Pileus Git - ~andy/linux/blob - 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 "compat.h"
  27 #include "hash.h"
  28 #include "ctree.h"
  29 #include "disk-io.h"
  30 #include "print-tree.h"
  31 #include "transaction.h"
  32 #include "volumes.h"
  33 #include "locking.h"
  34 #include "free-space-cache.h"
  35
  36 /* control flags for do_chunk_alloc's force field
  37  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  38  * if we really need one.
  39  *
  40  * CHUNK_ALLOC_FORCE means it must try to allocate one
  41  *
  42  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  43  * if we have very few chunks already allocated.  This is
  44  * used as part of the clustering code to help make sure
  45  * we have a good pool of storage to cluster in, without
  46  * filling the FS with empty chunks
  47  *
  48  */
  49 enum {
  50         CHUNK_ALLOC_NO_FORCE = 0,
  51         CHUNK_ALLOC_FORCE = 1,
  52         CHUNK_ALLOC_LIMITED = 2,
  53 };
  54
  55 /*
  56  * Control how reservations are dealt with.
  57  *
  58  * RESERVE_FREE - freeing a reservation.
  59  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  60  *   ENOSPC accounting
  61  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  62  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  63  */
  64 enum {
  65         RESERVE_FREE = 0,
  66         RESERVE_ALLOC = 1,
  67         RESERVE_ALLOC_NO_ACCOUNT = 2,
  68 };
  69
  70 static int update_block_group(struct btrfs_trans_handle *trans,
  71                               struct btrfs_root *root,
  72                               u64 bytenr, u64 num_bytes, int alloc);
  73 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  74                                 struct btrfs_root *root,
  75                                 u64 bytenr, u64 num_bytes, u64 parent,
  76                                 u64 root_objectid, u64 owner_objectid,
  77                                 u64 owner_offset, int refs_to_drop,
  78                                 struct btrfs_delayed_extent_op *extra_op);
  79 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  80                                     struct extent_buffer *leaf,
  81                                     struct btrfs_extent_item *ei);
  82 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  83                                       struct btrfs_root *root,
  84                                       u64 parent, u64 root_objectid,
  85                                       u64 flags, u64 owner, u64 offset,
  86                                       struct btrfs_key *ins, int ref_mod);
  87 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  88                                      struct btrfs_root *root,
  89                                      u64 parent, u64 root_objectid,
  90                                      u64 flags, struct btrfs_disk_key *key,
  91                                      int level, struct btrfs_key *ins);
  92 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  93                           struct btrfs_root *extent_root, u64 alloc_bytes,
  94                           u64 flags, int force);
  95 static int find_next_key(struct btrfs_path *path, int level,
  96                          struct btrfs_key *key);
  97 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
  98                             int dump_block_groups);
  99 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 100                                        u64 num_bytes, int reserve);
 101
 102 static noinline int
 103 block_group_cache_done(struct btrfs_block_group_cache *cache)
 104 {
 105         smp_mb();
 106         return cache->cached == BTRFS_CACHE_FINISHED;
 107 }
 108
 109 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 110 {
 111         return (cache->flags & bits) == bits;
 112 }
 113
 114 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 115 {
 116         atomic_inc(&cache->count);
 117 }
 118
 119 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 120 {
 121         if (atomic_dec_and_test(&cache->count)) {
 122                 WARN_ON(cache->pinned > 0);
 123                 WARN_ON(cache->reserved > 0);
 124                 WARN_ON(cache->reserved_pinned > 0);
 125                 kfree(cache->free_space_ctl);
 126                 kfree(cache);
 127         }
 128 }
 129
 130 /*
 131  * this adds the block group to the fs_info rb tree for the block group
 132  * cache
 133  */
 134 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 135                                 struct btrfs_block_group_cache *block_group)
 136 {
 137         struct rb_node **p;
 138         struct rb_node *parent = NULL;
 139         struct btrfs_block_group_cache *cache;
 140
 141         spin_lock(&info->block_group_cache_lock);
 142         p = &info->block_group_cache_tree.rb_node;
 143
 144         while (*p) {
 145                 parent = *p;
 146                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 147                                  cache_node);
 148                 if (block_group->key.objectid < cache->key.objectid) {
 149                         p = &(*p)->rb_left;
 150                 } else if (block_group->key.objectid > cache->key.objectid) {
 151                         p = &(*p)->rb_right;
 152                 } else {
 153                         spin_unlock(&info->block_group_cache_lock);
 154                         return -EEXIST;
 155                 }
 156         }
 157
 158         rb_link_node(&block_group->cache_node, parent, p);
 159         rb_insert_color(&block_group->cache_node,
 160                         &info->block_group_cache_tree);
 161         spin_unlock(&info->block_group_cache_lock);
 162
 163         return 0;
 164 }
 165
 166 /*
 167  * This will return the block group at or after bytenr if contains is 0, else
 168  * it will return the block group that contains the bytenr
 169  */
 170 static struct btrfs_block_group_cache *
 171 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 172                               int contains)
 173 {
 174         struct btrfs_block_group_cache *cache, *ret = NULL;
 175         struct rb_node *n;
 176         u64 end, start;
 177
 178         spin_lock(&info->block_group_cache_lock);
 179         n = info->block_group_cache_tree.rb_node;
 180
 181         while (n) {
 182                 cache = rb_entry(n, struct btrfs_block_group_cache,
 183                                  cache_node);
 184                 end = cache->key.objectid + cache->key.offset - 1;
 185                 start = cache->key.objectid;
 186
 187                 if (bytenr < start) {
 188                         if (!contains && (!ret || start < ret->key.objectid))
 189                                 ret = cache;
 190                         n = n->rb_left;
 191                 } else if (bytenr > start) {
 192                         if (contains && bytenr <= end) {
 193                                 ret = cache;
 194                                 break;
 195                         }
 196                         n = n->rb_right;
 197                 } else {
 198                         ret = cache;
 199                         break;
 200                 }
 201         }
 202         if (ret)
 203                 btrfs_get_block_group(ret);
 204         spin_unlock(&info->block_group_cache_lock);
 205
 206         return ret;
 207 }
 208
 209 static int add_excluded_extent(struct btrfs_root *root,
 210                                u64 start, u64 num_bytes)
 211 {
 212         u64 end = start + num_bytes - 1;
 213         set_extent_bits(&root->fs_info->freed_extents[0],
 214                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 215         set_extent_bits(&root->fs_info->freed_extents[1],
 216                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 217         return 0;
 218 }
 219
 220 static void free_excluded_extents(struct btrfs_root *root,
 221                                   struct btrfs_block_group_cache *cache)
 222 {
 223         u64 start, end;
 224
 225         start = cache->key.objectid;
 226         end = start + cache->key.offset - 1;
 227
 228         clear_extent_bits(&root->fs_info->freed_extents[0],
 229                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 230         clear_extent_bits(&root->fs_info->freed_extents[1],
 231                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 232 }
 233
 234 static int exclude_super_stripes(struct btrfs_root *root,
 235                                  struct btrfs_block_group_cache *cache)
 236 {
 237         u64 bytenr;
 238         u64 *logical;
 239         int stripe_len;
 240         int i, nr, ret;
 241
 242         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 243                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 244                 cache->bytes_super += stripe_len;
 245                 ret = add_excluded_extent(root, cache->key.objectid,
 246                                           stripe_len);
 247                 BUG_ON(ret);
 248         }
 249
 250         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 251                 bytenr = btrfs_sb_offset(i);
 252                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 253                                        cache->key.objectid, bytenr,
 254                                        0, &logical, &nr, &stripe_len);
 255                 BUG_ON(ret);
 256
 257                 while (nr--) {
 258                         cache->bytes_super += stripe_len;
 259                         ret = add_excluded_extent(root, logical[nr],
 260                                                   stripe_len);
 261                         BUG_ON(ret);
 262                 }
 263
 264                 kfree(logical);
 265         }
 266         return 0;
 267 }
 268
 269 static struct btrfs_caching_control *
 270 get_caching_control(struct btrfs_block_group_cache *cache)
 271 {
 272         struct btrfs_caching_control *ctl;
 273
 274         spin_lock(&cache->lock);
 275         if (cache->cached != BTRFS_CACHE_STARTED) {
 276                 spin_unlock(&cache->lock);
 277                 return NULL;
 278         }
 279
 280         /* We're loading it the fast way, so we don't have a caching_ctl. */
 281         if (!cache->caching_ctl) {
 282                 spin_unlock(&cache->lock);
 283                 return NULL;
 284         }
 285
 286         ctl = cache->caching_ctl;
 287         atomic_inc(&ctl->count);
 288         spin_unlock(&cache->lock);
 289         return ctl;
 290 }
 291
 292 static void put_caching_control(struct btrfs_caching_control *ctl)
 293 {
 294         if (atomic_dec_and_test(&ctl->count))
 295                 kfree(ctl);
 296 }
 297
 298 /*
 299  * this is only called by cache_block_group, since we could have freed extents
 300  * we need to check the pinned_extents for any extents that can't be used yet
 301  * since their free space will be released as soon as the transaction commits.
 302  */
 303 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 304                               struct btrfs_fs_info *info, u64 start, u64 end)
 305 {
 306         u64 extent_start, extent_end, size, total_added = 0;
 307         int ret;
 308
 309         while (start < end) {
 310                 ret = find_first_extent_bit(info->pinned_extents, start,
 311                                             &extent_start, &extent_end,
 312                                             EXTENT_DIRTY | EXTENT_UPTODATE);
 313                 if (ret)
 314                         break;
 315
 316                 if (extent_start <= start) {
 317                         start = extent_end + 1;
 318                 } else if (extent_start > start && extent_start < end) {
 319                         size = extent_start - start;
 320                         total_added += size;
 321                         ret = btrfs_add_free_space(block_group, start,
 322                                                    size);
 323                         BUG_ON(ret);
 324                         start = extent_end + 1;
 325                 } else {
 326                         break;
 327                 }
 328         }
 329
 330         if (start < end) {
 331                 size = end - start;
 332                 total_added += size;
 333                 ret = btrfs_add_free_space(block_group, start, size);
 334                 BUG_ON(ret);
 335         }
 336
 337         return total_added;
 338 }
 339
 340 static noinline void caching_thread(struct btrfs_work *work)
 341 {
 342         struct btrfs_block_group_cache *block_group;
 343         struct btrfs_fs_info *fs_info;
 344         struct btrfs_caching_control *caching_ctl;
 345         struct btrfs_root *extent_root;
 346         struct btrfs_path *path;
 347         struct extent_buffer *leaf;
 348         struct btrfs_key key;
 349         u64 total_found = 0;
 350         u64 last = 0;
 351         u32 nritems;
 352         int ret = 0;
 353
 354         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 355         block_group = caching_ctl->block_group;
 356         fs_info = block_group->fs_info;
 357         extent_root = fs_info->extent_root;
 358
 359         path = btrfs_alloc_path();
 360         if (!path)
 361                 goto out;
 362
 363         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 364
 365         /*
 366          * We don't want to deadlock with somebody trying to allocate a new
 367          * extent for the extent root while also trying to search the extent
 368          * root to add free space.  So we skip locking and search the commit
 369          * root, since its read-only
 370          */
 371         path->skip_locking = 1;
 372         path->search_commit_root = 1;
 373         path->reada = 1;
 374
 375         key.objectid = last;
 376         key.offset = 0;
 377         key.type = BTRFS_EXTENT_ITEM_KEY;
 378 again:
 379         mutex_lock(&caching_ctl->mutex);
 380         /* need to make sure the commit_root doesn't disappear */
 381         down_read(&fs_info->extent_commit_sem);
 382
 383         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 384         if (ret < 0)
 385                 goto err;
 386
 387         leaf = path->nodes[0];
 388         nritems = btrfs_header_nritems(leaf);
 389
 390         while (1) {
 391                 if (btrfs_fs_closing(fs_info) > 1) {
 392                         last = (u64)-1;
 393                         break;
 394                 }
 395
 396                 if (path->slots[0] < nritems) {
 397                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 398                 } else {
 399                         ret = find_next_key(path, 0, &key);
 400                         if (ret)
 401                                 break;
 402
 403                         if (need_resched() ||
 404                             btrfs_next_leaf(extent_root, path)) {
 405                                 caching_ctl->progress = last;
 406                                 btrfs_release_path(path);
 407                                 up_read(&fs_info->extent_commit_sem);
 408                                 mutex_unlock(&caching_ctl->mutex);
 409                                 cond_resched();
 410                                 goto again;
 411                         }
 412                         leaf = path->nodes[0];
 413                         nritems = btrfs_header_nritems(leaf);
 414                         continue;
 415                 }
 416
 417                 if (key.objectid < block_group->key.objectid) {
 418                         path->slots[0]++;
 419                         continue;
 420                 }
 421
 422                 if (key.objectid >= block_group->key.objectid +
 423                     block_group->key.offset)
 424                         break;
 425
 426                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 427                         total_found += add_new_free_space(block_group,
 428                                                           fs_info, last,
 429                                                           key.objectid);
 430                         last = key.objectid + key.offset;
 431
 432                         if (total_found > (1024 * 1024 * 2)) {
 433                                 total_found = 0;
 434                                 wake_up(&caching_ctl->wait);
 435                         }
 436                 }
 437                 path->slots[0]++;
 438         }
 439         ret = 0;
 440
 441         total_found += add_new_free_space(block_group, fs_info, last,
 442                                           block_group->key.objectid +
 443                                           block_group->key.offset);
 444         caching_ctl->progress = (u64)-1;
 445
 446         spin_lock(&block_group->lock);
 447         block_group->caching_ctl = NULL;
 448         block_group->cached = BTRFS_CACHE_FINISHED;
 449         spin_unlock(&block_group->lock);
 450
 451 err:
 452         btrfs_free_path(path);
 453         up_read(&fs_info->extent_commit_sem);
 454
 455         free_excluded_extents(extent_root, block_group);
 456
 457         mutex_unlock(&caching_ctl->mutex);
 458 out:
 459         wake_up(&caching_ctl->wait);
 460
 461         put_caching_control(caching_ctl);
 462         btrfs_put_block_group(block_group);
 463 }
 464
 465 static int cache_block_group(struct btrfs_block_group_cache *cache,
 466                              struct btrfs_trans_handle *trans,
 467                              struct btrfs_root *root,
 468                              int load_cache_only)
 469 {
 470         struct btrfs_fs_info *fs_info = cache->fs_info;
 471         struct btrfs_caching_control *caching_ctl;
 472         int ret = 0;
 473
 474         smp_mb();
 475         if (cache->cached != BTRFS_CACHE_NO)
 476                 return 0;
 477
 478         /*
 479          * We can't do the read from on-disk cache during a commit since we need
 480          * to have the normal tree locking.  Also if we are currently trying to
 481          * allocate blocks for the tree root we can't do the fast caching since
 482          * we likely hold important locks.
 483          */
 484         if (trans && (!trans->transaction->in_commit) &&
 485             (root && root != root->fs_info->tree_root)) {
 486                 spin_lock(&cache->lock);
 487                 if (cache->cached != BTRFS_CACHE_NO) {
 488                         spin_unlock(&cache->lock);
 489                         return 0;
 490                 }
 491                 cache->cached = BTRFS_CACHE_STARTED;
 492                 spin_unlock(&cache->lock);
 493
 494                 ret = load_free_space_cache(fs_info, cache);
 495
 496                 spin_lock(&cache->lock);
 497                 if (ret == 1) {
 498                         cache->cached = BTRFS_CACHE_FINISHED;
 499                         cache->last_byte_to_unpin = (u64)-1;
 500                 } else {
 501                         cache->cached = BTRFS_CACHE_NO;
 502                 }
 503                 spin_unlock(&cache->lock);
 504                 if (ret == 1) {
 505                         free_excluded_extents(fs_info->extent_root, cache);
 506                         return 0;
 507                 }
 508         }
 509
 510         if (load_cache_only)
 511                 return 0;
 512
 513         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 514         BUG_ON(!caching_ctl);
 515
 516         INIT_LIST_HEAD(&caching_ctl->list);
 517         mutex_init(&caching_ctl->mutex);
 518         init_waitqueue_head(&caching_ctl->wait);
 519         caching_ctl->block_group = cache;
 520         caching_ctl->progress = cache->key.objectid;
 521         /* one for caching kthread, one for caching block group list */
 522         atomic_set(&caching_ctl->count, 2);
 523         caching_ctl->work.func = caching_thread;
 524
 525         spin_lock(&cache->lock);
 526         if (cache->cached != BTRFS_CACHE_NO) {
 527                 spin_unlock(&cache->lock);
 528                 kfree(caching_ctl);
 529                 return 0;
 530         }
 531         cache->caching_ctl = caching_ctl;
 532         cache->cached = BTRFS_CACHE_STARTED;
 533         spin_unlock(&cache->lock);
 534
 535         down_write(&fs_info->extent_commit_sem);
 536         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 537         up_write(&fs_info->extent_commit_sem);
 538
 539         btrfs_get_block_group(cache);
 540
 541         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 542
 543         return ret;
 544 }
 545
 546 /*
 547  * return the block group that starts at or after bytenr
 548  */
 549 static struct btrfs_block_group_cache *
 550 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 551 {
 552         struct btrfs_block_group_cache *cache;
 553
 554         cache = block_group_cache_tree_search(info, bytenr, 0);
 555
 556         return cache;
 557 }
 558
 559 /*
 560  * return the block group that contains the given bytenr
 561  */
 562 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 563                                                  struct btrfs_fs_info *info,
 564                                                  u64 bytenr)
 565 {
 566         struct btrfs_block_group_cache *cache;
 567
 568         cache = block_group_cache_tree_search(info, bytenr, 1);
 569
 570         return cache;
 571 }
 572
 573 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 574                                                   u64 flags)
 575 {
 576         struct list_head *head = &info->space_info;
 577         struct btrfs_space_info *found;
 578
 579         flags &= BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_SYSTEM |
 580                  BTRFS_BLOCK_GROUP_METADATA;
 581
 582         rcu_read_lock();
 583         list_for_each_entry_rcu(found, head, list) {
 584                 if (found->flags & flags) {
 585                         rcu_read_unlock();
 586                         return found;
 587                 }
 588         }
 589         rcu_read_unlock();
 590         return NULL;
 591 }
 592
 593 /*
 594  * after adding space to the filesystem, we need to clear the full flags
 595  * on all the space infos.
 596  */
 597 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 598 {
 599         struct list_head *head = &info->space_info;
 600         struct btrfs_space_info *found;
 601
 602         rcu_read_lock();
 603         list_for_each_entry_rcu(found, head, list)
 604                 found->full = 0;
 605         rcu_read_unlock();
 606 }
 607
 608 static u64 div_factor(u64 num, int factor)
 609 {
 610         if (factor == 10)
 611                 return num;
 612         num *= factor;
 613         do_div(num, 10);
 614         return num;
 615 }
 616
 617 static u64 div_factor_fine(u64 num, int factor)
 618 {
 619         if (factor == 100)
 620                 return num;
 621         num *= factor;
 622         do_div(num, 100);
 623         return num;
 624 }
 625
 626 u64 btrfs_find_block_group(struct btrfs_root *root,
 627                            u64 search_start, u64 search_hint, int owner)
 628 {
 629         struct btrfs_block_group_cache *cache;
 630         u64 used;
 631         u64 last = max(search_hint, search_start);
 632         u64 group_start = 0;
 633         int full_search = 0;
 634         int factor = 9;
 635         int wrapped = 0;
 636 again:
 637         while (1) {
 638                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
 639                 if (!cache)
 640                         break;
 641
 642                 spin_lock(&cache->lock);
 643                 last = cache->key.objectid + cache->key.offset;
 644                 used = btrfs_block_group_used(&cache->item);
 645
 646                 if ((full_search || !cache->ro) &&
 647                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
 648                         if (used + cache->pinned + cache->reserved <
 649                             div_factor(cache->key.offset, factor)) {
 650                                 group_start = cache->key.objectid;
 651                                 spin_unlock(&cache->lock);
 652                                 btrfs_put_block_group(cache);
 653                                 goto found;
 654                         }
 655                 }
 656                 spin_unlock(&cache->lock);
 657                 btrfs_put_block_group(cache);
 658                 cond_resched();
 659         }
 660         if (!wrapped) {
 661                 last = search_start;
 662                 wrapped = 1;
 663                 goto again;
 664         }
 665         if (!full_search && factor < 10) {
 666                 last = search_start;
 667                 full_search = 1;
 668                 factor = 10;
 669                 goto again;
 670         }
 671 found:
 672         return group_start;
 673 }
 674
 675 /* simple helper to search for an existing extent at a given offset */
 676 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 677 {
 678         int ret;
 679         struct btrfs_key key;
 680         struct btrfs_path *path;
 681
 682         path = btrfs_alloc_path();
 683         if (!path)
 684                 return -ENOMEM;
 685
 686         key.objectid = start;
 687         key.offset = len;
 688         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 689         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 690                                 0, 0);
 691         btrfs_free_path(path);
 692         return ret;
 693 }
 694
 695 /*
 696  * helper function to lookup reference count and flags of extent.
 697  *
 698  * the head node for delayed ref is used to store the sum of all the
 699  * reference count modifications queued up in the rbtree. the head
 700  * node may also store the extent flags to set. This way you can check
 701  * to see what the reference count and extent flags would be if all of
 702  * the delayed refs are not processed.
 703  */
 704 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 705                              struct btrfs_root *root, u64 bytenr,
 706                              u64 num_bytes, u64 *refs, u64 *flags)
 707 {
 708         struct btrfs_delayed_ref_head *head;
 709         struct btrfs_delayed_ref_root *delayed_refs;
 710         struct btrfs_path *path;
 711         struct btrfs_extent_item *ei;
 712         struct extent_buffer *leaf;
 713         struct btrfs_key key;
 714         u32 item_size;
 715         u64 num_refs;
 716         u64 extent_flags;
 717         int ret;
 718
 719         path = btrfs_alloc_path();
 720         if (!path)
 721                 return -ENOMEM;
 722
 723         key.objectid = bytenr;
 724         key.type = BTRFS_EXTENT_ITEM_KEY;
 725         key.offset = num_bytes;
 726         if (!trans) {
 727                 path->skip_locking = 1;
 728                 path->search_commit_root = 1;
 729         }
 730 again:
 731         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 732                                 &key, path, 0, 0);
 733         if (ret < 0)
 734                 goto out_free;
 735
 736         if (ret == 0) {
 737                 leaf = path->nodes[0];
 738                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 739                 if (item_size >= sizeof(*ei)) {
 740                         ei = btrfs_item_ptr(leaf, path->slots[0],
 741                                             struct btrfs_extent_item);
 742                         num_refs = btrfs_extent_refs(leaf, ei);
 743                         extent_flags = btrfs_extent_flags(leaf, ei);
 744                 } else {
 745 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 746                         struct btrfs_extent_item_v0 *ei0;
 747                         BUG_ON(item_size != sizeof(*ei0));
 748                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 749                                              struct btrfs_extent_item_v0);
 750                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 751                         /* FIXME: this isn't correct for data */
 752                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 753 #else
 754                         BUG();
 755 #endif
 756                 }
 757                 BUG_ON(num_refs == 0);
 758         } else {
 759                 num_refs = 0;
 760                 extent_flags = 0;
 761                 ret = 0;
 762         }
 763
 764         if (!trans)
 765                 goto out;
 766
 767         delayed_refs = &trans->transaction->delayed_refs;
 768         spin_lock(&delayed_refs->lock);
 769         head = btrfs_find_delayed_ref_head(trans, bytenr);
 770         if (head) {
 771                 if (!mutex_trylock(&head->mutex)) {
 772                         atomic_inc(&head->node.refs);
 773                         spin_unlock(&delayed_refs->lock);
 774
 775                         btrfs_release_path(path);
 776
 777                         /*
 778                          * Mutex was contended, block until it's released and try
 779                          * again
 780                          */
 781                         mutex_lock(&head->mutex);
 782                         mutex_unlock(&head->mutex);
 783                         btrfs_put_delayed_ref(&head->node);
 784                         goto again;
 785                 }
 786                 if (head->extent_op && head->extent_op->update_flags)
 787                         extent_flags |= head->extent_op->flags_to_set;
 788                 else
 789                         BUG_ON(num_refs == 0);
 790
 791                 num_refs += head->node.ref_mod;
 792                 mutex_unlock(&head->mutex);
 793         }
 794         spin_unlock(&delayed_refs->lock);
 795 out:
 796         WARN_ON(num_refs == 0);
 797         if (refs)
 798                 *refs = num_refs;
 799         if (flags)
 800                 *flags = extent_flags;
 801 out_free:
 802         btrfs_free_path(path);
 803         return ret;
 804 }
 805
 806 /*
 807  * Back reference rules.  Back refs have three main goals:
 808  *
 809  * 1) differentiate between all holders of references to an extent so that
 810  *    when a reference is dropped we can make sure it was a valid reference
 811  *    before freeing the extent.
 812  *
 813  * 2) Provide enough information to quickly find the holders of an extent
 814  *    if we notice a given block is corrupted or bad.
 815  *
 816  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 817  *    maintenance.  This is actually the same as #2, but with a slightly
 818  *    different use case.
 819  *
 820  * There are two kinds of back refs. The implicit back refs is optimized
 821  * for pointers in non-shared tree blocks. For a given pointer in a block,
 822  * back refs of this kind provide information about the block's owner tree
 823  * and the pointer's key. These information allow us to find the block by
 824  * b-tree searching. The full back refs is for pointers in tree blocks not
 825  * referenced by their owner trees. The location of tree block is recorded
 826  * in the back refs. Actually the full back refs is generic, and can be
 827  * used in all cases the implicit back refs is used. The major shortcoming
 828  * of the full back refs is its overhead. Every time a tree block gets
 829  * COWed, we have to update back refs entry for all pointers in it.
 830  *
 831  * For a newly allocated tree block, we use implicit back refs for
 832  * pointers in it. This means most tree related operations only involve
 833  * implicit back refs. For a tree block created in old transaction, the
 834  * only way to drop a reference to it is COW it. So we can detect the
 835  * event that tree block loses its owner tree's reference and do the
 836  * back refs conversion.
 837  *
 838  * When a tree block is COW'd through a tree, there are four cases:
 839  *
 840  * The reference count of the block is one and the tree is the block's
 841  * owner tree. Nothing to do in this case.
 842  *
 843  * The reference count of the block is one and the tree is not the
 844  * block's owner tree. In this case, full back refs is used for pointers
 845  * in the block. Remove these full back refs, add implicit back refs for
 846  * every pointers in the new block.
 847  *
 848  * The reference count of the block is greater than one and the tree is
 849  * the block's owner tree. In this case, implicit back refs is used for
 850  * pointers in the block. Add full back refs for every pointers in the
 851  * block, increase lower level extents' reference counts. The original
 852  * implicit back refs are entailed to the new block.
 853  *
 854  * The reference count of the block is greater than one and the tree is
 855  * not the block's owner tree. Add implicit back refs for every pointer in
 856  * the new block, increase lower level extents' reference count.
 857  *
 858  * Back Reference Key composing:
 859  *
 860  * The key objectid corresponds to the first byte in the extent,
 861  * The key type is used to differentiate between types of back refs.
 862  * There are different meanings of the key offset for different types
 863  * of back refs.
 864  *
 865  * File extents can be referenced by:
 866  *
 867  * - multiple snapshots, subvolumes, or different generations in one subvol
 868  * - different files inside a single subvolume
 869  * - different offsets inside a file (bookend extents in file.c)
 870  *
 871  * The extent ref structure for the implicit back refs has fields for:
 872  *
 873  * - Objectid of the subvolume root
 874  * - objectid of the file holding the reference
 875  * - original offset in the file
 876  * - how many bookend extents
 877  *
 878  * The key offset for the implicit back refs is hash of the first
 879  * three fields.
 880  *
 881  * The extent ref structure for the full back refs has field for:
 882  *
 883  * - number of pointers in the tree leaf
 884  *
 885  * The key offset for the implicit back refs is the first byte of
 886  * the tree leaf
 887  *
 888  * When a file extent is allocated, The implicit back refs is used.
 889  * the fields are filled in:
 890  *
 891  *     (root_key.objectid, inode objectid, offset in file, 1)
 892  *
 893  * When a file extent is removed file truncation, we find the
 894  * corresponding implicit back refs and check the following fields:
 895  *
 896  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 897  *
 898  * Btree extents can be referenced by:
 899  *
 900  * - Different subvolumes
 901  *
 902  * Both the implicit back refs and the full back refs for tree blocks
 903  * only consist of key. The key offset for the implicit back refs is
 904  * objectid of block's owner tree. The key offset for the full back refs
 905  * is the first byte of parent block.
 906  *
 907  * When implicit back refs is used, information about the lowest key and
 908  * level of the tree block are required. These information are stored in
 909  * tree block info structure.
