fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include "compat.h"
  28 #include "hash.h"
  29 #include "ctree.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "transaction.h"
  33 #include "volumes.h"
  34 #include "locking.h"
  35 #include "free-space-cache.h"
  36
  37 /* control flags for do_chunk_alloc's force field
  38  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  39  * if we really need one.
  40  *
  41  * CHUNK_ALLOC_FORCE means it must try to allocate one
  42  *
  43  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  44  * if we have very few chunks already allocated.  This is
  45  * used as part of the clustering code to help make sure
  46  * we have a good pool of storage to cluster in, without
  47  * filling the FS with empty chunks
  48  *
  49  */
  50 enum {
  51         CHUNK_ALLOC_NO_FORCE = 0,
  52         CHUNK_ALLOC_FORCE = 1,
  53         CHUNK_ALLOC_LIMITED = 2,
  54 };
  55
  56 /*
  57  * Control how reservations are dealt with.
  58  *
  59  * RESERVE_FREE - freeing a reservation.
  60  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  61  *   ENOSPC accounting
  62  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  63  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  64  */
  65 enum {
  66         RESERVE_FREE = 0,
  67         RESERVE_ALLOC = 1,
  68         RESERVE_ALLOC_NO_ACCOUNT = 2,
  69 };
  70
  71 static int update_block_group(struct btrfs_trans_handle *trans,
  72                               struct btrfs_root *root,
  73                               u64 bytenr, u64 num_bytes, int alloc);
  74 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  75                                 struct btrfs_root *root,
  76                                 u64 bytenr, u64 num_bytes, u64 parent,
  77                                 u64 root_objectid, u64 owner_objectid,
  78                                 u64 owner_offset, int refs_to_drop,
  79                                 struct btrfs_delayed_extent_op *extra_op);
  80 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  81                                     struct extent_buffer *leaf,
  82                                     struct btrfs_extent_item *ei);
  83 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  84                                       struct btrfs_root *root,
  85                                       u64 parent, u64 root_objectid,
  86                                       u64 flags, u64 owner, u64 offset,
  87                                       struct btrfs_key *ins, int ref_mod);
  88 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  89                                      struct btrfs_root *root,
  90                                      u64 parent, u64 root_objectid,
  91                                      u64 flags, struct btrfs_disk_key *key,
  92                                      int level, struct btrfs_key *ins);
  93 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  94                           struct btrfs_root *extent_root, u64 alloc_bytes,
  95                           u64 flags, int force);
  96 static int find_next_key(struct btrfs_path *path, int level,
  97                          struct btrfs_key *key);
  98 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
  99                             int dump_block_groups);
 100 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 101                                        u64 num_bytes, int reserve);
 102
 103 static noinline int
 104 block_group_cache_done(struct btrfs_block_group_cache *cache)
 105 {
 106         smp_mb();
 107         return cache->cached == BTRFS_CACHE_FINISHED;
 108 }
 109
 110 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 111 {
 112         return (cache->flags & bits) == bits;
 113 }
 114
 115 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 116 {
 117         atomic_inc(&cache->count);
 118 }
 119
 120 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 121 {
 122         if (atomic_dec_and_test(&cache->count)) {
 123                 WARN_ON(cache->pinned > 0);
 124                 WARN_ON(cache->reserved > 0);
 125                 kfree(cache->free_space_ctl);
 126                 kfree(cache);
 127         }
 128 }
 129
 130 /*
 131  * this adds the block group to the fs_info rb tree for the block group
 132  * cache
 133  */
 134 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 135                                 struct btrfs_block_group_cache *block_group)
 136 {
 137         struct rb_node **p;
 138         struct rb_node *parent = NULL;
 139         struct btrfs_block_group_cache *cache;
 140
 141         spin_lock(&info->block_group_cache_lock);
 142         p = &info->block_group_cache_tree.rb_node;
 143
 144         while (*p) {
 145                 parent = *p;
 146                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 147                                  cache_node);
 148                 if (block_group->key.objectid < cache->key.objectid) {
 149                         p = &(*p)->rb_left;
 150                 } else if (block_group->key.objectid > cache->key.objectid) {
 151                         p = &(*p)->rb_right;
 152                 } else {
 153                         spin_unlock(&info->block_group_cache_lock);
 154                         return -EEXIST;
 155                 }
 156         }
 157
 158         rb_link_node(&block_group->cache_node, parent, p);
 159         rb_insert_color(&block_group->cache_node,
 160                         &info->block_group_cache_tree);
 161         spin_unlock(&info->block_group_cache_lock);
 162
 163         return 0;
 164 }
 165
 166 /*
 167  * This will return the block group at or after bytenr if contains is 0, else
 168  * it will return the block group that contains the bytenr
 169  */
 170 static struct btrfs_block_group_cache *
 171 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 172                               int contains)
 173 {
 174         struct btrfs_block_group_cache *cache, *ret = NULL;
 175         struct rb_node *n;
 176         u64 end, start;
 177
 178         spin_lock(&info->block_group_cache_lock);
 179         n = info->block_group_cache_tree.rb_node;
 180
 181         while (n) {
 182                 cache = rb_entry(n, struct btrfs_block_group_cache,
 183                                  cache_node);
 184                 end = cache->key.objectid + cache->key.offset - 1;
 185                 start = cache->key.objectid;
 186
 187                 if (bytenr < start) {
 188                         if (!contains && (!ret || start < ret->key.objectid))
 189                                 ret = cache;
 190                         n = n->rb_left;
 191                 } else if (bytenr > start) {
 192                         if (contains && bytenr <= end) {
 193                                 ret = cache;
 194                                 break;
 195                         }
 196                         n = n->rb_right;
 197                 } else {
 198                         ret = cache;
 199                         break;
 200                 }
 201         }
 202         if (ret)
 203                 btrfs_get_block_group(ret);
 204         spin_unlock(&info->block_group_cache_lock);
 205
 206         return ret;
 207 }
 208
 209 static int add_excluded_extent(struct btrfs_root *root,
 210                                u64 start, u64 num_bytes)
 211 {
 212         u64 end = start + num_bytes - 1;
 213         set_extent_bits(&root->fs_info->freed_extents[0],
 214                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 215         set_extent_bits(&root->fs_info->freed_extents[1],
 216                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 217         return 0;
 218 }
 219
 220 static void free_excluded_extents(struct btrfs_root *root,
 221                                   struct btrfs_block_group_cache *cache)
 222 {
 223         u64 start, end;
 224
 225         start = cache->key.objectid;
 226         end = start + cache->key.offset - 1;
 227
 228         clear_extent_bits(&root->fs_info->freed_extents[0],
 229                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 230         clear_extent_bits(&root->fs_info->freed_extents[1],
 231                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 232 }
 233
 234 static int exclude_super_stripes(struct btrfs_root *root,
 235                                  struct btrfs_block_group_cache *cache)
 236 {
 237         u64 bytenr;
 238         u64 *logical;
 239         int stripe_len;
 240         int i, nr, ret;
 241
 242         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 243                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 244                 cache->bytes_super += stripe_len;
 245                 ret = add_excluded_extent(root, cache->key.objectid,
 246                                           stripe_len);
 247                 BUG_ON(ret);
 248         }
 249
 250         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 251                 bytenr = btrfs_sb_offset(i);
 252                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 253                                        cache->key.objectid, bytenr,
 254                                        0, &logical, &nr, &stripe_len);
 255                 BUG_ON(ret);
 256
 257                 while (nr--) {
 258                         cache->bytes_super += stripe_len;
 259                         ret = add_excluded_extent(root, logical[nr],
 260                                                   stripe_len);
 261                         BUG_ON(ret);
 262                 }
 263
 264                 kfree(logical);
 265         }
 266         return 0;
 267 }
 268
 269 static struct btrfs_caching_control *
 270 get_caching_control(struct btrfs_block_group_cache *cache)
 271 {
 272         struct btrfs_caching_control *ctl;
 273
 274         spin_lock(&cache->lock);
 275         if (cache->cached != BTRFS_CACHE_STARTED) {
 276                 spin_unlock(&cache->lock);
 277                 return NULL;
 278         }
 279
 280         /* We're loading it the fast way, so we don't have a caching_ctl. */
 281         if (!cache->caching_ctl) {
 282                 spin_unlock(&cache->lock);
 283                 return NULL;
 284         }
 285
 286         ctl = cache->caching_ctl;
 287         atomic_inc(&ctl->count);
 288         spin_unlock(&cache->lock);
 289         return ctl;
 290 }
 291
 292 static void put_caching_control(struct btrfs_caching_control *ctl)
 293 {
 294         if (atomic_dec_and_test(&ctl->count))
 295                 kfree(ctl);
 296 }
 297
 298 /*
 299  * this is only called by cache_block_group, since we could have freed extents
 300  * we need to check the pinned_extents for any extents that can't be used yet
 301  * since their free space will be released as soon as the transaction commits.
