fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include "compat.h"
  27 #include "hash.h"
  28 #include "ctree.h"
  29 #include "disk-io.h"
  30 #include "print-tree.h"
  31 #include "transaction.h"
  32 #include "volumes.h"
  33 #include "locking.h"
  34 #include "free-space-cache.h"
  35
  36 /* control flags for do_chunk_alloc's force field
  37  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  38  * if we really need one.
  39  *
  40  * CHUNK_ALLOC_FORCE means it must try to allocate one
  41  *
  42  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  43  * if we have very few chunks already allocated.  This is
  44  * used as part of the clustering code to help make sure
  45  * we have a good pool of storage to cluster in, without
  46  * filling the FS with empty chunks
  47  *
  48  */
  49 enum {
  50         CHUNK_ALLOC_NO_FORCE = 0,
  51         CHUNK_ALLOC_FORCE = 1,
  52         CHUNK_ALLOC_LIMITED = 2,
  53 };
  54
  55 /*
  56  * Control how reservations are dealt with.
  57  *
  58  * RESERVE_FREE - freeing a reservation.
  59  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  60  *   ENOSPC accounting
  61  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  62  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  63  */
  64 enum {
  65         RESERVE_FREE = 0,
  66         RESERVE_ALLOC = 1,
  67         RESERVE_ALLOC_NO_ACCOUNT = 2,
  68 };
  69
  70 static int update_block_group(struct btrfs_trans_handle *trans,
  71                               struct btrfs_root *root,
  72                               u64 bytenr, u64 num_bytes, int alloc);
  73 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  74                                 struct btrfs_root *root,
  75                                 u64 bytenr, u64 num_bytes, u64 parent,
  76                                 u64 root_objectid, u64 owner_objectid,
  77                                 u64 owner_offset, int refs_to_drop,
  78                                 struct btrfs_delayed_extent_op *extra_op);
  79 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  80                                     struct extent_buffer *leaf,
  81                                     struct btrfs_extent_item *ei);
  82 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  83                                       struct btrfs_root *root,
  84                                       u64 parent, u64 root_objectid,
  85                                       u64 flags, u64 owner, u64 offset,
  86                                       struct btrfs_key *ins, int ref_mod);
  87 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  88                                      struct btrfs_root *root,
  89                                      u64 parent, u64 root_objectid,
  90                                      u64 flags, struct btrfs_disk_key *key,
  91                                      int level, struct btrfs_key *ins);
  92 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  93                           struct btrfs_root *extent_root, u64 alloc_bytes,
  94                           u64 flags, int force);
  95 static int find_next_key(struct btrfs_path *path, int level,
  96                          struct btrfs_key *key);
  97 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
  98                             int dump_block_groups);
  99 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 100                                        u64 num_bytes, int reserve);
 101
 102 static noinline int
 103 block_group_cache_done(struct btrfs_block_group_cache *cache)
 104 {
 105         smp_mb();
 106         return cache->cached == BTRFS_CACHE_FINISHED;
 107 }
 108
 109 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 110 {
 111         return (cache->flags & bits) == bits;
 112 }
 113
 114 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 115 {
 116         atomic_inc(&cache->count);
 117 }
 118
 119 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 120 {
 121         if (atomic_dec_and_test(&cache->count)) {
 122                 WARN_ON(cache->pinned > 0);
 123                 WARN_ON(cache->reserved > 0);
 124                 kfree(cache->free_space_ctl);
 125                 kfree(cache);
 126         }
 127 }
 128
 129 /*
 130  * this adds the block group to the fs_info rb tree for the block group
 131  * cache
 132  */
 133 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 134                                 struct btrfs_block_group_cache *block_group)
 135 {
 136         struct rb_node **p;
 137         struct rb_node *parent = NULL;
 138         struct btrfs_block_group_cache *cache;
 139
 140         spin_lock(&info->block_group_cache_lock);
 141         p = &info->block_group_cache_tree.rb_node;
 142
 143         while (*p) {
 144                 parent = *p;
 145                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 146                                  cache_node);
 147                 if (block_group->key.objectid < cache->key.objectid) {
 148                         p = &(*p)->rb_left;
 149                 } else if (block_group->key.objectid > cache->key.objectid) {
 150                         p = &(*p)->rb_right;
 151                 } else {
 152                         spin_unlock(&info->block_group_cache_lock);
 153                         return -EEXIST;
 154                 }
 155         }
 156
 157         rb_link_node(&block_group->cache_node, parent, p);
 158         rb_insert_color(&block_group->cache_node,
 159                         &info->block_group_cache_tree);
 160         spin_unlock(&info->block_group_cache_lock);
 161
 162         return 0;
 163 }
 164
 165 /*
 166  * This will return the block group at or after bytenr if contains is 0, else
 167  * it will return the block group that contains the bytenr
 168  */
 169 static struct btrfs_block_group_cache *
 170 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 171                               int contains)
 172 {
 173         struct btrfs_block_group_cache *cache, *ret = NULL;
 174         struct rb_node *n;
 175         u64 end, start;
 176
 177         spin_lock(&info->block_group_cache_lock);
 178         n = info->block_group_cache_tree.rb_node;
 179
 180         while (n) {
 181                 cache = rb_entry(n, struct btrfs_block_group_cache,
 182                                  cache_node);
 183                 end = cache->key.objectid + cache->key.offset - 1;
 184                 start = cache->key.objectid;
 185
 186                 if (bytenr < start) {
 187                         if (!contains && (!ret || start < ret->key.objectid))
 188                                 ret = cache;
 189                         n = n->rb_left;
 190                 } else if (bytenr > start) {
 191                         if (contains && bytenr <= end) {
 192                                 ret = cache;
 193                                 break;
 194                         }
 195                         n = n->rb_right;
 196                 } else {
 197                         ret = cache;
 198                         break;
 199                 }
 200         }
 201         if (ret)
 202                 btrfs_get_block_group(ret);
 203         spin_unlock(&info->block_group_cache_lock);
 204
 205         return ret;
 206 }
 207
 208 static int add_excluded_extent(struct btrfs_root *root,
 209                                u64 start, u64 num_bytes)
 210 {
 211         u64 end = start + num_bytes - 1;
 212         set_extent_bits(&root->fs_info->freed_extents[0],
 213                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 214         set_extent_bits(&root->fs_info->freed_extents[1],
 215                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 216         return 0;
 217 }
 218
 219 static void free_excluded_extents(struct btrfs_root *root,
 220                                   struct btrfs_block_group_cache *cache)
 221 {
 222         u64 start, end;
 223
 224         start = cache->key.objectid;
 225         end = start + cache->key.offset - 1;
 226
 227         clear_extent_bits(&root->fs_info->freed_extents[0],
 228                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 229         clear_extent_bits(&root->fs_info->freed_extents[1],
 230                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 231 }
 232
 233 static int exclude_super_stripes(struct btrfs_root *root,
 234                                  struct btrfs_block_group_cache *cache)
 235 {
 236         u64 bytenr;
 237         u64 *logical;
 238         int stripe_len;
 239         int i, nr, ret;
 240
 241         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 242                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 243                 cache->bytes_super += stripe_len;
 244                 ret = add_excluded_extent(root, cache->key.objectid,
 245                                           stripe_len);
 246                 BUG_ON(ret);
 247         }
 248
 249         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 250                 bytenr = btrfs_sb_offset(i);
 251                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 252                                        cache->key.objectid, bytenr,
 253                                        0, &logical, &nr, &stripe_len);
 254                 BUG_ON(ret);
 255
 256                 while (nr--) {
 257                         cache->bytes_super += stripe_len;
 258                         ret = add_excluded_extent(root, logical[nr],
 259                                                   stripe_len);
 260                         BUG_ON(ret);
 261                 }
 262
 263                 kfree(logical);
 264         }
 265         return 0;
 266 }
 267
 268 static struct btrfs_caching_control *
 269 get_caching_control(struct btrfs_block_group_cache *cache)
 270 {
 271         struct btrfs_caching_control *ctl;
 272
 273         spin_lock(&cache->lock);
 274         if (cache->cached != BTRFS_CACHE_STARTED) {
 275                 spin_unlock(&cache->lock);
 276                 return NULL;
 277         }
 278
 279         /* We're loading it the fast way, so we don't have a caching_ctl. */
 280         if (!cache->caching_ctl) {
 281                 spin_unlock(&cache->lock);
 282                 return NULL;
 283         }
 284
 285         ctl = cache->caching_ctl;
 286         atomic_inc(&ctl->count);
 287         spin_unlock(&cache->lock);
 288         return ctl;
 289 }
 290
 291 static void put_caching_control(struct btrfs_caching_control *ctl)
 292 {
 293         if (atomic_dec_and_test(&ctl->count))
 294                 kfree(ctl);
 295 }
 296
 297 /*
 298  * this is only called by cache_block_group, since we could have freed extents
 299  * we need to check the pinned_extents for any extents that can't be used yet
 300  * since their free space will be released as soon as the transaction commits.
 301  */
 302 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 303                               struct btrfs_fs_info *info, u64 start, u64 end)
 304 {
 305         u64 extent_start, extent_end, size, total_added = 0;
 306         int ret;
 307
 308         while (start < end) {
 309                 ret = find_first_extent_bit(info->pinned_extents, start,
 310                                             &extent_start, &extent_end,
 311                                             EXTENT_DIRTY | EXTENT_UPTODATE);
 312                 if (ret)
 313                         break;
 314
 315                 if (extent_start <= start) {
 316                         start = extent_end + 1;
 317                 } else if (extent_start > start && extent_start < end) {
 318                         size = extent_start - start;
 319                         total_added += size;
 320                         ret = btrfs_add_free_space(block_group, start,
 321                                                    size);
 322                         BUG_ON(ret);
 323                         start = extent_end + 1;
 324                 } else {
 325                         break;
 326                 }
 327         }
 328
 329         if (start < end) {
 330                 size = end - start;
 331                 total_added += size;
 332                 ret = btrfs_add_free_space(block_group, start, size);
 333                 BUG_ON(ret);
 334         }
 335
 336         return total_added;
 337 }
 338
 339 static noinline void caching_thread(struct btrfs_work *work)
 340 {
 341         struct btrfs_block_group_cache *block_group;
 342         struct btrfs_fs_info *fs_info;
 343         struct btrfs_caching_control *caching_ctl;
 344         struct btrfs_root *extent_root;
 345         struct btrfs_path *path;
 346         struct extent_buffer *leaf;
 347         struct btrfs_key key;
 348         u64 total_found = 0;
 349         u64 last = 0;
 350         u32 nritems;
 351         int ret = 0;
 352
 353         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 354         block_group = caching_ctl->block_group;
 355         fs_info = block_group->fs_info;
 356         extent_root = fs_info->extent_root;
 357
 358         path = btrfs_alloc_path();
 359         if (!path)
 360                 goto out;
 361
 362         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 363
 364         /*
 365          * We don't want to deadlock with somebody trying to allocate a new
 366          * extent for the extent root while also trying to search the extent
 367          * root to add free space.  So we skip locking and search the commit
 368          * root, since its read-only
 369          */
 370         path->skip_locking = 1;
 371         path->search_commit_root = 1;
 372         path->reada = 1;
 373
 374         key.objectid = last;
 375         key.offset = 0;
 376         key.type = BTRFS_EXTENT_ITEM_KEY;
 377 again:
 378         mutex_lock(&caching_ctl->mutex);
 379         /* need to make sure the commit_root doesn't disappear */
 380         down_read(&fs_info->extent_commit_sem);
 381
 382         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 383         if (ret < 0)
 384                 goto err;
 385
 386         leaf = path->nodes[0];
 387         nritems = btrfs_header_nritems(leaf);
 388
 389         while (1) {
 390                 if (btrfs_fs_closing(fs_info) > 1) {
 391                         last = (u64)-1;
 392                         break;
 393                 }
 394
 395                 if (path->slots[0] < nritems) {
 396                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 397                 } else {
 398                         ret = find_next_key(path, 0, &key);
 399                         if (ret)
 400                                 break;
 401
 402                         if (need_resched() ||
 403                             btrfs_next_leaf(extent_root, path)) {
 404                                 caching_ctl->progress = last;
 405                                 btrfs_release_path(path);
 406                                 up_read(&fs_info->extent_commit_sem);
 407                                 mutex_unlock(&caching_ctl->mutex);
 408                                 cond_resched();
 409                                 goto again;
 410                         }
 411                         leaf = path->nodes[0];
 412                         nritems = btrfs_header_nritems(leaf);
 413                         continue;
 414                 }
 415
 416                 if (key.objectid < block_group->key.objectid) {
 417                         path->slots[0]++;
 418                         continue;
 419                 }
 420
 421                 if (key.objectid >= block_group->key.objectid +
 422                     block_group->key.offset)
 423                         break;
 424
 425                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 426                         total_found += add_new_free_space(block_group,
 427                                                           fs_info, last,
 428                                                           key.objectid);
 429                         last = key.objectid + key.offset;
 430
 431                         if (total_found > (1024 * 1024 * 2)) {
 432                                 total_found = 0;
 433                                 wake_up(&caching_ctl->wait);
 434                         }
 435                 }
 436                 path->slots[0]++;
 437         }
 438         ret = 0;
 439
 440         total_found += add_new_free_space(block_group, fs_info, last,
 441                                           block_group->key.objectid +
 442                                           block_group->key.offset);
 443         caching_ctl->progress = (u64)-1;
 444
 445         spin_lock(&block_group->lock);
 446         block_group->caching_ctl = NULL;
 447         block_group->cached = BTRFS_CACHE_FINISHED;
 448         spin_unlock(&block_group->lock);
 449
 450 err:
 451         btrfs_free_path(path);
 452         up_read(&fs_info->extent_commit_sem);
 453
 454         free_excluded_extents(extent_root, block_group);
 455
 456         mutex_unlock(&caching_ctl->mutex);
 457 out:
 458         wake_up(&caching_ctl->wait);
 459
 460         put_caching_control(caching_ctl);
 461         btrfs_put_block_group(block_group);
 462 }
 463
 464 static int cache_block_group(struct btrfs_block_group_cache *cache,
 465                              struct btrfs_trans_handle *trans,
 466                              struct btrfs_root *root,
 467                              int load_cache_only)
 468 {
 469         struct btrfs_fs_info *fs_info = cache->fs_info;
 470         struct btrfs_caching_control *caching_ctl;
 471         int ret = 0;
 472
 473         smp_mb();
 474         if (cache->cached != BTRFS_CACHE_NO)
 475                 return 0;
 476
 477         /*
 478          * We can't do the read from on-disk cache during a commit since we need
 479          * to have the normal tree locking.  Also if we are currently trying to
 480          * allocate blocks for the tree root we can't do the fast caching since
 481          * we likely hold important locks.
 482          */
 483         if (trans && (!trans->transaction->in_commit) &&
 484             (root && root != root->fs_info->tree_root) &&
 485             btrfs_test_opt(root, SPACE_CACHE)) {
 486                 spin_lock(&cache->lock);
 487                 if (cache->cached != BTRFS_CACHE_NO) {
 488                         spin_unlock(&cache->lock);
 489                         return 0;
 490                 }
 491                 cache->cached = BTRFS_CACHE_STARTED;
 492                 spin_unlock(&cache->lock);
 493
 494                 ret = load_free_space_cache(fs_info, cache);
 495
 496                 spin_lock(&cache->lock);
 497                 if (ret == 1) {
 498                         cache->cached = BTRFS_CACHE_FINISHED;
 499                         cache->last_byte_to_unpin = (u64)-1;
 500                 } else {
 501                         cache->cached = BTRFS_CACHE_NO;
 502                 }
 503                 spin_unlock(&cache->lock);
 504                 if (ret == 1) {
 505                         free_excluded_extents(fs_info->extent_root, cache);
 506                         return 0;
 507                 }
 508         }
 509
 510         if (load_cache_only)
 511                 return 0;
 512
 513         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 514         BUG_ON(!caching_ctl);
 515
 516         INIT_LIST_HEAD(&caching_ctl->list);
 517         mutex_init(&caching_ctl->mutex);
 518         init_waitqueue_head(&caching_ctl->wait);
 519         caching_ctl->block_group = cache;
 520         caching_ctl->progress = cache->key.objectid;
 521         /* one for caching kthread, one for caching block group list */
 522         atomic_set(&caching_ctl->count, 2);
 523         caching_ctl->work.func = caching_thread;
 524
 525         spin_lock(&cache->lock);
 526         if (cache->cached != BTRFS_CACHE_NO) {
 527                 spin_unlock(&cache->lock);
 528                 kfree(caching_ctl);
 529                 return 0;
 530         }
 531         cache->caching_ctl = caching_ctl;
 532         cache->cached = BTRFS_CACHE_STARTED;
 533         spin_unlock(&cache->lock);
 534
 535         down_write(&fs_info->extent_commit_sem);
 536         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 537         up_write(&fs_info->extent_commit_sem);
 538
 539         btrfs_get_block_group(cache);
 540
 541         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 542
 543         return ret;
 544 }
 545
 546 /*
 547  * return the block group that starts at or after bytenr
 548  */
 549 static struct btrfs_block_group_cache *
 550 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 551 {
 552         struct btrfs_block_group_cache *cache;
 553
 554         cache = block_group_cache_tree_search(info, bytenr, 0);
 555
 556         return cache;
 557 }
 558
 559 /*
 560  * return the block group that contains the given bytenr
 561  */
 562 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 563                                                  struct btrfs_fs_info *info,
 564                                                  u64 bytenr)
 565 {
 566         struct btrfs_block_group_cache *cache;
 567
 568         cache = block_group_cache_tree_search(info, bytenr, 1);
 569
 570         return cache;
 571 }
 572
 573 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 574                                                   u64 flags)
 575 {
 576         struct list_head *head = &info->space_info;
 577         struct btrfs_space_info *found;
 578
 579         flags &= BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_SYSTEM |
 580                  BTRFS_BLOCK_GROUP_METADATA;
 581
 582         rcu_read_lock();
 583         list_for_each_entry_rcu(found, head, list) {
 584                 if (found->flags & flags) {
 585                         rcu_read_unlock();
 586                         return found;
 587                 }
 588         }
 589         rcu_read_unlock();
 590         return NULL;
 591 }
 592
 593 /*
 594  * after adding space to the filesystem, we need to clear the full flags
 595  * on all the space infos.
 596  */
 597 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 598 {
 599         struct list_head *head = &info->space_info;
 600         struct btrfs_space_info *found;
 601
 602         rcu_read_lock();
 603         list_for_each_entry_rcu(found, head, list)
 604                 found->full = 0;
 605         rcu_read_unlock();
 606 }
 607
 608 static u64 div_factor(u64 num, int factor)
 609 {
 610         if (factor == 10)
 611                 return num;
 612         num *= factor;
 613         do_div(num, 10);
 614         return num;
 615 }
 616
 617 static u64 div_factor_fine(u64 num, int factor)
 618 {
 619         if (factor == 100)
 620                 return num;
 621         num *= factor;
 622         do_div(num, 100);
 623         return num;
 624 }
 625
 626 u64 btrfs_find_block_group(struct btrfs_root *root,
 627                            u64 search_start, u64 search_hint, int owner)
 628 {
 629         struct btrfs_block_group_cache *cache;
 630         u64 used;
 631         u64 last = max(search_hint, search_start);
 632         u64 group_start = 0;
 633         int full_search = 0;
 634         int factor = 9;
 635         int wrapped = 0;
 636 again:
 637         while (1) {
 638                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
 639                 if (!cache)
 640                         break;
 641
 642                 spin_lock(&cache->lock);
 643                 last = cache->key.objectid + cache->key.offset;
 644                 used = btrfs_block_group_used(&cache->item);
 645
 646                 if ((full_search || !cache->ro) &&
 647                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
 648                         if (used + cache->pinned + cache->reserved <
 649                             div_factor(cache->key.offset, factor)) {
 650                                 group_start = cache->key.objectid;
 651                                 spin_unlock(&cache->lock);
 652                                 btrfs_put_block_group(cache);
 653                                 goto found;
 654                         }
 655                 }
 656                 spin_unlock(&cache->lock);
 657                 btrfs_put_block_group(cache);
 658                 cond_resched();
 659         }
 660         if (!wrapped) {
 661                 last = search_start;
 662                 wrapped = 1;
 663                 goto again;
 664         }
 665         if (!full_search && factor < 10) {
 666                 last = search_start;
 667                 full_search = 1;
 668                 factor = 10;
 669                 goto again;
 670         }
 671 found:
 672         return group_start;
 673 }
 674
 675 /* simple helper to search for an existing extent at a given offset */
 676 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 677 {
 678         int ret;
 679         struct btrfs_key key;
 680         struct btrfs_path *path;
 681
 682         path = btrfs_alloc_path();
 683         if (!path)
 684                 return -ENOMEM;
 685
 686         key.objectid = start;
 687         key.offset = len;
 688         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 689         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 690                                 0, 0);
 691         btrfs_free_path(path);
 692         return ret;
 693 }
 694
 695 /*
 696  * helper function to lookup reference count and flags of extent.
 697  *
 698  * the head node for delayed ref is used to store the sum of all the
 699  * reference count modifications queued up in the rbtree. the head
 700  * node may also store the extent flags to set. This way you can check
 701  * to see what the reference count and extent flags would be if all of
 702  * the delayed refs are not processed.
 703  */
 704 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 705                              struct btrfs_root *root, u64 bytenr,
 706                              u64 num_bytes, u64 *refs, u64 *flags)
 707 {
 708         struct btrfs_delayed_ref_head *head;
 709         struct btrfs_delayed_ref_root *delayed_refs;
 710         struct btrfs_path *path;
 711         struct btrfs_extent_item *ei;
 712         struct extent_buffer *leaf;
 713         struct btrfs_key key;
 714         u32 item_size;
 715         u64 num_refs;
 716         u64 extent_flags;
 717         int ret;
 718
 719         path = btrfs_alloc_path();
 720         if (!path)
 721                 return -ENOMEM;
 722
 723         key.objectid = bytenr;
 724         key.type = BTRFS_EXTENT_ITEM_KEY;
 725         key.offset = num_bytes;
 726         if (!trans) {
 727                 path->skip_locking = 1;
 728                 path->search_commit_root = 1;
 729         }
 730 again:
 731         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 732                                 &key, path, 0, 0);
 733         if (ret < 0)
 734                 goto out_free;
 735
 736         if (ret == 0) {
 737                 leaf = path->nodes[0];
 738                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 739                 if (item_size >= sizeof(*ei)) {
 740                         ei = btrfs_item_ptr(leaf, path->slots[0],
 741                                             struct btrfs_extent_item);
 742                         num_refs = btrfs_extent_refs(leaf, ei);
 743                         extent_flags = btrfs_extent_flags(leaf, ei);
 744                 } else {
 745 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 746                         struct btrfs_extent_item_v0 *ei0;
 747                         BUG_ON(item_size != sizeof(*ei0));
 748                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 749                                              struct btrfs_extent_item_v0);
 750                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 751                         /* FIXME: this isn't correct for data */
 752                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 753 #else
 754                         BUG();
 755 #endif
 756                 }
 757                 BUG_ON(num_refs == 0);
 758         } else {
 759                 num_refs = 0;
 760                 extent_flags = 0;
 761                 ret = 0;
 762         }
 763
 764         if (!trans)
 765                 goto out;
 766
 767         delayed_refs = &trans->transaction->delayed_refs;
 768         spin_lock(&delayed_refs->lock);
 769         head = btrfs_find_delayed_ref_head(trans, bytenr);
 770         if (head) {
 771                 if (!mutex_trylock(&head->mutex)) {
 772                         atomic_inc(&head->node.refs);
 773                         spin_unlock(&delayed_refs->lock);
 774
 775                         btrfs_release_path(path);
 776
 777                         /*
 778                          * Mutex was contended, block until it's released and try
 779                          * again
 780                          */
 781                         mutex_lock(&head->mutex);
 782                         mutex_unlock(&head->mutex);
 783                         btrfs_put_delayed_ref(&head->node);
 784                         goto again;
 785                 }
 786                 if (head->extent_op && head->extent_op->update_flags)
 787                         extent_flags |= head->extent_op->flags_to_set;
 788                 else
 789                         BUG_ON(num_refs == 0);
 790
 791                 num_refs += head->node.ref_mod;
 792                 mutex_unlock(&head->mutex);
 793         }
 794         spin_unlock(&delayed_refs->lock);
 795 out:
 796         WARN_ON(num_refs == 0);
 797         if (refs)
 798                 *refs = num_refs;
 799         if (flags)
 800                 *flags = extent_flags;
 801 out_free:
 802         btrfs_free_path(path);
 803         return ret;
 804 }
 805
 806 /*
 807  * Back reference rules.  Back refs have three main goals:
 808  *
 809  * 1) differentiate between all holders of references to an extent so that
 810  *    when a reference is dropped we can make sure it was a valid reference
 811  *    before freeing the extent.
 812  *
 813  * 2) Provide enough information to quickly find the holders of an extent
 814  *    if we notice a given block is corrupted or bad.
 815  *
 816  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 817  *    maintenance.  This is actually the same as #2, but with a slightly
 818  *    different use case.
 819  *
 820  * There are two kinds of back refs. The implicit back refs is optimized
 821  * for pointers in non-shared tree blocks. For a given pointer in a block,
 822  * back refs of this kind provide information about the block's owner tree
 823  * and the pointer's key. These information allow us to find the block by
 824  * b-tree searching. The full back refs is for pointers in tree blocks not
 825  * referenced by their owner trees. The location of tree block is recorded
 826  * in the back refs. Actually the full back refs is generic, and can be
 827  * used in all cases the implicit back refs is used. The major shortcoming
 828  * of the full back refs is its overhead. Every time a tree block gets
 829  * COWed, we have to update back refs entry for all pointers in it.
 830  *
 831  * For a newly allocated tree block, we use implicit back refs for
 832  * pointers in it. This means most tree related operations only involve
 833  * implicit back refs. For a tree block created in old transaction, the
 834  * only way to drop a reference to it is COW it. So we can detect the
 835  * event that tree block loses its owner tree's reference and do the
 836  * back refs conversion.
 837  *
 838  * When a tree block is COW'd through a tree, there are four cases:
 839  *
 840  * The reference count of the block is one and the tree is the block's
 841  * owner tree. Nothing to do in this case.
 842  *
 843  * The reference count of the block is one and the tree is not the
 844  * block's owner tree. In this case, full back refs is used for pointers
 845  * in the block. Remove these full back refs, add implicit back refs for
 846  * every pointers in the new block.
 847  *
 848  * The reference count of the block is greater than one and the tree is
 849  * the block's owner tree. In this case, implicit back refs is used for
 850  * pointers in the block. Add full back refs for every pointers in the
 851  * block, increase lower level extents' reference counts. The original
 852  * implicit back refs are entailed to the new block.
 853  *
 854  * The reference count of the block is greater than one and the tree is
 855  * not the block's owner tree. Add implicit back refs for every pointer in
 856  * the new block, increase lower level extents' reference count.
 857  *
 858  * Back Reference Key composing:
 859  *
 860  * The key objectid corresponds to the first byte in the extent,
 861  * The key type is used to differentiate between types of back refs.
 862  * There are different meanings of the key offset for different types
 863  * of back refs.
 864  *
 865  * File extents can be referenced by:
 866  *
 867  * - multiple snapshots, subvolumes, or different generations in one subvol
 868  * - different files inside a single subvolume
 869  * - different offsets inside a file (bookend extents in file.c)
 870  *
 871  * The extent ref structure for the implicit back refs has fields for:
 872  *
 873  * - Objectid of the subvolume root
 874  * - objectid of the file holding the reference
 875  * - original offset in the file
 876  * - how many bookend extents
 877  *
 878  * The key offset for the implicit back refs is hash of the first
 879  * three fields.
 880  *
 881  * The extent ref structure for the full back refs has field for:
 882  *
 883  * - number of pointers in the tree leaf
 884  *
 885  * The key offset for the implicit back refs is the first byte of
 886  * the tree leaf
 887  *
 888  * When a file extent is allocated, The implicit back refs is used.
 889  * the fields are filled in:
 890  *
 891  *     (root_key.objectid, inode objectid, offset in file, 1)
 892  *
 893  * When a file extent is removed file truncation, we find the
 894  * corresponding implicit back refs and check the following fields:
 895  *
 896  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 897  *
 898  * Btree extents can be referenced by:
 899  *
 900  * - Different subvolumes
 901  *
 902  * Both the implicit back refs and the full back refs for tree blocks
 903  * only consist of key. The key offset for the implicit back refs is
 904  * objectid of block's owner tree. The key offset for the full back refs
 905  * is the first byte of parent block.
 906  *
 907  * When implicit back refs is used, information about the lowest key and
 908  * level of the tree block are required. These information are stored in
 909  * tree block info structure.