 910  */
 911
 912 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 913 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 914                                   struct btrfs_root *root,
 915                                   struct btrfs_path *path,
 916                                   u64 owner, u32 extra_size)
 917 {
 918         struct btrfs_extent_item *item;
 919         struct btrfs_extent_item_v0 *ei0;
 920         struct btrfs_extent_ref_v0 *ref0;
 921         struct btrfs_tree_block_info *bi;
 922         struct extent_buffer *leaf;
 923         struct btrfs_key key;
 924         struct btrfs_key found_key;
 925         u32 new_size = sizeof(*item);
 926         u64 refs;
 927         int ret;
 928
 929         leaf = path->nodes[0];
 930         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 931
 932         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 933         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 934                              struct btrfs_extent_item_v0);
 935         refs = btrfs_extent_refs_v0(leaf, ei0);
 936
 937         if (owner == (u64)-1) {
 938                 while (1) {
 939                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 940                                 ret = btrfs_next_leaf(root, path);
 941                                 if (ret < 0)
 942                                         return ret;
 943                                 BUG_ON(ret > 0);
 944                                 leaf = path->nodes[0];
 945                         }
 946                         btrfs_item_key_to_cpu(leaf, &found_key,
 947                                               path->slots[0]);
 948                         BUG_ON(key.objectid != found_key.objectid);
 949                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 950                                 path->slots[0]++;
 951                                 continue;
 952                         }
 953                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 954                                               struct btrfs_extent_ref_v0);
 955                         owner = btrfs_ref_objectid_v0(leaf, ref0);
 956                         break;
 957                 }
 958         }
 959         btrfs_release_path(path);
 960
 961         if (owner < BTRFS_FIRST_FREE_OBJECTID)
 962                 new_size += sizeof(*bi);
 963
 964         new_size -= sizeof(*ei0);
 965         ret = btrfs_search_slot(trans, root, &key, path,
 966                                 new_size + extra_size, 1);
 967         if (ret < 0)
 968                 return ret;
 969         BUG_ON(ret);
 970
 971         ret = btrfs_extend_item(trans, root, path, new_size);
 972
 973         leaf = path->nodes[0];
 974         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 975         btrfs_set_extent_refs(leaf, item, refs);
 976         /* FIXME: get real generation */
 977         btrfs_set_extent_generation(leaf, item, 0);
 978         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 979                 btrfs_set_extent_flags(leaf, item,
 980                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
 981                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
 982                 bi = (struct btrfs_tree_block_info *)(item + 1);
 983                 /* FIXME: get first key of the block */
 984                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
 985                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
 986         } else {
 987                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
 988         }
 989         btrfs_mark_buffer_dirty(leaf);
 990         return 0;
 991 }
 992 #endif
 993
 994 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
 995 {
 996         u32 high_crc = ~(u32)0;
 997         u32 low_crc = ~(u32)0;
 998         __le64 lenum;
 999
1000         lenum = cpu_to_le64(root_objectid);
1001         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1002         lenum = cpu_to_le64(owner);
1003         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1004         lenum = cpu_to_le64(offset);
1005         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1006
1007         return ((u64)high_crc << 31) ^ (u64)low_crc;
1008 }
1009
1010 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1011                                      struct btrfs_extent_data_ref *ref)
1012 {
1013         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1014                                     btrfs_extent_data_ref_objectid(leaf, ref),
1015                                     btrfs_extent_data_ref_offset(leaf, ref));
1016 }
1017
1018 static int match_extent_data_ref(struct extent_buffer *leaf,
1019                                  struct btrfs_extent_data_ref *ref,
1020                                  u64 root_objectid, u64 owner, u64 offset)
1021 {
1022         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1023             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1024             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1025                 return 0;
1026         return 1;
1027 }
1028
1029 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1030                                            struct btrfs_root *root,
1031                                            struct btrfs_path *path,
1032                                            u64 bytenr, u64 parent,
1033                                            u64 root_objectid,
1034                                            u64 owner, u64 offset)
1035 {
1036         struct btrfs_key key;
1037         struct btrfs_extent_data_ref *ref;
1038         struct extent_buffer *leaf;
1039         u32 nritems;
1040         int ret;
1041         int recow;
1042         int err = -ENOENT;
1043
1044         key.objectid = bytenr;
1045         if (parent) {
1046                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1047                 key.offset = parent;
1048         } else {
1049                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1050                 key.offset = hash_extent_data_ref(root_objectid,
1051                                                   owner, offset);
1052         }
1053 again:
1054         recow = 0;
1055         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1056         if (ret < 0) {
1057                 err = ret;
1058                 goto fail;
1059         }
1060
1061         if (parent) {
1062                 if (!ret)
1063                         return 0;
1064 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1065                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1066                 btrfs_release_path(path);
1067                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1068                 if (ret < 0) {
1069                         err = ret;
1070                         goto fail;
1071                 }
1072                 if (!ret)
1073                         return 0;
1074 #endif
1075                 goto fail;
1076         }
1077
1078         leaf = path->nodes[0];
1079         nritems = btrfs_header_nritems(leaf);
1080         while (1) {
1081                 if (path->slots[0] >= nritems) {
1082                         ret = btrfs_next_leaf(root, path);
1083                         if (ret < 0)
1084                                 err = ret;
1085                         if (ret)
1086                                 goto fail;
1087
1088                         leaf = path->nodes[0];
1089                         nritems = btrfs_header_nritems(leaf);
1090                         recow = 1;
1091                 }
1092
1093                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1094                 if (key.objectid != bytenr ||
1095                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1096                         goto fail;
1097
1098                 ref = btrfs_item_ptr(leaf, path->slots[0],
1099                                      struct btrfs_extent_data_ref);
1100
1101                 if (match_extent_data_ref(leaf, ref, root_objectid,
1102                                           owner, offset)) {
1103                         if (recow) {
1104                                 btrfs_release_path(path);
1105                                 goto again;
1106                         }
1107                         err = 0;
1108                         break;
1109                 }
1110                 path->slots[0]++;
1111         }
1112 fail:
1113         return err;
1114 }
1115
1116 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1117                                            struct btrfs_root *root,
1118                                            struct btrfs_path *path,
1119                                            u64 bytenr, u64 parent,
1120                                            u64 root_objectid, u64 owner,
1121                                            u64 offset, int refs_to_add)
1122 {
1123         struct btrfs_key key;
1124         struct extent_buffer *leaf;
1125         u32 size;
1126         u32 num_refs;
1127         int ret;
1128
1129         key.objectid = bytenr;
1130         if (parent) {
1131                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1132                 key.offset = parent;
1133                 size = sizeof(struct btrfs_shared_data_ref);
1134         } else {
1135                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1136                 key.offset = hash_extent_data_ref(root_objectid,
1137                                                   owner, offset);
1138                 size = sizeof(struct btrfs_extent_data_ref);
1139         }
1140
1141         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1142         if (ret && ret != -EEXIST)
1143                 goto fail;
1144
1145         leaf = path->nodes[0];
1146         if (parent) {
1147                 struct btrfs_shared_data_ref *ref;
1148                 ref = btrfs_item_ptr(leaf, path->slots[0],
1149                                      struct btrfs_shared_data_ref);
1150                 if (ret == 0) {
1151                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1152                 } else {
1153                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1154                         num_refs += refs_to_add;
1155                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1156                 }
1157         } else {
1158                 struct btrfs_extent_data_ref *ref;
1159                 while (ret == -EEXIST) {
1160                         ref = btrfs_item_ptr(leaf, path->slots[0],
1161                                              struct btrfs_extent_data_ref);
1162                         if (match_extent_data_ref(leaf, ref, root_objectid,
1163                                                   owner, offset))
1164                                 break;
1165                         btrfs_release_path(path);
1166                         key.offset++;
1167                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1168                                                       size);
1169                         if (ret && ret != -EEXIST)
1170                                 goto fail;
1171
1172                         leaf = path->nodes[0];
1173                 }
1174                 ref = btrfs_item_ptr(leaf, path->slots[0],
1175                                      struct btrfs_extent_data_ref);
1176                 if (ret == 0) {
1177                         btrfs_set_extent_data_ref_root(leaf, ref,
1178                                                        root_objectid);
1179                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1180                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1181                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1182                 } else {
1183                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1184                         num_refs += refs_to_add;
1185                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1186                 }
1187         }
1188         btrfs_mark_buffer_dirty(leaf);
1189         ret = 0;
1190 fail:
1191         btrfs_release_path(path);
1192         return ret;
1193 }
1194
1195 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1196                                            struct btrfs_root *root,
1197                                            struct btrfs_path *path,
1198                                            int refs_to_drop)
1199 {
1200         struct btrfs_key key;
1201         struct btrfs_extent_data_ref *ref1 = NULL;
1202         struct btrfs_shared_data_ref *ref2 = NULL;
1203         struct extent_buffer *leaf;
1204         u32 num_refs = 0;
1205         int ret = 0;
1206
1207         leaf = path->nodes[0];
1208         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1209
1210         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1211                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1212                                       struct btrfs_extent_data_ref);
1213                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1214         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1215                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1216                                       struct btrfs_shared_data_ref);
1217                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1218 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1219         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1220                 struct btrfs_extent_ref_v0 *ref0;
1221                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1222                                       struct btrfs_extent_ref_v0);
1223                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1224 #endif
1225         } else {
1226                 BUG();
1227         }
1228
1229         BUG_ON(num_refs < refs_to_drop);
1230         num_refs -= refs_to_drop;
1231
1232         if (num_refs == 0) {
1233                 ret = btrfs_del_item(trans, root, path);
1234         } else {
1235                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1236                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1237                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1238                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1239 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1240                 else {
1241                         struct btrfs_extent_ref_v0 *ref0;
1242                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1243                                         struct btrfs_extent_ref_v0);
1244                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1245                 }
1246 #endif
1247                 btrfs_mark_buffer_dirty(leaf);
1248         }
1249         return ret;
1250 }
1251
1252 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1253                                           struct btrfs_path *path,
1254                                           struct btrfs_extent_inline_ref *iref)
1255 {
1256         struct btrfs_key key;
1257         struct extent_buffer *leaf;
1258         struct btrfs_extent_data_ref *ref1;
1259         struct btrfs_shared_data_ref *ref2;
1260         u32 num_refs = 0;
1261
1262         leaf = path->nodes[0];
1263         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1264         if (iref) {
1265                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1266                     BTRFS_EXTENT_DATA_REF_KEY) {
1267                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1268                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1269                 } else {
1270                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1271                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1272                 }
1273         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1274                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1275                                       struct btrfs_extent_data_ref);
1276                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1277         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1278                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1279                                       struct btrfs_shared_data_ref);
1280                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1281 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1282         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1283                 struct btrfs_extent_ref_v0 *ref0;
1284                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1285                                       struct btrfs_extent_ref_v0);
1286                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1287 #endif
1288         } else {
1289                 WARN_ON(1);
1290         }
1291         return num_refs;
1292 }
1293
1294 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1295                                           struct btrfs_root *root,
1296                                           struct btrfs_path *path,
1297                                           u64 bytenr, u64 parent,
1298                                           u64 root_objectid)
1299 {
1300         struct btrfs_key key;
1301         int ret;
1302
1303         key.objectid = bytenr;
1304         if (parent) {
1305                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1306                 key.offset = parent;
1307         } else {
1308                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1309                 key.offset = root_objectid;
1310         }
1311
1312         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1313         if (ret > 0)
1314                 ret = -ENOENT;
1315 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1316         if (ret == -ENOENT && parent) {
1317                 btrfs_release_path(path);
1318                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1319                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1320                 if (ret > 0)
1321                         ret = -ENOENT;
1322         }
1323 #endif
1324         return ret;
1325 }
1326
1327 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1328                                           struct btrfs_root *root,
1329                                           struct btrfs_path *path,
1330                                           u64 bytenr, u64 parent,
1331                                           u64 root_objectid)
1332 {
1333         struct btrfs_key key;
1334         int ret;
1335
1336         key.objectid = bytenr;
1337         if (parent) {
1338                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1339                 key.offset = parent;
1340         } else {
1341                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1342                 key.offset = root_objectid;
1343         }
1344
1345         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1346         btrfs_release_path(path);
1347         return ret;
1348 }
1349
1350 static inline int extent_ref_type(u64 parent, u64 owner)
1351 {
1352         int type;
1353         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1354                 if (parent > 0)
1355                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1356                 else
1357                         type = BTRFS_TREE_BLOCK_REF_KEY;
1358         } else {
1359                 if (parent > 0)
1360                         type = BTRFS_SHARED_DATA_REF_KEY;
1361                 else
1362                         type = BTRFS_EXTENT_DATA_REF_KEY;
1363         }
1364         return type;
1365 }
1366
1367 static int find_next_key(struct btrfs_path *path, int level,
1368                          struct btrfs_key *key)
1369
1370 {
1371         for (; level < BTRFS_MAX_LEVEL; level++) {
1372                 if (!path->nodes[level])
1373                         break;
1374                 if (path->slots[level] + 1 >=
1375                     btrfs_header_nritems(path->nodes[level]))
1376                         continue;
1377                 if (level == 0)
1378                         btrfs_item_key_to_cpu(path->nodes[level], key,
1379                                               path->slots[level] + 1);
1380                 else
1381                         btrfs_node_key_to_cpu(path->nodes[level], key,
1382                                               path->slots[level] + 1);
1383                 return 0;
1384         }
1385         return 1;
1386 }
1387
1388 /*
1389  * look for inline back ref. if back ref is found, *ref_ret is set
1390  * to the address of inline back ref, and 0 is returned.
1391  *
1392  * if back ref isn't found, *ref_ret is set to the address where it
1393  * should be inserted, and -ENOENT is returned.
1394  *
1395  * if insert is true and there are too many inline back refs, the path
1396  * points to the extent item, and -EAGAIN is returned.
1397  *
1398  * NOTE: inline back refs are ordered in the same way that back ref
1399  *       items in the tree are ordered.
1400  */
1401 static noinline_for_stack
1402 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1403                                  struct btrfs_root *root,
1404                                  struct btrfs_path *path,
1405                                  struct btrfs_extent_inline_ref **ref_ret,
1406                                  u64 bytenr, u64 num_bytes,
1407                                  u64 parent, u64 root_objectid,
1408                                  u64 owner, u64 offset, int insert)
1409 {
1410         struct btrfs_key key;
1411         struct extent_buffer *leaf;
1412         struct btrfs_extent_item *ei;
1413         struct btrfs_extent_inline_ref *iref;
1414         u64 flags;
1415         u64 item_size;
1416         unsigned long ptr;
1417         unsigned long end;
1418         int extra_size;
1419         int type;
1420         int want;
1421         int ret;
1422         int err = 0;
1423
1424         key.objectid = bytenr;
1425         key.type = BTRFS_EXTENT_ITEM_KEY;
1426         key.offset = num_bytes;
1427
1428         want = extent_ref_type(parent, owner);
1429         if (insert) {
1430                 extra_size = btrfs_extent_inline_ref_size(want);
1431                 path->keep_locks = 1;
1432         } else
1433                 extra_size = -1;
1434         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1435         if (ret < 0) {
1436                 err = ret;
1437                 goto out;
1438         }
1439         BUG_ON(ret);
1440
1441         leaf = path->nodes[0];
1442         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1443 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1444         if (item_size < sizeof(*ei)) {
1445                 if (!insert) {
1446                         err = -ENOENT;
1447                         goto out;
1448                 }
1449                 ret = convert_extent_item_v0(trans, root, path, owner,
1450                                              extra_size);
1451                 if (ret < 0) {
1452                         err = ret;
1453                         goto out;
1454                 }
1455                 leaf = path->nodes[0];
1456                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1457         }
1458 #endif
1459         BUG_ON(item_size < sizeof(*ei));
1460
1461         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1462         flags = btrfs_extent_flags(leaf, ei);
1463
1464         ptr = (unsigned long)(ei + 1);
1465         end = (unsigned long)ei + item_size;
1466
1467         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1468                 ptr += sizeof(struct btrfs_tree_block_info);
1469                 BUG_ON(ptr > end);
1470         } else {
1471                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1472         }
1473
1474         err = -ENOENT;
1475         while (1) {
1476                 if (ptr >= end) {
1477                         WARN_ON(ptr > end);
1478                         break;
1479                 }
1480                 iref = (struct btrfs_extent_inline_ref *)ptr;
1481                 type = btrfs_extent_inline_ref_type(leaf, iref);
1482                 if (want < type)
1483                         break;
1484                 if (want > type) {
1485                         ptr += btrfs_extent_inline_ref_size(type);
1486                         continue;
1487                 }
1488
1489                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1490                         struct btrfs_extent_data_ref *dref;
1491                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1492                         if (match_extent_data_ref(leaf, dref, root_objectid,
1493                                                   owner, offset)) {
1494                                 err = 0;
1495                                 break;
1496                         }
1497                         if (hash_extent_data_ref_item(leaf, dref) <
1498                             hash_extent_data_ref(root_objectid, owner, offset))
1499                                 break;
1500                 } else {
1501                         u64 ref_offset;
1502                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1503                         if (parent > 0) {
1504                                 if (parent == ref_offset) {
1505                                         err = 0;
1506                                         break;
1507                                 }
1508                                 if (ref_offset < parent)
1509                                         break;
1510                         } else {
1511                                 if (root_objectid == ref_offset) {
1512                                         err = 0;
1513                                         break;
1514                                 }
1515                                 if (ref_offset < root_objectid)
1516                                         break;
1517                         }
1518                 }
1519                 ptr += btrfs_extent_inline_ref_size(type);
1520         }
1521         if (err == -ENOENT && insert) {
1522                 if (item_size + extra_size >=
1523                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1524                         err = -EAGAIN;
1525                         goto out;
1526                 }
1527                 /*
1528                  * To add new inline back ref, we have to make sure
1529                  * there is no corresponding back ref item.
1530                  * For simplicity, we just do not add new inline back
1531                  * ref if there is any kind of item for this block
1532                  */
1533                 if (find_next_key(path, 0, &key) == 0 &&
1534                     key.objectid == bytenr &&
1535                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1536                         err = -EAGAIN;
1537                         goto out;
1538                 }
1539         }
1540         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1541 out:
1542         if (insert) {
1543                 path->keep_locks = 0;
1544                 btrfs_unlock_up_safe(path, 1);
1545         }
1546         return err;
1547 }
1548
1549 /*
1550  * helper to add new inline back ref
1551  */
1552 static noinline_for_stack
1553 int setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1554                                 struct btrfs_root *root,
1555                                 struct btrfs_path *path,
1556                                 struct btrfs_extent_inline_ref *iref,
1557                                 u64 parent, u64 root_objectid,
1558                                 u64 owner, u64 offset, int refs_to_add,
1559                                 struct btrfs_delayed_extent_op *extent_op)
1560 {
1561         struct extent_buffer *leaf;
1562         struct btrfs_extent_item *ei;
1563         unsigned long ptr;
1564         unsigned long end;
1565         unsigned long item_offset;
1566         u64 refs;
1567         int size;
1568         int type;
1569         int ret;
1570
1571         leaf = path->nodes[0];
1572         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1573         item_offset = (unsigned long)iref - (unsigned long)ei;
1574
1575         type = extent_ref_type(parent, owner);
1576         size = btrfs_extent_inline_ref_size(type);
1577
1578         ret = btrfs_extend_item(trans, root, path, size);
1579
1580         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1581         refs = btrfs_extent_refs(leaf, ei);
1582         refs += refs_to_add;
1583         btrfs_set_extent_refs(leaf, ei, refs);
1584         if (extent_op)
1585                 __run_delayed_extent_op(extent_op, leaf, ei);
1586
1587         ptr = (unsigned long)ei + item_offset;
1588         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1589         if (ptr < end - size)
1590                 memmove_extent_buffer(leaf, ptr + size, ptr,
1591                                       end - size - ptr);
1592
1593         iref = (struct btrfs_extent_inline_ref *)ptr;
1594         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1595         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1596                 struct btrfs_extent_data_ref *dref;
1597                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1598                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1599                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1600                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1601                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1602         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1603                 struct btrfs_shared_data_ref *sref;
1604                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1605                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1606                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1607         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1608                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1609         } else {
1610                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1611         }
1612         btrfs_mark_buffer_dirty(leaf);
1613         return 0;
1614 }
1615
1616 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1617                                  struct btrfs_root *root,
1618                                  struct btrfs_path *path,
1619                                  struct btrfs_extent_inline_ref **ref_ret,
1620                                  u64 bytenr, u64 num_bytes, u64 parent,
1621                                  u64 root_objectid, u64 owner, u64 offset)
1622 {
1623         int ret;
1624
1625         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1626                                            bytenr, num_bytes, parent,
1627                                            root_objectid, owner, offset, 0);
1628         if (ret != -ENOENT)
1629                 return ret;
1630
1631         btrfs_release_path(path);
1632         *ref_ret = NULL;
1633
1634         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1635                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1636                                             root_objectid);
1637         } else {
1638                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1639                                              root_objectid, owner, offset);
1640         }
1641         return ret;
1642 }
1643
1644 /*
1645  * helper to update/remove inline back ref
1646  */
1647 static noinline_for_stack
1648 int update_inline_extent_backref(struct btrfs_trans_handle *trans,
1649                                  struct btrfs_root *root,
1650                                  struct btrfs_path *path,
1651                                  struct btrfs_extent_inline_ref *iref,
1652                                  int refs_to_mod,
1653                                  struct btrfs_delayed_extent_op *extent_op)
1654 {
1655         struct extent_buffer *leaf;
1656         struct btrfs_extent_item *ei;
1657         struct btrfs_extent_data_ref *dref = NULL;
1658         struct btrfs_shared_data_ref *sref = NULL;
1659         unsigned long ptr;
1660         unsigned long end;
1661         u32 item_size;
1662         int size;
1663         int type;
1664         int ret;
1665         u64 refs;
1666
1667         leaf = path->nodes[0];
1668         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1669         refs = btrfs_extent_refs(leaf, ei);
1670         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1671         refs += refs_to_mod;
1672         btrfs_set_extent_refs(leaf, ei, refs);
1673         if (extent_op)
1674                 __run_delayed_extent_op(extent_op, leaf, ei);
1675
1676         type = btrfs_extent_inline_ref_type(leaf, iref);
1677
1678         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1679                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1680                 refs = btrfs_extent_data_ref_count(leaf, dref);
1681         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1682                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1683                 refs = btrfs_shared_data_ref_count(leaf, sref);
1684         } else {
1685                 refs = 1;
1686                 BUG_ON(refs_to_mod != -1);
1687         }
1688
1689         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1690         refs += refs_to_mod;
1691
1692         if (refs > 0) {
1693                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1694                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1695                 else
1696                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1697         } else {
1698                 size =  btrfs_extent_inline_ref_size(type);
1699                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1700                 ptr = (unsigned long)iref;
1701                 end = (unsigned long)ei + item_size;
1702                 if (ptr + size < end)
1703                         memmove_extent_buffer(leaf, ptr, ptr + size,
1704                                               end - ptr - size);
1705                 item_size -= size;
1706                 ret = btrfs_truncate_item(trans, root, path, item_size, 1);
1707         }
1708         btrfs_mark_buffer_dirty(leaf);
1709         return 0;
1710 }
1711
1712 static noinline_for_stack
1713 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1714                                  struct btrfs_root *root,
1715                                  struct btrfs_path *path,
1716                                  u64 bytenr, u64 num_bytes, u64 parent,
1717                                  u64 root_objectid, u64 owner,
1718                                  u64 offset, int refs_to_add,
1719                                  struct btrfs_delayed_extent_op *extent_op)
1720 {
1721         struct btrfs_extent_inline_ref *iref;
1722         int ret;
1723
1724         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1725                                            bytenr, num_bytes, parent,
1726                                            root_objectid, owner, offset, 1);
1727         if (ret == 0) {
1728                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1729                 ret = update_inline_extent_backref(trans, root, path, iref,
1730                                                    refs_to_add, extent_op);
1731         } else if (ret == -ENOENT) {
1732                 ret = setup_inline_extent_backref(trans, root, path, iref,
1733                                                   parent, root_objectid,
1734                                                   owner, offset, refs_to_add,
1735                                                   extent_op);
1736         }
1737         return ret;
1738 }
1739
1740 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1741                                  struct btrfs_root *root,
1742                                  struct btrfs_path *path,
1743                                  u64 bytenr, u64 parent, u64 root_objectid,
1744                                  u64 owner, u64 offset, int refs_to_add)
1745 {
1746         int ret;
1747         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1748                 BUG_ON(refs_to_add != 1);
1749                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1750                                             parent, root_objectid);
1751         } else {
1752                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1753                                              parent, root_objectid,
1754                                              owner, offset, refs_to_add);
1755         }
1756         return ret;
1757 }
1758
1759 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1760                                  struct btrfs_root *root,
1761                                  struct btrfs_path *path,
1762                                  struct btrfs_extent_inline_ref *iref,
1763                                  int refs_to_drop, int is_data)
1764 {
1765         int ret;
1766
1767         BUG_ON(!is_data && refs_to_drop != 1);
1768         if (iref) {
1769                 ret = update_inline_extent_backref(trans, root, path, iref,
1770                                                    -refs_to_drop, NULL);
1771         } else if (is_data) {
1772                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1773         } else {
1774                 ret = btrfs_del_item(trans, root, path);
1775         }
1776         return ret;
1777 }
1778
1779 static int btrfs_issue_discard(struct block_device *bdev,
1780                                 u64 start, u64 len)
1781 {
1782         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1783 }
1784
1785 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1786                                 u64 num_bytes, u64 *actual_bytes)
1787 {
1788         int ret;
1789         u64 discarded_bytes = 0;
1790         struct btrfs_multi_bio *multi = NULL;
1791
1792
1793         /* Tell the block device(s) that the sectors can be discarded */
1794         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1795                               bytenr, &num_bytes, &multi, 0);
1796         if (!ret) {
1797                 struct btrfs_bio_stripe *stripe = multi->stripes;
1798                 int i;
1799
1800
1801                 for (i = 0; i < multi->num_stripes; i++, stripe++) {
1802                         if (!stripe->dev->can_discard)
1803                                 continue;
1804
1805                         ret = btrfs_issue_discard(stripe->dev->bdev,
1806                                                   stripe->physical,
1807                                                   stripe->length);
1808                         if (!ret)
1809                                 discarded_bytes += stripe->length;
1810                         else if (ret != -EOPNOTSUPP)
1811                                 break;
1812
1813                         /*
1814                          * Just in case we get back EOPNOTSUPP for some reason,
1815                          * just ignore the return value so we don't screw up
1816                          * people calling discard_extent.
