents(root, 0, 0);
3518                 return 0;
3519         }
3520
3521         max_reclaim = min(reserved, to_reclaim);
3522         nr_pages = max_t(unsigned long, nr_pages,
3523                          max_reclaim >> PAGE_CACHE_SHIFT);
3524         while (loops < 1024) {
3525                 /* have the flusher threads jump in and do some IO */
3526                 smp_mb();
3527                 nr_pages = min_t(unsigned long, nr_pages,
3528                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3529                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3530                                                 WB_REASON_FS_FREE_SPACE);
3531
3532                 spin_lock(&space_info->lock);
3533                 if (reserved > space_info->bytes_may_use)
3534                         reclaimed += reserved - space_info->bytes_may_use;
3535                 reserved = space_info->bytes_may_use;
3536                 spin_unlock(&space_info->lock);
3537
3538                 loops++;
3539
3540                 if (reserved == 0 || reclaimed >= max_reclaim)
3541                         break;
3542
3543                 if (trans && trans->transaction->blocked)
3544                         return -EAGAIN;
3545
3546                 if (wait_ordered && !trans) {
3547                         btrfs_wait_ordered_extents(root, 0, 0);
3548                 } else {
3549                         time_left = schedule_timeout_interruptible(1);
3550
3551                         /* We were interrupted, exit */
3552                         if (time_left)
3553                                 break;
3554                 }
3555
3556                 /* we've kicked the IO a few times, if anything has been freed,
3557                  * exit.  There is no sense in looping here for a long time
3558                  * when we really need to commit the transaction, or there are
3559                  * just too many writers without enough free space
3560                  */
3561
3562                 if (loops > 3) {
3563                         smp_mb();
3564                         if (progress != space_info->reservation_progress)
3565                                 break;
3566                 }
3567
3568         }
3569
3570         return reclaimed >= to_reclaim;
3571 }
3572
3573 /**
3574  * maybe_commit_transaction - possibly commit the transaction if its ok to
3575  * @root - the root we're allocating for
3576  * @bytes - the number of bytes we want to reserve
3577  * @force - force the commit
3578  *
3579  * This will check to make sure that committing the transaction will actually
3580  * get us somewhere and then commit the transaction if it does.  Otherwise it
3581  * will return -ENOSPC.
3582  */
3583 static int may_commit_transaction(struct btrfs_root *root,
3584                                   struct btrfs_space_info *space_info,
3585                                   u64 bytes, int force)
3586 {
3587         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3588         struct btrfs_trans_handle *trans;
3589
3590         trans = (struct btrfs_trans_handle *)current->journal_info;
3591         if (trans)
3592                 return -EAGAIN;
3593
3594         if (force)
3595                 goto commit;
3596
3597         /* See if there is enough pinned space to make this reservation */
3598         spin_lock(&space_info->lock);
3599         if (space_info->bytes_pinned >= bytes) {
3600                 spin_unlock(&space_info->lock);
3601                 goto commit;
3602         }
3603         spin_unlock(&space_info->lock);
3604
3605         /*
3606          * See if there is some space in the delayed insertion reservation for
3607          * this reservation.
3608          */
3609         if (space_info != delayed_rsv->space_info)
3610                 return -ENOSPC;
3611
3612         spin_lock(&space_info->lock);
3613         spin_lock(&delayed_rsv->lock);
3614         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3615                 spin_unlock(&delayed_rsv->lock);
3616                 spin_unlock(&space_info->lock);
3617                 return -ENOSPC;
3618         }
3619         spin_unlock(&delayed_rsv->lock);
3620         spin_unlock(&space_info->lock);
3621
3622 commit:
3623         trans = btrfs_join_transaction(root);
3624         if (IS_ERR(trans))
3625                 return -ENOSPC;
3626
3627         return btrfs_commit_transaction(trans, root);
3628 }
3629
3630 /**
3631  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3632  * @root - the root we're allocating for
3633  * @block_rsv - the block_rsv we're allocating for
3634  * @orig_bytes - the number of bytes we want
3635  * @flush - wether or not we can flush to make our reservation
3636  *
3637  * This will reserve orgi_bytes number of bytes from the space info associated
3638  * with the block_rsv.  If there is not enough space it will make an attempt to
3639  * flush out space to make room.  It will do this by flushing delalloc if
3640  * possible or committing the transaction.  If flush is 0 then no attempts to
3641  * regain reservations will be made and this will fail if there is not enough
3642  * space already.
3643  */
3644 static int reserve_metadata_bytes(struct btrfs_root *root,
3645                                   struct btrfs_block_rsv *block_rsv,
3646                                   u64 orig_bytes, int flush)
3647 {
3648         struct btrfs_space_info *space_info = block_rsv->space_info;
3649         u64 used;
3650         u64 num_bytes = orig_bytes;
3651         int retries = 0;
3652         int ret = 0;
3653         bool committed = false;
3654         bool flushing = false;
3655         bool wait_ordered = false;
3656
3657 again:
3658         ret = 0;
3659         spin_lock(&space_info->lock);
3660         /*
3661          * We only want to wait if somebody other than us is flushing and we are
3662          * actually alloed to flush.
3663          */
3664         while (flush && !flushing && space_info->flush) {
3665                 spin_unlock(&space_info->lock);
3666                 /*
3667                  * If we have a trans handle we can't wait because the flusher
3668                  * may have to commit the transaction, which would mean we would
3669                  * deadlock since we are waiting for the flusher to finish, but
3670                  * hold the current transaction open.
3671                  */
3672                 if (current->journal_info)
3673                         return -EAGAIN;
3674                 ret = wait_event_interruptible(space_info->wait,
3675                                                !space_info->flush);
3676                 /* Must have been interrupted, return */
3677                 if (ret)
3678                         return -EINTR;
3679
3680                 spin_lock(&space_info->lock);
3681         }
3682
3683         ret = -ENOSPC;
3684         used = space_info->bytes_used + space_info->bytes_reserved +
3685                 space_info->bytes_pinned + space_info->bytes_readonly +
3686                 space_info->bytes_may_use;
3687
3688         /*
3689          * The idea here is that we've not already over-reserved the block group
3690          * then we can go ahead and save our reservation first and then start
3691          * flushing if we need to.  Otherwise if we've already overcommitted
3692          * lets start flushing stuff first and then come back and try to make
3693          * our reservation.
3694          */
3695         if (used <= space_info->total_bytes) {
3696                 if (used + orig_bytes <= space_info->total_bytes) {
3697                         space_info->bytes_may_use += orig_bytes;
3698                         trace_btrfs_space_reservation(root->fs_info,
3699                                               "space_info",
3700                                               (u64)(unsigned long)space_info,
3701                                               orig_bytes, 1);
3702                         ret = 0;
3703                 } else {
3704                         /*
3705                          * Ok set num_bytes to orig_bytes since we aren't
3706                          * overocmmitted, this way we only try and reclaim what
3707                          * we need.
3708                          */
3709                         num_bytes = orig_bytes;
3710                 }
3711         } else {
3712                 /*
3713                  * Ok we're over committed, set num_bytes to the overcommitted
3714                  * amount plus the amount of bytes that we need for this
3715                  * reservation.
3716                  */
3717                 wait_ordered = true;
3718                 num_bytes = used - space_info->total_bytes +
3719                         (orig_bytes * (retries + 1));
3720         }
3721
3722         if (ret) {
3723                 u64 profile = btrfs_get_alloc_profile(root, 0);
3724                 u64 avail;
3725
3726                 /*
3727                  * If we have a lot of space that's pinned, don't bother doing
3728                  * the overcommit dance yet and just commit the transaction.
3729                  */
3730                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3731                 do_div(avail, 10);
3732                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3733                         space_info->flush = 1;
3734                         flushing = true;
3735                         spin_unlock(&space_info->lock);
3736                         ret = may_commit_transaction(root, space_info,
3737                                                      orig_bytes, 1);
3738                         if (ret)
3739                                 goto out;
3740                         committed = true;
3741                         goto again;
3742                 }
3743
3744                 spin_lock(&root->fs_info->free_chunk_lock);
3745                 avail = root->fs_info->free_chunk_space;
3746
3747                 /*
3748                  * If we have dup, raid1 or raid10 then only half of the free
3749                  * space is actually useable.
3750                  */
3751                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3752                                BTRFS_BLOCK_GROUP_RAID1 |
3753                                BTRFS_BLOCK_GROUP_RAID10))
3754                         avail >>= 1;
3755
3756                 /*
3757                  * If we aren't flushing don't let us overcommit too much, say
3758                  * 1/8th of the space.  If we can flush, let it overcommit up to
3759                  * 1/2 of the space.
3760                  */
3761                 if (flush)
3762                         avail >>= 3;
3763                 else
3764                         avail >>= 1;
3765                  spin_unlock(&root->fs_info->free_chunk_lock);
3766
3767                 if (used + num_bytes < space_info->total_bytes + avail) {
3768                         space_info->bytes_may_use += orig_bytes;
3769                         trace_btrfs_space_reservation(root->fs_info,
3770                                               "space_info",
3771                                               (u64)(unsigned long)space_info,
3772                                               orig_bytes, 1);
3773                         ret = 0;
3774                 } else {
3775                         wait_ordered = true;
3776                 }
3777         }
3778
3779         /*
3780          * Couldn't make our reservation, save our place so while we're trying
3781          * to reclaim space we can actually use it instead of somebody else
3782          * stealing it from us.
3783          */
3784         if (ret && flush) {
3785                 flushing = true;
3786                 space_info->flush = 1;
3787         }
3788
3789         spin_unlock(&space_info->lock);
3790
3791         if (!ret || !flush)
3792                 goto out;
3793
3794         /*
3795          * We do synchronous shrinking since we don't actually unreserve
3796          * metadata until after the IO is completed.
3797          */
3798         ret = shrink_delalloc(root, num_bytes, wait_ordered);
3799         if (ret < 0)
3800                 goto out;
3801
3802         ret = 0;
3803
3804         /*
3805          * So if we were overcommitted it's possible that somebody else flushed
3806          * out enough space and we simply didn't have enough space to reclaim,
3807          * so go back around and try again.
3808          */
3809         if (retries < 2) {
3810                 wait_ordered = true;
3811                 retries++;
3812                 goto again;
3813         }
3814
3815         ret = -ENOSPC;
3816         if (committed)
3817                 goto out;
3818
3819         ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3820         if (!ret) {
3821                 committed = true;
3822                 goto again;
3823         }
3824
3825 out:
3826         if (flushing) {
3827                 spin_lock(&space_info->lock);
3828                 space_info->flush = 0;
3829                 wake_up_all(&space_info->wait);
3830                 spin_unlock(&space_info->lock);
3831         }
3832         return ret;
3833 }
3834
3835 static struct btrfs_block_rsv *get_block_rsv(struct btrfs_trans_handle *trans,
3836                                              struct btrfs_root *root)
3837 {
3838         struct btrfs_block_rsv *block_rsv = NULL;
3839
3840         if (root->ref_cows || root == root->fs_info->csum_root)
3841                 block_rsv = trans->block_rsv;
3842
3843         if (!block_rsv)
3844                 block_rsv = root->block_rsv;
3845
3846         if (!block_rsv)
3847                 block_rsv = &root->fs_info->empty_block_rsv;
3848
3849         return block_rsv;
3850 }
3851
3852 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3853                                u64 num_bytes)
3854 {
3855         int ret = -ENOSPC;
3856         spin_lock(&block_rsv->lock);
3857         if (block_rsv->reserved >= num_bytes) {
3858                 block_rsv->reserved -= num_bytes;
3859                 if (block_rsv->reserved < block_rsv->size)
3860                         block_rsv->full = 0;
3861                 ret = 0;
3862         }
3863         spin_unlock(&block_rsv->lock);
3864         return ret;
3865 }
3866
3867 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3868                                 u64 num_bytes, int update_size)
3869 {
3870         spin_lock(&block_rsv->lock);
3871         block_rsv->reserved += num_bytes;
3872         if (update_size)
3873                 block_rsv->size += num_bytes;
3874         else if (block_rsv->reserved >= block_rsv->size)
3875                 block_rsv->full = 1;
3876         spin_unlock(&block_rsv->lock);
3877 }
3878
3879 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
3880                                     struct btrfs_block_rsv *block_rsv,
3881                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3882 {
3883         struct btrfs_space_info *space_info = block_rsv->space_info;
3884
3885         spin_lock(&block_rsv->lock);
3886         if (num_bytes == (u64)-1)
3887                 num_bytes = block_rsv->size;
3888         block_rsv->size -= num_bytes;
3889         if (block_rsv->reserved >= block_rsv->size) {
3890                 num_bytes = block_rsv->reserved - block_rsv->size;
3891                 block_rsv->reserved = block_rsv->size;
3892                 block_rsv->full = 1;
3893         } else {
3894                 num_bytes = 0;
3895         }
3896         spin_unlock(&block_rsv->lock);
3897
3898         if (num_bytes > 0) {
3899                 if (dest) {
3900                         spin_lock(&dest->lock);
3901                         if (!dest->full) {
3902                                 u64 bytes_to_add;
3903
3904                                 bytes_to_add = dest->size - dest->reserved;
3905                                 bytes_to_add = min(num_bytes, bytes_to_add);
3906                                 dest->reserved += bytes_to_add;
3907                                 if (dest->reserved >= dest->size)
3908                                         dest->full = 1;
3909                                 num_bytes -= bytes_to_add;
3910                         }
3911                         spin_unlock(&dest->lock);
3912                 }
3913                 if (num_bytes) {
3914                         spin_lock(&space_info->lock);
3915                         space_info->bytes_may_use -= num_bytes;
3916                         trace_btrfs_space_reservation(fs_info, "space_info",
3917                                               (u64)(unsigned long)space_info,
3918                                               num_bytes, 0);
3919                         space_info->reservation_progress++;
3920                         spin_unlock(&space_info->lock);
3921                 }
3922         }
3923 }
3924
3925 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3926                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3927 {
3928         int ret;
3929
3930         ret = block_rsv_use_bytes(src, num_bytes);
3931         if (ret)
3932                 return ret;
3933
3934         block_rsv_add_bytes(dst, num_bytes, 1);
3935         return 0;
3936 }
3937
3938 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3939 {
3940         memset(rsv, 0, sizeof(*rsv));
3941         spin_lock_init(&rsv->lock);
3942 }
3943
3944 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3945 {
3946         struct btrfs_block_rsv *block_rsv;
3947         struct btrfs_fs_info *fs_info = root->fs_info;
3948
3949         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3950         if (!block_rsv)
3951                 return NULL;
3952
3953         btrfs_init_block_rsv(block_rsv);
3954         block_rsv->space_info = __find_space_info(fs_info,
3955                                                   BTRFS_BLOCK_GROUP_METADATA);
3956         return block_rsv;
3957 }
3958
3959 void btrfs_free_block_rsv(struct btrfs_root *root,
3960                           struct btrfs_block_rsv *rsv)
3961 {
3962         btrfs_block_rsv_release(root, rsv, (u64)-1);
3963         kfree(rsv);
3964 }
3965
3966 static inline int __block_rsv_add(struct btrfs_root *root,
3967                                   struct btrfs_block_rsv *block_rsv,
3968                                   u64 num_bytes, int flush)
3969 {
3970         int ret;
3971
3972         if (num_bytes == 0)
3973                 return 0;
3974
3975         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
3976         if (!ret) {
3977                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3978                 return 0;
3979         }
3980
3981         return ret;
3982 }
3983
3984 int btrfs_block_rsv_add(struct btrfs_root *root,
3985                         struct btrfs_block_rsv *block_rsv,
3986                         u64 num_bytes)
3987 {
3988         return __block_rsv_add(root, block_rsv, num_bytes, 1);
3989 }
3990
3991 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
3992                                 struct btrfs_block_rsv *block_rsv,
3993                                 u64 num_bytes)
3994 {
3995         return __block_rsv_add(root, block_rsv, num_bytes, 0);
3996 }
3997
3998 int btrfs_block_rsv_check(struct btrfs_root *root,
3999                           struct btrfs_block_rsv *block_rsv, int min_factor)
4000 {
4001         u64 num_bytes = 0;
4002         int ret = -ENOSPC;
4003
4004         if (!block_rsv)
4005                 return 0;
4006
4007         spin_lock(&block_rsv->lock);
4008         num_bytes = div_factor(block_rsv->size, min_factor);
4009         if (block_rsv->reserved >= num_bytes)
4010                 ret = 0;
4011         spin_unlock(&block_rsv->lock);
4012
4013         return ret;
4014 }
4015
4016 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
4017                                            struct btrfs_block_rsv *block_rsv,
4018                                            u64 min_reserved, int flush)
4019 {
4020         u64 num_bytes = 0;
4021         int ret = -ENOSPC;
4022
4023         if (!block_rsv)
4024                 return 0;
4025
4026         spin_lock(&block_rsv->lock);
4027         num_bytes = min_reserved;
4028         if (block_rsv->reserved >= num_bytes)
4029                 ret = 0;
4030         else
4031                 num_bytes -= block_rsv->reserved;
4032         spin_unlock(&block_rsv->lock);
4033
4034         if (!ret)
4035                 return 0;
4036
4037         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4038         if (!ret) {
4039                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4040                 return 0;
4041         }
4042
4043         return ret;
4044 }
4045
4046 int btrfs_block_rsv_refill(struct btrfs_root *root,
4047                            struct btrfs_block_rsv *block_rsv,
4048                            u64 min_reserved)
4049 {
4050         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
4051 }
4052
4053 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
4054                                    struct btrfs_block_rsv *block_rsv,
4055                                    u64 min_reserved)
4056 {
4057         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
4058 }
4059
4060 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4061                             struct btrfs_block_rsv *dst_rsv,
4062                             u64 num_bytes)
4063 {
4064         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4065 }
4066
4067 void btrfs_block_rsv_release(struct btrfs_root *root,
4068                              struct btrfs_block_rsv *block_rsv,
4069                              u64 num_bytes)
4070 {
4071         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4072         if (global_rsv->full || global_rsv == block_rsv ||
4073             block_rsv->space_info != global_rsv->space_info)
4074                 global_rsv = NULL;
4075         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4076                                 num_bytes);
4077 }
4078
4079 /*
4080  * helper to calculate size of global block reservation.