 910  */
 911
 912 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 913 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 914                                   struct btrfs_root *root,
 915                                   struct btrfs_path *path,
 916                                   u64 owner, u32 extra_size)
 917 {
 918         struct btrfs_extent_item *item;
 919         struct btrfs_extent_item_v0 *ei0;
 920         struct btrfs_extent_ref_v0 *ref0;
 921         struct btrfs_tree_block_info *bi;
 922         struct extent_buffer *leaf;
 923         struct btrfs_key key;
 924         struct btrfs_key found_key;
 925         u32 new_size = sizeof(*item);
 926         u64 refs;
 927         int ret;
 928
 929         leaf = path->nodes[0];
 930         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 931
 932         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 933         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 934                              struct btrfs_extent_item_v0);
 935         refs = btrfs_extent_refs_v0(leaf, ei0);
 936
 937         if (owner == (u64)-1) {
 938                 while (1) {
 939                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 940                                 ret = btrfs_next_leaf(root, path);
 941                                 if (ret < 0)
 942                                         return ret;
 943                                 BUG_ON(ret > 0);
 944                                 leaf = path->nodes[0];
 945                         }
 946                         btrfs_item_key_to_cpu(leaf, &found_key,
 947                                               path->slots[0]);
 948                         BUG_ON(key.objectid != found_key.objectid);
 949                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 950                                 path->slots[0]++;
 951                                 continue;
 952                         }
 953                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 954                                               struct btrfs_extent_ref_v0);
 955                         owner = btrfs_ref_objectid_v0(leaf, ref0);
 956                         break;
 957                 }
 958         }
 959         btrfs_release_path(path);
 960
 961         if (owner < BTRFS_FIRST_FREE_OBJECTID)
 962                 new_size += sizeof(*bi);
 963
 964         new_size -= sizeof(*ei0);
 965         ret = btrfs_search_slot(trans, root, &key, path,
 966                                 new_size + extra_size, 1);
 967         if (ret < 0)
 968                 return ret;
 969         BUG_ON(ret);
 970
 971         ret = btrfs_extend_item(trans, root, path, new_size);
 972
 973         leaf = path->nodes[0];
 974         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 975         btrfs_set_extent_refs(leaf, item, refs);
 976         /* FIXME: get real generation */
 977         btrfs_set_extent_generation(leaf, item, 0);
 978         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 979                 btrfs_set_extent_flags(leaf, item,
 980                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
 981                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
 982                 bi = (struct btrfs_tree_block_info *)(item + 1);
 983                 /* FIXME: get first key of the block */
 984                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
 985                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
 986         } else {
 987                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
 988         }
 989         btrfs_mark_buffer_dirty(leaf);
 990         return 0;
 991 }
 992 #endif
 993
 994 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
 995 {
 996         u32 high_crc = ~(u32)0;
 997         u32 low_crc = ~(u32)0;
 998         __le64 lenum;
 999
1000         lenum = cpu_to_le64(root_objectid);
1001         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1002         lenum = cpu_to_le64(owner);
1003         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1004         lenum = cpu_to_le64(offset);
1005         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1006
1007         return ((u64)high_crc << 31) ^ (u64)low_crc;
1008 }
1009
1010 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1011                                      struct btrfs_extent_data_ref *ref)
1012 {
1013         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1014                                     btrfs_extent_data_ref_objectid(leaf, ref),
1015                                     btrfs_extent_data_ref_offset(leaf, ref));
1016 }
1017
1018 static int match_extent_data_ref(struct extent_buffer *leaf,
1019                                  struct btrfs_extent_data_ref *ref,
1020                                  u64 root_objectid, u64 owner, u64 offset)
1021 {
1022         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1023             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1024             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1025                 return 0;
1026         return 1;
1027 }
1028
1029 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1030                                            struct btrfs_root *root,
1031                                            struct btrfs_path *path,
1032                                            u64 bytenr, u64 parent,
1033                                            u64 root_objectid,
1034                                            u64 owner, u64 offset)
1035 {
1036         struct btrfs_key key;
1037         struct btrfs_extent_data_ref *ref;
1038         struct extent_buffer *leaf;
1039         u32 nritems;
1040         int ret;
1041         int recow;
1042         int err = -ENOENT;
1043
1044         key.objectid = bytenr;
1045         if (parent) {
1046                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1047                 key.offset = parent;
1048         } else {
1049                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1050                 key.offset = hash_extent_data_ref(root_objectid,
1051                                                   owner, offset);
1052         }
1053 again:
1054         recow = 0;
1055         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1056         if (ret < 0) {
1057                 err = ret;
1058                 goto fail;
1059         }
1060
1061         if (parent) {
1062                 if (!ret)
1063                         return 0;
1064 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1065                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1066                 btrfs_release_path(path);
1067                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1068                 if (ret < 0) {
1069                         err = ret;
1070                         goto fail;
1071                 }
1072                 if (!ret)
1073                         return 0;
1074 #endif
1075                 goto fail;
1076         }
1077
1078         leaf = path->nodes[0];
1079         nritems = btrfs_header_nritems(leaf);
1080         while (1) {
1081                 if (path->slots[0] >= nritems) {
1082                         ret = btrfs_next_leaf(root, path);
1083                         if (ret < 0)
1084                                 err = ret;
1085                         if (ret)
1086                                 goto fail;
1087
1088                         leaf = path->nodes[0];
1089                         nritems = btrfs_header_nritems(leaf);
1090                         recow = 1;
1091                 }
1092
1093                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1094                 if (key.objectid != bytenr ||
1095                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1096                         goto fail;
1097
1098                 ref = btrfs_item_ptr(leaf, path->slots[0],
1099                                      struct btrfs_extent_data_ref);
1100
1101                 if (match_extent_data_ref(leaf, ref, root_objectid,
1102                                           owner, offset)) {
1103                         if (recow) {
1104                                 btrfs_release_path(path);
1105                                 goto again;
1106                         }
1107                         err = 0;
1108                         break;
1109                 }
1110                 path->slots[0]++;
1111         }
1112 fail:
1113         return err;
1114 }
1115
1116 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1117                                            struct btrfs_root *root,
1118                                            struct btrfs_path *path,
1119                                            u64 bytenr, u64 parent,
1120                                            u64 root_objectid, u64 owner,
1121                                            u64 offset, int refs_to_add)
1122 {
1123         struct btrfs_key key;
1124         struct extent_buffer *leaf;
1125         u32 size;
1126         u32 num_refs;
1127         int ret;
1128
1129         key.objectid = bytenr;
1130         if (parent) {
1131                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1132                 key.offset = parent;
1133                 size = sizeof(struct btrfs_shared_data_ref);
1134         } else {
1135                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1136                 key.offset = hash_extent_data_ref(root_objectid,
1137                                                   owner, offset);
1138                 size = sizeof(struct btrfs_extent_data_ref);
1139         }
1140
1141         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1142         if (ret && ret != -EEXIST)
1143                 goto fail;
1144
1145         leaf = path->nodes[0];
1146         if (parent) {
1147                 struct btrfs_shared_data_ref *ref;
1148                 ref = btrfs_item_ptr(leaf, path->slots[0],
1149                                      struct btrfs_shared_data_ref);
1150                 if (ret == 0) {
1151                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1152                 } else {
1153                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1154                         num_refs += refs_to_add;
1155                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1156                 }
1157         } else {
1158                 struct btrfs_extent_data_ref *ref;
1159                 while (ret == -EEXIST) {
1160                         ref = btrfs_item_ptr(leaf, path->slots[0],
1161                                              struct btrfs_extent_data_ref);
1162                         if (match_extent_data_ref(leaf, ref, root_objectid,
1163                                                   owner, offset))
1164                                 break;
1165                         btrfs_release_path(path);
1166                         key.offset++;
1167                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1168                                                       size);
1169                         if (ret && ret != -EEXIST)
1170                                 goto fail;
1171
1172                         leaf = path->nodes[0];
1173                 }
1174                 ref = btrfs_item_ptr(leaf, path->slots[0],
1175                                      struct btrfs_extent_data_ref);
1176                 if (ret == 0) {
1177                         btrfs_set_extent_data_ref_root(leaf, ref,
1178                                                        root_objectid);
1179                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1180                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1181                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1182                 } else {
1183                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1184                         num_refs += refs_to_add;
1185                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1186                 }
1187         }
1188         btrfs_mark_buffer_dirty(leaf);
1189         ret = 0;
1190 fail:
1191         btrfs_release_path(path);
1192         return ret;
1193 }
1194
1195 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1196                                            struct btrfs_root *root,
1197                                            struct btrfs_path *path,
1198                                            int refs_to_drop)
1199 {
1200         struct btrfs_key key;
1201         struct btrfs_extent_data_ref *ref1 = NULL;
1202         struct btrfs_shared_data_ref *ref2 = NULL;
1203         struct extent_buffer *leaf;
1204         u32 num_refs = 0;
1205         int ret = 0;
1206
1207         leaf = path->nodes[0];
1208         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1209
1210         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1211                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1212                                       struct btrfs_extent_data_ref);
1213                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1214         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1215                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1216                                       struct btrfs_shared_data_ref);
1217                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1218 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1219         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1220                 struct btrfs_extent_ref_v0 *ref0;
1221                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1222                                       struct btrfs_extent_ref_v0);
1223                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1224 #endif
1225         } else {
1226                 BUG();
1227         }
1228
1229         BUG_ON(num_refs < refs_to_drop);
1230         num_refs -= refs_to_drop;
1231
1232         if (num_refs == 0) {
1233                 ret = btrfs_del_item(trans, root, path);
1234         } else {
1235                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1236                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1237                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1238                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1239 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1240                 else {
1241                         struct btrfs_extent_ref_v0 *ref0;
1242                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1243                                         struct btrfs_extent_ref_v0);
1244                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1245                 }
1246 #endif
1247                 btrfs_mark_buffer_dirty(leaf);
1248         }
1249         return ret;
1250 }
1251
1252 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1253                                           struct btrfs_path *path,
1254                                           struct btrfs_extent_inline_ref *iref)
1255 {
1256         struct btrfs_key key;
1257         struct extent_buffer *leaf;
1258         struct btrfs_extent_data_ref *ref1;
1259         struct btrfs_shared_data_ref *ref2;
1260         u32 num_refs = 0;
1261
1262         leaf = path->nodes[0];
1263         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1264         if (iref) {
1265                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1266                     BTRFS_EXTENT_DATA_REF_KEY) {
1267                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1268                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1269                 } else {
1270                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1271                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1272                 }
1273         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1274                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1275                                       struct btrfs_extent_data_ref);
1276                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1277         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1278                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1279                                       struct btrfs_shared_data_ref);
1280                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1281 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1282         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1283                 struct btrfs_extent_ref_v0 *ref0;
1284                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1285                                       struct btrfs_extent_ref_v0);
1286                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1287 #endif
1288         } else {
1289                 WARN_ON(1);
1290         }
1291         return num_refs;
1292 }
1293
1294 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1295                                           struct btrfs_root *root,
1296                                           struct btrfs_path *path,
1297                                           u64 bytenr, u64 parent,
1298                                           u64 root_objectid)
1299 {
1300         struct btrfs_key key;
1301         int ret;
1302
1303         key.objectid = bytenr;
1304         if (parent) {
1305                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1306                 key.offset = parent;
1307         } else {
1308                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1309                 key.offset = root_objectid;
1310         }
1311
1312         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1313         if (ret > 0)
1314                 ret = -ENOENT;
1315 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1316         if (ret == -ENOENT && parent) {
1317                 btrfs_release_path(path);
1318                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1319                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1320                 if (ret > 0)
1321                         ret = -ENOENT;
1322         }
1323 #endif
1324         return ret;
1325 }
1326
1327 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1328                                           struct btrfs_root *root,
1329                                           struct btrfs_path *path,
1330                                           u64 bytenr, u64 parent,
1331                                           u64 root_objectid)
1332 {
1333         struct btrfs_key key;
1334         int ret;
1335
1336         key.objectid = bytenr;
1337         if (parent) {
1338                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1339                 key.offset = parent;
1340         } else {
1341                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1342                 key.offset = root_objectid;
1343         }
1344
1345         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1346         btrfs_release_path(path);
1347         return ret;
1348 }
1349
1350 static inline int extent_ref_type(u64 parent, u64 owner)
1351 {
1352         int type;
1353         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1354                 if (parent > 0)
1355                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1356                 else
1357                         type = BTRFS_TREE_BLOCK_REF_KEY;
1358         } else {
1359                 if (parent > 0)
1360                         type = BTRFS_SHARED_DATA_REF_KEY;
1361                 else
1362                         type = BTRFS_EXTENT_DATA_REF_KEY;
1363         }
1364         return type;
1365 }
1366
1367 static int find_next_key(struct btrfs_path *path, int level,
1368                          struct btrfs_key *key)
1369
1370 {
1371         for (; level < BTRFS_MAX_LEVEL; level++) {
1372                 if (!path->nodes[level])
1373                         break;
1374                 if (path->slots[level] + 1 >=
1375                     btrfs_header_nritems(path->nodes[level]))
1376                         continue;
1377                 if (level == 0)
1378                         btrfs_item_key_to_cpu(path->nodes[level], key,
1379                                               path->slots[level] + 1);
1380                 else
1381                         btrfs_node_key_to_cpu(path->nodes[level], key,
1382                                               path->slots[level] + 1);
1383                 return 0;
1384         }
1385         return 1;
1386 }
1387
1388 /*
1389  * look for inline back ref. if back ref is found, *ref_ret is set
1390  * to the address of inline back ref, and 0 is returned.
1391  *
1392  * if back ref isn't found, *ref_ret is set to the address where it
1393  * should be inserted, and -ENOENT is returned.
1394  *
1395  * if insert is true and there are too many inline back refs, the path
1396  * points to the extent item, and -EAGAIN is returned.
1397  *
1398  * NOTE: inline back refs are ordered in the same way that back ref
1399  *       items in the tree are ordered.
1400  */
1401 static noinline_for_stack
1402 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1403                                  struct btrfs_root *root,
1404                                  struct btrfs_path *path,
1405                                  struct btrfs_extent_inline_ref **ref_ret,
1406                                  u64 bytenr, u64 num_bytes,
1407                                  u64 parent, u64 root_objectid,
1408                                  u64 owner, u64 offset, int insert)
1409 {
1410         struct btrfs_key key;
1411         struct extent_buffer *leaf;
1412         struct btrfs_extent_item *ei;
1413         struct btrfs_extent_inline_ref *iref;
1414         u64 flags;
1415         u64 item_size;
1416         unsigned long ptr;
1417         unsigned long end;
1418         int extra_size;
1419         int type;
1420         int want;
1421         int ret;
1422         int err = 0;
1423
1424         key.objectid = bytenr;
1425         key.type = BTRFS_EXTENT_ITEM_KEY;
1426         key.offset = num_bytes;
1427
1428         want = extent_ref_type(parent, owner);
1429         if (insert) {
1430                 extra_size = btrfs_extent_inline_ref_size(want);
1431                 path->keep_locks = 1;
1432         } else
1433                 extra_size = -1;
1434         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1435         if (ret < 0) {
1436                 err = ret;
1437                 goto out;
1438         }
1439         BUG_ON(ret);
1440
1441         leaf = path->nodes[0];
1442         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1443 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1444         if (item_size < sizeof(*ei)) {
1445                 if (!insert) {
1446                         err = -ENOENT;
1447                         goto out;
1448                 }
1449                 ret = convert_extent_item_v0(trans, root, path, owner,
1450                                              extra_size);
1451                 if (ret < 0) {
1452                         err = ret;
1453                         goto out;
1454                 }
1455                 leaf = path->nodes[0];
1456                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1457         }
1458 #endif
1459         BUG_ON(item_size < sizeof(*ei));
1460
1461         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1462         flags = btrfs_extent_flags(leaf, ei);
1463
1464         ptr = (unsigned long)(ei + 1);
1465         end = (unsigned long)ei + item_size;
1466
1467         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1468                 ptr += sizeof(struct btrfs_tree_block_info);
1469                 BUG_ON(ptr > end);
1470         } else {
1471                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1472         }
1473
1474         err = -ENOENT;
1475         while (1) {
1476                 if (ptr >= end) {
1477                         WARN_ON(ptr > end);
1478                         break;
1479                 }
1480                 iref = (struct btrfs_extent_inline_ref *)ptr;
1481                 type = btrfs_extent_inline_ref_type(leaf, iref);
1482                 if (want < type)
1483                         break;
1484                 if (want > type) {
1485                         ptr += btrfs_extent_inline_ref_size(type);
1486                         continue;
1487                 }
1488
1489                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1490                         struct btrfs_extent_data_ref *dref;
1491                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1492                         if (match_extent_data_ref(leaf, dref, root_objectid,
1493                                                   owner, offset)) {
1494                                 err = 0;
1495                                 break;
1496                         }
1497                         if (hash_extent_data_ref_item(leaf, dref) <
1498                             hash_extent_data_ref(root_objectid, owner, offset))
1499                                 break;
1500                 } else {
1501                         u64 ref_offset;
1502                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1503                         if (parent > 0) {
1504                                 if (parent == ref_offset) {
1505                                         err = 0;
1506                                         break;
1507                                 }
1508                                 if (ref_offset < parent)
1509                                         break;
1510                         } else {
1511                                 if (root_objectid == ref_offset) {
1512                                         err = 0;
1513                                         break;
1514                                 }
1515                                 if (ref_offset < root_objectid)
1516                                         break;
1517                         }
1518                 }
1519                 ptr += btrfs_extent_inline_ref_size(type);
1520         }
1521         if (err == -ENOENT && insert) {
1522                 if (item_size + extra_size >=
1523                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1524                         err = -EAGAIN;
1525                         goto out;
1526                 }
1527                 /*
1528                  * To add new inline back ref, we have to make sure
1529                  * there is no corresponding back ref item.
1530                  * For simplicity, we just do not add new inline back
1531                  * ref if there is any kind of item for this block
1532                  */
1533                 if (find_next_key(path, 0, &key) == 0 &&
1534                     key.objectid == bytenr &&
1535                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1536                         err = -EAGAIN;
1537                         goto out;
1538                 }
1539         }
1540         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1541 out:
1542         if (insert) {
1543                 path->keep_locks = 0;
1544                 btrfs_unlock_up_safe(path, 1);
1545         }
1546         return err;
1547 }
1548
1549 /*
1550  * helper to add new inline back ref
1551  */
1552 static noinline_for_stack
1553 int setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1554                                 struct btrfs_root *root,
1555                                 struct btrfs_path *path,
1556                                 struct btrfs_extent_inline_ref *iref,
1557                                 u64 parent, u64 root_objectid,
1558                                 u64 owner, u64 offset, int refs_to_add,
1559                                 struct btrfs_delayed_extent_op *extent_op)
1560 {
1561         struct extent_buffer *leaf;
1562         struct btrfs_extent_item *ei;
1563         unsigned long ptr;
1564         unsigned long end;
1565         unsigned long item_offset;
1566         u64 refs;
1567         int size;
1568         int type;
1569         int ret;
1570
1571         leaf = path->nodes[0];
1572         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1573         item_offset = (unsigned long)iref - (unsigned long)ei;
1574
1575         type = extent_ref_type(parent, owner);
1576         size = btrfs_extent_inline_ref_size(type);
1577
1578         ret = btrfs_extend_item(trans, root, path, size);
1579
1580         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1581         refs = btrfs_extent_refs(leaf, ei);
1582         refs += refs_to_add;
1583         btrfs_set_extent_refs(leaf, ei, refs);
1584         if (extent_op)
1585                 __run_delayed_extent_op(extent_op, leaf, ei);
1586
1587         ptr = (unsigned long)ei + item_offset;
1588         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1589         if (ptr < end - size)
1590                 memmove_extent_buffer(leaf, ptr + size, ptr,
1591                                       end - size - ptr);
1592
1593         iref = (struct btrfs_extent_inline_ref *)ptr;
1594         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1595         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1596                 struct btrfs_extent_data_ref *dref;
1597                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1598                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1599                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1600                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1601                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1602         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1603                 struct btrfs_shared_data_ref *sref;
1604                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1605                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1606                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1607         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1608                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1609         } else {
1610                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1611         }
1612         btrfs_mark_buffer_dirty(leaf);
1613         return 0;
1614 }
1615
1616 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1617                                  struct btrfs_root *root,
1618                                  struct btrfs_path *path,
1619                                  struct btrfs_extent_inline_ref **ref_ret,
1620                                  u64 bytenr, u64 num_bytes, u64 parent,
1621                                  u64 root_objectid, u64 owner, u64 offset)
1622 {
1623         int ret;
1624
1625         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1626                                            bytenr, num_bytes, parent,
1627                                            root_objectid, owner, offset, 0);
1628         if (ret != -ENOENT)
1629                 return ret;
1630
1631         btrfs_release_path(path);
1632         *ref_ret = NULL;
1633
1634         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1635                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1636                                             root_objectid);
1637         } else {
1638                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1639                                              root_objectid, owner, offset);
1640         }
1641         return ret;
1642 }
1643
1644 /*
1645  * helper to update/remove inline back ref
1646  */
1647 static noinline_for_stack
1648 int update_inline_extent_backref(struct btrfs_trans_handle *trans,
1649                                  struct btrfs_root *root,
1650                                  struct btrfs_path *path,
1651                                  struct btrfs_extent_inline_ref *iref,
1652                                  int refs_to_mod,
1653                                  struct btrfs_delayed_extent_op *extent_op)
1654 {
1655         struct extent_buffer *leaf;
1656         struct btrfs_extent_item *ei;
1657         struct btrfs_extent_data_ref *dref = NULL;
1658         struct btrfs_shared_data_ref *sref = NULL;
1659         unsigned long ptr;
1660         unsigned long end;
1661         u32 item_size;
1662         int size;
1663         int type;
1664         int ret;
1665         u64 refs;
1666
1667         leaf = path->nodes[0];
1668         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1669         refs = btrfs_extent_refs(leaf, ei);
1670         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1671         refs += refs_to_mod;
1672         btrfs_set_extent_refs(leaf, ei, refs);
1673         if (extent_op)
1674                 __run_delayed_extent_op(extent_op, leaf, ei);
1675
1676         type = btrfs_extent_inline_ref_type(leaf, iref);
1677
1678         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1679                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1680                 refs = btrfs_extent_data_ref_count(leaf, dref);
1681         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1682                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1683                 refs = btrfs_shared_data_ref_count(leaf, sref);
1684         } else {
1685                 refs = 1;
1686                 BUG_ON(refs_to_mod != -1);
1687         }
1688
1689         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1690         refs += refs_to_mod;
1691
1692         if (refs > 0) {
1693                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1694                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1695                 else
1696                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1697         } else {
1698                 size =  btrfs_extent_inline_ref_size(type);
1699                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1700                 ptr = (unsigned long)iref;
1701                 end = (unsigned long)ei + item_size;
1702                 if (ptr + size < end)
1703                         memmove_extent_buffer(leaf, ptr, ptr + size,
1704                                               end - ptr - size);
1705                 item_size -= size;
1706                 ret = btrfs_truncate_item(trans, root, path, item_size, 1);
1707         }
1708         btrfs_mark_buffer_dirty(leaf);
1709         return 0;
1710 }
1711
1712 static noinline_for_stack
1713 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1714                                  struct btrfs_root *root,
1715                                  struct btrfs_path *path,
1716                                  u64 bytenr, u64 num_bytes, u64 parent,
1717                                  u64 root_objectid, u64 owner,
1718                                  u64 offset, int refs_to_add,
1719                                  struct btrfs_delayed_extent_op *extent_op)
1720 {
1721         struct btrfs_extent_inline_ref *iref;
1722         int ret;
1723
1724         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1725                                            bytenr, num_bytes, parent,
1726                                            root_objectid, owner, offset, 1);
1727         if (ret == 0) {
1728                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1729                 ret = update_inline_extent_backref(trans, root, path, iref,
1730                                                    refs_to_add, extent_op);
1731         } else if (ret == -ENOENT) {
1732                 ret = setup_inline_extent_backref(trans, root, path, iref,
1733                                                   parent, root_objectid,
1734                                                   owner, offset, refs_to_add,
1735                                                   extent_op);
1736         }
1737         return ret;
1738 }
1739
1740 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1741                                  struct btrfs_root *root,
1742                                  struct btrfs_path *path,
1743                                  u64 bytenr, u64 parent, u64 root_objectid,
1744                                  u64 owner, u64 offset, int refs_to_add)
1745 {
1746         int ret;
1747         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1748                 BUG_ON(refs_to_add != 1);
1749                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1750                                             parent, root_objectid);
1751         } else {
1752                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1753                                              parent, root_objectid,
1754                                              owner, offset, refs_to_add);
1755         }
1756         return ret;
1757 }
1758
1759 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1760                                  struct btrfs_root *root,
1761                                  struct btrfs_path *path,
1762                                  struct btrfs_extent_inline_ref *iref,
1763                                  int refs_to_drop, int is_data)
1764 {
1765         int ret;
1766
1767         BUG_ON(!is_data && refs_to_drop != 1);
1768         if (iref) {
1769                 ret = update_inline_extent_backref(trans, root, path, iref,
1770                                                    -refs_to_drop, NULL);
1771         } else if (is_data) {
1772                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1773         } else {
1774                 ret = btrfs_del_item(trans, root, path);
1775         }
1776         return ret;
1777 }
1778
1779 static int btrfs_issue_discard(struct block_device *bdev,
1780                                 u64 start, u64 len)
1781 {
1782         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1783 }
1784
1785 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1786                                 u64 num_bytes, u64 *actual_bytes)
1787 {
1788         int ret;
1789         u64 discarded_bytes = 0;
1790         struct btrfs_multi_bio *multi = NULL;
1791
1792
1793         /* Tell the block device(s) that the sectors can be discarded */
1794         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1795                               bytenr, &num_bytes, &multi, 0);
1796         if (!ret) {
1797                 struct btrfs_bio_stripe *stripe = multi->stripes;
1798                 int i;
1799
1800
1801                 for (i = 0; i < multi->num_stripes; i++, stripe++) {
1802                         if (!stripe->dev->can_discard)
1803                                 continue;
1804
1805                         ret = btrfs_issue_discard(stripe->dev->bdev,
1806                                                   stripe->physical,
1807                                                   stripe->length);
1808                         if (!ret)
1809                                 discarded_bytes += stripe->length;
1810                         else if (ret != -EOPNOTSUPP)
1811                                 break;
1812
1813                         /*
1814                          * Just in case we get back EOPNOTSUPP for some reason,
1815                          * just ignore the return value so we don't screw up
1816                          * people calling discard_extent.
