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