 302  */
 303 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 304                               struct btrfs_fs_info *info, u64 start, u64 end)
 305 {
 306         u64 extent_start, extent_end, size, total_added = 0;
 307         int ret;
 308
 309         while (start < end) {
 310                 ret = find_first_extent_bit(info->pinned_extents, start,
 311                                             &extent_start, &extent_end,
 312                                             EXTENT_DIRTY | EXTENT_UPTODATE);
 313                 if (ret)
 314                         break;
 315
 316                 if (extent_start <= start) {
 317                         start = extent_end + 1;
 318                 } else if (extent_start > start && extent_start < end) {
 319                         size = extent_start - start;
 320                         total_added += size;
 321                         ret = btrfs_add_free_space(block_group, start,
 322                                                    size);
 323                         BUG_ON(ret);
 324                         start = extent_end + 1;
 325                 } else {
 326                         break;
 327                 }
 328         }
 329
 330         if (start < end) {
 331                 size = end - start;
 332                 total_added += size;
 333                 ret = btrfs_add_free_space(block_group, start, size);
 334                 BUG_ON(ret);
 335         }
 336
 337         return total_added;
 338 }
 339
 340 static noinline void caching_thread(struct btrfs_work *work)
 341 {
 342         struct btrfs_block_group_cache *block_group;
 343         struct btrfs_fs_info *fs_info;
 344         struct btrfs_caching_control *caching_ctl;
 345         struct btrfs_root *extent_root;
 346         struct btrfs_path *path;
 347         struct extent_buffer *leaf;
 348         struct btrfs_key key;
 349         u64 total_found = 0;
 350         u64 last = 0;
 351         u32 nritems;
 352         int ret = 0;
 353
 354         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 355         block_group = caching_ctl->block_group;
 356         fs_info = block_group->fs_info;
 357         extent_root = fs_info->extent_root;
 358
 359         path = btrfs_alloc_path();
 360         if (!path)
 361                 goto out;
 362
 363         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 364
 365         /*
 366          * We don't want to deadlock with somebody trying to allocate a new
 367          * extent for the extent root while also trying to search the extent
 368          * root to add free space.  So we skip locking and search the commit
 369          * root, since its read-only
 370          */
 371         path->skip_locking = 1;
 372         path->search_commit_root = 1;
 373         path->reada = 1;
 374
 375         key.objectid = last;
 376         key.offset = 0;
 377         key.type = BTRFS_EXTENT_ITEM_KEY;
 378 again:
 379         mutex_lock(&caching_ctl->mutex);
 380         /* need to make sure the commit_root doesn't disappear */
 381         down_read(&fs_info->extent_commit_sem);
 382
 383         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 384         if (ret < 0)
 385                 goto err;
 386
 387         leaf = path->nodes[0];
 388         nritems = btrfs_header_nritems(leaf);
 389
 390         while (1) {
 391                 if (btrfs_fs_closing(fs_info) > 1) {
 392                         last = (u64)-1;
 393                         break;
 394                 }
 395
 396                 if (path->slots[0] < nritems) {
 397                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 398                 } else {
 399                         ret = find_next_key(path, 0, &key);
 400                         if (ret)
 401                                 break;
 402
 403                         if (need_resched() ||
 404                             btrfs_next_leaf(extent_root, path)) {
 405                                 caching_ctl->progress = last;
 406                                 btrfs_release_path(path);
 407                                 up_read(&fs_info->extent_commit_sem);
 408                                 mutex_unlock(&caching_ctl->mutex);
 409                                 cond_resched();
 410                                 goto again;
 411                         }
 412                         leaf = path->nodes[0];
 413                         nritems = btrfs_header_nritems(leaf);
 414                         continue;
 415                 }
 416
 417                 if (key.objectid < block_group->key.objectid) {
 418                         path->slots[0]++;
 419                         continue;
 420                 }
 421
 422                 if (key.objectid >= block_group->key.objectid +
 423                     block_group->key.offset)
 424                         break;
 425
 426                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 427                         total_found += add_new_free_space(block_group,
 428                                                           fs_info, last,
 429                                                           key.objectid);
 430                         last = key.objectid + key.offset;
 431
 432                         if (total_found > (1024 * 1024 * 2)) {
 433                                 total_found = 0;
 434                                 wake_up(&caching_ctl->wait);
 435                         }
 436                 }
 437                 path->slots[0]++;
 438         }
 439         ret = 0;
 440
 441         total_found += add_new_free_space(block_group, fs_info, last,
 442                                           block_group->key.objectid +
 443                                           block_group->key.offset);
 444         caching_ctl->progress = (u64)-1;
 445
 446         spin_lock(&block_group->lock);
 447         block_group->caching_ctl = NULL;
 448         block_group->cached = BTRFS_CACHE_FINISHED;
 449         spin_unlock(&block_group->lock);
 450
 451 err:
 452         btrfs_free_path(path);
 453         up_read(&fs_info->extent_commit_sem);
 454
 455         free_excluded_extents(extent_root, block_group);
 456
 457         mutex_unlock(&caching_ctl->mutex);
 458 out:
 459         wake_up(&caching_ctl->wait);
 460
 461         put_caching_control(caching_ctl);
 462         btrfs_put_block_group(block_group);
 463 }
 464
 465 static int cache_block_group(struct btrfs_block_group_cache *cache,
 466                              struct btrfs_trans_handle *trans,
 467                              struct btrfs_root *root,
 468                              int load_cache_only)
 469 {
 470         struct btrfs_fs_info *fs_info = cache->fs_info;
 471         struct btrfs_caching_control *caching_ctl;
 472         int ret = 0;
 473
 474         smp_mb();
 475         if (cache->cached != BTRFS_CACHE_NO)
 476                 return 0;
 477
 478         /*
 479          * We can't do the read from on-disk cache during a commit since we need
 480          * to have the normal tree locking.  Also if we are currently trying to
 481          * allocate blocks for the tree root we can't do the fast caching since
 482          * we likely hold important locks.
 483          */
 484         if (trans && (!trans->transaction->in_commit) &&
 485             (root && root != root->fs_info->tree_root) &&
 486             btrfs_test_opt(root, SPACE_CACHE)) {
 487                 spin_lock(&cache->lock);
 488                 if (cache->cached != BTRFS_CACHE_NO) {
 489                         spin_unlock(&cache->lock);
 490                         return 0;
 491                 }
 492                 cache->cached = BTRFS_CACHE_STARTED;
 493                 spin_unlock(&cache->lock);
 494
 495                 ret = load_free_space_cache(fs_info, cache);
 496
 497                 spin_lock(&cache->lock);
 498                 if (ret == 1) {
 499                         cache->cached = BTRFS_CACHE_FINISHED;
 500                         cache->last_byte_to_unpin = (u64)-1;
 501                 } else {
 502                         cache->cached = BTRFS_CACHE_NO;
 503                 }
 504                 spin_unlock(&cache->lock);
 505                 if (ret == 1) {
 506                         free_excluded_extents(fs_info->extent_root, cache);
 507                         return 0;
 508                 }
 509         }
 510
 511         if (load_cache_only)
 512                 return 0;
 513
 514         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 515         BUG_ON(!caching_ctl);
 516
 517         INIT_LIST_HEAD(&caching_ctl->list);
 518         mutex_init(&caching_ctl->mutex);
 519         init_waitqueue_head(&caching_ctl->wait);
 520         caching_ctl->block_group = cache;
 521         caching_ctl->progress = cache->key.objectid;
 522         /* one for caching kthread, one for caching block group list */
 523         atomic_set(&caching_ctl->count, 2);
 524         caching_ctl->work.func = caching_thread;
 525
 526         spin_lock(&cache->lock);
 527         if (cache->cached != BTRFS_CACHE_NO) {
 528                 spin_unlock(&cache->lock);
 529                 kfree(caching_ctl);
 530                 return 0;
 531         }
 532         cache->caching_ctl = caching_ctl;
 533         cache->cached = BTRFS_CACHE_STARTED;
 534         spin_unlock(&cache->lock);
 535
 536         down_write(&fs_info->extent_commit_sem);
 537         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 538         up_write(&fs_info->extent_commit_sem);
 539
 540         btrfs_get_block_group(cache);
 541
 542         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 543
 544         return ret;
 545 }
 546
 547 /*
 548  * return the block group that starts at or after bytenr
 549  */
 550 static struct btrfs_block_group_cache *
 551 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 552 {
 553         struct btrfs_block_group_cache *cache;
 554
 555         cache = block_group_cache_tree_search(info, bytenr, 0);
 556
 557         return cache;
 558 }
 559
 560 /*
 561  * return the block group that contains the given bytenr
 562  */
 563 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 564                                                  struct btrfs_fs_info *info,
 565                                                  u64 bytenr)
 566 {
 567         struct btrfs_block_group_cache *cache;
 568
 569         cache = block_group_cache_tree_search(info, bytenr, 1);
 570
 571         return cache;
 572 }
 573
 574 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 575                                                   u64 flags)
 576 {
 577         struct list_head *head = &info->space_info;
 578         struct btrfs_space_info *found;
 579
 580         flags &= BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_SYSTEM |
 581                  BTRFS_BLOCK_GROUP_METADATA;
 582
 583         rcu_read_lock();
 584         list_for_each_entry_rcu(found, head, list) {
 585                 if (found->flags & flags) {
 586                         rcu_read_unlock();
 587                         return found;
 588                 }
 589         }
 590         rcu_read_unlock();
 591         return NULL;
 592 }
 593
 594 /*
 595  * after adding space to the filesystem, we need to clear the full flags
 596  * on all the space infos.
 597  */
 598 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 599 {
 600         struct list_head *head = &info->space_info;
 601         struct btrfs_space_info *found;
 602
 603         rcu_read_lock();
 604         list_for_each_entry_rcu(found, head, list)
 605                 found->full = 0;
 606         rcu_read_unlock();
 607 }
 608
 609 static u64 div_factor(u64 num, int factor)
 610 {
 611         if (factor == 10)
 612                 return num;
 613         num *= factor;
 614         do_div(num, 10);
 615         return num;
 616 }
 617
 618 static u64 div_factor_fine(u64 num, int factor)
 619 {
 620         if (factor == 100)
 621                 return num;
 622         num *= factor;
 623         do_div(num, 100);
 624         return num;
 625 }
 626
 627 u64 btrfs_find_block_group(struct btrfs_root *root,
 628                            u64 search_start, u64 search_hint, int owner)
 629 {
 630         struct btrfs_block_group_cache *cache;
 631         u64 used;
 632         u64 last = max(search_hint, search_start);
 633         u64 group_start = 0;
 634         int full_search = 0;
 635         int factor = 9;
 636         int wrapped = 0;
 637 again:
 638         while (1) {
 639                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
 640                 if (!cache)
 641                         break;
 642
 643                 spin_lock(&cache->lock);
 644                 last = cache->key.objectid + cache->key.offset;
 645                 used = btrfs_block_group_used(&cache->item);
 646
 647                 if ((full_search || !cache->ro) &&
 648                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
 649                         if (used + cache->pinned + cache->reserved <
 650                             div_factor(cache->key.offset, factor)) {
 651                                 group_start = cache->key.objectid;
 652                                 spin_unlock(&cache->lock);
 653                                 btrfs_put_block_group(cache);
 654                                 goto found;
 655                         }
 656                 }
 657                 spin_unlock(&cache->lock);
 658                 btrfs_put_block_group(cache);
 659                 cond_resched();
 660         }
 661         if (!wrapped) {
 662                 last = search_start;
 663                 wrapped = 1;
 664                 goto again;
 665         }
 666         if (!full_search && factor < 10) {
 667                 last = search_start;
 668                 full_search = 1;
 669                 factor = 10;
 670                 goto again;
 671         }
 672 found:
 673         return group_start;
 674 }
 675
 676 /* simple helper to search for an existing extent at a given offset */
 677 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 678 {
 679         int ret;
 680         struct btrfs_key key;
 681         struct btrfs_path *path;
 682
 683         path = btrfs_alloc_path();
 684         if (!path)
 685                 return -ENOMEM;
 686
 687         key.objectid = start;
 688         key.offset = len;
 689         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 690         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 691                                 0, 0);
 692         btrfs_free_path(path);
 693         return ret;
 694 }
 695
 696 /*
 697  * helper function to lookup reference count and flags of extent.
