fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include "compat.h"
  28 #include "hash.h"
  29 #include "ctree.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "transaction.h"
  33 #include "volumes.h"
  34 #include "locking.h"
  35 #include "free-space-cache.h"
  36
  37 /* control flags for do_chunk_alloc's force field
  38  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  39  * if we really need one.
  40  *
  41  * CHUNK_ALLOC_FORCE means it must try to allocate one
  42  *
  43  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  44  * if we have very few chunks already allocated.  This is
  45  * used as part of the clustering code to help make sure
  46  * we have a good pool of storage to cluster in, without
  47  * filling the FS with empty chunks
  48  *
  49  */
  50 enum {
  51         CHUNK_ALLOC_NO_FORCE = 0,
  52         CHUNK_ALLOC_FORCE = 1,
  53         CHUNK_ALLOC_LIMITED = 2,
  54 };
  55
  56 /*
  57  * Control how reservations are dealt with.
  58  *
  59  * RESERVE_FREE - freeing a reservation.
  60  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  61  *   ENOSPC accounting
  62  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  63  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  64  */
  65 enum {
  66         RESERVE_FREE = 0,
  67         RESERVE_ALLOC = 1,
  68         RESERVE_ALLOC_NO_ACCOUNT = 2,
  69 };
  70
  71 static int update_block_group(struct btrfs_trans_handle *trans,
  72                               struct btrfs_root *root,
  73                               u64 bytenr, u64 num_bytes, int alloc);
  74 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  75                                 struct btrfs_root *root,
  76                                 u64 bytenr, u64 num_bytes, u64 parent,
  77                                 u64 root_objectid, u64 owner_objectid,
  78                                 u64 owner_offset, int refs_to_drop,
  79                                 struct btrfs_delayed_extent_op *extra_op);
  80 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  81                                     struct extent_buffer *leaf,
  82                                     struct btrfs_extent_item *ei);
  83 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  84                                       struct btrfs_root *root,
  85                                       u64 parent, u64 root_objectid,
  86                                       u64 flags, u64 owner, u64 offset,
  87                                       struct btrfs_key *ins, int ref_mod);
  88 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  89                                      struct btrfs_root *root,
  90                                      u64 parent, u64 root_objectid,
  91                                      u64 flags, struct btrfs_disk_key *key,
  92                                      int level, struct btrfs_key *ins);
  93 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  94                           struct btrfs_root *extent_root, u64 alloc_bytes,
  95                           u64 flags, int force);
  96 static int find_next_key(struct btrfs_path *path, int level,
  97                          struct btrfs_key *key);
  98 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
  99                             int dump_block_groups);
 100 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 101                                        u64 num_bytes, int reserve);
 102
 103 static noinline int
 104 block_group_cache_done(struct btrfs_block_group_cache *cache)
 105 {
 106         smp_mb();
 107         return cache->cached == BTRFS_CACHE_FINISHED;
 108 }
 109
 110 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 111 {
 112         return (cache->flags & bits) == bits;
 113 }
 114
 115 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 116 {
 117         atomic_inc(&cache->count);
 118 }
 119
 120 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 121 {
 122         if (atomic_dec_and_test(&cache->count)) {
 123                 WARN_ON(cache->pinned > 0);
 124                 WARN_ON(cache->reserved > 0);
 125                 kfree(cache->free_space_ctl);
 126                 kfree(cache);
 127         }
 128 }
 129
 130 /*
 131  * this adds the block group to the fs_info rb tree for the block group
 132  * cache
 133  */
 134 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 135                                 struct btrfs_block_group_cache *block_group)
 136 {
 137         struct rb_node **p;
 138         struct rb_node *parent = NULL;
 139         struct btrfs_block_group_cache *cache;
 140
 141         spin_lock(&info->block_group_cache_lock);
 142         p = &info->block_group_cache_tree.rb_node;
 143
 144         while (*p) {
 145                 parent = *p;
 146                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 147                                  cache_node);
 148                 if (block_group->key.objectid < cache->key.objectid) {
 149                         p = &(*p)->rb_left;
 150                 } else if (block_group->key.objectid > cache->key.objectid) {
 151                         p = &(*p)->rb_right;
 152                 } else {
 153                         spin_unlock(&info->block_group_cache_lock);
 154                         return -EEXIST;
 155                 }
 156         }
 157
 158         rb_link_node(&block_group->cache_node, parent, p);
 159         rb_insert_color(&block_group->cache_node,
 160                         &info->block_group_cache_tree);
 161         spin_unlock(&info->block_group_cache_lock);
 162
 163         return 0;
 164 }
 165
 166 /*
 167  * This will return the block group at or after bytenr if contains is 0, else
 168  * it will return the block group that contains the bytenr
 169  */
 170 static struct btrfs_block_group_cache *
 171 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 172                               int contains)
 173 {
 174         struct btrfs_block_group_cache *cache, *ret = NULL;
 175         struct rb_node *n;
 176         u64 end, start;
 177
 178         spin_lock(&info->block_group_cache_lock);
 179         n = info->block_group_cache_tree.rb_node;
 180
 181         while (n) {
 182                 cache = rb_entry(n, struct btrfs_block_group_cache,
 183                                  cache_node);
 184                 end = cache->key.objectid + cache->key.offset - 1;
 185                 start = cache->key.objectid;
 186
 187                 if (bytenr < start) {
 188                         if (!contains && (!ret || start < ret->key.objectid))
 189                                 ret = cache;
 190                         n = n->rb_left;
 191                 } else if (bytenr > start) {
 192                         if (contains && bytenr <= end) {
 193                                 ret = cache;
 194                                 break;
 195                         }
 196                         n = n->rb_right;
 197                 } else {
 198                         ret = cache;
 199                         break;
 200                 }
 201         }
 202         if (ret)
 203                 btrfs_get_block_group(ret);
 204         spin_unlock(&info->block_group_cache_lock);
 205
 206         return ret;
 207 }
 208
 209 static int add_excluded_extent(struct btrfs_root *root,
 210                                u64 start, u64 num_bytes)
 211 {
 212         u64 end = start + num_bytes - 1;
 213         set_extent_bits(&root->fs_info->freed_extents[0],
 214                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 215         set_extent_bits(&root->fs_info->freed_extents[1],
 216                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 217         return 0;
 218 }
 219
 220 static void free_excluded_extents(struct btrfs_root *root,
 221                                   struct btrfs_block_group_cache *cache)
 222 {
 223         u64 start, end;
 224
 225         start = cache->key.objectid;
 226         end = start + cache->key.offset - 1;
 227
 228         clear_extent_bits(&root->fs_info->freed_extents[0],
 229                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 230         clear_extent_bits(&root->fs_info->freed_extents[1],
 231                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 232 }
 233
 234 static int exclude_super_stripes(struct btrfs_root *root,
 235                                  struct btrfs_block_group_cache *cache)
 236 {
 237         u64 bytenr;
 238         u64 *logical;
 239         int stripe_len;
 240         int i, nr, ret;
 241
 242         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 243                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 244                 cache->bytes_super += stripe_len;
 245                 ret = add_excluded_extent(root, cache->key.objectid,
 246                                           stripe_len);
 247                 BUG_ON(ret);
 248         }
 249
 250         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 251                 bytenr = btrfs_sb_offset(i);
 252                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 253                                        cache->key.objectid, bytenr,
 254                                        0, &logical, &nr, &stripe_len);
 255                 BUG_ON(ret);
 256
 257                 while (nr--) {
 258                         cache->bytes_super += stripe_len;
 259                         ret = add_excluded_extent(root, logical[nr],
 260                                                   stripe_len);
 261                         BUG_ON(ret);
 262                 }
 263
 264                 kfree(logical);
 265         }
 266         return 0;
 267 }
 268
 269 static struct btrfs_caching_control *
 270 get_caching_control(struct btrfs_block_group_cache *cache)
 271 {
 272         struct btrfs_caching_control *ctl;
 273
 274         spin_lock(&cache->lock);
 275         if (cache->cached != BTRFS_CACHE_STARTED) {
 276                 spin_unlock(&cache->lock);
 277                 return NULL;
 278         }
 279
 280         /* We're loading it the fast way, so we don't have a caching_ctl. */
 281         if (!cache->caching_ctl) {
 282                 spin_unlock(&cache->lock);
 283                 return NULL;
 284         }
 285
 286         ctl = cache->caching_ctl;
 287         atomic_inc(&ctl->count);
 288         spin_unlock(&cache->lock);
 289         return ctl;
 290 }
 291
 292 static void put_caching_control(struct btrfs_caching_control *ctl)
 293 {
 294         if (atomic_dec_and_test(&ctl->count))
 295                 kfree(ctl);
 296 }
 297
 298 /*
 299  * this is only called by cache_block_group, since we could have freed extents
 300  * we need to check the pinned_extents for any extents that can't be used yet
 301  * since their free space will be released as soon as the transaction commits.
 302  */
 303 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 304                               struct btrfs_fs_info *info, u64 start, u64 end)
 305 {
 306         u64 extent_start, extent_end, size, total_added = 0;
 307         int ret;
 308
 309         while (start < end) {
 310                 ret = find_first_extent_bit(info->pinned_extents, start,
 311                                             &extent_start, &extent_end,
 312                                             EXTENT_DIRTY | EXTENT_UPTODATE);
 313                 if (ret)
 314                         break;
 315
 316                 if (extent_start <= start) {
 317                         start = extent_end + 1;
 318                 } else if (extent_start > start && extent_start < end) {
 319                         size = extent_start - start;
 320                         total_added += size;
 321                         ret = btrfs_add_free_space(block_group, start,
 322                                                    size);
 323                         BUG_ON(ret);
 324                         start = extent_end + 1;
 325                 } else {
 326                         break;
 327                 }
 328         }
 329
 330         if (start < end) {
 331                 size = end - start;
 332                 total_added += size;
 333                 ret = btrfs_add_free_space(block_group, start, size);
 334                 BUG_ON(ret);
 335         }
 336
 337         return total_added;
 338 }
 339
 340 static noinline void caching_thread(struct btrfs_work *work)
 341 {
 342         struct btrfs_block_group_cache *block_group;
 343         struct btrfs_fs_info *fs_info;
 344         struct btrfs_caching_control *caching_ctl;
 345         struct btrfs_root *extent_root;
 346         struct btrfs_path *path;
 347         struct extent_buffer *leaf;
 348         struct btrfs_key key;
 349         u64 total_found = 0;
 350         u64 last = 0;
 351         u32 nritems;
 352         int ret = 0;
 353
 354         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 355         block_group = caching_ctl->block_group;
 356         fs_info = block_group->fs_info;
 357         extent_root = fs_info->extent_root;
 358
 359         path = btrfs_alloc_path();
 360         if (!path)
 361                 goto out;
 362
 363         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 364
 365         /*
 366          * We don't want to deadlock with somebody trying to allocate a new
 367          * extent for the extent root while also trying to search the extent
 368          * root to add free space.  So we skip locking and search the commit
 369          * root, since its read-only
 370          */
 371         path->skip_locking = 1;
 372         path->search_commit_root = 1;
 373         path->reada = 1;
 374
 375         key.objectid = last;
 376         key.offset = 0;
 377         key.type = BTRFS_EXTENT_ITEM_KEY;
 378 again:
 379         mutex_lock(&caching_ctl->mutex);
 380         /* need to make sure the commit_root doesn't disappear */
 381         down_read(&fs_info->extent_commit_sem);
 382
 383         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 384         if (ret < 0)
 385                 goto err;
 386
 387         leaf = path->nodes[0];
 388         nritems = btrfs_header_nritems(leaf);
 389
 390         while (1) {
 391                 if (btrfs_fs_closing(fs_info) > 1) {
 392                         last = (u64)-1;
 393                         break;
 394                 }
 395
 396                 if (path->slots[0] < nritems) {
 397                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 398                 } else {
 399                         ret = find_next_key(path, 0, &key);
 400                         if (ret)
 401                                 break;
 402
 403                         if (need_resched() ||
 404                             btrfs_next_leaf(extent_root, path)) {
 405                                 caching_ctl->progress = last;
 406                                 btrfs_release_path(path);
 407                                 up_read(&fs_info->extent_commit_sem);
 408                                 mutex_unlock(&caching_ctl->mutex);
 409                                 cond_resched();
 410                                 goto again;
 411                         }
 412                         leaf = path->nodes[0];
 413                         nritems = btrfs_header_nritems(leaf);
 414                         continue;
 415                 }
 416
 417                 if (key.objectid < block_group->key.objectid) {
 418                         path->slots[0]++;
 419                         continue;
 420                 }
 421
 422                 if (key.objectid >= block_group->key.objectid +
 423                     block_group->key.offset)
 424                         break;
 425
 426                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 427                         total_found += add_new_free_space(block_group,
 428                                                           fs_info, last,
 429                                                           key.objectid);
 430                         last = key.objectid + key.offset;
 431
 432                         if (total_found > (1024 * 1024 * 2)) {
 433                                 total_found = 0;
 434                                 wake_up(&caching_ctl->wait);
 435                         }
 436                 }
 437                 path->slots[0]++;
 438         }
 439         ret = 0;
 440
 441         total_found += add_new_free_space(block_group, fs_info, last,
 442                                           block_group->key.objectid +
 443                                           block_group->key.offset);
 444         caching_ctl->progress = (u64)-1;
 445
 446         spin_lock(&block_group->lock);
 447         block_group->caching_ctl = NULL;
 448         block_group->cached = BTRFS_CACHE_FINISHED;
 449         spin_unlock(&block_group->lock);
 450
 451 err:
 452         btrfs_free_path(path);
 453         up_read(&fs_info->extent_commit_sem);
 454
 455         free_excluded_extents(extent_root, block_group);
 456
 457         mutex_unlock(&caching_ctl->mutex);
 458 out:
 459         wake_up(&caching_ctl->wait);
 460
 461         put_caching_control(caching_ctl);
 462         btrfs_put_block_group(block_group);
 463 }
 464
 465 static int cache_block_group(struct btrfs_block_group_cache *cache,
 466                              struct btrfs_trans_handle *trans,
 467                              struct btrfs_root *root,
 468                              int load_cache_only)
 469 {
 470         DEFINE_WAIT(wait);
 471         struct btrfs_fs_info *fs_info = cache->fs_info;
 472         struct btrfs_caching_control *caching_ctl;
 473         int ret = 0;
 474
 475         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 476         BUG_ON(!caching_ctl);
 477
 478         INIT_LIST_HEAD(&caching_ctl->list);
 479         mutex_init(&caching_ctl->mutex);
 480         init_waitqueue_head(&caching_ctl->wait);
 481         caching_ctl->block_group = cache;
 482         caching_ctl->progress = cache->key.objectid;
 483         atomic_set(&caching_ctl->count, 1);
 484         caching_ctl->work.func = caching_thread;
 485
 486         spin_lock(&cache->lock);
 487         /*
 488          * This should be a rare occasion, but this could happen I think in the
 489          * case where one thread starts to load the space cache info, and then
 490          * some other thread starts a transaction commit which tries to do an
 491          * allocation while the other thread is still loading the space cache
 492          * info.  The previous loop should have kept us from choosing this block
 493          * group, but if we've moved to the state where we will wait on caching
 494          * block groups we need to first check if we're doing a fast load here,
 495          * so we can wait for it to finish, otherwise we could end up allocating
 496          * from a block group who's cache gets evicted for one reason or
 497          * another.
 498          */
 499         while (cache->cached == BTRFS_CACHE_FAST) {
 500                 struct btrfs_caching_control *ctl;
 501
 502                 ctl = cache->caching_ctl;
 503                 atomic_inc(&ctl->count);
 504                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 505                 spin_unlock(&cache->lock);
 506
 507                 schedule();
 508
 509                 finish_wait(&ctl->wait, &wait);
 510                 put_caching_control(ctl);
 511                 spin_lock(&cache->lock);
 512         }
 513
 514         if (cache->cached != BTRFS_CACHE_NO) {
 515                 spin_unlock(&cache->lock);
 516                 kfree(caching_ctl);
 517                 return 0;
 518         }
 519         WARN_ON(cache->caching_ctl);
 520         cache->caching_ctl = caching_ctl;
 521         cache->cached = BTRFS_CACHE_FAST;
 522         spin_unlock(&cache->lock);
 523
 524         /*
 525          * We can't do the read from on-disk cache during a commit since we need
 526          * to have the normal tree locking.  Also if we are currently trying to
 527          * allocate blocks for the tree root we can't do the fast caching since
 528          * we likely hold important locks.
 529          */
 530         if (trans && (!trans->transaction->in_commit) &&
 531             (root && root != root->fs_info->tree_root) &&
 532             btrfs_test_opt(root, SPACE_CACHE)) {
 533                 ret = load_free_space_cache(fs_info, cache);
 534
 535                 spin_lock(&cache->lock);
 536                 if (ret == 1) {
 537                         cache->caching_ctl = NULL;
 538                         cache->cached = BTRFS_CACHE_FINISHED;
 539                         cache->last_byte_to_unpin = (u64)-1;
 540                 } else {
 541                         if (load_cache_only) {
 542                                 cache->caching_ctl = NULL;
 543                                 cache->cached = BTRFS_CACHE_NO;
 544                         } else {
 545                                 cache->cached = BTRFS_CACHE_STARTED;
 546                         }
 547                 }
 548                 spin_unlock(&cache->lock);
 549                 wake_up(&caching_ctl->wait);
 550                 if (ret == 1) {
 551                         put_caching_control(caching_ctl);
 552                         free_excluded_extents(fs_info->extent_root, cache);
 553                         return 0;
 554                 }
 555         } else {
 556                 /*
 557                  * We are not going to do the fast caching, set cached to the
 558                  * appropriate value and wakeup any waiters.
 559                  */
 560                 spin_lock(&cache->lock);
 561                 if (load_cache_only) {
 562                         cache->caching_ctl = NULL;
 563                         cache->cached = BTRFS_CACHE_NO;
 564                 } else {
 565                         cache->cached = BTRFS_CACHE_STARTED;
 566                 }
 567                 spin_unlock(&cache->lock);
 568                 wake_up(&caching_ctl->wait);
 569         }
 570
 571         if (load_cache_only) {
 572                 put_caching_control(caching_ctl);
 573                 return 0;
 574         }
 575
 576         down_write(&fs_info->extent_commit_sem);
 577         atomic_inc(&caching_ctl->count);
 578         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 579         up_write(&fs_info->extent_commit_sem);
 580
 581         btrfs_get_block_group(cache);
 582
 583         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 584
 585         return ret;
 586 }
 587
 588 /*
 589  * return the block group that starts at or after bytenr
 590  */
 591 static struct btrfs_block_group_cache *
 592 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 593 {
 594         struct btrfs_block_group_cache *cache;
 595
 596         cache = block_group_cache_tree_search(info, bytenr, 0);
 597
 598         return cache;
 599 }
 600
 601 /*
 602  * return the block group that contains the given bytenr
 603  */
 604 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 605                                                  struct btrfs_fs_info *info,
 606                                                  u64 bytenr)
 607 {
 608         struct btrfs_block_group_cache *cache;
 609
 610         cache = block_group_cache_tree_search(info, bytenr, 1);
 611
 612         return cache;
 613 }
 614
 615 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 616                                                   u64 flags)
 617 {
 618         struct list_head *head = &info->space_info;
 619         struct btrfs_space_info *found;
 620
 621         flags &= BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_SYSTEM |
 622                  BTRFS_BLOCK_GROUP_METADATA;
 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         ret = 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 int 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         int ret;
1612
1613         leaf = path->nodes[0];
1614         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1615         item_offset = (unsigned long)iref - (unsigned long)ei;
1616
1617         type = extent_ref_type(parent, owner);
1618         size = btrfs_extent_inline_ref_size(type);
1619
1620         ret = btrfs_extend_item(trans, root, path, size);
1621
1622         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1623         refs = btrfs_extent_refs(leaf, ei);
1624         refs += refs_to_add;
1625         btrfs_set_extent_refs(leaf, ei, refs);
1626         if (extent_op)
1627                 __run_delayed_extent_op(extent_op, leaf, ei);
1628
1629         ptr = (unsigned long)ei + item_offset;
1630         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1631         if (ptr < end - size)
1632                 memmove_extent_buffer(leaf, ptr + size, ptr,
1633                                       end - size - ptr);
1634
1635         iref = (struct btrfs_extent_inline_ref *)ptr;
1636         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1637         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1638                 struct btrfs_extent_data_ref *dref;
1639                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1640                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1641                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1642                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1643                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1644         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1645                 struct btrfs_shared_data_ref *sref;
1646                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1647                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1648                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1649         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1650                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1651         } else {
1652                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1653         }
1654         btrfs_mark_buffer_dirty(leaf);
1655         return 0;
1656 }
1657
1658 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1659                                  struct btrfs_root *root,
1660                                  struct btrfs_path *path,
1661                                  struct btrfs_extent_inline_ref **ref_ret,
1662                                  u64 bytenr, u64 num_bytes, u64 parent,
1663                                  u64 root_objectid, u64 owner, u64 offset)
1664 {
1665         int ret;
1666
1667         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1668                                            bytenr, num_bytes, parent,
1669                                            root_objectid, owner, offset, 0);
1670         if (ret != -ENOENT)
1671                 return ret;
1672
1673         btrfs_release_path(path);
1674         *ref_ret = NULL;
1675
1676         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1677                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1678                                             root_objectid);
1679         } else {
1680                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1681                                              root_objectid, owner, offset);
1682         }
1683         return ret;
1684 }
1685
1686 /*
1687  * helper to update/remove inline back ref
1688  */
1689 static noinline_for_stack
1690 int update_inline_extent_backref(struct btrfs_trans_handle *trans,
1691                                  struct btrfs_root *root,
1692                                  struct btrfs_path *path,
1693                                  struct btrfs_extent_inline_ref *iref,
1694                                  int refs_to_mod,
1695                                  struct btrfs_delayed_extent_op *extent_op)
1696 {
1697         struct extent_buffer *leaf;
1698         struct btrfs_extent_item *ei;
1699         struct btrfs_extent_data_ref *dref = NULL;
1700         struct btrfs_shared_data_ref *sref = NULL;
1701         unsigned long ptr;
1702         unsigned long end;
1703         u32 item_size;
1704         int size;
1705         int type;
1706         int ret;
1707         u64 refs;
1708
1709         leaf = path->nodes[0];
1710         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1711         refs = btrfs_extent_refs(leaf, ei);
1712         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1713         refs += refs_to_mod;
1714         btrfs_set_extent_refs(leaf, ei, refs);
1715         if (extent_op)
1716                 __run_delayed_extent_op(extent_op, leaf, ei);
1717
1718         type = btrfs_extent_inline_ref_type(leaf, iref);
1719
1720         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1721                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1722                 refs = btrfs_extent_data_ref_count(leaf, dref);
1723         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1724                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1725                 refs = btrfs_shared_data_ref_count(leaf, sref);
1726         } else {
1727                 refs = 1;
1728                 BUG_ON(refs_to_mod != -1);
1729         }
1730
1731         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1732         refs += refs_to_mod;
1733
1734         if (refs > 0) {
1735                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1736                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1737                 else
1738                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1739         } else {
1740                 size =  btrfs_extent_inline_ref_size(type);
1741                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1742                 ptr = (unsigned long)iref;
1743                 end = (unsigned long)ei + item_size;
1744                 if (ptr + size < end)
1745                         memmove_extent_buffer(leaf, ptr, ptr + size,
1746                                               end - ptr - size);
1747                 item_size -= size;
1748                 ret = btrfs_truncate_item(trans, root, path, item_size, 1);
1749         }
1750         btrfs_mark_buffer_dirty(leaf);
1751         return 0;
1752 }
1753
1754 static noinline_for_stack
1755 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1756                                  struct btrfs_root *root,
1757                                  struct btrfs_path *path,
1758                                  u64 bytenr, u64 num_bytes, u64 parent,
1759                                  u64 root_objectid, u64 owner,
1760                                  u64 offset, int refs_to_add,
1761                                  struct btrfs_delayed_extent_op *extent_op)
1762 {
1763         struct btrfs_extent_inline_ref *iref;
1764         int ret;
1765
1766         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1767                                            bytenr, num_bytes, parent,
1768                                            root_objectid, owner, offset, 1);
1769         if (ret == 0) {
1770                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1771                 ret = update_inline_extent_backref(trans, root, path, iref,
1772                                                    refs_to_add, extent_op);
1773         } else if (ret == -ENOENT) {
1774                 ret = setup_inline_extent_backref(trans, root, path, iref,
1775                                                   parent, root_objectid,
1776                                                   owner, offset, refs_to_add,
1777                                                   extent_op);
1778         }
1779         return ret;
1780 }
1781
1782 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1783                                  struct btrfs_root *root,
1784                                  struct btrfs_path *path,
1785                                  u64 bytenr, u64 parent, u64 root_objectid,
1786                                  u64 owner, u64 offset, int refs_to_add)
1787 {
1788         int ret;
1789         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1790                 BUG_ON(refs_to_add != 1);
1791                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1792                                             parent, root_objectid);
1793         } else {
1794                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1795                                              parent, root_objectid,
1796                                              owner, offset, refs_to_add);
1797         }
1798         return ret;
1799 }
1800
1801 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1802                                  struct btrfs_root *root,
1803                                  struct btrfs_path *path,
1804                                  struct btrfs_extent_inline_ref *iref,
1805                                  int refs_to_drop, int is_data)
1806 {
1807         int ret;
1808
1809         BUG_ON(!is_data && refs_to_drop != 1);
1810         if (iref) {
1811                 ret = update_inline_extent_backref(trans, root, path, iref,
1812                                                    -refs_to_drop, NULL);
1813         } else if (is_data) {
1814                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1815         } else {
1816                 ret = btrfs_del_item(trans, root, path);
1817         }
1818         return ret;
1819 }
1820
1821 static int btrfs_issue_discard(struct block_device *bdev,
1822                                 u64 start, u64 len)
1823 {
1824         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1825 }
1826
1827 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1828                                 u64 num_bytes, u64 *actual_bytes)
1829 {
1830         int ret;
1831         u64 discarded_bytes = 0;
1832         struct btrfs_bio *bbio = NULL;
1833
1834
1835         /* Tell the block device(s) that the sectors can be discarded */
1836         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1837                               bytenr, &num_bytes, &bbio, 0);
1838         if (!ret) {
1839                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1840                 int i;
1841
1842
1843                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1844                         if (!stripe->dev->can_discard)
1845                                 continue;
1846
1847                         ret = btrfs_issue_discard(stripe->dev->bdev,
1848                                                   stripe->physical,
1849                                                   stripe->length);
1850                         if (!ret)
1851                                 discarded_bytes += stripe->length;
1852                         else if (ret != -EOPNOTSUPP)
1853                                 break;
1854
1855                         /*
1856                          * Just in case we get back EOPNOTSUPP for some reason,
1857                          * just ignore the return value so we don't screw up
1858                          * people calling discard_extent.