1817                          */
1818                         ret = 0;
1819                 }
1820                 kfree(multi);
1821         }
1822
1823         if (actual_bytes)
1824                 *actual_bytes = discarded_bytes;
1825
1826
1827         return ret;
1828 }
1829
1830 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1831                          struct btrfs_root *root,
1832                          u64 bytenr, u64 num_bytes, u64 parent,
1833                          u64 root_objectid, u64 owner, u64 offset)
1834 {
1835         int ret;
1836         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1837                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1838
1839         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1840                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
1841                                         parent, root_objectid, (int)owner,
1842                                         BTRFS_ADD_DELAYED_REF, NULL);
1843         } else {
1844                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
1845                                         parent, root_objectid, owner, offset,
1846                                         BTRFS_ADD_DELAYED_REF, NULL);
1847         }
1848         return ret;
1849 }
1850
1851 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1852                                   struct btrfs_root *root,
1853                                   u64 bytenr, u64 num_bytes,
1854                                   u64 parent, u64 root_objectid,
1855                                   u64 owner, u64 offset, int refs_to_add,
1856                                   struct btrfs_delayed_extent_op *extent_op)
1857 {
1858         struct btrfs_path *path;
1859         struct extent_buffer *leaf;
1860         struct btrfs_extent_item *item;
1861         u64 refs;
1862         int ret;
1863         int err = 0;
1864
1865         path = btrfs_alloc_path();
1866         if (!path)
1867                 return -ENOMEM;
1868
1869         path->reada = 1;
1870         path->leave_spinning = 1;
1871         /* this will setup the path even if it fails to insert the back ref */
1872         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1873                                            path, bytenr, num_bytes, parent,
1874                                            root_objectid, owner, offset,
1875                                            refs_to_add, extent_op);
1876         if (ret == 0)
1877                 goto out;
1878
1879         if (ret != -EAGAIN) {
1880                 err = ret;
1881                 goto out;
1882         }
1883
1884         leaf = path->nodes[0];
1885         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1886         refs = btrfs_extent_refs(leaf, item);
1887         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1888         if (extent_op)
1889                 __run_delayed_extent_op(extent_op, leaf, item);
1890
1891         btrfs_mark_buffer_dirty(leaf);
1892         btrfs_release_path(path);
1893
1894         path->reada = 1;
1895         path->leave_spinning = 1;
1896
1897         /* now insert the actual backref */
1898         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1899                                     path, bytenr, parent, root_objectid,
1900                                     owner, offset, refs_to_add);
1901         BUG_ON(ret);
1902 out:
1903         btrfs_free_path(path);
1904         return err;
1905 }
1906
1907 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1908                                 struct btrfs_root *root,
1909                                 struct btrfs_delayed_ref_node *node,
1910                                 struct btrfs_delayed_extent_op *extent_op,
1911                                 int insert_reserved)
1912 {
1913         int ret = 0;
1914         struct btrfs_delayed_data_ref *ref;
1915         struct btrfs_key ins;
1916         u64 parent = 0;
1917         u64 ref_root = 0;
1918         u64 flags = 0;
1919
1920         ins.objectid = node->bytenr;
1921         ins.offset = node->num_bytes;
1922         ins.type = BTRFS_EXTENT_ITEM_KEY;
1923
1924         ref = btrfs_delayed_node_to_data_ref(node);
1925         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1926                 parent = ref->parent;
1927         else
1928                 ref_root = ref->root;
1929
1930         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1931                 if (extent_op) {
1932                         BUG_ON(extent_op->update_key);
1933                         flags |= extent_op->flags_to_set;
1934                 }
1935                 ret = alloc_reserved_file_extent(trans, root,
1936                                                  parent, ref_root, flags,
1937                                                  ref->objectid, ref->offset,
1938                                                  &ins, node->ref_mod);
1939         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1940                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1941                                              node->num_bytes, parent,
1942                                              ref_root, ref->objectid,
1943                                              ref->offset, node->ref_mod,
1944                                              extent_op);
1945         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1946                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1947                                           node->num_bytes, parent,
1948                                           ref_root, ref->objectid,
1949                                           ref->offset, node->ref_mod,
1950                                           extent_op);
1951         } else {
1952                 BUG();
1953         }
1954         return ret;
1955 }
1956
1957 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1958                                     struct extent_buffer *leaf,
1959                                     struct btrfs_extent_item *ei)
1960 {
1961         u64 flags = btrfs_extent_flags(leaf, ei);
1962         if (extent_op->update_flags) {
1963                 flags |= extent_op->flags_to_set;
1964                 btrfs_set_extent_flags(leaf, ei, flags);
1965         }
1966
1967         if (extent_op->update_key) {
1968                 struct btrfs_tree_block_info *bi;
1969                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1970                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1971                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1972         }
1973 }
1974
1975 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1976                                  struct btrfs_root *root,
1977                                  struct btrfs_delayed_ref_node *node,
1978                                  struct btrfs_delayed_extent_op *extent_op)
1979 {
1980         struct btrfs_key key;
1981         struct btrfs_path *path;
1982         struct btrfs_extent_item *ei;
1983         struct extent_buffer *leaf;
1984         u32 item_size;
1985         int ret;
1986         int err = 0;
1987
1988         path = btrfs_alloc_path();
1989         if (!path)
1990                 return -ENOMEM;
1991
1992         key.objectid = node->bytenr;
1993         key.type = BTRFS_EXTENT_ITEM_KEY;
1994         key.offset = node->num_bytes;
1995
1996         path->reada = 1;
1997         path->leave_spinning = 1;
1998         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
1999                                 path, 0, 1);
2000         if (ret < 0) {
2001                 err = ret;
2002                 goto out;
2003         }
2004         if (ret > 0) {
2005                 err = -EIO;
2006                 goto out;
2007         }
2008
2009         leaf = path->nodes[0];
2010         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2011 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2012         if (item_size < sizeof(*ei)) {
2013                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2014                                              path, (u64)-1, 0);
2015                 if (ret < 0) {
2016                         err = ret;
2017                         goto out;
2018                 }
2019                 leaf = path->nodes[0];
2020                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2021         }
2022 #endif
2023         BUG_ON(item_size < sizeof(*ei));
2024         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2025         __run_delayed_extent_op(extent_op, leaf, ei);
2026
2027         btrfs_mark_buffer_dirty(leaf);
2028 out:
2029         btrfs_free_path(path);
2030         return err;
2031 }
2032
2033 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2034                                 struct btrfs_root *root,
2035                                 struct btrfs_delayed_ref_node *node,
2036                                 struct btrfs_delayed_extent_op *extent_op,
2037                                 int insert_reserved)
2038 {
2039         int ret = 0;
2040         struct btrfs_delayed_tree_ref *ref;
2041         struct btrfs_key ins;
2042         u64 parent = 0;
2043         u64 ref_root = 0;
2044
2045         ins.objectid = node->bytenr;
2046         ins.offset = node->num_bytes;
2047         ins.type = BTRFS_EXTENT_ITEM_KEY;
2048
2049         ref = btrfs_delayed_node_to_tree_ref(node);
2050         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2051                 parent = ref->parent;
2052         else
2053                 ref_root = ref->root;
2054
2055         BUG_ON(node->ref_mod != 1);
2056         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2057                 BUG_ON(!extent_op || !extent_op->update_flags ||
2058                        !extent_op->update_key);
2059                 ret = alloc_reserved_tree_block(trans, root,
2060                                                 parent, ref_root,
2061                                                 extent_op->flags_to_set,
2062                                                 &extent_op->key,
2063                                                 ref->level, &ins);
2064         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2065                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2066                                              node->num_bytes, parent, ref_root,
2067                                              ref->level, 0, 1, extent_op);
2068         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2069                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2070                                           node->num_bytes, parent, ref_root,
2071                                           ref->level, 0, 1, extent_op);
2072         } else {
2073                 BUG();
2074         }
2075         return ret;
2076 }
2077
2078 /* helper function to actually process a single delayed ref entry */
2079 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2080                                struct btrfs_root *root,
2081                                struct btrfs_delayed_ref_node *node,
2082                                struct btrfs_delayed_extent_op *extent_op,
2083                                int insert_reserved)
2084 {
2085         int ret;
2086         if (btrfs_delayed_ref_is_head(node)) {
2087                 struct btrfs_delayed_ref_head *head;
2088                 /*
2089                  * we've hit the end of the chain and we were supposed
2090                  * to insert this extent into the tree.  But, it got
2091                  * deleted before we ever needed to insert it, so all
2092                  * we have to do is clean up the accounting
2093                  */
2094                 BUG_ON(extent_op);
2095                 head = btrfs_delayed_node_to_head(node);
2096                 if (insert_reserved) {
2097                         btrfs_pin_extent(root, node->bytenr,
2098                                          node->num_bytes, 1);
2099                         if (head->is_data) {
2100                                 ret = btrfs_del_csums(trans, root,
2101                                                       node->bytenr,
2102                                                       node->num_bytes);
2103                                 BUG_ON(ret);
2104                         }
2105                 }
2106                 mutex_unlock(&head->mutex);
2107                 return 0;
2108         }
2109
2110         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2111             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2112                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2113                                            insert_reserved);
2114         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2115                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2116                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2117                                            insert_reserved);
2118         else
2119                 BUG();
2120         return ret;
2121 }
2122
2123 static noinline struct btrfs_delayed_ref_node *
2124 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2125 {
2126         struct rb_node *node;
2127         struct btrfs_delayed_ref_node *ref;
2128         int action = BTRFS_ADD_DELAYED_REF;
2129 again:
2130         /*
2131          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2132          * this prevents ref count from going down to zero when
2133          * there still are pending delayed ref.
2134          */
2135         node = rb_prev(&head->node.rb_node);
2136         while (1) {
2137                 if (!node)
2138                         break;
2139                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2140                                 rb_node);
2141                 if (ref->bytenr != head->node.bytenr)
2142                         break;
2143                 if (ref->action == action)
2144                         return ref;
2145                 node = rb_prev(node);
2146         }
2147         if (action == BTRFS_ADD_DELAYED_REF) {
2148                 action = BTRFS_DROP_DELAYED_REF;
2149                 goto again;
2150         }
2151         return NULL;
2152 }
2153
2154 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2155                                        struct btrfs_root *root,
2156                                        struct list_head *cluster)
2157 {
2158         struct btrfs_delayed_ref_root *delayed_refs;
2159         struct btrfs_delayed_ref_node *ref;
2160         struct btrfs_delayed_ref_head *locked_ref = NULL;
2161         struct btrfs_delayed_extent_op *extent_op;
2162         int ret;
2163         int count = 0;
2164         int must_insert_reserved = 0;
2165
2166         delayed_refs = &trans->transaction->delayed_refs;
2167         while (1) {
2168                 if (!locked_ref) {
2169                         /* pick a new head ref from the cluster list */
2170                         if (list_empty(cluster))
2171                                 break;
2172
2173                         locked_ref = list_entry(cluster->next,
2174                                      struct btrfs_delayed_ref_head, cluster);
2175
2176                         /* grab the lock that says we are going to process
2177                          * all the refs for this head */
2178                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2179
2180                         /*
2181                          * we may have dropped the spin lock to get the head
2182                          * mutex lock, and that might have given someone else
2183                          * time to free the head.  If that's true, it has been
2184                          * removed from our list and we can move on.
2185                          */
2186                         if (ret == -EAGAIN) {
2187                                 locked_ref = NULL;
2188                                 count++;
2189                                 continue;
2190                         }
2191                 }
2192
2193                 /*
2194                  * record the must insert reserved flag before we
2195                  * drop the spin lock.
2196                  */
2197                 must_insert_reserved = locked_ref->must_insert_reserved;
2198                 locked_ref->must_insert_reserved = 0;
2199
2200                 extent_op = locked_ref->extent_op;
2201                 locked_ref->extent_op = NULL;
2202
2203                 /*
2204                  * locked_ref is the head node, so we have to go one
2205                  * node back for any delayed ref updates
2206                  */
2207                 ref = select_delayed_ref(locked_ref);
2208                 if (!ref) {
2209                         /* All delayed refs have been processed, Go ahead
2210                          * and send the head node to run_one_delayed_ref,
2211                          * so that any accounting fixes can happen
2212                          */
2213                         ref = &locked_ref->node;
2214
2215                         if (extent_op && must_insert_reserved) {
2216                                 kfree(extent_op);
2217                                 extent_op = NULL;
2218                         }
2219
2220                         if (extent_op) {
2221                                 spin_unlock(&delayed_refs->lock);
2222
2223                                 ret = run_delayed_extent_op(trans, root,
2224                                                             ref, extent_op);
2225                                 BUG_ON(ret);
2226                                 kfree(extent_op);
2227
2228                                 cond_resched();
2229                                 spin_lock(&delayed_refs->lock);
2230                                 continue;
2231                         }
2232
2233                         list_del_init(&locked_ref->cluster);
2234                         locked_ref = NULL;
2235                 }
2236
2237                 ref->in_tree = 0;
2238                 rb_erase(&ref->rb_node, &delayed_refs->root);
2239                 delayed_refs->num_entries--;
2240
2241                 spin_unlock(&delayed_refs->lock);
2242
2243                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2244                                           must_insert_reserved);
2245                 BUG_ON(ret);
2246
2247                 btrfs_put_delayed_ref(ref);
2248                 kfree(extent_op);
2249                 count++;
2250
2251                 cond_resched();
2252                 spin_lock(&delayed_refs->lock);
2253         }
2254         return count;
2255 }
2256
2257 /*
2258  * this starts processing the delayed reference count updates and
2259  * extent insertions we have queued up so far.  count can be
2260  * 0, which means to process everything in the tree at the start
2261  * of the run (but not newly added entries), or it can be some target
2262  * number you'd like to process.
2263  */
2264 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2265                            struct btrfs_root *root, unsigned long count)
2266 {
2267         struct rb_node *node;
2268         struct btrfs_delayed_ref_root *delayed_refs;
2269         struct btrfs_delayed_ref_node *ref;
2270         struct list_head cluster;
2271         int ret;
2272         int run_all = count == (unsigned long)-1;
2273         int run_most = 0;
2274
2275         if (root == root->fs_info->extent_root)
2276                 root = root->fs_info->tree_root;
2277
2278         delayed_refs = &trans->transaction->delayed_refs;
2279         INIT_LIST_HEAD(&cluster);
2280 again:
2281         spin_lock(&delayed_refs->lock);
2282         if (count == 0) {
2283                 count = delayed_refs->num_entries * 2;
2284                 run_most = 1;
2285         }
2286         while (1) {
2287                 if (!(run_all || run_most) &&
2288                     delayed_refs->num_heads_ready < 64)
2289                         break;
2290
2291                 /*
2292                  * go find something we can process in the rbtree.  We start at
2293                  * the beginning of the tree, and then build a cluster
2294                  * of refs to process starting at the first one we are able to
2295                  * lock
2296                  */
2297                 ret = btrfs_find_ref_cluster(trans, &cluster,
2298                                              delayed_refs->run_delayed_start);
2299                 if (ret)
2300                         break;
2301
2302                 ret = run_clustered_refs(trans, root, &cluster);
2303                 BUG_ON(ret < 0);
2304
2305                 count -= min_t(unsigned long, ret, count);
2306
2307                 if (count == 0)
2308                         break;
2309         }
2310
2311         if (run_all) {
2312                 node = rb_first(&delayed_refs->root);
2313                 if (!node)
2314                         goto out;
2315                 count = (unsigned long)-1;
2316
2317                 while (node) {
2318                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2319                                        rb_node);
2320                         if (btrfs_delayed_ref_is_head(ref)) {
2321                                 struct btrfs_delayed_ref_head *head;
2322
2323                                 head = btrfs_delayed_node_to_head(ref);
2324                                 atomic_inc(&ref->refs);
2325
2326                                 spin_unlock(&delayed_refs->lock);
2327                                 /*
2328                                  * Mutex was contended, block until it's
2329                                  * released and try again
2330                                  */
2331                                 mutex_lock(&head->mutex);
2332                                 mutex_unlock(&head->mutex);
2333
2334                                 btrfs_put_delayed_ref(ref);
2335                                 cond_resched();
2336                                 goto again;
2337                         }
2338                         node = rb_next(node);
2339                 }
2340                 spin_unlock(&delayed_refs->lock);
2341                 schedule_timeout(1);
2342                 goto again;
2343         }
2344 out:
2345         spin_unlock(&delayed_refs->lock);
2346         return 0;
2347 }
2348
2349 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2350                                 struct btrfs_root *root,
2351                                 u64 bytenr, u64 num_bytes, u64 flags,
2352                                 int is_data)
2353 {
2354         struct btrfs_delayed_extent_op *extent_op;
2355         int ret;
2356
2357         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2358         if (!extent_op)
2359                 return -ENOMEM;
2360
2361         extent_op->flags_to_set = flags;
2362         extent_op->update_flags = 1;
2363         extent_op->update_key = 0;
2364         extent_op->is_data = is_data ? 1 : 0;
2365
2366         ret = btrfs_add_delayed_extent_op(trans, bytenr, num_bytes, extent_op);
2367         if (ret)
2368                 kfree(extent_op);
2369         return ret;
2370 }
2371
2372 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2373                                       struct btrfs_root *root,
2374                                       struct btrfs_path *path,
2375                                       u64 objectid, u64 offset, u64 bytenr)
2376 {
2377         struct btrfs_delayed_ref_head *head;
2378         struct btrfs_delayed_ref_node *ref;
2379         struct btrfs_delayed_data_ref *data_ref;
2380         struct btrfs_delayed_ref_root *delayed_refs;
2381         struct rb_node *node;
2382         int ret = 0;
2383
2384         ret = -ENOENT;
2385         delayed_refs = &trans->transaction->delayed_refs;
2386         spin_lock(&delayed_refs->lock);
2387         head = btrfs_find_delayed_ref_head(trans, bytenr);
2388         if (!head)
2389                 goto out;
2390
2391         if (!mutex_trylock(&head->mutex)) {
2392                 atomic_inc(&head->node.refs);
2393                 spin_unlock(&delayed_refs->lock);
2394
2395                 btrfs_release_path(path);
2396
2397                 /*
2398                  * Mutex was contended, block until it's released and let
2399                  * caller try again
2400                  */
2401                 mutex_lock(&head->mutex);
2402                 mutex_unlock(&head->mutex);
2403                 btrfs_put_delayed_ref(&head->node);
2404                 return -EAGAIN;
2405         }
2406
2407         node = rb_prev(&head->node.rb_node);
2408         if (!node)
2409                 goto out_unlock;
2410
2411         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2412
2413         if (ref->bytenr != bytenr)
2414                 goto out_unlock;
2415
2416         ret = 1;
2417         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2418                 goto out_unlock;
2419
2420         data_ref = btrfs_delayed_node_to_data_ref(ref);
2421
2422         node = rb_prev(node);
2423         if (node) {
2424                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2425                 if (ref->bytenr == bytenr)
2426                         goto out_unlock;
2427         }
2428
2429         if (data_ref->root != root->root_key.objectid ||
2430             data_ref->objectid != objectid || data_ref->offset != offset)
2431                 goto out_unlock;
2432
2433         ret = 0;
2434 out_unlock:
2435         mutex_unlock(&head->mutex);
2436 out:
2437         spin_unlock(&delayed_refs->lock);
2438         return ret;
2439 }
2440
2441 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2442                                         struct btrfs_root *root,
2443                                         struct btrfs_path *path,
2444                                         u64 objectid, u64 offset, u64 bytenr)
2445 {
2446         struct btrfs_root *extent_root = root->fs_info->extent_root;
2447         struct extent_buffer *leaf;
2448         struct btrfs_extent_data_ref *ref;
2449         struct btrfs_extent_inline_ref *iref;
2450         struct btrfs_extent_item *ei;
2451         struct btrfs_key key;
2452         u32 item_size;
2453         int ret;
2454
2455         key.objectid = bytenr;
2456         key.offset = (u64)-1;
2457         key.type = BTRFS_EXTENT_ITEM_KEY;
2458
2459         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2460         if (ret < 0)
2461                 goto out;
2462         BUG_ON(ret == 0);
2463
2464         ret = -ENOENT;
2465         if (path->slots[0] == 0)
2466                 goto out;
2467
2468         path->slots[0]--;
2469         leaf = path->nodes[0];
2470         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2471
2472         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2473                 goto out;
2474
2475         ret = 1;
2476         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2477 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2478         if (item_size < sizeof(*ei)) {
2479                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2480                 goto out;
2481         }
2482 #endif
2483         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2484
2485         if (item_size != sizeof(*ei) +
2486             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2487                 goto out;
2488
2489         if (btrfs_extent_generation(leaf, ei) <=
2490             btrfs_root_last_snapshot(&root->root_item))
2491                 goto out;
2492
2493         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2494         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2495             BTRFS_EXTENT_DATA_REF_KEY)
2496                 goto out;
2497
2498         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2499         if (btrfs_extent_refs(leaf, ei) !=
2500             btrfs_extent_data_ref_count(leaf, ref) ||
2501             btrfs_extent_data_ref_root(leaf, ref) !=
2502             root->root_key.objectid ||
2503             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2504             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2505                 goto out;
2506
2507         ret = 0;
2508 out:
2509         return ret;
2510 }
2511
2512 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2513                           struct btrfs_root *root,
2514                           u64 objectid, u64 offset, u64 bytenr)
2515 {
2516         struct btrfs_path *path;
2517         int ret;
2518         int ret2;
2519
2520         path = btrfs_alloc_path();
2521         if (!path)
2522                 return -ENOENT;
2523
2524         do {
2525                 ret = check_committed_ref(trans, root, path, objectid,
2526                                           offset, bytenr);
2527                 if (ret && ret != -ENOENT)
2528                         goto out;
2529
2530                 ret2 = check_delayed_ref(trans, root, path, objectid,
2531                                          offset, bytenr);
2532         } while (ret2 == -EAGAIN);
2533
2534         if (ret2 && ret2 != -ENOENT) {
2535                 ret = ret2;
2536                 goto out;
2537         }
2538
2539         if (ret != -ENOENT || ret2 != -ENOENT)
2540                 ret = 0;
2541 out:
2542         btrfs_free_path(path);
2543         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2544                 WARN_ON(ret > 0);
2545         return ret;
2546 }
2547
2548 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2549                            struct btrfs_root *root,
2550                            struct extent_buffer *buf,
2551                            int full_backref, int inc)
2552 {
2553         u64 bytenr;
2554         u64 num_bytes;
2555         u64 parent;
2556         u64 ref_root;
2557         u32 nritems;
2558         struct btrfs_key key;
2559         struct btrfs_file_extent_item *fi;
2560         int i;
2561         int level;
2562         int ret = 0;
2563         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2564                             u64, u64, u64, u64, u64, u64);
2565
2566         ref_root = btrfs_header_owner(buf);
2567         nritems = btrfs_header_nritems(buf);
2568         level = btrfs_header_level(buf);
2569
2570         if (!root->ref_cows && level == 0)
2571                 return 0;
2572
2573         if (inc)
2574                 process_func = btrfs_inc_extent_ref;
2575         else
2576                 process_func = btrfs_free_extent;
2577
2578         if (full_backref)
2579                 parent = buf->start;
2580         else
2581                 parent = 0;
2582
2583         for (i = 0; i < nritems; i++) {
2584                 if (level == 0) {
2585                         btrfs_item_key_to_cpu(buf, &key, i);
2586                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2587                                 continue;
2588                         fi = btrfs_item_ptr(buf, i,
2589                                             struct btrfs_file_extent_item);
2590                         if (btrfs_file_extent_type(buf, fi) ==
2591                             BTRFS_FILE_EXTENT_INLINE)
2592                                 continue;
2593                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2594                         if (bytenr == 0)
2595                                 continue;
2596
2597                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2598                         key.offset -= btrfs_file_extent_offset(buf, fi);
2599                         ret = process_func(trans, root, bytenr, num_bytes,
2600                                            parent, ref_root, key.objectid,
2601                                            key.offset);
2602                         if (ret)
2603                                 goto fail;
2604                 } else {
2605                         bytenr = btrfs_node_blockptr(buf, i);
2606                         num_bytes = btrfs_level_size(root, level - 1);
2607                         ret = process_func(trans, root, bytenr, num_bytes,
2608                                            parent, ref_root, level - 1, 0);
2609                         if (ret)
2610                                 goto fail;
2611                 }
2612         }
2613         return 0;
2614 fail:
2615         BUG();
2616         return ret;
2617 }
2618
2619 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2620                   struct extent_buffer *buf, int full_backref)
2621 {
2622         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2623 }
2624
2625 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2626                   struct extent_buffer *buf, int full_backref)
2627 {
2628         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2629 }
2630
2631 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2632                                  struct btrfs_root *root,
2633                                  struct btrfs_path *path,
2634                                  struct btrfs_block_group_cache *cache)
2635 {
2636         int ret;
2637         struct btrfs_root *extent_root = root->fs_info->extent_root;
2638         unsigned long bi;
2639         struct extent_buffer *leaf;
2640
2641         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2642         if (ret < 0)
2643                 goto fail;
2644         BUG_ON(ret);
2645
2646         leaf = path->nodes[0];
2647         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2648         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2649         btrfs_mark_buffer_dirty(leaf);
2650         btrfs_release_path(path);
2651 fail:
2652         if (ret)
2653                 return ret;
2654         return 0;
2655
2656 }
2657
2658 static struct btrfs_block_group_cache *
2659 next_block_group(struct btrfs_root *root,
2660                  struct btrfs_block_group_cache *cache)
2661 {
2662         struct rb_node *node;
2663         spin_lock(&root->fs_info->block_group_cache_lock);
2664         node = rb_next(&cache->cache_node);
2665         btrfs_put_block_group(cache);
2666         if (node) {
2667                 cache = rb_entry(node, struct btrfs_block_group_cache,
2668                                  cache_node);
2669                 btrfs_get_block_group(cache);
2670         } else
2671                 cache = NULL;
2672         spin_unlock(&root->fs_info->block_group_cache_lock);
2673         return cache;
2674 }
2675
2676 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2677                             struct btrfs_trans_handle *trans,
2678                             struct btrfs_path *path)
2679 {
2680         struct btrfs_root *root = block_group->fs_info->tree_root;
2681         struct inode *inode = NULL;
2682         u64 alloc_hint = 0;
2683         int dcs = BTRFS_DC_ERROR;
2684         int num_pages = 0;
2685         int retries = 0;
2686         int ret = 0;
2687
2688         /*
2689          * If this block group is smaller than 100 megs don't bother caching the
2690          * block group.
2691          */
2692         if (block_group->key.offset < (100 * 1024 * 1024)) {
2693                 spin_lock(&block_group->lock);
2694                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2695                 spin_unlock(&block_group->lock);
2696                 return 0;
2697         }
2698
2699 again:
2700         inode = lookup_free_space_inode(root, block_group, path);
2701         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2702                 ret = PTR_ERR(inode);
2703                 btrfs_release_path(path);
2704                 goto out;
2705         }
2706
2707         if (IS_ERR(inode)) {
2708                 BUG_ON(retries);
2709                 retries++;
2710
2711                 if (block_group->ro)
2712                         goto out_free;
2713
2714                 ret = create_free_space_inode(root, trans, block_group, path);
2715                 if (ret)
2716                         goto out_free;
2717                 goto again;
2718         }
2719
2720         /*
2721          * We want to set the generation to 0, that way if anything goes wrong
2722          * from here on out we know not to trust this cache when we load up next
2723          * time.
2724          */
2725         BTRFS_I(inode)->generation = 0;
2726         ret = btrfs_update_inode(trans, root, inode);
2727         WARN_ON(ret);
2728
2729         if (i_size_read(inode) > 0) {
2730                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2731                                                       inode);
2732                 if (ret)
2733                         goto out_put;
2734         }
2735
2736         spin_lock(&block_group->lock);
2737         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2738                 /* We're not cached, don't bother trying to write stuff out */
2739                 dcs = BTRFS_DC_WRITTEN;
2740                 spin_unlock(&block_group->lock);
2741                 goto out_put;
2742         }
2743         spin_unlock(&block_group->lock);
2744
2745         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2746         if (!num_pages)
2747                 num_pages = 1;
2748
2749         /*
2750          * Just to make absolutely sure we have enough space, we're going to
2751          * preallocate 12 pages worth of space for each block group.  In
2752          * practice we ought to use at most 8, but we need extra space so we can
2753          * add our header and have a terminator between the extents and the
2754          * bitmaps.
2755          */
2756         num_pages *= 16;
2757         num_pages *= PAGE_CACHE_SIZE;
2758
2759         ret = btrfs_check_data_free_space(inode, num_pages);
2760         if (ret)
2761                 goto out_put;
2762
2763         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2764                                               num_pages, num_pages,
2765                                               &alloc_hint);
2766         if (!ret)
2767                 dcs = BTRFS_DC_SETUP;
2768         btrfs_free_reserved_data_space(inode, num_pages);
2769 out_put:
2770         iput(inode);
2771 out_free:
2772         btrfs_release_path(path);
2773 out:
2774         spin_lock(&block_group->lock);
2775         block_group->disk_cache_state = dcs;
2776         spin_unlock(&block_group->lock);
2777
2778         return ret;
2779 }
2780
2781 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2782                                    struct btrfs_root *root)
2783 {
2784         struct btrfs_block_group_cache *cache;
2785         int err = 0;
2786         struct btrfs_path *path;
2787         u64 last = 0;
2788
2789         path = btrfs_alloc_path();
2790         if (!path)
2791                 return -ENOMEM;
2792
2793 again:
2794         while (1) {
2795                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2796                 while (cache) {
2797                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2798                                 break;
2799                         cache = next_block_group(root, cache);
2800                 }
2801                 if (!cache) {
2802                         if (last == 0)
2803                                 break;
2804                         last = 0;
2805                         continue;
2806                 }
2807                 err = cache_save_setup(cache, trans, path);
2808                 last = cache->key.objectid + cache->key.offset;
2809                 btrfs_put_block_group(cache);
2810         }
2811
2812         while (1) {
2813                 if (last == 0) {
2814                         err = btrfs_run_delayed_refs(trans, root,
2815                                                      (unsigned long)-1);
2816                         BUG_ON(err);
2817                 }
2818
2819                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2820                 while (cache) {
2821                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2822                                 btrfs_put_block_group(cache);
2823                                 goto again;
2824                         }
2825
2826                         if (cache->dirty)
2827                                 break;
2828                         cache = next_block_group(root, cache);
2829                 }
2830                 if (!cache) {
2831                         if (last == 0)
2832                                 break;
2833                         last = 0;
2834                         continue;
2835                 }
2836
2837                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
2838                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
2839                 cache->dirty = 0;
2840                 last = cache->key.objectid + cache->key.offset;
2841
2842                 err = write_one_cache_group(trans, root, path, cache);
2843                 BUG_ON(err);
2844                 btrfs_put_block_group(cache);
2845         }
2846
2847         while (1) {
2848                 /*
2849                  * I don't think this is needed since we're just marking our
2850                  * preallocated extent as written, but just in case it can't
2851                  * hurt.
2852                  */
2853                 if (last == 0) {
2854                         err = btrfs_run_delayed_refs(trans, root,
2855                                                      (unsigned long)-1);
2856                         BUG_ON(err);
2857                 }
2858
2859                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2860                 while (cache) {
2861                         /*
2862                          * Really this shouldn't happen, but it could if we
2863                          * couldn't write the entire preallocated extent and
2864                          * splitting the extent resulted in a new block.
2865                          */
2866                         if (cache->dirty) {
2867                                 btrfs_put_block_group(cache);
2868                                 goto again;
2869                         }
2870                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2871                                 break;
2872                         cache = next_block_group(root, cache);
2873                 }
2874                 if (!cache) {
2875                         if (last == 0)
2876                                 break;
2877                         last = 0;
2878                         continue;
2879                 }
2880
2881                 btrfs_write_out_cache(root, trans, cache, path);
2882
2883                 /*
2884                  * If we didn't have an error then the cache state is still
2885                  * NEED_WRITE, so we can set it to WRITTEN.
2886                  */
2887                 if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2888                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
2889                 last = cache->key.objectid + cache->key.offset;
2890                 btrfs_put_block_group(cache);
2891         }
2892
2893         btrfs_free_path(path);
2894         return 0;
2895 }
2896
2897 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
2898 {
2899         struct btrfs_block_group_cache *block_group;
2900         int readonly = 0;
2901
2902         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
2903         if (!block_group || block_group->ro)
2904                 readonly = 1;
2905         if (block_group)
2906                 btrfs_put_block_group(block_group);
2907         return readonly;
2908 }
2909
2910 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
2911                              u64 total_bytes, u64 bytes_used,
2912                              struct btrfs_space_info **space_info)
2913 {
2914         struct btrfs_space_info *found;
2915         int i;
2916         int factor;
2917
2918         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
2919                      BTRFS_BLOCK_GROUP_RAID10))
2920                 factor = 2;
2921         else
2922                 factor = 1;
2923
2924         found = __find_space_info(info, flags);
2925         if (found) {
2926                 spin_lock(&found->lock);
2927                 found->total_bytes += total_bytes;
2928                 found->disk_total += total_bytes * factor;
2929                 found->bytes_used += bytes_used;
2930                 found->disk_used += bytes_used * factor;
2931                 found->full = 0;
2932                 spin_unlock(&found->lock);
2933                 *space_info = found;
2934                 return 0;
2935         }
2936         found = kzalloc(sizeof(*found), GFP_NOFS);
2937         if (!found)
2938                 return -ENOMEM;
2939
2940         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
2941                 INIT_LIST_HEAD(&found->block_groups[i]);
2942         init_rwsem(&found->groups_sem);
2943         spin_lock_init(&found->lock);
2944         found->flags = flags & (BTRFS_BLOCK_GROUP_DATA |
2945                                 BTRFS_BLOCK_GROUP_SYSTEM |
2946                                 BTRFS_BLOCK_GROUP_METADATA);
2947         found->total_bytes = total_bytes;
2948         found->disk_total = total_bytes * factor;
2949         found->bytes_used = bytes_used;
2950         found->disk_used = bytes_used * factor;
2951         found->bytes_pinned = 0;
2952         found->bytes_reserved = 0;
2953         found->bytes_readonly = 0;
2954         found->bytes_may_use = 0;
2955         found->full = 0;
2956         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
2957         found->chunk_alloc = 0;
2958         found->flush = 0;
2959         init_waitqueue_head(&found->wait);
2960         *space_info = found;
2961         list_add_rcu(&found->list, &info->space_info);
2962         return 0;
2963 }
2964
2965 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
2966 {
2967         u64 extra_flags = flags & (BTRFS_BLOCK_GROUP_RAID0 |
2968                                    BTRFS_BLOCK_GROUP_RAID1 |
2969                                    BTRFS_BLOCK_GROUP_RAID10 |
2970                                    BTRFS_BLOCK_GROUP_DUP);
2971         if (extra_flags) {
2972                 if (flags & BTRFS_BLOCK_GROUP_DATA)
2973                         fs_info->avail_data_alloc_bits |= extra_flags;
2974                 if (flags & BTRFS_BLOCK_GROUP_METADATA)
2975                         fs_info->avail_metadata_alloc_bits |= extra_flags;
2976                 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
2977                         fs_info->avail_system_alloc_bits |= extra_flags;
2978         }
2979 }
2980
2981 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
2982 {
2983         /*
2984          * we add in the count of missing devices because we want
2985          * to make sure that any RAID levels on a degraded FS
2986          * continue to be honored.