1817                          */
1818                         ret = 0;
1819                 }
1820                 kfree(multi);
1821         }
1822
1823         if (actual_bytes)
1824                 *actual_bytes = discarded_bytes;
1825
1826
1827         return ret;
1828 }
1829
1830 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1831                          struct btrfs_root *root,
1832                          u64 bytenr, u64 num_bytes, u64 parent,
1833                          u64 root_objectid, u64 owner, u64 offset)
1834 {
1835         int ret;
1836         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1837                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1838
1839         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1840                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
1841                                         parent, root_objectid, (int)owner,
1842                                         BTRFS_ADD_DELAYED_REF, NULL);
1843         } else {
1844                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
1845                                         parent, root_objectid, owner, offset,
1846                                         BTRFS_ADD_DELAYED_REF, NULL);
1847         }
1848         return ret;
1849 }
1850
1851 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1852                                   struct btrfs_root *root,
1853                                   u64 bytenr, u64 num_bytes,
1854                                   u64 parent, u64 root_objectid,
1855                                   u64 owner, u64 offset, int refs_to_add,
1856                                   struct btrfs_delayed_extent_op *extent_op)
1857 {
1858         struct btrfs_path *path;
1859         struct extent_buffer *leaf;
1860         struct btrfs_extent_item *item;
1861         u64 refs;
1862         int ret;
1863         int err = 0;
1864
1865         path = btrfs_alloc_path();
1866         if (!path)
1867                 return -ENOMEM;
1868
1869         path->reada = 1;
1870         path->leave_spinning = 1;
1871         /* this will setup the path even if it fails to insert the back ref */
1872         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1873                                            path, bytenr, num_bytes, parent,
1874                                            root_objectid, owner, offset,
1875                                            refs_to_add, extent_op);
1876         if (ret == 0)
1877                 goto out;
1878
1879         if (ret != -EAGAIN) {
1880                 err = ret;
1881                 goto out;
1882         }
1883
1884         leaf = path->nodes[0];
1885         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1886         refs = btrfs_extent_refs(leaf, item);
1887         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1888         if (extent_op)
1889                 __run_delayed_extent_op(extent_op, leaf, item);
1890
1891         btrfs_mark_buffer_dirty(leaf);
1892         btrfs_release_path(path);
1893
1894         path->reada = 1;
1895         path->leave_spinning = 1;
1896
1897         /* now insert the actual backref */
1898         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1899                                     path, bytenr, parent, root_objectid,
1900                                     owner, offset, refs_to_add);
1901         BUG_ON(ret);
1902 out:
1903         btrfs_free_path(path);
1904         return err;
1905 }
1906
1907 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1908                                 struct btrfs_root *root,
1909                                 struct btrfs_delayed_ref_node *node,
1910                                 struct btrfs_delayed_extent_op *extent_op,
1911                                 int insert_reserved)
1912 {
1913         int ret = 0;
1914         struct btrfs_delayed_data_ref *ref;
1915         struct btrfs_key ins;
1916         u64 parent = 0;
1917         u64 ref_root = 0;
1918         u64 flags = 0;
1919
1920         ins.objectid = node->bytenr;
1921         ins.offset = node->num_bytes;
1922         ins.type = BTRFS_EXTENT_ITEM_KEY;
1923
1924         ref = btrfs_delayed_node_to_data_ref(node);
1925         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1926                 parent = ref->parent;
1927         else
1928                 ref_root = ref->root;
1929
1930         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1931                 if (extent_op) {
1932                         BUG_ON(extent_op->update_key);
1933                         flags |= extent_op->flags_to_set;
1934                 }
1935                 ret = alloc_reserved_file_extent(trans, root,
1936                                                  parent, ref_root, flags,
1937                                                  ref->objectid, ref->offset,
1938                                                  &ins, node->ref_mod);
1939         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1940                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1941                                              node->num_bytes, parent,
1942                                              ref_root, ref->objectid,
1943                                              ref->offset, node->ref_mod,
1944                                              extent_op);
1945         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1946                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1947                                           node->num_bytes, parent,
1948                                           ref_root, ref->objectid,
1949                                           ref->offset, node->ref_mod,
1950                                           extent_op);
1951         } else {
1952                 BUG();
1953         }
1954         return ret;
1955 }
1956
1957 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1958                                     struct extent_buffer *leaf,
1959                                     struct btrfs_extent_item *ei)
1960 {
1961         u64 flags = btrfs_extent_flags(leaf, ei);
1962         if (extent_op->update_flags) {
1963                 flags |= extent_op->flags_to_set;
1964                 btrfs_set_extent_flags(leaf, ei, flags);
1965         }
1966
1967         if (extent_op->update_key) {
1968                 struct btrfs_tree_block_info *bi;
1969                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1970                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1971                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1972         }
1973 }
1974
1975 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1976                                  struct btrfs_root *root,
1977                                  struct btrfs_delayed_ref_node *node,
1978                                  struct btrfs_delayed_extent_op *extent_op)
1979 {
1980         struct btrfs_key key;
1981         struct btrfs_path *path;
1982         struct btrfs_extent_item *ei;
1983         struct extent_buffer *leaf;
1984         u32 item_size;
1985         int ret;
1986         int err = 0;
1987
1988         path = btrfs_alloc_path();
1989         if (!path)
1990                 return -ENOMEM;
1991
1992         key.objectid = node->bytenr;
1993         key.type = BTRFS_EXTENT_ITEM_KEY;
1994         key.offset = node->num_bytes;
1995
1996         path->reada = 1;
1997         path->leave_spinning = 1;
1998         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
1999                                 path, 0, 1);
2000         if (ret < 0) {
2001                 err = ret;
2002                 goto out;
2003         }
2004         if (ret > 0) {
2005                 err = -EIO;
2006                 goto out;
2007         }
2008
2009         leaf = path->nodes[0];
2010         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2011 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2012         if (item_size < sizeof(*ei)) {
2013                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2014                                              path, (u64)-1, 0);
2015                 if (ret < 0) {
2016                         err = ret;
2017                         goto out;
2018                 }
2019                 leaf = path->nodes[0];
2020                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2021         }
2022 #endif
2023         BUG_ON(item_size < sizeof(*ei));
2024         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2025         __run_delayed_extent_op(extent_op, leaf, ei);
2026
2027         btrfs_mark_buffer_dirty(leaf);
2028 out:
2029         btrfs_free_path(path);
2030         return err;
2031 }
2032
2033 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2034                                 struct btrfs_root *root,
2035                                 struct btrfs_delayed_ref_node *node,
2036                                 struct btrfs_delayed_extent_op *extent_op,
2037                                 int insert_reserved)
2038 {
2039         int ret = 0;
2040         struct btrfs_delayed_tree_ref *ref;
2041         struct btrfs_key ins;
2042         u64 parent = 0;
2043         u64 ref_root = 0;
2044
2045         ins.objectid = node->bytenr;
2046         ins.offset = node->num_bytes;
2047         ins.type = BTRFS_EXTENT_ITEM_KEY;
2048
2049         ref = btrfs_delayed_node_to_tree_ref(node);
2050         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2051                 parent = ref->parent;
2052         else
2053                 ref_root = ref->root;
2054
2055         BUG_ON(node->ref_mod != 1);
2056         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2057                 BUG_ON(!extent_op || !extent_op->update_flags ||
2058                        !extent_op->update_key);
2059                 ret = alloc_reserved_tree_block(trans, root,
2060                                                 parent, ref_root,
2061                                                 extent_op->flags_to_set,
2062                                                 &extent_op->key,
2063                                                 ref->level, &ins);
2064         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2065                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2066                                              node->num_bytes, parent, ref_root,
2067                                              ref->level, 0, 1, extent_op);
2068         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2069                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2070                                           node->num_bytes, parent, ref_root,
2071                                           ref->level, 0, 1, extent_op);
2072         } else {
2073                 BUG();
2074         }
2075         return ret;
2076 }
2077
2078 /* helper function to actually process a single delayed ref entry */
2079 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2080                                struct btrfs_root *root,
2081                                struct btrfs_delayed_ref_node *node,
2082                                struct btrfs_delayed_extent_op *extent_op,
2083                                int insert_reserved)
2084 {
2085         int ret;
2086         if (btrfs_delayed_ref_is_head(node)) {
2087                 struct btrfs_delayed_ref_head *head;
2088                 /*
2089                  * we've hit the end of the chain and we were supposed
2090                  * to insert this extent into the tree.  But, it got
2091                  * deleted before we ever needed to insert it, so all
2092                  * we have to do is clean up the accounting
2093                  */
2094                 BUG_ON(extent_op);
2095                 head = btrfs_delayed_node_to_head(node);
2096                 if (insert_reserved) {
2097                         btrfs_pin_extent(root, node->bytenr,
2098                                          node->num_bytes, 1);
2099                         if (head->is_data) {
2100                                 ret = btrfs_del_csums(trans, root,
2101                                                       node->bytenr,
2102                                                       node->num_bytes);
2103                                 BUG_ON(ret);
2104                         }
2105                 }
2106                 mutex_unlock(&head->mutex);
2107                 return 0;
2108         }
2109
2110         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2111             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2112                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2113                                            insert_reserved);
2114         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2115                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2116                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2117                                            insert_reserved);
2118         else
2119                 BUG();
2120         return ret;
2121 }
2122
2123 static noinline struct btrfs_delayed_ref_node *
2124 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2125 {
2126         struct rb_node *node;
2127         struct btrfs_delayed_ref_node *ref;
2128         int action = BTRFS_ADD_DELAYED_REF;
2129 again:
2130         /*
2131          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2132          * this prevents ref count from going down to zero when
2133          * there still are pending delayed ref.
2134          */
2135         node = rb_prev(&head->node.rb_node);
2136         while (1) {
2137                 if (!node)
2138                         break;
2139                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2140                                 rb_node);
2141                 if (ref->bytenr != head->node.bytenr)
2142                         break;
2143                 if (ref->action == action)
2144                         return ref;
2145                 node = rb_prev(node);
2146         }
2147         if (action == BTRFS_ADD_DELAYED_REF) {
2148                 action = BTRFS_DROP_DELAYED_REF;
2149                 goto again;
2150         }
2151         return NULL;
2152 }
2153
2154 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2155                                        struct btrfs_root *root,
2156                                        struct list_head *cluster)
2157 {
2158         struct btrfs_delayed_ref_root *delayed_refs;
2159         struct btrfs_delayed_ref_node *ref;
2160         struct btrfs_delayed_ref_head *locked_ref = NULL;
2161         struct btrfs_delayed_extent_op *extent_op;
2162         int ret;
2163         int count = 0;
2164         int must_insert_reserved = 0;
2165
2166         delayed_refs = &trans->transaction->delayed_refs;
2167         while (1) {
2168                 if (!locked_ref) {
2169                         /* pick a new head ref from the cluster list */
2170                         if (list_empty(cluster))
2171                                 break;
2172
2173                         locked_ref = list_entry(cluster->next,
2174                                      struct btrfs_delayed_ref_head, cluster);
2175
2176                         /* grab the lock that says we are going to process
2177                          * all the refs for this head */
2178                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2179
2180                         /*
2181                          * we may have dropped the spin lock to get the head
2182                          * mutex lock, and that might have given someone else
2183                          * time to free the head.  If that's true, it has been
2184                          * removed from our list and we can move on.
2185                          */
2186                         if (ret == -EAGAIN) {
2187                                 locked_ref = NULL;
2188                                 count++;
2189                                 continue;
2190                         }
2191                 }
2192
2193                 /*
2194                  * record the must insert reserved flag before we
2195                  * drop the spin lock.
2196                  */
2197                 must_insert_reserved = locked_ref->must_insert_reserved;
2198                 locked_ref->must_insert_reserved = 0;
2199
2200                 extent_op = locked_ref->extent_op;
2201                 locked_ref->extent_op = NULL;
2202
2203                 /*
2204                  * locked_ref is the head node, so we have to go one
2205                  * node back for any delayed ref updates
2206                  */
2207                 ref = select_delayed_ref(locked_ref);
2208                 if (!ref) {
2209                         /* All delayed refs have been processed, Go ahead
2210                          * and send the head node to run_one_delayed_ref,
2211                          * so that any accounting fixes can happen
2212                          */
2213                         ref = &locked_ref->node;
2214
2215                         if (extent_op && must_insert_reserved) {
2216                                 kfree(extent_op);
2217                                 extent_op = NULL;
2218                         }
2219
2220                         if (extent_op) {
2221                                 spin_unlock(&delayed_refs->lock);
2222
2223                                 ret = run_delayed_extent_op(trans, root,
2224                                                             ref, extent_op);
2225                                 BUG_ON(ret);
2226                                 kfree(extent_op);
2227
2228                                 cond_resched();
2229                                 spin_lock(&delayed_refs->lock);
2230                                 continue;
2231                         }
2232
2233                         list_del_init(&locked_ref->cluster);
2234                         locked_ref = NULL;
2235                 }
2236
2237                 ref->in_tree = 0;
2238                 rb_erase(&ref->rb_node, &delayed_refs->root);
2239                 delayed_refs->num_entries--;
2240
2241                 spin_unlock(&delayed_refs->lock);
2242
2243                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2244                                           must_insert_reserved);
2245                 BUG_ON(ret);
2246
2247                 btrfs_put_delayed_ref(ref);
2248                 kfree(extent_op);
2249                 count++;
2250
2251                 cond_resched();
2252                 spin_lock(&delayed_refs->lock);
2253         }
2254         return count;
2255 }
2256
2257 /*
2258  * this starts processing the delayed reference count updates and
2259  * extent insertions we have queued up so far.  count can be
2260  * 0, which means to process everything in the tree at the start
2261  * of the run (but not newly added entries), or it can be some target
2262  * number you'd like to process.
2263  */
2264 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2265                            struct btrfs_root *root, unsigned long count)
2266 {
2267         struct rb_node *node;
2268         struct btrfs_delayed_ref_root *delayed_refs;
2269         struct btrfs_delayed_ref_node *ref;
2270         struct list_head cluster;
2271         int ret;
2272         int run_all = count == (unsigned long)-1;
2273         int run_most = 0;
2274
2275         if (root == root->fs_info->extent_root)
2276                 root = root->fs_info->tree_root;
2277
2278         delayed_refs = &trans->transaction->delayed_refs;
2279         INIT_LIST_HEAD(&cluster);
2280 again:
2281         spin_lock(&delayed_refs->lock);
2282         if (count == 0) {
2283                 count = delayed_refs->num_entries * 2;
2284                 run_most = 1;
2285         }
2286         while (1) {
2287                 if (!(run_all || run_most) &&
2288                     delayed_refs->num_heads_ready < 64)
2289                         break;
2290
2291                 /*
2292                  * go find something we can process in the rbtree.  We start at
2293                  * the beginning of the tree, and then build a cluster
2294                  * of refs to process starting at the first one we are able to
2295                  * lock
2296                  */
2297                 ret = btrfs_find_ref_cluster(trans, &cluster,
2298                                              delayed_refs->run_delayed_start);
2299                 if (ret)
2300                         break;
2301
2302                 ret = run_clustered_refs(trans, root, &cluster);
2303                 BUG_ON(ret < 0);
2304
2305                 count -= min_t(unsigned long, ret, count);
2306
2307                 if (count == 0)
2308                         break;
2309         }
2310
2311         if (run_all) {
2312                 node = rb_first(&delayed_refs->root);
2313                 if (!node)
2314                         goto out;
2315                 count = (unsigned long)-1;
2316
2317                 while (node) {
2318                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2319                                        rb_node);
2320                         if (btrfs_delayed_ref_is_head(ref)) {
2321                                 struct btrfs_delayed_ref_head *head;
2322
2323                                 head = btrfs_delayed_node_to_head(ref);
2324                                 atomic_inc(&ref->refs);
2325
2326                                 spin_unlock(&delayed_refs->lock);
2327                                 /*
2328                                  * Mutex was contended, block until it's
2329                                  * released and try again
2330                                  */
2331                                 mutex_lock(&head->mutex);
2332                                 mutex_unlock(&head->mutex);
2333
2334                                 btrfs_put_delayed_ref(ref);
2335                                 cond_resched();
2336                                 goto again;
2337                         }
2338                         node = rb_next(node);
2339                 }
2340                 spin_unlock(&delayed_refs->lock);
2341                 schedule_timeout(1);
2342                 goto again;
2343         }
2344 out:
2345         spin_unlock(&delayed_refs->lock);
2346         return 0;
2347 }
2348
2349 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2350                                 struct btrfs_root *root,
2351                                 u64 bytenr, u64 num_bytes, u64 flags,
2352                                 int is_data)
2353 {
2354         struct btrfs_delayed_extent_op *extent_op;
2355         int ret;
2356
2357         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2358         if (!extent_op)
2359                 return -ENOMEM;
2360
2361         extent_op->flags_to_set = flags;
2362         extent_op->update_flags = 1;
2363         extent_op->update_key = 0;
2364         extent_op->is_data = is_data ? 1 : 0;
2365
2366         ret = btrfs_add_delayed_extent_op(trans, bytenr, num_bytes, extent_op);
2367         if (ret)
2368                 kfree(extent_op);
2369         return ret;
2370 }
2371
2372 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2373                                       struct btrfs_root *root,
2374                                       struct btrfs_path *path,
2375                                       u64 objectid, u64 offset, u64 bytenr)
2376 {
2377         struct btrfs_delayed_ref_head *head;
2378         struct btrfs_delayed_ref_node *ref;
2379         struct btrfs_delayed_data_ref *data_ref;
2380         struct btrfs_delayed_ref_root *delayed_refs;
2381         struct rb_node *node;
2382         int ret = 0;
2383
2384         ret = -ENOENT;
2385         delayed_refs = &trans->transaction->delayed_refs;
2386         spin_lock(&delayed_refs->lock);
2387         head = btrfs_find_delayed_ref_head(trans, bytenr);
2388         if (!head)
2389                 goto out;
2390
2391         if (!mutex_trylock(&head->mutex)) {
2392                 atomic_inc(&head->node.refs);
2393                 spin_unlock(&delayed_refs->lock);
2394
2395                 btrfs_release_path(path);
2396
2397                 /*
2398                  * Mutex was contended, block until it's released and let
2399                  * caller try again
2400                  */
2401                 mutex_lock(&head->mutex);
2402                 mutex_unlock(&head->mutex);
2403                 btrfs_put_delayed_ref(&head->node);
2404                 return -EAGAIN;
2405         }
2406
2407         node = rb_prev(&head->node.rb_node);
2408         if (!node)
2409                 goto out_unlock;
2410
2411         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2412
2413         if (ref->bytenr != bytenr)
2414                 goto out_unlock;
2415
2416         ret = 1;
2417         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2418                 goto out_unlock;
2419
2420         data_ref = btrfs_delayed_node_to_data_ref(ref);
2421
2422         node = rb_prev(node);
2423         if (node) {
2424                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2425                 if (ref->bytenr == bytenr)
2426                         goto out_unlock;
2427         }
2428
2429         if (data_ref->root != root->root_key.objectid ||
2430             data_ref->objectid != objectid || data_ref->offset != offset)
2431                 goto out_unlock;
2432
2433         ret = 0;
2434 out_unlock:
2435         mutex_unlock(&head->mutex);
2436 out:
2437         spin_unlock(&delayed_refs->lock);
2438         return ret;
2439 }
2440
2441 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2442                                         struct btrfs_root *root,
2443                                         struct btrfs_path *path,
2444                                         u64 objectid, u64 offset, u64 bytenr)
2445 {
2446         struct btrfs_root *extent_root = root->fs_info->extent_root;
2447         struct extent_buffer *leaf;
2448         struct btrfs_extent_data_ref *ref;
2449         struct btrfs_extent_inline_ref *iref;
2450         struct btrfs_extent_item *ei;
2451         struct btrfs_key key;
2452         u32 item_size;
2453         int ret;
2454
2455         key.objectid = bytenr;
2456         key.offset = (u64)-1;
2457         key.type = BTRFS_EXTENT_ITEM_KEY;
2458
2459         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2460         if (ret < 0)
2461                 goto out;
2462         BUG_ON(ret == 0);
2463
2464         ret = -ENOENT;
2465         if (path->slots[0] == 0)
2466                 goto out;
2467
2468         path->slots[0]--;
2469         leaf = path->nodes[0];
2470         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2471
2472         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2473                 goto out;
2474
2475         ret = 1;
2476         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2477 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2478         if (item_size < sizeof(*ei)) {
2479                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2480                 goto out;
2481         }
2482 #endif
2483         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2484
2485         if (item_size != sizeof(*ei) +
2486             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2487                 goto out;
2488
2489         if (btrfs_extent_generation(leaf, ei) <=
2490             btrfs_root_last_snapshot(&root->root_item))
2491                 goto out;
2492
2493         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2494         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2495             BTRFS_EXTENT_DATA_REF_KEY)
2496                 goto out;
2497
2498         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2499         if (btrfs_extent_refs(leaf, ei) !=
2500             btrfs_extent_data_ref_count(leaf, ref) ||
2501             btrfs_extent_data_ref_root(leaf, ref) !=
2502             root->root_key.objectid ||
2503             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2504             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2505                 goto out;
2506
2507         ret = 0;
2508 out:
2509         return ret;
2510 }
2511
2512 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2513                           struct btrfs_root *root,
2514                           u64 objectid, u64 offset, u64 bytenr)
2515 {
2516         struct btrfs_path *path;
2517         int ret;
2518         int ret2;
2519
2520         path = btrfs_alloc_path();
2521         if (!path)
2522                 return -ENOENT;
2523
2524         do {
2525                 ret = check_committed_ref(trans, root, path, objectid,
2526                                           offset, bytenr);
2527                 if (ret && ret != -ENOENT)
2528                         goto out;
2529
2530                 ret2 = check_delayed_ref(trans, root, path, objectid,
2531                                          offset, bytenr);
2532         } while (ret2 == -EAGAIN);
2533
2534         if (ret2 && ret2 != -ENOENT) {
2535                 ret = ret2;
2536                 goto out;
2537         }
2538
2539         if (ret != -ENOENT || ret2 != -ENOENT)
2540                 ret = 0;
2541 out:
2542         btrfs_free_path(path);
2543         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2544                 WARN_ON(ret > 0);
2545         return ret;
2546 }
2547
2548 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2549                            struct btrfs_root *root,
2550                            struct extent_buffer *buf,
2551                            int full_backref, int inc)
2552 {
2553         u64 bytenr;
2554         u64 num_bytes;
2555         u64 parent;
2556         u64 ref_root;
2557         u32 nritems;
2558         struct btrfs_key key;
2559         struct btrfs_file_extent_item *fi;
2560         int i;
2561         int level;
2562         int ret = 0;
2563         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2564                             u64, u64, u64, u64, u64, u64);
2565
2566         ref_root = btrfs_header_owner(buf);
2567         nritems = btrfs_header_nritems(buf);
2568         level = btrfs_header_level(buf);
2569
2570         if (!root->ref_cows && level == 0)
2571                 return 0;
2572
2573         if (inc)
2574                 process_func = btrfs_inc_extent_ref;
2575         else
2576                 process_func = btrfs_free_extent;
2577
2578         if (full_backref)
2579                 parent = buf->start;
2580         else
2581                 parent = 0;
2582
2583         for (i = 0; i < nritems; i++) {
2584                 if (level == 0) {
2585                         btrfs_item_key_to_cpu(buf, &key, i);
2586                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2587                                 continue;
2588                         fi = btrfs_item_ptr(buf, i,
2589                                             struct btrfs_file_extent_item);
2590                         if (btrfs_file_extent_type(buf, fi) ==
2591                             BTRFS_FILE_EXTENT_INLINE)
2592                                 continue;
2593                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2594                         if (bytenr == 0)
2595                                 continue;
2596
2597                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2598                         key.offset -= btrfs_file_extent_offset(buf, fi);
2599                         ret = process_func(trans, root, bytenr, num_bytes,
2600                                            parent, ref_root, key.objectid,
2601                                            key.offset);
2602                         if (ret)
2603                                 goto fail;
2604                 } else {
2605                         bytenr = btrfs_node_blockptr(buf, i);
2606                         num_bytes = btrfs_level_size(root, level - 1);
2607                         ret = process_func(trans, root, bytenr, num_bytes,
2608                                            parent, ref_root, level - 1, 0);
2609                         if (ret)
2610                                 goto fail;
2611                 }
2612         }
2613         return 0;
2614 fail:
2615         BUG();
2616         return ret;
2617 }
2618
2619 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2620                   struct extent_buffer *buf, int full_backref)
2621 {
2622         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2623 }
2624
2625 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2626                   struct extent_buffer *buf, int full_backref)
2627 {
2628         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2629 }
2630
2631 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2632                                  struct btrfs_root *root,
2633                                  struct btrfs_path *path,
2634                                  struct btrfs_block_group_cache *cache)
2635 {
2636         int ret;
2637         struct btrfs_root *extent_root = root->fs_info->extent_root;
2638         unsigned long bi;
2639         struct extent_buffer *leaf;
2640
2641         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2642         if (ret < 0)
2643                 goto fail;
2644         BUG_ON(ret);
2645
2646         leaf = path->nodes[0];
2647         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2648         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2649         btrfs_mark_buffer_dirty(leaf);
2650         btrfs_release_path(path);
2651 fail:
2652         if (ret)
2653                 return ret;
2654         return 0;
2655
2656 }
2657
2658 static struct btrfs_block_group_cache *
2659 next_block_group(struct btrfs_root *root,
2660                  struct btrfs_block_group_cache *cache)
2661 {
2662         struct rb_node *node;
2663         spin_lock(&root->fs_info->block_group_cache_lock);
2664         node = rb_next(&cache->cache_node);
2665         btrfs_put_block_group(cache);
2666         if (node) {
2667                 cache = rb_entry(node, struct btrfs_block_group_cache,
2668                                  cache_node);
2669                 btrfs_get_block_group(cache);
2670         } else
2671                 cache = NULL;
2672         spin_unlock(&root->fs_info->block_group_cache_lock);
2673         return cache;
2674 }
2675
2676 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2677                             struct btrfs_trans_handle *trans,
2678                             struct btrfs_path *path)
2679 {
2680         struct btrfs_root *root = block_group->fs_info->tree_root;
2681         struct inode *inode = NULL;
2682         u64 alloc_hint = 0;
2683         int dcs = BTRFS_DC_ERROR;
2684         int num_pages = 0;
2685         int retries = 0;
2686         int ret = 0;
2687
2688         /*
2689          * If this block group is smaller than 100 megs don't bother caching the
2690          * block group.
2691          */
2692         if (block_group->key.offset < (100 * 1024 * 1024)) {
2693                 spin_lock(&block_group->lock);
2694                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2695                 spin_unlock(&block_group->lock);
2696                 return 0;
2697         }
2698
2699 again:
2700         inode = lookup_free_space_inode(root, block_group, path);
2701         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2702                 ret = PTR_ERR(inode);
2703                 btrfs_release_path(path);
2704                 goto out;
2705         }
2706
2707         if (IS_ERR(inode)) {
2708                 BUG_ON(retries);
2709                 retries++;
2710
2711                 if (block_group->ro)
2712                         goto out_free;
2713
2714                 ret = create_free_space_inode(root, trans, block_group, path);
2715                 if (ret)
2716                         goto out_free;
2717                 goto again;
2718         }
2719
2720         /* We've already setup this transaction, go ahead and exit */
2721         if (block_group->cache_generation == trans->transid &&
2722             i_size_read(inode)) {
2723                 dcs = BTRFS_DC_SETUP;
2724                 goto out_put;
2725         }
2726
2727         /*
2728          * We want to set the generation to 0, that way if anything goes wrong
2729          * from here on out we know not to trust this cache when we load up next
2730          * time.
2731          */
2732         BTRFS_I(inode)->generation = 0;
2733         ret = btrfs_update_inode(trans, root, inode);
2734         WARN_ON(ret);
2735
2736         if (i_size_read(inode) > 0) {
2737                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2738                                                       inode);
2739                 if (ret)
2740                         goto out_put;
2741         }
2742
2743         spin_lock(&block_group->lock);
2744         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2745                 /* We're not cached, don't bother trying to write stuff out */
2746                 dcs = BTRFS_DC_WRITTEN;
2747                 spin_unlock(&block_group->lock);
2748                 goto out_put;
2749         }
2750         spin_unlock(&block_group->lock);
2751
2752         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2753         if (!num_pages)
2754                 num_pages = 1;
2755
2756         /*
2757          * Just to make absolutely sure we have enough space, we're going to
2758          * preallocate 12 pages worth of space for each block group.  In
2759          * practice we ought to use at most 8, but we need extra space so we can
2760          * add our header and have a terminator between the extents and the
2761          * bitmaps.
2762          */
2763         num_pages *= 16;
2764         num_pages *= PAGE_CACHE_SIZE;
2765
2766         ret = btrfs_check_data_free_space(inode, num_pages);
2767         if (ret)
2768                 goto out_put;
2769
2770         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2771                                               num_pages, num_pages,
2772                                               &alloc_hint);
2773         if (!ret)
2774                 dcs = BTRFS_DC_SETUP;
2775         btrfs_free_reserved_data_space(inode, num_pages);
2776
2777 out_put:
2778         iput(inode);
2779 out_free:
2780         btrfs_release_path(path);
2781 out:
2782         spin_lock(&block_group->lock);
2783         if (!ret)
2784                 block_group->cache_generation = trans->transid;
2785         block_group->disk_cache_state = dcs;
2786         spin_unlock(&block_group->lock);
2787
2788         return ret;
2789 }
2790
2791 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2792                                    struct btrfs_root *root)
2793 {
2794         struct btrfs_block_group_cache *cache;
2795         int err = 0;
2796         struct btrfs_path *path;
2797         u64 last = 0;
2798
2799         path = btrfs_alloc_path();
2800         if (!path)
2801                 return -ENOMEM;
2802
2803 again:
2804         while (1) {
2805                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2806                 while (cache) {
2807                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2808                                 break;
2809                         cache = next_block_group(root, cache);
2810                 }
2811                 if (!cache) {
2812                         if (last == 0)
2813                                 break;
2814                         last = 0;
2815                         continue;
2816                 }
2817                 err = cache_save_setup(cache, trans, path);
2818                 last = cache->key.objectid + cache->key.offset;
2819                 btrfs_put_block_group(cache);
2820         }
2821
2822         while (1) {
2823                 if (last == 0) {
2824                         err = btrfs_run_delayed_refs(trans, root,
2825                                                      (unsigned long)-1);
2826                         BUG_ON(err);
2827                 }
2828
2829                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2830                 while (cache) {
2831                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2832                                 btrfs_put_block_group(cache);
2833                                 goto again;
2834                         }
2835
2836                         if (cache->dirty)
2837                                 break;
2838                         cache = next_block_group(root, cache);
2839                 }
2840                 if (!cache) {
2841                         if (last == 0)
2842                                 break;
2843                         last = 0;
2844                         continue;
2845                 }
2846
2847                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
2848                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
2849                 cache->dirty = 0;
2850                 last = cache->key.objectid + cache->key.offset;
2851
2852                 err = write_one_cache_group(trans, root, path, cache);
2853                 BUG_ON(err);
2854                 btrfs_put_block_group(cache);
2855         }
2856
2857         while (1) {
2858                 /*
2859                  * I don't think this is needed since we're just marking our
2860                  * preallocated extent as written, but just in case it can't
2861                  * hurt.
2862                  */
2863                 if (last == 0) {
2864                         err = btrfs_run_delayed_refs(trans, root,
2865                                                      (unsigned long)-1);
2866                         BUG_ON(err);
2867                 }
2868
2869                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2870                 while (cache) {
2871                         /*
2872                          * Really this shouldn't happen, but it could if we
2873                          * couldn't write the entire preallocated extent and
2874                          * splitting the extent resulted in a new block.
2875                          */
2876                         if (cache->dirty) {
2877                                 btrfs_put_block_group(cache);
2878                                 goto again;
2879                         }
2880                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2881                                 break;
2882                         cache = next_block_group(root, cache);
2883                 }
2884                 if (!cache) {
2885                         if (last == 0)
2886                                 break;
2887                         last = 0;
2888                         continue;
2889                 }
2890
2891                 btrfs_write_out_cache(root, trans, cache, path);
2892
2893                 /*
2894                  * If we didn't have an error then the cache state is still
2895                  * NEED_WRITE, so we can set it to WRITTEN.