1859                          */
1860                         ret = 0;
1861                 }
1862                 kfree(bbio);
1863         }
1864
1865         if (actual_bytes)
1866                 *actual_bytes = discarded_bytes;
1867
1868
1869         return ret;
1870 }
1871
1872 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1873                          struct btrfs_root *root,
1874                          u64 bytenr, u64 num_bytes, u64 parent,
1875                          u64 root_objectid, u64 owner, u64 offset)
1876 {
1877         int ret;
1878         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1879                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1880
1881         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1882                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
1883                                         parent, root_objectid, (int)owner,
1884                                         BTRFS_ADD_DELAYED_REF, NULL);
1885         } else {
1886                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
1887                                         parent, root_objectid, owner, offset,
1888                                         BTRFS_ADD_DELAYED_REF, NULL);
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                  * record the must insert reserved flag before we
2237                  * drop the spin lock.
2238                  */
2239                 must_insert_reserved = locked_ref->must_insert_reserved;
2240                 locked_ref->must_insert_reserved = 0;
2241
2242                 extent_op = locked_ref->extent_op;
2243                 locked_ref->extent_op = NULL;
2244
2245                 /*
2246                  * locked_ref is the head node, so we have to go one
2247                  * node back for any delayed ref updates
2248                  */
2249                 ref = select_delayed_ref(locked_ref);
2250                 if (!ref) {
2251                         /* All delayed refs have been processed, Go ahead
2252                          * and send the head node to run_one_delayed_ref,
2253                          * so that any accounting fixes can happen
2254                          */
2255                         ref = &locked_ref->node;
2256
2257                         if (extent_op && must_insert_reserved) {
2258                                 kfree(extent_op);
2259                                 extent_op = NULL;
2260                         }
2261
2262                         if (extent_op) {
2263                                 spin_unlock(&delayed_refs->lock);
2264
2265                                 ret = run_delayed_extent_op(trans, root,
2266                                                             ref, extent_op);
2267                                 BUG_ON(ret);
2268                                 kfree(extent_op);
2269
2270                                 goto next;
2271                         }
2272
2273                         list_del_init(&locked_ref->cluster);
2274                         locked_ref = NULL;
2275                 }
2276
2277                 ref->in_tree = 0;
2278                 rb_erase(&ref->rb_node, &delayed_refs->root);
2279                 delayed_refs->num_entries--;
2280
2281                 spin_unlock(&delayed_refs->lock);
2282
2283                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2284                                           must_insert_reserved);
2285                 BUG_ON(ret);
2286
2287                 btrfs_put_delayed_ref(ref);
2288                 kfree(extent_op);
2289                 count++;
2290 next:
2291                 do_chunk_alloc(trans, root->fs_info->extent_root,
2292                                2 * 1024 * 1024,
2293                                btrfs_get_alloc_profile(root, 0),
2294                                CHUNK_ALLOC_NO_FORCE);
2295                 cond_resched();
2296                 spin_lock(&delayed_refs->lock);
2297         }
2298         return count;
2299 }
2300
2301 /*
2302  * this starts processing the delayed reference count updates and
2303  * extent insertions we have queued up so far.  count can be
2304  * 0, which means to process everything in the tree at the start
2305  * of the run (but not newly added entries), or it can be some target
2306  * number you'd like to process.
2307  */
2308 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2309                            struct btrfs_root *root, unsigned long count)
2310 {
2311         struct rb_node *node;
2312         struct btrfs_delayed_ref_root *delayed_refs;
2313         struct btrfs_delayed_ref_node *ref;
2314         struct list_head cluster;
2315         int ret;
2316         int run_all = count == (unsigned long)-1;
2317         int run_most = 0;
2318
2319         if (root == root->fs_info->extent_root)
2320                 root = root->fs_info->tree_root;
2321
2322         do_chunk_alloc(trans, root->fs_info->extent_root,
2323                        2 * 1024 * 1024, btrfs_get_alloc_profile(root, 0),
2324                        CHUNK_ALLOC_NO_FORCE);
2325
2326         delayed_refs = &trans->transaction->delayed_refs;
2327         INIT_LIST_HEAD(&cluster);
2328 again:
2329         spin_lock(&delayed_refs->lock);
2330         if (count == 0) {
2331                 count = delayed_refs->num_entries * 2;
2332                 run_most = 1;
2333         }
2334         while (1) {
2335                 if (!(run_all || run_most) &&
2336                     delayed_refs->num_heads_ready < 64)
2337                         break;
2338
2339                 /*
2340                  * go find something we can process in the rbtree.  We start at
2341                  * the beginning of the tree, and then build a cluster
2342                  * of refs to process starting at the first one we are able to
2343                  * lock
2344                  */
2345                 ret = btrfs_find_ref_cluster(trans, &cluster,
2346                                              delayed_refs->run_delayed_start);
2347                 if (ret)
2348                         break;
2349
2350                 ret = run_clustered_refs(trans, root, &cluster);
2351                 BUG_ON(ret < 0);
2352
2353                 count -= min_t(unsigned long, ret, count);
2354
2355                 if (count == 0)
2356                         break;
2357         }
2358
2359         if (run_all) {
2360                 node = rb_first(&delayed_refs->root);
2361                 if (!node)
2362                         goto out;
2363                 count = (unsigned long)-1;
2364
2365                 while (node) {
2366                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2367                                        rb_node);
2368                         if (btrfs_delayed_ref_is_head(ref)) {
2369                                 struct btrfs_delayed_ref_head *head;
2370
2371                                 head = btrfs_delayed_node_to_head(ref);
2372                                 atomic_inc(&ref->refs);
2373
2374                                 spin_unlock(&delayed_refs->lock);
2375                                 /*
2376                                  * Mutex was contended, block until it's
2377                                  * released and try again
2378                                  */
2379                                 mutex_lock(&head->mutex);
2380                                 mutex_unlock(&head->mutex);
2381
2382                                 btrfs_put_delayed_ref(ref);
2383                                 cond_resched();
2384                                 goto again;
2385                         }
2386                         node = rb_next(node);
2387                 }
2388                 spin_unlock(&delayed_refs->lock);
2389                 schedule_timeout(1);
2390                 goto again;
2391         }
2392 out:
2393         spin_unlock(&delayed_refs->lock);
2394         return 0;
2395 }
2396
2397 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2398                                 struct btrfs_root *root,
2399                                 u64 bytenr, u64 num_bytes, u64 flags,
2400                                 int is_data)
2401 {
2402         struct btrfs_delayed_extent_op *extent_op;
2403         int ret;
2404
2405         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2406         if (!extent_op)
2407                 return -ENOMEM;
2408
2409         extent_op->flags_to_set = flags;
2410         extent_op->update_flags = 1;
2411         extent_op->update_key = 0;
2412         extent_op->is_data = is_data ? 1 : 0;
2413
2414         ret = btrfs_add_delayed_extent_op(trans, bytenr, num_bytes, extent_op);
2415         if (ret)
2416                 kfree(extent_op);
2417         return ret;
2418 }
2419
2420 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2421                                       struct btrfs_root *root,
2422                                       struct btrfs_path *path,
2423                                       u64 objectid, u64 offset, u64 bytenr)
2424 {
2425         struct btrfs_delayed_ref_head *head;
2426         struct btrfs_delayed_ref_node *ref;
2427         struct btrfs_delayed_data_ref *data_ref;
2428         struct btrfs_delayed_ref_root *delayed_refs;
2429         struct rb_node *node;
2430         int ret = 0;
2431
2432         ret = -ENOENT;
2433         delayed_refs = &trans->transaction->delayed_refs;
2434         spin_lock(&delayed_refs->lock);
2435         head = btrfs_find_delayed_ref_head(trans, bytenr);
2436         if (!head)
2437                 goto out;
2438
2439         if (!mutex_trylock(&head->mutex)) {
2440                 atomic_inc(&head->node.refs);
2441                 spin_unlock(&delayed_refs->lock);
2442
2443                 btrfs_release_path(path);
2444
2445                 /*
2446                  * Mutex was contended, block until it's released and let
2447                  * caller try again
2448                  */
2449                 mutex_lock(&head->mutex);
2450                 mutex_unlock(&head->mutex);
2451                 btrfs_put_delayed_ref(&head->node);
2452                 return -EAGAIN;
2453         }
2454
2455         node = rb_prev(&head->node.rb_node);
2456         if (!node)
2457                 goto out_unlock;
2458
2459         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2460
2461         if (ref->bytenr != bytenr)
2462                 goto out_unlock;
2463
2464         ret = 1;
2465         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2466                 goto out_unlock;
2467
2468         data_ref = btrfs_delayed_node_to_data_ref(ref);
2469
2470         node = rb_prev(node);
2471         if (node) {
2472                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2473                 if (ref->bytenr == bytenr)
2474                         goto out_unlock;
2475         }
2476
2477         if (data_ref->root != root->root_key.objectid ||
2478             data_ref->objectid != objectid || data_ref->offset != offset)
2479                 goto out_unlock;
2480
2481         ret = 0;
2482 out_unlock:
2483         mutex_unlock(&head->mutex);
2484 out:
2485         spin_unlock(&delayed_refs->lock);
2486         return ret;
2487 }
2488
2489 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2490                                         struct btrfs_root *root,
2491                                         struct btrfs_path *path,
2492                                         u64 objectid, u64 offset, u64 bytenr)
2493 {
2494         struct btrfs_root *extent_root = root->fs_info->extent_root;
2495         struct extent_buffer *leaf;
2496         struct btrfs_extent_data_ref *ref;
2497         struct btrfs_extent_inline_ref *iref;
2498         struct btrfs_extent_item *ei;
2499         struct btrfs_key key;
2500         u32 item_size;
2501         int ret;
2502
2503         key.objectid = bytenr;
2504         key.offset = (u64)-1;
2505         key.type = BTRFS_EXTENT_ITEM_KEY;
2506
2507         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2508         if (ret < 0)
2509                 goto out;
2510         BUG_ON(ret == 0);
2511
2512         ret = -ENOENT;
2513         if (path->slots[0] == 0)
2514                 goto out;
2515
2516         path->slots[0]--;
2517         leaf = path->nodes[0];
2518         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2519
2520         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2521                 goto out;
2522
2523         ret = 1;
2524         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2525 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2526         if (item_size < sizeof(*ei)) {
2527                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2528                 goto out;
2529         }
2530 #endif
2531         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2532
2533         if (item_size != sizeof(*ei) +
2534             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2535                 goto out;
2536
2537         if (btrfs_extent_generation(leaf, ei) <=
2538             btrfs_root_last_snapshot(&root->root_item))
2539                 goto out;
2540
2541         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2542         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2543             BTRFS_EXTENT_DATA_REF_KEY)
2544                 goto out;
2545
2546         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2547         if (btrfs_extent_refs(leaf, ei) !=
2548             btrfs_extent_data_ref_count(leaf, ref) ||
2549             btrfs_extent_data_ref_root(leaf, ref) !=
2550             root->root_key.objectid ||
2551             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2552             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2553                 goto out;
2554
2555         ret = 0;
2556 out:
2557         return ret;
2558 }
2559
2560 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2561                           struct btrfs_root *root,
2562                           u64 objectid, u64 offset, u64 bytenr)
2563 {
2564         struct btrfs_path *path;
2565         int ret;
2566         int ret2;
2567
2568         path = btrfs_alloc_path();
2569         if (!path)
2570                 return -ENOENT;
2571
2572         do {
2573                 ret = check_committed_ref(trans, root, path, objectid,
2574                                           offset, bytenr);
2575                 if (ret && ret != -ENOENT)
2576                         goto out;
2577
2578                 ret2 = check_delayed_ref(trans, root, path, objectid,
2579                                          offset, bytenr);
2580         } while (ret2 == -EAGAIN);
2581
2582         if (ret2 && ret2 != -ENOENT) {
2583                 ret = ret2;
2584                 goto out;
2585         }
2586
2587         if (ret != -ENOENT || ret2 != -ENOENT)
2588                 ret = 0;
2589 out:
2590         btrfs_free_path(path);
2591         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2592                 WARN_ON(ret > 0);
2593         return ret;
2594 }
2595
2596 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2597                            struct btrfs_root *root,
2598                            struct extent_buffer *buf,
2599                            int full_backref, int inc)
2600 {
2601         u64 bytenr;
2602         u64 num_bytes;
2603         u64 parent;
2604         u64 ref_root;
2605         u32 nritems;
2606         struct btrfs_key key;
2607         struct btrfs_file_extent_item *fi;
2608         int i;
2609         int level;
2610         int ret = 0;
2611         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2612                             u64, u64, u64, u64, u64, u64);
2613
2614         ref_root = btrfs_header_owner(buf);
2615         nritems = btrfs_header_nritems(buf);
2616         level = btrfs_header_level(buf);
2617
2618         if (!root->ref_cows && level == 0)
2619                 return 0;
2620
2621         if (inc)
2622                 process_func = btrfs_inc_extent_ref;
2623         else
2624                 process_func = btrfs_free_extent;
2625
2626         if (full_backref)
2627                 parent = buf->start;
2628         else
2629                 parent = 0;
2630
2631         for (i = 0; i < nritems; i++) {
2632                 if (level == 0) {
2633                         btrfs_item_key_to_cpu(buf, &key, i);
2634                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2635                                 continue;
2636                         fi = btrfs_item_ptr(buf, i,
2637                                             struct btrfs_file_extent_item);
2638                         if (btrfs_file_extent_type(buf, fi) ==
2639                             BTRFS_FILE_EXTENT_INLINE)
2640                                 continue;
2641                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2642                         if (bytenr == 0)
2643                                 continue;
2644
2645                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2646                         key.offset -= btrfs_file_extent_offset(buf, fi);
2647                         ret = process_func(trans, root, bytenr, num_bytes,
2648                                            parent, ref_root, key.objectid,
2649                                            key.offset);
2650                         if (ret)
2651                                 goto fail;
2652                 } else {
2653                         bytenr = btrfs_node_blockptr(buf, i);
2654                         num_bytes = btrfs_level_size(root, level - 1);
2655                         ret = process_func(trans, root, bytenr, num_bytes,
2656                                            parent, ref_root, level - 1, 0);
2657                         if (ret)
2658                                 goto fail;
2659                 }
2660         }
2661         return 0;
2662 fail:
2663         BUG();
2664         return ret;
2665 }
2666
2667 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2668                   struct extent_buffer *buf, int full_backref)
2669 {
2670         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2671 }
2672
2673 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2674                   struct extent_buffer *buf, int full_backref)
2675 {
2676         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2677 }
2678
2679 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2680                                  struct btrfs_root *root,
2681                                  struct btrfs_path *path,
2682                                  struct btrfs_block_group_cache *cache)
2683 {
2684         int ret;
2685         struct btrfs_root *extent_root = root->fs_info->extent_root;
2686         unsigned long bi;
2687         struct extent_buffer *leaf;
2688
2689         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2690         if (ret < 0)
2691                 goto fail;
2692         BUG_ON(ret);
2693
2694         leaf = path->nodes[0];
2695         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2696         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2697         btrfs_mark_buffer_dirty(leaf);
2698         btrfs_release_path(path);
2699 fail:
2700         if (ret)
2701                 return ret;
2702         return 0;
2703
2704 }
2705
2706 static struct btrfs_block_group_cache *
2707 next_block_group(struct btrfs_root *root,
2708                  struct btrfs_block_group_cache *cache)
2709 {
2710         struct rb_node *node;
2711         spin_lock(&root->fs_info->block_group_cache_lock);
2712         node = rb_next(&cache->cache_node);
2713         btrfs_put_block_group(cache);
2714         if (node) {
2715                 cache = rb_entry(node, struct btrfs_block_group_cache,
2716                                  cache_node);
2717                 btrfs_get_block_group(cache);
2718         } else
2719                 cache = NULL;
2720         spin_unlock(&root->fs_info->block_group_cache_lock);
2721         return cache;
2722 }
2723
2724 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2725                             struct btrfs_trans_handle *trans,
2726                             struct btrfs_path *path)
2727 {
2728         struct btrfs_root *root = block_group->fs_info->tree_root;
2729         struct inode *inode = NULL;
2730         u64 alloc_hint = 0;
2731         int dcs = BTRFS_DC_ERROR;
2732         int num_pages = 0;
2733         int retries = 0;
2734         int ret = 0;
2735
2736         /*
2737          * If this block group is smaller than 100 megs don't bother caching the
2738          * block group.
2739          */
2740         if (block_group->key.offset < (100 * 1024 * 1024)) {
2741                 spin_lock(&block_group->lock);
2742                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2743                 spin_unlock(&block_group->lock);
2744                 return 0;
2745         }
2746
2747 again:
2748         inode = lookup_free_space_inode(root, block_group, path);
2749         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2750                 ret = PTR_ERR(inode);
2751                 btrfs_release_path(path);
2752                 goto out;
2753         }
2754
2755         if (IS_ERR(inode)) {
2756                 BUG_ON(retries);
2757                 retries++;
2758
2759                 if (block_group->ro)
2760                         goto out_free;
2761
2762                 ret = create_free_space_inode(root, trans, block_group, path);
2763                 if (ret)
2764                         goto out_free;
2765                 goto again;
2766         }
2767
2768         /* We've already setup this transaction, go ahead and exit */
2769         if (block_group->cache_generation == trans->transid &&
2770             i_size_read(inode)) {
2771                 dcs = BTRFS_DC_SETUP;
2772                 goto out_put;
2773         }
2774
2775         /*
2776          * We want to set the generation to 0, that way if anything goes wrong
2777          * from here on out we know not to trust this cache when we load up next
2778          * time.
2779          */
2780         BTRFS_I(inode)->generation = 0;
2781         ret = btrfs_update_inode(trans, root, inode);
2782         WARN_ON(ret);
2783
2784         if (i_size_read(inode) > 0) {
2785                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2786                                                       inode);
2787                 if (ret)
2788                         goto out_put;
2789         }
2790
2791         spin_lock(&block_group->lock);
2792         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2793                 /* We're not cached, don't bother trying to write stuff out */
2794                 dcs = BTRFS_DC_WRITTEN;
2795                 spin_unlock(&block_group->lock);
2796                 goto out_put;
2797         }
2798         spin_unlock(&block_group->lock);
2799
2800         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2801         if (!num_pages)
2802                 num_pages = 1;
2803
2804         /*
2805          * Just to make absolutely sure we have enough space, we're going to
2806          * preallocate 12 pages worth of space for each block group.  In
2807          * practice we ought to use at most 8, but we need extra space so we can
2808          * add our header and have a terminator between the extents and the
2809          * bitmaps.
2810          */
2811         num_pages *= 16;
2812         num_pages *= PAGE_CACHE_SIZE;
2813
2814         ret = btrfs_check_data_free_space(inode, num_pages);
2815         if (ret)
2816                 goto out_put;
2817
2818         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2819                                               num_pages, num_pages,
2820                                               &alloc_hint);
2821         if (!ret)
2822                 dcs = BTRFS_DC_SETUP;
2823         btrfs_free_reserved_data_space(inode, num_pages);
2824
2825 out_put:
2826         iput(inode);
2827 out_free:
2828         btrfs_release_path(path);
2829 out:
2830         spin_lock(&block_group->lock);
2831         if (!ret && dcs == BTRFS_DC_SETUP)
2832                 block_group->cache_generation = trans->transid;
2833         block_group->disk_cache_state = dcs;
2834         spin_unlock(&block_group->lock);
2835
2836         return ret;
2837 }
2838
2839 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2840                                    struct btrfs_root *root)
2841 {
2842         struct btrfs_block_group_cache *cache;
2843         int err = 0;
2844         struct btrfs_path *path;
2845         u64 last = 0;
2846
2847         path = btrfs_alloc_path();
2848         if (!path)
2849                 return -ENOMEM;
2850
2851 again:
2852         while (1) {
2853                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2854                 while (cache) {
2855                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2856                                 break;
2857                         cache = next_block_group(root, cache);
2858                 }
2859                 if (!cache) {
2860                         if (last == 0)
2861                                 break;
2862                         last = 0;
2863                         continue;
2864                 }
2865                 err = cache_save_setup(cache, trans, path);
2866                 last = cache->key.objectid + cache->key.offset;
2867                 btrfs_put_block_group(cache);
2868         }
2869
2870         while (1) {
2871                 if (last == 0) {
2872                         err = btrfs_run_delayed_refs(trans, root,
2873                                                      (unsigned long)-1);
2874                         BUG_ON(err);
2875                 }
2876
2877                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2878                 while (cache) {
2879                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2880                                 btrfs_put_block_group(cache);
2881                                 goto again;
2882                         }
2883
2884                         if (cache->dirty)
2885                                 break;
2886                         cache = next_block_group(root, cache);
2887                 }
2888                 if (!cache) {
2889                         if (last == 0)
2890                                 break;
2891                         last = 0;
2892                         continue;
2893                 }
2894
2895                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
2896                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
2897                 cache->dirty = 0;
2898                 last = cache->key.objectid + cache->key.offset;
2899
2900                 err = write_one_cache_group(trans, root, path, cache);
2901                 BUG_ON(err);
2902                 btrfs_put_block_group(cache);
2903         }
2904
2905         while (1) {
2906                 /*
2907                  * I don't think this is needed since we're just marking our
2908                  * preallocated extent as written, but just in case it can't
2909                  * hurt.
2910                  */
2911                 if (last == 0) {
2912                         err = btrfs_run_delayed_refs(trans, root,
2913                                                      (unsigned long)-1);
2914                         BUG_ON(err);
2915                 }
2916
2917                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2918                 while (cache) {
2919                         /*
2920                          * Really this shouldn't happen, but it could if we
2921                          * couldn't write the entire preallocated extent and
2922                          * splitting the extent resulted in a new block.
2923                          */
2924                         if (cache->dirty) {
2925                                 btrfs_put_block_group(cache);
2926                                 goto again;
2927                         }
2928                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2929                                 break;
2930                         cache = next_block_group(root, cache);
2931                 }
2932                 if (!cache) {
2933                         if (last == 0)
2934                                 break;
2935                         last = 0;
2936                         continue;
2937                 }
2938
2939                 btrfs_write_out_cache(root, trans, cache, path);
2940
2941                 /*
2942                  * If we didn't have an error then the cache state is still
2943                  * NEED_WRITE, so we can set it to WRITTEN.