4081  * the desired value is sum of space used by extent tree,
4082  * checksum tree and root tree
4083  */
4084 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4085 {
4086         struct btrfs_space_info *sinfo;
4087         u64 num_bytes;
4088         u64 meta_used;
4089         u64 data_used;
4090         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4091
4092         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4093         spin_lock(&sinfo->lock);
4094         data_used = sinfo->bytes_used;
4095         spin_unlock(&sinfo->lock);
4096
4097         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4098         spin_lock(&sinfo->lock);
4099         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4100                 data_used = 0;
4101         meta_used = sinfo->bytes_used;
4102         spin_unlock(&sinfo->lock);
4103
4104         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4105                     csum_size * 2;
4106         num_bytes += div64_u64(data_used + meta_used, 50);
4107
4108         if (num_bytes * 3 > meta_used)
4109                 num_bytes = div64_u64(meta_used, 3) * 2;
4110
4111         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4112 }
4113
4114 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4115 {
4116         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4117         struct btrfs_space_info *sinfo = block_rsv->space_info;
4118         u64 num_bytes;
4119
4120         num_bytes = calc_global_metadata_size(fs_info);
4121
4122         spin_lock(&block_rsv->lock);
4123         spin_lock(&sinfo->lock);
4124
4125         block_rsv->size = num_bytes;
4126
4127         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4128                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4129                     sinfo->bytes_may_use;
4130
4131         if (sinfo->total_bytes > num_bytes) {
4132                 num_bytes = sinfo->total_bytes - num_bytes;
4133                 block_rsv->reserved += num_bytes;
4134                 sinfo->bytes_may_use += num_bytes;
4135                 trace_btrfs_space_reservation(fs_info, "space_info",
4136                                       (u64)(unsigned long)sinfo, num_bytes, 1);
4137         }
4138
4139         if (block_rsv->reserved >= block_rsv->size) {
4140                 num_bytes = block_rsv->reserved - block_rsv->size;
4141                 sinfo->bytes_may_use -= num_bytes;
4142                 trace_btrfs_space_reservation(fs_info, "space_info",
4143                                       (u64)(unsigned long)sinfo, num_bytes, 0);
4144                 sinfo->reservation_progress++;
4145                 block_rsv->reserved = block_rsv->size;
4146                 block_rsv->full = 1;
4147         }
4148
4149         spin_unlock(&sinfo->lock);
4150         spin_unlock(&block_rsv->lock);
4151 }
4152
4153 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4154 {
4155         struct btrfs_space_info *space_info;
4156
4157         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4158         fs_info->chunk_block_rsv.space_info = space_info;
4159
4160         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4161         fs_info->global_block_rsv.space_info = space_info;
4162         fs_info->delalloc_block_rsv.space_info = space_info;
4163         fs_info->trans_block_rsv.space_info = space_info;
4164         fs_info->empty_block_rsv.space_info = space_info;
4165         fs_info->delayed_block_rsv.space_info = space_info;
4166
4167         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4168         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4169         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4170         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4171         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4172
4173         update_global_block_rsv(fs_info);
4174 }
4175
4176 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4177 {
4178         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4179                                 (u64)-1);
4180         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4181         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4182         WARN_ON(fs_info->trans_block_rsv.size > 0);
4183         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4184         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4185         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4186         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4187         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4188 }
4189
4190 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4191                                   struct btrfs_root *root)
4192 {
4193         if (!trans->bytes_reserved)
4194                 return;
4195
4196         trace_btrfs_space_reservation(root->fs_info, "transaction",
4197                                       (u64)(unsigned long)trans,
4198                                       trans->bytes_reserved, 0);
4199         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4200         trans->bytes_reserved = 0;
4201 }
4202
4203 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4204                                   struct inode *inode)
4205 {
4206         struct btrfs_root *root = BTRFS_I(inode)->root;
4207         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4208         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4209
4210         /*
4211          * We need to hold space in order to delete our orphan item once we've
4212          * added it, so this takes the reservation so we can release it later
4213          * when we are truly done with the orphan item.
4214          */
4215         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4216         trace_btrfs_space_reservation(root->fs_info, "orphan",
4217                                       btrfs_ino(inode), num_bytes, 1);
4218         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4219 }
4220
4221 void btrfs_orphan_release_metadata(struct inode *inode)
4222 {
4223         struct btrfs_root *root = BTRFS_I(inode)->root;
4224         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4225         trace_btrfs_space_reservation(root->fs_info, "orphan",
4226                                       btrfs_ino(inode), num_bytes, 0);
4227         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4228 }
4229
4230 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4231                                 struct btrfs_pending_snapshot *pending)
4232 {
4233         struct btrfs_root *root = pending->root;
4234         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4235         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4236         /*
4237          * two for root back/forward refs, two for directory entries
4238          * and one for root of the snapshot.
4239          */
4240         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4241         dst_rsv->space_info = src_rsv->space_info;
4242         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4243 }
4244
4245 /**
4246  * drop_outstanding_extent - drop an outstanding extent
4247  * @inode: the inode we're dropping the extent for
4248  *
4249  * This is called when we are freeing up an outstanding extent, either called
4250  * after an error or after an extent is written.  This will return the number of
4251  * reserved extents that need to be freed.  This must be called with
4252  * BTRFS_I(inode)->lock held.
4253  */
4254 static unsigned drop_outstanding_extent(struct inode *inode)
4255 {
4256         unsigned drop_inode_space = 0;
4257         unsigned dropped_extents = 0;
4258
4259         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4260         BTRFS_I(inode)->outstanding_extents--;
4261
4262         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4263             BTRFS_I(inode)->delalloc_meta_reserved) {
4264                 drop_inode_space = 1;
4265                 BTRFS_I(inode)->delalloc_meta_reserved = 0;
4266         }
4267
4268         /*
4269          * If we have more or the same amount of outsanding extents than we have
4270          * reserved then we need to leave the reserved extents count alone.
4271          */
4272         if (BTRFS_I(inode)->outstanding_extents >=
4273             BTRFS_I(inode)->reserved_extents)
4274                 return drop_inode_space;
4275
4276         dropped_extents = BTRFS_I(inode)->reserved_extents -
4277                 BTRFS_I(inode)->outstanding_extents;
4278         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4279         return dropped_extents + drop_inode_space;
4280 }
4281
4282 /**
4283  * calc_csum_metadata_size - return the amount of metada space that must be
4284  *      reserved/free'd for the given bytes.
4285  * @inode: the inode we're manipulating
4286  * @num_bytes: the number of bytes in question
4287  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4288  *
4289  * This adjusts the number of csum_bytes in the inode and then returns the
4290  * correct amount of metadata that must either be reserved or freed.  We
4291  * calculate how many checksums we can fit into one leaf and then divide the
4292  * number of bytes that will need to be checksumed by this value to figure out
4293  * how many checksums will be required.  If we are adding bytes then the number
4294  * may go up and we will return the number of additional bytes that must be
4295  * reserved.  If it is going down we will return the number of bytes that must
4296  * be freed.
4297  *
4298  * This must be called with BTRFS_I(inode)->lock held.
4299  */
4300 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4301                                    int reserve)
4302 {
4303         struct btrfs_root *root = BTRFS_I(inode)->root;
4304         u64 csum_size;
4305         int num_csums_per_leaf;
4306         int num_csums;
4307         int old_csums;
4308
4309         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4310             BTRFS_I(inode)->csum_bytes == 0)
4311                 return 0;
4312
4313         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4314         if (reserve)
4315                 BTRFS_I(inode)->csum_bytes += num_bytes;
4316         else
4317                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4318         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4319         num_csums_per_leaf = (int)div64_u64(csum_size,
4320                                             sizeof(struct btrfs_csum_item) +
4321                                             sizeof(struct btrfs_disk_key));
4322         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4323         num_csums = num_csums + num_csums_per_leaf - 1;
4324         num_csums = num_csums / num_csums_per_leaf;
4325
4326         old_csums = old_csums + num_csums_per_leaf - 1;
4327         old_csums = old_csums / num_csums_per_leaf;
4328
4329         /* No change, no need to reserve more */
4330         if (old_csums == num_csums)
4331                 return 0;
4332
4333         if (reserve)
4334                 return btrfs_calc_trans_metadata_size(root,
4335                                                       num_csums - old_csums);
4336
4337         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4338 }
4339
4340 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4341 {
4342         struct btrfs_root *root = BTRFS_I(inode)->root;
4343         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4344         u64 to_reserve = 0;
4345         u64 csum_bytes;
4346         unsigned nr_extents = 0;
4347         int extra_reserve = 0;
4348         int flush = 1;
4349         int ret;
4350
4351         /* Need to be holding the i_mutex here if we aren't free space cache */
4352         if (btrfs_is_free_space_inode(root, inode))
4353                 flush = 0;
4354
4355         if (flush && btrfs_transaction_in_commit(root->fs_info))
4356                 schedule_timeout(1);
4357
4358         mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4359         num_bytes = ALIGN(num_bytes, root->sectorsize);
4360
4361         spin_lock(&BTRFS_I(inode)->lock);
4362         BTRFS_I(inode)->outstanding_extents++;
4363
4364         if (BTRFS_I(inode)->outstanding_extents >
4365             BTRFS_I(inode)->reserved_extents)
4366                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4367                         BTRFS_I(inode)->reserved_extents;
4368
4369         /*
4370          * Add an item to reserve for updating the inode when we complete the
4371          * delalloc io.
4372          */
4373         if (!BTRFS_I(inode)->delalloc_meta_reserved) {
4374                 nr_extents++;
4375                 extra_reserve = 1;
4376         }
4377
4378         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4379         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4380         csum_bytes = BTRFS_I(inode)->csum_bytes;
4381         spin_unlock(&BTRFS_I(inode)->lock);
4382
4383         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4384         if (ret) {
4385                 u64 to_free = 0;
4386                 unsigned dropped;
4387
4388                 spin_lock(&BTRFS_I(inode)->lock);
4389                 dropped = drop_outstanding_extent(inode);
4390                 /*
4391                  * If the inodes csum_bytes is the same as the original
4392                  * csum_bytes then we know we haven't raced with any free()ers
4393                  * so we can just reduce our inodes csum bytes and carry on.
4394                  * Otherwise we have to do the normal free thing to account for
4395                  * the case that the free side didn't free up its reserve
4396                  * because of this outstanding reservation.
4397                  */
4398                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4399                         calc_csum_metadata_size(inode, num_bytes, 0);
4400                 else
4401                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4402                 spin_unlock(&BTRFS_I(inode)->lock);
4403                 if (dropped)
4404                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4405
4406                 if (to_free) {
4407                         btrfs_block_rsv_release(root, block_rsv, to_free);
4408                         trace_btrfs_space_reservation(root->fs_info,
4409                                                       "delalloc",
4410                                                       btrfs_ino(inode),
4411                                                       to_free, 0);
4412                 }
4413                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4414                 return ret;
4415         }
4416
4417         spin_lock(&BTRFS_I(inode)->lock);
4418         if (extra_reserve) {
4419                 BTRFS_I(inode)->delalloc_meta_reserved = 1;
4420                 nr_extents--;
4421         }
4422         BTRFS_I(inode)->reserved_extents += nr_extents;
4423         spin_unlock(&BTRFS_I(inode)->lock);
4424         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4425
4426         if (to_reserve)
4427                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4428                                               btrfs_ino(inode), to_reserve, 1);
4429         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4430
4431         return 0;
4432 }
4433
4434 /**
4435  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4436  * @inode: the inode to release the reservation for
4437  * @num_bytes: the number of bytes we're releasing
4438  *
4439  * This will release the metadata reservation for an inode.  This can be called
4440  * once we complete IO for a given set of bytes to release their metadata
4441  * reservations.
4442  */
4443 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4444 {
4445         struct btrfs_root *root = BTRFS_I(inode)->root;
4446         u64 to_free = 0;
4447         unsigned dropped;
4448
4449         num_bytes = ALIGN(num_bytes, root->sectorsize);
4450         spin_lock(&BTRFS_I(inode)->lock);
4451         dropped = drop_outstanding_extent(inode);
4452
4453         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4454         spin_unlock(&BTRFS_I(inode)->lock);
4455         if (dropped > 0)
4456                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4457
4458         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4459                                       btrfs_ino(inode), to_free, 0);
4460         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4461                                 to_free);
4462 }
4463
4464 /**
4465  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4466  * @inode: inode we're writing to
4467  * @num_bytes: the number of bytes we want to allocate
4468  *
4469  * This will do the following things
4470  *
4471  * o reserve space in the data space info for num_bytes
4472  * o reserve space in the metadata space info based on number of outstanding
4473  *   extents and how much csums will be needed
4474  * o add to the inodes ->delalloc_bytes
4475  * o add it to the fs_info's delalloc inodes list.
4476  *
4477  * This will return 0 for success and -ENOSPC if there is no space left.
4478  */
4479 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4480 {
4481         int ret;
4482
4483         ret = btrfs_check_data_free_space(inode, num_bytes);
4484         if (ret)
4485                 return ret;
4486
4487         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4488         if (ret) {
4489                 btrfs_free_reserved_data_space(inode, num_bytes);
4490                 return ret;
4491         }
4492
4493         return 0;
4494 }
4495
4496 /**
4497  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4498  * @inode: inode we're releasing space for
4499  * @num_bytes: the number of bytes we want to free up
4500  *
4501  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4502  * called in the case that we don't need the metadata AND data reservations
4503  * anymore.  So if there is an error or we insert an inline extent.
4504  *
4505  * This function will release the metadata space that was not used and will
4506  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4507  * list if there are no delalloc bytes left.
4508  */
4509 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4510 {
4511         btrfs_delalloc_release_metadata(inode, num_bytes);
4512         btrfs_free_reserved_data_space(inode, num_bytes);
4513 }
4514
4515 static int update_block_group(struct btrfs_trans_handle *trans,
4516                               struct btrfs_root *root,
4517                               u64 bytenr, u64 num_bytes, int alloc)
4518 {
4519         struct btrfs_block_group_cache *cache = NULL;
4520         struct btrfs_fs_info *info = root->fs_info;
4521         u64 total = num_bytes;
4522         u64 old_val;
4523         u64 byte_in_group;
4524         int factor;
4525
4526         /* block accounting for super block */
4527         spin_lock(&info->delalloc_lock);
4528         old_val = btrfs_super_bytes_used(info->super_copy);
4529         if (alloc)
4530                 old_val += num_bytes;
4531         else
4532                 old_val -= num_bytes;
4533         btrfs_set_super_bytes_used(info->super_copy, old_val);
4534         spin_unlock(&info->delalloc_lock);
4535
4536         while (total) {
4537                 cache = btrfs_lookup_block_group(info, bytenr);
4538                 if (!cache)
4539                         return -1;
4540                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4541                                     BTRFS_BLOCK_GROUP_RAID1 |
4542                                     BTRFS_BLOCK_GROUP_RAID10))
4543                         factor = 2;
4544                 else
4545                         factor = 1;
4546                 /*
4547                  * If this block group has free space cache written out, we
4548                  * need to make sure to load it if we are removing space.  This
4549                  * is because we need the unpinning stage to actually add the
4550                  * space back to the block group, otherwise we will leak space.
4551                  */
4552                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4553                         cache_block_group(cache, trans, NULL, 1);
4554
4555                 byte_in_group = bytenr - cache->key.objectid;
4556                 WARN_ON(byte_in_group > cache->key.offset);
4557
4558                 spin_lock(&cache->space_info->lock);
4559                 spin_lock(&cache->lock);
4560
4561                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4562                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4563                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4564
4565                 cache->dirty = 1;
4566                 old_val = btrfs_block_group_used(&cache->item);
4567                 num_bytes = min(total, cache->key.offset - byte_in_group);
4568                 if (alloc) {
4569                         old_val += num_bytes;
4570                         btrfs_set_block_group_used(&cache->item, old_val);
4571                         cache->reserved -= num_bytes;
4572                         cache->space_info->bytes_reserved -= num_bytes;
4573                         cache->space_info->bytes_used += num_bytes;
4574                         cache->space_info->disk_used += num_bytes * factor;
4575                         spin_unlock(&cache->lock);
4576                         spin_unlock(&cache->space_info->lock);
4577                 } else {
4578                         old_val -= num_bytes;
4579                         btrfs_set_block_group_used(&cache->item, old_val);
4580                         cache->pinned += num_bytes;
4581                         cache->space_info->bytes_pinned += num_bytes;
4582                         cache->space_info->bytes_used -= num_bytes;
4583                         cache->space_info->disk_used -= num_bytes * factor;
4584                         spin_unlock(&cache->lock);
4585                         spin_unlock(&cache->space_info->lock);
4586
4587                         set_extent_dirty(info->pinned_extents,
4588                                          bytenr, bytenr + num_bytes - 1,
4589                                          GFP_NOFS | __GFP_NOFAIL);
4590                 }
4591                 btrfs_put_block_group(cache);
4592                 total -= num_bytes;
4593                 bytenr += num_bytes;
4594         }
4595         return 0;
4596 }
4597
4598 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4599 {
4600         struct btrfs_block_group_cache *cache;
4601         u64 bytenr;
4602
4603         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4604         if (!cache)
4605                 return 0;
4606
4607         bytenr = cache->key.objectid;
4608         btrfs_put_block_group(cache);
4609
4610         return bytenr;
4611 }
4612
4613 static int pin_down_extent(struct btrfs_root *root,
4614                            struct btrfs_block_group_cache *cache,
4615                            u64 bytenr, u64 num_bytes, int reserved)
4616 {
4617         spin_lock(&cache->space_info->lock);
4618         spin_lock(&cache->lock);
4619         cache->pinned += num_bytes;
4620         cache->space_info->bytes_pinned += num_bytes;
4621         if (reserved) {
4622                 cache->reserved -= num_bytes;
4623                 cache->space_info->bytes_reserved -= num_bytes;
4624         }
4625         spin_unlock(&cache->lock);
4626         spin_unlock(&cache->space_info->lock);
4627
4628         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4629                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4630         return 0;
4631 }
4632
4633 /*
4634  * this function must be called within transaction
4635  */
4636 int btrfs_pin_extent(struct btrfs_root *root,
4637                      u64 bytenr, u64 num_bytes, int reserved)
4638 {
4639         struct btrfs_block_group_cache *cache;
4640
4641         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4642         BUG_ON(!cache);
4643
4644         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4645
4646         btrfs_put_block_group(cache);
4647         return 0;
4648 }
4649
4650 /*
4651  * this function must be called within transaction
4652  */
4653 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4654                                     struct btrfs_root *root,
4655                                     u64 bytenr, u64 num_bytes)
4656 {
4657         struct btrfs_block_group_cache *cache;
4658
4659         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4660         BUG_ON(!cache);
4661
4662         /*
4663          * pull in the free space cache (if any) so that our pin
4664          * removes the free space from the cache.  We have load_only set
4665          * to one because the slow code to read in the free extents does check
4666          * the pinned extents.