2896                  */
2897                 if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2898                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
2899                 last = cache->key.objectid + cache->key.offset;
2900                 btrfs_put_block_group(cache);
2901         }
2902
2903         btrfs_free_path(path);
2904         return 0;
2905 }
2906
2907 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
2908 {
2909         struct btrfs_block_group_cache *block_group;
2910         int readonly = 0;
2911
2912         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
2913         if (!block_group || block_group->ro)
2914                 readonly = 1;
2915         if (block_group)
2916                 btrfs_put_block_group(block_group);
2917         return readonly;
2918 }
2919
2920 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
2921                              u64 total_bytes, u64 bytes_used,
2922                              struct btrfs_space_info **space_info)
2923 {
2924         struct btrfs_space_info *found;
2925         int i;
2926         int factor;
2927
2928         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
2929                      BTRFS_BLOCK_GROUP_RAID10))
2930                 factor = 2;
2931         else
2932                 factor = 1;
2933
2934         found = __find_space_info(info, flags);
2935         if (found) {
2936                 spin_lock(&found->lock);
2937                 found->total_bytes += total_bytes;
2938                 found->disk_total += total_bytes * factor;
2939                 found->bytes_used += bytes_used;
2940                 found->disk_used += bytes_used * factor;
2941                 found->full = 0;
2942                 spin_unlock(&found->lock);
2943                 *space_info = found;
2944                 return 0;
2945         }
2946         found = kzalloc(sizeof(*found), GFP_NOFS);
2947         if (!found)
2948                 return -ENOMEM;
2949
2950         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
2951                 INIT_LIST_HEAD(&found->block_groups[i]);
2952         init_rwsem(&found->groups_sem);
2953         spin_lock_init(&found->lock);
2954         found->flags = flags & (BTRFS_BLOCK_GROUP_DATA |
2955                                 BTRFS_BLOCK_GROUP_SYSTEM |
2956                                 BTRFS_BLOCK_GROUP_METADATA);
2957         found->total_bytes = total_bytes;
2958         found->disk_total = total_bytes * factor;
2959         found->bytes_used = bytes_used;
2960         found->disk_used = bytes_used * factor;
2961         found->bytes_pinned = 0;
2962         found->bytes_reserved = 0;
2963         found->bytes_readonly = 0;
2964         found->bytes_may_use = 0;
2965         found->full = 0;
2966         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
2967         found->chunk_alloc = 0;
2968         found->flush = 0;
2969         init_waitqueue_head(&found->wait);
2970         *space_info = found;
2971         list_add_rcu(&found->list, &info->space_info);
2972         return 0;
2973 }
2974
2975 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
2976 {
2977         u64 extra_flags = flags & (BTRFS_BLOCK_GROUP_RAID0 |
2978                                    BTRFS_BLOCK_GROUP_RAID1 |
2979                                    BTRFS_BLOCK_GROUP_RAID10 |
2980                                    BTRFS_BLOCK_GROUP_DUP);
2981         if (extra_flags) {
2982                 if (flags & BTRFS_BLOCK_GROUP_DATA)
2983                         fs_info->avail_data_alloc_bits |= extra_flags;
2984                 if (flags & BTRFS_BLOCK_GROUP_METADATA)
2985                         fs_info->avail_metadata_alloc_bits |= extra_flags;
2986                 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
2987                         fs_info->avail_system_alloc_bits |= extra_flags;
2988         }
2989 }
2990
2991 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
2992 {
2993         /*
2994          * we add in the count of missing devices because we want
2995          * to make sure that any RAID levels on a degraded FS
2996          * continue to be honored.
2997          */
2998         u64 num_devices = root->fs_info->fs_devices->rw_devices +
2999                 root->fs_info->fs_devices->missing_devices;
3000
3001         if (num_devices == 1)
3002                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3003         if (num_devices < 4)
3004                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3005
3006         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3007             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3008                       BTRFS_BLOCK_GROUP_RAID10))) {
3009                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3010         }
3011
3012         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3013             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3014                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3015         }
3016
3017         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3018             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3019              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3020              (flags & BTRFS_BLOCK_GROUP_DUP)))
3021                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3022         return flags;
3023 }
3024
3025 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3026 {
3027         if (flags & BTRFS_BLOCK_GROUP_DATA)
3028                 flags |= root->fs_info->avail_data_alloc_bits &
3029                          root->fs_info->data_alloc_profile;
3030         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3031                 flags |= root->fs_info->avail_system_alloc_bits &
3032                          root->fs_info->system_alloc_profile;
3033         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3034                 flags |= root->fs_info->avail_metadata_alloc_bits &
3035                          root->fs_info->metadata_alloc_profile;
3036         return btrfs_reduce_alloc_profile(root, flags);
3037 }
3038
3039 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3040 {
3041         u64 flags;
3042
3043         if (data)
3044                 flags = BTRFS_BLOCK_GROUP_DATA;
3045         else if (root == root->fs_info->chunk_root)
3046                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3047         else
3048                 flags = BTRFS_BLOCK_GROUP_METADATA;
3049
3050         return get_alloc_profile(root, flags);
3051 }
3052
3053 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3054 {
3055         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3056                                                        BTRFS_BLOCK_GROUP_DATA);
3057 }
3058
3059 /*
3060  * This will check the space that the inode allocates from to make sure we have
3061  * enough space for bytes.
3062  */
3063 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3064 {
3065         struct btrfs_space_info *data_sinfo;
3066         struct btrfs_root *root = BTRFS_I(inode)->root;
3067         u64 used;
3068         int ret = 0, committed = 0, alloc_chunk = 1;
3069
3070         /* make sure bytes are sectorsize aligned */
3071         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3072
3073         if (root == root->fs_info->tree_root ||
3074             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3075                 alloc_chunk = 0;
3076                 committed = 1;
3077         }
3078
3079         data_sinfo = BTRFS_I(inode)->space_info;
3080         if (!data_sinfo)
3081                 goto alloc;
3082
3083 again:
3084         /* make sure we have enough space to handle the data first */
3085         spin_lock(&data_sinfo->lock);
3086         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3087                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3088                 data_sinfo->bytes_may_use;
3089
3090         if (used + bytes > data_sinfo->total_bytes) {
3091                 struct btrfs_trans_handle *trans;
3092
3093                 /*
3094                  * if we don't have enough free bytes in this space then we need
3095                  * to alloc a new chunk.
3096                  */
3097                 if (!data_sinfo->full && alloc_chunk) {
3098                         u64 alloc_target;
3099
3100                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3101                         spin_unlock(&data_sinfo->lock);
3102 alloc:
3103                         alloc_target = btrfs_get_alloc_profile(root, 1);
3104                         trans = btrfs_join_transaction(root);
3105                         if (IS_ERR(trans))
3106                                 return PTR_ERR(trans);
3107
3108                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3109                                              bytes + 2 * 1024 * 1024,
3110                                              alloc_target,
3111                                              CHUNK_ALLOC_NO_FORCE);
3112                         btrfs_end_transaction(trans, root);
3113                         if (ret < 0) {
3114                                 if (ret != -ENOSPC)
3115                                         return ret;
3116                                 else
3117                                         goto commit_trans;
3118                         }
3119
3120                         if (!data_sinfo) {
3121                                 btrfs_set_inode_space_info(root, inode);
3122                                 data_sinfo = BTRFS_I(inode)->space_info;
3123                         }
3124                         goto again;
3125                 }
3126
3127                 /*
3128                  * If we have less pinned bytes than we want to allocate then
3129                  * don't bother committing the transaction, it won't help us.
3130                  */
3131                 if (data_sinfo->bytes_pinned < bytes)
3132                         committed = 1;
3133                 spin_unlock(&data_sinfo->lock);
3134
3135                 /* commit the current transaction and try again */
3136 commit_trans:
3137                 if (!committed &&
3138                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3139                         committed = 1;
3140                         trans = btrfs_join_transaction(root);
3141                         if (IS_ERR(trans))
3142                                 return PTR_ERR(trans);
3143                         ret = btrfs_commit_transaction(trans, root);
3144                         if (ret)
3145                                 return ret;
3146                         goto again;
3147                 }
3148
3149                 return -ENOSPC;
3150         }
3151         data_sinfo->bytes_may_use += bytes;
3152         spin_unlock(&data_sinfo->lock);
3153
3154         return 0;
3155 }
3156
3157 /*
3158  * Called if we need to clear a data reservation for this inode.
3159  */
3160 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3161 {
3162         struct btrfs_root *root = BTRFS_I(inode)->root;
3163         struct btrfs_space_info *data_sinfo;
3164
3165         /* make sure bytes are sectorsize aligned */
3166         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3167
3168         data_sinfo = BTRFS_I(inode)->space_info;
3169         spin_lock(&data_sinfo->lock);
3170         data_sinfo->bytes_may_use -= bytes;
3171         spin_unlock(&data_sinfo->lock);
3172 }
3173
3174 static void force_metadata_allocation(struct btrfs_fs_info *info)
3175 {
3176         struct list_head *head = &info->space_info;
3177         struct btrfs_space_info *found;
3178
3179         rcu_read_lock();
3180         list_for_each_entry_rcu(found, head, list) {
3181                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3182                         found->force_alloc = CHUNK_ALLOC_FORCE;
3183         }
3184         rcu_read_unlock();
3185 }
3186
3187 static int should_alloc_chunk(struct btrfs_root *root,
3188                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3189                               int force)
3190 {
3191         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3192         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3193         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3194         u64 thresh;
3195
3196         if (force == CHUNK_ALLOC_FORCE)
3197                 return 1;
3198
3199         /*
3200          * We need to take into account the global rsv because for all intents
3201          * and purposes it's used space.  Don't worry about locking the
3202          * global_rsv, it doesn't change except when the transaction commits.
3203          */
3204         num_allocated += global_rsv->size;
3205
3206         /*
3207          * in limited mode, we want to have some free space up to
3208          * about 1% of the FS size.
3209          */
3210         if (force == CHUNK_ALLOC_LIMITED) {
3211                 thresh = btrfs_super_total_bytes(&root->fs_info->super_copy);
3212                 thresh = max_t(u64, 64 * 1024 * 1024,
3213                                div_factor_fine(thresh, 1));
3214
3215                 if (num_bytes - num_allocated < thresh)
3216                         return 1;
3217         }
3218
3219         /*
3220          * we have two similar checks here, one based on percentage
3221          * and once based on a hard number of 256MB.  The idea
3222          * is that if we have a good amount of free
3223          * room, don't allocate a chunk.  A good mount is
3224          * less than 80% utilized of the chunks we have allocated,
3225          * or more than 256MB free
3226          */
3227         if (num_allocated + alloc_bytes + 256 * 1024 * 1024 < num_bytes)
3228                 return 0;
3229
3230         if (num_allocated + alloc_bytes < div_factor(num_bytes, 8))
3231                 return 0;
3232
3233         thresh = btrfs_super_total_bytes(&root->fs_info->super_copy);
3234
3235         /* 256MB or 5% of the FS */
3236         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 5));
3237
3238         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 3))
3239                 return 0;
3240         return 1;
3241 }
3242
3243 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3244                           struct btrfs_root *extent_root, u64 alloc_bytes,
3245                           u64 flags, int force)
3246 {
3247         struct btrfs_space_info *space_info;
3248         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3249         int wait_for_alloc = 0;
3250         int ret = 0;
3251
3252         flags = btrfs_reduce_alloc_profile(extent_root, flags);
3253
3254         space_info = __find_space_info(extent_root->fs_info, flags);
3255         if (!space_info) {
3256                 ret = update_space_info(extent_root->fs_info, flags,
3257                                         0, 0, &space_info);
3258                 BUG_ON(ret);
3259         }
3260         BUG_ON(!space_info);
3261
3262 again:
3263         spin_lock(&space_info->lock);
3264         if (space_info->force_alloc)
3265                 force = space_info->force_alloc;
3266         if (space_info->full) {
3267                 spin_unlock(&space_info->lock);
3268                 return 0;
3269         }
3270
3271         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3272                 spin_unlock(&space_info->lock);
3273                 return 0;
3274         } else if (space_info->chunk_alloc) {
3275                 wait_for_alloc = 1;
3276         } else {
3277                 space_info->chunk_alloc = 1;
3278         }
3279
3280         spin_unlock(&space_info->lock);
3281
3282         mutex_lock(&fs_info->chunk_mutex);
3283
3284         /*
3285          * The chunk_mutex is held throughout the entirety of a chunk
3286          * allocation, so once we've acquired the chunk_mutex we know that the
3287          * other guy is done and we need to recheck and see if we should
3288          * allocate.
3289          */
3290         if (wait_for_alloc) {
3291                 mutex_unlock(&fs_info->chunk_mutex);
3292                 wait_for_alloc = 0;
3293                 goto again;
3294         }
3295
3296         /*
3297          * If we have mixed data/metadata chunks we want to make sure we keep
3298          * allocating mixed chunks instead of individual chunks.
3299          */
3300         if (btrfs_mixed_space_info(space_info))
3301                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3302
3303         /*
3304          * if we're doing a data chunk, go ahead and make sure that
3305          * we keep a reasonable number of metadata chunks allocated in the
3306          * FS as well.
3307          */
3308         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3309                 fs_info->data_chunk_allocations++;
3310                 if (!(fs_info->data_chunk_allocations %
3311                       fs_info->metadata_ratio))
3312                         force_metadata_allocation(fs_info);
3313         }
3314
3315         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3316         if (ret < 0 && ret != -ENOSPC)
3317                 goto out;
3318
3319         spin_lock(&space_info->lock);
3320         if (ret)
3321                 space_info->full = 1;
3322         else
3323                 ret = 1;
3324
3325         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3326         space_info->chunk_alloc = 0;
3327         spin_unlock(&space_info->lock);
3328 out:
3329         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3330         return ret;
3331 }
3332
3333 /*
3334  * shrink metadata reservation for delalloc
3335  */
3336 static int shrink_delalloc(struct btrfs_trans_handle *trans,
3337                            struct btrfs_root *root, u64 to_reclaim,
3338                            bool wait_ordered)
3339 {
3340         struct btrfs_block_rsv *block_rsv;
3341         struct btrfs_space_info *space_info;
3342         u64 reserved;
3343         u64 max_reclaim;
3344         u64 reclaimed = 0;
3345         long time_left;
3346         int nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3347         int loops = 0;
3348         unsigned long progress;
3349
3350         block_rsv = &root->fs_info->delalloc_block_rsv;
3351         space_info = block_rsv->space_info;
3352
3353         smp_mb();
3354         reserved = space_info->bytes_may_use;
3355         progress = space_info->reservation_progress;
3356
3357         if (reserved == 0)
3358                 return 0;
3359
3360         smp_mb();
3361         if (root->fs_info->delalloc_bytes == 0) {
3362                 if (trans)
3363                         return 0;
3364                 btrfs_wait_ordered_extents(root, 0, 0);
3365                 return 0;
3366         }
3367
3368         max_reclaim = min(reserved, to_reclaim);
3369
3370         while (loops < 1024) {
3371                 /* have the flusher threads jump in and do some IO */
3372                 smp_mb();
3373                 nr_pages = min_t(unsigned long, nr_pages,
3374                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3375                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages);
3376
3377                 spin_lock(&space_info->lock);
3378                 if (reserved > space_info->bytes_may_use)
3379                         reclaimed += reserved - space_info->bytes_may_use;
3380                 reserved = space_info->bytes_may_use;
3381                 spin_unlock(&space_info->lock);
3382
3383                 loops++;
3384
3385                 if (reserved == 0 || reclaimed >= max_reclaim)
3386                         break;
3387
3388                 if (trans && trans->transaction->blocked)
3389                         return -EAGAIN;
3390
3391                 if (wait_ordered && !trans) {
3392                         btrfs_wait_ordered_extents(root, 0, 0);
3393                 } else {
3394                         time_left = schedule_timeout_interruptible(1);
3395
3396                         /* We were interrupted, exit */
3397                         if (time_left)
3398                                 break;
3399                 }
3400
3401                 /* we've kicked the IO a few times, if anything has been freed,
3402                  * exit.  There is no sense in looping here for a long time
3403                  * when we really need to commit the transaction, or there are
3404                  * just too many writers without enough free space
3405                  */
3406
3407                 if (loops > 3) {
3408                         smp_mb();
3409                         if (progress != space_info->reservation_progress)
3410                                 break;
3411                 }
3412
3413         }
3414
3415         return reclaimed >= to_reclaim;
3416 }
3417
3418 /**
3419  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3420  * @root - the root we're allocating for
3421  * @block_rsv - the block_rsv we're allocating for
3422  * @orig_bytes - the number of bytes we want
3423  * @flush - wether or not we can flush to make our reservation
3424  * @check - wether this is just to check if we have enough space or not
3425  *
3426  * This will reserve orgi_bytes number of bytes from the space info associated
3427  * with the block_rsv.  If there is not enough space it will make an attempt to
3428  * flush out space to make room.  It will do this by flushing delalloc if
3429  * possible or committing the transaction.  If flush is 0 then no attempts to
3430  * regain reservations will be made and this will fail if there is not enough
3431  * space already.
3432  */
3433 static int reserve_metadata_bytes(struct btrfs_root *root,
3434                                   struct btrfs_block_rsv *block_rsv,
3435                                   u64 orig_bytes, int flush, int check)
3436 {
3437         struct btrfs_space_info *space_info = block_rsv->space_info;
3438         struct btrfs_trans_handle *trans;
3439         u64 used;
3440         u64 num_bytes = orig_bytes;
3441         int retries = 0;
3442         int ret = 0;
3443         bool committed = false;
3444         bool flushing = false;
3445         bool wait_ordered = false;
3446
3447         trans = (struct btrfs_trans_handle *)current->journal_info;
3448 again:
3449         ret = 0;
3450         spin_lock(&space_info->lock);
3451         /*
3452          * We only want to wait if somebody other than us is flushing and we are
3453          * actually alloed to flush.
3454          */
3455         while (flush && !flushing && space_info->flush) {
3456                 spin_unlock(&space_info->lock);
3457                 /*
3458                  * If we have a trans handle we can't wait because the flusher
3459                  * may have to commit the transaction, which would mean we would
3460                  * deadlock since we are waiting for the flusher to finish, but
3461                  * hold the current transaction open.
3462                  */
3463                 if (trans)
3464                         return -EAGAIN;
3465                 ret = wait_event_interruptible(space_info->wait,
3466                                                !space_info->flush);
3467                 /* Must have been interrupted, return */
3468                 if (ret)
3469                         return -EINTR;
3470
3471                 spin_lock(&space_info->lock);
3472         }
3473
3474         ret = -ENOSPC;
3475         used = space_info->bytes_used + space_info->bytes_reserved +
3476                 space_info->bytes_pinned + space_info->bytes_readonly +
3477                 space_info->bytes_may_use;
3478
3479         /*
3480          * The idea here is that we've not already over-reserved the block group
3481          * then we can go ahead and save our reservation first and then start
3482          * flushing if we need to.  Otherwise if we've already overcommitted
3483          * lets start flushing stuff first and then come back and try to make
3484          * our reservation.
3485          */
3486         if (used <= space_info->total_bytes) {
3487                 if (used + orig_bytes <= space_info->total_bytes) {
3488                         space_info->bytes_may_use += orig_bytes;
3489                         ret = 0;
3490                 } else {
3491                         /*
3492                          * Ok set num_bytes to orig_bytes since we aren't
3493                          * overocmmitted, this way we only try and reclaim what
3494                          * we need.
3495                          */
3496                         num_bytes = orig_bytes;
3497                 }
3498         } else {
3499                 /*
3500                  * Ok we're over committed, set num_bytes to the overcommitted
3501                  * amount plus the amount of bytes that we need for this
3502                  * reservation.
3503                  */
3504                 wait_ordered = true;
3505                 num_bytes = used - space_info->total_bytes +
3506                         (orig_bytes * (retries + 1));
3507         }
3508
3509         if (ret && !check) {
3510                 u64 profile = btrfs_get_alloc_profile(root, 0);
3511                 u64 avail;
3512
3513                 spin_lock(&root->fs_info->free_chunk_lock);
3514                 avail = root->fs_info->free_chunk_space;
3515
3516                 /*
3517                  * If we have dup, raid1 or raid10 then only half of the free
3518                  * space is actually useable.
3519                  */
3520                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3521                                BTRFS_BLOCK_GROUP_RAID1 |
3522                                BTRFS_BLOCK_GROUP_RAID10))
3523                         avail >>= 1;
3524
3525                 /*
3526                  * If we aren't flushing don't let us overcommit too much, say
3527                  * 1/8th of the space.  If we can flush, let it overcommit up to
3528                  * 1/2 of the space.
3529                  */
3530                 if (flush)
3531                         avail >>= 3;
3532                 else
3533                         avail >>= 1;
3534                  spin_unlock(&root->fs_info->free_chunk_lock);
3535
3536                 if (used + num_bytes < space_info->total_bytes + avail) {
3537                         space_info->bytes_may_use += orig_bytes;
3538                         ret = 0;
3539                 } else {
3540                         wait_ordered = true;
3541                 }
3542         }
3543
3544         /*
3545          * Couldn't make our reservation, save our place so while we're trying
3546          * to reclaim space we can actually use it instead of somebody else
3547          * stealing it from us.
3548          */
3549         if (ret && flush) {
3550                 flushing = true;
3551                 space_info->flush = 1;
3552         }
3553
3554         spin_unlock(&space_info->lock);
3555
3556         if (!ret || !flush)
3557                 goto out;
3558
3559         /*
3560          * We do synchronous shrinking since we don't actually unreserve
3561          * metadata until after the IO is completed.
3562          */
3563         ret = shrink_delalloc(trans, root, num_bytes, wait_ordered);
3564         if (ret < 0)
3565                 goto out;
3566
3567         ret = 0;
3568
3569         /*
3570          * So if we were overcommitted it's possible that somebody else flushed
3571          * out enough space and we simply didn't have enough space to reclaim,
3572          * so go back around and try again.
3573          */
3574         if (retries < 2) {
3575                 wait_ordered = true;
3576                 retries++;
3577                 goto again;
3578         }
3579
3580         ret = -EAGAIN;
3581         if (trans)
3582                 goto out;
3583
3584         ret = -ENOSPC;
3585         if (committed)
3586                 goto out;
3587
3588         trans = btrfs_join_transaction(root);
3589         if (IS_ERR(trans))
3590                 goto out;
3591         ret = btrfs_commit_transaction(trans, root);
3592         if (!ret) {
3593                 trans = NULL;
3594                 committed = true;
3595                 goto again;
3596         }
3597
3598 out:
3599         if (flushing) {
3600                 spin_lock(&space_info->lock);
3601                 space_info->flush = 0;
3602                 wake_up_all(&space_info->wait);
3603                 spin_unlock(&space_info->lock);
3604         }
3605         return ret;
3606 }
3607
3608 static struct btrfs_block_rsv *get_block_rsv(struct btrfs_trans_handle *trans,
3609                                              struct btrfs_root *root)
3610 {
3611         struct btrfs_block_rsv *block_rsv = NULL;
3612
3613         if (root->ref_cows || root == root->fs_info->csum_root)
3614                 block_rsv = trans->block_rsv;
3615
3616         if (!block_rsv)
3617                 block_rsv = root->block_rsv;
3618
3619         if (!block_rsv)
3620                 block_rsv = &root->fs_info->empty_block_rsv;
3621
3622         return block_rsv;
3623 }
3624
3625 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3626                                u64 num_bytes)
3627 {
3628         int ret = -ENOSPC;
3629         spin_lock(&block_rsv->lock);
3630         if (block_rsv->reserved >= num_bytes) {
3631                 block_rsv->reserved -= num_bytes;
3632                 if (block_rsv->reserved < block_rsv->size)
3633                         block_rsv->full = 0;
3634                 ret = 0;
3635         }
3636         spin_unlock(&block_rsv->lock);
3637         return ret;
3638 }
3639
3640 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3641                                 u64 num_bytes, int update_size)
3642 {
3643         spin_lock(&block_rsv->lock);
3644         block_rsv->reserved += num_bytes;
3645         if (update_size)
3646                 block_rsv->size += num_bytes;
3647         else if (block_rsv->reserved >= block_rsv->size)
3648                 block_rsv->full = 1;
3649         spin_unlock(&block_rsv->lock);
3650 }
3651
3652 static void block_rsv_release_bytes(struct btrfs_block_rsv *block_rsv,
3653                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3654 {
3655         struct btrfs_space_info *space_info = block_rsv->space_info;
3656
3657         spin_lock(&block_rsv->lock);
3658         if (num_bytes == (u64)-1)
3659                 num_bytes = block_rsv->size;
3660         block_rsv->size -= num_bytes;
3661         if (block_rsv->reserved >= block_rsv->size) {
3662                 num_bytes = block_rsv->reserved - block_rsv->size;
3663                 block_rsv->reserved = block_rsv->size;
3664                 block_rsv->full = 1;
3665         } else {
3666                 num_bytes = 0;
3667         }
3668         spin_unlock(&block_rsv->lock);
3669
3670         if (num_bytes > 0) {
3671                 if (dest) {
3672                         spin_lock(&dest->lock);
3673                         if (!dest->full) {
3674                                 u64 bytes_to_add;
3675
3676                                 bytes_to_add = dest->size - dest->reserved;
3677                                 bytes_to_add = min(num_bytes, bytes_to_add);
3678                                 dest->reserved += bytes_to_add;
3679                                 if (dest->reserved >= dest->size)
3680                                         dest->full = 1;
3681                                 num_bytes -= bytes_to_add;
3682                         }
3683                         spin_unlock(&dest->lock);
3684                 }
3685                 if (num_bytes) {
3686                         spin_lock(&space_info->lock);
3687                         space_info->bytes_may_use -= num_bytes;
3688                         space_info->reservation_progress++;
3689                         spin_unlock(&space_info->lock);
3690                 }
3691         }
3692 }
3693
3694 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3695                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3696 {
3697         int ret;
3698
3699         ret = block_rsv_use_bytes(src, num_bytes);
3700         if (ret)
3701                 return ret;
3702
3703         block_rsv_add_bytes(dst, num_bytes, 1);
3704         return 0;
3705 }
3706
3707 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3708 {
3709         memset(rsv, 0, sizeof(*rsv));
3710         spin_lock_init(&rsv->lock);
3711 }
3712
3713 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3714 {
3715         struct btrfs_block_rsv *block_rsv;
3716         struct btrfs_fs_info *fs_info = root->fs_info;
3717
3718         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3719         if (!block_rsv)
3720                 return NULL;
3721
3722         btrfs_init_block_rsv(block_rsv);
3723         block_rsv->space_info = __find_space_info(fs_info,
3724                                                   BTRFS_BLOCK_GROUP_METADATA);
3725         return block_rsv;
3726 }
3727
3728 void btrfs_free_block_rsv(struct btrfs_root *root,
3729                           struct btrfs_block_rsv *rsv)
3730 {
3731         btrfs_block_rsv_release(root, rsv, (u64)-1);
3732         kfree(rsv);
3733 }
3734
3735 int btrfs_block_rsv_add(struct btrfs_root *root,
3736                         struct btrfs_block_rsv *block_rsv,
3737                         u64 num_bytes)
3738 {
3739         int ret;
3740
3741         if (num_bytes == 0)
3742                 return 0;
3743
3744         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, 1, 0);
3745         if (!ret) {
3746                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3747                 return 0;
3748         }
3749
3750         return ret;
3751 }
3752
3753 int btrfs_block_rsv_check(struct btrfs_root *root,
3754                           struct btrfs_block_rsv *block_rsv,
3755                           u64 min_reserved, int min_factor, int flush)
3756 {
3757         u64 num_bytes = 0;
3758         int ret = -ENOSPC;
3759
3760         if (!block_rsv)
3761                 return 0;
3762
3763         spin_lock(&block_rsv->lock);
3764         if (min_factor > 0)
3765                 num_bytes = div_factor(block_rsv->size, min_factor);
3766         if (min_reserved > num_bytes)
3767                 num_bytes = min_reserved;
3768
3769         if (block_rsv->reserved >= num_bytes)
3770                 ret = 0;
3771         else
3772                 num_bytes -= block_rsv->reserved;
3773         spin_unlock(&block_rsv->lock);
3774
3775         if (!ret)
3776                 return 0;
3777
3778         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush, !flush);
3779         if (!ret) {
3780                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
3781                 return 0;
3782         }
3783
3784         return ret;
3785 }
3786
3787 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
3788                             struct btrfs_block_rsv *dst_rsv,
3789                             u64 num_bytes)
3790 {
3791         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3792 }
3793
3794 void btrfs_block_rsv_release(struct btrfs_root *root,
3795                              struct btrfs_block_rsv *block_rsv,
3796                              u64 num_bytes)
3797 {
3798         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3799         if (global_rsv->full || global_rsv == block_rsv ||
3800             block_rsv->space_info != global_rsv->space_info)
3801                 global_rsv = NULL;
3802         block_rsv_release_bytes(block_rsv, global_rsv, num_bytes);
3803 }
3804
3805 /*
3806  * helper to calculate size of global block reservation.