2944                  */
2945                 if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2946                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
2947                 last = cache->key.objectid + cache->key.offset;
2948                 btrfs_put_block_group(cache);
2949         }
2950
2951         btrfs_free_path(path);
2952         return 0;
2953 }
2954
2955 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
2956 {
2957         struct btrfs_block_group_cache *block_group;
2958         int readonly = 0;
2959
2960         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
2961         if (!block_group || block_group->ro)
2962                 readonly = 1;
2963         if (block_group)
2964                 btrfs_put_block_group(block_group);
2965         return readonly;
2966 }
2967
2968 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
2969                              u64 total_bytes, u64 bytes_used,
2970                              struct btrfs_space_info **space_info)
2971 {
2972         struct btrfs_space_info *found;
2973         int i;
2974         int factor;
2975
2976         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
2977                      BTRFS_BLOCK_GROUP_RAID10))
2978                 factor = 2;
2979         else
2980                 factor = 1;
2981
2982         found = __find_space_info(info, flags);
2983         if (found) {
2984                 spin_lock(&found->lock);
2985                 found->total_bytes += total_bytes;
2986                 found->disk_total += total_bytes * factor;
2987                 found->bytes_used += bytes_used;
2988                 found->disk_used += bytes_used * factor;
2989                 found->full = 0;
2990                 spin_unlock(&found->lock);
2991                 *space_info = found;
2992                 return 0;
2993         }
2994         found = kzalloc(sizeof(*found), GFP_NOFS);
2995         if (!found)
2996                 return -ENOMEM;
2997
2998         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
2999                 INIT_LIST_HEAD(&found->block_groups[i]);
3000         init_rwsem(&found->groups_sem);
3001         spin_lock_init(&found->lock);
3002         found->flags = flags & (BTRFS_BLOCK_GROUP_DATA |
3003                                 BTRFS_BLOCK_GROUP_SYSTEM |
3004                                 BTRFS_BLOCK_GROUP_METADATA);
3005         found->total_bytes = total_bytes;
3006         found->disk_total = total_bytes * factor;
3007         found->bytes_used = bytes_used;
3008         found->disk_used = bytes_used * factor;
3009         found->bytes_pinned = 0;
3010         found->bytes_reserved = 0;
3011         found->bytes_readonly = 0;
3012         found->bytes_may_use = 0;
3013         found->full = 0;
3014         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3015         found->chunk_alloc = 0;
3016         found->flush = 0;
3017         init_waitqueue_head(&found->wait);
3018         *space_info = found;
3019         list_add_rcu(&found->list, &info->space_info);
3020         return 0;
3021 }
3022
3023 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3024 {
3025         u64 extra_flags = flags & (BTRFS_BLOCK_GROUP_RAID0 |
3026                                    BTRFS_BLOCK_GROUP_RAID1 |
3027                                    BTRFS_BLOCK_GROUP_RAID10 |
3028                                    BTRFS_BLOCK_GROUP_DUP);
3029         if (extra_flags) {
3030                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3031                         fs_info->avail_data_alloc_bits |= extra_flags;
3032                 if (flags & BTRFS_BLOCK_GROUP_METADATA)
3033                         fs_info->avail_metadata_alloc_bits |= extra_flags;
3034                 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3035                         fs_info->avail_system_alloc_bits |= extra_flags;
3036         }
3037 }
3038
3039 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3040 {
3041         /*
3042          * we add in the count of missing devices because we want
3043          * to make sure that any RAID levels on a degraded FS
3044          * continue to be honored.
3045          */
3046         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3047                 root->fs_info->fs_devices->missing_devices;
3048
3049         if (num_devices == 1)
3050                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3051         if (num_devices < 4)
3052                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3053
3054         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3055             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3056                       BTRFS_BLOCK_GROUP_RAID10))) {
3057                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3058         }
3059
3060         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3061             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3062                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3063         }
3064
3065         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3066             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3067              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3068              (flags & BTRFS_BLOCK_GROUP_DUP)))
3069                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3070         return flags;
3071 }
3072
3073 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3074 {
3075         if (flags & BTRFS_BLOCK_GROUP_DATA)
3076                 flags |= root->fs_info->avail_data_alloc_bits &
3077                          root->fs_info->data_alloc_profile;
3078         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3079                 flags |= root->fs_info->avail_system_alloc_bits &
3080                          root->fs_info->system_alloc_profile;
3081         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3082                 flags |= root->fs_info->avail_metadata_alloc_bits &
3083                          root->fs_info->metadata_alloc_profile;
3084         return btrfs_reduce_alloc_profile(root, flags);
3085 }
3086
3087 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3088 {
3089         u64 flags;
3090
3091         if (data)
3092                 flags = BTRFS_BLOCK_GROUP_DATA;
3093         else if (root == root->fs_info->chunk_root)
3094                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3095         else
3096                 flags = BTRFS_BLOCK_GROUP_METADATA;
3097
3098         return get_alloc_profile(root, flags);
3099 }
3100
3101 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3102 {
3103         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3104                                                        BTRFS_BLOCK_GROUP_DATA);
3105 }
3106
3107 /*
3108  * This will check the space that the inode allocates from to make sure we have
3109  * enough space for bytes.
3110  */
3111 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3112 {
3113         struct btrfs_space_info *data_sinfo;
3114         struct btrfs_root *root = BTRFS_I(inode)->root;
3115         u64 used;
3116         int ret = 0, committed = 0, alloc_chunk = 1;
3117
3118         /* make sure bytes are sectorsize aligned */
3119         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3120
3121         if (root == root->fs_info->tree_root ||
3122             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3123                 alloc_chunk = 0;
3124                 committed = 1;
3125         }
3126
3127         data_sinfo = BTRFS_I(inode)->space_info;
3128         if (!data_sinfo)
3129                 goto alloc;
3130
3131 again:
3132         /* make sure we have enough space to handle the data first */
3133         spin_lock(&data_sinfo->lock);
3134         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3135                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3136                 data_sinfo->bytes_may_use;
3137
3138         if (used + bytes > data_sinfo->total_bytes) {
3139                 struct btrfs_trans_handle *trans;
3140
3141                 /*
3142                  * if we don't have enough free bytes in this space then we need
3143                  * to alloc a new chunk.
3144                  */
3145                 if (!data_sinfo->full && alloc_chunk) {
3146                         u64 alloc_target;
3147
3148                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3149                         spin_unlock(&data_sinfo->lock);
3150 alloc:
3151                         alloc_target = btrfs_get_alloc_profile(root, 1);
3152                         trans = btrfs_join_transaction(root);
3153                         if (IS_ERR(trans))
3154                                 return PTR_ERR(trans);
3155
3156                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3157                                              bytes + 2 * 1024 * 1024,
3158                                              alloc_target,
3159                                              CHUNK_ALLOC_NO_FORCE);
3160                         btrfs_end_transaction(trans, root);
3161                         if (ret < 0) {
3162                                 if (ret != -ENOSPC)
3163                                         return ret;
3164                                 else
3165                                         goto commit_trans;
3166                         }
3167
3168                         if (!data_sinfo) {
3169                                 btrfs_set_inode_space_info(root, inode);
3170                                 data_sinfo = BTRFS_I(inode)->space_info;
3171                         }
3172                         goto again;
3173                 }
3174
3175                 /*
3176                  * If we have less pinned bytes than we want to allocate then
3177                  * don't bother committing the transaction, it won't help us.
3178                  */
3179                 if (data_sinfo->bytes_pinned < bytes)
3180                         committed = 1;
3181                 spin_unlock(&data_sinfo->lock);
3182
3183                 /* commit the current transaction and try again */
3184 commit_trans:
3185                 if (!committed &&
3186                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3187                         committed = 1;
3188                         trans = btrfs_join_transaction(root);
3189                         if (IS_ERR(trans))
3190                                 return PTR_ERR(trans);
3191                         ret = btrfs_commit_transaction(trans, root);
3192                         if (ret)
3193                                 return ret;
3194                         goto again;
3195                 }
3196
3197                 return -ENOSPC;
3198         }
3199         data_sinfo->bytes_may_use += bytes;
3200         spin_unlock(&data_sinfo->lock);
3201
3202         return 0;
3203 }
3204
3205 /*
3206  * Called if we need to clear a data reservation for this inode.
3207  */
3208 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3209 {
3210         struct btrfs_root *root = BTRFS_I(inode)->root;
3211         struct btrfs_space_info *data_sinfo;
3212
3213         /* make sure bytes are sectorsize aligned */
3214         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3215
3216         data_sinfo = BTRFS_I(inode)->space_info;
3217         spin_lock(&data_sinfo->lock);
3218         data_sinfo->bytes_may_use -= bytes;
3219         spin_unlock(&data_sinfo->lock);
3220 }
3221
3222 static void force_metadata_allocation(struct btrfs_fs_info *info)
3223 {
3224         struct list_head *head = &info->space_info;
3225         struct btrfs_space_info *found;
3226
3227         rcu_read_lock();
3228         list_for_each_entry_rcu(found, head, list) {
3229                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3230                         found->force_alloc = CHUNK_ALLOC_FORCE;
3231         }
3232         rcu_read_unlock();
3233 }
3234
3235 static int should_alloc_chunk(struct btrfs_root *root,
3236                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3237                               int force)
3238 {
3239         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3240         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3241         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3242         u64 thresh;
3243
3244         if (force == CHUNK_ALLOC_FORCE)
3245                 return 1;
3246
3247         /*
3248          * We need to take into account the global rsv because for all intents
3249          * and purposes it's used space.  Don't worry about locking the
3250          * global_rsv, it doesn't change except when the transaction commits.
3251          */
3252         num_allocated += global_rsv->size;
3253
3254         /*
3255          * in limited mode, we want to have some free space up to
3256          * about 1% of the FS size.
3257          */
3258         if (force == CHUNK_ALLOC_LIMITED) {
3259                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3260                 thresh = max_t(u64, 64 * 1024 * 1024,
3261                                div_factor_fine(thresh, 1));
3262
3263                 if (num_bytes - num_allocated < thresh)
3264                         return 1;
3265         }
3266
3267         /*
3268          * we have two similar checks here, one based on percentage
3269          * and once based on a hard number of 256MB.  The idea
3270          * is that if we have a good amount of free
3271          * room, don't allocate a chunk.  A good mount is
3272          * less than 80% utilized of the chunks we have allocated,
3273          * or more than 256MB free
3274          */
3275         if (num_allocated + alloc_bytes + 256 * 1024 * 1024 < num_bytes)
3276                 return 0;
3277
3278         if (num_allocated + alloc_bytes < div_factor(num_bytes, 8))
3279                 return 0;
3280
3281         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3282
3283         /* 256MB or 5% of the FS */
3284         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 5));
3285
3286         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 3))
3287                 return 0;
3288         return 1;
3289 }
3290
3291 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3292                           struct btrfs_root *extent_root, u64 alloc_bytes,
3293                           u64 flags, int force)
3294 {
3295         struct btrfs_space_info *space_info;
3296         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3297         int wait_for_alloc = 0;
3298         int ret = 0;
3299
3300         flags = btrfs_reduce_alloc_profile(extent_root, flags);
3301
3302         space_info = __find_space_info(extent_root->fs_info, flags);
3303         if (!space_info) {
3304                 ret = update_space_info(extent_root->fs_info, flags,
3305                                         0, 0, &space_info);
3306                 BUG_ON(ret);
3307         }
3308         BUG_ON(!space_info);
3309
3310 again:
3311         spin_lock(&space_info->lock);
3312         if (space_info->force_alloc)
3313                 force = space_info->force_alloc;
3314         if (space_info->full) {
3315                 spin_unlock(&space_info->lock);
3316                 return 0;
3317         }
3318
3319         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3320                 spin_unlock(&space_info->lock);
3321                 return 0;
3322         } else if (space_info->chunk_alloc) {
3323                 wait_for_alloc = 1;
3324         } else {
3325                 space_info->chunk_alloc = 1;
3326         }
3327
3328         spin_unlock(&space_info->lock);
3329
3330         mutex_lock(&fs_info->chunk_mutex);
3331
3332         /*
3333          * The chunk_mutex is held throughout the entirety of a chunk
3334          * allocation, so once we've acquired the chunk_mutex we know that the
3335          * other guy is done and we need to recheck and see if we should
3336          * allocate.
3337          */
3338         if (wait_for_alloc) {
3339                 mutex_unlock(&fs_info->chunk_mutex);
3340                 wait_for_alloc = 0;
3341                 goto again;
3342         }
3343
3344         /*
3345          * If we have mixed data/metadata chunks we want to make sure we keep
3346          * allocating mixed chunks instead of individual chunks.
3347          */
3348         if (btrfs_mixed_space_info(space_info))
3349                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3350
3351         /*
3352          * if we're doing a data chunk, go ahead and make sure that
3353          * we keep a reasonable number of metadata chunks allocated in the
3354          * FS as well.
3355          */
3356         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3357                 fs_info->data_chunk_allocations++;
3358                 if (!(fs_info->data_chunk_allocations %
3359                       fs_info->metadata_ratio))
3360                         force_metadata_allocation(fs_info);
3361         }
3362
3363         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3364         if (ret < 0 && ret != -ENOSPC)
3365                 goto out;
3366
3367         spin_lock(&space_info->lock);
3368         if (ret)
3369                 space_info->full = 1;
3370         else
3371                 ret = 1;
3372
3373         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3374         space_info->chunk_alloc = 0;
3375         spin_unlock(&space_info->lock);
3376 out:
3377         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3378         return ret;
3379 }
3380
3381 /*
3382  * shrink metadata reservation for delalloc
3383  */
3384 static int shrink_delalloc(struct btrfs_root *root, u64 to_reclaim,
3385                            bool wait_ordered)
3386 {
3387         struct btrfs_block_rsv *block_rsv;
3388         struct btrfs_space_info *space_info;
3389         struct btrfs_trans_handle *trans;
3390         u64 reserved;
3391         u64 max_reclaim;
3392         u64 reclaimed = 0;
3393         long time_left;
3394         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3395         int loops = 0;
3396         unsigned long progress;
3397
3398         trans = (struct btrfs_trans_handle *)current->journal_info;
3399         block_rsv = &root->fs_info->delalloc_block_rsv;
3400         space_info = block_rsv->space_info;
3401
3402         smp_mb();
3403         reserved = space_info->bytes_may_use;
3404         progress = space_info->reservation_progress;
3405
3406         if (reserved == 0)
3407                 return 0;
3408
3409         smp_mb();
3410         if (root->fs_info->delalloc_bytes == 0) {
3411                 if (trans)
3412                         return 0;
3413                 btrfs_wait_ordered_extents(root, 0, 0);
3414                 return 0;
3415         }
3416
3417         max_reclaim = min(reserved, to_reclaim);
3418         nr_pages = max_t(unsigned long, nr_pages,
3419                          max_reclaim >> PAGE_CACHE_SHIFT);
3420         while (loops < 1024) {
3421                 /* have the flusher threads jump in and do some IO */
3422                 smp_mb();
3423                 nr_pages = min_t(unsigned long, nr_pages,
3424                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3425                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3426                                                 WB_REASON_FS_FREE_SPACE);
3427
3428                 spin_lock(&space_info->lock);
3429                 if (reserved > space_info->bytes_may_use)
3430                         reclaimed += reserved - space_info->bytes_may_use;
3431                 reserved = space_info->bytes_may_use;
3432                 spin_unlock(&space_info->lock);
3433
3434                 loops++;
3435
3436                 if (reserved == 0 || reclaimed >= max_reclaim)
3437                         break;
3438
3439                 if (trans && trans->transaction->blocked)
3440                         return -EAGAIN;
3441
3442                 if (wait_ordered && !trans) {
3443                         btrfs_wait_ordered_extents(root, 0, 0);
3444                 } else {
3445                         time_left = schedule_timeout_interruptible(1);
3446
3447                         /* We were interrupted, exit */
3448                         if (time_left)
3449                                 break;
3450                 }
3451
3452                 /* we've kicked the IO a few times, if anything has been freed,
3453                  * exit.  There is no sense in looping here for a long time
3454                  * when we really need to commit the transaction, or there are
3455                  * just too many writers without enough free space
3456                  */
3457
3458                 if (loops > 3) {
3459                         smp_mb();
3460                         if (progress != space_info->reservation_progress)
3461                                 break;
3462                 }
3463
3464         }
3465
3466         return reclaimed >= to_reclaim;
3467 }
3468
3469 /**
3470  * maybe_commit_transaction - possibly commit the transaction if its ok to
3471  * @root - the root we're allocating for
3472  * @bytes - the number of bytes we want to reserve
3473  * @force - force the commit
3474  *
3475  * This will check to make sure that committing the transaction will actually
3476  * get us somewhere and then commit the transaction if it does.  Otherwise it
3477  * will return -ENOSPC.
3478  */
3479 static int may_commit_transaction(struct btrfs_root *root,
3480                                   struct btrfs_space_info *space_info,
3481                                   u64 bytes, int force)
3482 {
3483         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3484         struct btrfs_trans_handle *trans;
3485
3486         trans = (struct btrfs_trans_handle *)current->journal_info;
3487         if (trans)
3488                 return -EAGAIN;
3489
3490         if (force)
3491                 goto commit;
3492
3493         /* See if there is enough pinned space to make this reservation */
3494         spin_lock(&space_info->lock);
3495         if (space_info->bytes_pinned >= bytes) {
3496                 spin_unlock(&space_info->lock);
3497                 goto commit;
3498         }
3499         spin_unlock(&space_info->lock);
3500
3501         /*
3502          * See if there is some space in the delayed insertion reservation for
3503          * this reservation.
3504          */
3505         if (space_info != delayed_rsv->space_info)
3506                 return -ENOSPC;
3507
3508         spin_lock(&delayed_rsv->lock);
3509         if (delayed_rsv->size < bytes) {
3510                 spin_unlock(&delayed_rsv->lock);
3511                 return -ENOSPC;
3512         }
3513         spin_unlock(&delayed_rsv->lock);
3514
3515 commit:
3516         trans = btrfs_join_transaction(root);
3517         if (IS_ERR(trans))
3518                 return -ENOSPC;
3519
3520         return btrfs_commit_transaction(trans, root);
3521 }
3522
3523 /**
3524  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3525  * @root - the root we're allocating for
3526  * @block_rsv - the block_rsv we're allocating for
3527  * @orig_bytes - the number of bytes we want
3528  * @flush - wether or not we can flush to make our reservation
3529  *
3530  * This will reserve orgi_bytes number of bytes from the space info associated
3531  * with the block_rsv.  If there is not enough space it will make an attempt to
3532  * flush out space to make room.  It will do this by flushing delalloc if
3533  * possible or committing the transaction.  If flush is 0 then no attempts to
3534  * regain reservations will be made and this will fail if there is not enough
3535  * space already.
3536  */
3537 static int reserve_metadata_bytes(struct btrfs_root *root,
3538                                   struct btrfs_block_rsv *block_rsv,
3539                                   u64 orig_bytes, int flush)
3540 {
3541         struct btrfs_space_info *space_info = block_rsv->space_info;
3542         u64 used;
3543         u64 num_bytes = orig_bytes;
3544         int retries = 0;
3545         int ret = 0;
3546         bool committed = false;
3547         bool flushing = false;
3548         bool wait_ordered = false;
3549
3550 again:
3551         ret = 0;
3552         spin_lock(&space_info->lock);
3553         /*
3554          * We only want to wait if somebody other than us is flushing and we are
3555          * actually alloed to flush.
3556          */
3557         while (flush && !flushing && space_info->flush) {
3558                 spin_unlock(&space_info->lock);
3559                 /*
3560                  * If we have a trans handle we can't wait because the flusher
3561                  * may have to commit the transaction, which would mean we would
3562                  * deadlock since we are waiting for the flusher to finish, but
3563                  * hold the current transaction open.
3564                  */
3565                 if (current->journal_info)
3566                         return -EAGAIN;
3567                 ret = wait_event_interruptible(space_info->wait,
3568                                                !space_info->flush);
3569                 /* Must have been interrupted, return */
3570                 if (ret)
3571                         return -EINTR;
3572
3573                 spin_lock(&space_info->lock);
3574         }
3575
3576         ret = -ENOSPC;
3577         used = space_info->bytes_used + space_info->bytes_reserved +
3578                 space_info->bytes_pinned + space_info->bytes_readonly +
3579                 space_info->bytes_may_use;
3580
3581         /*
3582          * The idea here is that we've not already over-reserved the block group
3583          * then we can go ahead and save our reservation first and then start
3584          * flushing if we need to.  Otherwise if we've already overcommitted
3585          * lets start flushing stuff first and then come back and try to make
3586          * our reservation.
3587          */
3588         if (used <= space_info->total_bytes) {
3589                 if (used + orig_bytes <= space_info->total_bytes) {
3590                         space_info->bytes_may_use += orig_bytes;
3591                         ret = 0;
3592                 } else {
3593                         /*
3594                          * Ok set num_bytes to orig_bytes since we aren't
3595                          * overocmmitted, this way we only try and reclaim what
3596                          * we need.
3597                          */
3598                         num_bytes = orig_bytes;
3599                 }
3600         } else {
3601                 /*
3602                  * Ok we're over committed, set num_bytes to the overcommitted
3603                  * amount plus the amount of bytes that we need for this
3604                  * reservation.
3605                  */
3606                 wait_ordered = true;
3607                 num_bytes = used - space_info->total_bytes +
3608                         (orig_bytes * (retries + 1));
3609         }
3610
3611         if (ret) {
3612                 u64 profile = btrfs_get_alloc_profile(root, 0);
3613                 u64 avail;
3614
3615                 /*
3616                  * If we have a lot of space that's pinned, don't bother doing
3617                  * the overcommit dance yet and just commit the transaction.
3618                  */
3619                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3620                 do_div(avail, 10);
3621                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3622                         space_info->flush = 1;
3623                         flushing = true;
3624                         spin_unlock(&space_info->lock);
3625                         ret = may_commit_transaction(root, space_info,
3626                                                      orig_bytes, 1);
3627                         if (ret)
3628                                 goto out;
3629                         committed = true;
3630                         goto again;
3631                 }
3632
3633                 spin_lock(&root->fs_info->free_chunk_lock);
3634                 avail = root->fs_info->free_chunk_space;
3635
3636                 /*
3637                  * If we have dup, raid1 or raid10 then only half of the free
3638                  * space is actually useable.
3639                  */
3640                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3641                                BTRFS_BLOCK_GROUP_RAID1 |
3642                                BTRFS_BLOCK_GROUP_RAID10))
3643                         avail >>= 1;
3644
3645                 /*
3646                  * If we aren't flushing don't let us overcommit too much, say
3647                  * 1/8th of the space.  If we can flush, let it overcommit up to
3648                  * 1/2 of the space.
3649                  */
3650                 if (flush)
3651                         avail >>= 3;
3652                 else
3653                         avail >>= 1;
3654                  spin_unlock(&root->fs_info->free_chunk_lock);
3655
3656                 if (used + num_bytes < space_info->total_bytes + avail) {
3657                         space_info->bytes_may_use += orig_bytes;
3658                         ret = 0;
3659                 } else {
3660                         wait_ordered = true;
3661                 }
3662         }
3663
3664         /*
3665          * Couldn't make our reservation, save our place so while we're trying
3666          * to reclaim space we can actually use it instead of somebody else
3667          * stealing it from us.
3668          */
3669         if (ret && flush) {
3670                 flushing = true;
3671                 space_info->flush = 1;
3672         }
3673
3674         spin_unlock(&space_info->lock);
3675
3676         if (!ret || !flush)
3677                 goto out;
3678
3679         /*
3680          * We do synchronous shrinking since we don't actually unreserve
3681          * metadata until after the IO is completed.
3682          */
3683         ret = shrink_delalloc(root, num_bytes, wait_ordered);
3684         if (ret < 0)
3685                 goto out;
3686
3687         ret = 0;
3688
3689         /*
3690          * So if we were overcommitted it's possible that somebody else flushed
3691          * out enough space and we simply didn't have enough space to reclaim,
3692          * so go back around and try again.