2987          */
2988         u64 num_devices = root->fs_info->fs_devices->rw_devices +
2989                 root->fs_info->fs_devices->missing_devices;
2990
2991         if (num_devices == 1)
2992                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
2993         if (num_devices < 4)
2994                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
2995
2996         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
2997             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
2998                       BTRFS_BLOCK_GROUP_RAID10))) {
2999                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3000         }
3001
3002         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3003             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3004                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3005         }
3006
3007         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3008             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3009              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3010              (flags & BTRFS_BLOCK_GROUP_DUP)))
3011                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3012         return flags;
3013 }
3014
3015 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3016 {
3017         if (flags & BTRFS_BLOCK_GROUP_DATA)
3018                 flags |= root->fs_info->avail_data_alloc_bits &
3019                          root->fs_info->data_alloc_profile;
3020         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3021                 flags |= root->fs_info->avail_system_alloc_bits &
3022                          root->fs_info->system_alloc_profile;
3023         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3024                 flags |= root->fs_info->avail_metadata_alloc_bits &
3025                          root->fs_info->metadata_alloc_profile;
3026         return btrfs_reduce_alloc_profile(root, flags);
3027 }
3028
3029 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3030 {
3031         u64 flags;
3032
3033         if (data)
3034                 flags = BTRFS_BLOCK_GROUP_DATA;
3035         else if (root == root->fs_info->chunk_root)
3036                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3037         else
3038                 flags = BTRFS_BLOCK_GROUP_METADATA;
3039
3040         return get_alloc_profile(root, flags);
3041 }
3042
3043 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3044 {
3045         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3046                                                        BTRFS_BLOCK_GROUP_DATA);
3047 }
3048
3049 /*
3050  * This will check the space that the inode allocates from to make sure we have
3051  * enough space for bytes.
3052  */
3053 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3054 {
3055         struct btrfs_space_info *data_sinfo;
3056         struct btrfs_root *root = BTRFS_I(inode)->root;
3057         u64 used;
3058         int ret = 0, committed = 0, alloc_chunk = 1;
3059
3060         /* make sure bytes are sectorsize aligned */
3061         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3062
3063         if (root == root->fs_info->tree_root ||
3064             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3065                 alloc_chunk = 0;
3066                 committed = 1;
3067         }
3068
3069         data_sinfo = BTRFS_I(inode)->space_info;
3070         if (!data_sinfo)
3071                 goto alloc;
3072
3073 again:
3074         /* make sure we have enough space to handle the data first */
3075         spin_lock(&data_sinfo->lock);
3076         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3077                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3078                 data_sinfo->bytes_may_use;
3079
3080         if (used + bytes > data_sinfo->total_bytes) {
3081                 struct btrfs_trans_handle *trans;
3082
3083                 /*
3084                  * if we don't have enough free bytes in this space then we need
3085                  * to alloc a new chunk.
3086                  */
3087                 if (!data_sinfo->full && alloc_chunk) {
3088                         u64 alloc_target;
3089
3090                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3091                         spin_unlock(&data_sinfo->lock);
3092 alloc:
3093                         alloc_target = btrfs_get_alloc_profile(root, 1);
3094                         trans = btrfs_join_transaction(root);
3095                         if (IS_ERR(trans))
3096                                 return PTR_ERR(trans);
3097
3098                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3099                                              bytes + 2 * 1024 * 1024,
3100                                              alloc_target,
3101                                              CHUNK_ALLOC_NO_FORCE);
3102                         btrfs_end_transaction(trans, root);
3103                         if (ret < 0) {
3104                                 if (ret != -ENOSPC)
3105                                         return ret;
3106                                 else
3107                                         goto commit_trans;
3108                         }
3109
3110                         if (!data_sinfo) {
3111                                 btrfs_set_inode_space_info(root, inode);
3112                                 data_sinfo = BTRFS_I(inode)->space_info;
3113                         }
3114                         goto again;
3115                 }
3116
3117                 /*
3118                  * If we have less pinned bytes than we want to allocate then
3119                  * don't bother committing the transaction, it won't help us.
3120                  */
3121                 if (data_sinfo->bytes_pinned < bytes)
3122                         committed = 1;
3123                 spin_unlock(&data_sinfo->lock);
3124
3125                 /* commit the current transaction and try again */
3126 commit_trans:
3127                 if (!committed &&
3128                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3129                         committed = 1;
3130                         trans = btrfs_join_transaction(root);
3131                         if (IS_ERR(trans))
3132                                 return PTR_ERR(trans);
3133                         ret = btrfs_commit_transaction(trans, root);
3134                         if (ret)
3135                                 return ret;
3136                         goto again;
3137                 }
3138
3139                 return -ENOSPC;
3140         }
3141         data_sinfo->bytes_may_use += bytes;
3142         spin_unlock(&data_sinfo->lock);
3143
3144         return 0;
3145 }
3146
3147 /*
3148  * Called if we need to clear a data reservation for this inode.
3149  */
3150 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3151 {
3152         struct btrfs_root *root = BTRFS_I(inode)->root;
3153         struct btrfs_space_info *data_sinfo;
3154
3155         /* make sure bytes are sectorsize aligned */
3156         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3157
3158         data_sinfo = BTRFS_I(inode)->space_info;
3159         spin_lock(&data_sinfo->lock);
3160         data_sinfo->bytes_may_use -= bytes;
3161         spin_unlock(&data_sinfo->lock);
3162 }
3163
3164 static void force_metadata_allocation(struct btrfs_fs_info *info)
3165 {
3166         struct list_head *head = &info->space_info;
3167         struct btrfs_space_info *found;
3168
3169         rcu_read_lock();
3170         list_for_each_entry_rcu(found, head, list) {
3171                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3172                         found->force_alloc = CHUNK_ALLOC_FORCE;
3173         }
3174         rcu_read_unlock();
3175 }
3176
3177 static int should_alloc_chunk(struct btrfs_root *root,
3178                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3179                               int force)
3180 {
3181         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3182         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3183         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3184         u64 thresh;
3185
3186         if (force == CHUNK_ALLOC_FORCE)
3187                 return 1;
3188
3189         /*
3190          * We need to take into account the global rsv because for all intents
3191          * and purposes it's used space.  Don't worry about locking the
3192          * global_rsv, it doesn't change except when the transaction commits.
3193          */
3194         num_allocated += global_rsv->size;
3195
3196         /*
3197          * in limited mode, we want to have some free space up to
3198          * about 1% of the FS size.
3199          */
3200         if (force == CHUNK_ALLOC_LIMITED) {
3201                 thresh = btrfs_super_total_bytes(&root->fs_info->super_copy);
3202                 thresh = max_t(u64, 64 * 1024 * 1024,
3203                                div_factor_fine(thresh, 1));
3204
3205                 if (num_bytes - num_allocated < thresh)
3206                         return 1;
3207         }
3208
3209         /*
3210          * we have two similar checks here, one based on percentage
3211          * and once based on a hard number of 256MB.  The idea
3212          * is that if we have a good amount of free
3213          * room, don't allocate a chunk.  A good mount is
3214          * less than 80% utilized of the chunks we have allocated,
3215          * or more than 256MB free
3216          */
3217         if (num_allocated + alloc_bytes + 256 * 1024 * 1024 < num_bytes)
3218                 return 0;
3219
3220         if (num_allocated + alloc_bytes < div_factor(num_bytes, 8))
3221                 return 0;
3222
3223         thresh = btrfs_super_total_bytes(&root->fs_info->super_copy);
3224
3225         /* 256MB or 5% of the FS */
3226         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 5));
3227
3228         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 3))
3229                 return 0;
3230         return 1;
3231 }
3232
3233 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3234                           struct btrfs_root *extent_root, u64 alloc_bytes,
3235                           u64 flags, int force)
3236 {
3237         struct btrfs_space_info *space_info;
3238         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3239         int wait_for_alloc = 0;
3240         int ret = 0;
3241
3242         flags = btrfs_reduce_alloc_profile(extent_root, flags);
3243
3244         space_info = __find_space_info(extent_root->fs_info, flags);
3245         if (!space_info) {
3246                 ret = update_space_info(extent_root->fs_info, flags,
3247                                         0, 0, &space_info);
3248                 BUG_ON(ret);
3249         }
3250         BUG_ON(!space_info);
3251
3252 again:
3253         spin_lock(&space_info->lock);
3254         if (space_info->force_alloc)
3255                 force = space_info->force_alloc;
3256         if (space_info->full) {
3257                 spin_unlock(&space_info->lock);
3258                 return 0;
3259         }
3260
3261         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3262                 spin_unlock(&space_info->lock);
3263                 return 0;
3264         } else if (space_info->chunk_alloc) {
3265                 wait_for_alloc = 1;
3266         } else {
3267                 space_info->chunk_alloc = 1;
3268         }
3269
3270         spin_unlock(&space_info->lock);
3271
3272         mutex_lock(&fs_info->chunk_mutex);
3273
3274         /*
3275          * The chunk_mutex is held throughout the entirety of a chunk
3276          * allocation, so once we've acquired the chunk_mutex we know that the
3277          * other guy is done and we need to recheck and see if we should
3278          * allocate.
3279          */
3280         if (wait_for_alloc) {
3281                 mutex_unlock(&fs_info->chunk_mutex);
3282                 wait_for_alloc = 0;
3283                 goto again;
3284         }
3285
3286         /*
3287          * If we have mixed data/metadata chunks we want to make sure we keep
3288          * allocating mixed chunks instead of individual chunks.
3289          */
3290         if (btrfs_mixed_space_info(space_info))
3291                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3292
3293         /*
3294          * if we're doing a data chunk, go ahead and make sure that
3295          * we keep a reasonable number of metadata chunks allocated in the
3296          * FS as well.
3297          */
3298         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3299                 fs_info->data_chunk_allocations++;
3300                 if (!(fs_info->data_chunk_allocations %
3301                       fs_info->metadata_ratio))
3302                         force_metadata_allocation(fs_info);
3303         }
3304
3305         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3306         if (ret < 0 && ret != -ENOSPC)
3307                 goto out;
3308
3309         spin_lock(&space_info->lock);
3310         if (ret)
3311                 space_info->full = 1;
3312         else
3313                 ret = 1;
3314
3315         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3316         space_info->chunk_alloc = 0;
3317         spin_unlock(&space_info->lock);
3318 out:
3319         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3320         return ret;
3321 }
3322
3323 /*
3324  * shrink metadata reservation for delalloc
3325  */
3326 static int shrink_delalloc(struct btrfs_trans_handle *trans,
3327                            struct btrfs_root *root, u64 to_reclaim, int sync)
3328 {
3329         struct btrfs_block_rsv *block_rsv;
3330         struct btrfs_space_info *space_info;
3331         u64 reserved;
3332         u64 max_reclaim;
3333         u64 reclaimed = 0;
3334         long time_left;
3335         int nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3336         int loops = 0;
3337         unsigned long progress;
3338
3339         block_rsv = &root->fs_info->delalloc_block_rsv;
3340         space_info = block_rsv->space_info;
3341
3342         smp_mb();
3343         reserved = space_info->bytes_may_use;
3344         progress = space_info->reservation_progress;
3345
3346         if (reserved == 0)
3347                 return 0;
3348
3349         smp_mb();
3350         if (root->fs_info->delalloc_bytes == 0) {
3351                 if (trans)
3352                         return 0;
3353                 btrfs_wait_ordered_extents(root, 0, 0);
3354                 return 0;
3355         }
3356
3357         max_reclaim = min(reserved, to_reclaim);
3358
3359         while (loops < 1024) {
3360                 /* have the flusher threads jump in and do some IO */
3361                 smp_mb();
3362                 nr_pages = min_t(unsigned long, nr_pages,
3363                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3364                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages);
3365
3366                 spin_lock(&space_info->lock);
3367                 if (reserved > space_info->bytes_may_use)
3368                         reclaimed += reserved - space_info->bytes_may_use;
3369                 reserved = space_info->bytes_may_use;
3370                 spin_unlock(&space_info->lock);
3371
3372                 loops++;
3373
3374                 if (reserved == 0 || reclaimed >= max_reclaim)
3375                         break;
3376
3377                 if (trans && trans->transaction->blocked)
3378                         return -EAGAIN;
3379
3380                 time_left = schedule_timeout_interruptible(1);
3381
3382                 /* We were interrupted, exit */
3383                 if (time_left)
3384                         break;
3385
3386                 /* we've kicked the IO a few times, if anything has been freed,
3387                  * exit.  There is no sense in looping here for a long time
3388                  * when we really need to commit the transaction, or there are
3389                  * just too many writers without enough free space
3390                  */
3391
3392                 if (loops > 3) {
3393                         smp_mb();
3394                         if (progress != space_info->reservation_progress)
3395                                 break;
3396                 }
3397
3398         }
3399         if (reclaimed >= to_reclaim && !trans)
3400                 btrfs_wait_ordered_extents(root, 0, 0);
3401         return reclaimed >= to_reclaim;
3402 }
3403
3404 /*
3405  * Retries tells us how many times we've called reserve_metadata_bytes.  The
3406  * idea is if this is the first call (retries == 0) then we will add to our
3407  * reserved count if we can't make the allocation in order to hold our place
3408  * while we go and try and free up space.  That way for retries > 1 we don't try
3409  * and add space, we just check to see if the amount of unused space is >= the
3410  * total space, meaning that our reservation is valid.
3411  *
3412  * However if we don't intend to retry this reservation, pass -1 as retries so
3413  * that it short circuits this logic.
3414  */
3415 static int reserve_metadata_bytes(struct btrfs_trans_handle *trans,
3416                                   struct btrfs_root *root,
3417                                   struct btrfs_block_rsv *block_rsv,
3418                                   u64 orig_bytes, int flush)
3419 {
3420         struct btrfs_space_info *space_info = block_rsv->space_info;
3421         u64 unused;
3422         u64 num_bytes = orig_bytes;
3423         int retries = 0;
3424         int ret = 0;
3425         bool committed = false;
3426         bool flushing = false;
3427 again:
3428         ret = 0;
3429         spin_lock(&space_info->lock);
3430         /*
3431          * We only want to wait if somebody other than us is flushing and we are
3432          * actually alloed to flush.
3433          */
3434         while (flush && !flushing && space_info->flush) {
3435                 spin_unlock(&space_info->lock);
3436                 /*
3437                  * If we have a trans handle we can't wait because the flusher
3438                  * may have to commit the transaction, which would mean we would
3439                  * deadlock since we are waiting for the flusher to finish, but
3440                  * hold the current transaction open.
3441                  */
3442                 if (trans)
3443                         return -EAGAIN;
3444                 ret = wait_event_interruptible(space_info->wait,
3445                                                !space_info->flush);
3446                 /* Must have been interrupted, return */
3447                 if (ret)
3448                         return -EINTR;
3449
3450                 spin_lock(&space_info->lock);
3451         }
3452
3453         ret = -ENOSPC;
3454         unused = space_info->bytes_used + space_info->bytes_reserved +
3455                  space_info->bytes_pinned + space_info->bytes_readonly +
3456                  space_info->bytes_may_use;
3457
3458         /*
3459          * The idea here is that we've not already over-reserved the block group
3460          * then we can go ahead and save our reservation first and then start
3461          * flushing if we need to.  Otherwise if we've already overcommitted
3462          * lets start flushing stuff first and then come back and try to make
3463          * our reservation.
3464          */
3465         if (unused <= space_info->total_bytes) {
3466                 unused = space_info->total_bytes - unused;
3467                 if (unused >= num_bytes) {
3468                         space_info->bytes_may_use += orig_bytes;
3469                         ret = 0;
3470                 } else {
3471                         /*
3472                          * Ok set num_bytes to orig_bytes since we aren't
3473                          * overocmmitted, this way we only try and reclaim what
3474                          * we need.
3475                          */
3476                         num_bytes = orig_bytes;
3477                 }
3478         } else {
3479                 /*
3480                  * Ok we're over committed, set num_bytes to the overcommitted
3481                  * amount plus the amount of bytes that we need for this
3482                  * reservation.
3483                  */
3484                 num_bytes = unused - space_info->total_bytes +
3485                         (orig_bytes * (retries + 1));
3486         }
3487
3488         /*
3489          * Couldn't make our reservation, save our place so while we're trying
3490          * to reclaim space we can actually use it instead of somebody else
3491          * stealing it from us.
3492          */
3493         if (ret && flush) {
3494                 flushing = true;
3495                 space_info->flush = 1;
3496         }
3497
3498         spin_unlock(&space_info->lock);
3499
3500         if (!ret || !flush)
3501                 goto out;
3502
3503         /*
3504          * We do synchronous shrinking since we don't actually unreserve
3505          * metadata until after the IO is completed.
3506          */
3507         ret = shrink_delalloc(trans, root, num_bytes, 1);
3508         if (ret < 0)
3509                 goto out;
3510
3511         ret = 0;
3512
3513         /*
3514          * So if we were overcommitted it's possible that somebody else flushed
3515          * out enough space and we simply didn't have enough space to reclaim,
3516          * so go back around and try again.
3517          */
3518         if (retries < 2) {
3519                 retries++;
3520                 goto again;
3521         }
3522
3523         /*
3524          * Not enough space to be reclaimed, don't bother committing the
3525          * transaction.
3526          */
3527         spin_lock(&space_info->lock);
3528         if (space_info->bytes_pinned < orig_bytes)
3529                 ret = -ENOSPC;
3530         spin_unlock(&space_info->lock);
3531         if (ret)
3532                 goto out;
3533
3534         ret = -EAGAIN;
3535         if (trans)
3536                 goto out;
3537
3538         ret = -ENOSPC;
3539         if (committed)
3540                 goto out;
3541
3542         trans = btrfs_join_transaction(root);
3543         if (IS_ERR(trans))
3544                 goto out;
3545         ret = btrfs_commit_transaction(trans, root);
3546         if (!ret) {
3547                 trans = NULL;
3548                 committed = true;
3549                 goto again;
3550         }
3551
3552 out:
3553         if (flushing) {
3554                 spin_lock(&space_info->lock);
3555                 space_info->flush = 0;
3556                 wake_up_all(&space_info->wait);
3557                 spin_unlock(&space_info->lock);
3558         }
3559         return ret;
3560 }
3561
3562 static struct btrfs_block_rsv *get_block_rsv(struct btrfs_trans_handle *trans,
3563                                              struct btrfs_root *root)
3564 {
3565         struct btrfs_block_rsv *block_rsv;
3566         if (root->ref_cows)
3567                 block_rsv = trans->block_rsv;
3568         else
3569                 block_rsv = root->block_rsv;
3570
3571         if (!block_rsv)
3572                 block_rsv = &root->fs_info->empty_block_rsv;
3573
3574         return block_rsv;
3575 }
3576
3577 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3578                                u64 num_bytes)
3579 {
3580         int ret = -ENOSPC;
3581         spin_lock(&block_rsv->lock);
3582         if (block_rsv->reserved >= num_bytes) {
3583                 block_rsv->reserved -= num_bytes;
3584                 if (block_rsv->reserved < block_rsv->size)
3585                         block_rsv->full = 0;
3586                 ret = 0;
3587         }
3588         spin_unlock(&block_rsv->lock);
3589         return ret;
3590 }
3591
3592 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3593                                 u64 num_bytes, int update_size)
3594 {
3595         spin_lock(&block_rsv->lock);
3596         block_rsv->reserved += num_bytes;
3597         if (update_size)
3598                 block_rsv->size += num_bytes;
3599         else if (block_rsv->reserved >= block_rsv->size)
3600                 block_rsv->full = 1;
3601         spin_unlock(&block_rsv->lock);
3602 }
3603
3604 static void block_rsv_release_bytes(struct btrfs_block_rsv *block_rsv,
3605                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3606 {
3607         struct btrfs_space_info *space_info = block_rsv->space_info;
3608
3609         spin_lock(&block_rsv->lock);
3610         if (num_bytes == (u64)-1)
3611                 num_bytes = block_rsv->size;
3612         block_rsv->size -= num_bytes;
3613         if (block_rsv->reserved >= block_rsv->size) {
3614                 num_bytes = block_rsv->reserved - block_rsv->size;
3615                 block_rsv->reserved = block_rsv->size;
3616                 block_rsv->full = 1;
3617         } else {
3618                 num_bytes = 0;
3619         }
3620         spin_unlock(&block_rsv->lock);
3621
3622         if (num_bytes > 0) {
3623                 if (dest) {
3624                         spin_lock(&dest->lock);
3625                         if (!dest->full) {
3626                                 u64 bytes_to_add;
3627
3628                                 bytes_to_add = dest->size - dest->reserved;
3629                                 bytes_to_add = min(num_bytes, bytes_to_add);
3630                                 dest->reserved += bytes_to_add;
3631                                 if (dest->reserved >= dest->size)
3632                                         dest->full = 1;
3633                                 num_bytes -= bytes_to_add;
3634                         }
3635                         spin_unlock(&dest->lock);
3636                 }
3637                 if (num_bytes) {
3638                         spin_lock(&space_info->lock);
3639                         space_info->bytes_may_use -= num_bytes;
3640                         space_info->reservation_progress++;
3641                         spin_unlock(&space_info->lock);
3642                 }
3643         }
3644 }
3645
3646 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3647                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3648 {
3649         int ret;
3650
3651         ret = block_rsv_use_bytes(src, num_bytes);
3652         if (ret)
3653                 return ret;
3654
3655         block_rsv_add_bytes(dst, num_bytes, 1);
3656         return 0;
3657 }
3658
3659 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3660 {
3661         memset(rsv, 0, sizeof(*rsv));
3662         spin_lock_init(&rsv->lock);
3663         atomic_set(&rsv->usage, 1);
3664         rsv->priority = 6;
3665         INIT_LIST_HEAD(&rsv->list);
3666 }
3667
3668 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3669 {
3670         struct btrfs_block_rsv *block_rsv;
3671         struct btrfs_fs_info *fs_info = root->fs_info;
3672
3673         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3674         if (!block_rsv)
3675                 return NULL;
3676
3677         btrfs_init_block_rsv(block_rsv);
3678         block_rsv->space_info = __find_space_info(fs_info,
3679                                                   BTRFS_BLOCK_GROUP_METADATA);
3680         return block_rsv;
3681 }
3682
3683 void btrfs_free_block_rsv(struct btrfs_root *root,
3684                           struct btrfs_block_rsv *rsv)
3685 {
3686         if (rsv && atomic_dec_and_test(&rsv->usage)) {
3687                 btrfs_block_rsv_release(root, rsv, (u64)-1);
3688                 if (!rsv->durable)
3689                         kfree(rsv);
3690         }
3691 }
3692
3693 /*
3694  * make the block_rsv struct be able to capture freed space.
3695  * the captured space will re-add to the the block_rsv struct
3696  * after transaction commit
3697  */
3698 void btrfs_add_durable_block_rsv(struct btrfs_fs_info *fs_info,
3699                                  struct btrfs_block_rsv *block_rsv)
3700 {
3701         block_rsv->durable = 1;
3702         mutex_lock(&fs_info->durable_block_rsv_mutex);
3703         list_add_tail(&block_rsv->list, &fs_info->durable_block_rsv_list);
3704         mutex_unlock(&fs_info->durable_block_rsv_mutex);
3705 }
3706
3707 int btrfs_block_rsv_add(struct btrfs_trans_handle *trans,
3708                         struct btrfs_root *root,
3709                         struct btrfs_block_rsv *block_rsv,
3710                         u64 num_bytes)
3711 {
3712         int ret;
3713
3714         if (num_bytes == 0)
3715                 return 0;
3716
3717         ret = reserve_metadata_bytes(trans, root, block_rsv, num_bytes, 1);
3718         if (!ret) {
3719                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3720                 return 0;
3721         }
3722
3723         return ret;
3724 }
3725
3726 int btrfs_block_rsv_check(struct btrfs_trans_handle *trans,
3727                           struct btrfs_root *root,
3728                           struct btrfs_block_rsv *block_rsv,
3729                           u64 min_reserved, int min_factor)
3730 {
3731         u64 num_bytes = 0;
3732         int commit_trans = 0;
3733         int ret = -ENOSPC;
3734
3735         if (!block_rsv)
3736                 return 0;
3737
3738         spin_lock(&block_rsv->lock);
3739         if (min_factor > 0)
3740                 num_bytes = div_factor(block_rsv->size, min_factor);
3741         if (min_reserved > num_bytes)
3742                 num_bytes = min_reserved;
3743
3744         if (block_rsv->reserved >= num_bytes) {
3745                 ret = 0;
3746         } else {
3747                 num_bytes -= block_rsv->reserved;
3748                 if (block_rsv->durable &&
3749                     block_rsv->freed[0] + block_rsv->freed[1] >= num_bytes)
3750                         commit_trans = 1;
3751         }
3752         spin_unlock(&block_rsv->lock);
3753         if (!ret)
3754                 return 0;
3755
3756         if (block_rsv->refill_used) {
3757                 ret = reserve_metadata_bytes(trans, root, block_rsv,
3758                                              num_bytes, 0);
3759                 if (!ret) {
3760                         block_rsv_add_bytes(block_rsv, num_bytes, 0);
3761                         return 0;
3762                 }
3763         }
3764
3765         if (commit_trans) {
3766                 if (trans)
3767                         return -EAGAIN;
3768                 trans = btrfs_join_transaction(root);
3769                 BUG_ON(IS_ERR(trans));
3770                 ret = btrfs_commit_transaction(trans, root);
3771                 return 0;
3772         }
3773
3774         return -ENOSPC;
3775 }
3776
3777 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
3778                             struct btrfs_block_rsv *dst_rsv,
3779                             u64 num_bytes)
3780 {
3781         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3782 }
3783
3784 void btrfs_block_rsv_release(struct btrfs_root *root,
3785                              struct btrfs_block_rsv *block_rsv,
3786                              u64 num_bytes)
3787 {
3788         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3789         if (global_rsv->full || global_rsv == block_rsv ||
3790             block_rsv->space_info != global_rsv->space_info)
3791                 global_rsv = NULL;
3792         block_rsv_release_bytes(block_rsv, global_rsv, num_bytes);
3793 }
3794
3795 /*
3796  * helper to calculate size of global block reservation.
3797  * the desired value is sum of space used by extent tree,
3798  * checksum tree and root tree
3799  */
3800 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
3801 {
3802         struct btrfs_space_info *sinfo;
3803         u64 num_bytes;
3804         u64 meta_used;
3805         u64 data_used;
3806         int csum_size = btrfs_super_csum_size(&fs_info->super_copy);
3807
3808         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
3809         spin_lock(&sinfo->lock);
3810         data_used = sinfo->bytes_used;
3811         spin_unlock(&sinfo->lock);
3812
3813         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3814         spin_lock(&sinfo->lock);
3815         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
3816                 data_used = 0;
3817         meta_used = sinfo->bytes_used;
3818         spin_unlock(&sinfo->lock);
3819
3820         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
3821                     csum_size * 2;
3822         num_bytes += div64_u64(data_used + meta_used, 50);
3823
3824         if (num_bytes * 3 > meta_used)
3825                 num_bytes = div64_u64(meta_used, 3);
3826
3827         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
3828 }
3829
3830 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
3831 {
3832         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
3833         struct btrfs_space_info *sinfo = block_rsv->space_info;
3834         u64 num_bytes;
3835
3836         num_bytes = calc_global_metadata_size(fs_info);
3837
3838         spin_lock(&block_rsv->lock);
3839         spin_lock(&sinfo->lock);
3840
3841         block_rsv->size = num_bytes;
3842
3843         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
3844                     sinfo->bytes_reserved + sinfo->bytes_readonly +
3845                     sinfo->bytes_may_use;
3846
3847         if (sinfo->total_bytes > num_bytes) {
3848                 num_bytes = sinfo->total_bytes - num_bytes;
3849                 block_rsv->reserved += num_bytes;
3850                 sinfo->bytes_may_use += num_bytes;
3851         }
3852
3853         if (block_rsv->reserved >= block_rsv->size) {
3854                 num_bytes = block_rsv->reserved - block_rsv->size;
3855                 sinfo->bytes_may_use -= num_bytes;
3856                 sinfo->reservation_progress++;
3857                 block_rsv->reserved = block_rsv->size;
3858                 block_rsv->full = 1;
3859         }
3860
3861         spin_unlock(&sinfo->lock);
3862         spin_unlock(&block_rsv->lock);
3863 }
3864
3865 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
3866 {
3867         struct btrfs_space_info *space_info;
3868
3869         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3870         fs_info->chunk_block_rsv.space_info = space_info;
3871         fs_info->chunk_block_rsv.priority = 10;
3872
3873         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3874         fs_info->global_block_rsv.space_info = space_info;
3875         fs_info->global_block_rsv.priority = 10;
3876         fs_info->global_block_rsv.refill_used = 1;
3877         fs_info->delalloc_block_rsv.space_info = space_info;
3878         fs_info->trans_block_rsv.space_info = space_info;
3879         fs_info->empty_block_rsv.space_info = space_info;
3880         fs_info->empty_block_rsv.priority = 10;
3881
3882         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
3883         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
3884         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
3885         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
3886         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
3887
3888         btrfs_add_durable_block_rsv(fs_info, &fs_info->global_block_rsv);
3889
3890         btrfs_add_durable_block_rsv(fs_info, &fs_info->delalloc_block_rsv);
3891
3892         update_global_block_rsv(fs_info);
3893 }
3894
3895 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
3896 {
3897         block_rsv_release_bytes(&fs_info->global_block_rsv, NULL, (u64)-1);
3898         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
3899         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
3900         WARN_ON(fs_info->trans_block_rsv.size > 0);
3901         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
3902         WARN_ON(fs_info->chunk_block_rsv.size > 0);
3903         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
3904 }
3905
3906 int btrfs_truncate_reserve_metadata(struct btrfs_trans_handle *trans,
3907                                     struct btrfs_root *root,
3908                                     struct btrfs_block_rsv *rsv)
3909 {
3910         struct btrfs_block_rsv *trans_rsv = &root->fs_info->trans_block_rsv;
3911         u64 num_bytes;
3912         int ret;
3913
3914         /*
3915          * Truncate should be freeing data, but give us 2 items just in case it
3916          * needs to use some space.  We may want to be smarter about this in the
3917          * future.