4667          */
4668         cache_block_group(cache, trans, root, 1);
4669
4670         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4671
4672         /* remove us from the free space cache (if we're there at all) */
4673         btrfs_remove_free_space(cache, bytenr, num_bytes);
4674         btrfs_put_block_group(cache);
4675         return 0;
4676 }
4677
4678 /**
4679  * btrfs_update_reserved_bytes - update the block_group and space info counters
4680  * @cache:      The cache we are manipulating
4681  * @num_bytes:  The number of bytes in question
4682  * @reserve:    One of the reservation enums
4683  *
4684  * This is called by the allocator when it reserves space, or by somebody who is
4685  * freeing space that was never actually used on disk.  For example if you
4686  * reserve some space for a new leaf in transaction A and before transaction A
4687  * commits you free that leaf, you call this with reserve set to 0 in order to
4688  * clear the reservation.
4689  *
4690  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4691  * ENOSPC accounting.  For data we handle the reservation through clearing the
4692  * delalloc bits in the io_tree.  We have to do this since we could end up
4693  * allocating less disk space for the amount of data we have reserved in the
4694  * case of compression.
4695  *
4696  * If this is a reservation and the block group has become read only we cannot
4697  * make the reservation and return -EAGAIN, otherwise this function always
4698  * succeeds.
4699  */
4700 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4701                                        u64 num_bytes, int reserve)
4702 {
4703         struct btrfs_space_info *space_info = cache->space_info;
4704         int ret = 0;
4705         spin_lock(&space_info->lock);
4706         spin_lock(&cache->lock);
4707         if (reserve != RESERVE_FREE) {
4708                 if (cache->ro) {
4709                         ret = -EAGAIN;
4710                 } else {
4711                         cache->reserved += num_bytes;
4712                         space_info->bytes_reserved += num_bytes;
4713                         if (reserve == RESERVE_ALLOC) {
4714                                 trace_btrfs_space_reservation(cache->fs_info,
4715                                               "space_info",
4716                                               (u64)(unsigned long)space_info,
4717                                               num_bytes, 0);
4718                                 space_info->bytes_may_use -= num_bytes;
4719                         }
4720                 }
4721         } else {
4722                 if (cache->ro)
4723                         space_info->bytes_readonly += num_bytes;
4724                 cache->reserved -= num_bytes;
4725                 space_info->bytes_reserved -= num_bytes;
4726                 space_info->reservation_progress++;
4727         }
4728         spin_unlock(&cache->lock);
4729         spin_unlock(&space_info->lock);
4730         return ret;
4731 }
4732
4733 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4734                                 struct btrfs_root *root)
4735 {
4736         struct btrfs_fs_info *fs_info = root->fs_info;
4737         struct btrfs_caching_control *next;
4738         struct btrfs_caching_control *caching_ctl;
4739         struct btrfs_block_group_cache *cache;
4740
4741         down_write(&fs_info->extent_commit_sem);
4742
4743         list_for_each_entry_safe(caching_ctl, next,
4744                                  &fs_info->caching_block_groups, list) {
4745                 cache = caching_ctl->block_group;
4746                 if (block_group_cache_done(cache)) {
4747                         cache->last_byte_to_unpin = (u64)-1;
4748                         list_del_init(&caching_ctl->list);
4749                         put_caching_control(caching_ctl);
4750                 } else {
4751                         cache->last_byte_to_unpin = caching_ctl->progress;
4752                 }
4753         }
4754
4755         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4756                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4757         else
4758                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4759
4760         up_write(&fs_info->extent_commit_sem);
4761
4762         update_global_block_rsv(fs_info);
4763 }
4764
4765 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4766 {
4767         struct btrfs_fs_info *fs_info = root->fs_info;
4768         struct btrfs_block_group_cache *cache = NULL;
4769         u64 len;
4770
4771         while (start <= end) {
4772                 if (!cache ||
4773                     start >= cache->key.objectid + cache->key.offset) {
4774                         if (cache)
4775                                 btrfs_put_block_group(cache);
4776                         cache = btrfs_lookup_block_group(fs_info, start);
4777                         BUG_ON(!cache);
4778                 }
4779
4780                 len = cache->key.objectid + cache->key.offset - start;
4781                 len = min(len, end + 1 - start);
4782
4783                 if (start < cache->last_byte_to_unpin) {
4784                         len = min(len, cache->last_byte_to_unpin - start);
4785                         btrfs_add_free_space(cache, start, len);
4786                 }
4787
4788                 start += len;
4789
4790                 spin_lock(&cache->space_info->lock);
4791                 spin_lock(&cache->lock);
4792                 cache->pinned -= len;
4793                 cache->space_info->bytes_pinned -= len;
4794                 if (cache->ro)
4795                         cache->space_info->bytes_readonly += len;
4796                 spin_unlock(&cache->lock);
4797                 spin_unlock(&cache->space_info->lock);
4798         }
4799
4800         if (cache)
4801                 btrfs_put_block_group(cache);
4802         return 0;
4803 }
4804
4805 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4806                                struct btrfs_root *root)
4807 {
4808         struct btrfs_fs_info *fs_info = root->fs_info;
4809         struct extent_io_tree *unpin;
4810         u64 start;
4811         u64 end;
4812         int ret;
4813
4814         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4815                 unpin = &fs_info->freed_extents[1];
4816         else
4817                 unpin = &fs_info->freed_extents[0];
4818
4819         while (1) {
4820                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4821                                             EXTENT_DIRTY);
4822                 if (ret)
4823                         break;
4824
4825                 if (btrfs_test_opt(root, DISCARD))
4826                         ret = btrfs_discard_extent(root, start,
4827                                                    end + 1 - start, NULL);
4828
4829                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4830                 unpin_extent_range(root, start, end);
4831                 cond_resched();
4832         }
4833
4834         return 0;
4835 }
4836
4837 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4838                                 struct btrfs_root *root,
4839                                 u64 bytenr, u64 num_bytes, u64 parent,
4840                                 u64 root_objectid, u64 owner_objectid,
4841                                 u64 owner_offset, int refs_to_drop,
4842                                 struct btrfs_delayed_extent_op *extent_op)
4843 {
4844         struct btrfs_key key;
4845         struct btrfs_path *path;
4846         struct btrfs_fs_info *info = root->fs_info;
4847         struct btrfs_root *extent_root = info->extent_root;
4848         struct extent_buffer *leaf;
4849         struct btrfs_extent_item *ei;
4850         struct btrfs_extent_inline_ref *iref;
4851         int ret;
4852         int is_data;
4853         int extent_slot = 0;
4854         int found_extent = 0;
4855         int num_to_del = 1;
4856         u32 item_size;
4857         u64 refs;
4858
4859         path = btrfs_alloc_path();
4860         if (!path)
4861                 return -ENOMEM;
4862
4863         path->reada = 1;
4864         path->leave_spinning = 1;
4865
4866         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4867         BUG_ON(!is_data && refs_to_drop != 1);
4868
4869         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4870                                     bytenr, num_bytes, parent,
4871                                     root_objectid, owner_objectid,
4872                                     owner_offset);
4873         if (ret == 0) {
4874                 extent_slot = path->slots[0];
4875                 while (extent_slot >= 0) {
4876                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4877                                               extent_slot);
4878                         if (key.objectid != bytenr)
4879                                 break;
4880                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4881                             key.offset == num_bytes) {
4882                                 found_extent = 1;
4883                                 break;
4884                         }
4885                         if (path->slots[0] - extent_slot > 5)
4886                                 break;
4887                         extent_slot--;
4888                 }
4889 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4890                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4891                 if (found_extent && item_size < sizeof(*ei))
4892                         found_extent = 0;
4893 #endif
4894                 if (!found_extent) {
4895                         BUG_ON(iref);
4896                         ret = remove_extent_backref(trans, extent_root, path,
4897                                                     NULL, refs_to_drop,
4898                                                     is_data);
4899                         BUG_ON(ret);
4900                         btrfs_release_path(path);
4901                         path->leave_spinning = 1;
4902
4903                         key.objectid = bytenr;
4904                         key.type = BTRFS_EXTENT_ITEM_KEY;
4905                         key.offset = num_bytes;
4906
4907                         ret = btrfs_search_slot(trans, extent_root,
4908                                                 &key, path, -1, 1);
4909                         if (ret) {
4910                                 printk(KERN_ERR "umm, got %d back from search"
4911                                        ", was looking for %llu\n", ret,
4912                                        (unsigned long long)bytenr);
4913                                 if (ret > 0)
4914                                         btrfs_print_leaf(extent_root,
4915                                                          path->nodes[0]);
4916                         }
4917                         BUG_ON(ret);
4918                         extent_slot = path->slots[0];
4919                 }
4920         } else {
4921                 btrfs_print_leaf(extent_root, path->nodes[0]);
4922                 WARN_ON(1);
4923                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4924                        "parent %llu root %llu  owner %llu offset %llu\n",
4925                        (unsigned long long)bytenr,
4926                        (unsigned long long)parent,
4927                        (unsigned long long)root_objectid,
4928                        (unsigned long long)owner_objectid,
4929                        (unsigned long long)owner_offset);
4930         }
4931
4932         leaf = path->nodes[0];
4933         item_size = btrfs_item_size_nr(leaf, extent_slot);
4934 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4935         if (item_size < sizeof(*ei)) {
4936                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4937                 ret = convert_extent_item_v0(trans, extent_root, path,
4938                                              owner_objectid, 0);
4939                 BUG_ON(ret < 0);
4940
4941                 btrfs_release_path(path);
4942                 path->leave_spinning = 1;
4943
4944                 key.objectid = bytenr;
4945                 key.type = BTRFS_EXTENT_ITEM_KEY;
4946                 key.offset = num_bytes;
4947
4948                 ret = btrfs_search_slot(trans, extent_root, &key, path,
4949                                         -1, 1);
4950                 if (ret) {
4951                         printk(KERN_ERR "umm, got %d back from search"
4952                                ", was looking for %llu\n", ret,
4953                                (unsigned long long)bytenr);
4954                         btrfs_print_leaf(extent_root, path->nodes[0]);
4955                 }
4956                 BUG_ON(ret);
4957                 extent_slot = path->slots[0];
4958                 leaf = path->nodes[0];
4959                 item_size = btrfs_item_size_nr(leaf, extent_slot);
4960         }
4961 #endif
4962         BUG_ON(item_size < sizeof(*ei));
4963         ei = btrfs_item_ptr(leaf, extent_slot,
4964                             struct btrfs_extent_item);
4965         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
4966                 struct btrfs_tree_block_info *bi;
4967                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4968                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4969                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4970         }
4971
4972         refs = btrfs_extent_refs(leaf, ei);
4973         BUG_ON(refs < refs_to_drop);
4974         refs -= refs_to_drop;
4975
4976         if (refs > 0) {
4977                 if (extent_op)
4978                         __run_delayed_extent_op(extent_op, leaf, ei);
4979                 /*
4980                  * In the case of inline back ref, reference count will
4981                  * be updated by remove_extent_backref
4982                  */
4983                 if (iref) {
4984                         BUG_ON(!found_extent);
4985                 } else {
4986                         btrfs_set_extent_refs(leaf, ei, refs);
4987                         btrfs_mark_buffer_dirty(leaf);
4988                 }
4989                 if (found_extent) {
4990                         ret = remove_extent_backref(trans, extent_root, path,
4991                                                     iref, refs_to_drop,
4992                                                     is_data);
4993                         BUG_ON(ret);
4994                 }
4995         } else {
4996                 if (found_extent) {
4997                         BUG_ON(is_data && refs_to_drop !=
4998                                extent_data_ref_count(root, path, iref));
4999                         if (iref) {
5000                                 BUG_ON(path->slots[0] != extent_slot);
5001                         } else {
5002                                 BUG_ON(path->slots[0] != extent_slot + 1);
5003                                 path->slots[0] = extent_slot;
5004                                 num_to_del = 2;
5005                         }
5006                 }
5007
5008                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5009                                       num_to_del);
5010                 BUG_ON(ret);
5011                 btrfs_release_path(path);
5012
5013                 if (is_data) {
5014                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5015                         BUG_ON(ret);
5016                 } else {
5017                         invalidate_mapping_pages(info->btree_inode->i_mapping,
5018                              bytenr >> PAGE_CACHE_SHIFT,
5019                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
5020                 }
5021
5022                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
5023                 BUG_ON(ret);
5024         }
5025         btrfs_free_path(path);
5026         return ret;
5027 }
5028
5029 /*
5030  * when we free an block, it is possible (and likely) that we free the last
5031  * delayed ref for that extent as well.  This searches the delayed ref tree for
5032  * a given extent, and if there are no other delayed refs to be processed, it
5033  * removes it from the tree.
5034  */
5035 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5036                                       struct btrfs_root *root, u64 bytenr)
5037 {
5038         struct btrfs_delayed_ref_head *head;
5039         struct btrfs_delayed_ref_root *delayed_refs;
5040         struct btrfs_delayed_ref_node *ref;
5041         struct rb_node *node;
5042         int ret = 0;
5043
5044         delayed_refs = &trans->transaction->delayed_refs;
5045         spin_lock(&delayed_refs->lock);
5046         head = btrfs_find_delayed_ref_head(trans, bytenr);
5047         if (!head)
5048                 goto out;
5049
5050         node = rb_prev(&head->node.rb_node);
5051         if (!node)
5052                 goto out;
5053
5054         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5055
5056         /* there are still entries for this ref, we can't drop it */
5057         if (ref->bytenr == bytenr)
5058                 goto out;
5059
5060         if (head->extent_op) {
5061                 if (!head->must_insert_reserved)
5062                         goto out;
5063                 kfree(head->extent_op);
5064                 head->extent_op = NULL;
5065         }
5066
5067         /*
5068          * waiting for the lock here would deadlock.  If someone else has it
5069          * locked they are already in the process of dropping it anyway
5070          */
5071         if (!mutex_trylock(&head->mutex))
5072                 goto out;
5073
5074         /*
5075          * at this point we have a head with no other entries.  Go
5076          * ahead and process it.
5077          */
5078         head->node.in_tree = 0;
5079         rb_erase(&head->node.rb_node, &delayed_refs->root);
5080
5081         delayed_refs->num_entries--;
5082         if (waitqueue_active(&delayed_refs->seq_wait))
5083                 wake_up(&delayed_refs->seq_wait);
5084
5085         /*
5086          * we don't take a ref on the node because we're removing it from the
5087          * tree, so we just steal the ref the tree was holding.
5088          */
5089         delayed_refs->num_heads--;
5090         if (list_empty(&head->cluster))
5091                 delayed_refs->num_heads_ready--;
5092
5093         list_del_init(&head->cluster);
5094         spin_unlock(&delayed_refs->lock);
5095
5096         BUG_ON(head->extent_op);
5097         if (head->must_insert_reserved)
5098                 ret = 1;
5099
5100         mutex_unlock(&head->mutex);
5101         btrfs_put_delayed_ref(&head->node);
5102         return ret;
5103 out:
5104         spin_unlock(&delayed_refs->lock);
5105         return 0;
5106 }
5107
5108 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5109                            struct btrfs_root *root,
5110                            struct extent_buffer *buf,
5111                            u64 parent, int last_ref, int for_cow)
5112 {
5113         struct btrfs_block_group_cache *cache = NULL;
5114         int ret;
5115
5116         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5117                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5118                                         buf->start, buf->len,
5119                                         parent, root->root_key.objectid,
5120                                         btrfs_header_level(buf),
5121                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5122                 BUG_ON(ret);
5123         }
5124
5125         if (!last_ref)
5126                 return;
5127
5128         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5129
5130         if (btrfs_header_generation(buf) == trans->transid) {
5131                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5132                         ret = check_ref_cleanup(trans, root, buf->start);
5133                         if (!ret)
5134                                 goto out;
5135                 }
5136
5137                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5138                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5139                         goto out;
5140                 }
5141
5142                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5143
5144                 btrfs_add_free_space(cache, buf->start, buf->len);
5145                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5146         }
5147 out:
5148         /*
5149          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5150          * anymore.