3693          */
3694         if (retries < 2) {
3695                 wait_ordered = true;
3696                 retries++;
3697                 goto again;
3698         }
3699
3700         ret = -ENOSPC;
3701         if (committed)
3702                 goto out;
3703
3704         ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3705         if (!ret) {
3706                 committed = true;
3707                 goto again;
3708         }
3709
3710 out:
3711         if (flushing) {
3712                 spin_lock(&space_info->lock);
3713                 space_info->flush = 0;
3714                 wake_up_all(&space_info->wait);
3715                 spin_unlock(&space_info->lock);
3716         }
3717         return ret;
3718 }
3719
3720 static struct btrfs_block_rsv *get_block_rsv(struct btrfs_trans_handle *trans,
3721                                              struct btrfs_root *root)
3722 {
3723         struct btrfs_block_rsv *block_rsv = NULL;
3724
3725         if (root->ref_cows || root == root->fs_info->csum_root)
3726                 block_rsv = trans->block_rsv;
3727
3728         if (!block_rsv)
3729                 block_rsv = root->block_rsv;
3730
3731         if (!block_rsv)
3732                 block_rsv = &root->fs_info->empty_block_rsv;
3733
3734         return block_rsv;
3735 }
3736
3737 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3738                                u64 num_bytes)
3739 {
3740         int ret = -ENOSPC;
3741         spin_lock(&block_rsv->lock);
3742         if (block_rsv->reserved >= num_bytes) {
3743                 block_rsv->reserved -= num_bytes;
3744                 if (block_rsv->reserved < block_rsv->size)
3745                         block_rsv->full = 0;
3746                 ret = 0;
3747         }
3748         spin_unlock(&block_rsv->lock);
3749         return ret;
3750 }
3751
3752 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3753                                 u64 num_bytes, int update_size)
3754 {
3755         spin_lock(&block_rsv->lock);
3756         block_rsv->reserved += num_bytes;
3757         if (update_size)
3758                 block_rsv->size += num_bytes;
3759         else if (block_rsv->reserved >= block_rsv->size)
3760                 block_rsv->full = 1;
3761         spin_unlock(&block_rsv->lock);
3762 }
3763
3764 static void block_rsv_release_bytes(struct btrfs_block_rsv *block_rsv,
3765                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3766 {
3767         struct btrfs_space_info *space_info = block_rsv->space_info;
3768
3769         spin_lock(&block_rsv->lock);
3770         if (num_bytes == (u64)-1)
3771                 num_bytes = block_rsv->size;
3772         block_rsv->size -= num_bytes;
3773         if (block_rsv->reserved >= block_rsv->size) {
3774                 num_bytes = block_rsv->reserved - block_rsv->size;
3775                 block_rsv->reserved = block_rsv->size;
3776                 block_rsv->full = 1;
3777         } else {
3778                 num_bytes = 0;
3779         }
3780         spin_unlock(&block_rsv->lock);
3781
3782         if (num_bytes > 0) {
3783                 if (dest) {
3784                         spin_lock(&dest->lock);
3785                         if (!dest->full) {
3786                                 u64 bytes_to_add;
3787
3788                                 bytes_to_add = dest->size - dest->reserved;
3789                                 bytes_to_add = min(num_bytes, bytes_to_add);
3790                                 dest->reserved += bytes_to_add;
3791                                 if (dest->reserved >= dest->size)
3792                                         dest->full = 1;
3793                                 num_bytes -= bytes_to_add;
3794                         }
3795                         spin_unlock(&dest->lock);
3796                 }
3797                 if (num_bytes) {
3798                         spin_lock(&space_info->lock);
3799                         space_info->bytes_may_use -= num_bytes;
3800                         space_info->reservation_progress++;
3801                         spin_unlock(&space_info->lock);
3802                 }
3803         }
3804 }
3805
3806 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3807                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3808 {
3809         int ret;
3810
3811         ret = block_rsv_use_bytes(src, num_bytes);
3812         if (ret)
3813                 return ret;
3814
3815         block_rsv_add_bytes(dst, num_bytes, 1);
3816         return 0;
3817 }
3818
3819 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3820 {
3821         memset(rsv, 0, sizeof(*rsv));
3822         spin_lock_init(&rsv->lock);
3823 }
3824
3825 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3826 {
3827         struct btrfs_block_rsv *block_rsv;
3828         struct btrfs_fs_info *fs_info = root->fs_info;
3829
3830         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3831         if (!block_rsv)
3832                 return NULL;
3833
3834         btrfs_init_block_rsv(block_rsv);
3835         block_rsv->space_info = __find_space_info(fs_info,
3836                                                   BTRFS_BLOCK_GROUP_METADATA);
3837         return block_rsv;
3838 }
3839
3840 void btrfs_free_block_rsv(struct btrfs_root *root,
3841                           struct btrfs_block_rsv *rsv)
3842 {
3843         btrfs_block_rsv_release(root, rsv, (u64)-1);
3844         kfree(rsv);
3845 }
3846
3847 static inline int __block_rsv_add(struct btrfs_root *root,
3848                                   struct btrfs_block_rsv *block_rsv,
3849                                   u64 num_bytes, int flush)
3850 {
3851         int ret;
3852
3853         if (num_bytes == 0)
3854                 return 0;
3855
3856         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
3857         if (!ret) {
3858                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3859                 return 0;
3860         }
3861
3862         return ret;
3863 }
3864
3865 int btrfs_block_rsv_add(struct btrfs_root *root,
3866                         struct btrfs_block_rsv *block_rsv,
3867                         u64 num_bytes)
3868 {
3869         return __block_rsv_add(root, block_rsv, num_bytes, 1);
3870 }
3871
3872 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
3873                                 struct btrfs_block_rsv *block_rsv,
3874                                 u64 num_bytes)
3875 {
3876         return __block_rsv_add(root, block_rsv, num_bytes, 0);
3877 }
3878
3879 int btrfs_block_rsv_check(struct btrfs_root *root,
3880                           struct btrfs_block_rsv *block_rsv, int min_factor)
3881 {
3882         u64 num_bytes = 0;
3883         int ret = -ENOSPC;
3884
3885         if (!block_rsv)
3886                 return 0;
3887
3888         spin_lock(&block_rsv->lock);
3889         num_bytes = div_factor(block_rsv->size, min_factor);
3890         if (block_rsv->reserved >= num_bytes)
3891                 ret = 0;
3892         spin_unlock(&block_rsv->lock);
3893
3894         return ret;
3895 }
3896
3897 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
3898                                            struct btrfs_block_rsv *block_rsv,
3899                                            u64 min_reserved, int flush)
3900 {
3901         u64 num_bytes = 0;
3902         int ret = -ENOSPC;
3903
3904         if (!block_rsv)
3905                 return 0;
3906
3907         spin_lock(&block_rsv->lock);
3908         num_bytes = min_reserved;
3909         if (block_rsv->reserved >= num_bytes)
3910                 ret = 0;
3911         else
3912                 num_bytes -= block_rsv->reserved;
3913         spin_unlock(&block_rsv->lock);
3914
3915         if (!ret)
3916                 return 0;
3917
3918         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
3919         if (!ret) {
3920                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
3921                 return 0;
3922         }
3923
3924         return ret;
3925 }
3926
3927 int btrfs_block_rsv_refill(struct btrfs_root *root,
3928                            struct btrfs_block_rsv *block_rsv,
3929                            u64 min_reserved)
3930 {
3931         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
3932 }
3933
3934 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
3935                                    struct btrfs_block_rsv *block_rsv,
3936                                    u64 min_reserved)
3937 {
3938         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
3939 }
3940
3941 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
3942                             struct btrfs_block_rsv *dst_rsv,
3943                             u64 num_bytes)
3944 {
3945         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3946 }
3947
3948 void btrfs_block_rsv_release(struct btrfs_root *root,
3949                              struct btrfs_block_rsv *block_rsv,
3950                              u64 num_bytes)
3951 {
3952         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3953         if (global_rsv->full || global_rsv == block_rsv ||
3954             block_rsv->space_info != global_rsv->space_info)
3955                 global_rsv = NULL;
3956         block_rsv_release_bytes(block_rsv, global_rsv, num_bytes);
3957 }
3958
3959 /*
3960  * helper to calculate size of global block reservation.
3961  * the desired value is sum of space used by extent tree,
3962  * checksum tree and root tree
3963  */
3964 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
3965 {
3966         struct btrfs_space_info *sinfo;
3967         u64 num_bytes;
3968         u64 meta_used;
3969         u64 data_used;
3970         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
3971
3972         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
3973         spin_lock(&sinfo->lock);
3974         data_used = sinfo->bytes_used;
3975         spin_unlock(&sinfo->lock);
3976
3977         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3978         spin_lock(&sinfo->lock);
3979         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
3980                 data_used = 0;
3981         meta_used = sinfo->bytes_used;
3982         spin_unlock(&sinfo->lock);
3983
3984         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
3985                     csum_size * 2;
3986         num_bytes += div64_u64(data_used + meta_used, 50);
3987
3988         if (num_bytes * 3 > meta_used)
3989                 num_bytes = div64_u64(meta_used, 3);
3990
3991         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
3992 }
3993
3994 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
3995 {
3996         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
3997         struct btrfs_space_info *sinfo = block_rsv->space_info;
3998         u64 num_bytes;
3999
4000         num_bytes = calc_global_metadata_size(fs_info);
4001
4002         spin_lock(&block_rsv->lock);
4003         spin_lock(&sinfo->lock);
4004
4005         block_rsv->size = num_bytes;
4006
4007         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4008                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4009                     sinfo->bytes_may_use;
4010
4011         if (sinfo->total_bytes > num_bytes) {
4012                 num_bytes = sinfo->total_bytes - num_bytes;
4013                 block_rsv->reserved += num_bytes;
4014                 sinfo->bytes_may_use += num_bytes;
4015         }
4016
4017         if (block_rsv->reserved >= block_rsv->size) {
4018                 num_bytes = block_rsv->reserved - block_rsv->size;
4019                 sinfo->bytes_may_use -= num_bytes;
4020                 sinfo->reservation_progress++;
4021                 block_rsv->reserved = block_rsv->size;
4022                 block_rsv->full = 1;
4023         }
4024
4025         spin_unlock(&sinfo->lock);
4026         spin_unlock(&block_rsv->lock);
4027 }
4028
4029 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4030 {
4031         struct btrfs_space_info *space_info;
4032
4033         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4034         fs_info->chunk_block_rsv.space_info = space_info;
4035
4036         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4037         fs_info->global_block_rsv.space_info = space_info;
4038         fs_info->delalloc_block_rsv.space_info = space_info;
4039         fs_info->trans_block_rsv.space_info = space_info;
4040         fs_info->empty_block_rsv.space_info = space_info;
4041         fs_info->delayed_block_rsv.space_info = space_info;
4042
4043         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4044         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4045         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4046         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4047         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4048
4049         update_global_block_rsv(fs_info);
4050 }
4051
4052 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4053 {
4054         block_rsv_release_bytes(&fs_info->global_block_rsv, NULL, (u64)-1);
4055         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4056         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4057         WARN_ON(fs_info->trans_block_rsv.size > 0);
4058         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4059         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4060         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4061         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4062         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4063 }
4064
4065 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4066                                   struct btrfs_root *root)
4067 {
4068         if (!trans->bytes_reserved)
4069                 return;
4070
4071         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4072         trans->bytes_reserved = 0;
4073 }
4074
4075 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4076                                   struct inode *inode)
4077 {
4078         struct btrfs_root *root = BTRFS_I(inode)->root;
4079         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4080         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4081
4082         /*
4083          * We need to hold space in order to delete our orphan item once we've
4084          * added it, so this takes the reservation so we can release it later
4085          * when we are truly done with the orphan item.
4086          */
4087         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4088         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4089 }
4090
4091 void btrfs_orphan_release_metadata(struct inode *inode)
4092 {
4093         struct btrfs_root *root = BTRFS_I(inode)->root;
4094         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4095         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4096 }
4097
4098 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4099                                 struct btrfs_pending_snapshot *pending)
4100 {
4101         struct btrfs_root *root = pending->root;
4102         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4103         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4104         /*
4105          * two for root back/forward refs, two for directory entries
4106          * and one for root of the snapshot.
4107          */
4108         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4109         dst_rsv->space_info = src_rsv->space_info;
4110         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4111 }
4112
4113 /**
4114  * drop_outstanding_extent - drop an outstanding extent
4115  * @inode: the inode we're dropping the extent for
4116  *
4117  * This is called when we are freeing up an outstanding extent, either called
4118  * after an error or after an extent is written.  This will return the number of
4119  * reserved extents that need to be freed.  This must be called with
4120  * BTRFS_I(inode)->lock held.
4121  */
4122 static unsigned drop_outstanding_extent(struct inode *inode)
4123 {
4124         unsigned drop_inode_space = 0;
4125         unsigned dropped_extents = 0;
4126
4127         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4128         BTRFS_I(inode)->outstanding_extents--;
4129
4130         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4131             BTRFS_I(inode)->delalloc_meta_reserved) {
4132                 drop_inode_space = 1;
4133                 BTRFS_I(inode)->delalloc_meta_reserved = 0;
4134         }
4135
4136         /*
4137          * If we have more or the same amount of outsanding extents than we have
4138          * reserved then we need to leave the reserved extents count alone.
4139          */
4140         if (BTRFS_I(inode)->outstanding_extents >=
4141             BTRFS_I(inode)->reserved_extents)
4142                 return drop_inode_space;
4143
4144         dropped_extents = BTRFS_I(inode)->reserved_extents -
4145                 BTRFS_I(inode)->outstanding_extents;
4146         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4147         return dropped_extents + drop_inode_space;
4148 }
4149
4150 /**
4151  * calc_csum_metadata_size - return the amount of metada space that must be
4152  *      reserved/free'd for the given bytes.
4153  * @inode: the inode we're manipulating
4154  * @num_bytes: the number of bytes in question
4155  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4156  *
4157  * This adjusts the number of csum_bytes in the inode and then returns the
4158  * correct amount of metadata that must either be reserved or freed.  We
4159  * calculate how many checksums we can fit into one leaf and then divide the
4160  * number of bytes that will need to be checksumed by this value to figure out
4161  * how many checksums will be required.  If we are adding bytes then the number
4162  * may go up and we will return the number of additional bytes that must be
4163  * reserved.  If it is going down we will return the number of bytes that must
4164  * be freed.
4165  *
4166  * This must be called with BTRFS_I(inode)->lock held.
4167  */
4168 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4169                                    int reserve)
4170 {
4171         struct btrfs_root *root = BTRFS_I(inode)->root;
4172         u64 csum_size;
4173         int num_csums_per_leaf;
4174         int num_csums;
4175         int old_csums;
4176
4177         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4178             BTRFS_I(inode)->csum_bytes == 0)
4179                 return 0;
4180
4181         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4182         if (reserve)
4183                 BTRFS_I(inode)->csum_bytes += num_bytes;
4184         else
4185                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4186         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4187         num_csums_per_leaf = (int)div64_u64(csum_size,
4188                                             sizeof(struct btrfs_csum_item) +
4189                                             sizeof(struct btrfs_disk_key));
4190         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4191         num_csums = num_csums + num_csums_per_leaf - 1;
4192         num_csums = num_csums / num_csums_per_leaf;
4193
4194         old_csums = old_csums + num_csums_per_leaf - 1;
4195         old_csums = old_csums / num_csums_per_leaf;
4196
4197         /* No change, no need to reserve more */
4198         if (old_csums == num_csums)
4199                 return 0;
4200
4201         if (reserve)
4202                 return btrfs_calc_trans_metadata_size(root,
4203                                                       num_csums - old_csums);
4204
4205         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4206 }
4207
4208 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4209 {
4210         struct btrfs_root *root = BTRFS_I(inode)->root;
4211         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4212         u64 to_reserve = 0;
4213         u64 csum_bytes;
4214         unsigned nr_extents = 0;
4215         int extra_reserve = 0;
4216         int flush = 1;
4217         int ret;
4218
4219         /* Need to be holding the i_mutex here if we aren't free space cache */
4220         if (btrfs_is_free_space_inode(root, inode))
4221                 flush = 0;
4222         else
4223                 WARN_ON(!mutex_is_locked(&inode->i_mutex));
4224
4225         if (flush && btrfs_transaction_in_commit(root->fs_info))
4226                 schedule_timeout(1);
4227
4228         num_bytes = ALIGN(num_bytes, root->sectorsize);
4229
4230         spin_lock(&BTRFS_I(inode)->lock);
4231         BTRFS_I(inode)->outstanding_extents++;
4232
4233         if (BTRFS_I(inode)->outstanding_extents >
4234             BTRFS_I(inode)->reserved_extents)
4235                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4236                         BTRFS_I(inode)->reserved_extents;
4237
4238         /*
4239          * Add an item to reserve for updating the inode when we complete the
4240          * delalloc io.
4241          */
4242         if (!BTRFS_I(inode)->delalloc_meta_reserved) {
4243                 nr_extents++;
4244                 extra_reserve = 1;
4245         }
4246
4247         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4248         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4249         csum_bytes = BTRFS_I(inode)->csum_bytes;
4250         spin_unlock(&BTRFS_I(inode)->lock);
4251
4252         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4253         if (ret) {
4254                 u64 to_free = 0;
4255                 unsigned dropped;
4256
4257                 spin_lock(&BTRFS_I(inode)->lock);
4258                 dropped = drop_outstanding_extent(inode);
4259                 /*
4260                  * If the inodes csum_bytes is the same as the original
4261                  * csum_bytes then we know we haven't raced with any free()ers
4262                  * so we can just reduce our inodes csum bytes and carry on.
4263                  * Otherwise we have to do the normal free thing to account for
4264                  * the case that the free side didn't free up its reserve
4265                  * because of this outstanding reservation.
4266                  */
4267                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4268                         calc_csum_metadata_size(inode, num_bytes, 0);
4269                 else
4270                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4271                 spin_unlock(&BTRFS_I(inode)->lock);
4272                 if (dropped)
4273                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4274
4275                 if (to_free)
4276                         btrfs_block_rsv_release(root, block_rsv, to_free);
4277                 return ret;
4278         }
4279
4280         spin_lock(&BTRFS_I(inode)->lock);
4281         if (extra_reserve) {
4282                 BTRFS_I(inode)->delalloc_meta_reserved = 1;
4283                 nr_extents--;
4284         }
4285         BTRFS_I(inode)->reserved_extents += nr_extents;
4286         spin_unlock(&BTRFS_I(inode)->lock);
4287
4288         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4289
4290         return 0;
4291 }
4292
4293 /**
4294  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4295  * @inode: the inode to release the reservation for
4296  * @num_bytes: the number of bytes we're releasing
4297  *
4298  * This will release the metadata reservation for an inode.  This can be called
4299  * once we complete IO for a given set of bytes to release their metadata
4300  * reservations.
4301  */
4302 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4303 {
4304         struct btrfs_root *root = BTRFS_I(inode)->root;
4305         u64 to_free = 0;
4306         unsigned dropped;
4307
4308         num_bytes = ALIGN(num_bytes, root->sectorsize);
4309         spin_lock(&BTRFS_I(inode)->lock);
4310         dropped = drop_outstanding_extent(inode);
4311
4312         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4313         spin_unlock(&BTRFS_I(inode)->lock);
4314         if (dropped > 0)
4315                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4316
4317         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4318                                 to_free);
4319 }
4320
4321 /**
4322  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4323  * @inode: inode we're writing to
4324  * @num_bytes: the number of bytes we want to allocate
4325  *
4326  * This will do the following things
4327  *
4328  * o reserve space in the data space info for num_bytes
4329  * o reserve space in the metadata space info based on number of outstanding
4330  *   extents and how much csums will be needed
4331  * o add to the inodes ->delalloc_bytes
4332  * o add it to the fs_info's delalloc inodes list.
4333  *
4334  * This will return 0 for success and -ENOSPC if there is no space left.
4335  */
4336 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4337 {
4338         int ret;
4339
4340         ret = btrfs_check_data_free_space(inode, num_bytes);
4341         if (ret)
4342                 return ret;
4343
4344         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4345         if (ret) {
4346                 btrfs_free_reserved_data_space(inode, num_bytes);
4347                 return ret;
4348         }
4349
4350         return 0;
4351 }
4352
4353 /**
4354  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4355  * @inode: inode we're releasing space for
4356  * @num_bytes: the number of bytes we want to free up
4357  *
4358  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4359  * called in the case that we don't need the metadata AND data reservations
4360  * anymore.  So if there is an error or we insert an inline extent.
4361  *
4362  * This function will release the metadata space that was not used and will
4363  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4364  * list if there are no delalloc bytes left.
4365  */
4366 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4367 {
4368         btrfs_delalloc_release_metadata(inode, num_bytes);
4369         btrfs_free_reserved_data_space(inode, num_bytes);
4370 }
4371
4372 static int update_block_group(struct btrfs_trans_handle *trans,
4373                               struct btrfs_root *root,
4374                               u64 bytenr, u64 num_bytes, int alloc)
4375 {
4376         struct btrfs_block_group_cache *cache = NULL;
4377         struct btrfs_fs_info *info = root->fs_info;
4378         u64 total = num_bytes;
4379         u64 old_val;
4380         u64 byte_in_group;
4381         int factor;
4382
4383         /* block accounting for super block */
4384         spin_lock(&info->delalloc_lock);
4385         old_val = btrfs_super_bytes_used(info->super_copy);
4386         if (alloc)
4387                 old_val += num_bytes;
4388         else
4389                 old_val -= num_bytes;
4390         btrfs_set_super_bytes_used(info->super_copy, old_val);
4391         spin_unlock(&info->delalloc_lock);
4392
4393         while (total) {
4394                 cache = btrfs_lookup_block_group(info, bytenr);
4395                 if (!cache)
4396                         return -1;
4397                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4398                                     BTRFS_BLOCK_GROUP_RAID1 |
4399                                     BTRFS_BLOCK_GROUP_RAID10))
4400                         factor = 2;
4401                 else
4402                         factor = 1;
4403                 /*
4404                  * If this block group has free space cache written out, we
4405                  * need to make sure to load it if we are removing space.  This
4406                  * is because we need the unpinning stage to actually add the
4407                  * space back to the block group, otherwise we will leak space.
4408                  */
4409                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4410                         cache_block_group(cache, trans, NULL, 1);
4411
4412                 byte_in_group = bytenr - cache->key.objectid;
4413                 WARN_ON(byte_in_group > cache->key.offset);
4414
4415                 spin_lock(&cache->space_info->lock);
4416                 spin_lock(&cache->lock);
4417
4418                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4419                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4420                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4421
4422                 cache->dirty = 1;
4423                 old_val = btrfs_block_group_used(&cache->item);
4424                 num_bytes = min(total, cache->key.offset - byte_in_group);
4425                 if (alloc) {
4426                         old_val += num_bytes;
4427                         btrfs_set_block_group_used(&cache->item, old_val);
4428                         cache->reserved -= num_bytes;
4429                         cache->space_info->bytes_reserved -= num_bytes;
4430                         cache->space_info->bytes_used += num_bytes;
4431                         cache->space_info->disk_used += num_bytes * factor;
4432                         spin_unlock(&cache->lock);
4433                         spin_unlock(&cache->space_info->lock);
4434                 } else {
4435                         old_val -= num_bytes;
4436                         btrfs_set_block_group_used(&cache->item, old_val);
4437                         cache->pinned += num_bytes;
4438                         cache->space_info->bytes_pinned += num_bytes;
4439                         cache->space_info->bytes_used -= num_bytes;
4440                         cache->space_info->disk_used -= num_bytes * factor;
4441                         spin_unlock(&cache->lock);
4442                         spin_unlock(&cache->space_info->lock);
4443
4444                         set_extent_dirty(info->pinned_extents,
4445                                          bytenr, bytenr + num_bytes - 1,
4446                                          GFP_NOFS | __GFP_NOFAIL);
4447                 }
4448                 btrfs_put_block_group(cache);
4449                 total -= num_bytes;
4450                 bytenr += num_bytes;
4451         }
4452         return 0;
4453 }
4454
4455 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4456 {
4457         struct btrfs_block_group_cache *cache;
4458         u64 bytenr;
4459
4460         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4461         if (!cache)
4462                 return 0;
4463
4464         bytenr = cache->key.objectid;
4465         btrfs_put_block_group(cache);
4466
4467         return bytenr;
4468 }
4469
4470 static int pin_down_extent(struct btrfs_root *root,
4471                            struct btrfs_block_group_cache *cache,
4472                            u64 bytenr, u64 num_bytes, int reserved)
4473 {
4474         spin_lock(&cache->space_info->lock);
4475         spin_lock(&cache->lock);
4476         cache->pinned += num_bytes;
4477         cache->space_info->bytes_pinned += num_bytes;
4478         if (reserved) {
4479                 cache->reserved -= num_bytes;
4480                 cache->space_info->bytes_reserved -= num_bytes;
4481         }
4482         spin_unlock(&cache->lock);
4483         spin_unlock(&cache->space_info->lock);
4484
4485         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4486                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4487         return 0;
4488 }
4489
4490 /*
4491  * this function must be called within transaction
4492  */
4493 int btrfs_pin_extent(struct btrfs_root *root,
4494                      u64 bytenr, u64 num_bytes, int reserved)
4495 {
4496         struct btrfs_block_group_cache *cache;
4497
4498         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4499         BUG_ON(!cache);
4500
4501         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4502
4503         btrfs_put_block_group(cache);
4504         return 0;
4505 }
4506
4507 /*
4508  * this function must be called within transaction
4509  */
4510 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4511                                     struct btrfs_root *root,
4512                                     u64 bytenr, u64 num_bytes)
4513 {
4514         struct btrfs_block_group_cache *cache;
4515
4516         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4517         BUG_ON(!cache);
4518
4519         /*
4520          * pull in the free space cache (if any) so that our pin
4521          * removes the free space from the cache.  We have load_only set
4522          * to one because the slow code to read in the free extents does check
4523          * the pinned extents.
4524          */
4525         cache_block_group(cache, trans, root, 1);
4526
4527         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4528
4529         /* remove us from the free space cache (if we're there at all) */
4530         btrfs_remove_free_space(cache, bytenr, num_bytes);
4531         btrfs_put_block_group(cache);
4532         return 0;
4533 }
4534
4535 /**
4536  * btrfs_update_reserved_bytes - update the block_group and space info counters
4537  * @cache:      The cache we are manipulating
4538  * @num_bytes:  The number of bytes in question
4539  * @reserve:    One of the reservation enums
4540  *
4541  * This is called by the allocator when it reserves space, or by somebody who is
4542  * freeing space that was never actually used on disk.  For example if you
4543  * reserve some space for a new leaf in transaction A and before transaction A
4544  * commits you free that leaf, you call this with reserve set to 0 in order to
4545  * clear the reservation.
4546  *
4547  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4548  * ENOSPC accounting.  For data we handle the reservation through clearing the
4549  * delalloc bits in the io_tree.  We have to do this since we could end up
4550  * allocating less disk space for the amount of data we have reserved in the
4551  * case of compression.
4552  *
4553  * If this is a reservation and the block group has become read only we cannot
4554  * make the reservation and return -EAGAIN, otherwise this function always
4555  * succeeds.