3807  * the desired value is sum of space used by extent tree,
3808  * checksum tree and root tree
3809  */
3810 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
3811 {
3812         struct btrfs_space_info *sinfo;
3813         u64 num_bytes;
3814         u64 meta_used;
3815         u64 data_used;
3816         int csum_size = btrfs_super_csum_size(&fs_info->super_copy);
3817
3818         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
3819         spin_lock(&sinfo->lock);
3820         data_used = sinfo->bytes_used;
3821         spin_unlock(&sinfo->lock);
3822
3823         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3824         spin_lock(&sinfo->lock);
3825         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
3826                 data_used = 0;
3827         meta_used = sinfo->bytes_used;
3828         spin_unlock(&sinfo->lock);
3829
3830         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
3831                     csum_size * 2;
3832         num_bytes += div64_u64(data_used + meta_used, 50);
3833
3834         if (num_bytes * 3 > meta_used)
3835                 num_bytes = div64_u64(meta_used, 3);
3836
3837         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
3838 }
3839
3840 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
3841 {
3842         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
3843         struct btrfs_space_info *sinfo = block_rsv->space_info;
3844         u64 num_bytes;
3845
3846         num_bytes = calc_global_metadata_size(fs_info);
3847
3848         spin_lock(&block_rsv->lock);
3849         spin_lock(&sinfo->lock);
3850
3851         block_rsv->size = num_bytes;
3852
3853         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
3854                     sinfo->bytes_reserved + sinfo->bytes_readonly +
3855                     sinfo->bytes_may_use;
3856
3857         if (sinfo->total_bytes > num_bytes) {
3858                 num_bytes = sinfo->total_bytes - num_bytes;
3859                 block_rsv->reserved += num_bytes;
3860                 sinfo->bytes_may_use += num_bytes;
3861         }
3862
3863         if (block_rsv->reserved >= block_rsv->size) {
3864                 num_bytes = block_rsv->reserved - block_rsv->size;
3865                 sinfo->bytes_may_use -= num_bytes;
3866                 sinfo->reservation_progress++;
3867                 block_rsv->reserved = block_rsv->size;
3868                 block_rsv->full = 1;
3869         }
3870
3871         spin_unlock(&sinfo->lock);
3872         spin_unlock(&block_rsv->lock);
3873 }
3874
3875 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
3876 {
3877         struct btrfs_space_info *space_info;
3878
3879         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3880         fs_info->chunk_block_rsv.space_info = space_info;
3881
3882         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3883         fs_info->global_block_rsv.space_info = space_info;
3884         fs_info->delalloc_block_rsv.space_info = space_info;
3885         fs_info->trans_block_rsv.space_info = space_info;
3886         fs_info->empty_block_rsv.space_info = space_info;
3887
3888         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
3889         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
3890         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
3891         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
3892         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
3893
3894         update_global_block_rsv(fs_info);
3895 }
3896
3897 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
3898 {
3899         block_rsv_release_bytes(&fs_info->global_block_rsv, NULL, (u64)-1);
3900         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
3901         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
3902         WARN_ON(fs_info->trans_block_rsv.size > 0);
3903         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
3904         WARN_ON(fs_info->chunk_block_rsv.size > 0);
3905         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
3906 }
3907
3908 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
3909                                   struct btrfs_root *root)
3910 {
3911         struct btrfs_block_rsv *block_rsv;
3912
3913         if (!trans->bytes_reserved)
3914                 return;
3915
3916         block_rsv = &root->fs_info->trans_block_rsv;
3917         btrfs_block_rsv_release(root, block_rsv, trans->bytes_reserved);
3918         trans->bytes_reserved = 0;
3919 }
3920
3921 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
3922                                   struct inode *inode)
3923 {
3924         struct btrfs_root *root = BTRFS_I(inode)->root;
3925         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
3926         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
3927
3928         /*
3929          * We need to hold space in order to delete our orphan item once we've
3930          * added it, so this takes the reservation so we can release it later
3931          * when we are truly done with the orphan item.
3932          */
3933         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
3934         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3935 }
3936
3937 void btrfs_orphan_release_metadata(struct inode *inode)
3938 {
3939         struct btrfs_root *root = BTRFS_I(inode)->root;
3940         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
3941         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
3942 }
3943
3944 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
3945                                 struct btrfs_pending_snapshot *pending)
3946 {
3947         struct btrfs_root *root = pending->root;
3948         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
3949         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
3950         /*
3951          * two for root back/forward refs, two for directory entries
3952          * and one for root of the snapshot.
3953          */
3954         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
3955         dst_rsv->space_info = src_rsv->space_info;
3956         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3957 }
3958
3959 /**
3960  * drop_outstanding_extent - drop an outstanding extent
3961  * @inode: the inode we're dropping the extent for
3962  *
3963  * This is called when we are freeing up an outstanding extent, either called
3964  * after an error or after an extent is written.  This will return the number of
3965  * reserved extents that need to be freed.  This must be called with
3966  * BTRFS_I(inode)->lock held.
3967  */
3968 static unsigned drop_outstanding_extent(struct inode *inode)
3969 {
3970         unsigned dropped_extents = 0;
3971
3972         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
3973         BTRFS_I(inode)->outstanding_extents--;
3974
3975         /*
3976          * If we have more or the same amount of outsanding extents than we have
3977          * reserved then we need to leave the reserved extents count alone.
3978          */
3979         if (BTRFS_I(inode)->outstanding_extents >=
3980             BTRFS_I(inode)->reserved_extents)
3981                 return 0;
3982
3983         dropped_extents = BTRFS_I(inode)->reserved_extents -
3984                 BTRFS_I(inode)->outstanding_extents;
3985         BTRFS_I(inode)->reserved_extents -= dropped_extents;
3986         return dropped_extents;
3987 }
3988
3989 /**
3990  * calc_csum_metadata_size - return the amount of metada space that must be
3991  *      reserved/free'd for the given bytes.
3992  * @inode: the inode we're manipulating
3993  * @num_bytes: the number of bytes in question
3994  * @reserve: 1 if we are reserving space, 0 if we are freeing space
3995  *
3996  * This adjusts the number of csum_bytes in the inode and then returns the
3997  * correct amount of metadata that must either be reserved or freed.  We
3998  * calculate how many checksums we can fit into one leaf and then divide the
3999  * number of bytes that will need to be checksumed by this value to figure out
4000  * how many checksums will be required.  If we are adding bytes then the number
4001  * may go up and we will return the number of additional bytes that must be
4002  * reserved.  If it is going down we will return the number of bytes that must
4003  * be freed.
4004  *
4005  * This must be called with BTRFS_I(inode)->lock held.
4006  */
4007 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4008                                    int reserve)
4009 {
4010         struct btrfs_root *root = BTRFS_I(inode)->root;
4011         u64 csum_size;
4012         int num_csums_per_leaf;
4013         int num_csums;
4014         int old_csums;
4015
4016         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4017             BTRFS_I(inode)->csum_bytes == 0)
4018                 return 0;
4019
4020         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4021         if (reserve)
4022                 BTRFS_I(inode)->csum_bytes += num_bytes;
4023         else
4024                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4025         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4026         num_csums_per_leaf = (int)div64_u64(csum_size,
4027                                             sizeof(struct btrfs_csum_item) +
4028                                             sizeof(struct btrfs_disk_key));
4029         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4030         num_csums = num_csums + num_csums_per_leaf - 1;
4031         num_csums = num_csums / num_csums_per_leaf;
4032
4033         old_csums = old_csums + num_csums_per_leaf - 1;
4034         old_csums = old_csums / num_csums_per_leaf;
4035
4036         /* No change, no need to reserve more */
4037         if (old_csums == num_csums)
4038                 return 0;
4039
4040         if (reserve)
4041                 return btrfs_calc_trans_metadata_size(root,
4042                                                       num_csums - old_csums);
4043
4044         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4045 }
4046
4047 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4048 {
4049         struct btrfs_root *root = BTRFS_I(inode)->root;
4050         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4051         u64 to_reserve = 0;
4052         unsigned nr_extents = 0;
4053         int flush = 1;
4054         int ret;
4055
4056         if (btrfs_is_free_space_inode(root, inode))
4057                 flush = 0;
4058
4059         if (flush && btrfs_transaction_in_commit(root->fs_info))
4060                 schedule_timeout(1);
4061
4062         num_bytes = ALIGN(num_bytes, root->sectorsize);
4063
4064         spin_lock(&BTRFS_I(inode)->lock);
4065         BTRFS_I(inode)->outstanding_extents++;
4066
4067         if (BTRFS_I(inode)->outstanding_extents >
4068             BTRFS_I(inode)->reserved_extents) {
4069                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4070                         BTRFS_I(inode)->reserved_extents;
4071                 BTRFS_I(inode)->reserved_extents += nr_extents;
4072
4073                 to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4074         }
4075         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4076         spin_unlock(&BTRFS_I(inode)->lock);
4077
4078         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush, 0);
4079         if (ret) {
4080                 u64 to_free = 0;
4081                 unsigned dropped;
4082
4083                 spin_lock(&BTRFS_I(inode)->lock);
4084                 dropped = drop_outstanding_extent(inode);
4085                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4086                 spin_unlock(&BTRFS_I(inode)->lock);
4087                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4088
4089                 /*
4090                  * Somebody could have come in and twiddled with the
4091                  * reservation, so if we have to free more than we would have
4092                  * reserved from this reservation go ahead and release those
4093                  * bytes.
4094                  */
4095                 to_free -= to_reserve;
4096                 if (to_free)
4097                         btrfs_block_rsv_release(root, block_rsv, to_free);
4098                 return ret;
4099         }
4100
4101         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4102
4103         return 0;
4104 }
4105
4106 /**
4107  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4108  * @inode: the inode to release the reservation for
4109  * @num_bytes: the number of bytes we're releasing
4110  *
4111  * This will release the metadata reservation for an inode.  This can be called
4112  * once we complete IO for a given set of bytes to release their metadata
4113  * reservations.
4114  */
4115 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4116 {
4117         struct btrfs_root *root = BTRFS_I(inode)->root;
4118         u64 to_free = 0;
4119         unsigned dropped;
4120
4121         num_bytes = ALIGN(num_bytes, root->sectorsize);
4122         spin_lock(&BTRFS_I(inode)->lock);
4123         dropped = drop_outstanding_extent(inode);
4124
4125         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4126         spin_unlock(&BTRFS_I(inode)->lock);
4127         if (dropped > 0)
4128                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4129
4130         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4131                                 to_free);
4132 }
4133
4134 /**
4135  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4136  * @inode: inode we're writing to
4137  * @num_bytes: the number of bytes we want to allocate
4138  *
4139  * This will do the following things
4140  *
4141  * o reserve space in the data space info for num_bytes
4142  * o reserve space in the metadata space info based on number of outstanding
4143  *   extents and how much csums will be needed
4144  * o add to the inodes ->delalloc_bytes
4145  * o add it to the fs_info's delalloc inodes list.
4146  *
4147  * This will return 0 for success and -ENOSPC if there is no space left.
4148  */
4149 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4150 {
4151         int ret;
4152
4153         ret = btrfs_check_data_free_space(inode, num_bytes);
4154         if (ret)
4155                 return ret;
4156
4157         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4158         if (ret) {
4159                 btrfs_free_reserved_data_space(inode, num_bytes);
4160                 return ret;
4161         }
4162
4163         return 0;
4164 }
4165
4166 /**
4167  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4168  * @inode: inode we're releasing space for
4169  * @num_bytes: the number of bytes we want to free up
4170  *
4171  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4172  * called in the case that we don't need the metadata AND data reservations
4173  * anymore.  So if there is an error or we insert an inline extent.
4174  *
4175  * This function will release the metadata space that was not used and will
4176  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4177  * list if there are no delalloc bytes left.
4178  */
4179 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4180 {
4181         btrfs_delalloc_release_metadata(inode, num_bytes);
4182         btrfs_free_reserved_data_space(inode, num_bytes);
4183 }
4184
4185 static int update_block_group(struct btrfs_trans_handle *trans,
4186                               struct btrfs_root *root,
4187                               u64 bytenr, u64 num_bytes, int alloc)
4188 {
4189         struct btrfs_block_group_cache *cache = NULL;
4190         struct btrfs_fs_info *info = root->fs_info;
4191         u64 total = num_bytes;
4192         u64 old_val;
4193         u64 byte_in_group;
4194         int factor;
4195
4196         /* block accounting for super block */
4197         spin_lock(&info->delalloc_lock);
4198         old_val = btrfs_super_bytes_used(&info->super_copy);
4199         if (alloc)
4200                 old_val += num_bytes;
4201         else
4202                 old_val -= num_bytes;
4203         btrfs_set_super_bytes_used(&info->super_copy, old_val);
4204         spin_unlock(&info->delalloc_lock);
4205
4206         while (total) {
4207                 cache = btrfs_lookup_block_group(info, bytenr);
4208                 if (!cache)
4209                         return -1;
4210                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4211                                     BTRFS_BLOCK_GROUP_RAID1 |
4212                                     BTRFS_BLOCK_GROUP_RAID10))
4213                         factor = 2;
4214                 else
4215                         factor = 1;
4216                 /*
4217                  * If this block group has free space cache written out, we
4218                  * need to make sure to load it if we are removing space.  This
4219                  * is because we need the unpinning stage to actually add the
4220                  * space back to the block group, otherwise we will leak space.
4221                  */
4222                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4223                         cache_block_group(cache, trans, NULL, 1);
4224
4225                 byte_in_group = bytenr - cache->key.objectid;
4226                 WARN_ON(byte_in_group > cache->key.offset);
4227
4228                 spin_lock(&cache->space_info->lock);
4229                 spin_lock(&cache->lock);
4230
4231                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4232                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4233                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4234
4235                 cache->dirty = 1;
4236                 old_val = btrfs_block_group_used(&cache->item);
4237                 num_bytes = min(total, cache->key.offset - byte_in_group);
4238                 if (alloc) {
4239                         old_val += num_bytes;
4240                         btrfs_set_block_group_used(&cache->item, old_val);
4241                         cache->reserved -= num_bytes;
4242                         cache->space_info->bytes_reserved -= num_bytes;
4243                         cache->space_info->bytes_used += num_bytes;
4244                         cache->space_info->disk_used += num_bytes * factor;
4245                         spin_unlock(&cache->lock);
4246                         spin_unlock(&cache->space_info->lock);
4247                 } else {
4248                         old_val -= num_bytes;
4249                         btrfs_set_block_group_used(&cache->item, old_val);
4250                         cache->pinned += num_bytes;
4251                         cache->space_info->bytes_pinned += num_bytes;
4252                         cache->space_info->bytes_used -= num_bytes;
4253                         cache->space_info->disk_used -= num_bytes * factor;
4254                         spin_unlock(&cache->lock);
4255                         spin_unlock(&cache->space_info->lock);
4256
4257                         set_extent_dirty(info->pinned_extents,
4258                                          bytenr, bytenr + num_bytes - 1,
4259                                          GFP_NOFS | __GFP_NOFAIL);
4260                 }
4261                 btrfs_put_block_group(cache);
4262                 total -= num_bytes;
4263                 bytenr += num_bytes;
4264         }
4265         return 0;
4266 }
4267
4268 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4269 {
4270         struct btrfs_block_group_cache *cache;
4271         u64 bytenr;
4272
4273         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4274         if (!cache)
4275                 return 0;
4276
4277         bytenr = cache->key.objectid;
4278         btrfs_put_block_group(cache);
4279
4280         return bytenr;
4281 }
4282
4283 static int pin_down_extent(struct btrfs_root *root,
4284                            struct btrfs_block_group_cache *cache,
4285                            u64 bytenr, u64 num_bytes, int reserved)
4286 {
4287         spin_lock(&cache->space_info->lock);
4288         spin_lock(&cache->lock);
4289         cache->pinned += num_bytes;
4290         cache->space_info->bytes_pinned += num_bytes;
4291         if (reserved) {
4292                 cache->reserved -= num_bytes;
4293                 cache->space_info->bytes_reserved -= num_bytes;
4294         }
4295         spin_unlock(&cache->lock);
4296         spin_unlock(&cache->space_info->lock);
4297
4298         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4299                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4300         return 0;
4301 }
4302
4303 /*
4304  * this function must be called within transaction
4305  */
4306 int btrfs_pin_extent(struct btrfs_root *root,
4307                      u64 bytenr, u64 num_bytes, int reserved)
4308 {
4309         struct btrfs_block_group_cache *cache;
4310
4311         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4312         BUG_ON(!cache);
4313
4314         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4315
4316         btrfs_put_block_group(cache);
4317         return 0;
4318 }
4319
4320 /**
4321  * btrfs_update_reserved_bytes - update the block_group and space info counters
4322  * @cache:      The cache we are manipulating
4323  * @num_bytes:  The number of bytes in question
4324  * @reserve:    One of the reservation enums
4325  *
4326  * This is called by the allocator when it reserves space, or by somebody who is
4327  * freeing space that was never actually used on disk.  For example if you
4328  * reserve some space for a new leaf in transaction A and before transaction A
4329  * commits you free that leaf, you call this with reserve set to 0 in order to
4330  * clear the reservation.
4331  *
4332  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4333  * ENOSPC accounting.  For data we handle the reservation through clearing the
4334  * delalloc bits in the io_tree.  We have to do this since we could end up
4335  * allocating less disk space for the amount of data we have reserved in the
4336  * case of compression.
4337  *
4338  * If this is a reservation and the block group has become read only we cannot
4339  * make the reservation and return -EAGAIN, otherwise this function always
4340  * succeeds.
4341  */
4342 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4343                                        u64 num_bytes, int reserve)
4344 {
4345         struct btrfs_space_info *space_info = cache->space_info;
4346         int ret = 0;
4347         spin_lock(&space_info->lock);
4348         spin_lock(&cache->lock);
4349         if (reserve != RESERVE_FREE) {
4350                 if (cache->ro) {
4351                         ret = -EAGAIN;
4352                 } else {
4353                         cache->reserved += num_bytes;
4354                         space_info->bytes_reserved += num_bytes;
4355                         if (reserve == RESERVE_ALLOC) {
4356                                 BUG_ON(space_info->bytes_may_use < num_bytes);
4357                                 space_info->bytes_may_use -= num_bytes;
4358                         }
4359                 }
4360         } else {
4361                 if (cache->ro)
4362                         space_info->bytes_readonly += num_bytes;
4363                 cache->reserved -= num_bytes;
4364                 space_info->bytes_reserved -= num_bytes;
4365                 space_info->reservation_progress++;
4366         }
4367         spin_unlock(&cache->lock);
4368         spin_unlock(&space_info->lock);
4369         return ret;
4370 }
4371
4372 int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4373                                 struct btrfs_root *root)
4374 {
4375         struct btrfs_fs_info *fs_info = root->fs_info;
4376         struct btrfs_caching_control *next;
4377         struct btrfs_caching_control *caching_ctl;
4378         struct btrfs_block_group_cache *cache;
4379
4380         down_write(&fs_info->extent_commit_sem);
4381
4382         list_for_each_entry_safe(caching_ctl, next,
4383                                  &fs_info->caching_block_groups, list) {
4384                 cache = caching_ctl->block_group;
4385                 if (block_group_cache_done(cache)) {
4386                         cache->last_byte_to_unpin = (u64)-1;
4387                         list_del_init(&caching_ctl->list);
4388                         put_caching_control(caching_ctl);
4389                 } else {
4390                         cache->last_byte_to_unpin = caching_ctl->progress;
4391                 }
4392         }
4393
4394         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4395                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4396         else
4397                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4398
4399         up_write(&fs_info->extent_commit_sem);
4400
4401         update_global_block_rsv(fs_info);
4402         return 0;
4403 }
4404
4405 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4406 {
4407         struct btrfs_fs_info *fs_info = root->fs_info;
4408         struct btrfs_block_group_cache *cache = NULL;
4409         u64 len;
4410
4411         while (start <= end) {
4412                 if (!cache ||
4413                     start >= cache->key.objectid + cache->key.offset) {
4414                         if (cache)
4415                                 btrfs_put_block_group(cache);
4416                         cache = btrfs_lookup_block_group(fs_info, start);
4417                         BUG_ON(!cache);
4418                 }
4419
4420                 len = cache->key.objectid + cache->key.offset - start;
4421                 len = min(len, end + 1 - start);
4422
4423                 if (start < cache->last_byte_to_unpin) {
4424                         len = min(len, cache->last_byte_to_unpin - start);
4425                         btrfs_add_free_space(cache, start, len);
4426                 }
4427
4428                 start += len;
4429
4430                 spin_lock(&cache->space_info->lock);
4431                 spin_lock(&cache->lock);
4432                 cache->pinned -= len;
4433                 cache->space_info->bytes_pinned -= len;
4434                 if (cache->ro)
4435                         cache->space_info->bytes_readonly += len;
4436                 spin_unlock(&cache->lock);
4437                 spin_unlock(&cache->space_info->lock);
4438         }
4439
4440         if (cache)
4441                 btrfs_put_block_group(cache);
4442         return 0;
4443 }
4444
4445 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4446                                struct btrfs_root *root)
4447 {
4448         struct btrfs_fs_info *fs_info = root->fs_info;
4449         struct extent_io_tree *unpin;
4450         u64 start;
4451         u64 end;
4452         int ret;
4453
4454         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4455                 unpin = &fs_info->freed_extents[1];
4456         else
4457                 unpin = &fs_info->freed_extents[0];
4458
4459         while (1) {
4460                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4461                                             EXTENT_DIRTY);
4462                 if (ret)
4463                         break;
4464
4465                 if (btrfs_test_opt(root, DISCARD))
4466                         ret = btrfs_discard_extent(root, start,
4467                                                    end + 1 - start, NULL);
4468
4469                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4470                 unpin_extent_range(root, start, end);
4471                 cond_resched();
4472         }
4473
4474         return 0;
4475 }
4476
4477 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4478                                 struct btrfs_root *root,
4479                                 u64 bytenr, u64 num_bytes, u64 parent,
4480                                 u64 root_objectid, u64 owner_objectid,
4481                                 u64 owner_offset, int refs_to_drop,
4482                                 struct btrfs_delayed_extent_op *extent_op)
4483 {
4484         struct btrfs_key key;
4485         struct btrfs_path *path;
4486         struct btrfs_fs_info *info = root->fs_info;
4487         struct btrfs_root *extent_root = info->extent_root;
4488         struct extent_buffer *leaf;
4489         struct btrfs_extent_item *ei;
4490         struct btrfs_extent_inline_ref *iref;
4491         int ret;
4492         int is_data;
4493         int extent_slot = 0;
4494         int found_extent = 0;
4495         int num_to_del = 1;
4496         u32 item_size;
4497         u64 refs;
4498
4499         path = btrfs_alloc_path();
4500         if (!path)
4501                 return -ENOMEM;
4502
4503         path->reada = 1;
4504         path->leave_spinning = 1;
4505
4506         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4507         BUG_ON(!is_data && refs_to_drop != 1);
4508
4509         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4510                                     bytenr, num_bytes, parent,
4511                                     root_objectid, owner_objectid,
4512                                     owner_offset);
4513         if (ret == 0) {
4514                 extent_slot = path->slots[0];
4515                 while (extent_slot >= 0) {
4516                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4517                                               extent_slot);
4518                         if (key.objectid != bytenr)
4519                                 break;
4520                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4521                             key.offset == num_bytes) {
4522                                 found_extent = 1;
4523                                 break;
4524                         }
4525                         if (path->slots[0] - extent_slot > 5)
4526                                 break;
4527                         extent_slot--;
4528                 }
4529 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4530                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4531                 if (found_extent && item_size < sizeof(*ei))
4532                         found_extent = 0;
4533 #endif
4534                 if (!found_extent) {
4535                         BUG_ON(iref);
4536                         ret = remove_extent_backref(trans, extent_root, path,
4537                                                     NULL, refs_to_drop,
4538                                                     is_data);
4539                         BUG_ON(ret);
4540                         btrfs_release_path(path);
4541                         path->leave_spinning = 1;
4542
4543                         key.objectid = bytenr;
4544                         key.type = BTRFS_EXTENT_ITEM_KEY;
4545                         key.offset = num_bytes;
4546
4547                         ret = btrfs_search_slot(trans, extent_root,
4548                                                 &key, path, -1, 1);
4549                         if (ret) {
4550                                 printk(KERN_ERR "umm, got %d back from search"
4551                                        ", was looking for %llu\n", ret,
4552                                        (unsigned long long)bytenr);
4553                                 if (ret > 0)
4554                                         btrfs_print_leaf(extent_root,
4555                                                          path->nodes[0]);
4556                         }
4557                         BUG_ON(ret);
4558                         extent_slot = path->slots[0];
4559                 }
4560         } else {
4561                 btrfs_print_leaf(extent_root, path->nodes[0]);
4562                 WARN_ON(1);
4563                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4564                        "parent %llu root %llu  owner %llu offset %llu\n",
4565                        (unsigned long long)bytenr,
4566                        (unsigned long long)parent,
4567                        (unsigned long long)root_objectid,
4568                        (unsigned long long)owner_objectid,
4569                        (unsigned long long)owner_offset);
4570         }
4571
4572         leaf = path->nodes[0];
4573         item_size = btrfs_item_size_nr(leaf, extent_slot);
4574 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4575         if (item_size < sizeof(*ei)) {
4576                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4577                 ret = convert_extent_item_v0(trans, extent_root, path,
4578                                              owner_objectid, 0);
4579                 BUG_ON(ret < 0);
4580
4581                 btrfs_release_path(path);
4582                 path->leave_spinning = 1;
4583
4584                 key.objectid = bytenr;
4585                 key.type = BTRFS_EXTENT_ITEM_KEY;
4586                 key.offset = num_bytes;
4587
4588                 ret = btrfs_search_slot(trans, extent_root, &key, path,
4589                                         -1, 1);
4590                 if (ret) {
4591                         printk(KERN_ERR "umm, got %d back from search"
4592                                ", was looking for %llu\n", ret,
4593                                (unsigned long long)bytenr);
4594                         btrfs_print_leaf(extent_root, path->nodes[0]);
4595                 }
4596                 BUG_ON(ret);
4597                 extent_slot = path->slots[0];
4598                 leaf = path->nodes[0];
4599                 item_size = btrfs_item_size_nr(leaf, extent_slot);
4600         }
4601 #endif
4602         BUG_ON(item_size < sizeof(*ei));
4603         ei = btrfs_item_ptr(leaf, extent_slot,
4604                             struct btrfs_extent_item);
4605         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
4606                 struct btrfs_tree_block_info *bi;
4607                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4608                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4609                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4610         }
4611
4612         refs = btrfs_extent_refs(leaf, ei);
4613         BUG_ON(refs < refs_to_drop);
4614         refs -= refs_to_drop;
4615
4616         if (refs > 0) {
4617                 if (extent_op)
4618                         __run_delayed_extent_op(extent_op, leaf, ei);
4619                 /*
4620                  * In the case of inline back ref, reference count will
4621                  * be updated by remove_extent_backref
4622                  */
4623                 if (iref) {
4624                         BUG_ON(!found_extent);
4625                 } else {
4626                         btrfs_set_extent_refs(leaf, ei, refs);
4627                         btrfs_mark_buffer_dirty(leaf);
4628                 }
4629                 if (found_extent) {
4630                         ret = remove_extent_backref(trans, extent_root, path,
4631                                                     iref, refs_to_drop,
4632                                                     is_data);
4633                         BUG_ON(ret);
4634                 }
4635         } else {
4636                 if (found_extent) {
4637                         BUG_ON(is_data && refs_to_drop !=
4638                                extent_data_ref_count(root, path, iref));
4639                         if (iref) {
4640                                 BUG_ON(path->slots[0] != extent_slot);
4641                         } else {
4642                                 BUG_ON(path->slots[0] != extent_slot + 1);
4643                                 path->slots[0] = extent_slot;
4644                                 num_to_del = 2;
4645                         }
4646                 }
4647
4648                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
4649                                       num_to_del);
4650                 BUG_ON(ret);
4651                 btrfs_release_path(path);
4652
4653                 if (is_data) {
4654                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
4655                         BUG_ON(ret);
4656                 } else {
4657                         invalidate_mapping_pages(info->btree_inode->i_mapping,
4658                              bytenr >> PAGE_CACHE_SHIFT,
4659                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
4660                 }
4661
4662                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
4663                 BUG_ON(ret);
4664         }
4665         btrfs_free_path(path);
4666         return ret;
4667 }
4668
4669 /*
4670  * when we free an block, it is possible (and likely) that we free the last
4671  * delayed ref for that extent as well.  This searches the delayed ref tree for
4672  * a given extent, and if there are no other delayed refs to be processed, it
4673  * removes it from the tree.
4674  */
4675 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
4676                                       struct btrfs_root *root, u64 bytenr)
4677 {
4678         struct btrfs_delayed_ref_head *head;
4679         struct btrfs_delayed_ref_root *delayed_refs;
4680         struct btrfs_delayed_ref_node *ref;
4681         struct rb_node *node;
4682         int ret = 0;
4683
4684         delayed_refs = &trans->transaction->delayed_refs;
4685         spin_lock(&delayed_refs->lock);
4686         head = btrfs_find_delayed_ref_head(trans, bytenr);
4687         if (!head)
4688                 goto out;
4689
4690         node = rb_prev(&head->node.rb_node);
4691         if (!node)
4692                 goto out;
4693
4694         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
4695
4696         /* there are still entries for this ref, we can't drop it */
4697         if (ref->bytenr == bytenr)
4698                 goto out;
4699
4700         if (head->extent_op) {
4701                 if (!head->must_insert_reserved)
4702                         goto out;
4703                 kfree(head->extent_op);
4704                 head->extent_op = NULL;
4705         }
4706
4707         /*
4708          * waiting for the lock here would deadlock.  If someone else has it
4709          * locked they are already in the process of dropping it anyway
4710          */
4711         if (!mutex_trylock(&head->mutex))
4712                 goto out;
4713
4714         /*
4715          * at this point we have a head with no other entries.  Go
4716          * ahead and process it.
4717          */
4718         head->node.in_tree = 0;
4719         rb_erase(&head->node.rb_node, &delayed_refs->root);
4720
4721         delayed_refs->num_entries--;
4722
4723         /*
4724          * we don't take a ref on the node because we're removing it from the
4725          * tree, so we just steal the ref the tree was holding.