5151          */
5152         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5153         btrfs_put_block_group(cache);
5154 }
5155
5156 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5157                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5158                       u64 owner, u64 offset, int for_cow)
5159 {
5160         int ret;
5161         struct btrfs_fs_info *fs_info = root->fs_info;
5162
5163         /*
5164          * tree log blocks never actually go into the extent allocation
5165          * tree, just update pinning info and exit early.
5166          */
5167         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5168                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5169                 /* unlocks the pinned mutex */
5170                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5171                 ret = 0;
5172         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5173                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5174                                         num_bytes,
5175                                         parent, root_objectid, (int)owner,
5176                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5177                 BUG_ON(ret);
5178         } else {
5179                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5180                                                 num_bytes,
5181                                                 parent, root_objectid, owner,
5182                                                 offset, BTRFS_DROP_DELAYED_REF,
5183                                                 NULL, for_cow);
5184                 BUG_ON(ret);
5185         }
5186         return ret;
5187 }
5188
5189 static u64 stripe_align(struct btrfs_root *root, u64 val)
5190 {
5191         u64 mask = ((u64)root->stripesize - 1);
5192         u64 ret = (val + mask) & ~mask;
5193         return ret;
5194 }
5195
5196 /*
5197  * when we wait for progress in the block group caching, its because
5198  * our allocation attempt failed at least once.  So, we must sleep
5199  * and let some progress happen before we try again.
5200  *
5201  * This function will sleep at least once waiting for new free space to
5202  * show up, and then it will check the block group free space numbers
5203  * for our min num_bytes.  Another option is to have it go ahead
5204  * and look in the rbtree for a free extent of a given size, but this
5205  * is a good start.
5206  */
5207 static noinline int
5208 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5209                                 u64 num_bytes)
5210 {
5211         struct btrfs_caching_control *caching_ctl;
5212         DEFINE_WAIT(wait);
5213
5214         caching_ctl = get_caching_control(cache);
5215         if (!caching_ctl)
5216                 return 0;
5217
5218         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5219                    (cache->free_space_ctl->free_space >= num_bytes));
5220
5221         put_caching_control(caching_ctl);
5222         return 0;
5223 }
5224
5225 static noinline int
5226 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5227 {
5228         struct btrfs_caching_control *caching_ctl;
5229         DEFINE_WAIT(wait);
5230
5231         caching_ctl = get_caching_control(cache);
5232         if (!caching_ctl)
5233                 return 0;
5234
5235         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5236
5237         put_caching_control(caching_ctl);
5238         return 0;
5239 }
5240
5241 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5242 {
5243         int index;
5244         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
5245                 index = 0;
5246         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
5247                 index = 1;
5248         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
5249                 index = 2;
5250         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
5251                 index = 3;
5252         else
5253                 index = 4;
5254         return index;
5255 }
5256
5257 enum btrfs_loop_type {
5258         LOOP_FIND_IDEAL = 0,
5259         LOOP_CACHING_NOWAIT = 1,
5260         LOOP_CACHING_WAIT = 2,
5261         LOOP_ALLOC_CHUNK = 3,
5262         LOOP_NO_EMPTY_SIZE = 4,
5263 };
5264
5265 /*
5266  * walks the btree of allocated extents and find a hole of a given size.
5267  * The key ins is changed to record the hole:
5268  * ins->objectid == block start
5269  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5270  * ins->offset == number of blocks
5271  * Any available blocks before search_start are skipped.
5272  */
5273 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5274                                      struct btrfs_root *orig_root,
5275                                      u64 num_bytes, u64 empty_size,
5276                                      u64 search_start, u64 search_end,
5277                                      u64 hint_byte, struct btrfs_key *ins,
5278                                      u64 data)
5279 {
5280         int ret = 0;
5281         struct btrfs_root *root = orig_root->fs_info->extent_root;
5282         struct btrfs_free_cluster *last_ptr = NULL;
5283         struct btrfs_block_group_cache *block_group = NULL;
5284         struct btrfs_block_group_cache *used_block_group;
5285         int empty_cluster = 2 * 1024 * 1024;
5286         int allowed_chunk_alloc = 0;
5287         int done_chunk_alloc = 0;
5288         struct btrfs_space_info *space_info;
5289         int loop = 0;
5290         int index = 0;
5291         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5292                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5293         bool found_uncached_bg = false;
5294         bool failed_cluster_refill = false;
5295         bool failed_alloc = false;
5296         bool use_cluster = true;
5297         bool have_caching_bg = false;
5298         u64 ideal_cache_percent = 0;
5299         u64 ideal_cache_offset = 0;
5300
5301         WARN_ON(num_bytes < root->sectorsize);
5302         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5303         ins->objectid = 0;
5304         ins->offset = 0;
5305
5306         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5307
5308         space_info = __find_space_info(root->fs_info, data);
5309         if (!space_info) {
5310                 printk(KERN_ERR "No space info for %llu\n", data);
5311                 return -ENOSPC;
5312         }
5313
5314         /*
5315          * If the space info is for both data and metadata it means we have a
5316          * small filesystem and we can't use the clustering stuff.
5317          */
5318         if (btrfs_mixed_space_info(space_info))
5319                 use_cluster = false;
5320
5321         if (orig_root->ref_cows || empty_size)
5322                 allowed_chunk_alloc = 1;
5323
5324         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5325                 last_ptr = &root->fs_info->meta_alloc_cluster;
5326                 if (!btrfs_test_opt(root, SSD))
5327                         empty_cluster = 64 * 1024;
5328         }
5329
5330         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5331             btrfs_test_opt(root, SSD)) {
5332                 last_ptr = &root->fs_info->data_alloc_cluster;
5333         }
5334
5335         if (last_ptr) {
5336                 spin_lock(&last_ptr->lock);
5337                 if (last_ptr->block_group)
5338                         hint_byte = last_ptr->window_start;
5339                 spin_unlock(&last_ptr->lock);
5340         }
5341
5342         search_start = max(search_start, first_logical_byte(root, 0));
5343         search_start = max(search_start, hint_byte);
5344
5345         if (!last_ptr)
5346                 empty_cluster = 0;
5347
5348         if (search_start == hint_byte) {
5349 ideal_cache:
5350                 block_group = btrfs_lookup_block_group(root->fs_info,
5351                                                        search_start);
5352                 used_block_group = block_group;
5353                 /*
5354                  * we don't want to use the block group if it doesn't match our
5355                  * allocation bits, or if its not cached.
5356                  *
5357                  * However if we are re-searching with an ideal block group
5358                  * picked out then we don't care that the block group is cached.
5359                  */
5360                 if (block_group && block_group_bits(block_group, data) &&
5361                     (block_group->cached != BTRFS_CACHE_NO ||
5362                      search_start == ideal_cache_offset)) {
5363                         down_read(&space_info->groups_sem);
5364                         if (list_empty(&block_group->list) ||
5365                             block_group->ro) {
5366                                 /*
5367                                  * someone is removing this block group,
5368                                  * we can't jump into the have_block_group
5369                                  * target because our list pointers are not
5370                                  * valid
5371                                  */
5372                                 btrfs_put_block_group(block_group);
5373                                 up_read(&space_info->groups_sem);
5374                         } else {
5375                                 index = get_block_group_index(block_group);
5376                                 goto have_block_group;
5377                         }
5378                 } else if (block_group) {
5379                         btrfs_put_block_group(block_group);
5380                 }
5381         }
5382 search:
5383         have_caching_bg = false;
5384         down_read(&space_info->groups_sem);
5385         list_for_each_entry(block_group, &space_info->block_groups[index],
5386                             list) {
5387                 u64 offset;
5388                 int cached;
5389
5390                 used_block_group = block_group;
5391                 btrfs_get_block_group(block_group);
5392                 search_start = block_group->key.objectid;
5393
5394                 /*
5395                  * this can happen if we end up cycling through all the
5396                  * raid types, but we want to make sure we only allocate
5397                  * for the proper type.
5398                  */
5399                 if (!block_group_bits(block_group, data)) {
5400                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5401                                 BTRFS_BLOCK_GROUP_RAID1 |
5402                                 BTRFS_BLOCK_GROUP_RAID10;
5403
5404                         /*
5405                          * if they asked for extra copies and this block group
5406                          * doesn't provide them, bail.  This does allow us to
5407                          * fill raid0 from raid1.
5408                          */
5409                         if ((data & extra) && !(block_group->flags & extra))
5410                                 goto loop;
5411                 }
5412
5413 have_block_group:
5414                 cached = block_group_cache_done(block_group);
5415                 if (unlikely(!cached)) {
5416                         u64 free_percent;
5417
5418                         found_uncached_bg = true;
5419                         ret = cache_block_group(block_group, trans,
5420                                                 orig_root, 1);
5421                         if (block_group->cached == BTRFS_CACHE_FINISHED)
5422                                 goto alloc;
5423
5424                         free_percent = btrfs_block_group_used(&block_group->item);
5425                         free_percent *= 100;
5426                         free_percent = div64_u64(free_percent,
5427                                                  block_group->key.offset);
5428                         free_percent = 100 - free_percent;
5429                         if (free_percent > ideal_cache_percent &&
5430                             likely(!block_group->ro)) {
5431                                 ideal_cache_offset = block_group->key.objectid;
5432                                 ideal_cache_percent = free_percent;
5433                         }
5434
5435                         /*
5436                          * The caching workers are limited to 2 threads, so we
5437                          * can queue as much work as we care to.
5438                          */
5439                         if (loop > LOOP_FIND_IDEAL) {
5440                                 ret = cache_block_group(block_group, trans,
5441                                                         orig_root, 0);
5442                                 BUG_ON(ret);
5443                         }
5444
5445                         /*
5446                          * If loop is set for cached only, try the next block
5447                          * group.
5448                          */
5449                         if (loop == LOOP_FIND_IDEAL)
5450                                 goto loop;
5451                 }
5452
5453 alloc:
5454                 if (unlikely(block_group->ro))
5455                         goto loop;
5456
5457                 /*
5458                  * Ok we want to try and use the cluster allocator, so
5459                  * lets look there
5460                  */
5461                 if (last_ptr) {
5462                         /*
5463                          * the refill lock keeps out other
5464                          * people trying to start a new cluster
5465                          */
5466                         spin_lock(&last_ptr->refill_lock);
5467                         used_block_group = last_ptr->block_group;
5468                         if (used_block_group != block_group &&
5469                             (!used_block_group ||
5470                              used_block_group->ro ||
5471                              !block_group_bits(used_block_group, data))) {
5472                                 used_block_group = block_group;
5473                                 goto refill_cluster;
5474                         }
5475
5476                         if (used_block_group != block_group)
5477                                 btrfs_get_block_group(used_block_group);
5478
5479                         offset = btrfs_alloc_from_cluster(used_block_group,
5480                           last_ptr, num_bytes, used_block_group->key.objectid);
5481                         if (offset) {
5482                                 /* we have a block, we're done */
5483                                 spin_unlock(&last_ptr->refill_lock);
5484                                 trace_btrfs_reserve_extent_cluster(root,
5485                                         block_group, search_start, num_bytes);
5486                                 goto checks;
5487                         }
5488
5489                         WARN_ON(last_ptr->block_group != used_block_group);
5490                         if (used_block_group != block_group) {
5491                                 btrfs_put_block_group(used_block_group);
5492                                 used_block_group = block_group;
5493                         }
5494 refill_cluster:
5495                         BUG_ON(used_block_group != block_group);
5496                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5497                          * set up a new clusters, so lets just skip it
5498                          * and let the allocator find whatever block
5499                          * it can find.  If we reach this point, we
5500                          * will have tried the cluster allocator
5501                          * plenty of times and not have found
5502                          * anything, so we are likely way too
5503                          * fragmented for the clustering stuff to find
5504                          * anything.
5505                          *
5506                          * However, if the cluster is taken from the
5507                          * current block group, release the cluster
5508                          * first, so that we stand a better chance of
5509                          * succeeding in the unclustered
5510                          * allocation.  */
5511                         if (loop >= LOOP_NO_EMPTY_SIZE &&
5512                             last_ptr->block_group != block_group) {
5513                                 spin_unlock(&last_ptr->refill_lock);
5514                                 goto unclustered_alloc;
5515                         }
5516
5517                         /*
5518                          * this cluster didn't work out, free it and
5519                          * start over
5520                          */
5521                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5522
5523                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5524                                 spin_unlock(&last_ptr->refill_lock);
5525                                 goto unclustered_alloc;
5526                         }
5527
5528                         /* allocate a cluster in this block group */
5529                         ret = btrfs_find_space_cluster(trans, root,
5530                                                block_group, last_ptr,
5531                                                search_start, num_bytes,
5532                                                empty_cluster + empty_size);
5533                         if (ret == 0) {
5534                                 /*
5535                                  * now pull our allocation out of this
5536                                  * cluster
5537                                  */
5538                                 offset = btrfs_alloc_from_cluster(block_group,
5539                                                   last_ptr, num_bytes,
5540                                                   search_start);
5541                                 if (offset) {
5542                                         /* we found one, proceed */
5543                                         spin_unlock(&last_ptr->refill_lock);
5544                                         trace_btrfs_reserve_extent_cluster(root,
5545                                                 block_group, search_start,
5546                                                 num_bytes);
5547                                         goto checks;
5548                                 }
5549                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5550                                    && !failed_cluster_refill) {
5551                                 spin_unlock(&last_ptr->refill_lock);
5552
5553                                 failed_cluster_refill = true;
5554                                 wait_block_group_cache_progress(block_group,
5555                                        num_bytes + empty_cluster + empty_size);
5556                                 goto have_block_group;
5557                         }
5558
5559                         /*
5560                          * at this point we either didn't find a cluster
5561                          * or we weren't able to allocate a block from our
5562                          * cluster.  Free the cluster we've been trying
5563                          * to use, and go to the next block group
5564                          */
5565                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5566                         spin_unlock(&last_ptr->refill_lock);
5567                         goto loop;
5568                 }
5569
5570 unclustered_alloc:
5571                 spin_lock(&block_group->free_space_ctl->tree_lock);
5572                 if (cached &&
5573                     block_group->free_space_ctl->free_space <
5574                     num_bytes + empty_cluster + empty_size) {
5575                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5576                         goto loop;
5577                 }
5578                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5579
5580                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5581                                                     num_bytes, empty_size);
5582                 /*
5583                  * If we didn't find a chunk, and we haven't failed on this
5584                  * block group before, and this block group is in the middle of
5585                  * caching and we are ok with waiting, then go ahead and wait
5586                  * for progress to be made, and set failed_alloc to true.
5587                  *
5588                  * If failed_alloc is true then we've already waited on this
5589                  * block group once and should move on to the next block group.
5590                  */
5591                 if (!offset && !failed_alloc && !cached &&
5592                     loop > LOOP_CACHING_NOWAIT) {
5593                         wait_block_group_cache_progress(block_group,
5594                                                 num_bytes + empty_size);
5595                         failed_alloc = true;
5596                         goto have_block_group;
5597                 } else if (!offset) {
5598                         if (!cached)
5599                                 have_caching_bg = true;
5600                         goto loop;
5601                 }
5602 checks:
5603                 search_start = stripe_align(root, offset);
5604                 /* move on to the next group */
5605                 if (search_start + num_bytes >= search_end) {
5606                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5607                         goto loop;
5608                 }
5609
5610                 /* move on to the next group */
5611                 if (search_start + num_bytes >
5612                     used_block_group->key.objectid + used_block_group->key.offset) {
5613                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5614                         goto loop;
5615                 }
5616
5617                 if (offset < search_start)
5618                         btrfs_add_free_space(used_block_group, offset,
5619                                              search_start - offset);
5620                 BUG_ON(offset > search_start);
5621
5622                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
5623                                                   alloc_type);
5624                 if (ret == -EAGAIN) {
5625                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5626                         goto loop;
5627                 }
5628
5629                 /* we are all good, lets return */
5630                 ins->objectid = search_start;
5631                 ins->offset = num_bytes;
5632
5633                 trace_btrfs_reserve_extent(orig_root, block_group,
5634                                            search_start, num_bytes);
5635                 if (offset < search_start)
5636                         btrfs_add_free_space(used_block_group, offset,
5637                                              search_start - offset);
5638                 BUG_ON(offset > search_start);
5639                 if (used_block_group != block_group)
5640                         btrfs_put_block_group(used_block_group);
5641                 btrfs_put_block_group(block_group);
5642                 break;
5643 loop:
5644                 failed_cluster_refill = false;
5645                 failed_alloc = false;
5646                 BUG_ON(index != get_block_group_index(block_group));
5647                 if (used_block_group != block_group)
5648                         btrfs_put_block_group(used_block_group);
5649                 btrfs_put_block_group(block_group);
5650         }
5651         up_read(&space_info->groups_sem);
5652
5653         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
5654                 goto search;
5655
5656         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5657                 goto search;
5658
5659         /* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
5660          *                      for them to make caching progress.  Also
5661          *                      determine the best possible bg to cache
5662          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5663          *                      caching kthreads as we move along
5664          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5665          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5666          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5667          *                      again
5668          */
5669         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5670                 index = 0;
5671                 if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
5672                         found_uncached_bg = false;
5673                         loop++;
5674                         if (!ideal_cache_percent)
5675                                 goto search;
5676
5677                         /*
5678                          * 1 of the following 2 things have happened so far
5679                          *
5680                          * 1) We found an ideal block group for caching that
5681                          * is mostly full and will cache quickly, so we might
5682                          * as well wait for it.
5683                          *
5684                          * 2) We searched for cached only and we didn't find
5685                          * anything, and we didn't start any caching kthreads
5686                          * either, so chances are we will loop through and
5687                          * start a couple caching kthreads, and then come back
5688                          * around and just wait for them.  This will be slower
5689                          * because we will have 2 caching kthreads reading at
5690                          * the same time when we could have just started one
5691                          * and waited for it to get far enough to give us an
5692                          * allocation, so go ahead and go to the wait caching
5693                          * loop.