4556  */
4557 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4558                                        u64 num_bytes, int reserve)
4559 {
4560         struct btrfs_space_info *space_info = cache->space_info;
4561         int ret = 0;
4562         spin_lock(&space_info->lock);
4563         spin_lock(&cache->lock);
4564         if (reserve != RESERVE_FREE) {
4565                 if (cache->ro) {
4566                         ret = -EAGAIN;
4567                 } else {
4568                         cache->reserved += num_bytes;
4569                         space_info->bytes_reserved += num_bytes;
4570                         if (reserve == RESERVE_ALLOC) {
4571                                 BUG_ON(space_info->bytes_may_use < num_bytes);
4572                                 space_info->bytes_may_use -= num_bytes;
4573                         }
4574                 }
4575         } else {
4576                 if (cache->ro)
4577                         space_info->bytes_readonly += num_bytes;
4578                 cache->reserved -= num_bytes;
4579                 space_info->bytes_reserved -= num_bytes;
4580                 space_info->reservation_progress++;
4581         }
4582         spin_unlock(&cache->lock);
4583         spin_unlock(&space_info->lock);
4584         return ret;
4585 }
4586
4587 int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4588                                 struct btrfs_root *root)
4589 {
4590         struct btrfs_fs_info *fs_info = root->fs_info;
4591         struct btrfs_caching_control *next;
4592         struct btrfs_caching_control *caching_ctl;
4593         struct btrfs_block_group_cache *cache;
4594
4595         down_write(&fs_info->extent_commit_sem);
4596
4597         list_for_each_entry_safe(caching_ctl, next,
4598                                  &fs_info->caching_block_groups, list) {
4599                 cache = caching_ctl->block_group;
4600                 if (block_group_cache_done(cache)) {
4601                         cache->last_byte_to_unpin = (u64)-1;
4602                         list_del_init(&caching_ctl->list);
4603                         put_caching_control(caching_ctl);
4604                 } else {
4605                         cache->last_byte_to_unpin = caching_ctl->progress;
4606                 }
4607         }
4608
4609         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4610                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4611         else
4612                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4613
4614         up_write(&fs_info->extent_commit_sem);
4615
4616         update_global_block_rsv(fs_info);
4617         return 0;
4618 }
4619
4620 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4621 {
4622         struct btrfs_fs_info *fs_info = root->fs_info;
4623         struct btrfs_block_group_cache *cache = NULL;
4624         u64 len;
4625
4626         while (start <= end) {
4627                 if (!cache ||
4628                     start >= cache->key.objectid + cache->key.offset) {
4629                         if (cache)
4630                                 btrfs_put_block_group(cache);
4631                         cache = btrfs_lookup_block_group(fs_info, start);
4632                         BUG_ON(!cache);
4633                 }
4634
4635                 len = cache->key.objectid + cache->key.offset - start;
4636                 len = min(len, end + 1 - start);
4637
4638                 if (start < cache->last_byte_to_unpin) {
4639                         len = min(len, cache->last_byte_to_unpin - start);
4640                         btrfs_add_free_space(cache, start, len);
4641                 }
4642
4643                 start += len;
4644
4645                 spin_lock(&cache->space_info->lock);
4646                 spin_lock(&cache->lock);
4647                 cache->pinned -= len;
4648                 cache->space_info->bytes_pinned -= len;
4649                 if (cache->ro)
4650                         cache->space_info->bytes_readonly += len;
4651                 spin_unlock(&cache->lock);
4652                 spin_unlock(&cache->space_info->lock);
4653         }
4654
4655         if (cache)
4656                 btrfs_put_block_group(cache);
4657         return 0;
4658 }
4659
4660 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4661                                struct btrfs_root *root)
4662 {
4663         struct btrfs_fs_info *fs_info = root->fs_info;
4664         struct extent_io_tree *unpin;
4665         u64 start;
4666         u64 end;
4667         int ret;
4668
4669         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4670                 unpin = &fs_info->freed_extents[1];
4671         else
4672                 unpin = &fs_info->freed_extents[0];
4673
4674         while (1) {
4675                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4676                                             EXTENT_DIRTY);
4677                 if (ret)
4678                         break;
4679
4680                 if (btrfs_test_opt(root, DISCARD))
4681                         ret = btrfs_discard_extent(root, start,
4682                                                    end + 1 - start, NULL);
4683
4684                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4685                 unpin_extent_range(root, start, end);
4686                 cond_resched();
4687         }
4688
4689         return 0;
4690 }
4691
4692 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4693                                 struct btrfs_root *root,
4694                                 u64 bytenr, u64 num_bytes, u64 parent,
4695                                 u64 root_objectid, u64 owner_objectid,
4696                                 u64 owner_offset, int refs_to_drop,
4697                                 struct btrfs_delayed_extent_op *extent_op)
4698 {
4699         struct btrfs_key key;
4700         struct btrfs_path *path;
4701         struct btrfs_fs_info *info = root->fs_info;
4702         struct btrfs_root *extent_root = info->extent_root;
4703         struct extent_buffer *leaf;
4704         struct btrfs_extent_item *ei;
4705         struct btrfs_extent_inline_ref *iref;
4706         int ret;
4707         int is_data;
4708         int extent_slot = 0;
4709         int found_extent = 0;
4710         int num_to_del = 1;
4711         u32 item_size;
4712         u64 refs;
4713
4714         path = btrfs_alloc_path();
4715         if (!path)
4716                 return -ENOMEM;
4717
4718         path->reada = 1;
4719         path->leave_spinning = 1;
4720
4721         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4722         BUG_ON(!is_data && refs_to_drop != 1);
4723
4724         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4725                                     bytenr, num_bytes, parent,
4726                                     root_objectid, owner_objectid,
4727                                     owner_offset);
4728         if (ret == 0) {
4729                 extent_slot = path->slots[0];
4730                 while (extent_slot >= 0) {
4731                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4732                                               extent_slot);
4733                         if (key.objectid != bytenr)
4734                                 break;
4735                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4736                             key.offset == num_bytes) {
4737                                 found_extent = 1;
4738                                 break;
4739                         }
4740                         if (path->slots[0] - extent_slot > 5)
4741                                 break;
4742                         extent_slot--;
4743                 }
4744 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4745                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4746                 if (found_extent && item_size < sizeof(*ei))
4747                         found_extent = 0;
4748 #endif
4749                 if (!found_extent) {
4750                         BUG_ON(iref);
4751                         ret = remove_extent_backref(trans, extent_root, path,
4752                                                     NULL, refs_to_drop,
4753                                                     is_data);
4754                         BUG_ON(ret);
4755                         btrfs_release_path(path);
4756                         path->leave_spinning = 1;
4757
4758                         key.objectid = bytenr;
4759                         key.type = BTRFS_EXTENT_ITEM_KEY;
4760                         key.offset = num_bytes;
4761
4762                         ret = btrfs_search_slot(trans, extent_root,
4763                                                 &key, path, -1, 1);
4764                         if (ret) {
4765                                 printk(KERN_ERR "umm, got %d back from search"
4766                                        ", was looking for %llu\n", ret,
4767                                        (unsigned long long)bytenr);
4768                                 if (ret > 0)
4769                                         btrfs_print_leaf(extent_root,
4770                                                          path->nodes[0]);
4771                         }
4772                         BUG_ON(ret);
4773                         extent_slot = path->slots[0];
4774                 }
4775         } else {
4776                 btrfs_print_leaf(extent_root, path->nodes[0]);
4777                 WARN_ON(1);
4778                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4779                        "parent %llu root %llu  owner %llu offset %llu\n",
4780                        (unsigned long long)bytenr,
4781                        (unsigned long long)parent,
4782                        (unsigned long long)root_objectid,
4783                        (unsigned long long)owner_objectid,
4784                        (unsigned long long)owner_offset);
4785         }
4786
4787         leaf = path->nodes[0];
4788         item_size = btrfs_item_size_nr(leaf, extent_slot);
4789 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4790         if (item_size < sizeof(*ei)) {
4791                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4792                 ret = convert_extent_item_v0(trans, extent_root, path,
4793                                              owner_objectid, 0);
4794                 BUG_ON(ret < 0);
4795
4796                 btrfs_release_path(path);
4797                 path->leave_spinning = 1;
4798
4799                 key.objectid = bytenr;
4800                 key.type = BTRFS_EXTENT_ITEM_KEY;
4801                 key.offset = num_bytes;
4802
4803                 ret = btrfs_search_slot(trans, extent_root, &key, path,
4804                                         -1, 1);
4805                 if (ret) {
4806                         printk(KERN_ERR "umm, got %d back from search"
4807                                ", was looking for %llu\n", ret,
4808                                (unsigned long long)bytenr);
4809                         btrfs_print_leaf(extent_root, path->nodes[0]);
4810                 }
4811                 BUG_ON(ret);
4812                 extent_slot = path->slots[0];
4813                 leaf = path->nodes[0];
4814                 item_size = btrfs_item_size_nr(leaf, extent_slot);
4815         }
4816 #endif
4817         BUG_ON(item_size < sizeof(*ei));
4818         ei = btrfs_item_ptr(leaf, extent_slot,
4819                             struct btrfs_extent_item);
4820         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
4821                 struct btrfs_tree_block_info *bi;
4822                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4823                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4824                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4825         }
4826
4827         refs = btrfs_extent_refs(leaf, ei);
4828         BUG_ON(refs < refs_to_drop);
4829         refs -= refs_to_drop;
4830
4831         if (refs > 0) {
4832                 if (extent_op)
4833                         __run_delayed_extent_op(extent_op, leaf, ei);
4834                 /*
4835                  * In the case of inline back ref, reference count will
4836                  * be updated by remove_extent_backref
4837                  */
4838                 if (iref) {
4839                         BUG_ON(!found_extent);
4840                 } else {
4841                         btrfs_set_extent_refs(leaf, ei, refs);
4842                         btrfs_mark_buffer_dirty(leaf);
4843                 }
4844                 if (found_extent) {
4845                         ret = remove_extent_backref(trans, extent_root, path,
4846                                                     iref, refs_to_drop,
4847                                                     is_data);
4848                         BUG_ON(ret);
4849                 }
4850         } else {
4851                 if (found_extent) {
4852                         BUG_ON(is_data && refs_to_drop !=
4853                                extent_data_ref_count(root, path, iref));
4854                         if (iref) {
4855                                 BUG_ON(path->slots[0] != extent_slot);
4856                         } else {
4857                                 BUG_ON(path->slots[0] != extent_slot + 1);
4858                                 path->slots[0] = extent_slot;
4859                                 num_to_del = 2;
4860                         }
4861                 }
4862
4863                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
4864                                       num_to_del);
4865                 BUG_ON(ret);
4866                 btrfs_release_path(path);
4867
4868                 if (is_data) {
4869                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
4870                         BUG_ON(ret);
4871                 } else {
4872                         invalidate_mapping_pages(info->btree_inode->i_mapping,
4873                              bytenr >> PAGE_CACHE_SHIFT,
4874                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
4875                 }
4876
4877                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
4878                 BUG_ON(ret);
4879         }
4880         btrfs_free_path(path);
4881         return ret;
4882 }
4883
4884 /*
4885  * when we free an block, it is possible (and likely) that we free the last
4886  * delayed ref for that extent as well.  This searches the delayed ref tree for
4887  * a given extent, and if there are no other delayed refs to be processed, it
4888  * removes it from the tree.
4889  */
4890 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
4891                                       struct btrfs_root *root, u64 bytenr)
4892 {
4893         struct btrfs_delayed_ref_head *head;
4894         struct btrfs_delayed_ref_root *delayed_refs;
4895         struct btrfs_delayed_ref_node *ref;
4896         struct rb_node *node;
4897         int ret = 0;
4898
4899         delayed_refs = &trans->transaction->delayed_refs;
4900         spin_lock(&delayed_refs->lock);
4901         head = btrfs_find_delayed_ref_head(trans, bytenr);
4902         if (!head)
4903                 goto out;
4904
4905         node = rb_prev(&head->node.rb_node);
4906         if (!node)
4907                 goto out;
4908
4909         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
4910
4911         /* there are still entries for this ref, we can't drop it */
4912         if (ref->bytenr == bytenr)
4913                 goto out;
4914
4915         if (head->extent_op) {
4916                 if (!head->must_insert_reserved)
4917                         goto out;
4918                 kfree(head->extent_op);
4919                 head->extent_op = NULL;
4920         }
4921
4922         /*
4923          * waiting for the lock here would deadlock.  If someone else has it
4924          * locked they are already in the process of dropping it anyway
4925          */
4926         if (!mutex_trylock(&head->mutex))
4927                 goto out;
4928
4929         /*
4930          * at this point we have a head with no other entries.  Go
4931          * ahead and process it.
4932          */
4933         head->node.in_tree = 0;
4934         rb_erase(&head->node.rb_node, &delayed_refs->root);
4935
4936         delayed_refs->num_entries--;
4937
4938         /*
4939          * we don't take a ref on the node because we're removing it from the
4940          * tree, so we just steal the ref the tree was holding.
4941          */
4942         delayed_refs->num_heads--;
4943         if (list_empty(&head->cluster))
4944                 delayed_refs->num_heads_ready--;
4945
4946         list_del_init(&head->cluster);
4947         spin_unlock(&delayed_refs->lock);
4948
4949         BUG_ON(head->extent_op);
4950         if (head->must_insert_reserved)
4951                 ret = 1;
4952
4953         mutex_unlock(&head->mutex);
4954         btrfs_put_delayed_ref(&head->node);
4955         return ret;
4956 out:
4957         spin_unlock(&delayed_refs->lock);
4958         return 0;
4959 }
4960
4961 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
4962                            struct btrfs_root *root,
4963                            struct extent_buffer *buf,
4964                            u64 parent, int last_ref)
4965 {
4966         struct btrfs_block_group_cache *cache = NULL;
4967         int ret;
4968
4969         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4970                 ret = btrfs_add_delayed_tree_ref(trans, buf->start, buf->len,
4971                                                 parent, root->root_key.objectid,
4972                                                 btrfs_header_level(buf),
4973                                                 BTRFS_DROP_DELAYED_REF, NULL);
4974                 BUG_ON(ret);
4975         }
4976
4977         if (!last_ref)
4978                 return;
4979
4980         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
4981
4982         if (btrfs_header_generation(buf) == trans->transid) {
4983                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4984                         ret = check_ref_cleanup(trans, root, buf->start);
4985                         if (!ret)
4986                                 goto out;
4987                 }
4988
4989                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
4990                         pin_down_extent(root, cache, buf->start, buf->len, 1);
4991                         goto out;
4992                 }
4993
4994                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
4995
4996                 btrfs_add_free_space(cache, buf->start, buf->len);
4997                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
4998         }
4999 out:
5000         /*
5001          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5002          * anymore.
5003          */
5004         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5005         btrfs_put_block_group(cache);
5006 }
5007
5008 int btrfs_free_extent(struct btrfs_trans_handle *trans,
5009                       struct btrfs_root *root,
5010                       u64 bytenr, u64 num_bytes, u64 parent,
5011                       u64 root_objectid, u64 owner, u64 offset)
5012 {
5013         int ret;
5014
5015         /*
5016          * tree log blocks never actually go into the extent allocation
5017          * tree, just update pinning info and exit early.
5018          */
5019         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5020                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5021                 /* unlocks the pinned mutex */
5022                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5023                 ret = 0;
5024         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5025                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
5026                                         parent, root_objectid, (int)owner,
5027                                         BTRFS_DROP_DELAYED_REF, NULL);
5028                 BUG_ON(ret);
5029         } else {
5030                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
5031                                         parent, root_objectid, owner,
5032                                         offset, BTRFS_DROP_DELAYED_REF, NULL);
5033                 BUG_ON(ret);
5034         }
5035         return ret;
5036 }
5037
5038 static u64 stripe_align(struct btrfs_root *root, u64 val)
5039 {
5040         u64 mask = ((u64)root->stripesize - 1);
5041         u64 ret = (val + mask) & ~mask;
5042         return ret;
5043 }
5044
5045 /*
5046  * when we wait for progress in the block group caching, its because
5047  * our allocation attempt failed at least once.  So, we must sleep
5048  * and let some progress happen before we try again.
5049  *
5050  * This function will sleep at least once waiting for new free space to
5051  * show up, and then it will check the block group free space numbers
5052  * for our min num_bytes.  Another option is to have it go ahead
5053  * and look in the rbtree for a free extent of a given size, but this
5054  * is a good start.
5055  */
5056 static noinline int
5057 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5058                                 u64 num_bytes)
5059 {
5060         struct btrfs_caching_control *caching_ctl;
5061         DEFINE_WAIT(wait);
5062
5063         caching_ctl = get_caching_control(cache);
5064         if (!caching_ctl)
5065                 return 0;
5066
5067         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5068                    (cache->free_space_ctl->free_space >= num_bytes));
5069
5070         put_caching_control(caching_ctl);
5071         return 0;
5072 }
5073
5074 static noinline int
5075 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5076 {
5077         struct btrfs_caching_control *caching_ctl;
5078         DEFINE_WAIT(wait);
5079
5080         caching_ctl = get_caching_control(cache);
5081         if (!caching_ctl)
5082                 return 0;
5083
5084         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5085
5086         put_caching_control(caching_ctl);
5087         return 0;
5088 }
5089
5090 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5091 {
5092         int index;
5093         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
5094                 index = 0;
5095         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
5096                 index = 1;
5097         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
5098                 index = 2;
5099         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
5100                 index = 3;
5101         else
5102                 index = 4;
5103         return index;
5104 }
5105
5106 enum btrfs_loop_type {
5107         LOOP_FIND_IDEAL = 0,
5108         LOOP_CACHING_NOWAIT = 1,
5109         LOOP_CACHING_WAIT = 2,
5110         LOOP_ALLOC_CHUNK = 3,
5111         LOOP_NO_EMPTY_SIZE = 4,
5112 };
5113
5114 /*
5115  * walks the btree of allocated extents and find a hole of a given size.
5116  * The key ins is changed to record the hole:
5117  * ins->objectid == block start
5118  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5119  * ins->offset == number of blocks
5120  * Any available blocks before search_start are skipped.
5121  */
5122 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5123                                      struct btrfs_root *orig_root,
5124                                      u64 num_bytes, u64 empty_size,
5125                                      u64 search_start, u64 search_end,
5126                                      u64 hint_byte, struct btrfs_key *ins,
5127                                      u64 data)
5128 {
5129         int ret = 0;
5130         struct btrfs_root *root = orig_root->fs_info->extent_root;
5131         struct btrfs_free_cluster *last_ptr = NULL;
5132         struct btrfs_block_group_cache *block_group = NULL;
5133         struct btrfs_block_group_cache *used_block_group;
5134         int empty_cluster = 2 * 1024 * 1024;
5135         int allowed_chunk_alloc = 0;
5136         int done_chunk_alloc = 0;
5137         struct btrfs_space_info *space_info;
5138         int loop = 0;
5139         int index = 0;
5140         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5141                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5142         bool found_uncached_bg = false;
5143         bool failed_cluster_refill = false;
5144         bool failed_alloc = false;
5145         bool use_cluster = true;
5146         bool have_caching_bg = false;
5147         u64 ideal_cache_percent = 0;
5148         u64 ideal_cache_offset = 0;
5149
5150         WARN_ON(num_bytes < root->sectorsize);
5151         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5152         ins->objectid = 0;
5153         ins->offset = 0;
5154
5155         space_info = __find_space_info(root->fs_info, data);
5156         if (!space_info) {
5157                 printk(KERN_ERR "No space info for %llu\n", data);
5158                 return -ENOSPC;
5159         }
5160
5161         /*
5162          * If the space info is for both data and metadata it means we have a
5163          * small filesystem and we can't use the clustering stuff.
5164          */
5165         if (btrfs_mixed_space_info(space_info))
5166                 use_cluster = false;
5167
5168         if (orig_root->ref_cows || empty_size)
5169                 allowed_chunk_alloc = 1;
5170
5171         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5172                 last_ptr = &root->fs_info->meta_alloc_cluster;
5173                 if (!btrfs_test_opt(root, SSD))
5174                         empty_cluster = 64 * 1024;
5175         }
5176
5177         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5178             btrfs_test_opt(root, SSD)) {
5179                 last_ptr = &root->fs_info->data_alloc_cluster;
5180         }
5181
5182         if (last_ptr) {
5183                 spin_lock(&last_ptr->lock);
5184                 if (last_ptr->block_group)
5185                         hint_byte = last_ptr->window_start;
5186                 spin_unlock(&last_ptr->lock);
5187         }
5188
5189         search_start = max(search_start, first_logical_byte(root, 0));
5190         search_start = max(search_start, hint_byte);
5191
5192         if (!last_ptr)
5193                 empty_cluster = 0;
5194
5195         if (search_start == hint_byte) {
5196 ideal_cache:
5197                 block_group = btrfs_lookup_block_group(root->fs_info,
5198                                                        search_start);
5199                 used_block_group = block_group;
5200                 /*
5201                  * we don't want to use the block group if it doesn't match our
5202                  * allocation bits, or if its not cached.
5203                  *
5204                  * However if we are re-searching with an ideal block group
5205                  * picked out then we don't care that the block group is cached.
5206                  */
5207                 if (block_group && block_group_bits(block_group, data) &&
5208                     (block_group->cached != BTRFS_CACHE_NO ||
5209                      search_start == ideal_cache_offset)) {
5210                         down_read(&space_info->groups_sem);
5211                         if (list_empty(&block_group->list) ||
5212                             block_group->ro) {
5213                                 /*
5214                                  * someone is removing this block group,
5215                                  * we can't jump into the have_block_group
5216                                  * target because our list pointers are not
5217                                  * valid
5218                                  */
5219                                 btrfs_put_block_group(block_group);
5220                                 up_read(&space_info->groups_sem);
5221                         } else {
5222                                 index = get_block_group_index(block_group);
5223                                 goto have_block_group;
5224                         }
5225                 } else if (block_group) {
5226                         btrfs_put_block_group(block_group);
5227                 }
5228         }
5229 search:
5230         have_caching_bg = false;
5231         down_read(&space_info->groups_sem);
5232         list_for_each_entry(block_group, &space_info->block_groups[index],
5233                             list) {
5234                 u64 offset;
5235                 int cached;
5236
5237                 used_block_group = block_group;
5238                 btrfs_get_block_group(block_group);
5239                 search_start = block_group->key.objectid;
5240
5241                 /*
5242                  * this can happen if we end up cycling through all the
5243                  * raid types, but we want to make sure we only allocate
5244                  * for the proper type.
5245                  */
5246                 if (!block_group_bits(block_group, data)) {
5247                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5248                                 BTRFS_BLOCK_GROUP_RAID1 |
5249                                 BTRFS_BLOCK_GROUP_RAID10;
5250
5251                         /*
5252                          * if they asked for extra copies and this block group
5253                          * doesn't provide them, bail.  This does allow us to
5254                          * fill raid0 from raid1.
5255                          */
5256                         if ((data & extra) && !(block_group->flags & extra))
5257                                 goto loop;
5258                 }
5259
5260 have_block_group:
5261                 cached = block_group_cache_done(block_group);
5262                 if (unlikely(!cached)) {
5263                         u64 free_percent;
5264
5265                         found_uncached_bg = true;
5266                         ret = cache_block_group(block_group, trans,
5267                                                 orig_root, 1);
5268                         if (block_group->cached == BTRFS_CACHE_FINISHED)
5269                                 goto alloc;
5270
5271                         free_percent = btrfs_block_group_used(&block_group->item);
5272                         free_percent *= 100;
5273                         free_percent = div64_u64(free_percent,
5274                                                  block_group->key.offset);
5275                         free_percent = 100 - free_percent;
5276                         if (free_percent > ideal_cache_percent &&
5277                             likely(!block_group->ro)) {
5278                                 ideal_cache_offset = block_group->key.objectid;
5279                                 ideal_cache_percent = free_percent;
5280                         }
5281
5282                         /*
5283                          * The caching workers are limited to 2 threads, so we
5284                          * can queue as much work as we care to.
5285                          */
5286                         if (loop > LOOP_FIND_IDEAL) {
5287                                 ret = cache_block_group(block_group, trans,
5288                                                         orig_root, 0);
5289                                 BUG_ON(ret);
5290                         }
5291
5292                         /*
5293                          * If loop is set for cached only, try the next block
5294                          * group.