4726          */
4727         delayed_refs->num_heads--;
4728         if (list_empty(&head->cluster))
4729                 delayed_refs->num_heads_ready--;
4730
4731         list_del_init(&head->cluster);
4732         spin_unlock(&delayed_refs->lock);
4733
4734         BUG_ON(head->extent_op);
4735         if (head->must_insert_reserved)
4736                 ret = 1;
4737
4738         mutex_unlock(&head->mutex);
4739         btrfs_put_delayed_ref(&head->node);
4740         return ret;
4741 out:
4742         spin_unlock(&delayed_refs->lock);
4743         return 0;
4744 }
4745
4746 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
4747                            struct btrfs_root *root,
4748                            struct extent_buffer *buf,
4749                            u64 parent, int last_ref)
4750 {
4751         struct btrfs_block_group_cache *cache = NULL;
4752         int ret;
4753
4754         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4755                 ret = btrfs_add_delayed_tree_ref(trans, buf->start, buf->len,
4756                                                 parent, root->root_key.objectid,
4757                                                 btrfs_header_level(buf),
4758                                                 BTRFS_DROP_DELAYED_REF, NULL);
4759                 BUG_ON(ret);
4760         }
4761
4762         if (!last_ref)
4763                 return;
4764
4765         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
4766
4767         if (btrfs_header_generation(buf) == trans->transid) {
4768                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4769                         ret = check_ref_cleanup(trans, root, buf->start);
4770                         if (!ret)
4771                                 goto out;
4772                 }
4773
4774                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
4775                         pin_down_extent(root, cache, buf->start, buf->len, 1);
4776                         goto out;
4777                 }
4778
4779                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
4780
4781                 btrfs_add_free_space(cache, buf->start, buf->len);
4782                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
4783         }
4784 out:
4785         /*
4786          * Deleting the buffer, clear the corrupt flag since it doesn't matter
4787          * anymore.
4788          */
4789         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
4790         btrfs_put_block_group(cache);
4791 }
4792
4793 int btrfs_free_extent(struct btrfs_trans_handle *trans,
4794                       struct btrfs_root *root,
4795                       u64 bytenr, u64 num_bytes, u64 parent,
4796                       u64 root_objectid, u64 owner, u64 offset)
4797 {
4798         int ret;
4799
4800         /*
4801          * tree log blocks never actually go into the extent allocation
4802          * tree, just update pinning info and exit early.
4803          */
4804         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
4805                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
4806                 /* unlocks the pinned mutex */
4807                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
4808                 ret = 0;
4809         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
4810                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
4811                                         parent, root_objectid, (int)owner,
4812                                         BTRFS_DROP_DELAYED_REF, NULL);
4813                 BUG_ON(ret);
4814         } else {
4815                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
4816                                         parent, root_objectid, owner,
4817                                         offset, BTRFS_DROP_DELAYED_REF, NULL);
4818                 BUG_ON(ret);
4819         }
4820         return ret;
4821 }
4822
4823 static u64 stripe_align(struct btrfs_root *root, u64 val)
4824 {
4825         u64 mask = ((u64)root->stripesize - 1);
4826         u64 ret = (val + mask) & ~mask;
4827         return ret;
4828 }
4829
4830 /*
4831  * when we wait for progress in the block group caching, its because
4832  * our allocation attempt failed at least once.  So, we must sleep
4833  * and let some progress happen before we try again.
4834  *
4835  * This function will sleep at least once waiting for new free space to
4836  * show up, and then it will check the block group free space numbers
4837  * for our min num_bytes.  Another option is to have it go ahead
4838  * and look in the rbtree for a free extent of a given size, but this
4839  * is a good start.
4840  */
4841 static noinline int
4842 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
4843                                 u64 num_bytes)
4844 {
4845         struct btrfs_caching_control *caching_ctl;
4846         DEFINE_WAIT(wait);
4847
4848         caching_ctl = get_caching_control(cache);
4849         if (!caching_ctl)
4850                 return 0;
4851
4852         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
4853                    (cache->free_space_ctl->free_space >= num_bytes));
4854
4855         put_caching_control(caching_ctl);
4856         return 0;
4857 }
4858
4859 static noinline int
4860 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
4861 {
4862         struct btrfs_caching_control *caching_ctl;
4863         DEFINE_WAIT(wait);
4864
4865         caching_ctl = get_caching_control(cache);
4866         if (!caching_ctl)
4867                 return 0;
4868
4869         wait_event(caching_ctl->wait, block_group_cache_done(cache));
4870
4871         put_caching_control(caching_ctl);
4872         return 0;
4873 }
4874
4875 static int get_block_group_index(struct btrfs_block_group_cache *cache)
4876 {
4877         int index;
4878         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
4879                 index = 0;
4880         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
4881                 index = 1;
4882         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
4883                 index = 2;
4884         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
4885                 index = 3;
4886         else
4887                 index = 4;
4888         return index;
4889 }
4890
4891 enum btrfs_loop_type {
4892         LOOP_FIND_IDEAL = 0,
4893         LOOP_CACHING_NOWAIT = 1,
4894         LOOP_CACHING_WAIT = 2,
4895         LOOP_ALLOC_CHUNK = 3,
4896         LOOP_NO_EMPTY_SIZE = 4,
4897 };
4898
4899 /*
4900  * walks the btree of allocated extents and find a hole of a given size.
4901  * The key ins is changed to record the hole:
4902  * ins->objectid == block start
4903  * ins->flags = BTRFS_EXTENT_ITEM_KEY
4904  * ins->offset == number of blocks
4905  * Any available blocks before search_start are skipped.
4906  */
4907 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
4908                                      struct btrfs_root *orig_root,
4909                                      u64 num_bytes, u64 empty_size,
4910                                      u64 search_start, u64 search_end,
4911                                      u64 hint_byte, struct btrfs_key *ins,
4912                                      u64 data)
4913 {
4914         int ret = 0;
4915         struct btrfs_root *root = orig_root->fs_info->extent_root;
4916         struct btrfs_free_cluster *last_ptr = NULL;
4917         struct btrfs_block_group_cache *block_group = NULL;
4918         int empty_cluster = 2 * 1024 * 1024;
4919         int allowed_chunk_alloc = 0;
4920         int done_chunk_alloc = 0;
4921         struct btrfs_space_info *space_info;
4922         int last_ptr_loop = 0;
4923         int loop = 0;
4924         int index = 0;
4925         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
4926                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
4927         bool found_uncached_bg = false;
4928         bool failed_cluster_refill = false;
4929         bool failed_alloc = false;
4930         bool use_cluster = true;
4931         u64 ideal_cache_percent = 0;
4932         u64 ideal_cache_offset = 0;
4933
4934         WARN_ON(num_bytes < root->sectorsize);
4935         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
4936         ins->objectid = 0;
4937         ins->offset = 0;
4938
4939         space_info = __find_space_info(root->fs_info, data);
4940         if (!space_info) {
4941                 printk(KERN_ERR "No space info for %llu\n", data);
4942                 return -ENOSPC;
4943         }
4944
4945         /*
4946          * If the space info is for both data and metadata it means we have a
4947          * small filesystem and we can't use the clustering stuff.
4948          */
4949         if (btrfs_mixed_space_info(space_info))
4950                 use_cluster = false;
4951
4952         if (orig_root->ref_cows || empty_size)
4953                 allowed_chunk_alloc = 1;
4954
4955         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
4956                 last_ptr = &root->fs_info->meta_alloc_cluster;
4957                 if (!btrfs_test_opt(root, SSD))
4958                         empty_cluster = 64 * 1024;
4959         }
4960
4961         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
4962             btrfs_test_opt(root, SSD)) {
4963                 last_ptr = &root->fs_info->data_alloc_cluster;
4964         }
4965
4966         if (last_ptr) {
4967                 spin_lock(&last_ptr->lock);
4968                 if (last_ptr->block_group)
4969                         hint_byte = last_ptr->window_start;
4970                 spin_unlock(&last_ptr->lock);
4971         }
4972
4973         search_start = max(search_start, first_logical_byte(root, 0));
4974         search_start = max(search_start, hint_byte);
4975
4976         if (!last_ptr)
4977                 empty_cluster = 0;
4978
4979         if (search_start == hint_byte) {
4980 ideal_cache:
4981                 block_group = btrfs_lookup_block_group(root->fs_info,
4982                                                        search_start);
4983                 /*
4984                  * we don't want to use the block group if it doesn't match our
4985                  * allocation bits, or if its not cached.
4986                  *
4987                  * However if we are re-searching with an ideal block group
4988                  * picked out then we don't care that the block group is cached.
4989                  */
4990                 if (block_group && block_group_bits(block_group, data) &&
4991                     (block_group->cached != BTRFS_CACHE_NO ||
4992                      search_start == ideal_cache_offset)) {
4993                         down_read(&space_info->groups_sem);
4994                         if (list_empty(&block_group->list) ||
4995                             block_group->ro) {
4996                                 /*
4997                                  * someone is removing this block group,
4998                                  * we can't jump into the have_block_group
4999                                  * target because our list pointers are not
5000                                  * valid
5001                                  */
5002                                 btrfs_put_block_group(block_group);
5003                                 up_read(&space_info->groups_sem);
5004                         } else {
5005                                 index = get_block_group_index(block_group);
5006                                 goto have_block_group;
5007                         }
5008                 } else if (block_group) {
5009                         btrfs_put_block_group(block_group);
5010                 }
5011         }
5012 search:
5013         down_read(&space_info->groups_sem);
5014         list_for_each_entry(block_group, &space_info->block_groups[index],
5015                             list) {
5016                 u64 offset;
5017                 int cached;
5018
5019                 btrfs_get_block_group(block_group);
5020                 search_start = block_group->key.objectid;
5021
5022                 /*
5023                  * this can happen if we end up cycling through all the
5024                  * raid types, but we want to make sure we only allocate
5025                  * for the proper type.
5026                  */
5027                 if (!block_group_bits(block_group, data)) {
5028                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5029                                 BTRFS_BLOCK_GROUP_RAID1 |
5030                                 BTRFS_BLOCK_GROUP_RAID10;
5031
5032                         /*
5033                          * if they asked for extra copies and this block group
5034                          * doesn't provide them, bail.  This does allow us to
5035                          * fill raid0 from raid1.
5036                          */
5037                         if ((data & extra) && !(block_group->flags & extra))
5038                                 goto loop;
5039                 }
5040
5041 have_block_group:
5042                 if (unlikely(block_group->cached == BTRFS_CACHE_NO)) {
5043                         u64 free_percent;
5044
5045                         ret = cache_block_group(block_group, trans,
5046                                                 orig_root, 1);
5047                         if (block_group->cached == BTRFS_CACHE_FINISHED)
5048                                 goto have_block_group;
5049
5050                         free_percent = btrfs_block_group_used(&block_group->item);
5051                         free_percent *= 100;
5052                         free_percent = div64_u64(free_percent,
5053                                                  block_group->key.offset);
5054                         free_percent = 100 - free_percent;
5055                         if (free_percent > ideal_cache_percent &&
5056                             likely(!block_group->ro)) {
5057                                 ideal_cache_offset = block_group->key.objectid;
5058                                 ideal_cache_percent = free_percent;
5059                         }
5060
5061                         /*
5062                          * The caching workers are limited to 2 threads, so we
5063                          * can queue as much work as we care to.
5064                          */
5065                         if (loop > LOOP_FIND_IDEAL) {
5066                                 ret = cache_block_group(block_group, trans,
5067                                                         orig_root, 0);
5068                                 BUG_ON(ret);
5069                         }
5070                         found_uncached_bg = true;
5071
5072                         /*
5073                          * If loop is set for cached only, try the next block
5074                          * group.
5075                          */
5076                         if (loop == LOOP_FIND_IDEAL)
5077                                 goto loop;
5078                 }
5079
5080                 cached = block_group_cache_done(block_group);
5081                 if (unlikely(!cached))
5082                         found_uncached_bg = true;
5083
5084                 if (unlikely(block_group->ro))
5085                         goto loop;
5086
5087                 spin_lock(&block_group->free_space_ctl->tree_lock);
5088                 if (cached &&
5089                     block_group->free_space_ctl->free_space <
5090                     num_bytes + empty_size) {
5091                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5092                         goto loop;
5093                 }
5094                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5095
5096                 /*
5097                  * Ok we want to try and use the cluster allocator, so lets look
5098                  * there, unless we are on LOOP_NO_EMPTY_SIZE, since we will
5099                  * have tried the cluster allocator plenty of times at this
5100                  * point and not have found anything, so we are likely way too
5101                  * fragmented for the clustering stuff to find anything, so lets
5102                  * just skip it and let the allocator find whatever block it can
5103                  * find
5104                  */
5105                 if (last_ptr && loop < LOOP_NO_EMPTY_SIZE) {
5106                         /*
5107                          * the refill lock keeps out other
5108                          * people trying to start a new cluster
5109                          */
5110                         spin_lock(&last_ptr->refill_lock);
5111                         if (last_ptr->block_group &&
5112                             (last_ptr->block_group->ro ||
5113                             !block_group_bits(last_ptr->block_group, data))) {
5114                                 offset = 0;
5115                                 goto refill_cluster;
5116                         }
5117
5118                         offset = btrfs_alloc_from_cluster(block_group, last_ptr,
5119                                                  num_bytes, search_start);
5120                         if (offset) {
5121                                 /* we have a block, we're done */
5122                                 spin_unlock(&last_ptr->refill_lock);
5123                                 goto checks;
5124                         }
5125
5126                         spin_lock(&last_ptr->lock);
5127                         /*
5128                          * whoops, this cluster doesn't actually point to
5129                          * this block group.  Get a ref on the block
5130                          * group is does point to and try again
5131                          */
5132                         if (!last_ptr_loop && last_ptr->block_group &&
5133                             last_ptr->block_group != block_group &&
5134                             index <=
5135                                  get_block_group_index(last_ptr->block_group)) {
5136
5137                                 btrfs_put_block_group(block_group);
5138                                 block_group = last_ptr->block_group;
5139                                 btrfs_get_block_group(block_group);
5140                                 spin_unlock(&last_ptr->lock);
5141                                 spin_unlock(&last_ptr->refill_lock);
5142
5143                                 last_ptr_loop = 1;
5144                                 search_start = block_group->key.objectid;
5145                                 /*
5146                                  * we know this block group is properly
5147                                  * in the list because
5148                                  * btrfs_remove_block_group, drops the
5149                                  * cluster before it removes the block
5150                                  * group from the list
5151                                  */
5152                                 goto have_block_group;
5153                         }
5154                         spin_unlock(&last_ptr->lock);
5155 refill_cluster:
5156                         /*
5157                          * this cluster didn't work out, free it and
5158                          * start over
5159                          */
5160                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5161
5162                         last_ptr_loop = 0;
5163
5164                         /* allocate a cluster in this block group */
5165                         ret = btrfs_find_space_cluster(trans, root,
5166                                                block_group, last_ptr,
5167                                                offset, num_bytes,
5168                                                empty_cluster + empty_size);
5169                         if (ret == 0) {
5170                                 /*
5171                                  * now pull our allocation out of this
5172                                  * cluster
5173                                  */
5174                                 offset = btrfs_alloc_from_cluster(block_group,
5175                                                   last_ptr, num_bytes,
5176                                                   search_start);
5177                                 if (offset) {
5178                                         /* we found one, proceed */
5179                                         spin_unlock(&last_ptr->refill_lock);
5180                                         goto checks;
5181                                 }
5182                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5183                                    && !failed_cluster_refill) {
5184                                 spin_unlock(&last_ptr->refill_lock);
5185
5186                                 failed_cluster_refill = true;
5187                                 wait_block_group_cache_progress(block_group,
5188                                        num_bytes + empty_cluster + empty_size);
5189                                 goto have_block_group;
5190                         }
5191
5192                         /*
5193                          * at this point we either didn't find a cluster
5194                          * or we weren't able to allocate a block from our
5195                          * cluster.  Free the cluster we've been trying
5196                          * to use, and go to the next block group
5197                          */
5198                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5199                         spin_unlock(&last_ptr->refill_lock);
5200                         goto loop;
5201                 }
5202
5203                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5204                                                     num_bytes, empty_size);
5205                 /*
5206                  * If we didn't find a chunk, and we haven't failed on this
5207                  * block group before, and this block group is in the middle of
5208                  * caching and we are ok with waiting, then go ahead and wait
5209                  * for progress to be made, and set failed_alloc to true.
5210                  *
5211                  * If failed_alloc is true then we've already waited on this
5212                  * block group once and should move on to the next block group.
5213                  */
5214                 if (!offset && !failed_alloc && !cached &&
5215                     loop > LOOP_CACHING_NOWAIT) {
5216                         wait_block_group_cache_progress(block_group,
5217                                                 num_bytes + empty_size);
5218                         failed_alloc = true;
5219                         goto have_block_group;
5220                 } else if (!offset) {
5221                         goto loop;
5222                 }
5223 checks:
5224                 search_start = stripe_align(root, offset);
5225                 /* move on to the next group */
5226                 if (search_start + num_bytes >= search_end) {
5227                         btrfs_add_free_space(block_group, offset, num_bytes);
5228                         goto loop;
5229                 }
5230
5231                 /* move on to the next group */
5232                 if (search_start + num_bytes >
5233                     block_group->key.objectid + block_group->key.offset) {
5234                         btrfs_add_free_space(block_group, offset, num_bytes);
5235                         goto loop;
5236                 }
5237
5238                 ins->objectid = search_start;
5239                 ins->offset = num_bytes;
5240
5241                 if (offset < search_start)
5242                         btrfs_add_free_space(block_group, offset,
5243                                              search_start - offset);
5244                 BUG_ON(offset > search_start);
5245
5246                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
5247                                                   alloc_type);
5248                 if (ret == -EAGAIN) {
5249                         btrfs_add_free_space(block_group, offset, num_bytes);
5250                         goto loop;
5251                 }
5252
5253                 /* we are all good, lets return */
5254                 ins->objectid = search_start;
5255                 ins->offset = num_bytes;
5256
5257                 if (offset < search_start)
5258                         btrfs_add_free_space(block_group, offset,
5259                                              search_start - offset);
5260                 BUG_ON(offset > search_start);
5261                 btrfs_put_block_group(block_group);
5262                 break;
5263 loop:
5264                 failed_cluster_refill = false;
5265                 failed_alloc = false;
5266                 BUG_ON(index != get_block_group_index(block_group));
5267                 btrfs_put_block_group(block_group);
5268         }
5269         up_read(&space_info->groups_sem);
5270
5271         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5272                 goto search;
5273
5274         /* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
5275          *                      for them to make caching progress.  Also
5276          *                      determine the best possible bg to cache
5277          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5278          *                      caching kthreads as we move along
5279          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5280          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5281          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5282          *                      again
5283          */
5284         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5285                 index = 0;
5286                 if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
5287                         found_uncached_bg = false;
5288                         loop++;
5289                         if (!ideal_cache_percent)
5290                                 goto search;
5291
5292                         /*
5293                          * 1 of the following 2 things have happened so far
5294                          *
5295                          * 1) We found an ideal block group for caching that
5296                          * is mostly full and will cache quickly, so we might
5297                          * as well wait for it.
5298                          *
5299                          * 2) We searched for cached only and we didn't find
5300                          * anything, and we didn't start any caching kthreads
5301                          * either, so chances are we will loop through and
5302                          * start a couple caching kthreads, and then come back
5303                          * around and just wait for them.  This will be slower
5304                          * because we will have 2 caching kthreads reading at
5305                          * the same time when we could have just started one
5306                          * and waited for it to get far enough to give us an
5307                          * allocation, so go ahead and go to the wait caching
5308                          * loop.
5309                          */
5310                         loop = LOOP_CACHING_WAIT;
5311                         search_start = ideal_cache_offset;
5312                         ideal_cache_percent = 0;
5313                         goto ideal_cache;
5314                 } else if (loop == LOOP_FIND_IDEAL) {
5315                         /*
5316                          * Didn't find a uncached bg, wait on anything we find
5317                          * next.
5318                          */
5319                         loop = LOOP_CACHING_WAIT;
5320                         goto search;
5321                 }
5322
5323                 loop++;
5324
5325                 if (loop == LOOP_ALLOC_CHUNK) {
5326                        if (allowed_chunk_alloc) {
5327                                 ret = do_chunk_alloc(trans, root, num_bytes +
5328                                                      2 * 1024 * 1024, data,
5329                                                      CHUNK_ALLOC_LIMITED);
5330                                 allowed_chunk_alloc = 0;
5331                                 if (ret == 1)
5332                                         done_chunk_alloc = 1;
5333                         } else if (!done_chunk_alloc &&
5334                                    space_info->force_alloc ==
5335                                    CHUNK_ALLOC_NO_FORCE) {
5336                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5337                         }
5338
5339                        /*
5340                         * We didn't allocate a chunk, go ahead and drop the
5341                         * empty size and loop again.
5342                         */
5343                        if (!done_chunk_alloc)
5344                                loop = LOOP_NO_EMPTY_SIZE;
5345                 }
5346
5347                 if (loop == LOOP_NO_EMPTY_SIZE) {
5348                         empty_size = 0;
5349                         empty_cluster = 0;
5350                 }
5351
5352                 goto search;
5353         } else if (!ins->objectid) {
5354                 ret = -ENOSPC;
5355         } else if (ins->objectid) {
5356                 ret = 0;
5357         }
5358
5359         return ret;
5360 }
5361
5362 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5363                             int dump_block_groups)
5364 {
5365         struct btrfs_block_group_cache *cache;
5366         int index = 0;
5367
5368         spin_lock(&info->lock);
5369         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5370                (unsigned long long)info->flags,
5371                (unsigned long long)(info->total_bytes - info->bytes_used -
5372                                     info->bytes_pinned - info->bytes_reserved -
5373                                     info->bytes_readonly),
5374                (info->full) ? "" : "not ");
5375         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5376                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5377                (unsigned long long)info->total_bytes,
5378                (unsigned long long)info->bytes_used,
5379                (unsigned long long)info->bytes_pinned,
5380                (unsigned long long)info->bytes_reserved,
5381                (unsigned long long)info->bytes_may_use,
5382                (unsigned long long)info->bytes_readonly);
5383         spin_unlock(&info->lock);
5384
5385         if (!dump_block_groups)
5386                 return;
5387
5388         down_read(&info->groups_sem);
5389 again:
5390         list_for_each_entry(cache, &info->block_groups[index], list) {
5391                 spin_lock(&cache->lock);
5392                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5393                        "%llu pinned %llu reserved\n",
5394                        (unsigned long long)cache->key.objectid,
5395                        (unsigned long long)cache->key.offset,
5396                        (unsigned long long)btrfs_block_group_used(&cache->item),
5397                        (unsigned long long)cache->pinned,
5398                        (unsigned long long)cache->reserved);
5399                 btrfs_dump_free_space(cache, bytes);
5400                 spin_unlock(&cache->lock);
5401         }
5402         if (++index < BTRFS_NR_RAID_TYPES)
5403                 goto again;
5404         up_read(&info->groups_sem);
5405 }
5406
5407 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5408                          struct btrfs_root *root,
5409                          u64 num_bytes, u64 min_alloc_size,
5410                          u64 empty_size, u64 hint_byte,
5411                          u64 search_end, struct btrfs_key *ins,
5412                          u64 data)
5413 {
5414         int ret;
5415         u64 search_start = 0;
5416
5417         data = btrfs_get_alloc_profile(root, data);
5418 again:
5419         /*
5420          * the only place that sets empty_size is btrfs_realloc_node, which
5421          * is not called recursively on allocations
5422          */
5423         if (empty_size || root->ref_cows)
5424                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5425                                      num_bytes + 2 * 1024 * 1024, data,
5426                                      CHUNK_ALLOC_NO_FORCE);
5427
5428         WARN_ON(num_bytes < root->sectorsize);
5429         ret = find_free_extent(trans, root, num_bytes, empty_size,
5430                                search_start, search_end, hint_byte,
5431                                ins, data);
5432
5433         if (ret == -ENOSPC && num_bytes > min_alloc_size) {
5434                 num_bytes = num_bytes >> 1;
5435                 num_bytes = num_bytes & ~(root->sectorsize - 1);
5436                 num_bytes = max(num_bytes, min_alloc_size);
5437                 do_chunk_alloc(trans, root->fs_info->extent_root,
5438                                num_bytes, data, CHUNK_ALLOC_FORCE);
5439                 goto again;
5440         }
5441         if (ret == -ENOSPC && btrfs_test_opt(root, ENOSPC_DEBUG)) {
5442                 struct btrfs_space_info *sinfo;
5443
5444                 sinfo = __find_space_info(root->fs_info, data);
5445                 printk(KERN_ERR "btrfs allocation failed flags %llu, "
5446                        "wanted %llu\n", (unsigned long long)data,
5447                        (unsigned long long)num_bytes);
5448                 dump_space_info(sinfo, num_bytes, 1);
5449         }
5450
5451         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
5452
5453         return ret;
5454 }
5455
5456 int btrfs_free_reserved_extent(struct btrfs_root *root, u64 start, u64 len)
5457 {
5458         struct btrfs_block_group_cache *cache;
5459         int ret = 0;
5460
5461         cache = btrfs_lookup_block_group(root->fs_info, start);
5462         if (!cache) {
5463                 printk(KERN_ERR "Unable to find block group for %llu\n",
5464                        (unsigned long long)start);
5465                 return -ENOSPC;
5466         }
5467
5468         if (btrfs_test_opt(root, DISCARD))
5469                 ret = btrfs_discard_extent(root, start, len, NULL);
5470
5471         btrfs_add_free_space(cache, start, len);
5472         btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
5473         btrfs_put_block_group(cache);
5474
5475         trace_btrfs_reserved_extent_free(root, start, len);
5476
5477         return ret;
5478 }
5479
5480 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5481                                       struct btrfs_root *root,
5482                                       u64 parent, u64 root_objectid,
5483                                       u64 flags, u64 owner, u64 offset,
5484                                       struct btrfs_key *ins, int ref_mod)
5485 {
5486         int ret;
5487         struct btrfs_fs_info *fs_info = root->fs_info;
5488         struct btrfs_extent_item *extent_item;
5489         struct btrfs_extent_inline_ref *iref;
5490         struct btrfs_path *path;
5491         struct extent_buffer *leaf;
5492         int type;
5493         u32 size;
5494
5495         if (parent > 0)
5496                 type = BTRFS_SHARED_DATA_REF_KEY;
5497         else
5498                 type = BTRFS_EXTENT_DATA_REF_KEY;
5499
5500         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5501
5502         path = btrfs_alloc_path();
5503         if (!path)
5504                 return -ENOMEM;
5505
5506         path->leave_spinning = 1;
5507         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5508                                       ins, size);
5509         BUG_ON(ret);
5510
5511         leaf = path->nodes[0];
5512         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5513                                      struct btrfs_extent_item);
5514         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
5515         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5516         btrfs_set_extent_flags(leaf, extent_item,
5517                                flags | BTRFS_EXTENT_FLAG_DATA);
5518
5519         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
5520         btrfs_set_extent_inline_ref_type(leaf, iref, type);
5521         if (parent > 0) {
5522                 struct btrfs_shared_data_ref *ref;
5523                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
5524                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5525                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
5526         } else {
5527                 struct btrfs_extent_data_ref *ref;
5528                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
5529                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
5530                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
5531                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
5532                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
5533         }
5534
5535         btrfs_mark_buffer_dirty(path->nodes[0]);
5536         btrfs_free_path(path);
5537
5538         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5539         if (ret) {
5540                 printk(KERN_ERR "btrfs update block group failed for %llu "
5541                        "%llu\n", (unsigned long long)ins->objectid,
5542                        (unsigned long long)ins->offset);
5543                 BUG();
5544         }
5545         return ret;
5546 }
5547
5548 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
5549                                      struct btrfs_root *root,
5550                                      u64 parent, u64 root_objectid,
5551                                      u64 flags, struct btrfs_disk_key *key,
5552                                      int level, struct btrfs_key *ins)
5553 {
5554         int ret;
5555         struct btrfs_fs_info *fs_info = root->fs_info;
5556         struct btrfs_extent_item *extent_item;
5557         struct btrfs_tree_block_info *block_info;
5558         struct btrfs_extent_inline_ref *iref;
5559         struct btrfs_path *path;
5560         struct extent_buffer *leaf;
5561         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
5562
5563         path = btrfs_alloc_path();
5564         if (!path)
5565                 return -ENOMEM;
5566
5567         path->leave_spinning = 1;
5568         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5569                                       ins, size);
5570         BUG_ON(ret);
5571
5572         leaf = path->nodes[0];
5573         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5574                                      struct btrfs_extent_item);
5575         btrfs_set_extent_refs(leaf, extent_item, 1);
5576         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5577         btrfs_set_extent_flags(leaf, extent_item,
5578                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
5579         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
5580
5581         btrfs_set_tree_block_key(leaf, block_info, key);
5582         btrfs_set_tree_block_level(leaf, block_info, level);
5583
5584         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
5585         if (parent > 0) {
5586                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
5587                 btrfs_set_extent_inline_ref_type(leaf, iref,
5588                                                  BTRFS_SHARED_BLOCK_REF_KEY);
5589                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5590         } else {
5591                 btrfs_set_extent_inline_ref_type(leaf, iref,
5592                                                  BTRFS_TREE_BLOCK_REF_KEY);
5593                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
5594         }
5595
5596         btrfs_mark_buffer_dirty(leaf);
5597         btrfs_free_path(path);
5598
5599         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5600         if (ret) {
5601                 printk(KERN_ERR "btrfs update block group failed for %llu "
5602                        "%llu\n", (unsigned long long)ins->objectid,
5603                        (unsigned long long)ins->offset);
5604                 BUG();
5605         }
5606         return ret;
5607 }
5608
5609 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5610                                      struct btrfs_root *root,
5611                                      u64 root_objectid, u64 owner,
5612                                      u64 offset, struct btrfs_key *ins)
5613 {
5614         int ret;
5615
5616         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
5617
5618         ret = btrfs_add_delayed_data_ref(trans, ins->objectid, ins->offset,
5619                                          0, root_objectid, owner, offset,
5620                                          BTRFS_ADD_DELAYED_EXTENT, NULL);
5621         return ret;
5622 }
5623
5624 /*
5625  * this is used by the tree logging recovery code.  It records that
5626  * an extent has been allocated and makes sure to clear the free
5627  * space cache bits as well
5628  */
5629 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
5630                                    struct btrfs_root *root,
5631                                    u64 root_objectid, u64 owner, u64 offset,
5632                                    struct btrfs_key *ins)
5633 {
5634         int ret;
5635         struct btrfs_block_group_cache *block_group;
5636         struct btrfs_caching_control *caching_ctl;
5637         u64 start = ins->objectid;
5638         u64 num_bytes = ins->offset;
5639
5640         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
5641         cache_block_group(block_group, trans, NULL, 0);
5642         caching_ctl = get_caching_control(block_group);
5643
5644         if (!caching_ctl) {
5645                 BUG_ON(!block_group_cache_done(block_group));
5646                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5647                 BUG_ON(ret);
5648         } else {
5649                 mutex_lock(&caching_ctl->mutex);
5650
5651                 if (start >= caching_ctl->progress) {
5652                         ret = add_excluded_extent(root, start, num_bytes);
5653                         BUG_ON(ret);
5654                 } else if (start + num_bytes <= caching_ctl->progress) {
5655                         ret = btrfs_remove_free_space(block_group,
5656                                                       start, num_bytes);
5657                         BUG_ON(ret);
5658                 } else {
5659                         num_bytes = caching_ctl->progress - start;
5660                         ret = btrfs_remove_free_space(block_group,
5661                                                       start, num_bytes);
5662                         BUG_ON(ret);
5663
5664                         start = caching_ctl->progress;
5665                         num_bytes = ins->objectid + ins->offset -
5666                                     caching_ctl->progress;
5667                         ret = add_excluded_extent(root, start, num_bytes);
5668                         BUG_ON(ret);
5669                 }
5670
5671                 mutex_unlock(&caching_ctl->mutex);
5672                 put_caching_control(caching_ctl);
5673         }
5674
5675         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
5676                                           RESERVE_ALLOC_NO_ACCOUNT);
5677         BUG_ON(ret);
5678         btrfs_put_block_group(block_group);
5679         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
5680                                          0, owner, offset, ins, 1);
5681         return ret;
5682 }
5683
5684 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
5685                                             struct btrfs_root *root,
5686                                             u64 bytenr, u32 blocksize,
5687                                             int level)
5688 {
5689         struct extent_buffer *buf;
5690
5691         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
5692         if (!buf)
5693                 return ERR_PTR(-ENOMEM);
5694         btrfs_set_header_generation(buf, trans->transid);
5695         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
5696         btrfs_tree_lock(buf);
5697         clean_tree_block(trans, root, buf);
5698
5699         btrfs_set_lock_blocking(buf);
5700         btrfs_set_buffer_uptodate(buf);
5701
5702         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
5703                 /*
5704                  * we allow two log transactions at a time, use different
5705                  * EXENT bit to differentiate dirty pages.