5694                          */
5695                         loop = LOOP_CACHING_WAIT;
5696                         search_start = ideal_cache_offset;
5697                         ideal_cache_percent = 0;
5698                         goto ideal_cache;
5699                 } else if (loop == LOOP_FIND_IDEAL) {
5700                         /*
5701                          * Didn't find a uncached bg, wait on anything we find
5702                          * next.
5703                          */
5704                         loop = LOOP_CACHING_WAIT;
5705                         goto search;
5706                 }
5707
5708                 loop++;
5709
5710                 if (loop == LOOP_ALLOC_CHUNK) {
5711                        if (allowed_chunk_alloc) {
5712                                 ret = do_chunk_alloc(trans, root, num_bytes +
5713                                                      2 * 1024 * 1024, data,
5714                                                      CHUNK_ALLOC_LIMITED);
5715                                 allowed_chunk_alloc = 0;
5716                                 if (ret == 1)
5717                                         done_chunk_alloc = 1;
5718                         } else if (!done_chunk_alloc &&
5719                                    space_info->force_alloc ==
5720                                    CHUNK_ALLOC_NO_FORCE) {
5721                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5722                         }
5723
5724                        /*
5725                         * We didn't allocate a chunk, go ahead and drop the
5726                         * empty size and loop again.
5727                         */
5728                        if (!done_chunk_alloc)
5729                                loop = LOOP_NO_EMPTY_SIZE;
5730                 }
5731
5732                 if (loop == LOOP_NO_EMPTY_SIZE) {
5733                         empty_size = 0;
5734                         empty_cluster = 0;
5735                 }
5736
5737                 goto search;
5738         } else if (!ins->objectid) {
5739                 ret = -ENOSPC;
5740         } else if (ins->objectid) {
5741                 ret = 0;
5742         }
5743
5744         return ret;
5745 }
5746
5747 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5748                             int dump_block_groups)
5749 {
5750         struct btrfs_block_group_cache *cache;
5751         int index = 0;
5752
5753         spin_lock(&info->lock);
5754         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5755                (unsigned long long)info->flags,
5756                (unsigned long long)(info->total_bytes - info->bytes_used -
5757                                     info->bytes_pinned - info->bytes_reserved -
5758                                     info->bytes_readonly),
5759                (info->full) ? "" : "not ");
5760         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5761                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5762                (unsigned long long)info->total_bytes,
5763                (unsigned long long)info->bytes_used,
5764                (unsigned long long)info->bytes_pinned,
5765                (unsigned long long)info->bytes_reserved,
5766                (unsigned long long)info->bytes_may_use,
5767                (unsigned long long)info->bytes_readonly);
5768         spin_unlock(&info->lock);
5769
5770         if (!dump_block_groups)
5771                 return;
5772
5773         down_read(&info->groups_sem);
5774 again:
5775         list_for_each_entry(cache, &info->block_groups[index], list) {
5776                 spin_lock(&cache->lock);
5777                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5778                        "%llu pinned %llu reserved\n",
5779                        (unsigned long long)cache->key.objectid,
5780                        (unsigned long long)cache->key.offset,
5781                        (unsigned long long)btrfs_block_group_used(&cache->item),
5782                        (unsigned long long)cache->pinned,
5783                        (unsigned long long)cache->reserved);
5784                 btrfs_dump_free_space(cache, bytes);
5785                 spin_unlock(&cache->lock);
5786         }
5787         if (++index < BTRFS_NR_RAID_TYPES)
5788                 goto again;
5789         up_read(&info->groups_sem);
5790 }
5791
5792 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5793                          struct btrfs_root *root,
5794                          u64 num_bytes, u64 min_alloc_size,
5795                          u64 empty_size, u64 hint_byte,
5796                          u64 search_end, struct btrfs_key *ins,
5797                          u64 data)
5798 {
5799         bool final_tried = false;
5800         int ret;
5801         u64 search_start = 0;
5802
5803         data = btrfs_get_alloc_profile(root, data);
5804 again:
5805         /*
5806          * the only place that sets empty_size is btrfs_realloc_node, which
5807          * is not called recursively on allocations
5808          */
5809         if (empty_size || root->ref_cows)
5810                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5811                                      num_bytes + 2 * 1024 * 1024, data,
5812                                      CHUNK_ALLOC_NO_FORCE);
5813
5814         WARN_ON(num_bytes < root->sectorsize);
5815         ret = find_free_extent(trans, root, num_bytes, empty_size,
5816                                search_start, search_end, hint_byte,
5817                                ins, data);
5818
5819         if (ret == -ENOSPC) {
5820                 if (!final_tried) {
5821                         num_bytes = num_bytes >> 1;
5822                         num_bytes = num_bytes & ~(root->sectorsize - 1);
5823                         num_bytes = max(num_bytes, min_alloc_size);
5824                         do_chunk_alloc(trans, root->fs_info->extent_root,
5825                                        num_bytes, data, CHUNK_ALLOC_FORCE);
5826                         if (num_bytes == min_alloc_size)
5827                                 final_tried = true;
5828                         goto again;
5829                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
5830                         struct btrfs_space_info *sinfo;
5831
5832                         sinfo = __find_space_info(root->fs_info, data);
5833                         printk(KERN_ERR "btrfs allocation failed flags %llu, "
5834                                "wanted %llu\n", (unsigned long long)data,
5835                                (unsigned long long)num_bytes);
5836                         if (sinfo)
5837                                 dump_space_info(sinfo, num_bytes, 1);
5838                 }
5839         }
5840
5841         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
5842
5843         return ret;
5844 }
5845
5846 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
5847                                         u64 start, u64 len, int pin)
5848 {
5849         struct btrfs_block_group_cache *cache;
5850         int ret = 0;
5851
5852         cache = btrfs_lookup_block_group(root->fs_info, start);
5853         if (!cache) {
5854                 printk(KERN_ERR "Unable to find block group for %llu\n",
5855                        (unsigned long long)start);
5856                 return -ENOSPC;
5857         }
5858
5859         if (btrfs_test_opt(root, DISCARD))
5860                 ret = btrfs_discard_extent(root, start, len, NULL);
5861
5862         if (pin)
5863                 pin_down_extent(root, cache, start, len, 1);
5864         else {
5865                 btrfs_add_free_space(cache, start, len);
5866                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
5867         }
5868         btrfs_put_block_group(cache);
5869
5870         trace_btrfs_reserved_extent_free(root, start, len);
5871
5872         return ret;
5873 }
5874
5875 int btrfs_free_reserved_extent(struct btrfs_root *root,
5876                                         u64 start, u64 len)
5877 {
5878         return __btrfs_free_reserved_extent(root, start, len, 0);
5879 }
5880
5881 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
5882                                        u64 start, u64 len)
5883 {
5884         return __btrfs_free_reserved_extent(root, start, len, 1);
5885 }
5886
5887 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5888                                       struct btrfs_root *root,
5889                                       u64 parent, u64 root_objectid,
5890                                       u64 flags, u64 owner, u64 offset,
5891                                       struct btrfs_key *ins, int ref_mod)
5892 {
5893         int ret;
5894         struct btrfs_fs_info *fs_info = root->fs_info;
5895         struct btrfs_extent_item *extent_item;
5896         struct btrfs_extent_inline_ref *iref;
5897         struct btrfs_path *path;
5898         struct extent_buffer *leaf;
5899         int type;
5900         u32 size;
5901
5902         if (parent > 0)
5903                 type = BTRFS_SHARED_DATA_REF_KEY;
5904         else
5905                 type = BTRFS_EXTENT_DATA_REF_KEY;
5906
5907         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5908
5909         path = btrfs_alloc_path();
5910         if (!path)
5911                 return -ENOMEM;
5912
5913         path->leave_spinning = 1;
5914         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5915                                       ins, size);
5916         BUG_ON(ret);
5917
5918         leaf = path->nodes[0];
5919         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5920                                      struct btrfs_extent_item);
5921         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
5922         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5923         btrfs_set_extent_flags(leaf, extent_item,
5924                                flags | BTRFS_EXTENT_FLAG_DATA);
5925
5926         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
5927         btrfs_set_extent_inline_ref_type(leaf, iref, type);
5928         if (parent > 0) {
5929                 struct btrfs_shared_data_ref *ref;
5930                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
5931                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5932                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
5933         } else {
5934                 struct btrfs_extent_data_ref *ref;
5935                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
5936                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
5937                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
5938                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
5939                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
5940         }
5941
5942         btrfs_mark_buffer_dirty(path->nodes[0]);
5943         btrfs_free_path(path);
5944
5945         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5946         if (ret) {
5947                 printk(KERN_ERR "btrfs update block group failed for %llu "
5948                        "%llu\n", (unsigned long long)ins->objectid,
5949                        (unsigned long long)ins->offset);
5950                 BUG();
5951         }
5952         return ret;
5953 }
5954
5955 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
5956                                      struct btrfs_root *root,
5957                                      u64 parent, u64 root_objectid,
5958                                      u64 flags, struct btrfs_disk_key *key,
5959                                      int level, struct btrfs_key *ins)
5960 {
5961         int ret;
5962         struct btrfs_fs_info *fs_info = root->fs_info;
5963         struct btrfs_extent_item *extent_item;
5964         struct btrfs_tree_block_info *block_info;
5965         struct btrfs_extent_inline_ref *iref;
5966         struct btrfs_path *path;
5967         struct extent_buffer *leaf;
5968         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
5969
5970         path = btrfs_alloc_path();
5971         if (!path)
5972                 return -ENOMEM;
5973
5974         path->leave_spinning = 1;
5975         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5976                                       ins, size);
5977         BUG_ON(ret);
5978
5979         leaf = path->nodes[0];
5980         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5981                                      struct btrfs_extent_item);
5982         btrfs_set_extent_refs(leaf, extent_item, 1);
5983         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5984         btrfs_set_extent_flags(leaf, extent_item,
5985                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
5986         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
5987
5988         btrfs_set_tree_block_key(leaf, block_info, key);
5989         btrfs_set_tree_block_level(leaf, block_info, level);
5990
5991         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
5992         if (parent > 0) {
5993                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
5994                 btrfs_set_extent_inline_ref_type(leaf, iref,
5995                                                  BTRFS_SHARED_BLOCK_REF_KEY);
5996                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5997         } else {
5998                 btrfs_set_extent_inline_ref_type(leaf, iref,
5999                                                  BTRFS_TREE_BLOCK_REF_KEY);
6000                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6001         }
6002
6003         btrfs_mark_buffer_dirty(leaf);
6004         btrfs_free_path(path);
6005
6006         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
6007         if (ret) {
6008                 printk(KERN_ERR "btrfs update block group failed for %llu "
6009                        "%llu\n", (unsigned long long)ins->objectid,
6010                        (unsigned long long)ins->offset);
6011                 BUG();
6012         }
6013         return ret;
6014 }
6015
6016 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6017                                      struct btrfs_root *root,
6018                                      u64 root_objectid, u64 owner,
6019                                      u64 offset, struct btrfs_key *ins)
6020 {
6021         int ret;
6022
6023         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6024
6025         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6026                                          ins->offset, 0,
6027                                          root_objectid, owner, offset,
6028                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6029         return ret;
6030 }
6031
6032 /*
6033  * this is used by the tree logging recovery code.  It records that
6034  * an extent has been allocated and makes sure to clear the free
6035  * space cache bits as well
6036  */
6037 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6038                                    struct btrfs_root *root,
6039                                    u64 root_objectid, u64 owner, u64 offset,
6040                                    struct btrfs_key *ins)
6041 {
6042         int ret;
6043         struct btrfs_block_group_cache *block_group;
6044         struct btrfs_caching_control *caching_ctl;
6045         u64 start = ins->objectid;
6046         u64 num_bytes = ins->offset;
6047
6048         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6049         cache_block_group(block_group, trans, NULL, 0);
6050         caching_ctl = get_caching_control(block_group);
6051
6052         if (!caching_ctl) {
6053                 BUG_ON(!block_group_cache_done(block_group));
6054                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6055                 BUG_ON(ret);
6056         } else {
6057                 mutex_lock(&caching_ctl->mutex);
6058
6059                 if (start >= caching_ctl->progress) {
6060                         ret = add_excluded_extent(root, start, num_bytes);
6061                         BUG_ON(ret);
6062                 } else if (start + num_bytes <= caching_ctl->progress) {
6063                         ret = btrfs_remove_free_space(block_group,
6064                                                       start, num_bytes);
6065                         BUG_ON(ret);
6066                 } else {
6067                         num_bytes = caching_ctl->progress - start;
6068                         ret = btrfs_remove_free_space(block_group,
6069                                                       start, num_bytes);
6070                         BUG_ON(ret);
6071
6072                         start = caching_ctl->progress;
6073                         num_bytes = ins->objectid + ins->offset -
6074                                     caching_ctl->progress;
6075                         ret = add_excluded_extent(root, start, num_bytes);
6076                         BUG_ON(ret);
6077                 }
6078
6079                 mutex_unlock(&caching_ctl->mutex);
6080                 put_caching_control(caching_ctl);
6081         }
6082
6083         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6084                                           RESERVE_ALLOC_NO_ACCOUNT);
6085         BUG_ON(ret);
6086         btrfs_put_block_group(block_group);
6087         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6088                                          0, owner, offset, ins, 1);
6089         return ret;
6090 }
6091
6092 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
6093                                             struct btrfs_root *root,
6094                                             u64 bytenr, u32 blocksize,
6095                                             int level)
6096 {
6097         struct extent_buffer *buf;
6098
6099         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6100         if (!buf)
6101                 return ERR_PTR(-ENOMEM);
6102         btrfs_set_header_generation(buf, trans->transid);
6103         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6104         btrfs_tree_lock(buf);
6105         clean_tree_block(trans, root, buf);
6106
6107         btrfs_set_lock_blocking(buf);
6108         btrfs_set_buffer_uptodate(buf);
6109
6110         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6111                 /*
6112                  * we allow two log transactions at a time, use different
6113                  * EXENT bit to differentiate dirty pages.
6114                  */
6115                 if (root->log_transid % 2 == 0)
6116                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6117                                         buf->start + buf->len - 1, GFP_NOFS);
6118                 else
6119                         set_extent_new(&root->dirty_log_pages, buf->start,
6120                                         buf->start + buf->len - 1, GFP_NOFS);
6121         } else {
6122                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6123                          buf->start + buf->len - 1, GFP_NOFS);
6124         }
6125         trans->blocks_used++;
6126         /* this returns a buffer locked for blocking */
6127         return buf;
6128 }
6129
6130 static struct btrfs_block_rsv *
6131 use_block_rsv(struct btrfs_trans_handle *trans,
6132               struct btrfs_root *root, u32 blocksize)
6133 {
6134         struct btrfs_block_rsv *block_rsv;
6135         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6136         int ret;
6137
6138         block_rsv = get_block_rsv(trans, root);
6139
6140         if (block_rsv->size == 0) {
6141                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
6142                 /*
6143                  * If we couldn't reserve metadata bytes try and use some from
6144                  * the global reserve.
6145                  */
6146                 if (ret && block_rsv != global_rsv) {
6147                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6148                         if (!ret)
6149                                 return global_rsv;
6150                         return ERR_PTR(ret);
6151                 } else if (ret) {
6152                         return ERR_PTR(ret);
6153                 }
6154                 return block_rsv;
6155         }
6156
6157         ret = block_rsv_use_bytes(block_rsv, blocksize);
6158         if (!ret)
6159                 return block_rsv;
6160         if (ret) {
6161                 static DEFINE_RATELIMIT_STATE(_rs,
6162                                 DEFAULT_RATELIMIT_INTERVAL,
6163                                 /*DEFAULT_RATELIMIT_BURST*/ 2);
6164                 if (__ratelimit(&_rs)) {
6165                         printk(KERN_DEBUG "btrfs: block rsv returned %d\n", ret);
6166                         WARN_ON(1);
6167                 }
6168                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
6169                 if (!ret) {
6170                         return block_rsv;
6171                 } else if (ret && block_rsv != global_rsv) {
6172                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6173                         if (!ret)
6174                                 return global_rsv;
6175                 }
6176         }
6177
6178         return ERR_PTR(-ENOSPC);
6179 }
6180
6181 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6182                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6183 {
6184         block_rsv_add_bytes(block_rsv, blocksize, 0);
6185         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6186 }
6187
6188 /*
6189  * finds a free extent and does all the dirty work required for allocation
6190  * returns the key for the extent through ins, and a tree buffer for
6191  * the first block of the extent through buf.
6192  *
6193  * returns the tree buffer or NULL.