 698  *
 699  * the head node for delayed ref is used to store the sum of all the
 700  * reference count modifications queued up in the rbtree. the head
 701  * node may also store the extent flags to set. This way you can check
 702  * to see what the reference count and extent flags would be if all of
 703  * the delayed refs are not processed.
 704  */
 705 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 706                              struct btrfs_root *root, u64 bytenr,
 707                              u64 num_bytes, u64 *refs, u64 *flags)
 708 {
 709         struct btrfs_delayed_ref_head *head;
 710         struct btrfs_delayed_ref_root *delayed_refs;
 711         struct btrfs_path *path;
 712         struct btrfs_extent_item *ei;
 713         struct extent_buffer *leaf;
 714         struct btrfs_key key;
 715         u32 item_size;
 716         u64 num_refs;
 717         u64 extent_flags;
 718         int ret;
 719
 720         path = btrfs_alloc_path();
 721         if (!path)
 722                 return -ENOMEM;
 723
 724         key.objectid = bytenr;
 725         key.type = BTRFS_EXTENT_ITEM_KEY;
 726         key.offset = num_bytes;
 727         if (!trans) {
 728                 path->skip_locking = 1;
 729                 path->search_commit_root = 1;
 730         }
 731 again:
 732         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 733                                 &key, path, 0, 0);
 734         if (ret < 0)
 735                 goto out_free;
 736
 737         if (ret == 0) {
 738                 leaf = path->nodes[0];
 739                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 740                 if (item_size >= sizeof(*ei)) {
 741                         ei = btrfs_item_ptr(leaf, path->slots[0],
 742                                             struct btrfs_extent_item);
 743                         num_refs = btrfs_extent_refs(leaf, ei);
 744                         extent_flags = btrfs_extent_flags(leaf, ei);
 745                 } else {
 746 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 747                         struct btrfs_extent_item_v0 *ei0;
 748                         BUG_ON(item_size != sizeof(*ei0));
 749                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 750                                              struct btrfs_extent_item_v0);
 751                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 752                         /* FIXME: this isn't correct for data */
 753                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 754 #else
 755                         BUG();
 756 #endif
 757                 }
 758                 BUG_ON(num_refs == 0);
 759         } else {
 760                 num_refs = 0;
 761                 extent_flags = 0;
 762                 ret = 0;
 763         }
 764
 765         if (!trans)
 766                 goto out;
 767
 768         delayed_refs = &trans->transaction->delayed_refs;
 769         spin_lock(&delayed_refs->lock);
 770         head = btrfs_find_delayed_ref_head(trans, bytenr);
 771         if (head) {
 772                 if (!mutex_trylock(&head->mutex)) {
 773                         atomic_inc(&head->node.refs);
 774                         spin_unlock(&delayed_refs->lock);
 775
 776                         btrfs_release_path(path);
 777
 778                         /*
 779                          * Mutex was contended, block until it's released and try
 780                          * again
 781                          */
 782                         mutex_lock(&head->mutex);
 783                         mutex_unlock(&head->mutex);
 784                         btrfs_put_delayed_ref(&head->node);
 785                         goto again;
 786                 }
 787                 if (head->extent_op && head->extent_op->update_flags)
 788                         extent_flags |= head->extent_op->flags_to_set;
 789                 else
 790                         BUG_ON(num_refs == 0);
 791
 792                 num_refs += head->node.ref_mod;
 793                 mutex_unlock(&head->mutex);
 794         }
 795         spin_unlock(&delayed_refs->lock);
 796 out:
 797         WARN_ON(num_refs == 0);
 798         if (refs)
 799                 *refs = num_refs;
 800         if (flags)
 801                 *flags = extent_flags;
 802 out_free:
 803         btrfs_free_path(path);
 804         return ret;
 805 }
 806
 807 /*
 808  * Back reference rules.  Back refs have three main goals:
 809  *
 810  * 1) differentiate between all holders of references to an extent so that
 811  *    when a reference is dropped we can make sure it was a valid reference
 812  *    before freeing the extent.
 813  *
 814  * 2) Provide enough information to quickly find the holders of an extent
 815  *    if we notice a given block is corrupted or bad.
 816  *
 817  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 818  *    maintenance.  This is actually the same as #2, but with a slightly
 819  *    different use case.
 820  *
 821  * There are two kinds of back refs. The implicit back refs is optimized
 822  * for pointers in non-shared tree blocks. For a given pointer in a block,
 823  * back refs of this kind provide information about the block's owner tree
 824  * and the pointer's key. These information allow us to find the block by
 825  * b-tree searching. The full back refs is for pointers in tree blocks not
 826  * referenced by their owner trees. The location of tree block is recorded
 827  * in the back refs. Actually the full back refs is generic, and can be
 828  * used in all cases the implicit back refs is used. The major shortcoming
 829  * of the full back refs is its overhead. Every time a tree block gets
 830  * COWed, we have to update back refs entry for all pointers in it.
 831  *
 832  * For a newly allocated tree block, we use implicit back refs for
 833  * pointers in it. This means most tree related operations only involve
 834  * implicit back refs. For a tree block created in old transaction, the
 835  * only way to drop a reference to it is COW it. So we can detect the
 836  * event that tree block loses its owner tree's reference and do the
 837  * back refs conversion.
 838  *
 839  * When a tree block is COW'd through a tree, there are four cases:
 840  *
 841  * The reference count of the block is one and the tree is the block's
 842  * owner tree. Nothing to do in this case.
 843  *
 844  * The reference count of the block is one and the tree is not the
 845  * block's owner tree. In this case, full back refs is used for pointers
 846  * in the block. Remove these full back refs, add implicit back refs for
 847  * every pointers in the new block.
 848  *
 849  * The reference count of the block is greater than one and the tree is
 850  * the block's owner tree. In this case, implicit back refs is used for
 851  * pointers in the block. Add full back refs for every pointers in the
 852  * block, increase lower level extents' reference counts. The original
 853  * implicit back refs are entailed to the new block.
 854  *
 855  * The reference count of the block is greater than one and the tree is
 856  * not the block's owner tree. Add implicit back refs for every pointer in
 857  * the new block, increase lower level extents' reference count.
 858  *
 859  * Back Reference Key composing:
 860  *
 861  * The key objectid corresponds to the first byte in the extent,
 862  * The key type is used to differentiate between types of back refs.
 863  * There are different meanings of the key offset for different types
 864  * of back refs.
 865  *
 866  * File extents can be referenced by:
 867  *
 868  * - multiple snapshots, subvolumes, or different generations in one subvol
 869  * - different files inside a single subvolume
 870  * - different offsets inside a file (bookend extents in file.c)
 871  *
 872  * The extent ref structure for the implicit back refs has fields for:
 873  *
 874  * - Objectid of the subvolume root
 875  * - objectid of the file holding the reference
 876  * - original offset in the file
 877  * - how many bookend extents
 878  *
 879  * The key offset for the implicit back refs is hash of the first
 880  * three fields.
 881  *
 882  * The extent ref structure for the full back refs has field for:
 883  *
 884  * - number of pointers in the tree leaf
 885  *
 886  * The key offset for the implicit back refs is the first byte of
 887  * the tree leaf
 888  *
 889  * When a file extent is allocated, The implicit back refs is used.
 890  * the fields are filled in:
 891  *
 892  *     (root_key.objectid, inode objectid, offset in file, 1)
 893  *
 894  * When a file extent is removed file truncation, we find the
 895  * corresponding implicit back refs and check the following fields:
 896  *
 897  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 898  *
 899  * Btree extents can be referenced by:
 900  *
 901  * - Different subvolumes
 902  *
 903  * Both the implicit back refs and the full back refs for tree blocks
 904  * only consist of key. The key offset for the implicit back refs is
 905  * objectid of block's owner tree. The key offset for the full back refs
 906  * is the first byte of parent block.
 907  *
 908  * When implicit back refs is used, information about the lowest key and
 909  * level of the tree block are required. These information are stored in
 910  * tree block info structure.