5295                          */
5296                         if (loop == LOOP_FIND_IDEAL)
5297                                 goto loop;
5298                 }
5299
5300 alloc:
5301                 if (unlikely(block_group->ro))
5302                         goto loop;
5303
5304                 spin_lock(&block_group->free_space_ctl->tree_lock);
5305                 if (cached &&
5306                     block_group->free_space_ctl->free_space <
5307                     num_bytes + empty_cluster + empty_size) {
5308                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5309                         goto loop;
5310                 }
5311                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5312
5313                 /*
5314                  * Ok we want to try and use the cluster allocator, so
5315                  * lets look there
5316                  */
5317                 if (last_ptr) {
5318                         /*
5319                          * the refill lock keeps out other
5320                          * people trying to start a new cluster
5321                          */
5322                         spin_lock(&last_ptr->refill_lock);
5323                         used_block_group = last_ptr->block_group;
5324                         if (used_block_group != block_group &&
5325                             (!used_block_group ||
5326                              used_block_group->ro ||
5327                              !block_group_bits(used_block_group, data))) {
5328                                 used_block_group = block_group;
5329                                 goto refill_cluster;
5330                         }
5331
5332                         if (used_block_group != block_group)
5333                                 btrfs_get_block_group(used_block_group);
5334
5335                         offset = btrfs_alloc_from_cluster(used_block_group,
5336                           last_ptr, num_bytes, used_block_group->key.objectid);
5337                         if (offset) {
5338                                 /* we have a block, we're done */
5339                                 spin_unlock(&last_ptr->refill_lock);
5340                                 goto checks;
5341                         }
5342
5343                         WARN_ON(last_ptr->block_group != used_block_group);
5344                         if (used_block_group != block_group) {
5345                                 btrfs_put_block_group(used_block_group);
5346                                 used_block_group = block_group;
5347                         }
5348 refill_cluster:
5349                         BUG_ON(used_block_group != block_group);
5350                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5351                          * set up a new clusters, so lets just skip it
5352                          * and let the allocator find whatever block
5353                          * it can find.  If we reach this point, we
5354                          * will have tried the cluster allocator
5355                          * plenty of times and not have found
5356                          * anything, so we are likely way too
5357                          * fragmented for the clustering stuff to find
5358                          * anything.  */
5359                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5360                                 spin_unlock(&last_ptr->refill_lock);
5361                                 goto unclustered_alloc;
5362                         }
5363
5364                         /*
5365                          * this cluster didn't work out, free it and
5366                          * start over
5367                          */
5368                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5369
5370                         /* allocate a cluster in this block group */
5371                         ret = btrfs_find_space_cluster(trans, root,
5372                                                block_group, last_ptr,
5373                                                search_start, num_bytes,
5374                                                empty_cluster + empty_size);
5375                         if (ret == 0) {
5376                                 /*
5377                                  * now pull our allocation out of this
5378                                  * cluster
5379                                  */
5380                                 offset = btrfs_alloc_from_cluster(block_group,
5381                                                   last_ptr, num_bytes,
5382                                                   search_start);
5383                                 if (offset) {
5384                                         /* we found one, proceed */
5385                                         spin_unlock(&last_ptr->refill_lock);
5386                                         goto checks;
5387                                 }
5388                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5389                                    && !failed_cluster_refill) {
5390                                 spin_unlock(&last_ptr->refill_lock);
5391
5392                                 failed_cluster_refill = true;
5393                                 wait_block_group_cache_progress(block_group,
5394                                        num_bytes + empty_cluster + empty_size);
5395                                 goto have_block_group;
5396                         }
5397
5398                         /*
5399                          * at this point we either didn't find a cluster
5400                          * or we weren't able to allocate a block from our
5401                          * cluster.  Free the cluster we've been trying
5402                          * to use, and go to the next block group
5403                          */
5404                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5405                         spin_unlock(&last_ptr->refill_lock);
5406                         goto loop;
5407                 }
5408
5409 unclustered_alloc:
5410                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5411                                                     num_bytes, empty_size);
5412                 /*
5413                  * If we didn't find a chunk, and we haven't failed on this
5414                  * block group before, and this block group is in the middle of
5415                  * caching and we are ok with waiting, then go ahead and wait
5416                  * for progress to be made, and set failed_alloc to true.
5417                  *
5418                  * If failed_alloc is true then we've already waited on this
5419                  * block group once and should move on to the next block group.
5420                  */
5421                 if (!offset && !failed_alloc && !cached &&
5422                     loop > LOOP_CACHING_NOWAIT) {
5423                         wait_block_group_cache_progress(block_group,
5424                                                 num_bytes + empty_size);
5425                         failed_alloc = true;
5426                         goto have_block_group;
5427                 } else if (!offset) {
5428                         if (!cached)
5429                                 have_caching_bg = true;
5430                         goto loop;
5431                 }
5432 checks:
5433                 search_start = stripe_align(root, offset);
5434                 /* move on to the next group */
5435                 if (search_start + num_bytes >= search_end) {
5436                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5437                         goto loop;
5438                 }
5439
5440                 /* move on to the next group */
5441                 if (search_start + num_bytes >
5442                     used_block_group->key.objectid + used_block_group->key.offset) {
5443                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5444                         goto loop;
5445                 }
5446
5447                 ins->objectid = search_start;
5448                 ins->offset = num_bytes;
5449
5450                 if (offset < search_start)
5451                         btrfs_add_free_space(used_block_group, offset,
5452                                              search_start - offset);
5453                 BUG_ON(offset > search_start);
5454
5455                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
5456                                                   alloc_type);
5457                 if (ret == -EAGAIN) {
5458                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5459                         goto loop;
5460                 }
5461
5462                 /* we are all good, lets return */
5463                 ins->objectid = search_start;
5464                 ins->offset = num_bytes;
5465
5466                 if (offset < search_start)
5467                         btrfs_add_free_space(used_block_group, offset,
5468                                              search_start - offset);
5469                 BUG_ON(offset > search_start);
5470                 if (used_block_group != block_group)
5471                         btrfs_put_block_group(used_block_group);
5472                 btrfs_put_block_group(block_group);
5473                 break;
5474 loop:
5475                 failed_cluster_refill = false;
5476                 failed_alloc = false;
5477                 BUG_ON(index != get_block_group_index(block_group));
5478                 if (used_block_group != block_group)
5479                         btrfs_put_block_group(used_block_group);
5480                 btrfs_put_block_group(block_group);
5481         }
5482         up_read(&space_info->groups_sem);
5483
5484         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
5485                 goto search;
5486
5487         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5488                 goto search;
5489
5490         /* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
5491          *                      for them to make caching progress.  Also
5492          *                      determine the best possible bg to cache
5493          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5494          *                      caching kthreads as we move along
5495          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5496          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5497          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5498          *                      again
5499          */
5500         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5501                 index = 0;
5502                 if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
5503                         found_uncached_bg = false;
5504                         loop++;
5505                         if (!ideal_cache_percent)
5506                                 goto search;
5507
5508                         /*
5509                          * 1 of the following 2 things have happened so far
5510                          *
5511                          * 1) We found an ideal block group for caching that
5512                          * is mostly full and will cache quickly, so we might
5513                          * as well wait for it.
5514                          *
5515                          * 2) We searched for cached only and we didn't find
5516                          * anything, and we didn't start any caching kthreads
5517                          * either, so chances are we will loop through and
5518                          * start a couple caching kthreads, and then come back
5519                          * around and just wait for them.  This will be slower
5520                          * because we will have 2 caching kthreads reading at
5521                          * the same time when we could have just started one
5522                          * and waited for it to get far enough to give us an
5523                          * allocation, so go ahead and go to the wait caching
5524                          * loop.
5525                          */
5526                         loop = LOOP_CACHING_WAIT;
5527                         search_start = ideal_cache_offset;
5528                         ideal_cache_percent = 0;
5529                         goto ideal_cache;
5530                 } else if (loop == LOOP_FIND_IDEAL) {
5531                         /*
5532                          * Didn't find a uncached bg, wait on anything we find
5533                          * next.
5534                          */
5535                         loop = LOOP_CACHING_WAIT;
5536                         goto search;
5537                 }
5538
5539                 loop++;
5540
5541                 if (loop == LOOP_ALLOC_CHUNK) {
5542                        if (allowed_chunk_alloc) {
5543                                 ret = do_chunk_alloc(trans, root, num_bytes +
5544                                                      2 * 1024 * 1024, data,
5545                                                      CHUNK_ALLOC_LIMITED);
5546                                 allowed_chunk_alloc = 0;
5547                                 if (ret == 1)
5548                                         done_chunk_alloc = 1;
5549                         } else if (!done_chunk_alloc &&
5550                                    space_info->force_alloc ==
5551                                    CHUNK_ALLOC_NO_FORCE) {
5552                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5553                         }
5554
5555                        /*
5556                         * We didn't allocate a chunk, go ahead and drop the
5557                         * empty size and loop again.
5558                         */
5559                        if (!done_chunk_alloc)
5560                                loop = LOOP_NO_EMPTY_SIZE;
5561                 }
5562
5563                 if (loop == LOOP_NO_EMPTY_SIZE) {
5564                         empty_size = 0;
5565                         empty_cluster = 0;
5566                 }
5567
5568                 goto search;
5569         } else if (!ins->objectid) {
5570                 ret = -ENOSPC;
5571         } else if (ins->objectid) {
5572                 ret = 0;
5573         }
5574
5575         return ret;
5576 }
5577
5578 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5579                             int dump_block_groups)
5580 {
5581         struct btrfs_block_group_cache *cache;
5582         int index = 0;
5583
5584         spin_lock(&info->lock);
5585         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5586                (unsigned long long)info->flags,
5587                (unsigned long long)(info->total_bytes - info->bytes_used -
5588                                     info->bytes_pinned - info->bytes_reserved -
5589                                     info->bytes_readonly),
5590                (info->full) ? "" : "not ");
5591         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5592                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5593                (unsigned long long)info->total_bytes,
5594                (unsigned long long)info->bytes_used,
5595                (unsigned long long)info->bytes_pinned,
5596                (unsigned long long)info->bytes_reserved,
5597                (unsigned long long)info->bytes_may_use,
5598                (unsigned long long)info->bytes_readonly);
5599         spin_unlock(&info->lock);
5600
5601         if (!dump_block_groups)
5602                 return;
5603
5604         down_read(&info->groups_sem);
5605 again:
5606         list_for_each_entry(cache, &info->block_groups[index], list) {
5607                 spin_lock(&cache->lock);
5608                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5609                        "%llu pinned %llu reserved\n",
5610                        (unsigned long long)cache->key.objectid,
5611                        (unsigned long long)cache->key.offset,
5612                        (unsigned long long)btrfs_block_group_used(&cache->item),
5613                        (unsigned long long)cache->pinned,
5614                        (unsigned long long)cache->reserved);
5615                 btrfs_dump_free_space(cache, bytes);
5616                 spin_unlock(&cache->lock);
5617         }
5618         if (++index < BTRFS_NR_RAID_TYPES)
5619                 goto again;
5620         up_read(&info->groups_sem);
5621 }
5622
5623 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5624                          struct btrfs_root *root,
5625                          u64 num_bytes, u64 min_alloc_size,
5626                          u64 empty_size, u64 hint_byte,
5627                          u64 search_end, struct btrfs_key *ins,
5628                          u64 data)
5629 {
5630         int ret;
5631         u64 search_start = 0;
5632
5633         data = btrfs_get_alloc_profile(root, data);
5634 again:
5635         /*
5636          * the only place that sets empty_size is btrfs_realloc_node, which
5637          * is not called recursively on allocations
5638          */
5639         if (empty_size || root->ref_cows)
5640                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5641                                      num_bytes + 2 * 1024 * 1024, data,
5642                                      CHUNK_ALLOC_NO_FORCE);
5643
5644         WARN_ON(num_bytes < root->sectorsize);
5645         ret = find_free_extent(trans, root, num_bytes, empty_size,
5646                                search_start, search_end, hint_byte,
5647                                ins, data);
5648
5649         if (ret == -ENOSPC && num_bytes > min_alloc_size) {
5650                 num_bytes = num_bytes >> 1;
5651                 num_bytes = num_bytes & ~(root->sectorsize - 1);
5652                 num_bytes = max(num_bytes, min_alloc_size);
5653                 do_chunk_alloc(trans, root->fs_info->extent_root,
5654                                num_bytes, data, CHUNK_ALLOC_FORCE);
5655                 goto again;
5656         }
5657         if (ret == -ENOSPC && btrfs_test_opt(root, ENOSPC_DEBUG)) {
5658                 struct btrfs_space_info *sinfo;
5659
5660                 sinfo = __find_space_info(root->fs_info, data);
5661                 printk(KERN_ERR "btrfs allocation failed flags %llu, "
5662                        "wanted %llu\n", (unsigned long long)data,
5663                        (unsigned long long)num_bytes);
5664                 dump_space_info(sinfo, num_bytes, 1);
5665         }
5666
5667         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
5668
5669         return ret;
5670 }
5671
5672 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
5673                                         u64 start, u64 len, int pin)
5674 {
5675         struct btrfs_block_group_cache *cache;
5676         int ret = 0;
5677
5678         cache = btrfs_lookup_block_group(root->fs_info, start);
5679         if (!cache) {
5680                 printk(KERN_ERR "Unable to find block group for %llu\n",
5681                        (unsigned long long)start);
5682                 return -ENOSPC;
5683         }
5684
5685         if (btrfs_test_opt(root, DISCARD))
5686                 ret = btrfs_discard_extent(root, start, len, NULL);
5687
5688         if (pin)
5689                 pin_down_extent(root, cache, start, len, 1);
5690         else {
5691                 btrfs_add_free_space(cache, start, len);
5692                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
5693         }
5694         btrfs_put_block_group(cache);
5695
5696         trace_btrfs_reserved_extent_free(root, start, len);
5697
5698         return ret;
5699 }
5700
5701 int btrfs_free_reserved_extent(struct btrfs_root *root,
5702                                         u64 start, u64 len)
5703 {
5704         return __btrfs_free_reserved_extent(root, start, len, 0);
5705 }
5706
5707 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
5708                                        u64 start, u64 len)
5709 {
5710         return __btrfs_free_reserved_extent(root, start, len, 1);
5711 }
5712
5713 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5714                                       struct btrfs_root *root,
5715                                       u64 parent, u64 root_objectid,
5716                                       u64 flags, u64 owner, u64 offset,
5717                                       struct btrfs_key *ins, int ref_mod)
5718 {
5719         int ret;
5720         struct btrfs_fs_info *fs_info = root->fs_info;
5721         struct btrfs_extent_item *extent_item;
5722         struct btrfs_extent_inline_ref *iref;
5723         struct btrfs_path *path;
5724         struct extent_buffer *leaf;
5725         int type;
5726         u32 size;
5727
5728         if (parent > 0)
5729                 type = BTRFS_SHARED_DATA_REF_KEY;
5730         else
5731                 type = BTRFS_EXTENT_DATA_REF_KEY;
5732
5733         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5734
5735         path = btrfs_alloc_path();
5736         if (!path)
5737                 return -ENOMEM;
5738
5739         path->leave_spinning = 1;
5740         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5741                                       ins, size);
5742         BUG_ON(ret);
5743
5744         leaf = path->nodes[0];
5745         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5746                                      struct btrfs_extent_item);
5747         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
5748         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5749         btrfs_set_extent_flags(leaf, extent_item,
5750                                flags | BTRFS_EXTENT_FLAG_DATA);
5751
5752         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
5753         btrfs_set_extent_inline_ref_type(leaf, iref, type);
5754         if (parent > 0) {
5755                 struct btrfs_shared_data_ref *ref;
5756                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
5757                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5758                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
5759         } else {
5760                 struct btrfs_extent_data_ref *ref;
5761                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
5762                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
5763                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
5764                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
5765                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
5766         }
5767
5768         btrfs_mark_buffer_dirty(path->nodes[0]);
5769         btrfs_free_path(path);
5770
5771         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5772         if (ret) {
5773                 printk(KERN_ERR "btrfs update block group failed for %llu "
5774                        "%llu\n", (unsigned long long)ins->objectid,
5775                        (unsigned long long)ins->offset);
5776                 BUG();
5777         }
5778         return ret;
5779 }
5780
5781 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
5782                                      struct btrfs_root *root,
5783                                      u64 parent, u64 root_objectid,
5784                                      u64 flags, struct btrfs_disk_key *key,
5785                                      int level, struct btrfs_key *ins)
5786 {
5787         int ret;
5788         struct btrfs_fs_info *fs_info = root->fs_info;
5789         struct btrfs_extent_item *extent_item;
5790         struct btrfs_tree_block_info *block_info;
5791         struct btrfs_extent_inline_ref *iref;
5792         struct btrfs_path *path;
5793         struct extent_buffer *leaf;
5794         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
5795
5796         path = btrfs_alloc_path();
5797         if (!path)
5798                 return -ENOMEM;
5799
5800         path->leave_spinning = 1;
5801         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5802                                       ins, size);
5803         BUG_ON(ret);
5804
5805         leaf = path->nodes[0];
5806         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5807                                      struct btrfs_extent_item);
5808         btrfs_set_extent_refs(leaf, extent_item, 1);
5809         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5810         btrfs_set_extent_flags(leaf, extent_item,
5811                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
5812         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
5813
5814         btrfs_set_tree_block_key(leaf, block_info, key);
5815         btrfs_set_tree_block_level(leaf, block_info, level);
5816
5817         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
5818         if (parent > 0) {
5819                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
5820                 btrfs_set_extent_inline_ref_type(leaf, iref,
5821                                                  BTRFS_SHARED_BLOCK_REF_KEY);
5822                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5823         } else {
5824                 btrfs_set_extent_inline_ref_type(leaf, iref,
5825                                                  BTRFS_TREE_BLOCK_REF_KEY);
5826                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
5827         }
5828
5829         btrfs_mark_buffer_dirty(leaf);
5830         btrfs_free_path(path);
5831
5832         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5833         if (ret) {
5834                 printk(KERN_ERR "btrfs update block group failed for %llu "
5835                        "%llu\n", (unsigned long long)ins->objectid,
5836                        (unsigned long long)ins->offset);
5837                 BUG();
5838         }
5839         return ret;
5840 }
5841
5842 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5843                                      struct btrfs_root *root,
5844                                      u64 root_objectid, u64 owner,
5845                                      u64 offset, struct btrfs_key *ins)
5846 {
5847         int ret;
5848
5849         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
5850
5851         ret = btrfs_add_delayed_data_ref(trans, ins->objectid, ins->offset,
5852                                          0, root_objectid, owner, offset,
5853                                          BTRFS_ADD_DELAYED_EXTENT, NULL);
5854         return ret;
5855 }
5856
5857 /*
5858  * this is used by the tree logging recovery code.  It records that
5859  * an extent has been allocated and makes sure to clear the free
5860  * space cache bits as well
5861  */
5862 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
5863                                    struct btrfs_root *root,
5864                                    u64 root_objectid, u64 owner, u64 offset,
5865                                    struct btrfs_key *ins)
5866 {
5867         int ret;
5868         struct btrfs_block_group_cache *block_group;
5869         struct btrfs_caching_control *caching_ctl;
5870         u64 start = ins->objectid;
5871         u64 num_bytes = ins->offset;
5872
5873         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
5874         cache_block_group(block_group, trans, NULL, 0);
5875         caching_ctl = get_caching_control(block_group);
5876
5877         if (!caching_ctl) {
5878                 BUG_ON(!block_group_cache_done(block_group));
5879                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5880                 BUG_ON(ret);
5881         } else {
5882                 mutex_lock(&caching_ctl->mutex);
5883
5884                 if (start >= caching_ctl->progress) {
5885                         ret = add_excluded_extent(root, start, num_bytes);
5886                         BUG_ON(ret);
5887                 } else if (start + num_bytes <= caching_ctl->progress) {
5888                         ret = btrfs_remove_free_space(block_group,
5889                                                       start, num_bytes);
5890                         BUG_ON(ret);
5891                 } else {
5892                         num_bytes = caching_ctl->progress - start;
5893                         ret = btrfs_remove_free_space(block_group,
5894                                                       start, num_bytes);
5895                         BUG_ON(ret);
5896
5897                         start = caching_ctl->progress;
5898                         num_bytes = ins->objectid + ins->offset -
5899                                     caching_ctl->progress;
5900                         ret = add_excluded_extent(root, start, num_bytes);
5901                         BUG_ON(ret);
5902                 }
5903
5904                 mutex_unlock(&caching_ctl->mutex);
5905                 put_caching_control(caching_ctl);
5906         }
5907
5908         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
5909                                           RESERVE_ALLOC_NO_ACCOUNT);
5910         BUG_ON(ret);
5911         btrfs_put_block_group(block_group);
5912         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
5913                                          0, owner, offset, ins, 1);
5914         return ret;
5915 }
5916
5917 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
5918                                             struct btrfs_root *root,
5919                                             u64 bytenr, u32 blocksize,
5920                                             int level)
5921 {
5922         struct extent_buffer *buf;
5923
5924         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
5925         if (!buf)
5926                 return ERR_PTR(-ENOMEM);
5927         btrfs_set_header_generation(buf, trans->transid);
5928         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
5929         btrfs_tree_lock(buf);
5930         clean_tree_block(trans, root, buf);
5931
5932         btrfs_set_lock_blocking(buf);
5933         btrfs_set_buffer_uptodate(buf);
5934
5935         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
5936                 /*
5937                  * we allow two log transactions at a time, use different
5938                  * EXENT bit to differentiate dirty pages.
5939                  */
5940                 if (root->log_transid % 2 == 0)
5941                         set_extent_dirty(&root->dirty_log_pages, buf->start,
5942                                         buf->start + buf->len - 1, GFP_NOFS);
5943                 else
5944                         set_extent_new(&root->dirty_log_pages, buf->start,
5945                                         buf->start + buf->len - 1, GFP_NOFS);
5946         } else {
5947                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
5948                          buf->start + buf->len - 1, GFP_NOFS);
5949         }
5950         trans->blocks_used++;
5951         /* this returns a buffer locked for blocking */
5952         return buf;
5953 }
5954
5955 static struct btrfs_block_rsv *
5956 use_block_rsv(struct btrfs_trans_handle *trans,
5957               struct btrfs_root *root, u32 blocksize)
5958 {
5959         struct btrfs_block_rsv *block_rsv;
5960         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5961         int ret;
5962
5963         block_rsv = get_block_rsv(trans, root);
5964
5965         if (block_rsv->size == 0) {
5966                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
5967                 /*
5968                  * If we couldn't reserve metadata bytes try and use some from
5969                  * the global reserve.
5970                  */
5971                 if (ret && block_rsv != global_rsv) {
5972                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5973                         if (!ret)
5974                                 return global_rsv;
5975                         return ERR_PTR(ret);
5976                 } else if (ret) {
5977                         return ERR_PTR(ret);
5978                 }
5979                 return block_rsv;
5980         }
5981
5982         ret = block_rsv_use_bytes(block_rsv, blocksize);
5983         if (!ret)
5984                 return block_rsv;
5985         if (ret) {
5986                 static DEFINE_RATELIMIT_STATE(_rs,
5987                                 DEFAULT_RATELIMIT_INTERVAL,
5988                                 /*DEFAULT_RATELIMIT_BURST*/ 2);
5989                 if (__ratelimit(&_rs)) {
5990                         printk(KERN_DEBUG "btrfs: block rsv returned %d\n", ret);
5991                         WARN_ON(1);
5992                 }
5993                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
5994                 if (!ret) {
5995                         return block_rsv;
5996                 } else if (ret && block_rsv != global_rsv) {
5997                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5998                         if (!ret)
5999                                 return global_rsv;
6000                 }
6001         }
6002
6003         return ERR_PTR(-ENOSPC);
6004 }
6005
6006 static void unuse_block_rsv(struct btrfs_block_rsv *block_rsv, u32 blocksize)
6007 {
6008         block_rsv_add_bytes(block_rsv, blocksize, 0);
6009         block_rsv_release_bytes(block_rsv, NULL, 0);
6010 }
6011
6012 /*
6013  * finds a free extent and does all the dirty work required for allocation
6014  * returns the key for the extent through ins, and a tree buffer for
6015  * the first block of the extent through buf.
6016  *
6017  * returns the tree buffer or NULL.