6194  */
6195 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6196                                         struct btrfs_root *root, u32 blocksize,
6197                                         u64 parent, u64 root_objectid,
6198                                         struct btrfs_disk_key *key, int level,
6199                                         u64 hint, u64 empty_size, int for_cow)
6200 {
6201         struct btrfs_key ins;
6202         struct btrfs_block_rsv *block_rsv;
6203         struct extent_buffer *buf;
6204         u64 flags = 0;
6205         int ret;
6206
6207
6208         block_rsv = use_block_rsv(trans, root, blocksize);
6209         if (IS_ERR(block_rsv))
6210                 return ERR_CAST(block_rsv);
6211
6212         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6213                                    empty_size, hint, (u64)-1, &ins, 0);
6214         if (ret) {
6215                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6216                 return ERR_PTR(ret);
6217         }
6218
6219         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6220                                     blocksize, level);
6221         BUG_ON(IS_ERR(buf));
6222
6223         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6224                 if (parent == 0)
6225                         parent = ins.objectid;
6226                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6227         } else
6228                 BUG_ON(parent > 0);
6229
6230         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6231                 struct btrfs_delayed_extent_op *extent_op;
6232                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
6233                 BUG_ON(!extent_op);
6234                 if (key)
6235                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6236                 else
6237                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6238                 extent_op->flags_to_set = flags;
6239                 extent_op->update_key = 1;
6240                 extent_op->update_flags = 1;
6241                 extent_op->is_data = 0;
6242
6243                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6244                                         ins.objectid,
6245                                         ins.offset, parent, root_objectid,
6246                                         level, BTRFS_ADD_DELAYED_EXTENT,
6247                                         extent_op, for_cow);
6248                 BUG_ON(ret);
6249         }
6250         return buf;
6251 }
6252
6253 struct walk_control {
6254         u64 refs[BTRFS_MAX_LEVEL];
6255         u64 flags[BTRFS_MAX_LEVEL];
6256         struct btrfs_key update_progress;
6257         int stage;
6258         int level;
6259         int shared_level;
6260         int update_ref;
6261         int keep_locks;
6262         int reada_slot;
6263         int reada_count;
6264         int for_reloc;
6265 };
6266
6267 #define DROP_REFERENCE  1
6268 #define UPDATE_BACKREF  2
6269
6270 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6271                                      struct btrfs_root *root,
6272                                      struct walk_control *wc,
6273                                      struct btrfs_path *path)
6274 {
6275         u64 bytenr;
6276         u64 generation;
6277         u64 refs;
6278         u64 flags;
6279         u32 nritems;
6280         u32 blocksize;
6281         struct btrfs_key key;
6282         struct extent_buffer *eb;
6283         int ret;
6284         int slot;
6285         int nread = 0;
6286
6287         if (path->slots[wc->level] < wc->reada_slot) {
6288                 wc->reada_count = wc->reada_count * 2 / 3;
6289                 wc->reada_count = max(wc->reada_count, 2);
6290         } else {
6291                 wc->reada_count = wc->reada_count * 3 / 2;
6292                 wc->reada_count = min_t(int, wc->reada_count,
6293                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6294         }
6295
6296         eb = path->nodes[wc->level];
6297         nritems = btrfs_header_nritems(eb);
6298         blocksize = btrfs_level_size(root, wc->level - 1);
6299
6300         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6301                 if (nread >= wc->reada_count)
6302                         break;
6303
6304                 cond_resched();
6305                 bytenr = btrfs_node_blockptr(eb, slot);
6306                 generation = btrfs_node_ptr_generation(eb, slot);
6307
6308                 if (slot == path->slots[wc->level])
6309                         goto reada;
6310
6311                 if (wc->stage == UPDATE_BACKREF &&
6312                     generation <= root->root_key.offset)
6313                         continue;
6314
6315                 /* We don't lock the tree block, it's OK to be racy here */
6316                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6317                                                &refs, &flags);
6318                 BUG_ON(ret);
6319                 BUG_ON(refs == 0);
6320
6321                 if (wc->stage == DROP_REFERENCE) {
6322                         if (refs == 1)
6323                                 goto reada;
6324
6325                         if (wc->level == 1 &&
6326                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6327                                 continue;
6328                         if (!wc->update_ref ||
6329                             generation <= root->root_key.offset)
6330                                 continue;
6331                         btrfs_node_key_to_cpu(eb, &key, slot);
6332                         ret = btrfs_comp_cpu_keys(&key,
6333                                                   &wc->update_progress);
6334                         if (ret < 0)
6335                                 continue;
6336                 } else {
6337                         if (wc->level == 1 &&
6338                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6339                                 continue;
6340                 }
6341 reada:
6342                 ret = readahead_tree_block(root, bytenr, blocksize,
6343                                            generation);
6344                 if (ret)
6345                         break;
6346                 nread++;
6347         }
6348         wc->reada_slot = slot;
6349 }
6350
6351 /*
6352  * hepler to process tree block while walking down the tree.
6353  *
6354  * when wc->stage == UPDATE_BACKREF, this function updates
6355  * back refs for pointers in the block.
6356  *
6357  * NOTE: return value 1 means we should stop walking down.
6358  */
6359 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6360                                    struct btrfs_root *root,
6361                                    struct btrfs_path *path,
6362                                    struct walk_control *wc, int lookup_info)
6363 {
6364         int level = wc->level;
6365         struct extent_buffer *eb = path->nodes[level];
6366         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6367         int ret;
6368
6369         if (wc->stage == UPDATE_BACKREF &&
6370             btrfs_header_owner(eb) != root->root_key.objectid)
6371                 return 1;
6372
6373         /*
6374          * when reference count of tree block is 1, it won't increase
6375          * again. once full backref flag is set, we never clear it.
6376          */
6377         if (lookup_info &&
6378             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6379              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6380                 BUG_ON(!path->locks[level]);
6381                 ret = btrfs_lookup_extent_info(trans, root,
6382                                                eb->start, eb->len,
6383                                                &wc->refs[level],
6384                                                &wc->flags[level]);
6385                 BUG_ON(ret);
6386                 BUG_ON(wc->refs[level] == 0);
6387         }
6388
6389         if (wc->stage == DROP_REFERENCE) {
6390                 if (wc->refs[level] > 1)
6391                         return 1;
6392
6393                 if (path->locks[level] && !wc->keep_locks) {
6394                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6395                         path->locks[level] = 0;
6396                 }
6397                 return 0;
6398         }
6399
6400         /* wc->stage == UPDATE_BACKREF */
6401         if (!(wc->flags[level] & flag)) {
6402                 BUG_ON(!path->locks[level]);
6403                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
6404                 BUG_ON(ret);
6405                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
6406                 BUG_ON(ret);
6407                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6408                                                   eb->len, flag, 0);
6409                 BUG_ON(ret);
6410                 wc->flags[level] |= flag;
6411         }
6412
6413         /*
6414          * the block is shared by multiple trees, so it's not good to
6415          * keep the tree lock
6416          */
6417         if (path->locks[level] && level > 0) {
6418                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6419                 path->locks[level] = 0;
6420         }
6421         return 0;
6422 }
6423
6424 /*
6425  * hepler to process tree block pointer.
6426  *
6427  * when wc->stage == DROP_REFERENCE, this function checks
6428  * reference count of the block pointed to. if the block
6429  * is shared and we need update back refs for the subtree
6430  * rooted at the block, this function changes wc->stage to
6431  * UPDATE_BACKREF. if the block is shared and there is no
6432  * need to update back, this function drops the reference
6433  * to the block.
6434  *
6435  * NOTE: return value 1 means we should stop walking down.
6436  */
6437 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6438                                  struct btrfs_root *root,
6439                                  struct btrfs_path *path,
6440                                  struct walk_control *wc, int *lookup_info)
6441 {
6442         u64 bytenr;
6443         u64 generation;
6444         u64 parent;
6445         u32 blocksize;
6446         struct btrfs_key key;
6447         struct extent_buffer *next;
6448         int level = wc->level;
6449         int reada = 0;
6450         int ret = 0;
6451
6452         generation = btrfs_node_ptr_generation(path->nodes[level],
6453                                                path->slots[level]);
6454         /*
6455          * if the lower level block was created before the snapshot
6456          * was created, we know there is no need to update back refs
6457          * for the subtree
6458          */
6459         if (wc->stage == UPDATE_BACKREF &&
6460             generation <= root->root_key.offset) {
6461                 *lookup_info = 1;
6462                 return 1;
6463         }
6464
6465         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6466         blocksize = btrfs_level_size(root, level - 1);
6467
6468         next = btrfs_find_tree_block(root, bytenr, blocksize);
6469         if (!next) {
6470                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6471                 if (!next)
6472                         return -ENOMEM;
6473                 reada = 1;
6474         }
6475         btrfs_tree_lock(next);
6476         btrfs_set_lock_blocking(next);
6477
6478         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6479                                        &wc->refs[level - 1],
6480                                        &wc->flags[level - 1]);
6481         BUG_ON(ret);
6482         BUG_ON(wc->refs[level - 1] == 0);
6483         *lookup_info = 0;
6484
6485         if (wc->stage == DROP_REFERENCE) {
6486                 if (wc->refs[level - 1] > 1) {
6487                         if (level == 1 &&
6488                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6489                                 goto skip;
6490
6491                         if (!wc->update_ref ||
6492                             generation <= root->root_key.offset)
6493                                 goto skip;
6494
6495                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6496                                               path->slots[level]);
6497                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6498                         if (ret < 0)
6499                                 goto skip;
6500
6501                         wc->stage = UPDATE_BACKREF;
6502                         wc->shared_level = level - 1;
6503                 }
6504         } else {
6505                 if (level == 1 &&
6506                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6507                         goto skip;
6508         }
6509
6510         if (!btrfs_buffer_uptodate(next, generation)) {
6511                 btrfs_tree_unlock(next);
6512                 free_extent_buffer(next);
6513                 next = NULL;
6514                 *lookup_info = 1;
6515         }
6516
6517         if (!next) {
6518                 if (reada && level == 1)
6519                         reada_walk_down(trans, root, wc, path);
6520                 next = read_tree_block(root, bytenr, blocksize, generation);
6521                 if (!next)
6522                         return -EIO;
6523                 btrfs_tree_lock(next);
6524                 btrfs_set_lock_blocking(next);
6525         }
6526
6527         level--;
6528         BUG_ON(level != btrfs_header_level(next));
6529         path->nodes[level] = next;
6530         path->slots[level] = 0;
6531         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6532         wc->level = level;
6533         if (wc->level == 1)
6534                 wc->reada_slot = 0;
6535         return 0;
6536 skip:
6537         wc->refs[level - 1] = 0;
6538         wc->flags[level - 1] = 0;
6539         if (wc->stage == DROP_REFERENCE) {
6540                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6541                         parent = path->nodes[level]->start;
6542                 } else {
6543                         BUG_ON(root->root_key.objectid !=
6544                                btrfs_header_owner(path->nodes[level]));
6545                         parent = 0;
6546                 }
6547
6548                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6549                                 root->root_key.objectid, level - 1, 0, 0);
6550                 BUG_ON(ret);
6551         }
6552         btrfs_tree_unlock(next);
6553         free_extent_buffer(next);
6554         *lookup_info = 1;
6555         return 1;
6556 }
6557
6558 /*
6559  * hepler to process tree block while walking up the tree.
6560  *
6561  * when wc->stage == DROP_REFERENCE, this function drops
6562  * reference count on the block.
6563  *
6564  * when wc->stage == UPDATE_BACKREF, this function changes
6565  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6566  * to UPDATE_BACKREF previously while processing the block.
6567  *
6568  * NOTE: return value 1 means we should stop walking up.
6569  */
6570 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6571                                  struct btrfs_root *root,
6572                                  struct btrfs_path *path,
6573                                  struct walk_control *wc)
6574 {
6575         int ret;
6576         int level = wc->level;
6577         struct extent_buffer *eb = path->nodes[level];
6578         u64 parent = 0;
6579
6580         if (wc->stage == UPDATE_BACKREF) {
6581                 BUG_ON(wc->shared_level < level);
6582                 if (level < wc->shared_level)
6583                         goto out;
6584
6585                 ret = find_next_key(path, level + 1, &wc->update_progress);
6586                 if (ret > 0)
6587                         wc->update_ref = 0;
6588
6589                 wc->stage = DROP_REFERENCE;
6590                 wc->shared_level = -1;
6591                 path->slots[level] = 0;
6592
6593                 /*
6594                  * check reference count again if the block isn't locked.
6595                  * we should start walking down the tree again if reference
6596                  * count is one.
6597                  */
6598                 if (!path->locks[level]) {
6599                         BUG_ON(level == 0);
6600                         btrfs_tree_lock(eb);
6601                         btrfs_set_lock_blocking(eb);
6602                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6603
6604                         ret = btrfs_lookup_extent_info(trans, root,
6605                                                        eb->start, eb->len,
6606                                                        &wc->refs[level],
6607                                                        &wc->flags[level]);
6608                         BUG_ON(ret);
6609                         BUG_ON(wc->refs[level] == 0);
6610                         if (wc->refs[level] == 1) {
6611                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6612                                 return 1;
6613                         }
6614                 }
6615         }
6616
6617         /* wc->stage == DROP_REFERENCE */
6618         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6619
6620         if (wc->refs[level] == 1) {
6621                 if (level == 0) {
6622                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6623                                 ret = btrfs_dec_ref(trans, root, eb, 1,
6624                                                     wc->for_reloc);
6625                         else
6626                                 ret = btrfs_dec_ref(trans, root, eb, 0,
6627                                                     wc->for_reloc);
6628                         BUG_ON(ret);
6629                 }
6630                 /* make block locked assertion in clean_tree_block happy */
6631                 if (!path->locks[level] &&
6632                     btrfs_header_generation(eb) == trans->transid) {
6633                         btrfs_tree_lock(eb);
6634                         btrfs_set_lock_blocking(eb);
6635                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6636                 }
6637                 clean_tree_block(trans, root, eb);
6638         }
6639
6640         if (eb == root->node) {
6641                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6642                         parent = eb->start;
6643                 else
6644                         BUG_ON(root->root_key.objectid !=
6645                                btrfs_header_owner(eb));
6646         } else {
6647                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6648                         parent = path->nodes[level + 1]->start;
6649                 else
6650                         BUG_ON(root->root_key.objectid !=
6651                                btrfs_header_owner(path->nodes[level + 1]));
6652         }
6653
6654         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1, 0);
6655 out:
6656         wc->refs[level] = 0;
6657         wc->flags[level] = 0;
6658         return 0;
6659 }
6660
6661 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6662                                    struct btrfs_root *root,
6663                                    struct btrfs_path *path,
6664                                    struct walk_control *wc)
6665 {
6666         int level = wc->level;
6667         int lookup_info = 1;
6668         int ret;
6669
6670         while (level >= 0) {
6671                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6672                 if (ret > 0)
6673                         break;
6674
6675                 if (level == 0)
6676                         break;
6677
6678                 if (path->slots[level] >=
6679                     btrfs_header_nritems(path->nodes[level]))
6680                         break;
6681
6682                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6683                 if (ret > 0) {
6684                         path->slots[level]++;
6685                         continue;
6686                 } else if (ret < 0)
6687                         return ret;
6688                 level = wc->level;
6689         }
6690         return 0;
6691 }
6692
6693 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6694                                  struct btrfs_root *root,
6695                                  struct btrfs_path *path,
6696                                  struct walk_control *wc, int max_level)
6697 {
6698         int level = wc->level;
6699         int ret;
6700
6701         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6702         while (level < max_level && path->nodes[level]) {
6703                 wc->level = level;
6704                 if (path->slots[level] + 1 <
6705                     btrfs_header_nritems(path->nodes[level])) {
6706                         path->slots[level]++;
6707                         return 0;
6708                 } else {
6709                         ret = walk_up_proc(trans, root, path, wc);
6710                         if (ret > 0)
6711                                 return 0;
6712
6713                         if (path->locks[level]) {
6714                                 btrfs_tree_unlock_rw(path->nodes[level],
6715                                                      path->locks[level]);
6716                                 path->locks[level] = 0;
6717                         }
6718                         free_extent_buffer(path->nodes[level]);
6719                         path->nodes[level] = NULL;
6720                         level++;
6721                 }
6722         }
6723         return 1;
6724 }
6725
6726 /*
6727  * drop a subvolume tree.
6728  *
6729  * this function traverses the tree freeing any blocks that only
6730  * referenced by the tree.
6731  *
6732  * when a shared tree block is found. this function decreases its
6733  * reference count by one. if update_ref is true, this function
6734  * also make sure backrefs for the shared block and all lower level
6735  * blocks are properly updated.