3918          */
3919         num_bytes = btrfs_calc_trans_metadata_size(root, 2);
3920
3921         /* We already have enough bytes, just return */
3922         if (rsv->reserved >= num_bytes)
3923                 return 0;
3924
3925         num_bytes -= rsv->reserved;
3926
3927         /*
3928          * You should have reserved enough space before hand to do this, so this
3929          * should not fail.
3930          */
3931         ret = block_rsv_migrate_bytes(trans_rsv, rsv, num_bytes);
3932         BUG_ON(ret);
3933
3934         return 0;
3935 }
3936
3937 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
3938                                   struct btrfs_root *root)
3939 {
3940         if (!trans->bytes_reserved)
3941                 return;
3942
3943         BUG_ON(trans->block_rsv != &root->fs_info->trans_block_rsv);
3944         btrfs_block_rsv_release(root, trans->block_rsv,
3945                                 trans->bytes_reserved);
3946         trans->bytes_reserved = 0;
3947 }
3948
3949 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
3950                                   struct inode *inode)
3951 {
3952         struct btrfs_root *root = BTRFS_I(inode)->root;
3953         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
3954         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
3955
3956         /*
3957          * We need to hold space in order to delete our orphan item once we've
3958          * added it, so this takes the reservation so we can release it later
3959          * when we are truly done with the orphan item.
3960          */
3961         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
3962         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3963 }
3964
3965 void btrfs_orphan_release_metadata(struct inode *inode)
3966 {
3967         struct btrfs_root *root = BTRFS_I(inode)->root;
3968         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
3969         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
3970 }
3971
3972 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
3973                                 struct btrfs_pending_snapshot *pending)
3974 {
3975         struct btrfs_root *root = pending->root;
3976         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
3977         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
3978         /*
3979          * two for root back/forward refs, two for directory entries
3980          * and one for root of the snapshot.
3981          */
3982         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
3983         dst_rsv->space_info = src_rsv->space_info;
3984         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3985 }
3986
3987 /**
3988  * drop_outstanding_extent - drop an outstanding extent
3989  * @inode: the inode we're dropping the extent for
3990  *
3991  * This is called when we are freeing up an outstanding extent, either called
3992  * after an error or after an extent is written.  This will return the number of
3993  * reserved extents that need to be freed.  This must be called with
3994  * BTRFS_I(inode)->lock held.
3995  */
3996 static unsigned drop_outstanding_extent(struct inode *inode)
3997 {
3998         unsigned dropped_extents = 0;
3999
4000         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4001         BTRFS_I(inode)->outstanding_extents--;
4002
4003         /*
4004          * If we have more or the same amount of outsanding extents than we have
4005          * reserved then we need to leave the reserved extents count alone.
4006          */
4007         if (BTRFS_I(inode)->outstanding_extents >=
4008             BTRFS_I(inode)->reserved_extents)
4009                 return 0;
4010
4011         dropped_extents = BTRFS_I(inode)->reserved_extents -
4012                 BTRFS_I(inode)->outstanding_extents;
4013         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4014         return dropped_extents;
4015 }
4016
4017 /**
4018  * calc_csum_metadata_size - return the amount of metada space that must be
4019  *      reserved/free'd for the given bytes.
4020  * @inode: the inode we're manipulating
4021  * @num_bytes: the number of bytes in question
4022  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4023  *
4024  * This adjusts the number of csum_bytes in the inode and then returns the
4025  * correct amount of metadata that must either be reserved or freed.  We
4026  * calculate how many checksums we can fit into one leaf and then divide the
4027  * number of bytes that will need to be checksumed by this value to figure out
4028  * how many checksums will be required.  If we are adding bytes then the number
4029  * may go up and we will return the number of additional bytes that must be
4030  * reserved.  If it is going down we will return the number of bytes that must
4031  * be freed.
4032  *
4033  * This must be called with BTRFS_I(inode)->lock held.
4034  */
4035 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4036                                    int reserve)
4037 {
4038         struct btrfs_root *root = BTRFS_I(inode)->root;
4039         u64 csum_size;
4040         int num_csums_per_leaf;
4041         int num_csums;
4042         int old_csums;
4043
4044         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4045             BTRFS_I(inode)->csum_bytes == 0)
4046                 return 0;
4047
4048         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4049         if (reserve)
4050                 BTRFS_I(inode)->csum_bytes += num_bytes;
4051         else
4052                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4053         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4054         num_csums_per_leaf = (int)div64_u64(csum_size,
4055                                             sizeof(struct btrfs_csum_item) +
4056                                             sizeof(struct btrfs_disk_key));
4057         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4058         num_csums = num_csums + num_csums_per_leaf - 1;
4059         num_csums = num_csums / num_csums_per_leaf;
4060
4061         old_csums = old_csums + num_csums_per_leaf - 1;
4062         old_csums = old_csums / num_csums_per_leaf;
4063
4064         /* No change, no need to reserve more */
4065         if (old_csums == num_csums)
4066                 return 0;
4067
4068         if (reserve)
4069                 return btrfs_calc_trans_metadata_size(root,
4070                                                       num_csums - old_csums);
4071
4072         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4073 }
4074
4075 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4076 {
4077         struct btrfs_root *root = BTRFS_I(inode)->root;
4078         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4079         u64 to_reserve = 0;
4080         unsigned nr_extents = 0;
4081         int ret;
4082
4083         if (btrfs_transaction_in_commit(root->fs_info))
4084                 schedule_timeout(1);
4085
4086         num_bytes = ALIGN(num_bytes, root->sectorsize);
4087
4088         spin_lock(&BTRFS_I(inode)->lock);
4089         BTRFS_I(inode)->outstanding_extents++;
4090
4091         if (BTRFS_I(inode)->outstanding_extents >
4092             BTRFS_I(inode)->reserved_extents) {
4093                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4094                         BTRFS_I(inode)->reserved_extents;
4095                 BTRFS_I(inode)->reserved_extents += nr_extents;
4096
4097                 to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4098         }
4099         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4100         spin_unlock(&BTRFS_I(inode)->lock);
4101
4102         ret = reserve_metadata_bytes(NULL, root, block_rsv, to_reserve, 1);
4103         if (ret) {
4104                 unsigned dropped;
4105                 /*
4106                  * We don't need the return value since our reservation failed,
4107                  * we just need to clean up our counter.
4108                  */
4109                 spin_lock(&BTRFS_I(inode)->lock);
4110                 dropped = drop_outstanding_extent(inode);
4111                 WARN_ON(dropped > 1);
4112                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4113                 spin_unlock(&BTRFS_I(inode)->lock);
4114                 return ret;
4115         }
4116
4117         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4118
4119         return 0;
4120 }
4121
4122 /**
4123  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4124  * @inode: the inode to release the reservation for
4125  * @num_bytes: the number of bytes we're releasing
4126  *
4127  * This will release the metadata reservation for an inode.  This can be called
4128  * once we complete IO for a given set of bytes to release their metadata
4129  * reservations.
4130  */
4131 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4132 {
4133         struct btrfs_root *root = BTRFS_I(inode)->root;
4134         u64 to_free = 0;
4135         unsigned dropped;
4136
4137         num_bytes = ALIGN(num_bytes, root->sectorsize);
4138         spin_lock(&BTRFS_I(inode)->lock);
4139         dropped = drop_outstanding_extent(inode);
4140
4141         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4142         spin_unlock(&BTRFS_I(inode)->lock);
4143         if (dropped > 0)
4144                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4145
4146         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4147                                 to_free);
4148 }
4149
4150 /**
4151  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4152  * @inode: inode we're writing to
4153  * @num_bytes: the number of bytes we want to allocate
4154  *
4155  * This will do the following things
4156  *
4157  * o reserve space in the data space info for num_bytes
4158  * o reserve space in the metadata space info based on number of outstanding
4159  *   extents and how much csums will be needed
4160  * o add to the inodes ->delalloc_bytes
4161  * o add it to the fs_info's delalloc inodes list.
4162  *
4163  * This will return 0 for success and -ENOSPC if there is no space left.
4164  */
4165 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4166 {
4167         int ret;
4168
4169         ret = btrfs_check_data_free_space(inode, num_bytes);
4170         if (ret)
4171                 return ret;
4172
4173         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4174         if (ret) {
4175                 btrfs_free_reserved_data_space(inode, num_bytes);
4176                 return ret;
4177         }
4178
4179         return 0;
4180 }
4181
4182 /**
4183  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4184  * @inode: inode we're releasing space for
4185  * @num_bytes: the number of bytes we want to free up
4186  *
4187  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4188  * called in the case that we don't need the metadata AND data reservations
4189  * anymore.  So if there is an error or we insert an inline extent.
4190  *
4191  * This function will release the metadata space that was not used and will
4192  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4193  * list if there are no delalloc bytes left.
4194  */
4195 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4196 {
4197         btrfs_delalloc_release_metadata(inode, num_bytes);
4198         btrfs_free_reserved_data_space(inode, num_bytes);
4199 }
4200
4201 static int update_block_group(struct btrfs_trans_handle *trans,
4202                               struct btrfs_root *root,
4203                               u64 bytenr, u64 num_bytes, int alloc)
4204 {
4205         struct btrfs_block_group_cache *cache = NULL;
4206         struct btrfs_fs_info *info = root->fs_info;
4207         u64 total = num_bytes;
4208         u64 old_val;
4209         u64 byte_in_group;
4210         int factor;
4211
4212         /* block accounting for super block */
4213         spin_lock(&info->delalloc_lock);
4214         old_val = btrfs_super_bytes_used(&info->super_copy);
4215         if (alloc)
4216                 old_val += num_bytes;
4217         else
4218                 old_val -= num_bytes;
4219         btrfs_set_super_bytes_used(&info->super_copy, old_val);
4220         spin_unlock(&info->delalloc_lock);
4221
4222         while (total) {
4223                 cache = btrfs_lookup_block_group(info, bytenr);
4224                 if (!cache)
4225                         return -1;
4226                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4227                                     BTRFS_BLOCK_GROUP_RAID1 |
4228                                     BTRFS_BLOCK_GROUP_RAID10))
4229                         factor = 2;
4230                 else
4231                         factor = 1;
4232                 /*
4233                  * If this block group has free space cache written out, we
4234                  * need to make sure to load it if we are removing space.  This
4235                  * is because we need the unpinning stage to actually add the
4236                  * space back to the block group, otherwise we will leak space.
4237                  */
4238                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4239                         cache_block_group(cache, trans, NULL, 1);
4240
4241                 byte_in_group = bytenr - cache->key.objectid;
4242                 WARN_ON(byte_in_group > cache->key.offset);
4243
4244                 spin_lock(&cache->space_info->lock);
4245                 spin_lock(&cache->lock);
4246
4247                 if (btrfs_super_cache_generation(&info->super_copy) != 0 &&
4248                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4249                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4250
4251                 cache->dirty = 1;
4252                 old_val = btrfs_block_group_used(&cache->item);
4253                 num_bytes = min(total, cache->key.offset - byte_in_group);
4254                 if (alloc) {
4255                         old_val += num_bytes;
4256                         btrfs_set_block_group_used(&cache->item, old_val);
4257                         cache->reserved -= num_bytes;
4258                         cache->space_info->bytes_reserved -= num_bytes;
4259                         cache->space_info->bytes_used += num_bytes;
4260                         cache->space_info->disk_used += num_bytes * factor;
4261                         spin_unlock(&cache->lock);
4262                         spin_unlock(&cache->space_info->lock);
4263                 } else {
4264                         old_val -= num_bytes;
4265                         btrfs_set_block_group_used(&cache->item, old_val);
4266                         cache->pinned += num_bytes;
4267                         cache->space_info->bytes_pinned += num_bytes;
4268                         cache->space_info->bytes_used -= num_bytes;
4269                         cache->space_info->disk_used -= num_bytes * factor;
4270                         spin_unlock(&cache->lock);
4271                         spin_unlock(&cache->space_info->lock);
4272
4273                         set_extent_dirty(info->pinned_extents,
4274                                          bytenr, bytenr + num_bytes - 1,
4275                                          GFP_NOFS | __GFP_NOFAIL);
4276                 }
4277                 btrfs_put_block_group(cache);
4278                 total -= num_bytes;
4279                 bytenr += num_bytes;
4280         }
4281         return 0;
4282 }
4283
4284 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4285 {
4286         struct btrfs_block_group_cache *cache;
4287         u64 bytenr;
4288
4289         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4290         if (!cache)
4291                 return 0;
4292
4293         bytenr = cache->key.objectid;
4294         btrfs_put_block_group(cache);
4295
4296         return bytenr;
4297 }
4298
4299 static int pin_down_extent(struct btrfs_root *root,
4300                            struct btrfs_block_group_cache *cache,
4301                            u64 bytenr, u64 num_bytes, int reserved)
4302 {
4303         spin_lock(&cache->space_info->lock);
4304         spin_lock(&cache->lock);
4305         cache->pinned += num_bytes;
4306         cache->space_info->bytes_pinned += num_bytes;
4307         if (reserved) {
4308                 cache->reserved -= num_bytes;
4309                 cache->space_info->bytes_reserved -= num_bytes;
4310         }
4311         spin_unlock(&cache->lock);
4312         spin_unlock(&cache->space_info->lock);
4313
4314         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4315                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4316         return 0;
4317 }
4318
4319 /*
4320  * this function must be called within transaction
4321  */
4322 int btrfs_pin_extent(struct btrfs_root *root,
4323                      u64 bytenr, u64 num_bytes, int reserved)
4324 {
4325         struct btrfs_block_group_cache *cache;
4326
4327         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4328         BUG_ON(!cache);
4329
4330         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4331
4332         btrfs_put_block_group(cache);
4333         return 0;
4334 }
4335
4336 /**
4337  * btrfs_update_reserved_bytes - update the block_group and space info counters
4338  * @cache:      The cache we are manipulating
4339  * @num_bytes:  The number of bytes in question
4340  * @reserve:    One of the reservation enums
4341  *
4342  * This is called by the allocator when it reserves space, or by somebody who is
4343  * freeing space that was never actually used on disk.  For example if you
4344  * reserve some space for a new leaf in transaction A and before transaction A
4345  * commits you free that leaf, you call this with reserve set to 0 in order to
4346  * clear the reservation.
4347  *
4348  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4349  * ENOSPC accounting.  For data we handle the reservation through clearing the
4350  * delalloc bits in the io_tree.  We have to do this since we could end up
4351  * allocating less disk space for the amount of data we have reserved in the
4352  * case of compression.
4353  *
4354  * If this is a reservation and the block group has become read only we cannot
4355  * make the reservation and return -EAGAIN, otherwise this function always
4356  * succeeds.
4357  */
4358 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4359                                        u64 num_bytes, int reserve)
4360 {
4361         struct btrfs_space_info *space_info = cache->space_info;
4362         int ret = 0;
4363         spin_lock(&space_info->lock);
4364         spin_lock(&cache->lock);
4365         if (reserve != RESERVE_FREE) {
4366                 if (cache->ro) {
4367                         ret = -EAGAIN;
4368                 } else {
4369                         cache->reserved += num_bytes;
4370                         space_info->bytes_reserved += num_bytes;
4371                         if (reserve == RESERVE_ALLOC) {
4372                                 BUG_ON(space_info->bytes_may_use < num_bytes);
4373                                 space_info->bytes_may_use -= num_bytes;
4374                         }
4375                 }
4376         } else {
4377                 if (cache->ro)
4378                         space_info->bytes_readonly += num_bytes;
4379                 cache->reserved -= num_bytes;
4380                 space_info->bytes_reserved -= num_bytes;
4381                 space_info->reservation_progress++;
4382         }
4383         spin_unlock(&cache->lock);
4384         spin_unlock(&space_info->lock);
4385         return ret;
4386 }
4387
4388 int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4389                                 struct btrfs_root *root)
4390 {
4391         struct btrfs_fs_info *fs_info = root->fs_info;
4392         struct btrfs_caching_control *next;
4393         struct btrfs_caching_control *caching_ctl;
4394         struct btrfs_block_group_cache *cache;
4395
4396         down_write(&fs_info->extent_commit_sem);
4397
4398         list_for_each_entry_safe(caching_ctl, next,
4399                                  &fs_info->caching_block_groups, list) {
4400                 cache = caching_ctl->block_group;
4401                 if (block_group_cache_done(cache)) {
4402                         cache->last_byte_to_unpin = (u64)-1;
4403                         list_del_init(&caching_ctl->list);
4404                         put_caching_control(caching_ctl);
4405                 } else {
4406                         cache->last_byte_to_unpin = caching_ctl->progress;
4407                 }
4408         }
4409
4410         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4411                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4412         else
4413                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4414
4415         up_write(&fs_info->extent_commit_sem);
4416
4417         update_global_block_rsv(fs_info);
4418         return 0;
4419 }
4420
4421 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4422 {
4423         struct btrfs_fs_info *fs_info = root->fs_info;
4424         struct btrfs_block_group_cache *cache = NULL;
4425         u64 len;
4426
4427         while (start <= end) {
4428                 if (!cache ||
4429                     start >= cache->key.objectid + cache->key.offset) {
4430                         if (cache)
4431                                 btrfs_put_block_group(cache);
4432                         cache = btrfs_lookup_block_group(fs_info, start);
4433                         BUG_ON(!cache);
4434                 }
4435
4436                 len = cache->key.objectid + cache->key.offset - start;
4437                 len = min(len, end + 1 - start);
4438
4439                 if (start < cache->last_byte_to_unpin) {
4440                         len = min(len, cache->last_byte_to_unpin - start);
4441                         btrfs_add_free_space(cache, start, len);
4442                 }
4443
4444                 start += len;
4445
4446                 spin_lock(&cache->space_info->lock);
4447                 spin_lock(&cache->lock);
4448                 cache->pinned -= len;
4449                 cache->space_info->bytes_pinned -= len;
4450                 if (cache->ro) {
4451                         cache->space_info->bytes_readonly += len;
4452                 } else if (cache->reserved_pinned > 0) {
4453                         len = min(len, cache->reserved_pinned);
4454                         cache->reserved_pinned -= len;
4455                         cache->space_info->bytes_may_use += len;
4456                 }
4457                 spin_unlock(&cache->lock);
4458                 spin_unlock(&cache->space_info->lock);
4459         }
4460
4461         if (cache)
4462                 btrfs_put_block_group(cache);
4463         return 0;
4464 }
4465
4466 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4467                                struct btrfs_root *root)
4468 {
4469         struct btrfs_fs_info *fs_info = root->fs_info;
4470         struct extent_io_tree *unpin;
4471         struct btrfs_block_rsv *block_rsv;
4472         struct btrfs_block_rsv *next_rsv;
4473         u64 start;
4474         u64 end;
4475         int idx;
4476         int ret;
4477
4478         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4479                 unpin = &fs_info->freed_extents[1];
4480         else
4481                 unpin = &fs_info->freed_extents[0];
4482
4483         while (1) {
4484                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4485                                             EXTENT_DIRTY);
4486                 if (ret)
4487                         break;
4488
4489                 if (btrfs_test_opt(root, DISCARD))
4490                         ret = btrfs_discard_extent(root, start,
4491                                                    end + 1 - start, NULL);
4492
4493                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4494                 unpin_extent_range(root, start, end);
4495                 cond_resched();
4496         }
4497
4498         mutex_lock(&fs_info->durable_block_rsv_mutex);
4499         list_for_each_entry_safe(block_rsv, next_rsv,
4500                                  &fs_info->durable_block_rsv_list, list) {
4501
4502                 idx = trans->transid & 0x1;
4503                 if (block_rsv->freed[idx] > 0) {
4504                         block_rsv_add_bytes(block_rsv,
4505                                             block_rsv->freed[idx], 0);
4506                         block_rsv->freed[idx] = 0;
4507                 }
4508                 if (atomic_read(&block_rsv->usage) == 0) {
4509                         btrfs_block_rsv_release(root, block_rsv, (u64)-1);
4510
4511                         if (block_rsv->freed[0] == 0 &&
4512                             block_rsv->freed[1] == 0) {
4513                                 list_del_init(&block_rsv->list);
4514                                 kfree(block_rsv);
4515                         }
4516                 } else {
4517                         btrfs_block_rsv_release(root, block_rsv, 0);
4518                 }
4519         }
4520         mutex_unlock(&fs_info->durable_block_rsv_mutex);
4521
4522         return 0;
4523 }
4524
4525 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4526                                 struct btrfs_root *root,
4527                                 u64 bytenr, u64 num_bytes, u64 parent,
4528                                 u64 root_objectid, u64 owner_objectid,
4529                                 u64 owner_offset, int refs_to_drop,
4530                                 struct btrfs_delayed_extent_op *extent_op)
4531 {
4532         struct btrfs_key key;
4533         struct btrfs_path *path;
4534         struct btrfs_fs_info *info = root->fs_info;
4535         struct btrfs_root *extent_root = info->extent_root;
4536         struct extent_buffer *leaf;
4537         struct btrfs_extent_item *ei;
4538         struct btrfs_extent_inline_ref *iref;
4539         int ret;
4540         int is_data;
4541         int extent_slot = 0;
4542         int found_extent = 0;
4543         int num_to_del = 1;
4544         u32 item_size;
4545         u64 refs;
4546
4547         path = btrfs_alloc_path();
4548         if (!path)
4549                 return -ENOMEM;
4550
4551         path->reada = 1;
4552         path->leave_spinning = 1;
4553
4554         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4555         BUG_ON(!is_data && refs_to_drop != 1);
4556
4557         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4558                                     bytenr, num_bytes, parent,
4559                                     root_objectid, owner_objectid,
4560                                     owner_offset);
4561         if (ret == 0) {
4562                 extent_slot = path->slots[0];
4563                 while (extent_slot >= 0) {
4564                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4565                                               extent_slot);
4566                         if (key.objectid != bytenr)
4567                                 break;
4568                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4569                             key.offset == num_bytes) {
4570                                 found_extent = 1;
4571                                 break;
4572                         }
4573                         if (path->slots[0] - extent_slot > 5)
4574                                 break;
4575                         extent_slot--;
4576                 }
4577 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4578                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4579                 if (found_extent && item_size < sizeof(*ei))
4580                         found_extent = 0;
4581 #endif
4582                 if (!found_extent) {
4583                         BUG_ON(iref);
4584                         ret = remove_extent_backref(trans, extent_root, path,
4585                                                     NULL, refs_to_drop,
4586                                                     is_data);
4587                         BUG_ON(ret);
4588                         btrfs_release_path(path);
4589                         path->leave_spinning = 1;
4590
4591                         key.objectid = bytenr;
4592                         key.type = BTRFS_EXTENT_ITEM_KEY;
4593                         key.offset = num_bytes;
4594
4595                         ret = btrfs_search_slot(trans, extent_root,
4596                                                 &key, path, -1, 1);
4597                         if (ret) {
4598                                 printk(KERN_ERR "umm, got %d back from search"
4599                                        ", was looking for %llu\n", ret,
4600                                        (unsigned long long)bytenr);
4601                                 if (ret > 0)
4602                                         btrfs_print_leaf(extent_root,
4603                                                          path->nodes[0]);
4604                         }
4605                         BUG_ON(ret);
4606                         extent_slot = path->slots[0];
4607                 }
4608         } else {
4609                 btrfs_print_leaf(extent_root, path->nodes[0]);
4610                 WARN_ON(1);
4611                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4612                        "parent %llu root %llu  owner %llu offset %llu\n",
4613                        (unsigned long long)bytenr,
4614                        (unsigned long long)parent,
4615                        (unsigned long long)root_objectid,
4616                        (unsigned long long)owner_objectid,
4617                        (unsigned long long)owner_offset);
4618         }
4619
4620         leaf = path->nodes[0];
4621         item_size = btrfs_item_size_nr(leaf, extent_slot);
4622 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4623         if (item_size < sizeof(*ei)) {
4624                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4625                 ret = convert_extent_item_v0(trans, extent_root, path,
4626                                              owner_objectid, 0);
4627                 BUG_ON(ret < 0);
4628
4629                 btrfs_release_path(path);
4630                 path->leave_spinning = 1;
4631
4632                 key.objectid = bytenr;
4633                 key.type = BTRFS_EXTENT_ITEM_KEY;
4634                 key.offset = num_bytes;
4635
4636                 ret = btrfs_search_slot(trans, extent_root, &key, path,
4637                                         -1, 1);
4638                 if (ret) {
4639                         printk(KERN_ERR "umm, got %d back from search"
4640                                ", was looking for %llu\n", ret,
4641                                (unsigned long long)bytenr);
4642                         btrfs_print_leaf(extent_root, path->nodes[0]);
4643                 }
4644                 BUG_ON(ret);
4645                 extent_slot = path->slots[0];
4646                 leaf = path->nodes[0];
4647                 item_size = btrfs_item_size_nr(leaf, extent_slot);
4648         }
4649 #endif
4650         BUG_ON(item_size < sizeof(*ei));
4651         ei = btrfs_item_ptr(leaf, extent_slot,
4652                             struct btrfs_extent_item);
4653         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
4654                 struct btrfs_tree_block_info *bi;
4655                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4656                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4657                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4658         }
4659
4660         refs = btrfs_extent_refs(leaf, ei);
4661         BUG_ON(refs < refs_to_drop);
4662         refs -= refs_to_drop;
4663
4664         if (refs > 0) {
4665                 if (extent_op)
4666                         __run_delayed_extent_op(extent_op, leaf, ei);
4667                 /*
4668                  * In the case of inline back ref, reference count will
4669                  * be updated by remove_extent_backref
4670                  */
4671                 if (iref) {
4672                         BUG_ON(!found_extent);
4673                 } else {
4674                         btrfs_set_extent_refs(leaf, ei, refs);
4675                         btrfs_mark_buffer_dirty(leaf);
4676                 }
4677                 if (found_extent) {
4678                         ret = remove_extent_backref(trans, extent_root, path,
4679                                                     iref, refs_to_drop,
4680                                                     is_data);
4681                         BUG_ON(ret);
4682                 }
4683         } else {
4684                 if (found_extent) {
4685                         BUG_ON(is_data && refs_to_drop !=
4686                                extent_data_ref_count(root, path, iref));
4687                         if (iref) {
4688                                 BUG_ON(path->slots[0] != extent_slot);
4689                         } else {
4690                                 BUG_ON(path->slots[0] != extent_slot + 1);
4691                                 path->slots[0] = extent_slot;
4692                                 num_to_del = 2;
4693                         }
4694                 }
4695
4696                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
4697                                       num_to_del);
4698                 BUG_ON(ret);
4699                 btrfs_release_path(path);
4700
4701                 if (is_data) {
4702                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
4703                         BUG_ON(ret);
4704                 } else {
4705                         invalidate_mapping_pages(info->btree_inode->i_mapping,
4706                              bytenr >> PAGE_CACHE_SHIFT,
4707                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
4708                 }
4709
4710                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
4711                 BUG_ON(ret);
4712         }
4713         btrfs_free_path(path);
4714         return ret;
4715 }
4716
4717 /*
4718  * when we free an block, it is possible (and likely) that we free the last
4719  * delayed ref for that extent as well.  This searches the delayed ref tree for
4720  * a given extent, and if there are no other delayed refs to be processed, it
4721  * removes it from the tree.
4722  */
4723 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
4724                                       struct btrfs_root *root, u64 bytenr)
4725 {
4726         struct btrfs_delayed_ref_head *head;
4727         struct btrfs_delayed_ref_root *delayed_refs;
4728         struct btrfs_delayed_ref_node *ref;
4729         struct rb_node *node;
4730         int ret = 0;
4731
4732         delayed_refs = &trans->transaction->delayed_refs;
4733         spin_lock(&delayed_refs->lock);
4734         head = btrfs_find_delayed_ref_head(trans, bytenr);
4735         if (!head)
4736                 goto out;
4737
4738         node = rb_prev(&head->node.rb_node);
4739         if (!node)
4740                 goto out;
4741
4742         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
4743
4744         /* there are still entries for this ref, we can't drop it */
4745         if (ref->bytenr == bytenr)
4746                 goto out;
4747
4748         if (head->extent_op) {
4749                 if (!head->must_insert_reserved)
4750                         goto out;
4751                 kfree(head->extent_op);
4752                 head->extent_op = NULL;
4753         }
4754
4755         /*
4756          * waiting for the lock here would deadlock.  If someone else has it
4757          * locked they are already in the process of dropping it anyway
4758          */
4759         if (!mutex_trylock(&head->mutex))
4760                 goto out;
4761
4762         /*
4763          * at this point we have a head with no other entries.  Go
4764          * ahead and process it.
4765          */
4766         head->node.in_tree = 0;
4767         rb_erase(&head->node.rb_node, &delayed_refs->root);
4768
4769         delayed_refs->num_entries--;
4770
4771         /*
4772          * we don't take a ref on the node because we're removing it from the
4773          * tree, so we just steal the ref the tree was holding.