 911  */
 912
 913 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 914 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 915                                   struct btrfs_root *root,
 916                                   struct btrfs_path *path,
 917                                   u64 owner, u32 extra_size)
 918 {
 919         struct btrfs_extent_item *item;
 920         struct btrfs_extent_item_v0 *ei0;
 921         struct btrfs_extent_ref_v0 *ref0;
 922         struct btrfs_tree_block_info *bi;
 923         struct extent_buffer *leaf;
 924         struct btrfs_key key;
 925         struct btrfs_key found_key;
 926         u32 new_size = sizeof(*item);
 927         u64 refs;
 928         int ret;
 929
 930         leaf = path->nodes[0];
 931         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 932
 933         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 934         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 935                              struct btrfs_extent_item_v0);
 936         refs = btrfs_extent_refs_v0(leaf, ei0);
 937
 938         if (owner == (u64)-1) {
 939                 while (1) {
 940                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 941                                 ret = btrfs_next_leaf(root, path);
 942                                 if (ret < 0)
 943                                         return ret;
 944                                 BUG_ON(ret > 0);
 945                                 leaf = path->nodes[0];
 946                         }
 947                         btrfs_item_key_to_cpu(leaf, &found_key,
 948                                               path->slots[0]);
 949                         BUG_ON(key.objectid != found_key.objectid);
 950                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 951                                 path->slots[0]++;
 952                                 continue;
 953                         }
 954                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 955                                               struct btrfs_extent_ref_v0);
 956                         owner = btrfs_ref_objectid_v0(leaf, ref0);
 957                         break;
 958                 }
 959         }
 960         btrfs_release_path(path);
 961
 962         if (owner < BTRFS_FIRST_FREE_OBJECTID)
 963                 new_size += sizeof(*bi);
 964
 965         new_size -= sizeof(*ei0);
 966         ret = btrfs_search_slot(trans, root, &key, path,
 967                                 new_size + extra_size, 1);
 968         if (ret < 0)
 969                 return ret;
 970         BUG_ON(ret);
 971
 972         ret = btrfs_extend_item(trans, root, path, new_size);
 973
 974         leaf = path->nodes[0];
 975         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
 976         btrfs_set_extent_refs(leaf, item, refs);
 977         /* FIXME: get real generation */
 978         btrfs_set_extent_generation(leaf, item, 0);
 979         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
 980                 btrfs_set_extent_flags(leaf, item,
 981                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
 982                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
 983                 bi = (struct btrfs_tree_block_info *)(item + 1);
 984                 /* FIXME: get first key of the block */
 985                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
 986                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
 987         } else {
 988                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
 989         }
 990         btrfs_mark_buffer_dirty(leaf);
 991         return 0;
 992 }
 993 #endif
 994
 995 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
 996 {
 997         u32 high_crc = ~(u32)0;
 998         u32 low_crc = ~(u32)0;
 999         __le64 lenum;
1000
1001         lenum = cpu_to_le64(root_objectid);
1002         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1003         lenum = cpu_to_le64(owner);
1004         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1005         lenum = cpu_to_le64(offset);
1006         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1007
1008         return ((u64)high_crc << 31) ^ (u64)low_crc;
1009 }
1010
1011 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1012                                      struct btrfs_extent_data_ref *ref)
1013 {
1014         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1015                                     btrfs_extent_data_ref_objectid(leaf, ref),
1016                                     btrfs_extent_data_ref_offset(leaf, ref));
1017 }
1018
1019 static int match_extent_data_ref(struct extent_buffer *leaf,
1020                                  struct btrfs_extent_data_ref *ref,
1021                                  u64 root_objectid, u64 owner, u64 offset)
1022 {
1023         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1024             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1025             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1026                 return 0;
1027         return 1;
1028 }
1029
1030 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1031                                            struct btrfs_root *root,
1032                                            struct btrfs_path *path,
1033                                            u64 bytenr, u64 parent,
1034                                            u64 root_objectid,
1035                                            u64 owner, u64 offset)
1036 {
1037         struct btrfs_key key;
1038         struct btrfs_extent_data_ref *ref;
1039         struct extent_buffer *leaf;
1040         u32 nritems;
1041         int ret;
1042         int recow;
1043         int err = -ENOENT;
1044
1045         key.objectid = bytenr;
1046         if (parent) {
1047                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1048                 key.offset = parent;
1049         } else {
1050                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1051                 key.offset = hash_extent_data_ref(root_objectid,
1052                                                   owner, offset);
1053         }
1054 again:
1055         recow = 0;
1056         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1057         if (ret < 0) {
1058                 err = ret;
1059                 goto fail;
1060         }
1061
1062         if (parent) {
1063                 if (!ret)
1064                         return 0;
1065 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1066                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1067                 btrfs_release_path(path);
1068                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1069                 if (ret < 0) {
1070                         err = ret;
1071                         goto fail;
1072                 }
1073                 if (!ret)
1074                         return 0;
1075 #endif
1076                 goto fail;
1077         }
1078
1079         leaf = path->nodes[0];
1080         nritems = btrfs_header_nritems(leaf);
1081         while (1) {
1082                 if (path->slots[0] >= nritems) {
1083                         ret = btrfs_next_leaf(root, path);
1084                         if (ret < 0)
1085                                 err = ret;
1086                         if (ret)
1087                                 goto fail;
1088
1089                         leaf = path->nodes[0];
1090                         nritems = btrfs_header_nritems(leaf);
1091                         recow = 1;
1092                 }
1093
1094                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1095                 if (key.objectid != bytenr ||
1096                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1097                         goto fail;
1098
1099                 ref = btrfs_item_ptr(leaf, path->slots[0],
1100                                      struct btrfs_extent_data_ref);
1101
1102                 if (match_extent_data_ref(leaf, ref, root_objectid,
1103                                           owner, offset)) {
1104                         if (recow) {
1105                                 btrfs_release_path(path);
1106                                 goto again;
1107                         }
1108                         err = 0;
1109                         break;
1110                 }
1111                 path->slots[0]++;
1112         }
1113 fail:
1114         return err;
1115 }
1116
1117 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1118                                            struct btrfs_root *root,
1119                                            struct btrfs_path *path,
1120                                            u64 bytenr, u64 parent,
1121                                            u64 root_objectid, u64 owner,
1122                                            u64 offset, int refs_to_add)
1123 {
1124         struct btrfs_key key;
1125         struct extent_buffer *leaf;
1126         u32 size;
1127         u32 num_refs;
1128         int ret;
1129
1130         key.objectid = bytenr;
1131         if (parent) {
1132                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1133                 key.offset = parent;
1134                 size = sizeof(struct btrfs_shared_data_ref);
1135         } else {
1136                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1137                 key.offset = hash_extent_data_ref(root_objectid,
1138                                                   owner, offset);
1139                 size = sizeof(struct btrfs_extent_data_ref);
1140         }
1141
1142         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1143         if (ret && ret != -EEXIST)
1144                 goto fail;
1145
1146         leaf = path->nodes[0];
1147         if (parent) {
1148                 struct btrfs_shared_data_ref *ref;
1149                 ref = btrfs_item_ptr(leaf, path->slots[0],
1150                                      struct btrfs_shared_data_ref);
1151                 if (ret == 0) {
1152                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1153                 } else {
1154                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1155                         num_refs += refs_to_add;
1156                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1157                 }
1158         } else {
1159                 struct btrfs_extent_data_ref *ref;
1160                 while (ret == -EEXIST) {
1161                         ref = btrfs_item_ptr(leaf, path->slots[0],
1162                                              struct btrfs_extent_data_ref);
1163                         if (match_extent_data_ref(leaf, ref, root_objectid,
1164                                                   owner, offset))
1165                                 break;
1166                         btrfs_release_path(path);
1167                         key.offset++;
1168                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1169                                                       size);
1170                         if (ret && ret != -EEXIST)
1171                                 goto fail;
1172
1173                         leaf = path->nodes[0];
1174                 }
1175                 ref = btrfs_item_ptr(leaf, path->slots[0],
1176                                      struct btrfs_extent_data_ref);
1177                 if (ret == 0) {
1178                         btrfs_set_extent_data_ref_root(leaf, ref,
1179                                                        root_objectid);
1180                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1181                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1182                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1183                 } else {
1184                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1185                         num_refs += refs_to_add;
1186                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1187                 }
1188         }
1189         btrfs_mark_buffer_dirty(leaf);
1190         ret = 0;
1191 fail:
1192         btrfs_release_path(path);
1193         return ret;
1194 }
1195
1196 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1197                                            struct btrfs_root *root,
1198                                            struct btrfs_path *path,
1199                                            int refs_to_drop)
1200 {
1201         struct btrfs_key key;
1202         struct btrfs_extent_data_ref *ref1 = NULL;
1203         struct btrfs_shared_data_ref *ref2 = NULL;
1204         struct extent_buffer *leaf;
1205         u32 num_refs = 0;
1206         int ret = 0;
1207
1208         leaf = path->nodes[0];
1209         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1210
1211         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1212                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1213                                       struct btrfs_extent_data_ref);
1214                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1215         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1216                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1217                                       struct btrfs_shared_data_ref);
1218                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1219 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1220         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1221                 struct btrfs_extent_ref_v0 *ref0;
1222                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1223                                       struct btrfs_extent_ref_v0);
1224                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1225 #endif
1226         } else {
1227                 BUG();
1228         }
1229
1230         BUG_ON(num_refs < refs_to_drop);
1231         num_refs -= refs_to_drop;
1232
1233         if (num_refs == 0) {
1234                 ret = btrfs_del_item(trans, root, path);
1235         } else {
1236                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1237                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1238                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1239                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1240 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1241                 else {
1242                         struct btrfs_extent_ref_v0 *ref0;
1243                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1244                                         struct btrfs_extent_ref_v0);
1245                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1246                 }
1247 #endif
1248                 btrfs_mark_buffer_dirty(leaf);
1249         }
1250         return ret;
1251 }
1252
1253 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1254                                           struct btrfs_path *path,
1255                                           struct btrfs_extent_inline_ref *iref)
1256 {
1257         struct btrfs_key key;
1258         struct extent_buffer *leaf;
1259         struct btrfs_extent_data_ref *ref1;
1260         struct btrfs_shared_data_ref *ref2;
1261         u32 num_refs = 0;
1262
1263         leaf = path->nodes[0];
1264         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1265         if (iref) {
1266                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1267                     BTRFS_EXTENT_DATA_REF_KEY) {
1268                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1269                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1270                 } else {
1271                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1272                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1273                 }
1274         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1275                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1276                                       struct btrfs_extent_data_ref);
1277                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1278         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1279                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1280                                       struct btrfs_shared_data_ref);
1281                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1282 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1283         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1284                 struct btrfs_extent_ref_v0 *ref0;
1285                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1286                                       struct btrfs_extent_ref_v0);
1287                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1288 #endif
1289         } else {
1290                 WARN_ON(1);
1291         }
1292         return num_refs;
1293 }
1294
1295 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1296                                           struct btrfs_root *root,
1297                                           struct btrfs_path *path,
1298                                           u64 bytenr, u64 parent,
1299                                           u64 root_objectid)
1300 {
1301         struct btrfs_key key;
1302         int ret;
1303
1304         key.objectid = bytenr;
1305         if (parent) {
1306                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1307                 key.offset = parent;
1308         } else {
1309                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1310                 key.offset = root_objectid;
1311         }
1312
1313         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1314         if (ret > 0)
1315                 ret = -ENOENT;
1316 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1317         if (ret == -ENOENT && parent) {
1318                 btrfs_release_path(path);
1319                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1320                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1321                 if (ret > 0)
1322                         ret = -ENOENT;
1323         }
1324 #endif
1325         return ret;
1326 }
1327
1328 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1329                                           struct btrfs_root *root,
1330                                           struct btrfs_path *path,
1331                                           u64 bytenr, u64 parent,
1332                                           u64 root_objectid)
1333 {
1334         struct btrfs_key key;
1335         int ret;
1336
1337         key.objectid = bytenr;
1338         if (parent) {
1339                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1340                 key.offset = parent;
1341         } else {
1342                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1343                 key.offset = root_objectid;
1344         }
1345
1346         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1347         btrfs_release_path(path);
1348         return ret;
1349 }
1350
1351 static inline int extent_ref_type(u64 parent, u64 owner)
1352 {
1353         int type;
1354         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1355                 if (parent > 0)
1356                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1357                 else
1358                         type = BTRFS_TREE_BLOCK_REF_KEY;
1359         } else {
1360                 if (parent > 0)
1361                         type = BTRFS_SHARED_DATA_REF_KEY;
1362                 else
1363                         type = BTRFS_EXTENT_DATA_REF_KEY;
1364         }
1365         return type;
1366 }
1367
1368 static int find_next_key(struct btrfs_path *path, int level,
1369                          struct btrfs_key *key)
1370
1371 {
1372         for (; level < BTRFS_MAX_LEVEL; level++) {
1373                 if (!path->nodes[level])
1374                         break;
1375                 if (path->slots[level] + 1 >=
1376                     btrfs_header_nritems(path->nodes[level]))
1377                         continue;
1378                 if (level == 0)
1379                         btrfs_item_key_to_cpu(path->nodes[level], key,
1380                                               path->slots[level] + 1);
1381                 else
1382                         btrfs_node_key_to_cpu(path->nodes[level], key,
1383                                               path->slots[level] + 1);
1384                 return 0;
1385         }
1386         return 1;
1387 }
1388
1389 /*
1390  * look for inline back ref. if back ref is found, *ref_ret is set
1391  * to the address of inline back ref, and 0 is returned.