6018  */
6019 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6020                                         struct btrfs_root *root, u32 blocksize,
6021                                         u64 parent, u64 root_objectid,
6022                                         struct btrfs_disk_key *key, int level,
6023                                         u64 hint, u64 empty_size)
6024 {
6025         struct btrfs_key ins;
6026         struct btrfs_block_rsv *block_rsv;
6027         struct extent_buffer *buf;
6028         u64 flags = 0;
6029         int ret;
6030
6031
6032         block_rsv = use_block_rsv(trans, root, blocksize);
6033         if (IS_ERR(block_rsv))
6034                 return ERR_CAST(block_rsv);
6035
6036         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6037                                    empty_size, hint, (u64)-1, &ins, 0);
6038         if (ret) {
6039                 unuse_block_rsv(block_rsv, blocksize);
6040                 return ERR_PTR(ret);
6041         }
6042
6043         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6044                                     blocksize, level);
6045         BUG_ON(IS_ERR(buf));
6046
6047         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6048                 if (parent == 0)
6049                         parent = ins.objectid;
6050                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6051         } else
6052                 BUG_ON(parent > 0);
6053
6054         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6055                 struct btrfs_delayed_extent_op *extent_op;
6056                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
6057                 BUG_ON(!extent_op);
6058                 if (key)
6059                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6060                 else
6061                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6062                 extent_op->flags_to_set = flags;
6063                 extent_op->update_key = 1;
6064                 extent_op->update_flags = 1;
6065                 extent_op->is_data = 0;
6066
6067                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
6068                                         ins.offset, parent, root_objectid,
6069                                         level, BTRFS_ADD_DELAYED_EXTENT,
6070                                         extent_op);
6071                 BUG_ON(ret);
6072         }
6073         return buf;
6074 }
6075
6076 struct walk_control {
6077         u64 refs[BTRFS_MAX_LEVEL];
6078         u64 flags[BTRFS_MAX_LEVEL];
6079         struct btrfs_key update_progress;
6080         int stage;
6081         int level;
6082         int shared_level;
6083         int update_ref;
6084         int keep_locks;
6085         int reada_slot;
6086         int reada_count;
6087 };
6088
6089 #define DROP_REFERENCE  1
6090 #define UPDATE_BACKREF  2
6091
6092 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6093                                      struct btrfs_root *root,
6094                                      struct walk_control *wc,
6095                                      struct btrfs_path *path)
6096 {
6097         u64 bytenr;
6098         u64 generation;
6099         u64 refs;
6100         u64 flags;
6101         u32 nritems;
6102         u32 blocksize;
6103         struct btrfs_key key;
6104         struct extent_buffer *eb;
6105         int ret;
6106         int slot;
6107         int nread = 0;
6108
6109         if (path->slots[wc->level] < wc->reada_slot) {
6110                 wc->reada_count = wc->reada_count * 2 / 3;
6111                 wc->reada_count = max(wc->reada_count, 2);
6112         } else {
6113                 wc->reada_count = wc->reada_count * 3 / 2;
6114                 wc->reada_count = min_t(int, wc->reada_count,
6115                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6116         }
6117
6118         eb = path->nodes[wc->level];
6119         nritems = btrfs_header_nritems(eb);
6120         blocksize = btrfs_level_size(root, wc->level - 1);
6121
6122         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6123                 if (nread >= wc->reada_count)
6124                         break;
6125
6126                 cond_resched();
6127                 bytenr = btrfs_node_blockptr(eb, slot);
6128                 generation = btrfs_node_ptr_generation(eb, slot);
6129
6130                 if (slot == path->slots[wc->level])
6131                         goto reada;
6132
6133                 if (wc->stage == UPDATE_BACKREF &&
6134                     generation <= root->root_key.offset)
6135                         continue;
6136
6137                 /* We don't lock the tree block, it's OK to be racy here */
6138                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6139                                                &refs, &flags);
6140                 BUG_ON(ret);
6141                 BUG_ON(refs == 0);
6142
6143                 if (wc->stage == DROP_REFERENCE) {
6144                         if (refs == 1)
6145                                 goto reada;
6146
6147                         if (wc->level == 1 &&
6148                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6149                                 continue;
6150                         if (!wc->update_ref ||
6151                             generation <= root->root_key.offset)
6152                                 continue;
6153                         btrfs_node_key_to_cpu(eb, &key, slot);
6154                         ret = btrfs_comp_cpu_keys(&key,
6155                                                   &wc->update_progress);
6156                         if (ret < 0)
6157                                 continue;
6158                 } else {
6159                         if (wc->level == 1 &&
6160                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6161                                 continue;
6162                 }
6163 reada:
6164                 ret = readahead_tree_block(root, bytenr, blocksize,
6165                                            generation);
6166                 if (ret)
6167                         break;
6168                 nread++;
6169         }
6170         wc->reada_slot = slot;
6171 }
6172
6173 /*
6174  * hepler to process tree block while walking down the tree.
6175  *
6176  * when wc->stage == UPDATE_BACKREF, this function updates
6177  * back refs for pointers in the block.
6178  *
6179  * NOTE: return value 1 means we should stop walking down.
6180  */
6181 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6182                                    struct btrfs_root *root,
6183                                    struct btrfs_path *path,
6184                                    struct walk_control *wc, int lookup_info)
6185 {
6186         int level = wc->level;
6187         struct extent_buffer *eb = path->nodes[level];
6188         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6189         int ret;
6190
6191         if (wc->stage == UPDATE_BACKREF &&
6192             btrfs_header_owner(eb) != root->root_key.objectid)
6193                 return 1;
6194
6195         /*
6196          * when reference count of tree block is 1, it won't increase
6197          * again. once full backref flag is set, we never clear it.
6198          */
6199         if (lookup_info &&
6200             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6201              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6202                 BUG_ON(!path->locks[level]);
6203                 ret = btrfs_lookup_extent_info(trans, root,
6204                                                eb->start, eb->len,
6205                                                &wc->refs[level],
6206                                                &wc->flags[level]);
6207                 BUG_ON(ret);
6208                 BUG_ON(wc->refs[level] == 0);
6209         }
6210
6211         if (wc->stage == DROP_REFERENCE) {
6212                 if (wc->refs[level] > 1)
6213                         return 1;
6214
6215                 if (path->locks[level] && !wc->keep_locks) {
6216                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6217                         path->locks[level] = 0;
6218                 }
6219                 return 0;
6220         }
6221
6222         /* wc->stage == UPDATE_BACKREF */
6223         if (!(wc->flags[level] & flag)) {
6224                 BUG_ON(!path->locks[level]);
6225                 ret = btrfs_inc_ref(trans, root, eb, 1);
6226                 BUG_ON(ret);
6227                 ret = btrfs_dec_ref(trans, root, eb, 0);
6228                 BUG_ON(ret);
6229                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6230                                                   eb->len, flag, 0);
6231                 BUG_ON(ret);
6232                 wc->flags[level] |= flag;
6233         }
6234
6235         /*
6236          * the block is shared by multiple trees, so it's not good to
6237          * keep the tree lock
6238          */
6239         if (path->locks[level] && level > 0) {
6240                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6241                 path->locks[level] = 0;
6242         }
6243         return 0;
6244 }
6245
6246 /*
6247  * hepler to process tree block pointer.
6248  *
6249  * when wc->stage == DROP_REFERENCE, this function checks
6250  * reference count of the block pointed to. if the block
6251  * is shared and we need update back refs for the subtree
6252  * rooted at the block, this function changes wc->stage to
6253  * UPDATE_BACKREF. if the block is shared and there is no
6254  * need to update back, this function drops the reference
6255  * to the block.
6256  *
6257  * NOTE: return value 1 means we should stop walking down.
6258  */
6259 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6260                                  struct btrfs_root *root,
6261                                  struct btrfs_path *path,
6262                                  struct walk_control *wc, int *lookup_info)
6263 {
6264         u64 bytenr;
6265         u64 generation;
6266         u64 parent;
6267         u32 blocksize;
6268         struct btrfs_key key;
6269         struct extent_buffer *next;
6270         int level = wc->level;
6271         int reada = 0;
6272         int ret = 0;
6273
6274         generation = btrfs_node_ptr_generation(path->nodes[level],
6275                                                path->slots[level]);
6276         /*
6277          * if the lower level block was created before the snapshot
6278          * was created, we know there is no need to update back refs
6279          * for the subtree
6280          */
6281         if (wc->stage == UPDATE_BACKREF &&
6282             generation <= root->root_key.offset) {
6283                 *lookup_info = 1;
6284                 return 1;
6285         }
6286
6287         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6288         blocksize = btrfs_level_size(root, level - 1);
6289
6290         next = btrfs_find_tree_block(root, bytenr, blocksize);
6291         if (!next) {
6292                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6293                 if (!next)
6294                         return -ENOMEM;
6295                 reada = 1;
6296         }
6297         btrfs_tree_lock(next);
6298         btrfs_set_lock_blocking(next);
6299
6300         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6301                                        &wc->refs[level - 1],
6302                                        &wc->flags[level - 1]);
6303         BUG_ON(ret);
6304         BUG_ON(wc->refs[level - 1] == 0);
6305         *lookup_info = 0;
6306
6307         if (wc->stage == DROP_REFERENCE) {
6308                 if (wc->refs[level - 1] > 1) {
6309                         if (level == 1 &&
6310                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6311                                 goto skip;
6312
6313                         if (!wc->update_ref ||
6314                             generation <= root->root_key.offset)
6315                                 goto skip;
6316
6317                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6318                                               path->slots[level]);
6319                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6320                         if (ret < 0)
6321                                 goto skip;
6322
6323                         wc->stage = UPDATE_BACKREF;
6324                         wc->shared_level = level - 1;
6325                 }
6326         } else {
6327                 if (level == 1 &&
6328                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6329                         goto skip;
6330         }
6331
6332         if (!btrfs_buffer_uptodate(next, generation)) {
6333                 btrfs_tree_unlock(next);
6334                 free_extent_buffer(next);
6335                 next = NULL;
6336                 *lookup_info = 1;
6337         }
6338
6339         if (!next) {
6340                 if (reada && level == 1)
6341                         reada_walk_down(trans, root, wc, path);
6342                 next = read_tree_block(root, bytenr, blocksize, generation);
6343                 if (!next)
6344                         return -EIO;
6345                 btrfs_tree_lock(next);
6346                 btrfs_set_lock_blocking(next);
6347         }
6348
6349         level--;
6350         BUG_ON(level != btrfs_header_level(next));
6351         path->nodes[level] = next;
6352         path->slots[level] = 0;
6353         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6354         wc->level = level;
6355         if (wc->level == 1)
6356                 wc->reada_slot = 0;
6357         return 0;
6358 skip:
6359         wc->refs[level - 1] = 0;
6360         wc->flags[level - 1] = 0;
6361         if (wc->stage == DROP_REFERENCE) {
6362                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6363                         parent = path->nodes[level]->start;
6364                 } else {
6365                         BUG_ON(root->root_key.objectid !=
6366                                btrfs_header_owner(path->nodes[level]));
6367                         parent = 0;
6368                 }
6369
6370                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6371                                         root->root_key.objectid, level - 1, 0);
6372                 BUG_ON(ret);
6373         }
6374         btrfs_tree_unlock(next);
6375         free_extent_buffer(next);
6376         *lookup_info = 1;
6377         return 1;
6378 }
6379
6380 /*
6381  * hepler to process tree block while walking up the tree.
6382  *
6383  * when wc->stage == DROP_REFERENCE, this function drops
6384  * reference count on the block.
6385  *
6386  * when wc->stage == UPDATE_BACKREF, this function changes
6387  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6388  * to UPDATE_BACKREF previously while processing the block.
6389  *
6390  * NOTE: return value 1 means we should stop walking up.
6391  */
6392 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6393                                  struct btrfs_root *root,
6394                                  struct btrfs_path *path,
6395                                  struct walk_control *wc)
6396 {
6397         int ret;
6398         int level = wc->level;
6399         struct extent_buffer *eb = path->nodes[level];
6400         u64 parent = 0;
6401
6402         if (wc->stage == UPDATE_BACKREF) {
6403                 BUG_ON(wc->shared_level < level);
6404                 if (level < wc->shared_level)
6405                         goto out;
6406
6407                 ret = find_next_key(path, level + 1, &wc->update_progress);
6408                 if (ret > 0)
6409                         wc->update_ref = 0;
6410
6411                 wc->stage = DROP_REFERENCE;
6412                 wc->shared_level = -1;
6413                 path->slots[level] = 0;
6414
6415                 /*
6416                  * check reference count again if the block isn't locked.
6417                  * we should start walking down the tree again if reference
6418                  * count is one.
6419                  */
6420                 if (!path->locks[level]) {
6421                         BUG_ON(level == 0);
6422                         btrfs_tree_lock(eb);
6423                         btrfs_set_lock_blocking(eb);
6424                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6425
6426                         ret = btrfs_lookup_extent_info(trans, root,
6427                                                        eb->start, eb->len,
6428                                                        &wc->refs[level],
6429                                                        &wc->flags[level]);
6430                         BUG_ON(ret);
6431                         BUG_ON(wc->refs[level] == 0);
6432                         if (wc->refs[level] == 1) {
6433                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6434                                 return 1;
6435                         }
6436                 }
6437         }
6438
6439         /* wc->stage == DROP_REFERENCE */
6440         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6441
6442         if (wc->refs[level] == 1) {
6443                 if (level == 0) {
6444                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6445                                 ret = btrfs_dec_ref(trans, root, eb, 1);
6446                         else
6447                                 ret = btrfs_dec_ref(trans, root, eb, 0);
6448                         BUG_ON(ret);
6449                 }
6450                 /* make block locked assertion in clean_tree_block happy */
6451                 if (!path->locks[level] &&
6452                     btrfs_header_generation(eb) == trans->transid) {
6453                         btrfs_tree_lock(eb);
6454                         btrfs_set_lock_blocking(eb);
6455                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6456                 }
6457                 clean_tree_block(trans, root, eb);
6458         }
6459
6460         if (eb == root->node) {
6461                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6462                         parent = eb->start;
6463                 else
6464                         BUG_ON(root->root_key.objectid !=
6465                                btrfs_header_owner(eb));
6466         } else {
6467                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6468                         parent = path->nodes[level + 1]->start;
6469                 else
6470                         BUG_ON(root->root_key.objectid !=
6471                                btrfs_header_owner(path->nodes[level + 1]));
6472         }
6473
6474         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6475 out:
6476         wc->refs[level] = 0;
6477         wc->flags[level] = 0;
6478         return 0;
6479 }
6480
6481 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6482                                    struct btrfs_root *root,
6483                                    struct btrfs_path *path,
6484                                    struct walk_control *wc)
6485 {
6486         int level = wc->level;
6487         int lookup_info = 1;
6488         int ret;
6489
6490         while (level >= 0) {
6491                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6492                 if (ret > 0)
6493                         break;
6494
6495                 if (level == 0)
6496                         break;
6497
6498                 if (path->slots[level] >=
6499                     btrfs_header_nritems(path->nodes[level]))
6500                         break;
6501
6502                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6503                 if (ret > 0) {
6504                         path->slots[level]++;
6505                         continue;
6506                 } else if (ret < 0)
6507                         return ret;
6508                 level = wc->level;
6509         }
6510         return 0;
6511 }
6512
6513 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6514                                  struct btrfs_root *root,
6515                                  struct btrfs_path *path,
6516                                  struct walk_control *wc, int max_level)
6517 {
6518         int level = wc->level;
6519         int ret;
6520
6521         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6522         while (level < max_level && path->nodes[level]) {
6523                 wc->level = level;
6524                 if (path->slots[level] + 1 <
6525                     btrfs_header_nritems(path->nodes[level])) {
6526                         path->slots[level]++;
6527                         return 0;
6528                 } else {
6529                         ret = walk_up_proc(trans, root, path, wc);
6530                         if (ret > 0)
6531                                 return 0;
6532
6533                         if (path->locks[level]) {
6534                                 btrfs_tree_unlock_rw(path->nodes[level],
6535                                                      path->locks[level]);
6536                                 path->locks[level] = 0;
6537                         }
6538                         free_extent_buffer(path->nodes[level]);
6539                         path->nodes[level] = NULL;
6540                         level++;
6541                 }
6542         }
6543         return 1;
6544 }
6545
6546 /*
6547  * drop a subvolume tree.
6548  *
6549  * this function traverses the tree freeing any blocks that only
6550  * referenced by the tree.
6551  *
6552  * when a shared tree block is found. this function decreases its
6553  * reference count by one. if update_ref is true, this function
6554  * also make sure backrefs for the shared block and all lower level
6555  * blocks are properly updated.
6556  */
6557 void btrfs_drop_snapshot(struct btrfs_root *root,
6558                          struct btrfs_block_rsv *block_rsv, int update_ref)
6559 {
6560         struct btrfs_path *path;
6561         struct btrfs_trans_handle *trans;
6562         struct btrfs_root *tree_root = root->fs_info->tree_root;
6563         struct btrfs_root_item *root_item = &root->root_item;
6564         struct walk_control *wc;
6565         struct btrfs_key key;
6566         int err = 0;
6567         int ret;
6568         int level;
6569
6570         path = btrfs_alloc_path();
6571         if (!path) {
6572                 err = -ENOMEM;
6573                 goto out;
6574         }
6575
6576         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6577         if (!wc) {
6578                 btrfs_free_path(path);
6579                 err = -ENOMEM;
6580                 goto out;
6581         }
6582
6583         trans = btrfs_start_transaction(tree_root, 0);
6584         BUG_ON(IS_ERR(trans));
6585
6586         if (block_rsv)
6587                 trans->block_rsv = block_rsv;
6588
6589         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6590                 level = btrfs_header_level(root->node);
6591                 path->nodes[level] = btrfs_lock_root_node(root);
6592                 btrfs_set_lock_blocking(path->nodes[level]);
6593                 path->slots[level] = 0;
6594                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6595                 memset(&wc->update_progress, 0,
6596                        sizeof(wc->update_progress));
6597         } else {
6598                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6599                 memcpy(&wc->update_progress, &key,
6600                        sizeof(wc->update_progress));
6601
6602                 level = root_item->drop_level;
6603                 BUG_ON(level == 0);
6604                 path->lowest_level = level;
6605                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6606                 path->lowest_level = 0;
6607                 if (ret < 0) {
6608                         err = ret;
6609                         goto out_free;
6610                 }
6611                 WARN_ON(ret > 0);
6612
6613                 /*
6614                  * unlock our path, this is safe because only this
6615                  * function is allowed to delete this snapshot
6616                  */
6617                 btrfs_unlock_up_safe(path, 0);
6618
6619                 level = btrfs_header_level(root->node);
6620                 while (1) {
6621                         btrfs_tree_lock(path->nodes[level]);
6622                         btrfs_set_lock_blocking(path->nodes[level]);
6623
6624                         ret = btrfs_lookup_extent_info(trans, root,
6625                                                 path->nodes[level]->start,
6626                                                 path->nodes[level]->len,
6627                                                 &wc->refs[level],
6628                                                 &wc->flags[level]);
6629                         BUG_ON(ret);
6630                         BUG_ON(wc->refs[level] == 0);
6631
6632                         if (level == root_item->drop_level)
6633                                 break;
6634
6635                         btrfs_tree_unlock(path->nodes[level]);
6636                         WARN_ON(wc->refs[level] != 1);
6637                         level--;
6638                 }
6639         }
6640
6641         wc->level = level;
6642         wc->shared_level = -1;
6643         wc->stage = DROP_REFERENCE;
6644         wc->update_ref = update_ref;
6645         wc->keep_locks = 0;
6646         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6647
6648         while (1) {
6649                 ret = walk_down_tree(trans, root, path, wc);
6650                 if (ret < 0) {
6651                         err = ret;
6652                         break;
6653                 }
6654
6655                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
6656                 if (ret < 0) {
6657                         err = ret;
6658                         break;
6659                 }
6660
6661                 if (ret > 0) {
6662                         BUG_ON(wc->stage != DROP_REFERENCE);
6663                         break;
6664                 }
6665
6666                 if (wc->stage == DROP_REFERENCE) {
6667                         level = wc->level;
6668                         btrfs_node_key(path->nodes[level],
6669                                        &root_item->drop_progress,
6670                                        path->slots[level]);
6671                         root_item->drop_level = level;
6672                 }
6673
6674                 BUG_ON(wc->level == 0);
6675                 if (btrfs_should_end_transaction(trans, tree_root)) {
6676                         ret = btrfs_update_root(trans, tree_root,
6677                                                 &root->root_key,
6678                                                 root_item);
6679                         BUG_ON(ret);
6680
6681                         btrfs_end_transaction_throttle(trans, tree_root);
6682                         trans = btrfs_start_transaction(tree_root, 0);
6683                         BUG_ON(IS_ERR(trans));
6684                         if (block_rsv)
6685                                 trans->block_rsv = block_rsv;
6686                 }
6687         }
6688         btrfs_release_path(path);
6689         BUG_ON(err);
6690
6691         ret = btrfs_del_root(trans, tree_root, &root->root_key);
6692         BUG_ON(ret);
6693
6694         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
6695                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
6696                                            NULL, NULL);
6697                 BUG_ON(ret < 0);
6698                 if (ret > 0) {
6699                         /* if we fail to delete the orphan item this time
6700                          * around, it'll get picked up the next time.
6701                          *
6702                          * The most common failure here is just -ENOENT.
6703                          */
6704                         btrfs_del_orphan_item(trans, tree_root,
6705                                               root->root_key.objectid);
6706                 }
6707         }
6708
6709         if (root->in_radix) {
6710                 btrfs_free_fs_root(tree_root->fs_info, root);
6711         } else {
6712                 free_extent_buffer(root->node);
6713                 free_extent_buffer(root->commit_root);
6714                 kfree(root);
6715         }
6716 out_free:
6717         btrfs_end_transaction_throttle(trans, tree_root);
6718         kfree(wc);
6719         btrfs_free_path(path);
6720 out:
6721         if (err)
6722                 btrfs_std_error(root->fs_info, err);
6723         return;
6724 }
6725
6726 /*
6727  * drop subtree rooted at tree block 'node'.
6728  *
6729  * NOTE: this function will unlock and release tree block 'node'
6730  */
6731 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
6732                         struct btrfs_root *root,
6733                         struct extent_buffer *node,
6734                         struct extent_buffer *parent)
6735 {
6736         struct btrfs_path *path;
6737         struct walk_control *wc;
6738         int level;
6739         int parent_level;
6740         int ret = 0;
6741         int wret;
6742
6743         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
6744
6745         path = btrfs_alloc_path();
6746         if (!path)
6747                 return -ENOMEM;
6748
6749         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6750         if (!wc) {
6751                 btrfs_free_path(path);
6752                 return -ENOMEM;
6753         }
6754
6755         btrfs_assert_tree_locked(parent);
6756         parent_level = btrfs_header_level(parent);
6757         extent_buffer_get(parent);
6758         path->nodes[parent_level] = parent;
6759         path->slots[parent_level] = btrfs_header_nritems(parent);
6760
6761         btrfs_assert_tree_locked(node);
6762         level = btrfs_header_level(node);
6763         path->nodes[level] = node;
6764         path->slots[level] = 0;
6765         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6766
6767         wc->refs[parent_level] = 1;
6768         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6769         wc->level = level;
6770         wc->shared_level = -1;
6771         wc->stage = DROP_REFERENCE;
6772         wc->update_ref = 0;
6773         wc->keep_locks = 1;
6774         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6775
6776         while (1) {
6777                 wret = walk_down_tree(trans, root, path, wc);
6778                 if (wret < 0) {
6779                         ret = wret;
6780                         break;
6781                 }
6782
6783                 wret = walk_up_tree(trans, root, path, wc, parent_level);
6784                 if (wret < 0)
6785                         ret = wret;
6786                 if (wret != 0)
6787                         break;
6788         }
6789
6790         kfree(wc);
6791         btrfs_free_path(path);
6792         return ret;
6793 }
6794
6795 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
6796 {
6797         u64 num_devices;
6798         u64 stripped = BTRFS_BLOCK_GROUP_RAID0 |
6799                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
6800
6801         /*
6802          * we add in the count of missing devices because we want
6803          * to make sure that any RAID levels on a degraded FS
6804          * continue to be honored.
6805          */
6806         num_devices = root->fs_info->fs_devices->rw_devices +
6807                 root->fs_info->fs_devices->missing_devices;
6808
6809         if (num_devices == 1) {
6810                 stripped |= BTRFS_BLOCK_GROUP_DUP;
6811                 stripped = flags & ~stripped;
6812
6813                 /* turn raid0 into single device chunks */
6814                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
6815                         return stripped;
6816
6817                 /* turn mirroring into duplication */
6818                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
6819                              BTRFS_BLOCK_GROUP_RAID10))
6820                         return stripped | BTRFS_BLOCK_GROUP_DUP;
6821                 return flags;
6822         } else {
6823                 /* they already had raid on here, just return */
6824                 if (flags & stripped)
6825                         return flags;
6826
6827                 stripped |= BTRFS_BLOCK_GROUP_DUP;
6828                 stripped = flags & ~stripped;
6829
6830                 /* switch duplicated blocks with raid1 */
6831                 if (flags & BTRFS_BLOCK_GROUP_DUP)
6832                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
6833
6834                 /* turn single device chunks into raid0 */
6835                 return stripped | BTRFS_BLOCK_GROUP_RAID0;
6836         }
6837         return flags;
6838 }
6839
6840 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
6841 {
6842         struct btrfs_space_info *sinfo = cache->space_info;
6843         u64 num_bytes;
6844         u64 min_allocable_bytes;
6845         int ret = -ENOSPC;
6846
6847
6848         /*
6849          * We need some metadata space and system metadata space for
6850          * allocating chunks in some corner cases until we force to set
6851          * it to be readonly.
6852          */
6853         if ((sinfo->flags &
6854              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
6855             !force)
6856                 min_allocable_bytes = 1 * 1024 * 1024;
6857         else
6858                 min_allocable_bytes = 0;
6859
6860         spin_lock(&sinfo->lock);
6861         spin_lock(&cache->lock);
6862
6863         if (cache->ro) {
6864                 ret = 0;
6865                 goto out;
6866         }
6867
6868         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
6869                     cache->bytes_super - btrfs_block_group_used(&cache->item);
6870
6871         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
6872             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
6873             min_allocable_bytes <= sinfo->total_bytes) {
6874                 sinfo->bytes_readonly += num_bytes;
6875                 cache->ro = 1;
6876                 ret = 0;
6877         }
6878 out:
6879         spin_unlock(&cache->lock);
6880         spin_unlock(&sinfo->lock);
6881         return ret;
6882 }
6883
6884 int btrfs_set_block_group_ro(struct btrfs_root *root,
6885                              struct btrfs_block_group_cache *cache)
6886
6887 {
6888         struct btrfs_trans_handle *trans;
6889         u64 alloc_flags;
6890         int ret;
6891
6892         BUG_ON(cache->ro);
6893
6894         trans = btrfs_join_transaction(root);
6895         BUG_ON(IS_ERR(trans));
6896
6897         alloc_flags = update_block_group_flags(root, cache->flags);
6898         if (alloc_flags != cache->flags)
6899                 do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6900                                CHUNK_ALLOC_FORCE);
6901
6902         ret = set_block_group_ro(cache, 0);
6903         if (!ret)
6904                 goto out;
6905         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
6906         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6907                              CHUNK_ALLOC_FORCE);
6908         if (ret < 0)
6909                 goto out;
6910         ret = set_block_group_ro(cache, 0);
6911 out:
6912         btrfs_end_transaction(trans, root);
6913         return ret;
6914 }
6915
6916 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
6917                             struct btrfs_root *root, u64 type)
6918 {
6919         u64 alloc_flags = get_alloc_profile(root, type);
6920         return do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6921                               CHUNK_ALLOC_FORCE);
6922 }
6923
6924 /*
6925  * helper to account the unused space of all the readonly block group in the
6926  * list. takes mirrors into account.
