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