1392  *
1393  * if back ref isn't found, *ref_ret is set to the address where it
1394  * should be inserted, and -ENOENT is returned.
1395  *
1396  * if insert is true and there are too many inline back refs, the path
1397  * points to the extent item, and -EAGAIN is returned.
1398  *
1399  * NOTE: inline back refs are ordered in the same way that back ref
1400  *       items in the tree are ordered.
1401  */
1402 static noinline_for_stack
1403 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1404                                  struct btrfs_root *root,
1405                                  struct btrfs_path *path,
1406                                  struct btrfs_extent_inline_ref **ref_ret,
1407                                  u64 bytenr, u64 num_bytes,
1408                                  u64 parent, u64 root_objectid,
1409                                  u64 owner, u64 offset, int insert)
1410 {
1411         struct btrfs_key key;
1412         struct extent_buffer *leaf;
1413         struct btrfs_extent_item *ei;
1414         struct btrfs_extent_inline_ref *iref;
1415         u64 flags;
1416         u64 item_size;
1417         unsigned long ptr;
1418         unsigned long end;
1419         int extra_size;
1420         int type;
1421         int want;
1422         int ret;
1423         int err = 0;
1424
1425         key.objectid = bytenr;
1426         key.type = BTRFS_EXTENT_ITEM_KEY;
1427         key.offset = num_bytes;
1428
1429         want = extent_ref_type(parent, owner);
1430         if (insert) {
1431                 extra_size = btrfs_extent_inline_ref_size(want);
1432                 path->keep_locks = 1;
1433         } else
1434                 extra_size = -1;
1435         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1436         if (ret < 0) {
1437                 err = ret;
1438                 goto out;
1439         }
1440         BUG_ON(ret);
1441
1442         leaf = path->nodes[0];
1443         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1444 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1445         if (item_size < sizeof(*ei)) {
1446                 if (!insert) {
1447                         err = -ENOENT;
1448                         goto out;
1449                 }
1450                 ret = convert_extent_item_v0(trans, root, path, owner,
1451                                              extra_size);
1452                 if (ret < 0) {
1453                         err = ret;
1454                         goto out;
1455                 }
1456                 leaf = path->nodes[0];
1457                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1458         }
1459 #endif
1460         BUG_ON(item_size < sizeof(*ei));
1461
1462         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1463         flags = btrfs_extent_flags(leaf, ei);
1464
1465         ptr = (unsigned long)(ei + 1);
1466         end = (unsigned long)ei + item_size;
1467
1468         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1469                 ptr += sizeof(struct btrfs_tree_block_info);
1470                 BUG_ON(ptr > end);
1471         } else {
1472                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1473         }
1474
1475         err = -ENOENT;
1476         while (1) {
1477                 if (ptr >= end) {
1478                         WARN_ON(ptr > end);
1479                         break;
1480                 }
1481                 iref = (struct btrfs_extent_inline_ref *)ptr;
1482                 type = btrfs_extent_inline_ref_type(leaf, iref);
1483                 if (want < type)
1484                         break;
1485                 if (want > type) {
1486                         ptr += btrfs_extent_inline_ref_size(type);
1487                         continue;
1488                 }
1489
1490                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1491                         struct btrfs_extent_data_ref *dref;
1492                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1493                         if (match_extent_data_ref(leaf, dref, root_objectid,
1494                                                   owner, offset)) {
1495                                 err = 0;
1496                                 break;
1497                         }
1498                         if (hash_extent_data_ref_item(leaf, dref) <
1499                             hash_extent_data_ref(root_objectid, owner, offset))
1500                                 break;
1501                 } else {
1502                         u64 ref_offset;
1503                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1504                         if (parent > 0) {
1505                                 if (parent == ref_offset) {
1506                                         err = 0;
1507                                         break;
1508                                 }
1509                                 if (ref_offset < parent)
1510                                         break;
1511                         } else {
1512                                 if (root_objectid == ref_offset) {
1513                                         err = 0;
1514                                         break;
1515                                 }
1516                                 if (ref_offset < root_objectid)
1517                                         break;
1518                         }
1519                 }
1520                 ptr += btrfs_extent_inline_ref_size(type);
1521         }
1522         if (err == -ENOENT && insert) {
1523                 if (item_size + extra_size >=
1524                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1525                         err = -EAGAIN;
1526                         goto out;
1527                 }
1528                 /*
1529                  * To add new inline back ref, we have to make sure
1530                  * there is no corresponding back ref item.
1531                  * For simplicity, we just do not add new inline back
1532                  * ref if there is any kind of item for this block
1533                  */
1534                 if (find_next_key(path, 0, &key) == 0 &&
1535                     key.objectid == bytenr &&
1536                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1537                         err = -EAGAIN;
1538                         goto out;
1539                 }
1540         }
1541         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1542 out:
1543         if (insert) {
1544                 path->keep_locks = 0;
1545                 btrfs_unlock_up_safe(path, 1);
1546         }
1547         return err;
1548 }
1549
1550 /*
1551  * helper to add new inline back ref
1552  */
1553 static noinline_for_stack
1554 int setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1555                                 struct btrfs_root *root,
1556                                 struct btrfs_path *path,
1557                                 struct btrfs_extent_inline_ref *iref,
1558                                 u64 parent, u64 root_objectid,
1559                                 u64 owner, u64 offset, int refs_to_add,
1560                                 struct btrfs_delayed_extent_op *extent_op)
1561 {
1562         struct extent_buffer *leaf;
1563         struct btrfs_extent_item *ei;
1564         unsigned long ptr;
1565         unsigned long end;
1566         unsigned long item_offset;
1567         u64 refs;
1568         int size;
1569         int type;
1570         int ret;
1571
1572         leaf = path->nodes[0];
1573         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1574         item_offset = (unsigned long)iref - (unsigned long)ei;
1575
1576         type = extent_ref_type(parent, owner);
1577         size = btrfs_extent_inline_ref_size(type);
1578
1579         ret = btrfs_extend_item(trans, root, path, size);
1580
1581         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1582         refs = btrfs_extent_refs(leaf, ei);
1583         refs += refs_to_add;
1584         btrfs_set_extent_refs(leaf, ei, refs);
1585         if (extent_op)
1586                 __run_delayed_extent_op(extent_op, leaf, ei);
1587
1588         ptr = (unsigned long)ei + item_offset;
1589         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1590         if (ptr < end - size)
1591                 memmove_extent_buffer(leaf, ptr + size, ptr,
1592                                       end - size - ptr);
1593
1594         iref = (struct btrfs_extent_inline_ref *)ptr;
1595         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1596         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1597                 struct btrfs_extent_data_ref *dref;
1598                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1599                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1600                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1601                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1602                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1603         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1604                 struct btrfs_shared_data_ref *sref;
1605                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1606                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1607                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1608         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1609                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1610         } else {
1611                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1612         }
1613         btrfs_mark_buffer_dirty(leaf);
1614         return 0;
1615 }
1616
1617 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1618                                  struct btrfs_root *root,
1619                                  struct btrfs_path *path,
1620                                  struct btrfs_extent_inline_ref **ref_ret,
1621                                  u64 bytenr, u64 num_bytes, u64 parent,
1622                                  u64 root_objectid, u64 owner, u64 offset)
1623 {
1624         int ret;
1625
1626         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1627                                            bytenr, num_bytes, parent,
1628                                            root_objectid, owner, offset, 0);
1629         if (ret != -ENOENT)
1630                 return ret;
1631
1632         btrfs_release_path(path);
1633         *ref_ret = NULL;
1634
1635         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1636                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1637                                             root_objectid);
1638         } else {
1639                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1640                                              root_objectid, owner, offset);
1641         }
1642         return ret;
1643 }
1644
1645 /*
1646  * helper to update/remove inline back ref
1647  */
1648 static noinline_for_stack
1649 int update_inline_extent_backref(struct btrfs_trans_handle *trans,
1650                                  struct btrfs_root *root,
1651                                  struct btrfs_path *path,
1652                                  struct btrfs_extent_inline_ref *iref,
1653                                  int refs_to_mod,
1654                                  struct btrfs_delayed_extent_op *extent_op)
1655 {
1656         struct extent_buffer *leaf;
1657         struct btrfs_extent_item *ei;
1658         struct btrfs_extent_data_ref *dref = NULL;
1659         struct btrfs_shared_data_ref *sref = NULL;
1660         unsigned long ptr;
1661         unsigned long end;
1662         u32 item_size;
1663         int size;
1664         int type;
1665         int ret;
1666         u64 refs;
1667
1668         leaf = path->nodes[0];
1669         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1670         refs = btrfs_extent_refs(leaf, ei);
1671         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1672         refs += refs_to_mod;
1673         btrfs_set_extent_refs(leaf, ei, refs);
1674         if (extent_op)
1675                 __run_delayed_extent_op(extent_op, leaf, ei);
1676
1677         type = btrfs_extent_inline_ref_type(leaf, iref);
1678
1679         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1680                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1681                 refs = btrfs_extent_data_ref_count(leaf, dref);
1682         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1683                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1684                 refs = btrfs_shared_data_ref_count(leaf, sref);
1685         } else {
1686                 refs = 1;
1687                 BUG_ON(refs_to_mod != -1);
1688         }
1689
1690         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1691         refs += refs_to_mod;
1692
1693         if (refs > 0) {
1694                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1695                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1696                 else
1697                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1698         } else {
1699                 size =  btrfs_extent_inline_ref_size(type);
1700                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1701                 ptr = (unsigned long)iref;
1702                 end = (unsigned long)ei + item_size;
1703                 if (ptr + size < end)
1704                         memmove_extent_buffer(leaf, ptr, ptr + size,
1705                                               end - ptr - size);
1706                 item_size -= size;
1707                 ret = btrfs_truncate_item(trans, root, path, item_size, 1);
1708         }
1709         btrfs_mark_buffer_dirty(leaf);
1710         return 0;
1711 }
1712
1713 static noinline_for_stack
1714 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1715                                  struct btrfs_root *root,
1716                                  struct btrfs_path *path,
1717                                  u64 bytenr, u64 num_bytes, u64 parent,
1718                                  u64 root_objectid, u64 owner,
1719                                  u64 offset, int refs_to_add,
1720                                  struct btrfs_delayed_extent_op *extent_op)
1721 {
1722         struct btrfs_extent_inline_ref *iref;
1723         int ret;
1724
1725         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1726                                            bytenr, num_bytes, parent,
1727                                            root_objectid, owner, offset, 1);
1728         if (ret == 0) {
1729                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1730                 ret = update_inline_extent_backref(trans, root, path, iref,
1731                                                    refs_to_add, extent_op);
1732         } else if (ret == -ENOENT) {
1733                 ret = setup_inline_extent_backref(trans, root, path, iref,
1734                                                   parent, root_objectid,
1735                                                   owner, offset, refs_to_add,
1736                                                   extent_op);
1737         }
1738         return ret;
1739 }
1740
1741 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1742                                  struct btrfs_root *root,
1743                                  struct btrfs_path *path,
1744                                  u64 bytenr, u64 parent, u64 root_objectid,
1745                                  u64 owner, u64 offset, int refs_to_add)
1746 {
1747         int ret;
1748         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1749                 BUG_ON(refs_to_add != 1);
1750                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1751                                             parent, root_objectid);
1752         } else {
1753                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1754                                              parent, root_objectid,
1755                                              owner, offset, refs_to_add);