4774          */
4775         delayed_refs->num_heads--;
4776         if (list_empty(&head->cluster))
4777                 delayed_refs->num_heads_ready--;
4778
4779         list_del_init(&head->cluster);
4780         spin_unlock(&delayed_refs->lock);
4781
4782         BUG_ON(head->extent_op);
4783         if (head->must_insert_reserved)
4784                 ret = 1;
4785
4786         mutex_unlock(&head->mutex);
4787         btrfs_put_delayed_ref(&head->node);
4788         return ret;
4789 out:
4790         spin_unlock(&delayed_refs->lock);
4791         return 0;
4792 }
4793
4794 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
4795                            struct btrfs_root *root,
4796                            struct extent_buffer *buf,
4797                            u64 parent, int last_ref)
4798 {
4799         struct btrfs_block_rsv *block_rsv;
4800         struct btrfs_block_group_cache *cache = NULL;
4801         int ret;
4802
4803         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4804                 ret = btrfs_add_delayed_tree_ref(trans, buf->start, buf->len,
4805                                                 parent, root->root_key.objectid,
4806                                                 btrfs_header_level(buf),
4807                                                 BTRFS_DROP_DELAYED_REF, NULL);
4808                 BUG_ON(ret);
4809         }
4810
4811         if (!last_ref)
4812                 return;
4813
4814         block_rsv = get_block_rsv(trans, root);
4815         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
4816         if (block_rsv->space_info != cache->space_info)
4817                 goto out;
4818
4819         if (btrfs_header_generation(buf) == trans->transid) {
4820                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4821                         ret = check_ref_cleanup(trans, root, buf->start);
4822                         if (!ret)
4823                                 goto pin;
4824                 }
4825
4826                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
4827                         pin_down_extent(root, cache, buf->start, buf->len, 1);
4828                         goto pin;
4829                 }
4830
4831                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
4832
4833                 btrfs_add_free_space(cache, buf->start, buf->len);
4834                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
4835
4836                 goto out;
4837         }
4838 pin:
4839         if (block_rsv->durable && !cache->ro) {
4840                 ret = 0;
4841                 spin_lock(&cache->lock);
4842                 if (!cache->ro) {
4843                         cache->reserved_pinned += buf->len;
4844                         ret = 1;
4845                 }
4846                 spin_unlock(&cache->lock);
4847
4848                 if (ret) {
4849                         spin_lock(&block_rsv->lock);
4850                         block_rsv->freed[trans->transid & 0x1] += buf->len;
4851                         spin_unlock(&block_rsv->lock);
4852                 }
4853         }
4854 out:
4855         /*
4856          * Deleting the buffer, clear the corrupt flag since it doesn't matter
4857          * anymore.
4858          */
4859         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
4860         btrfs_put_block_group(cache);
4861 }
4862
4863 int btrfs_free_extent(struct btrfs_trans_handle *trans,
4864                       struct btrfs_root *root,
4865                       u64 bytenr, u64 num_bytes, u64 parent,
4866                       u64 root_objectid, u64 owner, u64 offset)
4867 {
4868         int ret;
4869
4870         /*
4871          * tree log blocks never actually go into the extent allocation
4872          * tree, just update pinning info and exit early.
4873          */
4874         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
4875                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
4876                 /* unlocks the pinned mutex */
4877                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
4878                 ret = 0;
4879         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
4880                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
4881                                         parent, root_objectid, (int)owner,
4882                                         BTRFS_DROP_DELAYED_REF, NULL);
4883                 BUG_ON(ret);
4884         } else {
4885                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
4886                                         parent, root_objectid, owner,
4887                                         offset, BTRFS_DROP_DELAYED_REF, NULL);
4888                 BUG_ON(ret);
4889         }
4890         return ret;
4891 }
4892
4893 static u64 stripe_align(struct btrfs_root *root, u64 val)
4894 {
4895         u64 mask = ((u64)root->stripesize - 1);
4896         u64 ret = (val + mask) & ~mask;
4897         return ret;
4898 }
4899
4900 /*
4901  * when we wait for progress in the block group caching, its because
4902  * our allocation attempt failed at least once.  So, we must sleep
4903  * and let some progress happen before we try again.
4904  *
4905  * This function will sleep at least once waiting for new free space to
4906  * show up, and then it will check the block group free space numbers
4907  * for our min num_bytes.  Another option is to have it go ahead
4908  * and look in the rbtree for a free extent of a given size, but this
4909  * is a good start.
4910  */
4911 static noinline int
4912 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
4913                                 u64 num_bytes)
4914 {
4915         struct btrfs_caching_control *caching_ctl;
4916         DEFINE_WAIT(wait);
4917
4918         caching_ctl = get_caching_control(cache);
4919         if (!caching_ctl)
4920                 return 0;
4921
4922         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
4923                    (cache->free_space_ctl->free_space >= num_bytes));
4924
4925         put_caching_control(caching_ctl);
4926         return 0;
4927 }
4928
4929 static noinline int
4930 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
4931 {
4932         struct btrfs_caching_control *caching_ctl;
4933         DEFINE_WAIT(wait);
4934
4935         caching_ctl = get_caching_control(cache);
4936         if (!caching_ctl)
4937                 return 0;
4938
4939         wait_event(caching_ctl->wait, block_group_cache_done(cache));
4940
4941         put_caching_control(caching_ctl);
4942         return 0;
4943 }
4944
4945 static int get_block_group_index(struct btrfs_block_group_cache *cache)
4946 {
4947         int index;
4948         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
4949                 index = 0;
4950         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
4951                 index = 1;
4952         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
4953                 index = 2;
4954         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
4955                 index = 3;
4956         else
4957                 index = 4;
4958         return index;
4959 }
4960
4961 enum btrfs_loop_type {
4962         LOOP_FIND_IDEAL = 0,
4963         LOOP_CACHING_NOWAIT = 1,
4964         LOOP_CACHING_WAIT = 2,
4965         LOOP_ALLOC_CHUNK = 3,
4966         LOOP_NO_EMPTY_SIZE = 4,
4967 };
4968
4969 /*
4970  * walks the btree of allocated extents and find a hole of a given size.
4971  * The key ins is changed to record the hole:
4972  * ins->objectid == block start
4973  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4974  * ins->offset == number of blocks
4975  * Any available blocks before search_start are skipped.
4976  */
4977 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
4978                                      struct btrfs_root *orig_root,
4979                                      u64 num_bytes, u64 empty_size,
4980                                      u64 search_start, u64 search_end,
4981                                      u64 hint_byte, struct btrfs_key *ins,
4982                                      u64 data)
4983 {
4984         int ret = 0;
4985         struct btrfs_root *root = orig_root->fs_info->extent_root;
4986         struct btrfs_free_cluster *last_ptr = NULL;
4987         struct btrfs_block_group_cache *block_group = NULL;
4988         int empty_cluster = 2 * 1024 * 1024;
4989         int allowed_chunk_alloc = 0;
4990         int done_chunk_alloc = 0;
4991         struct btrfs_space_info *space_info;
4992         int last_ptr_loop = 0;
4993         int loop = 0;
4994         int index = 0;
4995         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
4996                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
4997         bool found_uncached_bg = false;
4998         bool failed_cluster_refill = false;
4999         bool failed_alloc = false;
5000         bool use_cluster = true;
5001         u64 ideal_cache_percent = 0;
5002         u64 ideal_cache_offset = 0;
5003
5004         WARN_ON(num_bytes < root->sectorsize);
5005         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5006         ins->objectid = 0;
5007         ins->offset = 0;
5008
5009         space_info = __find_space_info(root->fs_info, data);
5010         if (!space_info) {
5011                 printk(KERN_ERR "No space info for %llu\n", data);
5012                 return -ENOSPC;
5013         }
5014
5015         /*
5016          * If the space info is for both data and metadata it means we have a
5017          * small filesystem and we can't use the clustering stuff.
5018          */
5019         if (btrfs_mixed_space_info(space_info))
5020                 use_cluster = false;
5021
5022         if (orig_root->ref_cows || empty_size)
5023                 allowed_chunk_alloc = 1;
5024
5025         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5026                 last_ptr = &root->fs_info->meta_alloc_cluster;
5027                 if (!btrfs_test_opt(root, SSD))
5028                         empty_cluster = 64 * 1024;
5029         }
5030
5031         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5032             btrfs_test_opt(root, SSD)) {
5033                 last_ptr = &root->fs_info->data_alloc_cluster;
5034         }
5035
5036         if (last_ptr) {
5037                 spin_lock(&last_ptr->lock);
5038                 if (last_ptr->block_group)
5039                         hint_byte = last_ptr->window_start;
5040                 spin_unlock(&last_ptr->lock);
5041         }
5042
5043         search_start = max(search_start, first_logical_byte(root, 0));
5044         search_start = max(search_start, hint_byte);
5045
5046         if (!last_ptr)
5047                 empty_cluster = 0;
5048
5049         if (search_start == hint_byte) {
5050 ideal_cache:
5051                 block_group = btrfs_lookup_block_group(root->fs_info,
5052                                                        search_start);
5053                 /*
5054                  * we don't want to use the block group if it doesn't match our
5055                  * allocation bits, or if its not cached.
5056                  *
5057                  * However if we are re-searching with an ideal block group
5058                  * picked out then we don't care that the block group is cached.
5059                  */
5060                 if (block_group && block_group_bits(block_group, data) &&
5061                     (block_group->cached != BTRFS_CACHE_NO ||
5062                      search_start == ideal_cache_offset)) {
5063                         down_read(&space_info->groups_sem);
5064                         if (list_empty(&block_group->list) ||
5065                             block_group->ro) {
5066                                 /*
5067                                  * someone is removing this block group,
5068                                  * we can't jump into the have_block_group
5069                                  * target because our list pointers are not
5070                                  * valid
5071                                  */
5072                                 btrfs_put_block_group(block_group);
5073                                 up_read(&space_info->groups_sem);
5074                         } else {
5075                                 index = get_block_group_index(block_group);
5076                                 goto have_block_group;
5077                         }
5078                 } else if (block_group) {
5079                         btrfs_put_block_group(block_group);
5080                 }
5081         }
5082 search:
5083         down_read(&space_info->groups_sem);
5084         list_for_each_entry(block_group, &space_info->block_groups[index],
5085                             list) {
5086                 u64 offset;
5087                 int cached;
5088
5089                 btrfs_get_block_group(block_group);
5090                 search_start = block_group->key.objectid;
5091
5092                 /*
5093                  * this can happen if we end up cycling through all the
5094                  * raid types, but we want to make sure we only allocate
5095                  * for the proper type.
5096                  */
5097                 if (!block_group_bits(block_group, data)) {
5098                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5099                                 BTRFS_BLOCK_GROUP_RAID1 |
5100                                 BTRFS_BLOCK_GROUP_RAID10;
5101
5102                         /*
5103                          * if they asked for extra copies and this block group
5104                          * doesn't provide them, bail.  This does allow us to
5105                          * fill raid0 from raid1.
5106                          */
5107                         if ((data & extra) && !(block_group->flags & extra))
5108                                 goto loop;
5109                 }
5110
5111 have_block_group:
5112                 if (unlikely(block_group->cached == BTRFS_CACHE_NO)) {
5113                         u64 free_percent;
5114
5115                         ret = cache_block_group(block_group, trans,
5116                                                 orig_root, 1);
5117                         if (block_group->cached == BTRFS_CACHE_FINISHED)
5118                                 goto have_block_group;
5119
5120                         free_percent = btrfs_block_group_used(&block_group->item);
5121                         free_percent *= 100;
5122                         free_percent = div64_u64(free_percent,
5123                                                  block_group->key.offset);
5124                         free_percent = 100 - free_percent;
5125                         if (free_percent > ideal_cache_percent &&
5126                             likely(!block_group->ro)) {
5127                                 ideal_cache_offset = block_group->key.objectid;
5128                                 ideal_cache_percent = free_percent;
5129                         }
5130
5131                         /*
5132                          * The caching workers are limited to 2 threads, so we
5133                          * can queue as much work as we care to.
5134                          */
5135                         if (loop > LOOP_FIND_IDEAL) {
5136                                 ret = cache_block_group(block_group, trans,
5137                                                         orig_root, 0);
5138                                 BUG_ON(ret);
5139                         }
5140                         found_uncached_bg = true;
5141
5142                         /*
5143                          * If loop is set for cached only, try the next block
5144                          * group.
5145                          */
5146                         if (loop == LOOP_FIND_IDEAL)
5147                                 goto loop;
5148                 }
5149
5150                 cached = block_group_cache_done(block_group);
5151                 if (unlikely(!cached))
5152                         found_uncached_bg = true;
5153
5154                 if (unlikely(block_group->ro))
5155                         goto loop;
5156
5157                 spin_lock(&block_group->free_space_ctl->tree_lock);
5158                 if (cached &&
5159                     block_group->free_space_ctl->free_space <
5160                     num_bytes + empty_size) {
5161                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5162                         goto loop;
5163                 }
5164                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5165
5166                 /*
5167                  * Ok we want to try and use the cluster allocator, so lets look
5168                  * there, unless we are on LOOP_NO_EMPTY_SIZE, since we will
5169                  * have tried the cluster allocator plenty of times at this
5170                  * point and not have found anything, so we are likely way too
5171                  * fragmented for the clustering stuff to find anything, so lets
5172                  * just skip it and let the allocator find whatever block it can
5173                  * find
5174                  */
5175                 if (last_ptr && loop < LOOP_NO_EMPTY_SIZE) {
5176                         /*
5177                          * the refill lock keeps out other
5178                          * people trying to start a new cluster
5179                          */
5180                         spin_lock(&last_ptr->refill_lock);
5181                         if (last_ptr->block_group &&
5182                             (last_ptr->block_group->ro ||
5183                             !block_group_bits(last_ptr->block_group, data))) {
5184                                 offset = 0;
5185                                 goto refill_cluster;
5186                         }
5187
5188                         offset = btrfs_alloc_from_cluster(block_group, last_ptr,
5189                                                  num_bytes, search_start);
5190                         if (offset) {
5191                                 /* we have a block, we're done */
5192                                 spin_unlock(&last_ptr->refill_lock);
5193                                 goto checks;
5194                         }
5195
5196                         spin_lock(&last_ptr->lock);
5197                         /*
5198                          * whoops, this cluster doesn't actually point to
5199                          * this block group.  Get a ref on the block
5200                          * group is does point to and try again
5201                          */
5202                         if (!last_ptr_loop && last_ptr->block_group &&
5203                             last_ptr->block_group != block_group &&
5204                             index <=
5205                                  get_block_group_index(last_ptr->block_group)) {
5206
5207                                 btrfs_put_block_group(block_group);
5208                                 block_group = last_ptr->block_group;
5209                                 btrfs_get_block_group(block_group);
5210                                 spin_unlock(&last_ptr->lock);
5211                                 spin_unlock(&last_ptr->refill_lock);
5212
5213                                 last_ptr_loop = 1;
5214                                 search_start = block_group->key.objectid;
5215                                 /*
5216                                  * we know this block group is properly
5217                                  * in the list because
5218                                  * btrfs_remove_block_group, drops the
5219                                  * cluster before it removes the block
5220                                  * group from the list
5221                                  */
5222                                 goto have_block_group;
5223                         }
5224                         spin_unlock(&last_ptr->lock);
5225 refill_cluster:
5226                         /*
5227                          * this cluster didn't work out, free it and
5228                          * start over
5229                          */
5230                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5231
5232                         last_ptr_loop = 0;
5233
5234                         /* allocate a cluster in this block group */
5235                         ret = btrfs_find_space_cluster(trans, root,
5236                                                block_group, last_ptr,
5237                                                offset, num_bytes,
5238                                                empty_cluster + empty_size);
5239                         if (ret == 0) {
5240                                 /*
5241                                  * now pull our allocation out of this
5242                                  * cluster
5243                                  */
5244                                 offset = btrfs_alloc_from_cluster(block_group,
5245                                                   last_ptr, num_bytes,
5246                                                   search_start);
5247                                 if (offset) {
5248                                         /* we found one, proceed */
5249                                         spin_unlock(&last_ptr->refill_lock);
5250                                         goto checks;
5251                                 }
5252                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5253                                    && !failed_cluster_refill) {
5254                                 spin_unlock(&last_ptr->refill_lock);
5255
5256                                 failed_cluster_refill = true;
5257                                 wait_block_group_cache_progress(block_group,
5258                                        num_bytes + empty_cluster + empty_size);
5259                                 goto have_block_group;
5260                         }
5261
5262                         /*
5263                          * at this point we either didn't find a cluster
5264                          * or we weren't able to allocate a block from our
5265                          * cluster.  Free the cluster we've been trying
5266                          * to use, and go to the next block group
5267                          */
5268                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5269                         spin_unlock(&last_ptr->refill_lock);
5270                         goto loop;
5271                 }
5272
5273                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5274                                                     num_bytes, empty_size);
5275                 /*
5276                  * If we didn't find a chunk, and we haven't failed on this
5277                  * block group before, and this block group is in the middle of
5278                  * caching and we are ok with waiting, then go ahead and wait
5279                  * for progress to be made, and set failed_alloc to true.
5280                  *
5281                  * If failed_alloc is true then we've already waited on this
5282                  * block group once and should move on to the next block group.
5283                  */
5284                 if (!offset && !failed_alloc && !cached &&
5285                     loop > LOOP_CACHING_NOWAIT) {
5286                         wait_block_group_cache_progress(block_group,
5287                                                 num_bytes + empty_size);
5288                         failed_alloc = true;
5289                         goto have_block_group;
5290                 } else if (!offset) {
5291                         goto loop;
5292                 }
5293 checks:
5294                 search_start = stripe_align(root, offset);
5295                 /* move on to the next group */
5296                 if (search_start + num_bytes >= search_end) {
5297                         btrfs_add_free_space(block_group, offset, num_bytes);
5298                         goto loop;
5299                 }
5300
5301                 /* move on to the next group */
5302                 if (search_start + num_bytes >
5303                     block_group->key.objectid + block_group->key.offset) {
5304                         btrfs_add_free_space(block_group, offset, num_bytes);
5305                         goto loop;
5306                 }
5307
5308                 ins->objectid = search_start;
5309                 ins->offset = num_bytes;
5310
5311                 if (offset < search_start)
5312                         btrfs_add_free_space(block_group, offset,
5313                                              search_start - offset);
5314                 BUG_ON(offset > search_start);
5315
5316                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
5317                                                   alloc_type);
5318                 if (ret == -EAGAIN) {
5319                         btrfs_add_free_space(block_group, offset, num_bytes);
5320                         goto loop;
5321                 }
5322
5323                 /* we are all good, lets return */
5324                 ins->objectid = search_start;
5325                 ins->offset = num_bytes;
5326
5327                 if (offset < search_start)
5328                         btrfs_add_free_space(block_group, offset,
5329                                              search_start - offset);
5330                 BUG_ON(offset > search_start);
5331                 btrfs_put_block_group(block_group);
5332                 break;
5333 loop:
5334                 failed_cluster_refill = false;
5335                 failed_alloc = false;
5336                 BUG_ON(index != get_block_group_index(block_group));
5337                 btrfs_put_block_group(block_group);
5338         }
5339         up_read(&space_info->groups_sem);
5340
5341         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5342                 goto search;
5343
5344         /* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
5345          *                      for them to make caching progress.  Also
5346          *                      determine the best possible bg to cache
5347          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5348          *                      caching kthreads as we move along
5349          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5350          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5351          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5352          *                      again
5353          */
5354         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5355                 index = 0;
5356                 if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
5357                         found_uncached_bg = false;
5358                         loop++;
5359                         if (!ideal_cache_percent)
5360                                 goto search;
5361
5362                         /*
5363                          * 1 of the following 2 things have happened so far
5364                          *
5365                          * 1) We found an ideal block group for caching that
5366                          * is mostly full and will cache quickly, so we might
5367                          * as well wait for it.
5368                          *
5369                          * 2) We searched for cached only and we didn't find
5370                          * anything, and we didn't start any caching kthreads
5371                          * either, so chances are we will loop through and
5372                          * start a couple caching kthreads, and then come back
5373                          * around and just wait for them.  This will be slower
5374                          * because we will have 2 caching kthreads reading at
5375                          * the same time when we could have just started one
5376                          * and waited for it to get far enough to give us an
5377                          * allocation, so go ahead and go to the wait caching
5378                          * loop.
5379                          */
5380                         loop = LOOP_CACHING_WAIT;
5381                         search_start = ideal_cache_offset;
5382                         ideal_cache_percent = 0;
5383                         goto ideal_cache;
5384                 } else if (loop == LOOP_FIND_IDEAL) {
5385                         /*
5386                          * Didn't find a uncached bg, wait on anything we find
5387                          * next.
5388                          */
5389                         loop = LOOP_CACHING_WAIT;
5390                         goto search;
5391                 }
5392
5393                 loop++;
5394
5395                 if (loop == LOOP_ALLOC_CHUNK) {
5396                        if (allowed_chunk_alloc) {
5397                                 ret = do_chunk_alloc(trans, root, num_bytes +
5398                                                      2 * 1024 * 1024, data,
5399                                                      CHUNK_ALLOC_LIMITED);
5400                                 allowed_chunk_alloc = 0;
5401                                 if (ret == 1)
5402                                         done_chunk_alloc = 1;
5403                         } else if (!done_chunk_alloc &&
5404                                    space_info->force_alloc ==
5405                                    CHUNK_ALLOC_NO_FORCE) {
5406                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5407                         }
5408
5409                        /*
5410                         * We didn't allocate a chunk, go ahead and drop the
5411                         * empty size and loop again.
5412                         */
5413                        if (!done_chunk_alloc)
5414                                loop = LOOP_NO_EMPTY_SIZE;
5415                 }
5416
5417                 if (loop == LOOP_NO_EMPTY_SIZE) {
5418                         empty_size = 0;
5419                         empty_cluster = 0;
5420                 }
5421
5422                 goto search;
5423         } else if (!ins->objectid) {
5424                 ret = -ENOSPC;
5425         } else if (ins->objectid) {
5426                 ret = 0;
5427         }
5428
5429         return ret;
5430 }
5431
5432 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5433                             int dump_block_groups)
5434 {
5435         struct btrfs_block_group_cache *cache;
5436         int index = 0;
5437
5438         spin_lock(&info->lock);
5439         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5440                (unsigned long long)info->flags,
5441                (unsigned long long)(info->total_bytes - info->bytes_used -
5442                                     info->bytes_pinned - info->bytes_reserved -
5443                                     info->bytes_readonly),
5444                (info->full) ? "" : "not ");
5445         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5446                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5447                (unsigned long long)info->total_bytes,
5448                (unsigned long long)info->bytes_used,
5449                (unsigned long long)info->bytes_pinned,
5450                (unsigned long long)info->bytes_reserved,
5451                (unsigned long long)info->bytes_may_use,
5452                (unsigned long long)info->bytes_readonly);
5453         spin_unlock(&info->lock);
5454
5455         if (!dump_block_groups)
5456                 return;
5457
5458         down_read(&info->groups_sem);
5459 again:
5460         list_for_each_entry(cache, &info->block_groups[index], list) {
5461                 spin_lock(&cache->lock);
5462                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5463                        "%llu pinned %llu reserved\n",
5464                        (unsigned long long)cache->key.objectid,
5465                        (unsigned long long)cache->key.offset,
5466                        (unsigned long long)btrfs_block_group_used(&cache->item),
5467                        (unsigned long long)cache->pinned,
5468                        (unsigned long long)cache->reserved);
5469                 btrfs_dump_free_space(cache, bytes);
5470                 spin_unlock(&cache->lock);
5471         }
5472         if (++index < BTRFS_NR_RAID_TYPES)
5473                 goto again;
5474         up_read(&info->groups_sem);
5475 }
5476
5477 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5478                          struct btrfs_root *root,
5479                          u64 num_bytes, u64 min_alloc_size,
5480                          u64 empty_size, u64 hint_byte,
5481                          u64 search_end, struct btrfs_key *ins,
5482                          u64 data)
5483 {
5484         int ret;
5485         u64 search_start = 0;
5486
5487         data = btrfs_get_alloc_profile(root, data);
5488 again:
5489         /*
5490          * the only place that sets empty_size is btrfs_realloc_node, which
5491          * is not called recursively on allocations
5492          */
5493         if (empty_size || root->ref_cows)
5494                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5495                                      num_bytes + 2 * 1024 * 1024, data,
5496                                      CHUNK_ALLOC_NO_FORCE);
5497
5498         WARN_ON(num_bytes < root->sectorsize);
5499         ret = find_free_extent(trans, root, num_bytes, empty_size,
5500                                search_start, search_end, hint_byte,
5501                                ins, data);
5502
5503         if (ret == -ENOSPC && num_bytes > min_alloc_size) {
5504                 num_bytes = num_bytes >> 1;
5505                 num_bytes = num_bytes & ~(root->sectorsize - 1);
5506                 num_bytes = max(num_bytes, min_alloc_size);
5507                 do_chunk_alloc(trans, root->fs_info->extent_root,
5508                                num_bytes, data, CHUNK_ALLOC_FORCE);
5509                 goto again;
5510         }
5511         if (ret == -ENOSPC && btrfs_test_opt(root, ENOSPC_DEBUG)) {
5512                 struct btrfs_space_info *sinfo;
5513
5514                 sinfo = __find_space_info(root->fs_info, data);
5515                 printk(KERN_ERR "btrfs allocation failed flags %llu, "
5516                        "wanted %llu\n", (unsigned long long)data,
5517                        (unsigned long long)num_bytes);
5518                 dump_space_info(sinfo, num_bytes, 1);
5519         }
5520
5521         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
5522
5523         return ret;
5524 }
5525
5526 int btrfs_free_reserved_extent(struct btrfs_root *root, u64 start, u64 len)
5527 {
5528         struct btrfs_block_group_cache *cache;
5529         int ret = 0;
5530
5531         cache = btrfs_lookup_block_group(root->fs_info, start);
5532         if (!cache) {
5533                 printk(KERN_ERR "Unable to find block group for %llu\n",
5534                        (unsigned long long)start);
5535                 return -ENOSPC;
5536         }
5537
5538         if (btrfs_test_opt(root, DISCARD))
5539                 ret = btrfs_discard_extent(root, start, len, NULL);
5540
5541         btrfs_add_free_space(cache, start, len);
5542         btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
5543         btrfs_put_block_group(cache);
5544
5545         trace_btrfs_reserved_extent_free(root, start, len);
5546
5547         return ret;
5548 }
5549
5550 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5551                                       struct btrfs_root *root,
5552                                       u64 parent, u64 root_objectid,
5553                                       u64 flags, u64 owner, u64 offset,
5554                                       struct btrfs_key *ins, int ref_mod)
5555 {
5556         int ret;
5557         struct btrfs_fs_info *fs_info = root->fs_info;
5558         struct btrfs_extent_item *extent_item;
5559         struct btrfs_extent_inline_ref *iref;
5560         struct btrfs_path *path;
5561         struct extent_buffer *leaf;
5562         int type;
5563         u32 size;
5564
5565         if (parent > 0)
5566                 type = BTRFS_SHARED_DATA_REF_KEY;
5567         else
5568                 type = BTRFS_EXTENT_DATA_REF_KEY;
5569
5570         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5571
5572         path = btrfs_alloc_path();
5573         if (!path)
5574                 return -ENOMEM;
5575
5576         path->leave_spinning = 1;
5577         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5578                                       ins, size);
5579         BUG_ON(ret);
5580
5581         leaf = path->nodes[0];
5582         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5583                                      struct btrfs_extent_item);
5584         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
5585         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5586         btrfs_set_extent_flags(leaf, extent_item,
5587                                flags | BTRFS_EXTENT_FLAG_DATA);
5588
5589         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
5590         btrfs_set_extent_inline_ref_type(leaf, iref, type);
5591         if (parent > 0) {
5592                 struct btrfs_shared_data_ref *ref;
5593                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
5594                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5595                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
5596         } else {
5597                 struct btrfs_extent_data_ref *ref;
5598                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
5599                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
5600                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
5601                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
5602                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
5603         }
5604
5605         btrfs_mark_buffer_dirty(path->nodes[0]);
5606         btrfs_free_path(path);
5607
5608         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5609         if (ret) {
5610                 printk(KERN_ERR "btrfs update block group failed for %llu "
5611                        "%llu\n", (unsigned long long)ins->objectid,
5612                        (unsigned long long)ins->offset);
5613                 BUG();
5614         }
5615         return ret;
5616 }
5617
5618 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
5619                                      struct btrfs_root *root,
5620                                      u64 parent, u64 root_objectid,
5621                                      u64 flags, struct btrfs_disk_key *key,
5622                                      int level, struct btrfs_key *ins)
5623 {
5624         int ret;
5625         struct btrfs_fs_info *fs_info = root->fs_info;
5626         struct btrfs_extent_item *extent_item;
5627         struct btrfs_tree_block_info *block_info;
5628         struct btrfs_extent_inline_ref *iref;
5629         struct btrfs_path *path;
5630         struct extent_buffer *leaf;
5631         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
5632
5633         path = btrfs_alloc_path();
5634         if (!path)
5635                 return -ENOMEM;
5636
5637         path->leave_spinning = 1;
5638         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5639                                       ins, size);
5640         BUG_ON(ret);
5641
5642         leaf = path->nodes[0];
5643         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5644                                      struct btrfs_extent_item);
5645         btrfs_set_extent_refs(leaf, extent_item, 1);
5646         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5647         btrfs_set_extent_flags(leaf, extent_item,
5648                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
5649         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
5650
5651         btrfs_set_tree_block_key(leaf, block_info, key);
5652         btrfs_set_tree_block_level(leaf, block_info, level);
5653
5654         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
5655         if (parent > 0) {
5656                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
5657                 btrfs_set_extent_inline_ref_type(leaf, iref,
5658                                                  BTRFS_SHARED_BLOCK_REF_KEY);
5659                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5660         } else {
5661                 btrfs_set_extent_inline_ref_type(leaf, iref,
5662                                                  BTRFS_TREE_BLOCK_REF_KEY);
5663                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
5664         }
5665
5666         btrfs_mark_buffer_dirty(leaf);
5667         btrfs_free_path(path);
5668
5669         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5670         if (ret) {
5671                 printk(KERN_ERR "btrfs update block group failed for %llu "
5672                        "%llu\n", (unsigned long long)ins->objectid,
5673                        (unsigned long long)ins->offset);
5674                 BUG();
5675         }
5676         return ret;
5677 }
5678
5679 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5680                                      struct btrfs_root *root,
5681                                      u64 root_objectid, u64 owner,
5682                                      u64 offset, struct btrfs_key *ins)
5683 {
5684         int ret;
5685
5686         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
5687
5688         ret = btrfs_add_delayed_data_ref(trans, ins->objectid, ins->offset,
5689                                          0, root_objectid, owner, offset,
5690                                          BTRFS_ADD_DELAYED_EXTENT, NULL);
5691         return ret;
5692 }
5693
5694 /*
5695  * this is used by the tree logging recovery code.  It records that
5696  * an extent has been allocated and makes sure to clear the free
5697  * space cache bits as well
5698  */
5699 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
5700                                    struct btrfs_root *root,
5701                                    u64 root_objectid, u64 owner, u64 offset,
5702                                    struct btrfs_key *ins)
5703 {
5704         int ret;
5705         struct btrfs_block_group_cache *block_group;
5706         struct btrfs_caching_control *caching_ctl;
5707         u64 start = ins->objectid;
5708         u64 num_bytes = ins->offset;
5709
5710         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
5711         cache_block_group(block_group, trans, NULL, 0);
5712         caching_ctl = get_caching_control(block_group);
5713
5714         if (!caching_ctl) {
5715                 BUG_ON(!block_group_cache_done(block_group));
5716                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5717                 BUG_ON(ret);
5718         } else {
5719                 mutex_lock(&caching_ctl->mutex);
5720
5721                 if (start >= caching_ctl->progress) {
5722                         ret = add_excluded_extent(root, start, num_bytes);
5723                         BUG_ON(ret);
5724                 } else if (start + num_bytes <= caching_ctl->progress) {
5725                         ret = btrfs_remove_free_space(block_group,
5726                                                       start, num_bytes);
5727                         BUG_ON(ret);
5728                 } else {
5729                         num_bytes = caching_ctl->progress - start;
5730                         ret = btrfs_remove_free_space(block_group,
5731                                                       start, num_bytes);
5732                         BUG_ON(ret);
5733
5734                         start = caching_ctl->progress;
5735                         num_bytes = ins->objectid + ins->offset -
5736                                     caching_ctl->progress;
5737                         ret = add_excluded_extent(root, start, num_bytes);
5738                         BUG_ON(ret);
5739                 }
5740
5741                 mutex_unlock(&caching_ctl->mutex);
5742                 put_caching_control(caching_ctl);
5743         }
5744
5745         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
5746                                           RESERVE_ALLOC_NO_ACCOUNT);
5747         BUG_ON(ret);
5748         btrfs_put_block_group(block_group);
5749         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
5750                                          0, owner, offset, ins, 1);
5751         return ret;
5752 }
5753
5754 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
5755                                             struct btrfs_root *root,
5756                                             u64 bytenr, u32 blocksize,
5757                                             int level)
5758 {
5759         struct extent_buffer *buf;
5760
5761         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
5762         if (!buf)
5763                 return ERR_PTR(-ENOMEM);
5764         btrfs_set_header_generation(buf, trans->transid);
5765         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
5766         btrfs_tree_lock(buf);
5767         clean_tree_block(trans, root, buf);
5768
5769         btrfs_set_lock_blocking(buf);
5770         btrfs_set_buffer_uptodate(buf);
5771
5772         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
5773                 /*
5774                  * we allow two log transactions at a time, use different
5775                  * EXENT bit to differentiate dirty pages.