6736  */
6737 void btrfs_drop_snapshot(struct btrfs_root *root,
6738                          struct btrfs_block_rsv *block_rsv, int update_ref,
6739                          int for_reloc)
6740 {
6741         struct btrfs_path *path;
6742         struct btrfs_trans_handle *trans;
6743         struct btrfs_root *tree_root = root->fs_info->tree_root;
6744         struct btrfs_root_item *root_item = &root->root_item;
6745         struct walk_control *wc;
6746         struct btrfs_key key;
6747         int err = 0;
6748         int ret;
6749         int level;
6750
6751         path = btrfs_alloc_path();
6752         if (!path) {
6753                 err = -ENOMEM;
6754                 goto out;
6755         }
6756
6757         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6758         if (!wc) {
6759                 btrfs_free_path(path);
6760                 err = -ENOMEM;
6761                 goto out;
6762         }
6763
6764         trans = btrfs_start_transaction(tree_root, 0);
6765         BUG_ON(IS_ERR(trans));
6766
6767         if (block_rsv)
6768                 trans->block_rsv = block_rsv;
6769
6770         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6771                 level = btrfs_header_level(root->node);
6772                 path->nodes[level] = btrfs_lock_root_node(root);
6773                 btrfs_set_lock_blocking(path->nodes[level]);
6774                 path->slots[level] = 0;
6775                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6776                 memset(&wc->update_progress, 0,
6777                        sizeof(wc->update_progress));
6778         } else {
6779                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6780                 memcpy(&wc->update_progress, &key,
6781                        sizeof(wc->update_progress));
6782
6783                 level = root_item->drop_level;
6784                 BUG_ON(level == 0);
6785                 path->lowest_level = level;
6786                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6787                 path->lowest_level = 0;
6788                 if (ret < 0) {
6789                         err = ret;
6790                         goto out_free;
6791                 }
6792                 WARN_ON(ret > 0);
6793
6794                 /*
6795                  * unlock our path, this is safe because only this
6796                  * function is allowed to delete this snapshot
6797                  */
6798                 btrfs_unlock_up_safe(path, 0);
6799
6800                 level = btrfs_header_level(root->node);
6801                 while (1) {
6802                         btrfs_tree_lock(path->nodes[level]);
6803                         btrfs_set_lock_blocking(path->nodes[level]);
6804
6805                         ret = btrfs_lookup_extent_info(trans, root,
6806                                                 path->nodes[level]->start,
6807                                                 path->nodes[level]->len,
6808                                                 &wc->refs[level],
6809                                                 &wc->flags[level]);
6810                         BUG_ON(ret);
6811                         BUG_ON(wc->refs[level] == 0);
6812
6813                         if (level == root_item->drop_level)
6814                                 break;
6815
6816                         btrfs_tree_unlock(path->nodes[level]);
6817                         WARN_ON(wc->refs[level] != 1);
6818                         level--;
6819                 }
6820         }
6821
6822         wc->level = level;
6823         wc->shared_level = -1;
6824         wc->stage = DROP_REFERENCE;
6825         wc->update_ref = update_ref;
6826         wc->keep_locks = 0;
6827         wc->for_reloc = for_reloc;
6828         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6829
6830         while (1) {
6831                 ret = walk_down_tree(trans, root, path, wc);
6832                 if (ret < 0) {
6833                         err = ret;
6834                         break;
6835                 }
6836
6837                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
6838                 if (ret < 0) {
6839                         err = ret;
6840                         break;
6841                 }
6842
6843                 if (ret > 0) {
6844                         BUG_ON(wc->stage != DROP_REFERENCE);
6845                         break;
6846                 }
6847
6848                 if (wc->stage == DROP_REFERENCE) {
6849                         level = wc->level;
6850                         btrfs_node_key(path->nodes[level],
6851                                        &root_item->drop_progress,
6852                                        path->slots[level]);
6853                         root_item->drop_level = level;
6854                 }
6855
6856                 BUG_ON(wc->level == 0);
6857                 if (btrfs_should_end_transaction(trans, tree_root)) {
6858                         ret = btrfs_update_root(trans, tree_root,
6859                                                 &root->root_key,
6860                                                 root_item);
6861                         BUG_ON(ret);
6862
6863                         btrfs_end_transaction_throttle(trans, tree_root);
6864                         trans = btrfs_start_transaction(tree_root, 0);
6865                         BUG_ON(IS_ERR(trans));
6866                         if (block_rsv)
6867                                 trans->block_rsv = block_rsv;
6868                 }
6869         }
6870         btrfs_release_path(path);
6871         BUG_ON(err);
6872
6873         ret = btrfs_del_root(trans, tree_root, &root->root_key);
6874         BUG_ON(ret);
6875
6876         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
6877                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
6878                                            NULL, NULL);
6879                 BUG_ON(ret < 0);
6880                 if (ret > 0) {
6881                         /* if we fail to delete the orphan item this time
6882                          * around, it'll get picked up the next time.
6883                          *
6884                          * The most common failure here is just -ENOENT.
6885                          */
6886                         btrfs_del_orphan_item(trans, tree_root,
6887                                               root->root_key.objectid);
6888                 }
6889         }
6890
6891         if (root->in_radix) {
6892                 btrfs_free_fs_root(tree_root->fs_info, root);
6893         } else {
6894                 free_extent_buffer(root->node);
6895                 free_extent_buffer(root->commit_root);
6896                 kfree(root);
6897         }
6898 out_free:
6899         btrfs_end_transaction_throttle(trans, tree_root);
6900         kfree(wc);
6901         btrfs_free_path(path);
6902 out:
6903         if (err)
6904                 btrfs_std_error(root->fs_info, err);
6905         return;
6906 }
6907
6908 /*
6909  * drop subtree rooted at tree block 'node'.
6910  *
6911  * NOTE: this function will unlock and release tree block 'node'
6912  * only used by relocation code
6913  */
6914 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
6915                         struct btrfs_root *root,
6916                         struct extent_buffer *node,
6917                         struct extent_buffer *parent)
6918 {
6919         struct btrfs_path *path;
6920         struct walk_control *wc;
6921         int level;
6922         int parent_level;
6923         int ret = 0;
6924         int wret;
6925
6926         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
6927
6928         path = btrfs_alloc_path();
6929         if (!path)
6930                 return -ENOMEM;
6931
6932         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6933         if (!wc) {
6934                 btrfs_free_path(path);
6935                 return -ENOMEM;
6936         }
6937
6938         btrfs_assert_tree_locked(parent);
6939         parent_level = btrfs_header_level(parent);
6940         extent_buffer_get(parent);
6941         path->nodes[parent_level] = parent;
6942         path->slots[parent_level] = btrfs_header_nritems(parent);
6943
6944         btrfs_assert_tree_locked(node);
6945         level = btrfs_header_level(node);
6946         path->nodes[level] = node;
6947         path->slots[level] = 0;
6948         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6949
6950         wc->refs[parent_level] = 1;
6951         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6952         wc->level = level;
6953         wc->shared_level = -1;
6954         wc->stage = DROP_REFERENCE;
6955         wc->update_ref = 0;
6956         wc->keep_locks = 1;
6957         wc->for_reloc = 1;
6958         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6959
6960         while (1) {
6961                 wret = walk_down_tree(trans, root, path, wc);
6962                 if (wret < 0) {
6963                         ret = wret;
6964                         break;
6965                 }
6966
6967                 wret = walk_up_tree(trans, root, path, wc, parent_level);
6968                 if (wret < 0)
6969                         ret = wret;
6970                 if (wret != 0)
6971                         break;
6972         }
6973
6974         kfree(wc);
6975         btrfs_free_path(path);
6976         return ret;
6977 }
6978
6979 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
6980 {
6981         u64 num_devices;
6982         u64 stripped = BTRFS_BLOCK_GROUP_RAID0 |
6983                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
6984
6985         if (root->fs_info->balance_ctl) {
6986                 struct btrfs_balance_control *bctl = root->fs_info->balance_ctl;
6987                 u64 tgt = 0;
6988
6989                 /* pick restriper's target profile and return */
6990                 if (flags & BTRFS_BLOCK_GROUP_DATA &&
6991                     bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
6992                         tgt = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
6993                 } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
6994                            bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
6995                         tgt = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
6996                 } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
6997                            bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
6998                         tgt = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
6999                 }
7000
7001                 if (tgt) {
7002                         /* extended -> chunk profile */
7003                         tgt &= ~BTRFS_AVAIL_ALLOC_BIT_SINGLE;
7004                         return tgt;
7005                 }
7006         }
7007
7008         /*
7009          * we add in the count of missing devices because we want
7010          * to make sure that any RAID levels on a degraded FS
7011          * continue to be honored.
7012          */
7013         num_devices = root->fs_info->fs_devices->rw_devices +
7014                 root->fs_info->fs_devices->missing_devices;
7015
7016         if (num_devices == 1) {
7017                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7018                 stripped = flags & ~stripped;
7019
7020                 /* turn raid0 into single device chunks */
7021                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7022                         return stripped;
7023
7024                 /* turn mirroring into duplication */
7025                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7026                              BTRFS_BLOCK_GROUP_RAID10))
7027                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7028                 return flags;
7029         } else {
7030                 /* they already had raid on here, just return */
7031                 if (flags & stripped)
7032                         return flags;
7033
7034                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7035                 stripped = flags & ~stripped;
7036
7037                 /* switch duplicated blocks with raid1 */
7038                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7039                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7040
7041                 /* turn single device chunks into raid0 */
7042                 return stripped | BTRFS_BLOCK_GROUP_RAID0;
7043         }
7044         return flags;
7045 }
7046
7047 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7048 {
7049         struct btrfs_space_info *sinfo = cache->space_info;
7050         u64 num_bytes;
7051         u64 min_allocable_bytes;
7052         int ret = -ENOSPC;
7053
7054
7055         /*
7056          * We need some metadata space and system metadata space for
7057          * allocating chunks in some corner cases until we force to set
7058          * it to be readonly.
7059          */
7060         if ((sinfo->flags &
7061              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7062             !force)
7063                 min_allocable_bytes = 1 * 1024 * 1024;
7064         else
7065                 min_allocable_bytes = 0;
7066
7067         spin_lock(&sinfo->lock);
7068         spin_lock(&cache->lock);
7069
7070         if (cache->ro) {
7071                 ret = 0;
7072                 goto out;
7073         }
7074
7075         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7076                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7077
7078         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7079             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7080             min_allocable_bytes <= sinfo->total_bytes) {
7081                 sinfo->bytes_readonly += num_bytes;
7082                 cache->ro = 1;
7083                 ret = 0;
7084         }
7085 out:
7086         spin_unlock(&cache->lock);
7087         spin_unlock(&sinfo->lock);
7088         return ret;
7089 }
7090
7091 int btrfs_set_block_group_ro(struct btrfs_root *root,
7092                              struct btrfs_block_group_cache *cache)
7093
7094 {
7095         struct btrfs_trans_handle *trans;
7096         u64 alloc_flags;
7097         int ret;
7098
7099         BUG_ON(cache->ro);
7100
7101         trans = btrfs_join_transaction(root);
7102         BUG_ON(IS_ERR(trans));
7103
7104         alloc_flags = update_block_group_flags(root, cache->flags);
7105         if (alloc_flags != cache->flags)
7106                 do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7107                                CHUNK_ALLOC_FORCE);
7108
7109         ret = set_block_group_ro(cache, 0);
7110         if (!ret)
7111                 goto out;
7112         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7113         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7114                              CHUNK_ALLOC_FORCE);
7115         if (ret < 0)
7116                 goto out;
7117         ret = set_block_group_ro(cache, 0);
7118 out:
7119         btrfs_end_transaction(trans, root);
7120         return ret;
7121 }
7122
7123 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7124                             struct btrfs_root *root, u64 type)
7125 {
7126         u64 alloc_flags = get_alloc_profile(root, type);
7127         return do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7128                               CHUNK_ALLOC_FORCE);
7129 }
7130
7131 /*
7132  * helper to account the unused space of all the readonly block group in the
7133  * list. takes mirrors into account.
7134  */
7135 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7136 {
7137         struct btrfs_block_group_cache *block_group;
7138         u64 free_bytes = 0;
7139         int factor;
7140
7141         list_for_each_entry(block_group, groups_list, list) {
7142                 spin_lock(&block_group->lock);
7143
7144                 if (!block_group->ro) {
7145                         spin_unlock(&block_group->lock);
7146                         continue;
7147                 }
7148
7149                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7150                                           BTRFS_BLOCK_GROUP_RAID10 |
7151                                           BTRFS_BLOCK_GROUP_DUP))
7152                         factor = 2;
7153                 else
7154                         factor = 1;
7155
7156                 free_bytes += (block_group->key.offset -
7157                                btrfs_block_group_used(&block_group->item)) *
7158                                factor;
7159
7160                 spin_unlock(&block_group->lock);
7161         }
7162
7163         return free_bytes;
7164 }
7165
7166 /*
7167  * helper to account the unused space of all the readonly block group in the
7168  * space_info. takes mirrors into account.
7169  */
7170 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7171 {
7172         int i;
7173         u64 free_bytes = 0;
7174
7175         spin_lock(&sinfo->lock);
7176
7177         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7178                 if (!list_empty(&sinfo->block_groups[i]))
7179                         free_bytes += __btrfs_get_ro_block_group_free_space(
7180                                                 &sinfo->block_groups[i]);
7181
7182         spin_unlock(&sinfo->lock);
7183
7184         return free_bytes;
7185 }
7186
7187 void btrfs_set_block_group_rw(struct btrfs_root *root,
7188                               struct btrfs_block_group_cache *cache)
7189 {
7190         struct btrfs_space_info *sinfo = cache->space_info;
7191         u64 num_bytes;
7192
7193         BUG_ON(!cache->ro);
7194
7195         spin_lock(&sinfo->lock);
7196         spin_lock(&cache->lock);
7197         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7198                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7199         sinfo->bytes_readonly -= num_bytes;
7200         cache->ro = 0;
7201         spin_unlock(&cache->lock);
7202         spin_unlock(&sinfo->lock);
7203 }
7204
7205 /*
7206  * checks to see if its even possible to relocate this block group.
7207  *
7208  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7209  * ok to go ahead and try.
7210  */
7211 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7212 {
7213         struct btrfs_block_group_cache *block_group;
7214         struct btrfs_space_info *space_info;
7215         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7216         struct btrfs_device *device;
7217         u64 min_free;
7218         u64 dev_min = 1;
7219         u64 dev_nr = 0;
7220         int index;
7221         int full = 0;
7222         int ret = 0;
7223
7224         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7225
7226         /* odd, couldn't find the block group, leave it alone */
7227         if (!block_group)
7228                 return -1;
7229
7230         min_free = btrfs_block_group_used(&block_group->item);
7231
7232         /* no bytes used, we're good */
7233         if (!min_free)
7234                 goto out;
7235
7236         space_info = block_group->space_info;
7237         spin_lock(&space_info->lock);
7238
7239         full = space_info->full;
7240
7241         /*
7242          * if this is the last block group we have in this space, we can't
7243          * relocate it unless we're able to allocate a new chunk below.
7244          *
7245          * Otherwise, we need to make sure we have room in the space to handle
7246          * all of the extents from this block group.  If we can, we're good
7247          */
7248         if ((space_info->total_bytes != block_group->key.offset) &&
7249             (space_info->bytes_used + space_info->bytes_reserved +
7250              space_info->bytes_pinned + space_info->bytes_readonly +
7251              min_free < space_info->total_bytes)) {
7252                 spin_unlock(&space_info->lock);
7253                 goto out;
7254         }
7255         spin_unlock(&space_info->lock);
7256
7257         /*
7258          * ok we don't have enough space, but maybe we have free space on our
7259          * devices to allocate new chunks for relocation, so loop through our
7260          * alloc devices and guess if we have enough space.  However, if we
7261          * were marked as full, then we know there aren't enough chunks, and we
7262          * can just return.
7263          */
7264         ret = -1;
7265         if (full)
7266                 goto out;
7267
7268         /*
7269          * index:
7270          *      0: raid10
7271          *      1: raid1
7272          *      2: dup
7273          *      3: raid0
7274          *      4: single
7275          */
7276         index = get_block_group_index(block_group);
7277         if (index == 0) {
7278                 dev_min = 4;
7279                 /* Divide by 2 */
7280                 min_free >>= 1;
7281         } else if (index == 1) {
7282                 dev_min = 2;
7283         } else if (index == 2) {
7284                 /* Multiply by 2 */
7285                 min_free <<= 1;
7286         } else if (index == 3) {
7287                 dev_min = fs_devices->rw_devices;
7288                 do_div(min_free, dev_min);
7289         }
7290
7291         mutex_lock(&root->fs_info->chunk_mutex);
7292         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7293                 u64 dev_offset;
7294
7295                 /*
7296                  * check to make sure we can actually find a chunk with enough
7297                  * space to fit our block group in.
7298                  */
7299                 if (device->total_bytes > device->bytes_used + min_free) {
7300                         ret = find_free_dev_extent(device, min_free,
7301                                                    &dev_offset, NULL);
7302                         if (!ret)
7303                                 dev_nr++;
7304
7305                         if (dev_nr >= dev_min)
7306                                 break;
7307
7308                         ret = -1;
7309                 }
7310         }
7311         mutex_unlock(&root->fs_info->chunk_mutex);
7312 out:
7313         btrfs_put_block_group(block_group);
7314         return ret;
7315 }
7316
7317 static int find_first_block_group(struct btrfs_root *root,
7318                 struct btrfs_path *path, struct btrfs_key *key)
7319 {
7320         int ret = 0;
7321         struct btrfs_key found_key;
7322         struct extent_buffer *leaf;
7323         int slot;
7324
7325         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7326         if (ret < 0)
7327                 goto out;
7328
7329         while (1) {
7330                 slot = path->slots[0];
7331                 leaf = path->nodes[0];
7332                 if (slot >= btrfs_header_nritems(leaf)) {
7333                         ret = btrfs_next_leaf(root, path);
7334                         if (ret == 0)
7335                                 continue;
7336                         if (ret < 0)
7337                                 goto out;
7338                         break;
7339                 }
7340                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7341
7342                 if (found_key.objectid >= key->objectid &&
7343                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7344                         ret = 0;
7345                         goto out;
7346                 }
7347                 path->slots[0]++;
7348         }
7349 out:
7350         return ret;
7351 }
7352
7353 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7354 {
7355         struct btrfs_block_group_cache *block_group;
7356         u64 last = 0;
7357
7358         while (1) {
7359                 struct inode *inode;
7360
7361                 block_group = btrfs_lookup_first_block_group(info, last);
7362                 while (block_group) {
7363                         spin_lock(&block_group->lock);
7364                         if (block_group->iref)
7365                                 break;
7366                         spin_unlock(&block_group->lock);
7367                         block_group = next_block_group(info->tree_root,
7368                                                        block_group);
7369                 }
7370                 if (!block_group) {
7371                         if (last == 0)
7372                                 break;
7373                         last = 0;
7374                         continue;
7375                 }
7376
7377                 inode = block_group->inode;
7378                 block_group->iref = 0;
7379                 block_group->inode = NULL;
7380                 spin_unlock(&block_group->lock);
7381                 iput(inode);
7382                 last = block_group->key.objectid + block_group->key.offset;
7383                 btrfs_put_block_group(block_group);
7384         }
7385 }
7386
7387 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7388 {
7389         struct btrfs_block_group_cache *block_group;
7390         struct btrfs_space_info *space_info;
7391         struct btrfs_caching_control *caching_ctl;
7392         struct rb_node *n;
7393
7394         down_write(&info->extent_commit_sem);
7395         while (!list_empty(&info->caching_block_groups)) {
7396                 caching_ctl = list_entry(info->caching_block_groups.next,
7397                                          struct btrfs_caching_control, list);
7398                 list_del(&caching_ctl->list);
7399                 put_caching_control(caching_ctl);
7400         }
7401         up_write(&info->extent_commit_sem);
7402
7403         spin_lock(&info->block_group_cache_lock);
7404         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7405                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7406                                        cache_node);
7407                 rb_erase(&block_group->cache_node,
7408                          &info->block_group_cache_tree);
7409                 spin_unlock(&info->block_group_cache_lock);
7410
7411                 down_write(&block_group->space_info->groups_sem);
7412                 list_del(&block_group->list);
7413                 up_write(&block_group->space_info->groups_sem);
7414
7415                 if (block_group->cached == BTRFS_CACHE_STARTED)
7416                         wait_block_group_cache_done(block_group);
7417
7418                 /*
7419                  * We haven't cached this block group, which means we could
7420                  * possibly have excluded extents on this block group.
