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