5706                  */
5707                 if (root->log_transid % 2 == 0)
5708                         set_extent_dirty(&root->dirty_log_pages, buf->start,
5709                                         buf->start + buf->len - 1, GFP_NOFS);
5710                 else
5711                         set_extent_new(&root->dirty_log_pages, buf->start,
5712                                         buf->start + buf->len - 1, GFP_NOFS);
5713         } else {
5714                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
5715                          buf->start + buf->len - 1, GFP_NOFS);
5716         }
5717         trans->blocks_used++;
5718         /* this returns a buffer locked for blocking */
5719         return buf;
5720 }
5721
5722 static struct btrfs_block_rsv *
5723 use_block_rsv(struct btrfs_trans_handle *trans,
5724               struct btrfs_root *root, u32 blocksize)
5725 {
5726         struct btrfs_block_rsv *block_rsv;
5727         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5728         int ret;
5729
5730         block_rsv = get_block_rsv(trans, root);
5731
5732         if (block_rsv->size == 0) {
5733                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0, 0);
5734                 /*
5735                  * If we couldn't reserve metadata bytes try and use some from
5736                  * the global reserve.
5737                  */
5738                 if (ret && block_rsv != global_rsv) {
5739                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5740                         if (!ret)
5741                                 return global_rsv;
5742                         return ERR_PTR(ret);
5743                 } else if (ret) {
5744                         return ERR_PTR(ret);
5745                 }
5746                 return block_rsv;
5747         }
5748
5749         ret = block_rsv_use_bytes(block_rsv, blocksize);
5750         if (!ret)
5751                 return block_rsv;
5752         if (ret) {
5753                 WARN_ON(1);
5754                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0, 0);
5755                 if (!ret) {
5756                         return block_rsv;
5757                 } else if (ret && block_rsv != global_rsv) {
5758                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5759                         if (!ret)
5760                                 return global_rsv;
5761                 }
5762         }
5763
5764         return ERR_PTR(-ENOSPC);
5765 }
5766
5767 static void unuse_block_rsv(struct btrfs_block_rsv *block_rsv, u32 blocksize)
5768 {
5769         block_rsv_add_bytes(block_rsv, blocksize, 0);
5770         block_rsv_release_bytes(block_rsv, NULL, 0);
5771 }
5772
5773 /*
5774  * finds a free extent and does all the dirty work required for allocation
5775  * returns the key for the extent through ins, and a tree buffer for
5776  * the first block of the extent through buf.
5777  *
5778  * returns the tree buffer or NULL.
5779  */
5780 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
5781                                         struct btrfs_root *root, u32 blocksize,
5782                                         u64 parent, u64 root_objectid,
5783                                         struct btrfs_disk_key *key, int level,
5784                                         u64 hint, u64 empty_size)
5785 {
5786         struct btrfs_key ins;
5787         struct btrfs_block_rsv *block_rsv;
5788         struct extent_buffer *buf;
5789         u64 flags = 0;
5790         int ret;
5791
5792
5793         block_rsv = use_block_rsv(trans, root, blocksize);
5794         if (IS_ERR(block_rsv))
5795                 return ERR_CAST(block_rsv);
5796
5797         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
5798                                    empty_size, hint, (u64)-1, &ins, 0);
5799         if (ret) {
5800                 unuse_block_rsv(block_rsv, blocksize);
5801                 return ERR_PTR(ret);
5802         }
5803
5804         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
5805                                     blocksize, level);
5806         BUG_ON(IS_ERR(buf));
5807
5808         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
5809                 if (parent == 0)
5810                         parent = ins.objectid;
5811                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
5812         } else
5813                 BUG_ON(parent > 0);
5814
5815         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
5816                 struct btrfs_delayed_extent_op *extent_op;
5817                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
5818                 BUG_ON(!extent_op);
5819                 if (key)
5820                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
5821                 else
5822                         memset(&extent_op->key, 0, sizeof(extent_op->key));
5823                 extent_op->flags_to_set = flags;
5824                 extent_op->update_key = 1;
5825                 extent_op->update_flags = 1;
5826                 extent_op->is_data = 0;
5827
5828                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
5829                                         ins.offset, parent, root_objectid,
5830                                         level, BTRFS_ADD_DELAYED_EXTENT,
5831                                         extent_op);
5832                 BUG_ON(ret);
5833         }
5834         return buf;
5835 }
5836
5837 struct walk_control {
5838         u64 refs[BTRFS_MAX_LEVEL];
5839         u64 flags[BTRFS_MAX_LEVEL];
5840         struct btrfs_key update_progress;
5841         int stage;
5842         int level;
5843         int shared_level;
5844         int update_ref;
5845         int keep_locks;
5846         int reada_slot;
5847         int reada_count;
5848 };
5849
5850 #define DROP_REFERENCE  1
5851 #define UPDATE_BACKREF  2
5852
5853 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
5854                                      struct btrfs_root *root,
5855                                      struct walk_control *wc,
5856                                      struct btrfs_path *path)
5857 {
5858         u64 bytenr;
5859         u64 generation;
5860         u64 refs;
5861         u64 flags;
5862         u32 nritems;
5863         u32 blocksize;
5864         struct btrfs_key key;
5865         struct extent_buffer *eb;
5866         int ret;
5867         int slot;
5868         int nread = 0;
5869
5870         if (path->slots[wc->level] < wc->reada_slot) {
5871                 wc->reada_count = wc->reada_count * 2 / 3;
5872                 wc->reada_count = max(wc->reada_count, 2);
5873         } else {
5874                 wc->reada_count = wc->reada_count * 3 / 2;
5875                 wc->reada_count = min_t(int, wc->reada_count,
5876                                         BTRFS_NODEPTRS_PER_BLOCK(root));
5877         }
5878
5879         eb = path->nodes[wc->level];
5880         nritems = btrfs_header_nritems(eb);
5881         blocksize = btrfs_level_size(root, wc->level - 1);
5882
5883         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
5884                 if (nread >= wc->reada_count)
5885                         break;
5886
5887                 cond_resched();
5888                 bytenr = btrfs_node_blockptr(eb, slot);
5889                 generation = btrfs_node_ptr_generation(eb, slot);
5890
5891                 if (slot == path->slots[wc->level])
5892                         goto reada;
5893
5894                 if (wc->stage == UPDATE_BACKREF &&
5895                     generation <= root->root_key.offset)
5896                         continue;
5897
5898                 /* We don't lock the tree block, it's OK to be racy here */
5899                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
5900                                                &refs, &flags);
5901                 BUG_ON(ret);
5902                 BUG_ON(refs == 0);
5903
5904                 if (wc->stage == DROP_REFERENCE) {
5905                         if (refs == 1)
5906                                 goto reada;
5907
5908                         if (wc->level == 1 &&
5909                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5910                                 continue;
5911                         if (!wc->update_ref ||
5912                             generation <= root->root_key.offset)
5913                                 continue;
5914                         btrfs_node_key_to_cpu(eb, &key, slot);
5915                         ret = btrfs_comp_cpu_keys(&key,
5916                                                   &wc->update_progress);
5917                         if (ret < 0)
5918                                 continue;
5919                 } else {
5920                         if (wc->level == 1 &&
5921                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
5922                                 continue;
5923                 }
5924 reada:
5925                 ret = readahead_tree_block(root, bytenr, blocksize,
5926                                            generation);
5927                 if (ret)
5928                         break;
5929                 nread++;
5930         }
5931         wc->reada_slot = slot;
5932 }
5933
5934 /*
5935  * hepler to process tree block while walking down the tree.
5936  *
5937  * when wc->stage == UPDATE_BACKREF, this function updates
5938  * back refs for pointers in the block.
5939  *
5940  * NOTE: return value 1 means we should stop walking down.
5941  */
5942 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
5943                                    struct btrfs_root *root,
5944                                    struct btrfs_path *path,
5945                                    struct walk_control *wc, int lookup_info)
5946 {
5947         int level = wc->level;
5948         struct extent_buffer *eb = path->nodes[level];
5949         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
5950         int ret;
5951
5952         if (wc->stage == UPDATE_BACKREF &&
5953             btrfs_header_owner(eb) != root->root_key.objectid)
5954                 return 1;
5955
5956         /*
5957          * when reference count of tree block is 1, it won't increase
5958          * again. once full backref flag is set, we never clear it.
5959          */
5960         if (lookup_info &&
5961             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
5962              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
5963                 BUG_ON(!path->locks[level]);
5964                 ret = btrfs_lookup_extent_info(trans, root,
5965                                                eb->start, eb->len,
5966                                                &wc->refs[level],
5967                                                &wc->flags[level]);
5968                 BUG_ON(ret);
5969                 BUG_ON(wc->refs[level] == 0);
5970         }
5971
5972         if (wc->stage == DROP_REFERENCE) {
5973                 if (wc->refs[level] > 1)
5974                         return 1;
5975
5976                 if (path->locks[level] && !wc->keep_locks) {
5977                         btrfs_tree_unlock_rw(eb, path->locks[level]);
5978                         path->locks[level] = 0;
5979                 }
5980                 return 0;
5981         }
5982
5983         /* wc->stage == UPDATE_BACKREF */
5984         if (!(wc->flags[level] & flag)) {
5985                 BUG_ON(!path->locks[level]);
5986                 ret = btrfs_inc_ref(trans, root, eb, 1);
5987                 BUG_ON(ret);
5988                 ret = btrfs_dec_ref(trans, root, eb, 0);
5989                 BUG_ON(ret);
5990                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
5991                                                   eb->len, flag, 0);
5992                 BUG_ON(ret);
5993                 wc->flags[level] |= flag;
5994         }
5995
5996         /*
5997          * the block is shared by multiple trees, so it's not good to
5998          * keep the tree lock
5999          */
6000         if (path->locks[level] && level > 0) {
6001                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6002                 path->locks[level] = 0;
6003         }
6004         return 0;
6005 }
6006
6007 /*
6008  * hepler to process tree block pointer.
6009  *
6010  * when wc->stage == DROP_REFERENCE, this function checks
6011  * reference count of the block pointed to. if the block
6012  * is shared and we need update back refs for the subtree
6013  * rooted at the block, this function changes wc->stage to
6014  * UPDATE_BACKREF. if the block is shared and there is no
6015  * need to update back, this function drops the reference
6016  * to the block.
6017  *
6018  * NOTE: return value 1 means we should stop walking down.
6019  */
6020 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6021                                  struct btrfs_root *root,
6022                                  struct btrfs_path *path,
6023                                  struct walk_control *wc, int *lookup_info)
6024 {
6025         u64 bytenr;
6026         u64 generation;
6027         u64 parent;
6028         u32 blocksize;
6029         struct btrfs_key key;
6030         struct extent_buffer *next;
6031         int level = wc->level;
6032         int reada = 0;
6033         int ret = 0;
6034
6035         generation = btrfs_node_ptr_generation(path->nodes[level],
6036                                                path->slots[level]);
6037         /*
6038          * if the lower level block was created before the snapshot
6039          * was created, we know there is no need to update back refs
6040          * for the subtree
6041          */
6042         if (wc->stage == UPDATE_BACKREF &&
6043             generation <= root->root_key.offset) {
6044                 *lookup_info = 1;
6045                 return 1;
6046         }
6047
6048         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6049         blocksize = btrfs_level_size(root, level - 1);
6050
6051         next = btrfs_find_tree_block(root, bytenr, blocksize);
6052         if (!next) {
6053                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6054                 if (!next)
6055                         return -ENOMEM;
6056                 reada = 1;
6057         }
6058         btrfs_tree_lock(next);
6059         btrfs_set_lock_blocking(next);
6060
6061         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6062                                        &wc->refs[level - 1],
6063                                        &wc->flags[level - 1]);
6064         BUG_ON(ret);
6065         BUG_ON(wc->refs[level - 1] == 0);
6066         *lookup_info = 0;
6067
6068         if (wc->stage == DROP_REFERENCE) {
6069                 if (wc->refs[level - 1] > 1) {
6070                         if (level == 1 &&
6071                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6072                                 goto skip;
6073
6074                         if (!wc->update_ref ||
6075                             generation <= root->root_key.offset)
6076                                 goto skip;
6077
6078                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6079                                               path->slots[level]);
6080                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6081                         if (ret < 0)
6082                                 goto skip;
6083
6084                         wc->stage = UPDATE_BACKREF;
6085                         wc->shared_level = level - 1;
6086                 }
6087         } else {
6088                 if (level == 1 &&
6089                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6090                         goto skip;
6091         }
6092
6093         if (!btrfs_buffer_uptodate(next, generation)) {
6094                 btrfs_tree_unlock(next);
6095                 free_extent_buffer(next);
6096                 next = NULL;
6097                 *lookup_info = 1;
6098         }
6099
6100         if (!next) {
6101                 if (reada && level == 1)
6102                         reada_walk_down(trans, root, wc, path);
6103                 next = read_tree_block(root, bytenr, blocksize, generation);
6104                 if (!next)
6105                         return -EIO;
6106                 btrfs_tree_lock(next);
6107                 btrfs_set_lock_blocking(next);
6108         }
6109
6110         level--;
6111         BUG_ON(level != btrfs_header_level(next));
6112         path->nodes[level] = next;
6113         path->slots[level] = 0;
6114         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6115         wc->level = level;
6116         if (wc->level == 1)
6117                 wc->reada_slot = 0;
6118         return 0;
6119 skip:
6120         wc->refs[level - 1] = 0;
6121         wc->flags[level - 1] = 0;
6122         if (wc->stage == DROP_REFERENCE) {
6123                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6124                         parent = path->nodes[level]->start;
6125                 } else {
6126                         BUG_ON(root->root_key.objectid !=
6127                                btrfs_header_owner(path->nodes[level]));
6128                         parent = 0;
6129                 }
6130
6131                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6132                                         root->root_key.objectid, level - 1, 0);
6133                 BUG_ON(ret);
6134         }
6135         btrfs_tree_unlock(next);
6136         free_extent_buffer(next);
6137         *lookup_info = 1;
6138         return 1;
6139 }
6140
6141 /*
6142  * hepler to process tree block while walking up the tree.
6143  *
6144  * when wc->stage == DROP_REFERENCE, this function drops
6145  * reference count on the block.
6146  *
6147  * when wc->stage == UPDATE_BACKREF, this function changes
6148  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6149  * to UPDATE_BACKREF previously while processing the block.
6150  *
6151  * NOTE: return value 1 means we should stop walking up.
6152  */
6153 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6154                                  struct btrfs_root *root,
6155                                  struct btrfs_path *path,
6156                                  struct walk_control *wc)
6157 {
6158         int ret;
6159         int level = wc->level;
6160         struct extent_buffer *eb = path->nodes[level];
6161         u64 parent = 0;
6162
6163         if (wc->stage == UPDATE_BACKREF) {
6164                 BUG_ON(wc->shared_level < level);
6165                 if (level < wc->shared_level)
6166                         goto out;
6167
6168                 ret = find_next_key(path, level + 1, &wc->update_progress);
6169                 if (ret > 0)
6170                         wc->update_ref = 0;
6171
6172                 wc->stage = DROP_REFERENCE;
6173                 wc->shared_level = -1;
6174                 path->slots[level] = 0;
6175
6176                 /*
6177                  * check reference count again if the block isn't locked.
6178                  * we should start walking down the tree again if reference
6179                  * count is one.
6180                  */
6181                 if (!path->locks[level]) {
6182                         BUG_ON(level == 0);
6183                         btrfs_tree_lock(eb);
6184                         btrfs_set_lock_blocking(eb);
6185                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6186
6187                         ret = btrfs_lookup_extent_info(trans, root,
6188                                                        eb->start, eb->len,
6189                                                        &wc->refs[level],
6190                                                        &wc->flags[level]);
6191                         BUG_ON(ret);
6192                         BUG_ON(wc->refs[level] == 0);
6193                         if (wc->refs[level] == 1) {
6194                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6195                                 return 1;
6196                         }
6197                 }
6198         }
6199
6200         /* wc->stage == DROP_REFERENCE */
6201         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6202
6203         if (wc->refs[level] == 1) {
6204                 if (level == 0) {
6205                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6206                                 ret = btrfs_dec_ref(trans, root, eb, 1);
6207                         else
6208                                 ret = btrfs_dec_ref(trans, root, eb, 0);
6209                         BUG_ON(ret);
6210                 }
6211                 /* make block locked assertion in clean_tree_block happy */
6212                 if (!path->locks[level] &&
6213                     btrfs_header_generation(eb) == trans->transid) {
6214                         btrfs_tree_lock(eb);
6215                         btrfs_set_lock_blocking(eb);
6216                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6217                 }
6218                 clean_tree_block(trans, root, eb);
6219         }
6220
6221         if (eb == root->node) {
6222                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6223                         parent = eb->start;
6224                 else
6225                         BUG_ON(root->root_key.objectid !=
6226                                btrfs_header_owner(eb));
6227         } else {
6228                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6229                         parent = path->nodes[level + 1]->start;
6230                 else
6231                         BUG_ON(root->root_key.objectid !=
6232                                btrfs_header_owner(path->nodes[level + 1]));
6233         }
6234
6235         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6236 out:
6237         wc->refs[level] = 0;
6238         wc->flags[level] = 0;
6239         return 0;
6240 }
6241
6242 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6243                                    struct btrfs_root *root,
6244                                    struct btrfs_path *path,
6245                                    struct walk_control *wc)
6246 {
6247         int level = wc->level;
6248         int lookup_info = 1;
6249         int ret;
6250
6251         while (level >= 0) {
6252                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6253                 if (ret > 0)
6254                         break;
6255
6256                 if (level == 0)
6257                         break;
6258
6259                 if (path->slots[level] >=
6260                     btrfs_header_nritems(path->nodes[level]))
6261                         break;
6262
6263                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6264                 if (ret > 0) {
6265                         path->slots[level]++;
6266                         continue;
6267                 } else if (ret < 0)
6268                         return ret;
6269                 level = wc->level;
6270         }
6271         return 0;
6272 }
6273
6274 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6275                                  struct btrfs_root *root,
6276                                  struct btrfs_path *path,
6277                                  struct walk_control *wc, int max_level)
6278 {
6279         int level = wc->level;
6280         int ret;
6281
6282         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6283         while (level < max_level && path->nodes[level]) {
6284                 wc->level = level;
6285                 if (path->slots[level] + 1 <
6286                     btrfs_header_nritems(path->nodes[level])) {
6287                         path->slots[level]++;
6288                         return 0;
6289                 } else {
6290                         ret = walk_up_proc(trans, root, path, wc);
6291                         if (ret > 0)
6292                                 return 0;
6293
6294                         if (path->locks[level]) {
6295                                 btrfs_tree_unlock_rw(path->nodes[level],
6296                                                      path->locks[level]);
6297                                 path->locks[level] = 0;
6298                         }
6299                         free_extent_buffer(path->nodes[level]);
6300                         path->nodes[level] = NULL;
6301                         level++;
6302                 }
6303         }
6304         return 1;
6305 }
6306
6307 /*
6308  * drop a subvolume tree.
6309  *
6310  * this function traverses the tree freeing any blocks that only
6311  * referenced by the tree.
6312  *
6313  * when a shared tree block is found. this function decreases its
6314  * reference count by one. if update_ref is true, this function
6315  * also make sure backrefs for the shared block and all lower level
6316  * blocks are properly updated.
6317  */
6318 void btrfs_drop_snapshot(struct btrfs_root *root,
6319                          struct btrfs_block_rsv *block_rsv, int update_ref)
6320 {
6321         struct btrfs_path *path;
6322         struct btrfs_trans_handle *trans;
6323         struct btrfs_root *tree_root = root->fs_info->tree_root;
6324         struct btrfs_root_item *root_item = &root->root_item;
6325         struct walk_control *wc;
6326         struct btrfs_key key;
6327         int err = 0;
6328         int ret;
6329         int level;
6330
6331         path = btrfs_alloc_path();
6332         if (!path) {
6333                 err = -ENOMEM;
6334                 goto out;
6335         }
6336
6337         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6338         if (!wc) {
6339                 btrfs_free_path(path);
6340                 err = -ENOMEM;
6341                 goto out;
6342         }
6343
6344         trans = btrfs_start_transaction(tree_root, 0);
6345         BUG_ON(IS_ERR(trans));
6346
6347         if (block_rsv)
6348                 trans->block_rsv = block_rsv;
6349
6350         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6351                 level = btrfs_header_level(root->node);
6352                 path->nodes[level] = btrfs_lock_root_node(root);
6353                 btrfs_set_lock_blocking(path->nodes[level]);
6354                 path->slots[level] = 0;
6355                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6356                 memset(&wc->update_progress, 0,
6357                        sizeof(wc->update_progress));
6358         } else {
6359                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6360                 memcpy(&wc->update_progress, &key,
6361                        sizeof(wc->update_progress));
6362
6363                 level = root_item->drop_level;
6364                 BUG_ON(level == 0);
6365                 path->lowest_level = level;
6366                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6367                 path->lowest_level = 0;
6368                 if (ret < 0) {
6369                         err = ret;
6370                         goto out_free;
6371                 }
6372                 WARN_ON(ret > 0);
6373
6374                 /*
6375                  * unlock our path, this is safe because only this
6376                  * function is allowed to delete this snapshot
6377                  */
6378                 btrfs_unlock_up_safe(path, 0);
6379
6380                 level = btrfs_header_level(root->node);
6381                 while (1) {
6382                         btrfs_tree_lock(path->nodes[level]);
6383                         btrfs_set_lock_blocking(path->nodes[level]);
6384
6385                         ret = btrfs_lookup_extent_info(trans, root,
6386                                                 path->nodes[level]->start,
6387                                                 path->nodes[level]->len,
6388                                                 &wc->refs[level],
6389                                                 &wc->flags[level]);
6390                         BUG_ON(ret);
6391                         BUG_ON(wc->refs[level] == 0);
6392
6393                         if (level == root_item->drop_level)
6394                                 break;
6395
6396                         btrfs_tree_unlock(path->nodes[level]);
6397                         WARN_ON(wc->refs[level] != 1);
6398                         level--;
6399                 }
6400         }
6401
6402         wc->level = level;
6403         wc->shared_level = -1;
6404         wc->stage = DROP_REFERENCE;
6405         wc->update_ref = update_ref;
6406         wc->keep_locks = 0;
6407         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6408
6409         while (1) {
6410                 ret = walk_down_tree(trans, root, path, wc);
6411                 if (ret < 0) {
6412                         err = ret;
6413                         break;
6414                 }
6415
6416                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
6417                 if (ret < 0) {
6418                         err = ret;
6419                         break;
6420                 }
6421
6422                 if (ret > 0) {
6423                         BUG_ON(wc->stage != DROP_REFERENCE);
6424                         break;
6425                 }
6426
6427                 if (wc->stage == DROP_REFERENCE) {
6428                         level = wc->level;
6429                         btrfs_node_key(path->nodes[level],
6430                                        &root_item->drop_progress,
6431                                        path->slots[level]);
6432                         root_item->drop_level = level;
6433                 }
6434
6435                 BUG_ON(wc->level == 0);
6436                 if (btrfs_should_end_transaction(trans, tree_root)) {
6437                         ret = btrfs_update_root(trans, tree_root,
6438                                                 &root->root_key,
6439                                                 root_item);
6440                         BUG_ON(ret);
6441
6442                         btrfs_end_transaction_throttle(trans, tree_root);
6443                         trans = btrfs_start_transaction(tree_root, 0);
6444                         BUG_ON(IS_ERR(trans));
6445                         if (block_rsv)
6446                                 trans->block_rsv = block_rsv;
6447                 }
6448         }
6449         btrfs_release_path(path);
6450         BUG_ON(err);
6451
6452         ret = btrfs_del_root(trans, tree_root, &root->root_key);
6453         BUG_ON(ret);
6454
6455         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
6456                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
6457                                            NULL, NULL);
6458                 BUG_ON(ret < 0);
6459                 if (ret > 0) {
6460                         /* if we fail to delete the orphan item this time
6461                          * around, it'll get picked up the next time.
6462                          *
6463                          * The most common failure here is just -ENOENT.
6464                          */
6465                         btrfs_del_orphan_item(trans, tree_root,
6466                                               root->root_key.objectid);
6467                 }
6468         }
6469
6470         if (root->in_radix) {
6471                 btrfs_free_fs_root(tree_root->fs_info, root);
6472         } else {
6473                 free_extent_buffer(root->node);
6474                 free_extent_buffer(root->commit_root);
6475                 kfree(root);
6476         }
6477 out_free:
6478         btrfs_end_transaction_throttle(trans, tree_root);
6479         kfree(wc);
6480         btrfs_free_path(path);
6481 out:
6482         if (err)
6483                 btrfs_std_error(root->fs_info, err);
6484         return;
6485 }
6486
6487 /*
6488  * drop subtree rooted at tree block 'node'.
6489  *
6490  * NOTE: this function will unlock and release tree block 'node'
6491  */
6492 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
6493                         struct btrfs_root *root,
6494                         struct extent_buffer *node,
6495                         struct extent_buffer *parent)
6496 {
6497         struct btrfs_path *path;
6498         struct walk_control *wc;
6499         int level;
6500         int parent_level;
6501         int ret = 0;
6502         int wret;
6503
6504         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
6505
6506         path = btrfs_alloc_path();
6507         if (!path)
6508                 return -ENOMEM;
6509
6510         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6511         if (!wc) {
6512                 btrfs_free_path(path);
6513                 return -ENOMEM;
6514         }
6515
6516         btrfs_assert_tree_locked(parent);
6517         parent_level = btrfs_header_level(parent);
6518         extent_buffer_get(parent);
6519         path->nodes[parent_level] = parent;
6520         path->slots[parent_level] = btrfs_header_nritems(parent);
6521
6522         btrfs_assert_tree_locked(node);
6523         level = btrfs_header_level(node);
6524         path->nodes[level] = node;
6525         path->slots[level] = 0;
6526         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6527
6528         wc->refs[parent_level] = 1;
6529         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6530         wc->level = level;
6531         wc->shared_level = -1;
6532         wc->stage = DROP_REFERENCE;
6533         wc->update_ref = 0;
6534         wc->keep_locks = 1;
6535         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6536
6537         while (1) {
6538                 wret = walk_down_tree(trans, root, path, wc);
6539                 if (wret < 0) {
6540                         ret = wret;
6541                         break;
6542                 }
6543
6544                 wret = walk_up_tree(trans, root, path, wc, parent_level);
6545                 if (wret < 0)
6546                         ret = wret;
6547                 if (wret != 0)
6548                         break;
6549         }
6550
6551         kfree(wc);
6552         btrfs_free_path(path);
6553         return ret;
6554 }
6555
6556 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
6557 {
6558         u64 num_devices;
6559         u64 stripped = BTRFS_BLOCK_GROUP_RAID0 |
6560                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
6561
6562         /*
6563          * we add in the count of missing devices because we want
6564          * to make sure that any RAID levels on a degraded FS
6565          * continue to be honored.