5776                  */
5777                 if (root->log_transid % 2 == 0)
5778                         set_extent_dirty(&root->dirty_log_pages, buf->start,
5779                                         buf->start + buf->len - 1, GFP_NOFS);
5780                 else
5781                         set_extent_new(&root->dirty_log_pages, buf->start,
5782                                         buf->start + buf->len - 1, GFP_NOFS);
5783         } else {
5784                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
5785                          buf->start + buf->len - 1, GFP_NOFS);
5786         }
5787         trans->blocks_used++;
5788         /* this returns a buffer locked for blocking */
5789         return buf;
5790 }
5791
5792 static struct btrfs_block_rsv *
5793 use_block_rsv(struct btrfs_trans_handle *trans,
5794               struct btrfs_root *root, u32 blocksize)
5795 {
5796         struct btrfs_block_rsv *block_rsv;
5797         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5798         int ret;
5799
5800         block_rsv = get_block_rsv(trans, root);
5801
5802         if (block_rsv->size == 0) {
5803                 ret = reserve_metadata_bytes(trans, root, block_rsv,
5804                                              blocksize, 0);
5805                 /*
5806                  * If we couldn't reserve metadata bytes try and use some from
5807                  * the global reserve.
5808                  */
5809                 if (ret && block_rsv != global_rsv) {
5810                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5811                         if (!ret)
5812                                 return global_rsv;
5813                         return ERR_PTR(ret);
5814                 } else if (ret) {
5815                         return ERR_PTR(ret);
5816                 }
5817                 return block_rsv;
5818         }
5819
5820         ret = block_rsv_use_bytes(block_rsv, blocksize);
5821         if (!ret)
5822                 return block_rsv;
5823         if (ret) {
5824                 WARN_ON(1);
5825                 ret = reserve_metadata_bytes(trans, root, block_rsv, blocksize,
5826                                              0);
5827                 if (!ret) {
5828                         spin_lock(&block_rsv->lock);
5829                         block_rsv->size += blocksize;
5830                         spin_unlock(&block_rsv->lock);
5831                         return block_rsv;
5832                 } else if (ret && block_rsv != global_rsv) {
5833                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5834                         if (!ret)
5835                                 return global_rsv;
5836                 }
5837         }
5838
5839         return ERR_PTR(-ENOSPC);
5840 }
5841
5842 static void unuse_block_rsv(struct btrfs_block_rsv *block_rsv, u32 blocksize)
5843 {
5844         block_rsv_add_bytes(block_rsv, blocksize, 0);
5845         block_rsv_release_bytes(block_rsv, NULL, 0);
5846 }
5847
5848 /*
5849  * finds a free extent and does all the dirty work required for allocation
5850  * returns the key for the extent through ins, and a tree buffer for
5851  * the first block of the extent through buf.
5852  *
5853  * returns the tree buffer or NULL.
5854  */
5855 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
5856                                         struct btrfs_root *root, u32 blocksize,
5857                                         u64 parent, u64 root_objectid,
5858                                         struct btrfs_disk_key *key, int level,
5859                                         u64 hint, u64 empty_size)
5860 {
5861         struct btrfs_key ins;
5862         struct btrfs_block_rsv *block_rsv;
5863         struct extent_buffer *buf;
5864         u64 flags = 0;
5865         int ret;
5866
5867
5868         block_rsv = use_block_rsv(trans, root, blocksize);
5869         if (IS_ERR(block_rsv))
5870                 return ERR_CAST(block_rsv);
5871
5872         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
5873                                    empty_size, hint, (u64)-1, &ins, 0);
5874         if (ret) {
5875                 unuse_block_rsv(block_rsv, blocksize);
5876                 return ERR_PTR(ret);
5877         }
5878
5879         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
5880                                     blocksize, level);
5881         BUG_ON(IS_ERR(buf));
5882
5883         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
5884                 if (parent == 0)
5885                         parent = ins.objectid;
5886                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
5887         } else
5888                 BUG_ON(parent > 0);
5889
5890         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
5891                 struct btrfs_delayed_extent_op *extent_op;
5892                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
5893                 BUG_ON(!extent_op);
5894                 if (key)
5895                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
5896                 else
5897                         memset(&extent_op->key, 0, sizeof(extent_op->key));
5898                 extent_op->flags_to_set = flags;
5899                 extent_op->update_key = 1;
5900                 extent_op->update_flags = 1;
5901                 extent_op->is_data = 0;
5902
5903                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
5904                                         ins.offset, parent, root_objectid,
5905                                         level, BTRFS_ADD_DELAYED_EXTENT,
5906                                         extent_op);
5907                 BUG_ON(ret);
5908         }
5909         return buf;
5910 }
5911
5912 struct walk_control {
5913         u64 refs[BTRFS_MAX_LEVEL];
5914         u64 flags[BTRFS_MAX_LEVEL];
5915         struct btrfs_key update_progress;
5916         int stage;
5917         int level;
5918         int shared_level;
5919         int update_ref;
5920         int keep_locks;
5921         int reada_slot;
5922         int reada_count;
5923 };
5924
5925 #define DROP_REFERENCE  1
5926 #define UPDATE_BACKREF  2
5927
5928 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
5929                                      struct btrfs_root *root,
5930                                      struct walk_control *wc,
5931                                      struct btrfs_path *path)
5932 {
5933         u64 bytenr;
5934         u64 generation;
5935         u64 refs;
5936         u64 flags;
5937         u32 nritems;
5938         u32 blocksize;
5939         struct btrfs_key key;
5940         struct extent_buffer *eb;
5941         int ret;
5942         int slot;
5943         int nread = 0;
5944
5945         if (path->slots[wc->level] < wc->reada_slot) {
5946                 wc->reada_count = wc->reada_count * 2 / 3;
5947                 wc->reada_count = max(wc->reada_count, 2);
5948         } else {
5949                 wc->reada_count = wc->reada_count * 3 / 2;
5950                 wc->reada_count = min_t(int, wc->reada_count,
5951                                         BTRFS_NODEPTRS_PER_BLOCK(root));
5952         }
5953
5954         eb = path->nodes[wc->level];
5955         nritems = btrfs_header_nritems(eb);
5956         blocksize = btrfs_level_size(root, wc->level - 1);
5957
5958         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
5959                 if (nread >= wc->reada_count)
5960                         break;
5961
5962                 cond_resched();
5963                 bytenr = btrfs_node_blockptr(eb, slot);
5964                 generation = btrfs_node_ptr_generation(eb, slot);
5965
5966                 if (slot == path->slots[wc->level])
5967                         goto reada;
5968
5969                 if (wc->stage == UPDATE_BACKREF &&
5970                     generation <= root->root_key.offset)
5971                         continue;
5972
5973                 /* We don't lock the tree block, it's OK to be racy here */
5974                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
5975                                                &refs, &flags);
5976                 BUG_ON(ret);
5977                 BUG_ON(refs == 0);
5978
5979                 if (wc->stage == DROP_REFERENCE) {
5980                         if (refs == 1)
5981                                 goto reada;
5982
5983                         if (wc->level == 1 &&
5984                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5985                                 continue;
5986                         if (!wc->update_ref ||
5987                             generation <= root->root_key.offset)
5988                                 continue;
5989                         btrfs_node_key_to_cpu(eb, &key, slot);
5990                         ret = btrfs_comp_cpu_keys(&key,
5991                                                   &wc->update_progress);
5992                         if (ret < 0)
5993                                 continue;
5994                 } else {
5995                         if (wc->level == 1 &&
5996                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5997                                 continue;
5998                 }
5999 reada:
6000                 ret = readahead_tree_block(root, bytenr, blocksize,
6001                                            generation);
6002                 if (ret)
6003                         break;
6004                 nread++;
6005         }
6006         wc->reada_slot = slot;
6007 }
6008
6009 /*
6010  * hepler to process tree block while walking down the tree.
6011  *
6012  * when wc->stage == UPDATE_BACKREF, this function updates
6013  * back refs for pointers in the block.
6014  *
6015  * NOTE: return value 1 means we should stop walking down.
6016  */
6017 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6018                                    struct btrfs_root *root,
6019                                    struct btrfs_path *path,
6020                                    struct walk_control *wc, int lookup_info)
6021 {
6022         int level = wc->level;
6023         struct extent_buffer *eb = path->nodes[level];
6024         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6025         int ret;
6026
6027         if (wc->stage == UPDATE_BACKREF &&
6028             btrfs_header_owner(eb) != root->root_key.objectid)
6029                 return 1;
6030
6031         /*
6032          * when reference count of tree block is 1, it won't increase
6033          * again. once full backref flag is set, we never clear it.
6034          */
6035         if (lookup_info &&
6036             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6037              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6038                 BUG_ON(!path->locks[level]);
6039                 ret = btrfs_lookup_extent_info(trans, root,
6040                                                eb->start, eb->len,
6041                                                &wc->refs[level],
6042                                                &wc->flags[level]);
6043                 BUG_ON(ret);
6044                 BUG_ON(wc->refs[level] == 0);
6045         }
6046
6047         if (wc->stage == DROP_REFERENCE) {
6048                 if (wc->refs[level] > 1)
6049                         return 1;
6050
6051                 if (path->locks[level] && !wc->keep_locks) {
6052                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6053                         path->locks[level] = 0;
6054                 }
6055                 return 0;
6056         }
6057
6058         /* wc->stage == UPDATE_BACKREF */
6059         if (!(wc->flags[level] & flag)) {
6060                 BUG_ON(!path->locks[level]);
6061                 ret = btrfs_inc_ref(trans, root, eb, 1);
6062                 BUG_ON(ret);
6063                 ret = btrfs_dec_ref(trans, root, eb, 0);
6064                 BUG_ON(ret);
6065                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6066                                                   eb->len, flag, 0);
6067                 BUG_ON(ret);
6068                 wc->flags[level] |= flag;
6069         }
6070
6071         /*
6072          * the block is shared by multiple trees, so it's not good to
6073          * keep the tree lock
6074          */
6075         if (path->locks[level] && level > 0) {
6076                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6077                 path->locks[level] = 0;
6078         }
6079         return 0;
6080 }
6081
6082 /*
6083  * hepler to process tree block pointer.
6084  *
6085  * when wc->stage == DROP_REFERENCE, this function checks
6086  * reference count of the block pointed to. if the block
6087  * is shared and we need update back refs for the subtree
6088  * rooted at the block, this function changes wc->stage to
6089  * UPDATE_BACKREF. if the block is shared and there is no
6090  * need to update back, this function drops the reference
6091  * to the block.
6092  *
6093  * NOTE: return value 1 means we should stop walking down.
6094  */
6095 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6096                                  struct btrfs_root *root,
6097                                  struct btrfs_path *path,
6098                                  struct walk_control *wc, int *lookup_info)
6099 {
6100         u64 bytenr;
6101         u64 generation;
6102         u64 parent;
6103         u32 blocksize;
6104         struct btrfs_key key;
6105         struct extent_buffer *next;
6106         int level = wc->level;
6107         int reada = 0;
6108         int ret = 0;
6109
6110         generation = btrfs_node_ptr_generation(path->nodes[level],
6111                                                path->slots[level]);
6112         /*
6113          * if the lower level block was created before the snapshot
6114          * was created, we know there is no need to update back refs
6115          * for the subtree
6116          */
6117         if (wc->stage == UPDATE_BACKREF &&
6118             generation <= root->root_key.offset) {
6119                 *lookup_info = 1;
6120                 return 1;
6121         }
6122
6123         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6124         blocksize = btrfs_level_size(root, level - 1);
6125
6126         next = btrfs_find_tree_block(root, bytenr, blocksize);
6127         if (!next) {
6128                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6129                 if (!next)
6130                         return -ENOMEM;
6131                 reada = 1;
6132         }
6133         btrfs_tree_lock(next);
6134         btrfs_set_lock_blocking(next);
6135
6136         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6137                                        &wc->refs[level - 1],
6138                                        &wc->flags[level - 1]);
6139         BUG_ON(ret);
6140         BUG_ON(wc->refs[level - 1] == 0);
6141         *lookup_info = 0;
6142
6143         if (wc->stage == DROP_REFERENCE) {
6144                 if (wc->refs[level - 1] > 1) {
6145                         if (level == 1 &&
6146                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6147                                 goto skip;
6148
6149                         if (!wc->update_ref ||
6150                             generation <= root->root_key.offset)
6151                                 goto skip;
6152
6153                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6154                                               path->slots[level]);
6155                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6156                         if (ret < 0)
6157                                 goto skip;
6158
6159                         wc->stage = UPDATE_BACKREF;
6160                         wc->shared_level = level - 1;
6161                 }
6162         } else {
6163                 if (level == 1 &&
6164                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6165                         goto skip;
6166         }
6167
6168         if (!btrfs_buffer_uptodate(next, generation)) {
6169                 btrfs_tree_unlock(next);
6170                 free_extent_buffer(next);
6171                 next = NULL;
6172                 *lookup_info = 1;
6173         }
6174
6175         if (!next) {
6176                 if (reada && level == 1)
6177                         reada_walk_down(trans, root, wc, path);
6178                 next = read_tree_block(root, bytenr, blocksize, generation);
6179                 if (!next)
6180                         return -EIO;
6181                 btrfs_tree_lock(next);
6182                 btrfs_set_lock_blocking(next);
6183         }
6184
6185         level--;
6186         BUG_ON(level != btrfs_header_level(next));
6187         path->nodes[level] = next;
6188         path->slots[level] = 0;
6189         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6190         wc->level = level;
6191         if (wc->level == 1)
6192                 wc->reada_slot = 0;
6193         return 0;
6194 skip:
6195         wc->refs[level - 1] = 0;
6196         wc->flags[level - 1] = 0;
6197         if (wc->stage == DROP_REFERENCE) {
6198                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6199                         parent = path->nodes[level]->start;
6200                 } else {
6201                         BUG_ON(root->root_key.objectid !=
6202                                btrfs_header_owner(path->nodes[level]));
6203                         parent = 0;
6204                 }
6205
6206                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6207                                         root->root_key.objectid, level - 1, 0);
6208                 BUG_ON(ret);
6209         }
6210         btrfs_tree_unlock(next);
6211         free_extent_buffer(next);
6212         *lookup_info = 1;
6213         return 1;
6214 }
6215
6216 /*
6217  * hepler to process tree block while walking up the tree.
6218  *
6219  * when wc->stage == DROP_REFERENCE, this function drops
6220  * reference count on the block.
6221  *
6222  * when wc->stage == UPDATE_BACKREF, this function changes
6223  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6224  * to UPDATE_BACKREF previously while processing the block.
6225  *
6226  * NOTE: return value 1 means we should stop walking up.
6227  */
6228 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6229                                  struct btrfs_root *root,
6230                                  struct btrfs_path *path,
6231                                  struct walk_control *wc)
6232 {
6233         int ret;
6234         int level = wc->level;
6235         struct extent_buffer *eb = path->nodes[level];
6236         u64 parent = 0;
6237
6238         if (wc->stage == UPDATE_BACKREF) {
6239                 BUG_ON(wc->shared_level < level);
6240                 if (level < wc->shared_level)
6241                         goto out;
6242
6243                 ret = find_next_key(path, level + 1, &wc->update_progress);
6244                 if (ret > 0)
6245                         wc->update_ref = 0;
6246
6247                 wc->stage = DROP_REFERENCE;
6248                 wc->shared_level = -1;
6249                 path->slots[level] = 0;
6250
6251                 /*
6252                  * check reference count again if the block isn't locked.
6253                  * we should start walking down the tree again if reference
6254                  * count is one.
6255                  */
6256                 if (!path->locks[level]) {
6257                         BUG_ON(level == 0);
6258                         btrfs_tree_lock(eb);
6259                         btrfs_set_lock_blocking(eb);
6260                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6261
6262                         ret = btrfs_lookup_extent_info(trans, root,
6263                                                        eb->start, eb->len,
6264                                                        &wc->refs[level],
6265                                                        &wc->flags[level]);
6266                         BUG_ON(ret);
6267                         BUG_ON(wc->refs[level] == 0);
6268                         if (wc->refs[level] == 1) {
6269                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6270                                 return 1;
6271                         }
6272                 }
6273         }
6274
6275         /* wc->stage == DROP_REFERENCE */
6276         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6277
6278         if (wc->refs[level] == 1) {
6279                 if (level == 0) {
6280                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6281                                 ret = btrfs_dec_ref(trans, root, eb, 1);
6282                         else
6283                                 ret = btrfs_dec_ref(trans, root, eb, 0);
6284                         BUG_ON(ret);
6285                 }
6286                 /* make block locked assertion in clean_tree_block happy */
6287                 if (!path->locks[level] &&
6288                     btrfs_header_generation(eb) == trans->transid) {
6289                         btrfs_tree_lock(eb);
6290                         btrfs_set_lock_blocking(eb);
6291                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6292                 }
6293                 clean_tree_block(trans, root, eb);
6294         }
6295
6296         if (eb == root->node) {
6297                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6298                         parent = eb->start;
6299                 else
6300                         BUG_ON(root->root_key.objectid !=
6301                                btrfs_header_owner(eb));
6302         } else {
6303                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6304                         parent = path->nodes[level + 1]->start;
6305                 else
6306                         BUG_ON(root->root_key.objectid !=
6307                                btrfs_header_owner(path->nodes[level + 1]));
6308         }
6309
6310         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6311 out:
6312         wc->refs[level] = 0;
6313         wc->flags[level] = 0;
6314         return 0;
6315 }
6316
6317 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6318                                    struct btrfs_root *root,
6319                                    struct btrfs_path *path,
6320                                    struct walk_control *wc)
6321 {
6322         int level = wc->level;
6323         int lookup_info = 1;
6324         int ret;
6325
6326         while (level >= 0) {
6327                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6328                 if (ret > 0)
6329                         break;
6330
6331                 if (level == 0)
6332                         break;
6333
6334                 if (path->slots[level] >=
6335                     btrfs_header_nritems(path->nodes[level]))
6336                         break;
6337
6338                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6339                 if (ret > 0) {
6340                         path->slots[level]++;
6341                         continue;
6342                 } else if (ret < 0)
6343                         return ret;
6344                 level = wc->level;
6345         }
6346         return 0;
6347 }
6348
6349 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6350                                  struct btrfs_root *root,
6351                                  struct btrfs_path *path,
6352                                  struct walk_control *wc, int max_level)
6353 {
6354         int level = wc->level;
6355         int ret;
6356
6357         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6358         while (level < max_level && path->nodes[level]) {
6359                 wc->level = level;
6360                 if (path->slots[level] + 1 <
6361                     btrfs_header_nritems(path->nodes[level])) {
6362                         path->slots[level]++;
6363                         return 0;
6364                 } else {
6365                         ret = walk_up_proc(trans, root, path, wc);
6366                         if (ret > 0)
6367                                 return 0;
6368
6369                         if (path->locks[level]) {
6370                                 btrfs_tree_unlock_rw(path->nodes[level],
6371                                                      path->locks[level]);
6372                                 path->locks[level] = 0;
6373                         }
6374                         free_extent_buffer(path->nodes[level]);
6375                         path->nodes[level] = NULL;
6376                         level++;
6377                 }
6378         }
6379         return 1;
6380 }
6381
6382 /*
6383  * drop a subvolume tree.
6384  *
6385  * this function traverses the tree freeing any blocks that only
6386  * referenced by the tree.
6387  *
6388  * when a shared tree block is found. this function decreases its
6389  * reference count by one. if update_ref is true, this function
6390  * also make sure backrefs for the shared block and all lower level
6391  * blocks are properly updated.
6392  */
6393 void btrfs_drop_snapshot(struct btrfs_root *root,
6394                          struct btrfs_block_rsv *block_rsv, int update_ref)
6395 {
6396         struct btrfs_path *path;
6397         struct btrfs_trans_handle *trans;
6398         struct btrfs_root *tree_root = root->fs_info->tree_root;
6399         struct btrfs_root_item *root_item = &root->root_item;
6400         struct walk_control *wc;
6401         struct btrfs_key key;
6402         int err = 0;
6403         int ret;
6404         int level;
6405
6406         path = btrfs_alloc_path();
6407         if (!path) {
6408                 err = -ENOMEM;
6409                 goto out;
6410         }
6411
6412         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6413         if (!wc) {
6414                 btrfs_free_path(path);
6415                 err = -ENOMEM;
6416                 goto out;
6417         }
6418
6419         trans = btrfs_start_transaction(tree_root, 0);
6420         BUG_ON(IS_ERR(trans));
6421
6422         if (block_rsv)
6423                 trans->block_rsv = block_rsv;
6424
6425         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6426                 level = btrfs_header_level(root->node);
6427                 path->nodes[level] = btrfs_lock_root_node(root);
6428                 btrfs_set_lock_blocking(path->nodes[level]);
6429                 path->slots[level] = 0;
6430                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6431                 memset(&wc->update_progress, 0,
6432                        sizeof(wc->update_progress));
6433         } else {
6434                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6435                 memcpy(&wc->update_progress, &key,
6436                        sizeof(wc->update_progress));
6437
6438                 level = root_item->drop_level;
6439                 BUG_ON(level == 0);
6440                 path->lowest_level = level;
6441                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6442                 path->lowest_level = 0;
6443                 if (ret < 0) {
6444                         err = ret;
6445                         goto out_free;
6446                 }
6447                 WARN_ON(ret > 0);
6448
6449                 /*
6450                  * unlock our path, this is safe because only this
6451                  * function is allowed to delete this snapshot
6452                  */
6453                 btrfs_unlock_up_safe(path, 0);
6454
6455                 level = btrfs_header_level(root->node);
6456                 while (1) {
6457                         btrfs_tree_lock(path->nodes[level]);
6458                         btrfs_set_lock_blocking(path->nodes[level]);
6459
6460                         ret = btrfs_lookup_extent_info(trans, root,
6461                                                 path->nodes[level]->start,
6462                                                 path->nodes[level]->len,
6463                                                 &wc->refs[level],
6464                                                 &wc->flags[level]);
6465                         BUG_ON(ret);
6466                         BUG_ON(wc->refs[level] == 0);
6467
6468                         if (level == root_item->drop_level)
6469                                 break;
6470
6471                         btrfs_tree_unlock(path->nodes[level]);
6472                         WARN_ON(wc->refs[level] != 1);
6473                         level--;
6474                 }
6475         }
6476
6477         wc->level = level;
6478         wc->shared_level = -1;
6479         wc->stage = DROP_REFERENCE;
6480         wc->update_ref = update_ref;
6481         wc->keep_locks = 0;
6482         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6483
6484         while (1) {
6485                 ret = walk_down_tree(trans, root, path, wc);
6486                 if (ret < 0) {
6487                         err = ret;
6488                         break;
6489                 }
6490
6491                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
6492                 if (ret < 0) {
6493                         err = ret;
6494                         break;
6495                 }
6496
6497                 if (ret > 0) {
6498                         BUG_ON(wc->stage != DROP_REFERENCE);
6499                         break;
6500                 }
6501
6502                 if (wc->stage == DROP_REFERENCE) {
6503                         level = wc->level;
6504                         btrfs_node_key(path->nodes[level],
6505                                        &root_item->drop_progress,
6506                                        path->slots[level]);
6507                         root_item->drop_level = level;
6508                 }
6509
6510                 BUG_ON(wc->level == 0);
6511                 if (btrfs_should_end_transaction(trans, tree_root)) {
6512                         ret = btrfs_update_root(trans, tree_root,
6513                                                 &root->root_key,
6514                                                 root_item);
6515                         BUG_ON(ret);
6516
6517                         btrfs_end_transaction_throttle(trans, tree_root);
6518                         trans = btrfs_start_transaction(tree_root, 0);
6519                         BUG_ON(IS_ERR(trans));
6520                         if (block_rsv)
6521                                 trans->block_rsv = block_rsv;
6522                 }
6523         }
6524         btrfs_release_path(path);
6525         BUG_ON(err);
6526
6527         ret = btrfs_del_root(trans, tree_root, &root->root_key);
6528         BUG_ON(ret);
6529
6530         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
6531                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
6532                                            NULL, NULL);
6533                 BUG_ON(ret < 0);
6534                 if (ret > 0) {
6535                         /* if we fail to delete the orphan item this time
6536                          * around, it'll get picked up the next time.
6537                          *
6538                          * The most common failure here is just -ENOENT.
6539                          */
6540                         btrfs_del_orphan_item(trans, tree_root,
6541                                               root->root_key.objectid);
6542                 }
6543         }
6544
6545         if (root->in_radix) {
6546                 btrfs_free_fs_root(tree_root->fs_info, root);
6547         } else {
6548                 free_extent_buffer(root->node);
6549                 free_extent_buffer(root->commit_root);
6550                 kfree(root);
6551         }
6552 out_free:
6553         btrfs_end_transaction_throttle(trans, tree_root);
6554         kfree(wc);
6555         btrfs_free_path(path);
6556 out:
6557         if (err)
6558                 btrfs_std_error(root->fs_info, err);
6559         return;
6560 }
6561
6562 /*
6563  * drop subtree rooted at tree block 'node'.