6927  */
6928 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
6929 {
6930         struct btrfs_block_group_cache *block_group;
6931         u64 free_bytes = 0;
6932         int factor;
6933
6934         list_for_each_entry(block_group, groups_list, list) {
6935                 spin_lock(&block_group->lock);
6936
6937                 if (!block_group->ro) {
6938                         spin_unlock(&block_group->lock);
6939                         continue;
6940                 }
6941
6942                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
6943                                           BTRFS_BLOCK_GROUP_RAID10 |
6944                                           BTRFS_BLOCK_GROUP_DUP))
6945                         factor = 2;
6946                 else
6947                         factor = 1;
6948
6949                 free_bytes += (block_group->key.offset -
6950                                btrfs_block_group_used(&block_group->item)) *
6951                                factor;
6952
6953                 spin_unlock(&block_group->lock);
6954         }
6955
6956         return free_bytes;
6957 }
6958
6959 /*
6960  * helper to account the unused space of all the readonly block group in the
6961  * space_info. takes mirrors into account.
6962  */
6963 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
6964 {
6965         int i;
6966         u64 free_bytes = 0;
6967
6968         spin_lock(&sinfo->lock);
6969
6970         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
6971                 if (!list_empty(&sinfo->block_groups[i]))
6972                         free_bytes += __btrfs_get_ro_block_group_free_space(
6973                                                 &sinfo->block_groups[i]);
6974
6975         spin_unlock(&sinfo->lock);
6976
6977         return free_bytes;
6978 }
6979
6980 int btrfs_set_block_group_rw(struct btrfs_root *root,
6981                               struct btrfs_block_group_cache *cache)
6982 {
6983         struct btrfs_space_info *sinfo = cache->space_info;
6984         u64 num_bytes;
6985
6986         BUG_ON(!cache->ro);
6987
6988         spin_lock(&sinfo->lock);
6989         spin_lock(&cache->lock);
6990         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
6991                     cache->bytes_super - btrfs_block_group_used(&cache->item);
6992         sinfo->bytes_readonly -= num_bytes;
6993         cache->ro = 0;
6994         spin_unlock(&cache->lock);
6995         spin_unlock(&sinfo->lock);
6996         return 0;
6997 }
6998
6999 /*
7000  * checks to see if its even possible to relocate this block group.
7001  *
7002  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7003  * ok to go ahead and try.
7004  */
7005 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7006 {
7007         struct btrfs_block_group_cache *block_group;
7008         struct btrfs_space_info *space_info;
7009         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7010         struct btrfs_device *device;
7011         u64 min_free;
7012         u64 dev_min = 1;
7013         u64 dev_nr = 0;
7014         int index;
7015         int full = 0;
7016         int ret = 0;
7017
7018         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7019
7020         /* odd, couldn't find the block group, leave it alone */
7021         if (!block_group)
7022                 return -1;
7023
7024         min_free = btrfs_block_group_used(&block_group->item);
7025
7026         /* no bytes used, we're good */
7027         if (!min_free)
7028                 goto out;
7029
7030         space_info = block_group->space_info;
7031         spin_lock(&space_info->lock);
7032
7033         full = space_info->full;
7034
7035         /*
7036          * if this is the last block group we have in this space, we can't
7037          * relocate it unless we're able to allocate a new chunk below.
7038          *
7039          * Otherwise, we need to make sure we have room in the space to handle
7040          * all of the extents from this block group.  If we can, we're good
7041          */
7042         if ((space_info->total_bytes != block_group->key.offset) &&
7043             (space_info->bytes_used + space_info->bytes_reserved +
7044              space_info->bytes_pinned + space_info->bytes_readonly +
7045              min_free < space_info->total_bytes)) {
7046                 spin_unlock(&space_info->lock);
7047                 goto out;
7048         }
7049         spin_unlock(&space_info->lock);
7050
7051         /*
7052          * ok we don't have enough space, but maybe we have free space on our
7053          * devices to allocate new chunks for relocation, so loop through our
7054          * alloc devices and guess if we have enough space.  However, if we
7055          * were marked as full, then we know there aren't enough chunks, and we
7056          * can just return.
7057          */
7058         ret = -1;
7059         if (full)
7060                 goto out;
7061
7062         /*
7063          * index:
7064          *      0: raid10
7065          *      1: raid1
7066          *      2: dup
7067          *      3: raid0
7068          *      4: single
7069          */
7070         index = get_block_group_index(block_group);
7071         if (index == 0) {
7072                 dev_min = 4;
7073                 /* Divide by 2 */
7074                 min_free >>= 1;
7075         } else if (index == 1) {
7076                 dev_min = 2;
7077         } else if (index == 2) {
7078                 /* Multiply by 2 */
7079                 min_free <<= 1;
7080         } else if (index == 3) {
7081                 dev_min = fs_devices->rw_devices;
7082                 do_div(min_free, dev_min);
7083         }
7084
7085         mutex_lock(&root->fs_info->chunk_mutex);
7086         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7087                 u64 dev_offset;
7088
7089                 /*
7090                  * check to make sure we can actually find a chunk with enough
7091                  * space to fit our block group in.
7092                  */
7093                 if (device->total_bytes > device->bytes_used + min_free) {
7094                         ret = find_free_dev_extent(NULL, device, min_free,
7095                                                    &dev_offset, NULL);
7096                         if (!ret)
7097                                 dev_nr++;
7098
7099                         if (dev_nr >= dev_min)
7100                                 break;
7101
7102                         ret = -1;
7103                 }
7104         }
7105         mutex_unlock(&root->fs_info->chunk_mutex);
7106 out:
7107         btrfs_put_block_group(block_group);
7108         return ret;
7109 }
7110
7111 static int find_first_block_group(struct btrfs_root *root,
7112                 struct btrfs_path *path, struct btrfs_key *key)
7113 {
7114         int ret = 0;
7115         struct btrfs_key found_key;
7116         struct extent_buffer *leaf;
7117         int slot;
7118
7119         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7120         if (ret < 0)
7121                 goto out;
7122
7123         while (1) {
7124                 slot = path->slots[0];
7125                 leaf = path->nodes[0];
7126                 if (slot >= btrfs_header_nritems(leaf)) {
7127                         ret = btrfs_next_leaf(root, path);
7128                         if (ret == 0)
7129                                 continue;
7130                         if (ret < 0)
7131                                 goto out;
7132                         break;
7133                 }
7134                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7135
7136                 if (found_key.objectid >= key->objectid &&
7137                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7138                         ret = 0;
7139                         goto out;
7140                 }
7141                 path->slots[0]++;
7142         }
7143 out:
7144         return ret;
7145 }
7146
7147 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7148 {
7149         struct btrfs_block_group_cache *block_group;
7150         u64 last = 0;
7151
7152         while (1) {
7153                 struct inode *inode;
7154
7155                 block_group = btrfs_lookup_first_block_group(info, last);
7156                 while (block_group) {
7157                         spin_lock(&block_group->lock);
7158                         if (block_group->iref)
7159                                 break;
7160                         spin_unlock(&block_group->lock);
7161                         block_group = next_block_group(info->tree_root,
7162                                                        block_group);
7163                 }
7164                 if (!block_group) {
7165                         if (last == 0)
7166                                 break;
7167                         last = 0;
7168                         continue;
7169                 }
7170
7171                 inode = block_group->inode;
7172                 block_group->iref = 0;
7173                 block_group->inode = NULL;
7174                 spin_unlock(&block_group->lock);
7175                 iput(inode);
7176                 last = block_group->key.objectid + block_group->key.offset;
7177                 btrfs_put_block_group(block_group);
7178         }
7179 }
7180
7181 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7182 {
7183         struct btrfs_block_group_cache *block_group;
7184         struct btrfs_space_info *space_info;
7185         struct btrfs_caching_control *caching_ctl;
7186         struct rb_node *n;
7187
7188         down_write(&info->extent_commit_sem);
7189         while (!list_empty(&info->caching_block_groups)) {
7190                 caching_ctl = list_entry(info->caching_block_groups.next,
7191                                          struct btrfs_caching_control, list);
7192                 list_del(&caching_ctl->list);
7193                 put_caching_control(caching_ctl);
7194         }
7195         up_write(&info->extent_commit_sem);
7196
7197         spin_lock(&info->block_group_cache_lock);
7198         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7199                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7200                                        cache_node);
7201                 rb_erase(&block_group->cache_node,
7202                          &info->block_group_cache_tree);
7203                 spin_unlock(&info->block_group_cache_lock);
7204
7205                 down_write(&block_group->space_info->groups_sem);
7206                 list_del(&block_group->list);
7207                 up_write(&block_group->space_info->groups_sem);
7208
7209                 if (block_group->cached == BTRFS_CACHE_STARTED)
7210                         wait_block_group_cache_done(block_group);
7211
7212                 /*
7213                  * We haven't cached this block group, which means we could
7214                  * possibly have excluded extents on this block group.
7215                  */
7216                 if (block_group->cached == BTRFS_CACHE_NO)
7217                         free_excluded_extents(info->extent_root, block_group);
7218
7219                 btrfs_remove_free_space_cache(block_group);
7220                 btrfs_put_block_group(block_group);
7221
7222                 spin_lock(&info->block_group_cache_lock);
7223         }
7224         spin_unlock(&info->block_group_cache_lock);
7225
7226         /* now that all the block groups are freed, go through and
7227          * free all the space_info structs.  This is only called during
7228          * the final stages of unmount, and so we know nobody is
7229          * using them.  We call synchronize_rcu() once before we start,
7230          * just to be on the safe side.
7231          */
7232         synchronize_rcu();
7233
7234         release_global_block_rsv(info);
7235
7236         while(!list_empty(&info->space_info)) {
7237                 space_info = list_entry(info->space_info.next,
7238                                         struct btrfs_space_info,
7239                                         list);
7240                 if (space_info->bytes_pinned > 0 ||
7241                     space_info->bytes_reserved > 0 ||
7242                     space_info->bytes_may_use > 0) {
7243                         WARN_ON(1);
7244                         dump_space_info(space_info, 0, 0);
7245                 }
7246                 list_del(&space_info->list);
7247                 kfree(space_info);
7248         }
7249         return 0;
7250 }
7251
7252 static void __link_block_group(struct btrfs_space_info *space_info,
7253                                struct btrfs_block_group_cache *cache)
7254 {
7255         int index = get_block_group_index(cache);
7256
7257         down_write(&space_info->groups_sem);
7258         list_add_tail(&cache->list, &space_info->block_groups[index]);
7259         up_write(&space_info->groups_sem);
7260 }
7261
7262 int btrfs_read_block_groups(struct btrfs_root *root)
7263 {
7264         struct btrfs_path *path;
7265         int ret;
7266         struct btrfs_block_group_cache *cache;
7267         struct btrfs_fs_info *info = root->fs_info;
7268         struct btrfs_space_info *space_info;
7269         struct btrfs_key key;
7270         struct btrfs_key found_key;
7271         struct extent_buffer *leaf;
7272         int need_clear = 0;
7273         u64 cache_gen;
7274
7275         root = info->extent_root;
7276         key.objectid = 0;
7277         key.offset = 0;
7278         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7279         path = btrfs_alloc_path();
7280         if (!path)
7281                 return -ENOMEM;
7282         path->reada = 1;
7283
7284         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7285         if (btrfs_test_opt(root, SPACE_CACHE) &&
7286             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7287                 need_clear = 1;
7288         if (btrfs_test_opt(root, CLEAR_CACHE))
7289                 need_clear = 1;
7290
7291         while (1) {
7292                 ret = find_first_block_group(root, path, &key);
7293                 if (ret > 0)
7294                         break;
7295                 if (ret != 0)
7296                         goto error;
7297                 leaf = path->nodes[0];
7298                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7299                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7300                 if (!cache) {
7301                         ret = -ENOMEM;
7302                         goto error;
7303                 }
7304                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7305                                                 GFP_NOFS);
7306                 if (!cache->free_space_ctl) {
7307                         kfree(cache);
7308                         ret = -ENOMEM;
7309                         goto error;
7310                 }
7311
7312                 atomic_set(&cache->count, 1);
7313                 spin_lock_init(&cache->lock);
7314                 cache->fs_info = info;
7315                 INIT_LIST_HEAD(&cache->list);
7316                 INIT_LIST_HEAD(&cache->cluster_list);
7317
7318                 if (need_clear)
7319                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7320
7321                 read_extent_buffer(leaf, &cache->item,
7322                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7323                                    sizeof(cache->item));
7324                 memcpy(&cache->key, &found_key, sizeof(found_key));
7325
7326                 key.objectid = found_key.objectid + found_key.offset;
7327                 btrfs_release_path(path);
7328                 cache->flags = btrfs_block_group_flags(&cache->item);
7329                 cache->sectorsize = root->sectorsize;
7330
7331                 btrfs_init_free_space_ctl(cache);
7332
7333                 /*
7334                  * We need to exclude the super stripes now so that the space
7335                  * info has super bytes accounted for, otherwise we'll think
7336                  * we have more space than we actually do.
7337                  */
7338                 exclude_super_stripes(root, cache);
7339
7340                 /*
7341                  * check for two cases, either we are full, and therefore
7342                  * don't need to bother with the caching work since we won't
7343                  * find any space, or we are empty, and we can just add all
7344                  * the space in and be done with it.  This saves us _alot_ of
7345                  * time, particularly in the full case.
7346                  */
7347                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7348                         cache->last_byte_to_unpin = (u64)-1;
7349                         cache->cached = BTRFS_CACHE_FINISHED;
7350                         free_excluded_extents(root, cache);
7351                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7352                         cache->last_byte_to_unpin = (u64)-1;
7353                         cache->cached = BTRFS_CACHE_FINISHED;
7354                         add_new_free_space(cache, root->fs_info,
7355                                            found_key.objectid,
7356                                            found_key.objectid +
7357                                            found_key.offset);
7358                         free_excluded_extents(root, cache);
7359                 }
7360
7361                 ret = update_space_info(info, cache->flags, found_key.offset,
7362                                         btrfs_block_group_used(&cache->item),
7363                                         &space_info);
7364                 BUG_ON(ret);
7365                 cache->space_info = space_info;
7366                 spin_lock(&cache->space_info->lock);
7367                 cache->space_info->bytes_readonly += cache->bytes_super;
7368                 spin_unlock(&cache->space_info->lock);
7369
7370                 __link_block_group(space_info, cache);
7371
7372                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7373                 BUG_ON(ret);
7374
7375                 set_avail_alloc_bits(root->fs_info, cache->flags);
7376                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7377                         set_block_group_ro(cache, 1);
7378         }
7379
7380         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7381                 if (!(get_alloc_profile(root, space_info->flags) &
7382                       (BTRFS_BLOCK_GROUP_RAID10 |
7383                        BTRFS_BLOCK_GROUP_RAID1 |
7384                        BTRFS_BLOCK_GROUP_DUP)))
7385                         continue;
7386                 /*
7387                  * avoid allocating from un-mirrored block group if there are
7388                  * mirrored block groups.
7389                  */
7390                 list_for_each_entry(cache, &space_info->block_groups[3], list)
7391                         set_block_group_ro(cache, 1);
7392                 list_for_each_entry(cache, &space_info->block_groups[4], list)
7393                         set_block_group_ro(cache, 1);
7394         }
7395
7396         init_global_block_rsv(info);
7397         ret = 0;
7398 error:
7399         btrfs_free_path(path);
7400         return ret;
7401 }
7402
7403 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7404                            struct btrfs_root *root, u64 bytes_used,
7405                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7406                            u64 size)
7407 {
7408         int ret;
7409         struct btrfs_root *extent_root;
7410         struct btrfs_block_group_cache *cache;
7411
7412         extent_root = root->fs_info->extent_root;
7413
7414         root->fs_info->last_trans_log_full_commit = trans->transid;
7415
7416         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7417         if (!cache)
7418                 return -ENOMEM;
7419         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7420                                         GFP_NOFS);
7421         if (!cache->free_space_ctl) {
7422                 kfree(cache);
7423                 return -ENOMEM;
7424         }
7425
7426         cache->key.objectid = chunk_offset;
7427         cache->key.offset = size;
7428         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7429         cache->sectorsize = root->sectorsize;
7430         cache->fs_info = root->fs_info;
7431
7432         atomic_set(&cache->count, 1);
7433         spin_lock_init(&cache->lock);
7434         INIT_LIST_HEAD(&cache->list);
7435         INIT_LIST_HEAD(&cache->cluster_list);
7436
7437         btrfs_init_free_space_ctl(cache);
7438
7439         btrfs_set_block_group_used(&cache->item, bytes_used);
7440         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7441         cache->flags = type;
7442         btrfs_set_block_group_flags(&cache->item, type);
7443
7444         cache->last_byte_to_unpin = (u64)-1;
7445         cache->cached = BTRFS_CACHE_FINISHED;
7446         exclude_super_stripes(root, cache);
7447
7448         add_new_free_space(cache, root->fs_info, chunk_offset,
7449                            chunk_offset + size);
7450
7451         free_excluded_extents(root, cache);
7452
7453         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7454                                 &cache->space_info);
7455         BUG_ON(ret);
7456
7457         spin_lock(&cache->space_info->lock);
7458         cache->space_info->bytes_readonly += cache->bytes_super;
7459         spin_unlock(&cache->space_info->lock);
7460
7461         __link_block_group(cache->space_info, cache);
7462
7463         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7464         BUG_ON(ret);
7465
7466         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
7467                                 sizeof(cache->item));
7468         BUG_ON(ret);
7469
7470         set_avail_alloc_bits(extent_root->fs_info, type);
7471
7472         return 0;
7473 }
7474
7475 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7476                              struct btrfs_root *root, u64 group_start)
7477 {
7478         struct btrfs_path *path;
7479         struct btrfs_block_group_cache *block_group;
7480         struct btrfs_free_cluster *cluster;
7481         struct btrfs_root *tree_root = root->fs_info->tree_root;
7482         struct btrfs_key key;
7483         struct inode *inode;
7484         int ret;
7485         int factor;
7486
7487         root = root->fs_info->extent_root;
7488
7489         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7490         BUG_ON(!block_group);
7491         BUG_ON(!block_group->ro);
7492
7493         /*
7494          * Free the reserved super bytes from this block group before
7495          * remove it.
7496          */
7497         free_excluded_extents(root, block_group);
7498
7499         memcpy(&key, &block_group->key, sizeof(key));
7500         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7501                                   BTRFS_BLOCK_GROUP_RAID1 |
7502                                   BTRFS_BLOCK_GROUP_RAID10))
7503                 factor = 2;
7504         else
7505                 factor = 1;
7506
7507         /* make sure this block group isn't part of an allocation cluster */
7508         cluster = &root->fs_info->data_alloc_cluster;
7509         spin_lock(&cluster->refill_lock);
7510         btrfs_return_cluster_to_free_space(block_group, cluster);
7511         spin_unlock(&cluster->refill_lock);
7512
7513         /*
7514          * make sure this block group isn't part of a metadata
7515          * allocation cluster
7516          */
7517         cluster = &root->fs_info->meta_alloc_cluster;
7518         spin_lock(&cluster->refill_lock);
7519         btrfs_return_cluster_to_free_space(block_group, cluster);
7520         spin_unlock(&cluster->refill_lock);
7521
7522         path = btrfs_alloc_path();
7523         if (!path) {
7524                 ret = -ENOMEM;
7525                 goto out;
7526         }
7527
7528         inode = lookup_free_space_inode(tree_root, block_group, path);
7529         if (!IS_ERR(inode)) {
7530                 ret = btrfs_orphan_add(trans, inode);
7531                 BUG_ON(ret);
7532                 clear_nlink(inode);
7533                 /* One for the block groups ref */
7534                 spin_lock(&block_group->lock);
7535                 if (block_group->iref) {
7536                         block_group->iref = 0;
7537                         block_group->inode = NULL;
7538                         spin_unlock(&block_group->lock);
7539                         iput(inode);
7540                 } else {
7541                         spin_unlock(&block_group->lock);
7542                 }
7543                 /* One for our lookup ref */
7544                 btrfs_add_delayed_iput(inode);
7545         }
7546
7547         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
7548         key.offset = block_group->key.objectid;
7549         key.type = 0;
7550
7551         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
7552         if (ret < 0)
7553                 goto out;
7554         if (ret > 0)
7555                 btrfs_release_path(path);
7556         if (ret == 0) {
7557                 ret = btrfs_del_item(trans, tree_root, path);
7558                 if (ret)
7559                         goto out;
7560                 btrfs_release_path(path);
7561         }
7562
7563         spin_lock(&root->fs_info->block_group_cache_lock);
7564         rb_erase(&block_group->cache_node,
7565                  &root->fs_info->block_group_cache_tree);
7566         spin_unlock(&root->fs_info->block_group_cache_lock);
7567
7568         down_write(&block_group->space_info->groups_sem);
7569         /*
7570          * we must use list_del_init so people can check to see if they
7571          * are still on the list after taking the semaphore
7572          */
7573         list_del_init(&block_group->list);
7574         up_write(&block_group->space_info->groups_sem);
7575
7576         if (block_group->cached == BTRFS_CACHE_STARTED)
7577                 wait_block_group_cache_done(block_group);
7578
7579         btrfs_remove_free_space_cache(block_group);
7580
7581         spin_lock(&block_group->space_info->lock);
7582         block_group->space_info->total_bytes -= block_group->key.offset;
7583         block_group->space_info->bytes_readonly -= block_group->key.offset;
7584         block_group->space_info->disk_total -= block_group->key.offset * factor;
7585         spin_unlock(&block_group->space_info->lock);
7586
7587         memcpy(&key, &block_group->key, sizeof(key));
7588
7589         btrfs_clear_space_info_full(root->fs_info);
7590
7591         btrfs_put_block_group(block_group);
7592         btrfs_put_block_group(block_group);
7593
7594         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
7595         if (ret > 0)
7596                 ret = -EIO;
7597         if (ret < 0)
7598                 goto out;
7599
7600         ret = btrfs_del_item(trans, root, path);
7601 out:
7602         btrfs_free_path(path);
7603         return ret;
7604 }
7605
7606 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
7607 {
7608         struct btrfs_space_info *space_info;
7609         struct btrfs_super_block *disk_super;
7610         u64 features;
7611         u64 flags;
7612         int mixed = 0;
7613         int ret;
7614
7615         disk_super = fs_info->super_copy;
7616         if (!btrfs_super_root(disk_super))
7617                 return 1;
7618
7619         features = btrfs_super_incompat_flags(disk_super);
7620         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
7621                 mixed = 1;
7622
7623         flags = BTRFS_BLOCK_GROUP_SYSTEM;
7624         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7625         if (ret)
7626                 goto out;
7627
7628         if (mixed) {
7629                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
7630                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7631         } else {
7632                 flags = BTRFS_BLOCK_GROUP_METADATA;
7633                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7634                 if (ret)
7635                         goto out;
7636
7637                 flags = BTRFS_BLOCK_GROUP_DATA;
7638                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7639         }
7640 out:
7641         return ret;
7642 }
7643
7644 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
7645 {
7646         return unpin_extent_range(root, start, end);
7647 }
7648
7649 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
7650                                u64 num_bytes, u64 *actual_bytes)
7651 {
7652         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
7653 }
7654
7655 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
7656 {
7657         struct btrfs_fs_info *fs_info = root->fs_info;
7658         struct btrfs_block_group_cache *cache = NULL;
7659         u64 group_trimmed;
7660         u64 start;
7661         u64 end;
7662         u64 trimmed = 0;
7663         int ret = 0;
7664
7665         cache = btrfs_lookup_block_group(fs_info, range->start);
7666
7667         while (cache) {
7668                 if (cache->key.objectid >= (range->start + range->len)) {
7669                         btrfs_put_block_group(cache);
7670                         break;
7671                 }
7672
7673                 start = max(range->start, cache->key.objectid);
7674                 end = min(range->start + range->len,
7675                                 cache->key.objectid + cache->key.offset);
7676
7677                 if (end - start >= range->minlen) {
7678                         if (!block_group_cache_done(cache)) {
7679                                 ret = cache_block_group(cache, NULL, root, 0);
7680                                 if (!ret)
7681                                         wait_block_group_cache_done(cache);
7682                         }
7683                         ret = btrfs_trim_block_group(cache,
7684                                                      &group_trimmed,
7685                                                      start,
7686                                                      end,
7687                                                      range->minlen);
7688
7689                         trimmed += group_trimmed;
7690                         if (ret) {
7691                                 btrfs_put_block_group(cache);
7692                                 break;
7693                         }
7694                 }
7695
7696                 cache = next_block_group(fs_info->tree_root, cache);
7697         }
7698
7699         range->len = trimmed;
7700         return ret;
7701 }