6566          */
6567         num_devices = root->fs_info->fs_devices->rw_devices +
6568                 root->fs_info->fs_devices->missing_devices;
6569
6570         if (num_devices == 1) {
6571                 stripped |= BTRFS_BLOCK_GROUP_DUP;
6572                 stripped = flags & ~stripped;
6573
6574                 /* turn raid0 into single device chunks */
6575                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
6576                         return stripped;
6577
6578                 /* turn mirroring into duplication */
6579                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
6580                              BTRFS_BLOCK_GROUP_RAID10))
6581                         return stripped | BTRFS_BLOCK_GROUP_DUP;
6582                 return flags;
6583         } else {
6584                 /* they already had raid on here, just return */
6585                 if (flags & stripped)
6586                         return flags;
6587
6588                 stripped |= BTRFS_BLOCK_GROUP_DUP;
6589                 stripped = flags & ~stripped;
6590
6591                 /* switch duplicated blocks with raid1 */
6592                 if (flags & BTRFS_BLOCK_GROUP_DUP)
6593                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
6594
6595                 /* turn single device chunks into raid0 */
6596                 return stripped | BTRFS_BLOCK_GROUP_RAID0;
6597         }
6598         return flags;
6599 }
6600
6601 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
6602 {
6603         struct btrfs_space_info *sinfo = cache->space_info;
6604         u64 num_bytes;
6605         u64 min_allocable_bytes;
6606         int ret = -ENOSPC;
6607
6608
6609         /*
6610          * We need some metadata space and system metadata space for
6611          * allocating chunks in some corner cases until we force to set
6612          * it to be readonly.
6613          */
6614         if ((sinfo->flags &
6615              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
6616             !force)
6617                 min_allocable_bytes = 1 * 1024 * 1024;
6618         else
6619                 min_allocable_bytes = 0;
6620
6621         spin_lock(&sinfo->lock);
6622         spin_lock(&cache->lock);
6623
6624         if (cache->ro) {
6625                 ret = 0;
6626                 goto out;
6627         }
6628
6629         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
6630                     cache->bytes_super - btrfs_block_group_used(&cache->item);
6631
6632         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
6633             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
6634             min_allocable_bytes <= sinfo->total_bytes) {
6635                 sinfo->bytes_readonly += num_bytes;
6636                 cache->ro = 1;
6637                 ret = 0;
6638         }
6639 out:
6640         spin_unlock(&cache->lock);
6641         spin_unlock(&sinfo->lock);
6642         return ret;
6643 }
6644
6645 int btrfs_set_block_group_ro(struct btrfs_root *root,
6646                              struct btrfs_block_group_cache *cache)
6647
6648 {
6649         struct btrfs_trans_handle *trans;
6650         u64 alloc_flags;
6651         int ret;
6652
6653         BUG_ON(cache->ro);
6654
6655         trans = btrfs_join_transaction(root);
6656         BUG_ON(IS_ERR(trans));
6657
6658         alloc_flags = update_block_group_flags(root, cache->flags);
6659         if (alloc_flags != cache->flags)
6660                 do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6661                                CHUNK_ALLOC_FORCE);
6662
6663         ret = set_block_group_ro(cache, 0);
6664         if (!ret)
6665                 goto out;
6666         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
6667         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6668                              CHUNK_ALLOC_FORCE);
6669         if (ret < 0)
6670                 goto out;
6671         ret = set_block_group_ro(cache, 0);
6672 out:
6673         btrfs_end_transaction(trans, root);
6674         return ret;
6675 }
6676
6677 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
6678                             struct btrfs_root *root, u64 type)
6679 {
6680         u64 alloc_flags = get_alloc_profile(root, type);
6681         return do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6682                               CHUNK_ALLOC_FORCE);
6683 }
6684
6685 /*
6686  * helper to account the unused space of all the readonly block group in the
6687  * list. takes mirrors into account.
6688  */
6689 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
6690 {
6691         struct btrfs_block_group_cache *block_group;
6692         u64 free_bytes = 0;
6693         int factor;
6694
6695         list_for_each_entry(block_group, groups_list, list) {
6696                 spin_lock(&block_group->lock);
6697
6698                 if (!block_group->ro) {
6699                         spin_unlock(&block_group->lock);
6700                         continue;
6701                 }
6702
6703                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
6704                                           BTRFS_BLOCK_GROUP_RAID10 |
6705                                           BTRFS_BLOCK_GROUP_DUP))
6706                         factor = 2;
6707                 else
6708                         factor = 1;
6709
6710                 free_bytes += (block_group->key.offset -
6711                                btrfs_block_group_used(&block_group->item)) *
6712                                factor;
6713
6714                 spin_unlock(&block_group->lock);
6715         }
6716
6717         return free_bytes;
6718 }
6719
6720 /*
6721  * helper to account the unused space of all the readonly block group in the
6722  * space_info. takes mirrors into account.
6723  */
6724 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
6725 {
6726         int i;
6727         u64 free_bytes = 0;
6728
6729         spin_lock(&sinfo->lock);
6730
6731         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
6732                 if (!list_empty(&sinfo->block_groups[i]))
6733                         free_bytes += __btrfs_get_ro_block_group_free_space(
6734                                                 &sinfo->block_groups[i]);
6735
6736         spin_unlock(&sinfo->lock);
6737
6738         return free_bytes;
6739 }
6740
6741 int btrfs_set_block_group_rw(struct btrfs_root *root,
6742                               struct btrfs_block_group_cache *cache)
6743 {
6744         struct btrfs_space_info *sinfo = cache->space_info;
6745         u64 num_bytes;
6746
6747         BUG_ON(!cache->ro);
6748
6749         spin_lock(&sinfo->lock);
6750         spin_lock(&cache->lock);
6751         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
6752                     cache->bytes_super - btrfs_block_group_used(&cache->item);
6753         sinfo->bytes_readonly -= num_bytes;
6754         cache->ro = 0;
6755         spin_unlock(&cache->lock);
6756         spin_unlock(&sinfo->lock);
6757         return 0;
6758 }
6759
6760 /*
6761  * checks to see if its even possible to relocate this block group.
6762  *
6763  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
6764  * ok to go ahead and try.
6765  */
6766 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
6767 {
6768         struct btrfs_block_group_cache *block_group;
6769         struct btrfs_space_info *space_info;
6770         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
6771         struct btrfs_device *device;
6772         u64 min_free;
6773         u64 dev_min = 1;
6774         u64 dev_nr = 0;
6775         int index;
6776         int full = 0;
6777         int ret = 0;
6778
6779         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
6780
6781         /* odd, couldn't find the block group, leave it alone */
6782         if (!block_group)
6783                 return -1;
6784
6785         min_free = btrfs_block_group_used(&block_group->item);
6786
6787         /* no bytes used, we're good */
6788         if (!min_free)
6789                 goto out;
6790
6791         space_info = block_group->space_info;
6792         spin_lock(&space_info->lock);
6793
6794         full = space_info->full;
6795
6796         /*
6797          * if this is the last block group we have in this space, we can't
6798          * relocate it unless we're able to allocate a new chunk below.
6799          *
6800          * Otherwise, we need to make sure we have room in the space to handle
6801          * all of the extents from this block group.  If we can, we're good
6802          */
6803         if ((space_info->total_bytes != block_group->key.offset) &&
6804             (space_info->bytes_used + space_info->bytes_reserved +
6805              space_info->bytes_pinned + space_info->bytes_readonly +
6806              min_free < space_info->total_bytes)) {
6807                 spin_unlock(&space_info->lock);
6808                 goto out;
6809         }
6810         spin_unlock(&space_info->lock);
6811
6812         /*
6813          * ok we don't have enough space, but maybe we have free space on our
6814          * devices to allocate new chunks for relocation, so loop through our
6815          * alloc devices and guess if we have enough space.  However, if we
6816          * were marked as full, then we know there aren't enough chunks, and we
6817          * can just return.
6818          */
6819         ret = -1;
6820         if (full)
6821                 goto out;
6822
6823         /*
6824          * index:
6825          *      0: raid10
6826          *      1: raid1
6827          *      2: dup
6828          *      3: raid0
6829          *      4: single
6830          */
6831         index = get_block_group_index(block_group);
6832         if (index == 0) {
6833                 dev_min = 4;
6834                 /* Divide by 2 */
6835                 min_free >>= 1;
6836         } else if (index == 1) {
6837                 dev_min = 2;
6838         } else if (index == 2) {
6839                 /* Multiply by 2 */
6840                 min_free <<= 1;
6841         } else if (index == 3) {
6842                 dev_min = fs_devices->rw_devices;
6843                 do_div(min_free, dev_min);
6844         }
6845
6846         mutex_lock(&root->fs_info->chunk_mutex);
6847         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
6848                 u64 dev_offset;
6849
6850                 /*
6851                  * check to make sure we can actually find a chunk with enough
6852                  * space to fit our block group in.
6853                  */
6854                 if (device->total_bytes > device->bytes_used + min_free) {
6855                         ret = find_free_dev_extent(NULL, device, min_free,
6856                                                    &dev_offset, NULL);
6857                         if (!ret)
6858                                 dev_nr++;
6859
6860                         if (dev_nr >= dev_min)
6861                                 break;
6862
6863                         ret = -1;
6864                 }
6865         }
6866         mutex_unlock(&root->fs_info->chunk_mutex);
6867 out:
6868         btrfs_put_block_group(block_group);
6869         return ret;
6870 }
6871
6872 static int find_first_block_group(struct btrfs_root *root,
6873                 struct btrfs_path *path, struct btrfs_key *key)
6874 {
6875         int ret = 0;
6876         struct btrfs_key found_key;
6877         struct extent_buffer *leaf;
6878         int slot;
6879
6880         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
6881         if (ret < 0)
6882                 goto out;
6883
6884         while (1) {
6885                 slot = path->slots[0];
6886                 leaf = path->nodes[0];
6887                 if (slot >= btrfs_header_nritems(leaf)) {
6888                         ret = btrfs_next_leaf(root, path);
6889                         if (ret == 0)
6890                                 continue;
6891                         if (ret < 0)
6892                                 goto out;
6893                         break;
6894                 }
6895                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
6896
6897                 if (found_key.objectid >= key->objectid &&
6898                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
6899                         ret = 0;
6900                         goto out;
6901                 }
6902                 path->slots[0]++;
6903         }
6904 out:
6905         return ret;
6906 }
6907
6908 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
6909 {
6910         struct btrfs_block_group_cache *block_group;
6911         u64 last = 0;
6912
6913         while (1) {
6914                 struct inode *inode;
6915
6916                 block_group = btrfs_lookup_first_block_group(info, last);
6917                 while (block_group) {
6918                         spin_lock(&block_group->lock);
6919                         if (block_group->iref)
6920                                 break;
6921                         spin_unlock(&block_group->lock);
6922                         block_group = next_block_group(info->tree_root,
6923                                                        block_group);
6924                 }
6925                 if (!block_group) {
6926                         if (last == 0)
6927                                 break;
6928                         last = 0;
6929                         continue;
6930                 }
6931
6932                 inode = block_group->inode;
6933                 block_group->iref = 0;
6934                 block_group->inode = NULL;
6935                 spin_unlock(&block_group->lock);
6936                 iput(inode);
6937                 last = block_group->key.objectid + block_group->key.offset;
6938                 btrfs_put_block_group(block_group);
6939         }
6940 }
6941
6942 int btrfs_free_block_groups(struct btrfs_fs_info *info)
6943 {
6944         struct btrfs_block_group_cache *block_group;
6945         struct btrfs_space_info *space_info;
6946         struct btrfs_caching_control *caching_ctl;
6947         struct rb_node *n;
6948
6949         down_write(&info->extent_commit_sem);
6950         while (!list_empty(&info->caching_block_groups)) {
6951                 caching_ctl = list_entry(info->caching_block_groups.next,
6952                                          struct btrfs_caching_control, list);
6953                 list_del(&caching_ctl->list);
6954                 put_caching_control(caching_ctl);
6955         }
6956         up_write(&info->extent_commit_sem);
6957
6958         spin_lock(&info->block_group_cache_lock);
6959         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
6960                 block_group = rb_entry(n, struct btrfs_block_group_cache,
6961                                        cache_node);
6962                 rb_erase(&block_group->cache_node,
6963                          &info->block_group_cache_tree);
6964                 spin_unlock(&info->block_group_cache_lock);
6965
6966                 down_write(&block_group->space_info->groups_sem);
6967                 list_del(&block_group->list);
6968                 up_write(&block_group->space_info->groups_sem);
6969
6970                 if (block_group->cached == BTRFS_CACHE_STARTED)
6971                         wait_block_group_cache_done(block_group);
6972
6973                 /*
6974                  * We haven't cached this block group, which means we could
6975                  * possibly have excluded extents on this block group.
6976                  */
6977                 if (block_group->cached == BTRFS_CACHE_NO)
6978                         free_excluded_extents(info->extent_root, block_group);
6979
6980                 btrfs_remove_free_space_cache(block_group);
6981                 btrfs_put_block_group(block_group);
6982
6983                 spin_lock(&info->block_group_cache_lock);
6984         }
6985         spin_unlock(&info->block_group_cache_lock);
6986
6987         /* now that all the block groups are freed, go through and
6988          * free all the space_info structs.  This is only called during
6989          * the final stages of unmount, and so we know nobody is
6990          * using them.  We call synchronize_rcu() once before we start,
6991          * just to be on the safe side.
6992          */
6993         synchronize_rcu();
6994
6995         release_global_block_rsv(info);
6996
6997         while(!list_empty(&info->space_info)) {
6998                 space_info = list_entry(info->space_info.next,
6999                                         struct btrfs_space_info,
7000                                         list);
7001                 if (space_info->bytes_pinned > 0 ||
7002                     space_info->bytes_reserved > 0 ||
7003                     space_info->bytes_may_use > 0) {
7004                         WARN_ON(1);
7005                         dump_space_info(space_info, 0, 0);
7006                 }
7007                 list_del(&space_info->list);
7008                 kfree(space_info);
7009         }
7010         return 0;
7011 }
7012
7013 static void __link_block_group(struct btrfs_space_info *space_info,
7014                                struct btrfs_block_group_cache *cache)
7015 {
7016         int index = get_block_group_index(cache);
7017
7018         down_write(&space_info->groups_sem);
7019         list_add_tail(&cache->list, &space_info->block_groups[index]);
7020         up_write(&space_info->groups_sem);
7021 }
7022
7023 int btrfs_read_block_groups(struct btrfs_root *root)
7024 {
7025         struct btrfs_path *path;
7026         int ret;
7027         struct btrfs_block_group_cache *cache;
7028         struct btrfs_fs_info *info = root->fs_info;
7029         struct btrfs_space_info *space_info;
7030         struct btrfs_key key;
7031         struct btrfs_key found_key;
7032         struct extent_buffer *leaf;
7033         int need_clear = 0;
7034         u64 cache_gen;
7035
7036         root = info->extent_root;
7037         key.objectid = 0;
7038         key.offset = 0;
7039         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7040         path = btrfs_alloc_path();
7041         if (!path)
7042                 return -ENOMEM;
7043         path->reada = 1;
7044
7045         cache_gen = btrfs_super_cache_generation(&root->fs_info->super_copy);
7046         if (btrfs_test_opt(root, SPACE_CACHE) &&
7047             btrfs_super_generation(&root->fs_info->super_copy) != cache_gen)
7048                 need_clear = 1;
7049         if (btrfs_test_opt(root, CLEAR_CACHE))
7050                 need_clear = 1;
7051
7052         while (1) {
7053                 ret = find_first_block_group(root, path, &key);
7054                 if (ret > 0)
7055                         break;
7056                 if (ret != 0)
7057                         goto error;
7058                 leaf = path->nodes[0];
7059                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7060                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7061                 if (!cache) {
7062                         ret = -ENOMEM;
7063                         goto error;
7064                 }
7065                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7066                                                 GFP_NOFS);
7067                 if (!cache->free_space_ctl) {
7068                         kfree(cache);
7069                         ret = -ENOMEM;
7070                         goto error;
7071                 }
7072
7073                 atomic_set(&cache->count, 1);
7074                 spin_lock_init(&cache->lock);
7075                 cache->fs_info = info;
7076                 INIT_LIST_HEAD(&cache->list);
7077                 INIT_LIST_HEAD(&cache->cluster_list);
7078
7079                 if (need_clear)
7080                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7081
7082                 read_extent_buffer(leaf, &cache->item,
7083                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7084                                    sizeof(cache->item));
7085                 memcpy(&cache->key, &found_key, sizeof(found_key));
7086
7087                 key.objectid = found_key.objectid + found_key.offset;
7088                 btrfs_release_path(path);
7089                 cache->flags = btrfs_block_group_flags(&cache->item);
7090                 cache->sectorsize = root->sectorsize;
7091
7092                 btrfs_init_free_space_ctl(cache);
7093
7094                 /*
7095                  * We need to exclude the super stripes now so that the space
7096                  * info has super bytes accounted for, otherwise we'll think
7097                  * we have more space than we actually do.
7098                  */
7099                 exclude_super_stripes(root, cache);
7100
7101                 /*
7102                  * check for two cases, either we are full, and therefore
7103                  * don't need to bother with the caching work since we won't
7104                  * find any space, or we are empty, and we can just add all
7105                  * the space in and be done with it.  This saves us _alot_ of
7106                  * time, particularly in the full case.
7107                  */
7108                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7109                         cache->last_byte_to_unpin = (u64)-1;
7110                         cache->cached = BTRFS_CACHE_FINISHED;
7111                         free_excluded_extents(root, cache);
7112                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7113                         cache->last_byte_to_unpin = (u64)-1;
7114                         cache->cached = BTRFS_CACHE_FINISHED;
7115                         add_new_free_space(cache, root->fs_info,
7116                                            found_key.objectid,
7117                                            found_key.objectid +
7118                                            found_key.offset);
7119                         free_excluded_extents(root, cache);
7120                 }
7121
7122                 ret = update_space_info(info, cache->flags, found_key.offset,
7123                                         btrfs_block_group_used(&cache->item),
7124                                         &space_info);
7125                 BUG_ON(ret);
7126                 cache->space_info = space_info;
7127                 spin_lock(&cache->space_info->lock);
7128                 cache->space_info->bytes_readonly += cache->bytes_super;
7129                 spin_unlock(&cache->space_info->lock);
7130
7131                 __link_block_group(space_info, cache);
7132
7133                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7134                 BUG_ON(ret);
7135
7136                 set_avail_alloc_bits(root->fs_info, cache->flags);
7137                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7138                         set_block_group_ro(cache, 1);
7139         }
7140
7141         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7142                 if (!(get_alloc_profile(root, space_info->flags) &
7143                       (BTRFS_BLOCK_GROUP_RAID10 |
7144                        BTRFS_BLOCK_GROUP_RAID1 |
7145                        BTRFS_BLOCK_GROUP_DUP)))
7146                         continue;
7147                 /*
7148                  * avoid allocating from un-mirrored block group if there are
7149                  * mirrored block groups.
7150                  */
7151                 list_for_each_entry(cache, &space_info->block_groups[3], list)
7152                         set_block_group_ro(cache, 1);
7153                 list_for_each_entry(cache, &space_info->block_groups[4], list)
7154                         set_block_group_ro(cache, 1);
7155         }
7156
7157         init_global_block_rsv(info);
7158         ret = 0;
7159 error:
7160         btrfs_free_path(path);
7161         return ret;
7162 }
7163
7164 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7165                            struct btrfs_root *root, u64 bytes_used,
7166                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7167                            u64 size)
7168 {
7169         int ret;
7170         struct btrfs_root *extent_root;
7171         struct btrfs_block_group_cache *cache;
7172
7173         extent_root = root->fs_info->extent_root;
7174
7175         root->fs_info->last_trans_log_full_commit = trans->transid;
7176
7177         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7178         if (!cache)
7179                 return -ENOMEM;
7180         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7181                                         GFP_NOFS);
7182         if (!cache->free_space_ctl) {
7183                 kfree(cache);
7184                 return -ENOMEM;
7185         }
7186
7187         cache->key.objectid = chunk_offset;
7188         cache->key.offset = size;
7189         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7190         cache->sectorsize = root->sectorsize;
7191         cache->fs_info = root->fs_info;
7192
7193         atomic_set(&cache->count, 1);
7194         spin_lock_init(&cache->lock);
7195         INIT_LIST_HEAD(&cache->list);
7196         INIT_LIST_HEAD(&cache->cluster_list);
7197
7198         btrfs_init_free_space_ctl(cache);
7199
7200         btrfs_set_block_group_used(&cache->item, bytes_used);
7201         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7202         cache->flags = type;
7203         btrfs_set_block_group_flags(&cache->item, type);
7204
7205         cache->last_byte_to_unpin = (u64)-1;
7206         cache->cached = BTRFS_CACHE_FINISHED;
7207         exclude_super_stripes(root, cache);
7208
7209         add_new_free_space(cache, root->fs_info, chunk_offset,
7210                            chunk_offset + size);
7211
7212         free_excluded_extents(root, cache);
7213
7214         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7215                                 &cache->space_info);
7216         BUG_ON(ret);
7217
7218         spin_lock(&cache->space_info->lock);
7219         cache->space_info->bytes_readonly += cache->bytes_super;
7220         spin_unlock(&cache->space_info->lock);
7221
7222         __link_block_group(cache->space_info, cache);
7223
7224         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7225         BUG_ON(ret);
7226
7227         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
7228                                 sizeof(cache->item));
7229         BUG_ON(ret);
7230
7231         set_avail_alloc_bits(extent_root->fs_info, type);
7232
7233         return 0;
7234 }
7235
7236 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7237                              struct btrfs_root *root, u64 group_start)
7238 {
7239         struct btrfs_path *path;
7240         struct btrfs_block_group_cache *block_group;
7241         struct btrfs_free_cluster *cluster;
7242         struct btrfs_root *tree_root = root->fs_info->tree_root;
7243         struct btrfs_key key;
7244         struct inode *inode;
7245         int ret;
7246         int factor;
7247
7248         root = root->fs_info->extent_root;
7249
7250         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7251         BUG_ON(!block_group);
7252         BUG_ON(!block_group->ro);
7253
7254         /*
7255          * Free the reserved super bytes from this block group before
7256          * remove it.
7257          */
7258         free_excluded_extents(root, block_group);
7259
7260         memcpy(&key, &block_group->key, sizeof(key));
7261         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7262                                   BTRFS_BLOCK_GROUP_RAID1 |
7263                                   BTRFS_BLOCK_GROUP_RAID10))
7264                 factor = 2;
7265         else
7266                 factor = 1;
7267
7268         /* make sure this block group isn't part of an allocation cluster */
7269         cluster = &root->fs_info->data_alloc_cluster;
7270         spin_lock(&cluster->refill_lock);
7271         btrfs_return_cluster_to_free_space(block_group, cluster);
7272         spin_unlock(&cluster->refill_lock);
7273
7274         /*
7275          * make sure this block group isn't part of a metadata
7276          * allocation cluster
7277          */
7278         cluster = &root->fs_info->meta_alloc_cluster;
7279         spin_lock(&cluster->refill_lock);
7280         btrfs_return_cluster_to_free_space(block_group, cluster);
7281         spin_unlock(&cluster->refill_lock);
7282
7283         path = btrfs_alloc_path();
7284         if (!path) {
7285                 ret = -ENOMEM;
7286                 goto out;
7287         }
7288
7289         inode = lookup_free_space_inode(root, block_group, path);
7290         if (!IS_ERR(inode)) {
7291                 ret = btrfs_orphan_add(trans, inode);
7292                 BUG_ON(ret);
7293                 clear_nlink(inode);
7294                 /* One for the block groups ref */
7295                 spin_lock(&block_group->lock);
7296                 if (block_group->iref) {
7297                         block_group->iref = 0;
7298                         block_group->inode = NULL;
7299                         spin_unlock(&block_group->lock);
7300                         iput(inode);
7301                 } else {
7302                         spin_unlock(&block_group->lock);
7303                 }
7304                 /* One for our lookup ref */
7305                 btrfs_add_delayed_iput(inode);
7306         }
7307
7308         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
7309         key.offset = block_group->key.objectid;
7310         key.type = 0;
7311
7312         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
7313         if (ret < 0)
7314                 goto out;
7315         if (ret > 0)
7316                 btrfs_release_path(path);
7317         if (ret == 0) {
7318                 ret = btrfs_del_item(trans, tree_root, path);
7319                 if (ret)
7320                         goto out;
7321                 btrfs_release_path(path);
7322         }
7323
7324         spin_lock(&root->fs_info->block_group_cache_lock);
7325         rb_erase(&block_group->cache_node,
7326                  &root->fs_info->block_group_cache_tree);
7327         spin_unlock(&root->fs_info->block_group_cache_lock);
7328
7329         down_write(&block_group->space_info->groups_sem);
7330         /*
7331          * we must use list_del_init so people can check to see if they
7332          * are still on the list after taking the semaphore
7333          */
7334         list_del_init(&block_group->list);
7335         up_write(&block_group->space_info->groups_sem);
7336
7337         if (block_group->cached == BTRFS_CACHE_STARTED)
7338                 wait_block_group_cache_done(block_group);
7339
7340         btrfs_remove_free_space_cache(block_group);
7341
7342         spin_lock(&block_group->space_info->lock);
7343         block_group->space_info->total_bytes -= block_group->key.offset;
7344         block_group->space_info->bytes_readonly -= block_group->key.offset;
7345         block_group->space_info->disk_total -= block_group->key.offset * factor;
7346         spin_unlock(&block_group->space_info->lock);
7347
7348         memcpy(&key, &block_group->key, sizeof(key));
7349
7350         btrfs_clear_space_info_full(root->fs_info);
7351
7352         btrfs_put_block_group(block_group);
7353         btrfs_put_block_group(block_group);
7354
7355         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
7356         if (ret > 0)
7357                 ret = -EIO;
7358         if (ret < 0)
7359                 goto out;
7360
7361         ret = btrfs_del_item(trans, root, path);
7362 out:
7363         btrfs_free_path(path);
7364         return ret;
7365 }
7366
7367 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
7368 {
7369         struct btrfs_space_info *space_info;
7370         struct btrfs_super_block *disk_super;
7371         u64 features;
7372         u64 flags;
7373         int mixed = 0;
7374         int ret;
7375
7376         disk_super = &fs_info->super_copy;
7377         if (!btrfs_super_root(disk_super))
7378                 return 1;
7379
7380         features = btrfs_super_incompat_flags(disk_super);
7381         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
7382                 mixed = 1;
7383
7384         flags = BTRFS_BLOCK_GROUP_SYSTEM;
7385         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7386         if (ret)
7387                 goto out;
7388
7389         if (mixed) {
7390                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
7391                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7392         } else {
7393                 flags = BTRFS_BLOCK_GROUP_METADATA;
7394                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7395                 if (ret)
7396                         goto out;
7397
7398                 flags = BTRFS_BLOCK_GROUP_DATA;
7399                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7400         }
7401 out:
7402         return ret;
7403 }
7404
7405 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
7406 {
7407         return unpin_extent_range(root, start, end);
7408 }
7409
7410 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
7411                                u64 num_bytes, u64 *actual_bytes)
7412 {
7413         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
7414 }
7415
7416 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
7417 {
7418         struct btrfs_fs_info *fs_info = root->fs_info;
7419         struct btrfs_block_group_cache *cache = NULL;
7420         u64 group_trimmed;
7421         u64 start;
7422         u64 end;
7423         u64 trimmed = 0;
7424         int ret = 0;
7425
7426         cache = btrfs_lookup_block_group(fs_info, range->start);
7427
7428         while (cache) {
7429                 if (cache->key.objectid >= (range->start + range->len)) {
7430                         btrfs_put_block_group(cache);
7431                         break;
7432                 }
7433
7434                 start = max(range->start, cache->key.objectid);
7435                 end = min(range->start + range->len,
7436                                 cache->key.objectid + cache->key.offset);
7437
7438                 if (end - start >= range->minlen) {
7439                         if (!block_group_cache_done(cache)) {
7440                                 ret = cache_block_group(cache, NULL, root, 0);
7441                                 if (!ret)
7442                                         wait_block_group_cache_done(cache);
7443                         }
7444                         ret = btrfs_trim_block_group(cache,
7445                                                      &group_trimmed,
7446                                                      start,
7447                                                      end,
7448                                                      range->minlen);
7449
7450                         trimmed += group_trimmed;
7451                         if (ret) {
7452                                 btrfs_put_block_group(cache);
7453                                 break;
7454                         }
7455                 }
7456
7457                 cache = next_block_group(fs_info->tree_root, cache);
7458         }
7459
7460         range->len = trimmed;
7461         return ret;
7462 }