6564  *
6565  * NOTE: this function will unlock and release tree block 'node'
6566  */
6567 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
6568                         struct btrfs_root *root,
6569                         struct extent_buffer *node,
6570                         struct extent_buffer *parent)
6571 {
6572         struct btrfs_path *path;
6573         struct walk_control *wc;
6574         int level;
6575         int parent_level;
6576         int ret = 0;
6577         int wret;
6578
6579         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
6580
6581         path = btrfs_alloc_path();
6582         if (!path)
6583                 return -ENOMEM;
6584
6585         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6586         if (!wc) {
6587                 btrfs_free_path(path);
6588                 return -ENOMEM;
6589         }
6590
6591         btrfs_assert_tree_locked(parent);
6592         parent_level = btrfs_header_level(parent);
6593         extent_buffer_get(parent);
6594         path->nodes[parent_level] = parent;
6595         path->slots[parent_level] = btrfs_header_nritems(parent);
6596
6597         btrfs_assert_tree_locked(node);
6598         level = btrfs_header_level(node);
6599         path->nodes[level] = node;
6600         path->slots[level] = 0;
6601         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6602
6603         wc->refs[parent_level] = 1;
6604         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6605         wc->level = level;
6606         wc->shared_level = -1;
6607         wc->stage = DROP_REFERENCE;
6608         wc->update_ref = 0;
6609         wc->keep_locks = 1;
6610         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6611
6612         while (1) {
6613                 wret = walk_down_tree(trans, root, path, wc);
6614                 if (wret < 0) {
6615                         ret = wret;
6616                         break;
6617                 }
6618
6619                 wret = walk_up_tree(trans, root, path, wc, parent_level);
6620                 if (wret < 0)
6621                         ret = wret;
6622                 if (wret != 0)
6623                         break;
6624         }
6625
6626         kfree(wc);
6627         btrfs_free_path(path);
6628         return ret;
6629 }
6630
6631 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
6632 {
6633         u64 num_devices;
6634         u64 stripped = BTRFS_BLOCK_GROUP_RAID0 |
6635                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
6636
6637         /*
6638          * we add in the count of missing devices because we want
6639          * to make sure that any RAID levels on a degraded FS
6640          * continue to be honored.
6641          */
6642         num_devices = root->fs_info->fs_devices->rw_devices +
6643                 root->fs_info->fs_devices->missing_devices;
6644
6645         if (num_devices == 1) {
6646                 stripped |= BTRFS_BLOCK_GROUP_DUP;
6647                 stripped = flags & ~stripped;
6648
6649                 /* turn raid0 into single device chunks */
6650                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
6651                         return stripped;
6652
6653                 /* turn mirroring into duplication */
6654                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
6655                              BTRFS_BLOCK_GROUP_RAID10))
6656                         return stripped | BTRFS_BLOCK_GROUP_DUP;
6657                 return flags;
6658         } else {
6659                 /* they already had raid on here, just return */
6660                 if (flags & stripped)
6661                         return flags;
6662
6663                 stripped |= BTRFS_BLOCK_GROUP_DUP;
6664                 stripped = flags & ~stripped;
6665
6666                 /* switch duplicated blocks with raid1 */
6667                 if (flags & BTRFS_BLOCK_GROUP_DUP)
6668                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
6669
6670                 /* turn single device chunks into raid0 */
6671                 return stripped | BTRFS_BLOCK_GROUP_RAID0;
6672         }
6673         return flags;
6674 }
6675
6676 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
6677 {
6678         struct btrfs_space_info *sinfo = cache->space_info;
6679         u64 num_bytes;
6680         u64 min_allocable_bytes;
6681         int ret = -ENOSPC;
6682
6683
6684         /*
6685          * We need some metadata space and system metadata space for
6686          * allocating chunks in some corner cases until we force to set
6687          * it to be readonly.
6688          */
6689         if ((sinfo->flags &
6690              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
6691             !force)
6692                 min_allocable_bytes = 1 * 1024 * 1024;
6693         else
6694                 min_allocable_bytes = 0;
6695
6696         spin_lock(&sinfo->lock);
6697         spin_lock(&cache->lock);
6698
6699         if (cache->ro) {
6700                 ret = 0;
6701                 goto out;
6702         }
6703
6704         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
6705                     cache->bytes_super - btrfs_block_group_used(&cache->item);
6706
6707         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
6708             sinfo->bytes_may_use + sinfo->bytes_readonly +
6709             cache->reserved_pinned + num_bytes + min_allocable_bytes <=
6710             sinfo->total_bytes) {
6711                 sinfo->bytes_readonly += num_bytes;
6712                 sinfo->bytes_may_use += cache->reserved_pinned;
6713                 cache->reserved_pinned = 0;
6714                 cache->ro = 1;
6715                 ret = 0;
6716         }
6717 out:
6718         spin_unlock(&cache->lock);
6719         spin_unlock(&sinfo->lock);
6720         return ret;
6721 }
6722
6723 int btrfs_set_block_group_ro(struct btrfs_root *root,
6724                              struct btrfs_block_group_cache *cache)
6725
6726 {
6727         struct btrfs_trans_handle *trans;
6728         u64 alloc_flags;
6729         int ret;
6730
6731         BUG_ON(cache->ro);
6732
6733         trans = btrfs_join_transaction(root);
6734         BUG_ON(IS_ERR(trans));
6735
6736         alloc_flags = update_block_group_flags(root, cache->flags);
6737         if (alloc_flags != cache->flags)
6738                 do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6739                                CHUNK_ALLOC_FORCE);
6740
6741         ret = set_block_group_ro(cache, 0);
6742         if (!ret)
6743                 goto out;
6744         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
6745         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6746                              CHUNK_ALLOC_FORCE);
6747         if (ret < 0)
6748                 goto out;
6749         ret = set_block_group_ro(cache, 0);
6750 out:
6751         btrfs_end_transaction(trans, root);
6752         return ret;
6753 }
6754
6755 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
6756                             struct btrfs_root *root, u64 type)
6757 {
6758         u64 alloc_flags = get_alloc_profile(root, type);
6759         return do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6760                               CHUNK_ALLOC_FORCE);
6761 }
6762
6763 /*
6764  * helper to account the unused space of all the readonly block group in the
6765  * list. takes mirrors into account.
6766  */
6767 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
6768 {
6769         struct btrfs_block_group_cache *block_group;
6770         u64 free_bytes = 0;
6771         int factor;
6772
6773         list_for_each_entry(block_group, groups_list, list) {
6774                 spin_lock(&block_group->lock);
6775
6776                 if (!block_group->ro) {
6777                         spin_unlock(&block_group->lock);
6778                         continue;
6779                 }
6780
6781                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
6782                                           BTRFS_BLOCK_GROUP_RAID10 |
6783                                           BTRFS_BLOCK_GROUP_DUP))
6784                         factor = 2;
6785                 else
6786                         factor = 1;
6787
6788                 free_bytes += (block_group->key.offset -
6789                                btrfs_block_group_used(&block_group->item)) *
6790                                factor;
6791
6792                 spin_unlock(&block_group->lock);
6793         }
6794
6795         return free_bytes;
6796 }
6797
6798 /*
6799  * helper to account the unused space of all the readonly block group in the
6800  * space_info. takes mirrors into account.
6801  */
6802 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
6803 {
6804         int i;
6805         u64 free_bytes = 0;
6806
6807         spin_lock(&sinfo->lock);
6808
6809         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
6810                 if (!list_empty(&sinfo->block_groups[i]))
6811                         free_bytes += __btrfs_get_ro_block_group_free_space(
6812                                                 &sinfo->block_groups[i]);
6813
6814         spin_unlock(&sinfo->lock);
6815
6816         return free_bytes;
6817 }
6818
6819 int btrfs_set_block_group_rw(struct btrfs_root *root,
6820                               struct btrfs_block_group_cache *cache)
6821 {
6822         struct btrfs_space_info *sinfo = cache->space_info;
6823         u64 num_bytes;
6824
6825         BUG_ON(!cache->ro);
6826
6827         spin_lock(&sinfo->lock);
6828         spin_lock(&cache->lock);
6829         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
6830                     cache->bytes_super - btrfs_block_group_used(&cache->item);
6831         sinfo->bytes_readonly -= num_bytes;
6832         cache->ro = 0;
6833         spin_unlock(&cache->lock);
6834         spin_unlock(&sinfo->lock);
6835         return 0;
6836 }
6837
6838 /*
6839  * checks to see if its even possible to relocate this block group.
6840  *
6841  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
6842  * ok to go ahead and try.
6843  */
6844 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
6845 {
6846         struct btrfs_block_group_cache *block_group;
6847         struct btrfs_space_info *space_info;
6848         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
6849         struct btrfs_device *device;
6850         u64 min_free;
6851         u64 dev_min = 1;
6852         u64 dev_nr = 0;
6853         int index;
6854         int full = 0;
6855         int ret = 0;
6856
6857         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
6858
6859         /* odd, couldn't find the block group, leave it alone */
6860         if (!block_group)
6861                 return -1;
6862
6863         min_free = btrfs_block_group_used(&block_group->item);
6864
6865         /* no bytes used, we're good */
6866         if (!min_free)
6867                 goto out;
6868
6869         space_info = block_group->space_info;
6870         spin_lock(&space_info->lock);
6871
6872         full = space_info->full;
6873
6874         /*
6875          * if this is the last block group we have in this space, we can't
6876          * relocate it unless we're able to allocate a new chunk below.
6877          *
6878          * Otherwise, we need to make sure we have room in the space to handle
6879          * all of the extents from this block group.  If we can, we're good
6880          */
6881         if ((space_info->total_bytes != block_group->key.offset) &&
6882             (space_info->bytes_used + space_info->bytes_reserved +
6883              space_info->bytes_pinned + space_info->bytes_readonly +
6884              min_free < space_info->total_bytes)) {
6885                 spin_unlock(&space_info->lock);
6886                 goto out;
6887         }
6888         spin_unlock(&space_info->lock);
6889
6890         /*
6891          * ok we don't have enough space, but maybe we have free space on our
6892          * devices to allocate new chunks for relocation, so loop through our
6893          * alloc devices and guess if we have enough space.  However, if we
6894          * were marked as full, then we know there aren't enough chunks, and we
6895          * can just return.
6896          */
6897         ret = -1;
6898         if (full)
6899                 goto out;
6900
6901         /*
6902          * index:
6903          *      0: raid10
6904          *      1: raid1
6905          *      2: dup
6906          *      3: raid0
6907          *      4: single
6908          */
6909         index = get_block_group_index(block_group);
6910         if (index == 0) {
6911                 dev_min = 4;
6912                 /* Divide by 2 */
6913                 min_free >>= 1;
6914         } else if (index == 1) {
6915                 dev_min = 2;
6916         } else if (index == 2) {
6917                 /* Multiply by 2 */
6918                 min_free <<= 1;
6919         } else if (index == 3) {
6920                 dev_min = fs_devices->rw_devices;
6921                 do_div(min_free, dev_min);
6922         }
6923
6924         mutex_lock(&root->fs_info->chunk_mutex);
6925         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
6926                 u64 dev_offset;
6927
6928                 /*
6929                  * check to make sure we can actually find a chunk with enough
6930                  * space to fit our block group in.
6931                  */
6932                 if (device->total_bytes > device->bytes_used + min_free) {
6933                         ret = find_free_dev_extent(NULL, device, min_free,
6934                                                    &dev_offset, NULL);
6935                         if (!ret)
6936                                 dev_nr++;
6937
6938                         if (dev_nr >= dev_min)
6939                                 break;
6940
6941                         ret = -1;
6942                 }
6943         }
6944         mutex_unlock(&root->fs_info->chunk_mutex);
6945 out:
6946         btrfs_put_block_group(block_group);
6947         return ret;
6948 }
6949
6950 static int find_first_block_group(struct btrfs_root *root,
6951                 struct btrfs_path *path, struct btrfs_key *key)
6952 {
6953         int ret = 0;
6954         struct btrfs_key found_key;
6955         struct extent_buffer *leaf;
6956         int slot;
6957
6958         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
6959         if (ret < 0)
6960                 goto out;
6961
6962         while (1) {
6963                 slot = path->slots[0];
6964                 leaf = path->nodes[0];
6965                 if (slot >= btrfs_header_nritems(leaf)) {
6966                         ret = btrfs_next_leaf(root, path);
6967                         if (ret == 0)
6968                                 continue;
6969                         if (ret < 0)
6970                                 goto out;
6971                         break;
6972                 }
6973                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
6974
6975                 if (found_key.objectid >= key->objectid &&
6976                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
6977                         ret = 0;
6978                         goto out;
6979                 }
6980                 path->slots[0]++;
6981         }
6982 out:
6983         return ret;
6984 }
6985
6986 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
6987 {
6988         struct btrfs_block_group_cache *block_group;
6989         u64 last = 0;
6990
6991         while (1) {
6992                 struct inode *inode;
6993
6994                 block_group = btrfs_lookup_first_block_group(info, last);
6995                 while (block_group) {
6996                         spin_lock(&block_group->lock);
6997                         if (block_group->iref)
6998                                 break;
6999                         spin_unlock(&block_group->lock);
7000                         block_group = next_block_group(info->tree_root,
7001                                                        block_group);
7002                 }
7003                 if (!block_group) {
7004                         if (last == 0)
7005                                 break;
7006                         last = 0;
7007                         continue;
7008                 }
7009
7010                 inode = block_group->inode;
7011                 block_group->iref = 0;
7012                 block_group->inode = NULL;
7013                 spin_unlock(&block_group->lock);
7014                 iput(inode);
7015                 last = block_group->key.objectid + block_group->key.offset;
7016                 btrfs_put_block_group(block_group);
7017         }
7018 }
7019
7020 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7021 {
7022         struct btrfs_block_group_cache *block_group;
7023         struct btrfs_space_info *space_info;
7024         struct btrfs_caching_control *caching_ctl;
7025         struct rb_node *n;
7026
7027         down_write(&info->extent_commit_sem);
7028         while (!list_empty(&info->caching_block_groups)) {
7029                 caching_ctl = list_entry(info->caching_block_groups.next,
7030                                          struct btrfs_caching_control, list);
7031                 list_del(&caching_ctl->list);
7032                 put_caching_control(caching_ctl);
7033         }
7034         up_write(&info->extent_commit_sem);
7035
7036         spin_lock(&info->block_group_cache_lock);
7037         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7038                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7039                                        cache_node);
7040                 rb_erase(&block_group->cache_node,
7041                          &info->block_group_cache_tree);
7042                 spin_unlock(&info->block_group_cache_lock);
7043
7044                 down_write(&block_group->space_info->groups_sem);
7045                 list_del(&block_group->list);
7046                 up_write(&block_group->space_info->groups_sem);
7047
7048                 if (block_group->cached == BTRFS_CACHE_STARTED)
7049                         wait_block_group_cache_done(block_group);
7050
7051                 /*
7052                  * We haven't cached this block group, which means we could
7053                  * possibly have excluded extents on this block group.
7054                  */
7055                 if (block_group->cached == BTRFS_CACHE_NO)
7056                         free_excluded_extents(info->extent_root, block_group);
7057
7058                 btrfs_remove_free_space_cache(block_group);
7059                 btrfs_put_block_group(block_group);
7060
7061                 spin_lock(&info->block_group_cache_lock);
7062         }
7063         spin_unlock(&info->block_group_cache_lock);
7064
7065         /* now that all the block groups are freed, go through and
7066          * free all the space_info structs.  This is only called during
7067          * the final stages of unmount, and so we know nobody is
7068          * using them.  We call synchronize_rcu() once before we start,
7069          * just to be on the safe side.
7070          */
7071         synchronize_rcu();
7072
7073         release_global_block_rsv(info);
7074
7075         while(!list_empty(&info->space_info)) {
7076                 space_info = list_entry(info->space_info.next,
7077                                         struct btrfs_space_info,
7078                                         list);
7079                 if (space_info->bytes_pinned > 0 ||
7080                     space_info->bytes_reserved > 0 ||
7081                     space_info->bytes_may_use > 0) {
7082                         WARN_ON(1);
7083                         dump_space_info(space_info, 0, 0);
7084                 }
7085                 list_del(&space_info->list);
7086                 kfree(space_info);
7087         }
7088         return 0;
7089 }
7090
7091 static void __link_block_group(struct btrfs_space_info *space_info,
7092                                struct btrfs_block_group_cache *cache)
7093 {
7094         int index = get_block_group_index(cache);
7095
7096         down_write(&space_info->groups_sem);
7097         list_add_tail(&cache->list, &space_info->block_groups[index]);
7098         up_write(&space_info->groups_sem);
7099 }
7100
7101 int btrfs_read_block_groups(struct btrfs_root *root)
7102 {
7103         struct btrfs_path *path;
7104         int ret;
7105         struct btrfs_block_group_cache *cache;
7106         struct btrfs_fs_info *info = root->fs_info;
7107         struct btrfs_space_info *space_info;
7108         struct btrfs_key key;
7109         struct btrfs_key found_key;
7110         struct extent_buffer *leaf;
7111         int need_clear = 0;
7112         u64 cache_gen;
7113
7114         root = info->extent_root;
7115         key.objectid = 0;
7116         key.offset = 0;
7117         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7118         path = btrfs_alloc_path();
7119         if (!path)
7120                 return -ENOMEM;
7121         path->reada = 1;
7122
7123         cache_gen = btrfs_super_cache_generation(&root->fs_info->super_copy);
7124         if (cache_gen != 0 &&
7125             btrfs_super_generation(&root->fs_info->super_copy) != cache_gen)
7126                 need_clear = 1;
7127         if (btrfs_test_opt(root, CLEAR_CACHE))
7128                 need_clear = 1;
7129         if (!btrfs_test_opt(root, SPACE_CACHE) && cache_gen)
7130                 printk(KERN_INFO "btrfs: disk space caching is enabled\n");
7131
7132         while (1) {
7133                 ret = find_first_block_group(root, path, &key);
7134                 if (ret > 0)
7135                         break;
7136                 if (ret != 0)
7137                         goto error;
7138                 leaf = path->nodes[0];
7139                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7140                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7141                 if (!cache) {
7142                         ret = -ENOMEM;
7143                         goto error;
7144                 }
7145                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7146                                                 GFP_NOFS);
7147                 if (!cache->free_space_ctl) {
7148                         kfree(cache);
7149                         ret = -ENOMEM;
7150                         goto error;
7151                 }
7152
7153                 atomic_set(&cache->count, 1);
7154                 spin_lock_init(&cache->lock);
7155                 cache->fs_info = info;
7156                 INIT_LIST_HEAD(&cache->list);
7157                 INIT_LIST_HEAD(&cache->cluster_list);
7158
7159                 if (need_clear)
7160                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7161
7162                 read_extent_buffer(leaf, &cache->item,
7163                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7164                                    sizeof(cache->item));
7165                 memcpy(&cache->key, &found_key, sizeof(found_key));
7166
7167                 key.objectid = found_key.objectid + found_key.offset;
7168                 btrfs_release_path(path);
7169                 cache->flags = btrfs_block_group_flags(&cache->item);
7170                 cache->sectorsize = root->sectorsize;
7171
7172                 btrfs_init_free_space_ctl(cache);
7173
7174                 /*
7175                  * We need to exclude the super stripes now so that the space
7176                  * info has super bytes accounted for, otherwise we'll think
7177                  * we have more space than we actually do.
7178                  */
7179                 exclude_super_stripes(root, cache);
7180
7181                 /*
7182                  * check for two cases, either we are full, and therefore
7183                  * don't need to bother with the caching work since we won't
7184                  * find any space, or we are empty, and we can just add all
7185                  * the space in and be done with it.  This saves us _alot_ of
7186                  * time, particularly in the full case.
7187                  */
7188                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7189                         cache->last_byte_to_unpin = (u64)-1;
7190                         cache->cached = BTRFS_CACHE_FINISHED;
7191                         free_excluded_extents(root, cache);
7192                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7193                         cache->last_byte_to_unpin = (u64)-1;
7194                         cache->cached = BTRFS_CACHE_FINISHED;
7195                         add_new_free_space(cache, root->fs_info,
7196                                            found_key.objectid,
7197                                            found_key.objectid +
7198                                            found_key.offset);
7199                         free_excluded_extents(root, cache);
7200                 }
7201
7202                 ret = update_space_info(info, cache->flags, found_key.offset,
7203                                         btrfs_block_group_used(&cache->item),
7204                                         &space_info);
7205                 BUG_ON(ret);
7206                 cache->space_info = space_info;
7207                 spin_lock(&cache->space_info->lock);
7208                 cache->space_info->bytes_readonly += cache->bytes_super;
7209                 spin_unlock(&cache->space_info->lock);
7210
7211                 __link_block_group(space_info, cache);
7212
7213                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7214                 BUG_ON(ret);
7215
7216                 set_avail_alloc_bits(root->fs_info, cache->flags);
7217                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7218                         set_block_group_ro(cache, 1);
7219         }
7220
7221         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7222                 if (!(get_alloc_profile(root, space_info->flags) &
7223                       (BTRFS_BLOCK_GROUP_RAID10 |
7224                        BTRFS_BLOCK_GROUP_RAID1 |
7225                        BTRFS_BLOCK_GROUP_DUP)))
7226                         continue;
7227                 /*
7228                  * avoid allocating from un-mirrored block group if there are
7229                  * mirrored block groups.
7230                  */
7231                 list_for_each_entry(cache, &space_info->block_groups[3], list)
7232                         set_block_group_ro(cache, 1);
7233                 list_for_each_entry(cache, &space_info->block_groups[4], list)
7234                         set_block_group_ro(cache, 1);
7235         }
7236
7237         init_global_block_rsv(info);
7238         ret = 0;
7239 error:
7240         btrfs_free_path(path);
7241         return ret;
7242 }
7243
7244 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7245                            struct btrfs_root *root, u64 bytes_used,
7246                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7247                            u64 size)
7248 {
7249         int ret;
7250         struct btrfs_root *extent_root;
7251         struct btrfs_block_group_cache *cache;
7252
7253         extent_root = root->fs_info->extent_root;
7254
7255         root->fs_info->last_trans_log_full_commit = trans->transid;
7256
7257         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7258         if (!cache)
7259                 return -ENOMEM;
7260         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7261                                         GFP_NOFS);
7262         if (!cache->free_space_ctl) {
7263                 kfree(cache);
7264                 return -ENOMEM;
7265         }
7266
7267         cache->key.objectid = chunk_offset;
7268         cache->key.offset = size;
7269         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7270         cache->sectorsize = root->sectorsize;
7271         cache->fs_info = root->fs_info;
7272
7273         atomic_set(&cache->count, 1);
7274         spin_lock_init(&cache->lock);
7275         INIT_LIST_HEAD(&cache->list);
7276         INIT_LIST_HEAD(&cache->cluster_list);
7277
7278         btrfs_init_free_space_ctl(cache);
7279
7280         btrfs_set_block_group_used(&cache->item, bytes_used);
7281         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7282         cache->flags = type;
7283         btrfs_set_block_group_flags(&cache->item, type);
7284
7285         cache->last_byte_to_unpin = (u64)-1;
7286         cache->cached = BTRFS_CACHE_FINISHED;
7287         exclude_super_stripes(root, cache);
7288
7289         add_new_free_space(cache, root->fs_info, chunk_offset,
7290                            chunk_offset + size);
7291
7292         free_excluded_extents(root, cache);
7293
7294         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7295                                 &cache->space_info);
7296         BUG_ON(ret);
7297
7298         spin_lock(&cache->space_info->lock);
7299         cache->space_info->bytes_readonly += cache->bytes_super;
7300         spin_unlock(&cache->space_info->lock);
7301
7302         __link_block_group(cache->space_info, cache);
7303
7304         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7305         BUG_ON(ret);
7306
7307         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
7308                                 sizeof(cache->item));
7309         BUG_ON(ret);
7310
7311         set_avail_alloc_bits(extent_root->fs_info, type);
7312
7313         return 0;
7314 }
7315
7316 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7317                              struct btrfs_root *root, u64 group_start)
7318 {
7319         struct btrfs_path *path;
7320         struct btrfs_block_group_cache *block_group;
7321         struct btrfs_free_cluster *cluster;
7322         struct btrfs_root *tree_root = root->fs_info->tree_root;
7323         struct btrfs_key key;
7324         struct inode *inode;
7325         int ret;
7326         int factor;
7327
7328         root = root->fs_info->extent_root;
7329
7330         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7331         BUG_ON(!block_group);
7332         BUG_ON(!block_group->ro);
7333
7334         /*
7335          * Free the reserved super bytes from this block group before
7336          * remove it.
7337          */
7338         free_excluded_extents(root, block_group);
7339
7340         memcpy(&key, &block_group->key, sizeof(key));
7341         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7342                                   BTRFS_BLOCK_GROUP_RAID1 |
7343                                   BTRFS_BLOCK_GROUP_RAID10))
7344                 factor = 2;
7345         else
7346                 factor = 1;
7347
7348         /* make sure this block group isn't part of an allocation cluster */
7349         cluster = &root->fs_info->data_alloc_cluster;
7350         spin_lock(&cluster->refill_lock);
7351         btrfs_return_cluster_to_free_space(block_group, cluster);
7352         spin_unlock(&cluster->refill_lock);
7353
7354         /*
7355          * make sure this block group isn't part of a metadata
7356          * allocation cluster
7357          */
7358         cluster = &root->fs_info->meta_alloc_cluster;
7359         spin_lock(&cluster->refill_lock);
7360         btrfs_return_cluster_to_free_space(block_group, cluster);
7361         spin_unlock(&cluster->refill_lock);
7362
7363         path = btrfs_alloc_path();
7364         if (!path) {
7365                 ret = -ENOMEM;
7366                 goto out;
7367         }
7368
7369         inode = lookup_free_space_inode(root, block_group, path);
7370         if (!IS_ERR(inode)) {
7371                 ret = btrfs_orphan_add(trans, inode);
7372                 BUG_ON(ret);
7373                 clear_nlink(inode);
7374                 /* One for the block groups ref */
7375                 spin_lock(&block_group->lock);
7376                 if (block_group->iref) {
7377                         block_group->iref = 0;
7378                         block_group->inode = NULL;
7379                         spin_unlock(&block_group->lock);
7380                         iput(inode);
7381                 } else {
7382                         spin_unlock(&block_group->lock);
7383                 }
7384                 /* One for our lookup ref */
7385                 iput(inode);
7386         }
7387
7388         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
7389         key.offset = block_group->key.objectid;
7390         key.type = 0;
7391
7392         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
7393         if (ret < 0)
7394                 goto out;
7395         if (ret > 0)
7396                 btrfs_release_path(path);
7397         if (ret == 0) {
7398                 ret = btrfs_del_item(trans, tree_root, path);
7399                 if (ret)
7400                         goto out;
7401                 btrfs_release_path(path);
7402         }
7403
7404         spin_lock(&root->fs_info->block_group_cache_lock);
7405         rb_erase(&block_group->cache_node,
7406                  &root->fs_info->block_group_cache_tree);
7407         spin_unlock(&root->fs_info->block_group_cache_lock);
7408
7409         down_write(&block_group->space_info->groups_sem);
7410         /*
7411          * we must use list_del_init so people can check to see if they
7412          * are still on the list after taking the semaphore
7413          */
7414         list_del_init(&block_group->list);
7415         up_write(&block_group->space_info->groups_sem);
7416
7417         if (block_group->cached == BTRFS_CACHE_STARTED)
7418                 wait_block_group_cache_done(block_group);
7419
7420         btrfs_remove_free_space_cache(block_group);
7421
7422         spin_lock(&block_group->space_info->lock);
7423         block_group->space_info->total_bytes -= block_group->key.offset;
7424         block_group->space_info->bytes_readonly -= block_group->key.offset;
7425         block_group->space_info->disk_total -= block_group->key.offset * factor;
7426         spin_unlock(&block_group->space_info->lock);
7427
7428         memcpy(&key, &block_group->key, sizeof(key));
7429
7430         btrfs_clear_space_info_full(root->fs_info);
7431
7432         btrfs_put_block_group(block_group);
7433         btrfs_put_block_group(block_group);
7434
7435         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
7436         if (ret > 0)
7437                 ret = -EIO;
7438         if (ret < 0)
7439                 goto out;
7440
7441         ret = btrfs_del_item(trans, root, path);
7442 out:
7443         btrfs_free_path(path);
7444         return ret;
7445 }
7446
7447 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
7448 {
7449         struct btrfs_space_info *space_info;
7450         struct btrfs_super_block *disk_super;
7451         u64 features;
7452         u64 flags;
7453         int mixed = 0;
7454         int ret;
7455
7456         disk_super = &fs_info->super_copy;
7457         if (!btrfs_super_root(disk_super))
7458                 return 1;
7459
7460         features = btrfs_super_incompat_flags(disk_super);
7461         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
7462                 mixed = 1;
7463
7464         flags = BTRFS_BLOCK_GROUP_SYSTEM;
7465         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7466         if (ret)
7467                 goto out;
7468
7469         if (mixed) {
7470                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
7471                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7472         } else {
7473                 flags = BTRFS_BLOCK_GROUP_METADATA;
7474                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7475                 if (ret)
7476                         goto out;
7477
7478                 flags = BTRFS_BLOCK_GROUP_DATA;
7479                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7480         }
7481 out:
7482         return ret;
7483 }
7484
7485 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
7486 {
7487         return unpin_extent_range(root, start, end);
7488 }
7489
7490 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
7491                                u64 num_bytes, u64 *actual_bytes)
7492 {
7493         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
7494 }
7495
7496 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
7497 {
7498         struct btrfs_fs_info *fs_info = root->fs_info;
7499         struct btrfs_block_group_cache *cache = NULL;
7500         u64 group_trimmed;
7501         u64 start;
7502         u64 end;
7503         u64 trimmed = 0;
7504         int ret = 0;
7505
7506         cache = btrfs_lookup_block_group(fs_info, range->start);
7507
7508         while (cache) {
7509                 if (cache->key.objectid >= (range->start + range->len)) {
7510                         btrfs_put_block_group(cache);
7511                         break;
7512                 }
7513
7514                 start = max(range->start, cache->key.objectid);
7515                 end = min(range->start + range->len,
7516                                 cache->key.objectid + cache->key.offset);
7517
7518                 if (end - start >= range->minlen) {
7519                         if (!block_group_cache_done(cache)) {
7520                                 ret = cache_block_group(cache, NULL, root, 0);
7521                                 if (!ret)
7522                                         wait_block_group_cache_done(cache);
7523                         }
7524                         ret = btrfs_trim_block_group(cache,
7525                                                      &group_trimmed,
7526                                                      start,
7527                                                      end,
7528                                                      range->minlen);
7529
7530                         trimmed += group_trimmed;
7531                         if (ret) {
7532                                 btrfs_put_block_group(cache);
7533                                 break;
7534                         }
7535                 }
7536
7537                 cache = next_block_group(fs_info->tree_root, cache);
7538         }
7539
7540         range->len = trimmed;
7541         return ret;
7542 }