1756         }
1757         return ret;
1758 }
1759
1760 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1761                                  struct btrfs_root *root,
1762                                  struct btrfs_path *path,
1763                                  struct btrfs_extent_inline_ref *iref,
1764                                  int refs_to_drop, int is_data)
1765 {
1766         int ret;
1767
1768         BUG_ON(!is_data && refs_to_drop != 1);
1769         if (iref) {
1770                 ret = update_inline_extent_backref(trans, root, path, iref,
1771                                                    -refs_to_drop, NULL);
1772         } else if (is_data) {
1773                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1774         } else {
1775                 ret = btrfs_del_item(trans, root, path);
1776         }
1777         return ret;
1778 }
1779
1780 static int btrfs_issue_discard(struct block_device *bdev,
1781                                 u64 start, u64 len)
1782 {
1783         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1784 }
1785
1786 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1787                                 u64 num_bytes, u64 *actual_bytes)
1788 {
1789         int ret;
1790         u64 discarded_bytes = 0;
1791         struct btrfs_bio *bbio = NULL;
1792
1793
1794         /* Tell the block device(s) that the sectors can be discarded */
1795         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1796                               bytenr, &num_bytes, &bbio, 0);
1797         if (!ret) {
1798                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1799                 int i;
1800
1801
1802                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1803                         if (!stripe->dev->can_discard)
1804                                 continue;
1805
1806                         ret = btrfs_issue_discard(stripe->dev->bdev,
1807                                                   stripe->physical,
1808                                                   stripe->length);
1809                         if (!ret)
1810                                 discarded_bytes += stripe->length;
1811                         else if (ret != -EOPNOTSUPP)
1812                                 break;
1813
1814                         /*
1815                          * Just in case we get back EOPNOTSUPP for some reason,
1816                          * just ignore the return value so we don't screw up
1817                          * people calling discard_extent.
1818                          */
1819                         ret = 0;
1820                 }
1821                 kfree(bbio);
1822         }
1823
1824         if (actual_bytes)
1825                 *actual_bytes = discarded_bytes;
1826
1827
1828         return ret;
1829 }
1830
1831 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1832                          struct btrfs_root *root,
1833                          u64 bytenr, u64 num_bytes, u64 parent,
1834                          u64 root_objectid, u64 owner, u64 offset)
1835 {
1836         int ret;
1837         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1838                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1839
1840         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1841                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
1842                                         parent, root_objectid, (int)owner,
1843                                         BTRFS_ADD_DELAYED_REF, NULL);
1844         } else {
1845                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
1846                                         parent, root_objectid, owner, offset,
1847                                         BTRFS_ADD_DELAYED_REF, NULL);
1848         }
1849         return ret;
1850 }
1851
1852 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1853                                   struct btrfs_root *root,
1854                                   u64 bytenr, u64 num_bytes,
1855                                   u64 parent, u64 root_objectid,
1856                                   u64 owner, u64 offset, int refs_to_add,
1857                                   struct btrfs_delayed_extent_op *extent_op)
1858 {
1859         struct btrfs_path *path;
1860         struct extent_buffer *leaf;
1861         struct btrfs_extent_item *item;
1862         u64 refs;
1863         int ret;
1864         int err = 0;
1865
1866         path = btrfs_alloc_path();
1867         if (!path)
1868                 return -ENOMEM;
1869
1870         path->reada = 1;
1871         path->leave_spinning = 1;
1872         /* this will setup the path even if it fails to insert the back ref */
1873         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1874                                            path, bytenr, num_bytes, parent,
1875                                            root_objectid, owner, offset,
1876                                            refs_to_add, extent_op);
1877         if (ret == 0)
1878                 goto out;
1879
1880         if (ret != -EAGAIN) {
1881                 err = ret;
1882                 goto out;
1883         }
1884
1885         leaf = path->nodes[0];
1886         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1887         refs = btrfs_extent_refs(leaf, item);
1888         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1889         if (extent_op)
1890                 __run_delayed_extent_op(extent_op, leaf, item);
1891
1892         btrfs_mark_buffer_dirty(leaf);
1893         btrfs_release_path(path);
1894
1895         path->reada = 1;
1896         path->leave_spinning = 1;
1897
1898         /* now insert the actual backref */
1899         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1900                                     path, bytenr, parent, root_objectid,
1901                                     owner, offset, refs_to_add);
1902         BUG_ON(ret);
1903 out:
1904         btrfs_free_path(path);
1905         return err;
1906 }
1907
1908 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1909                                 struct btrfs_root *root,
1910                                 struct btrfs_delayed_ref_node *node,
1911                                 struct btrfs_delayed_extent_op *extent_op,
1912                                 int insert_reserved)
1913 {
1914         int ret = 0;
1915         struct btrfs_delayed_data_ref *ref;
1916         struct btrfs_key ins;
1917         u64 parent = 0;
1918         u64 ref_root = 0;
1919         u64 flags = 0;
1920
1921         ins.objectid = node->bytenr;
1922         ins.offset = node->num_bytes;
1923         ins.type = BTRFS_EXTENT_ITEM_KEY;
1924
1925         ref = btrfs_delayed_node_to_data_ref(node);
1926         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1927                 parent = ref->parent;
1928         else
1929                 ref_root = ref->root;
1930
1931         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1932                 if (extent_op) {
1933                         BUG_ON(extent_op->update_key);
1934                         flags |= extent_op->flags_to_set;
1935                 }
1936                 ret = alloc_reserved_file_extent(trans, root,
1937                                                  parent, ref_root, flags,
1938                                                  ref->objectid, ref->offset,
1939                                                  &ins, node->ref_mod);
1940         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1941                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1942                                              node->num_bytes, parent,
1943                                              ref_root, ref->objectid,
1944                                              ref->offset, node->ref_mod,
1945                                              extent_op);
1946         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1947                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1948                                           node->num_bytes, parent,
1949                                           ref_root, ref->objectid,
1950                                           ref->offset, node->ref_mod,
1951                                           extent_op);
1952         } else {
1953                 BUG();
1954         }
1955         return ret;
1956 }
1957
1958 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
1959                                     struct extent_buffer *leaf,
1960                                     struct btrfs_extent_item *ei)
1961 {
1962         u64 flags = btrfs_extent_flags(leaf, ei);
1963         if (extent_op->update_flags) {
1964                 flags |= extent_op->flags_to_set;
1965                 btrfs_set_extent_flags(leaf, ei, flags);
1966         }
1967
1968         if (extent_op->update_key) {
1969                 struct btrfs_tree_block_info *bi;
1970                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
1971                 bi = (struct btrfs_tree_block_info *)(ei + 1);
1972                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
1973         }
1974 }
1975
1976 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
1977                                  struct btrfs_root *root,
1978                                  struct btrfs_delayed_ref_node *node,
1979                                  struct btrfs_delayed_extent_op *extent_op)
1980 {
1981         struct btrfs_key key;
1982         struct btrfs_path *path;
1983         struct btrfs_extent_item *ei;
1984         struct extent_buffer *leaf;
1985         u32 item_size;
1986         int ret;
1987         int err = 0;
1988
1989         path = btrfs_alloc_path();
1990         if (!path)
1991                 return -ENOMEM;
1992
1993         key.objectid = node->bytenr;
1994         key.type = BTRFS_EXTENT_ITEM_KEY;
1995         key.offset = node->num_bytes;
1996
1997         path->reada = 1;
1998         path->leave_spinning = 1;
1999         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2000                                 path, 0, 1);
2001         if (ret < 0) {
2002                 err = ret;
2003                 goto out;
2004         }
2005         if (ret > 0) {
2006                 err = -EIO;
2007                 goto out;
2008         }
2009
2010         leaf = path->nodes[0];
2011         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2012 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2013         if (item_size < sizeof(*ei)) {
2014                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2015                                              path, (u64)-1, 0);
2016                 if (ret < 0) {
2017                         err = ret;
2018                         goto out;
2019                 }
2020                 leaf = path->nodes[0];
2021                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2022         }
2023 #endif
2024         BUG_ON(item_size < sizeof(*ei));
2025         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2026         __run_delayed_extent_op(extent_op, leaf, ei);
2027
2028         btrfs_mark_buffer_dirty(leaf);
2029 out:
2030         btrfs_free_path(path);
2031         return err;
2032 }
2033
2034 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2035                                 struct btrfs_root *root,
2036                                 struct btrfs_delayed_ref_node *node,
2037                                 struct btrfs_delayed_extent_op *extent_op,
2038                                 int insert_reserved)
2039 {
2040         int ret = 0;
2041         struct btrfs_delayed_tree_ref *ref;
2042         struct btrfs_key ins;
2043         u64 parent = 0;
2044         u64 ref_root = 0;
2045
2046         ins.objectid = node->bytenr;
2047         ins.offset = node->num_bytes;
2048         ins.type = BTRFS_EXTENT_ITEM_KEY;
2049
2050         ref = btrfs_delayed_node_to_tree_ref(node);
2051         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2052                 parent = ref->parent;
2053         else
2054                 ref_root = ref->root;
2055
2056         BUG_ON(node->ref_mod != 1);
2057         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2058                 BUG_ON(!extent_op || !extent_op->update_flags ||
2059                        !extent_op->update_key);
2060                 ret = alloc_reserved_tree_block(trans, root,
2061                                                 parent, ref_root,
2062                                                 extent_op->flags_to_set,
2063                                                 &extent_op->key,
2064                                                 ref->level, &ins);
2065         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2066                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2067                                              node->num_bytes, parent, ref_root,
2068                                              ref->level, 0, 1, extent_op);
2069         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2070                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2071                                           node->num_bytes, parent, ref_root,
2072                                           ref->level, 0, 1, extent_op);
2073         } else {
2074                 BUG();
2075         }
2076         return ret;
2077 }
2078
2079 /* helper function to actually process a single delayed ref entry */
2080 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2081                                struct btrfs_root *root,
2082                                struct btrfs_delayed_ref_node *node,
2083                                struct btrfs_delayed_extent_op *extent_op,
2084                                int insert_reserved)
2085 {
2086         int ret;
2087         if (btrfs_delayed_ref_is_head(node)) {
2088                 struct btrfs_delayed_ref_head *head;
2089                 /*
2090                  * we've hit the end of the chain and we were supposed
2091                  * to insert this extent into the tree.  But, it got
2092                  * deleted before we ever needed to insert it, so all
2093                  * we have to do is clean up the accounting
2094                  */
2095                 BUG_ON(extent_op);
2096                 head = btrfs_delayed_node_to_head(node);
2097                 if (insert_reserved) {
2098                         btrfs_pin_extent(root, node->bytenr,
2099                                          node->num_bytes, 1);
2100                         if (head->is_data) {
2101                                 ret = btrfs_del_csums(trans, root,
2102                                                       node->bytenr,
2103                                                       node->num_bytes);
2104                                 BUG_ON(ret);
2105                         }
2106                 }
2107                 mutex_unlock(&head->mutex);
2108                 return 0;
2109         }
2110
2111         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2112             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2113                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2114                                            insert_reserved);
2115         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2116                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2117                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2118                                            insert_reserved);
2119         else
2120                 BUG();
2121         return ret;
2122 }
2123
2124 static noinline struct btrfs_delayed_ref_node *
2125 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2126 {
2127         struct rb_node *node;
2128         struct btrfs_delayed_ref_node *ref;
2129         int action = BTRFS_ADD_DELAYED_REF;
2130 again:
2131         /*
2132          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2133          * this prevents ref count from going down to zero when
2134          * there still are pending delayed ref.