7421                  */
7422                 if (block_group->cached == BTRFS_CACHE_NO)
7423                         free_excluded_extents(info->extent_root, block_group);
7424
7425                 btrfs_remove_free_space_cache(block_group);
7426                 btrfs_put_block_group(block_group);
7427
7428                 spin_lock(&info->block_group_cache_lock);
7429         }
7430         spin_unlock(&info->block_group_cache_lock);
7431
7432         /* now that all the block groups are freed, go through and
7433          * free all the space_info structs.  This is only called during
7434          * the final stages of unmount, and so we know nobody is
7435          * using them.  We call synchronize_rcu() once before we start,
7436          * just to be on the safe side.
7437          */
7438         synchronize_rcu();
7439
7440         release_global_block_rsv(info);
7441
7442         while(!list_empty(&info->space_info)) {
7443                 space_info = list_entry(info->space_info.next,
7444                                         struct btrfs_space_info,
7445                                         list);
7446                 if (space_info->bytes_pinned > 0 ||
7447                     space_info->bytes_reserved > 0 ||
7448                     space_info->bytes_may_use > 0) {
7449                         WARN_ON(1);
7450                         dump_space_info(space_info, 0, 0);
7451                 }
7452                 list_del(&space_info->list);
7453                 kfree(space_info);
7454         }
7455         return 0;
7456 }
7457
7458 static void __link_block_group(struct btrfs_space_info *space_info,
7459                                struct btrfs_block_group_cache *cache)
7460 {
7461         int index = get_block_group_index(cache);
7462
7463         down_write(&space_info->groups_sem);
7464         list_add_tail(&cache->list, &space_info->block_groups[index]);
7465         up_write(&space_info->groups_sem);
7466 }
7467
7468 int btrfs_read_block_groups(struct btrfs_root *root)
7469 {
7470         struct btrfs_path *path;
7471         int ret;
7472         struct btrfs_block_group_cache *cache;
7473         struct btrfs_fs_info *info = root->fs_info;
7474         struct btrfs_space_info *space_info;
7475         struct btrfs_key key;
7476         struct btrfs_key found_key;
7477         struct extent_buffer *leaf;
7478         int need_clear = 0;
7479         u64 cache_gen;
7480
7481         root = info->extent_root;
7482         key.objectid = 0;
7483         key.offset = 0;
7484         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7485         path = btrfs_alloc_path();
7486         if (!path)
7487                 return -ENOMEM;
7488         path->reada = 1;
7489
7490         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7491         if (btrfs_test_opt(root, SPACE_CACHE) &&
7492             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7493                 need_clear = 1;
7494         if (btrfs_test_opt(root, CLEAR_CACHE))
7495                 need_clear = 1;
7496
7497         while (1) {
7498                 ret = find_first_block_group(root, path, &key);
7499                 if (ret > 0)
7500                         break;
7501                 if (ret != 0)
7502                         goto error;
7503                 leaf = path->nodes[0];
7504                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7505                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7506                 if (!cache) {
7507                         ret = -ENOMEM;
7508                         goto error;
7509                 }
7510                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7511                                                 GFP_NOFS);
7512                 if (!cache->free_space_ctl) {
7513                         kfree(cache);
7514                         ret = -ENOMEM;
7515                         goto error;
7516                 }
7517
7518                 atomic_set(&cache->count, 1);
7519                 spin_lock_init(&cache->lock);
7520                 cache->fs_info = info;
7521                 INIT_LIST_HEAD(&cache->list);
7522                 INIT_LIST_HEAD(&cache->cluster_list);
7523
7524                 if (need_clear)
7525                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7526
7527                 read_extent_buffer(leaf, &cache->item,
7528                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7529                                    sizeof(cache->item));
7530                 memcpy(&cache->key, &found_key, sizeof(found_key));
7531
7532                 key.objectid = found_key.objectid + found_key.offset;
7533                 btrfs_release_path(path);
7534                 cache->flags = btrfs_block_group_flags(&cache->item);
7535                 cache->sectorsize = root->sectorsize;
7536
7537                 btrfs_init_free_space_ctl(cache);
7538
7539                 /*
7540                  * We need to exclude the super stripes now so that the space
7541                  * info has super bytes accounted for, otherwise we'll think
7542                  * we have more space than we actually do.
7543                  */
7544                 exclude_super_stripes(root, cache);
7545
7546                 /*
7547                  * check for two cases, either we are full, and therefore
7548                  * don't need to bother with the caching work since we won't
7549                  * find any space, or we are empty, and we can just add all
7550                  * the space in and be done with it.  This saves us _alot_ of
7551                  * time, particularly in the full case.
7552                  */
7553                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7554                         cache->last_byte_to_unpin = (u64)-1;
7555                         cache->cached = BTRFS_CACHE_FINISHED;
7556                         free_excluded_extents(root, cache);
7557                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7558                         cache->last_byte_to_unpin = (u64)-1;
7559                         cache->cached = BTRFS_CACHE_FINISHED;
7560                         add_new_free_space(cache, root->fs_info,
7561                                            found_key.objectid,
7562                                            found_key.objectid +
7563                                            found_key.offset);
7564                         free_excluded_extents(root, cache);
7565                 }
7566
7567                 ret = update_space_info(info, cache->flags, found_key.offset,
7568                                         btrfs_block_group_used(&cache->item),
7569                                         &space_info);
7570                 BUG_ON(ret);
7571                 cache->space_info = space_info;
7572                 spin_lock(&cache->space_info->lock);
7573                 cache->space_info->bytes_readonly += cache->bytes_super;
7574                 spin_unlock(&cache->space_info->lock);
7575
7576                 __link_block_group(space_info, cache);
7577
7578                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7579                 BUG_ON(ret);
7580
7581                 set_avail_alloc_bits(root->fs_info, cache->flags);
7582                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7583                         set_block_group_ro(cache, 1);
7584         }
7585
7586         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7587                 if (!(get_alloc_profile(root, space_info->flags) &
7588                       (BTRFS_BLOCK_GROUP_RAID10 |
7589                        BTRFS_BLOCK_GROUP_RAID1 |
7590                        BTRFS_BLOCK_GROUP_DUP)))
7591                         continue;
7592                 /*
7593                  * avoid allocating from un-mirrored block group if there are
7594                  * mirrored block groups.
7595                  */
7596                 list_for_each_entry(cache, &space_info->block_groups[3], list)
7597                         set_block_group_ro(cache, 1);
7598                 list_for_each_entry(cache, &space_info->block_groups[4], list)
7599                         set_block_group_ro(cache, 1);
7600         }
7601
7602         init_global_block_rsv(info);
7603         ret = 0;
7604 error:
7605         btrfs_free_path(path);
7606         return ret;
7607 }
7608
7609 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7610                            struct btrfs_root *root, u64 bytes_used,
7611                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7612                            u64 size)
7613 {
7614         int ret;
7615         struct btrfs_root *extent_root;
7616         struct btrfs_block_group_cache *cache;
7617
7618         extent_root = root->fs_info->extent_root;
7619
7620         root->fs_info->last_trans_log_full_commit = trans->transid;
7621
7622         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7623         if (!cache)
7624                 return -ENOMEM;
7625         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7626                                         GFP_NOFS);
7627         if (!cache->free_space_ctl) {
7628                 kfree(cache);
7629                 return -ENOMEM;
7630         }
7631
7632         cache->key.objectid = chunk_offset;
7633         cache->key.offset = size;
7634         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7635         cache->sectorsize = root->sectorsize;
7636         cache->fs_info = root->fs_info;
7637
7638         atomic_set(&cache->count, 1);
7639         spin_lock_init(&cache->lock);
7640         INIT_LIST_HEAD(&cache->list);
7641         INIT_LIST_HEAD(&cache->cluster_list);
7642
7643         btrfs_init_free_space_ctl(cache);
7644
7645         btrfs_set_block_group_used(&cache->item, bytes_used);
7646         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7647         cache->flags = type;
7648         btrfs_set_block_group_flags(&cache->item, type);
7649
7650         cache->last_byte_to_unpin = (u64)-1;
7651         cache->cached = BTRFS_CACHE_FINISHED;
7652         exclude_super_stripes(root, cache);
7653
7654         add_new_free_space(cache, root->fs_info, chunk_offset,
7655                            chunk_offset + size);
7656
7657         free_excluded_extents(root, cache);
7658
7659         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7660                                 &cache->space_info);
7661         BUG_ON(ret);
7662         update_global_block_rsv(root->fs_info);
7663
7664         spin_lock(&cache->space_info->lock);
7665         cache->space_info->bytes_readonly += cache->bytes_super;
7666         spin_unlock(&cache->space_info->lock);
7667
7668         __link_block_group(cache->space_info, cache);
7669
7670         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7671         BUG_ON(ret);
7672
7673         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
7674                                 sizeof(cache->item));
7675         BUG_ON(ret);
7676
7677         set_avail_alloc_bits(extent_root->fs_info, type);
7678
7679         return 0;
7680 }
7681
7682 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
7683 {
7684         u64 extra_flags = flags & BTRFS_BLOCK_GROUP_PROFILE_MASK;
7685
7686         /* chunk -> extended profile */
7687         if (extra_flags == 0)
7688                 extra_flags = BTRFS_AVAIL_ALLOC_BIT_SINGLE;
7689
7690         if (flags & BTRFS_BLOCK_GROUP_DATA)
7691                 fs_info->avail_data_alloc_bits &= ~extra_flags;
7692         if (flags & BTRFS_BLOCK_GROUP_METADATA)
7693                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
7694         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
7695                 fs_info->avail_system_alloc_bits &= ~extra_flags;
7696 }
7697
7698 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7699                              struct btrfs_root *root, u64 group_start)
7700 {
7701         struct btrfs_path *path;
7702         struct btrfs_block_group_cache *block_group;
7703         struct btrfs_free_cluster *cluster;
7704         struct btrfs_root *tree_root = root->fs_info->tree_root;
7705         struct btrfs_key key;
7706         struct inode *inode;
7707         int ret;
7708         int index;
7709         int factor;
7710
7711         root = root->fs_info->extent_root;
7712
7713         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7714         BUG_ON(!block_group);
7715         BUG_ON(!block_group->ro);
7716
7717         /*
7718          * Free the reserved super bytes from this block group before
7719          * remove it.
7720          */
7721         free_excluded_extents(root, block_group);
7722
7723         memcpy(&key, &block_group->key, sizeof(key));
7724         index = get_block_group_index(block_group);
7725         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7726                                   BTRFS_BLOCK_GROUP_RAID1 |
7727                                   BTRFS_BLOCK_GROUP_RAID10))
7728                 factor = 2;
7729         else
7730                 factor = 1;
7731
7732         /* make sure this block group isn't part of an allocation cluster */
7733         cluster = &root->fs_info->data_alloc_cluster;
7734         spin_lock(&cluster->refill_lock);
7735         btrfs_return_cluster_to_free_space(block_group, cluster);
7736         spin_unlock(&cluster->refill_lock);
7737
7738         /*
7739          * make sure this block group isn't part of a metadata
7740          * allocation cluster
7741          */
7742         cluster = &root->fs_info->meta_alloc_cluster;
7743         spin_lock(&cluster->refill_lock);
7744         btrfs_return_cluster_to_free_space(block_group, cluster);
7745         spin_unlock(&cluster->refill_lock);
7746
7747         path = btrfs_alloc_path();
7748         if (!path) {
7749                 ret = -ENOMEM;
7750                 goto out;
7751         }
7752
7753         inode = lookup_free_space_inode(tree_root, block_group, path);
7754         if (!IS_ERR(inode)) {
7755                 ret = btrfs_orphan_add(trans, inode);
7756                 BUG_ON(ret);
7757                 clear_nlink(inode);
7758                 /* One for the block groups ref */
7759                 spin_lock(&block_group->lock);
7760                 if (block_group->iref) {
7761                         block_group->iref = 0;
7762                         block_group->inode = NULL;
7763                         spin_unlock(&block_group->lock);
7764                         iput(inode);
7765                 } else {
7766                         spin_unlock(&block_group->lock);
7767                 }
7768                 /* One for our lookup ref */
7769                 btrfs_add_delayed_iput(inode);
7770         }
7771
7772         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
7773         key.offset = block_group->key.objectid;
7774         key.type = 0;
7775
7776         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
7777         if (ret < 0)
7778                 goto out;
7779         if (ret > 0)
7780                 btrfs_release_path(path);
7781         if (ret == 0) {
7782                 ret = btrfs_del_item(trans, tree_root, path);
7783                 if (ret)
7784                         goto out;
7785                 btrfs_release_path(path);
7786         }
7787
7788         spin_lock(&root->fs_info->block_group_cache_lock);
7789         rb_erase(&block_group->cache_node,
7790                  &root->fs_info->block_group_cache_tree);
7791         spin_unlock(&root->fs_info->block_group_cache_lock);
7792
7793         down_write(&block_group->space_info->groups_sem);
7794         /*
7795          * we must use list_del_init so people can check to see if they
7796          * are still on the list after taking the semaphore
7797          */
7798         list_del_init(&block_group->list);
7799         if (list_empty(&block_group->space_info->block_groups[index]))
7800                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
7801         up_write(&block_group->space_info->groups_sem);
7802
7803         if (block_group->cached == BTRFS_CACHE_STARTED)
7804                 wait_block_group_cache_done(block_group);
7805
7806         btrfs_remove_free_space_cache(block_group);
7807
7808         spin_lock(&block_group->space_info->lock);
7809         block_group->space_info->total_bytes -= block_group->key.offset;
7810         block_group->space_info->bytes_readonly -= block_group->key.offset;
7811         block_group->space_info->disk_total -= block_group->key.offset * factor;
7812         spin_unlock(&block_group->space_info->lock);
7813
7814         memcpy(&key, &block_group->key, sizeof(key));
7815
7816         btrfs_clear_space_info_full(root->fs_info);
7817
7818         btrfs_put_block_group(block_group);
7819         btrfs_put_block_group(block_group);
7820
7821         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
7822         if (ret > 0)
7823                 ret = -EIO;
7824         if (ret < 0)
7825                 goto out;
7826
7827         ret = btrfs_del_item(trans, root, path);
7828 out:
7829         btrfs_free_path(path);
7830         return ret;
7831 }
7832
7833 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
7834 {
7835         struct btrfs_space_info *space_info;
7836         struct btrfs_super_block *disk_super;
7837         u64 features;
7838         u64 flags;
7839         int mixed = 0;
7840         int ret;
7841
7842         disk_super = fs_info->super_copy;
7843         if (!btrfs_super_root(disk_super))
7844                 return 1;
7845
7846         features = btrfs_super_incompat_flags(disk_super);
7847         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
7848                 mixed = 1;
7849
7850         flags = BTRFS_BLOCK_GROUP_SYSTEM;
7851         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7852         if (ret)
7853                 goto out;
7854
7855         if (mixed) {
7856                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
7857                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7858         } else {
7859                 flags = BTRFS_BLOCK_GROUP_METADATA;
7860                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7861                 if (ret)
7862                         goto out;
7863
7864                 flags = BTRFS_BLOCK_GROUP_DATA;
7865                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7866         }
7867 out:
7868         return ret;
7869 }
7870
7871 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
7872 {
7873         return unpin_extent_range(root, start, end);
7874 }
7875
7876 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
7877                                u64 num_bytes, u64 *actual_bytes)
7878 {
7879         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
7880 }
7881
7882 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
7883 {
7884         struct btrfs_fs_info *fs_info = root->fs_info;
7885         struct btrfs_block_group_cache *cache = NULL;
7886         u64 group_trimmed;
7887         u64 start;
7888         u64 end;
7889         u64 trimmed = 0;
7890         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
7891         int ret = 0;
7892
7893         /*
7894          * try to trim all FS space, our block group may start from non-zero.
7895          */
7896         if (range->len == total_bytes)
7897                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
7898         else
7899                 cache = btrfs_lookup_block_group(fs_info, range->start);
7900
7901         while (cache) {
7902                 if (cache->key.objectid >= (range->start + range->len)) {
7903                         btrfs_put_block_group(cache);
7904                         break;
7905                 }
7906
7907                 start = max(range->start, cache->key.objectid);
7908                 end = min(range->start + range->len,
7909                                 cache->key.objectid + cache->key.offset);
7910
7911                 if (end - start >= range->minlen) {
7912                         if (!block_group_cache_done(cache)) {
7913                                 ret = cache_block_group(cache, NULL, root, 0);
7914                                 if (!ret)
7915                                         wait_block_group_cache_done(cache);
7916                         }
7917                         ret = btrfs_trim_block_group(cache,
7918                                                      &group_trimmed,
7919                                                      start,
7920                                                      end,
7921                                                      range->minlen);
7922
7923                         trimmed += group_trimmed;
7924                         if (ret) {
7925                                 btrfs_put_block_group(cache);
7926                                 break;
7927                         }
7928                 }
7929
7930                 cache = next_block_group(fs_info->tree_root, cache);
7931         }
7932
7933         range->len = trimmed;
7934         return ret;
7935 }