2135          */
2136         node = rb_prev(&head->node.rb_node);
2137         while (1) {
2138                 if (!node)
2139                         break;
2140                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2141                                 rb_node);
2142                 if (ref->bytenr != head->node.bytenr)
2143                         break;
2144                 if (ref->action == action)
2145                         return ref;
2146                 node = rb_prev(node);
2147         }
2148         if (action == BTRFS_ADD_DELAYED_REF) {
2149                 action = BTRFS_DROP_DELAYED_REF;
2150                 goto again;
2151         }
2152         return NULL;
2153 }
2154
2155 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2156                                        struct btrfs_root *root,
2157                                        struct list_head *cluster)
2158 {
2159         struct btrfs_delayed_ref_root *delayed_refs;
2160         struct btrfs_delayed_ref_node *ref;
2161         struct btrfs_delayed_ref_head *locked_ref = NULL;
2162         struct btrfs_delayed_extent_op *extent_op;
2163         int ret;
2164         int count = 0;
2165         int must_insert_reserved = 0;
2166
2167         delayed_refs = &trans->transaction->delayed_refs;
2168         while (1) {
2169                 if (!locked_ref) {
2170                         /* pick a new head ref from the cluster list */
2171                         if (list_empty(cluster))
2172                                 break;
2173
2174                         locked_ref = list_entry(cluster->next,
2175                                      struct btrfs_delayed_ref_head, cluster);
2176
2177                         /* grab the lock that says we are going to process
2178                          * all the refs for this head */
2179                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2180
2181                         /*
2182                          * we may have dropped the spin lock to get the head
2183                          * mutex lock, and that might have given someone else
2184                          * time to free the head.  If that's true, it has been
2185                          * removed from our list and we can move on.
2186                          */
2187                         if (ret == -EAGAIN) {
2188                                 locked_ref = NULL;
2189                                 count++;
2190                                 continue;
2191                         }
2192                 }
2193
2194                 /*
2195                  * record the must insert reserved flag before we
2196                  * drop the spin lock.
2197                  */
2198                 must_insert_reserved = locked_ref->must_insert_reserved;
2199                 locked_ref->must_insert_reserved = 0;
2200
2201                 extent_op = locked_ref->extent_op;
2202                 locked_ref->extent_op = NULL;
2203
2204                 /*
2205                  * locked_ref is the head node, so we have to go one
2206                  * node back for any delayed ref updates
2207                  */
2208                 ref = select_delayed_ref(locked_ref);
2209                 if (!ref) {
2210                         /* All delayed refs have been processed, Go ahead
2211                          * and send the head node to run_one_delayed_ref,
2212                          * so that any accounting fixes can happen
2213                          */
2214                         ref = &locked_ref->node;
2215
2216                         if (extent_op && must_insert_reserved) {
2217                                 kfree(extent_op);
2218                                 extent_op = NULL;
2219                         }
2220
2221                         if (extent_op) {
2222                                 spin_unlock(&delayed_refs->lock);
2223
2224                                 ret = run_delayed_extent_op(trans, root,
2225                                                             ref, extent_op);
2226                                 BUG_ON(ret);
2227                                 kfree(extent_op);
2228
2229                                 cond_resched();
2230                                 spin_lock(&delayed_refs->lock);
2231                                 continue;
2232                         }
2233
2234                         list_del_init(&locked_ref->cluster);
2235                         locked_ref = NULL;
2236                 }
2237
2238                 ref->in_tree = 0;
2239                 rb_erase(&ref->rb_node, &delayed_refs->root);
2240                 delayed_refs->num_entries--;
2241
2242                 spin_unlock(&delayed_refs->lock);
2243
2244                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2245                                           must_insert_reserved);
2246                 BUG_ON(ret);
2247
2248                 btrfs_put_delayed_ref(ref);
2249                 kfree(extent_op);
2250                 count++;
2251
2252                 cond_resched();
2253                 spin_lock(&delayed_refs->lock);
2254         }
2255         return count;
2256 }
2257
2258 /*
2259  * this starts processing the delayed reference count updates and
2260  * extent insertions we have queued up so far.  count can be
2261  * 0, which means to process everything in the tree at the start
2262  * of the run (but not newly added entries), or it can be some target
2263  * number you'd like to process.
2264  */
2265 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2266                            struct btrfs_root *root, unsigned long count)
2267 {
2268         struct rb_node *node;
2269         struct btrfs_delayed_ref_root *delayed_refs;
2270         struct btrfs_delayed_ref_node *ref;
2271         struct list_head cluster;
2272         int ret;
2273         int run_all = count == (unsigned long)-1;
2274         int run_most = 0;
2275
2276         if (root == root->fs_info->extent_root)
2277                 root = root->fs_info->tree_root;
2278
2279         delayed_refs = &trans->transaction->delayed_refs;
2280         INIT_LIST_HEAD(&cluster);
2281 again:
2282         spin_lock(&delayed_refs->lock);
2283         if (count == 0) {
2284                 count = delayed_refs->num_entries * 2;
2285                 run_most = 1;
2286         }
2287         while (1) {
2288                 if (!(run_all || run_most) &&
2289                     delayed_refs->num_heads_ready < 64)
2290                         break;
2291
2292                 /*
2293                  * go find something we can process in the rbtree.  We start at
2294                  * the beginning of the tree, and then build a cluster
2295                  * of refs to process starting at the first one we are able to
2296                  * lock
2297                  */
2298                 ret = btrfs_find_ref_cluster(trans, &cluster,
2299                                              delayed_refs->run_delayed_start);
2300                 if (ret)
2301                         break;
2302
2303                 ret = run_clustered_refs(trans, root, &cluster);
2304                 BUG_ON(ret < 0);
2305
2306                 count -= min_t(unsigned long, ret, count);
2307
2308                 if (count == 0)
2309                         break;
2310         }
2311
2312         if (run_all) {
2313                 node = rb_first(&delayed_refs->root);
2314                 if (!node)
2315                         goto out;
2316                 count = (unsigned long)-1;
2317
2318                 while (node) {
2319                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2320                                        rb_node);
2321                         if (btrfs_delayed_ref_is_head(ref)) {
2322                                 struct btrfs_delayed_ref_head *head;
2323
2324                                 head = btrfs_delayed_node_to_head(ref);
2325                                 atomic_inc(&ref->refs);
2326
2327                                 spin_unlock(&delayed_refs->lock);
2328                                 /*
2329                                  * Mutex was contended, block until it's
2330                                  * released and try again
2331                                  */
2332                                 mutex_lock(&head->mutex);
2333                                 mutex_unlock(&head->mutex);
2334
2335                                 btrfs_put_delayed_ref(ref);
2336                                 cond_resched();
2337                                 goto again;
2338                         }
2339                         node = rb_next(node);
2340                 }
2341                 spin_unlock(&delayed_refs->lock);
2342                 schedule_timeout(1);
2343                 goto again;
2344         }
2345 out:
2346         spin_unlock(&delayed_refs->lock);
2347         return 0;
2348 }
2349
2350 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2351                                 struct btrfs_root *root,
2352                                 u64 bytenr, u64 num_bytes, u64 flags,
2353                                 int is_data)
2354 {
2355         struct btrfs_delayed_extent_op *extent_op;
2356         int ret;
2357
2358         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2359         if (!extent_op)
2360                 return -ENOMEM;
2361
2362         extent_op->flags_to_set = flags;
2363         extent_op->update_flags = 1;
2364         extent_op->update_key = 0;
2365         extent_op->is_data = is_data ? 1 : 0;
2366
2367         ret = btrfs_add_delayed_extent_op(trans, bytenr, num_bytes, extent_op);
2368         if (ret)
2369                 kfree(extent_op);
2370         return ret;
2371 }
2372
2373 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2374                                       struct btrfs_root *root,
2375                                       struct btrfs_path *path,
2376                                       u64 objectid, u64 offset, u64 bytenr)
2377 {
2378         struct btrfs_delayed_ref_head *head;
2379         struct btrfs_delayed_ref_node *ref;
2380         struct btrfs_delayed_data_ref *data_ref;
2381         struct btrfs_delayed_ref_root *delayed_refs;
2382         struct rb_node *node;
2383         int ret = 0;
2384
2385         ret = -ENOENT;
2386         delayed_refs = &trans->transaction->delayed_refs;
2387         spin_lock(&delayed_refs->lock);
2388         head = btrfs_find_delayed_ref_head(trans, bytenr);
2389         if (!head)
2390                 goto out;
2391
2392         if (!mutex_trylock(&head->mutex)) {
2393                 atomic_inc(&head->node.refs);
2394                 spin_unlock(&delayed_refs->lock);
2395
2396                 btrfs_release_path(path);
2397
2398                 /*
2399                  * Mutex was contended, block until it's released and let
2400                  * caller try again
2401                  */
2402                 mutex_lock(&head->mutex);
2403                 mutex_unlock(&head->mutex);
2404                 btrfs_put_delayed_ref(&head->node);
2405                 return -EAGAIN;
2406         }
2407
2408         node = rb_prev(&head->node.rb_node);
2409         if (!node)
2410                 goto out_unlock;
2411
2412         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2413
2414         if (ref->bytenr != bytenr)
2415                 goto out_unlock;
2416
2417         ret = 1;
2418         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2419                 goto out_unlock;
2420
2421         data_ref = btrfs_delayed_node_to_data_ref(ref);
2422
2423         node = rb_prev(node);
2424         if (node) {
2425                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2426                 if (ref->bytenr == bytenr)
2427                         goto out_unlock;
2428         }
2429
2430         if (data_ref->root != root->root_key.objectid ||
2431             data_ref->objectid != objectid || data_ref->offset != offset)
2432                 goto out_unlock;
2433
2434         ret = 0;
2435 out_unlock:
2436         mutex_unlock(&head->mutex);
2437 out:
2438         spin_unlock(&delayed_refs->lock);
2439         return ret;
2440 }
2441
2442 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2443                                         struct btrfs_root *root,
2444                                         struct btrfs_path *path,
2445                                         u64 objectid, u64 offset, u64 bytenr)
2446 {
2447         struct btrfs_root *extent_root = root->fs_info->extent_root;
2448         struct extent_buffer *leaf;
2449         struct btrfs_extent_data_ref *ref;
2450         struct btrfs_extent_inline_ref *iref;
2451         struct btrfs_extent_item *ei;
2452         struct btrfs_key key;
2453         u32 item_size;
2454         int ret;
2455
2456         key.objectid = bytenr;
2457         key.offset = (u64)-1;
2458         key.type = BTRFS_EXTENT_ITEM_KEY;
2459
2460         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2461         if (ret < 0)
2462                 goto out;
2463         BUG_ON(ret == 0);
2464
2465         ret = -ENOENT;
2466         if (path->slots[0] == 0)
2467                 goto out;
2468
2469         path->slots[0]--;
2470         leaf = path->nodes[0];
2471         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2472
2473         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2474                 goto out;
2475
2476         ret = 1;
2477         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2478 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2479         if (item_size < sizeof(*ei)) {
2480                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2481                 goto out;
2482         }
2483 #endif
2484         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2485
2486         if (item_size != sizeof(*ei) +
2487             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2488                 goto out;
2489
2490         if (btrfs_extent_generation(leaf, ei) <=
2491             btrfs_root_last_snapshot(&root->root_item))
2492                 goto out;
2493
2494         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2495         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2496             BTRFS_EXTENT_DATA_REF_KEY)
2497                 goto out;
2498
2499         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2500         if (btrfs_extent_refs(leaf, ei) !=
2501             btrfs_extent_data_ref_count(leaf, ref) ||
2502             btrfs_extent_data_ref_root(leaf, ref) !=
2503             root->root_key.objectid ||
2504             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2505             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2506                 goto out;
2507
2508         ret = 0;
2509 out:
2510         return ret;
2511 }
2512
2513 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2514                           struct btrfs_root *root,
2515                           u64 objectid, u64 offset, u64 bytenr)
2516 {
2517         struct btrfs_path *path;
2518         int ret;
2519         int ret2;
2520
2521         path = btrfs_alloc_path();
2522         if (!path)
2523                 return -ENOENT;
2524
2525         do {
2526                 ret = check_committed_ref(trans, root, path, objectid,
2527                                           offset, bytenr);
2528                 if (ret && ret != -ENOENT)
2529                         goto out;
2530
2531                 ret2 = check_delayed_ref(trans, root, path, objectid,
2532                                          offset, bytenr);
2533         } while (ret2 == -EAGAIN);
2534
2535         if (ret2 && ret2 != -ENOENT) {
2536                 ret = ret2;
2537                 goto out;
2538         }
2539
2540         if (ret != -ENOENT || ret2 != -ENOENT)
2541                 ret = 0;
2542 out:
2543         btrfs_free_path(path);
2544         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2545                 WARN_ON(ret > 0);
2546         return ret;
2547 }
2548
2549 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2550                            struct btrfs_root *root,
2551                            struct extent_buffer *buf,
2552                            int full_backref, int inc)
2553 {
2554         u64 bytenr;
2555         u64 num_bytes;
2556         u64 parent;
2557         u64 ref_root;
2558         u32 nritems;
2559         struct btrfs_key key;
2560         struct btrfs_file_extent_item *fi;
2561         int i;
2562         int level;
2563         int ret = 0;
2564         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2565                             u64, u64, u64, u64, u64, u64);
2566
2567         ref_root = btrfs_header_owner(buf);
2568         nritems = btrfs_header_nritems(buf);
2569         level = btrfs_header_level(buf);
2570
2571         if (!root->ref_cows && level == 0)
2572                 return 0;
2573
2574         if (inc)
2575                 process_func = btrfs_inc_extent_ref;
2576         else
2577                 process_func = btrfs_free_extent;
2578
2579         if (full_backref)
2580                 parent = buf->start;
2581         else
2582                 parent = 0;
2583
2584         for (i = 0; i < nritems; i++) {
2585                 if (level == 0) {
2586                         btrfs_item_key_to_cpu(buf, &key, i);
2587                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2588                                 continue;
2589                         fi = btrfs_item_ptr(buf, i,
2590                                             struct btrfs_file_extent_item);
2591                         if (btrfs_file_extent_type(buf, fi) ==
2592                             BTRFS_FILE_EXTENT_INLINE)
2593                                 continue;
2594                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2595                         if (bytenr == 0)
2596                                 continue;
2597
2598                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2599                         key.offset -= btrfs_file_extent_offset(buf, fi);
2600                         ret = process_func(trans, root, bytenr, num_bytes,
2601                                            parent, ref_root, key.objectid,
2602                                            key.offset);
2603                         if (ret)
2604                                 goto fail;
2605                 } else {
2606                         bytenr = btrfs_node_blockptr(buf, i);
2607                         num_bytes = btrfs_level_size(root, level - 1);
2608                         ret = process_func(trans, root, bytenr, num_bytes,
2609                                            parent, ref_root, level - 1, 0);
2610                         if (ret)
2611                                 goto fail;
2612                 }
2613         }
2614         return 0;
2615 fail:
2616         BUG();
2617         return ret;
2618 }
2619
2620 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2621                   struct extent_buffer *buf, int full_backref)
2622 {
2623         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
2624 }
2625
2626 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2627                   struct extent_buffer *buf, int full_backref)
2628 {
2629         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
2630 }
2631
2632 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2633                                  struct btrfs_root *root,
2634                                  struct btrfs_path *path,
2635                                  struct btrfs_block_group_cache *cache)
2636 {
2637         int ret;
2638         struct btrfs_root *extent_root = root->fs_info->extent_root;
2639         unsigned long bi;
2640         struct extent_buffer *leaf;
2641
2642         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2643         if (ret < 0)
2644                 goto fail;
2645         BUG_ON(ret);
2646
2647         leaf = path->nodes[0];
2648         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2649         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2650         btrfs_mark_buffer_dirty(leaf);
2651         btrfs_release_path(path);
2652 fail:
2653         if (ret)
2654                 return ret;
2655         return 0;
2656
2657 }
2658
2659 static struct btrfs_block_group_cache *
2660 next_block_group(struct btrfs_root *root,
2661                  struct btrfs_block_group_cache *cache)
2662 {
2663         struct rb_node *node;
2664         spin_lock(&root->fs_info->block_group_cache_lock);
2665         node = rb_next(&cache->cache_node);
2666         btrfs_put_block_group(cache);
2667         if (node) {
2668                 cache = rb_entry(node, struct btrfs_block_group_cache,
2669                                  cache_node);
2670                 btrfs_get_block_group(cache);
2671         } else
2672                 cache = NULL;
2673         spin_unlock(&root->fs_info->block_group_cache_lock);
2674         return cache;
2675 }
2676
2677 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2678                             struct btrfs_trans_handle *trans,
2679                             struct btrfs_path *path)
2680 {
2681         struct btrfs_root *root = block_group->fs_info->tree_root;
2682         struct inode *inode = NULL;
2683         u64 alloc_hint = 0;
2684         int dcs = BTRFS_DC_ERROR;
2685         int num_pages = 0;
2686         int retries = 0;
2687         int ret = 0;
2688
2689         /*
2690          * If this block group is smaller than 100 megs don't bother caching the
2691          * block group.
2692          */
2693         if (block_group->key.offset < (100 * 1024 * 1024)) {
2694                 spin_lock(&block_group->lock);
2695                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2696                 spin_unlock(&block_group->lock);
2697                 return 0;
2698         }
2699
2700 again:
2701         inode = lookup_free_space_inode(root, block_group, path);
2702         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2703                 ret = PTR_ERR(inode);
2704                 btrfs_release_path(path);
2705                 goto out;
2706         }
2707
2708         if (IS_ERR(inode)) {
2709                 BUG_ON(retries);
2710                 retries++;
2711
2712                 if (block_group->ro)
2713                         goto out_free;
2714
2715                 ret = create_free_space_inode(root, trans, block_group, path);
2716                 if (ret)
2717                         goto out_free;
2718                 goto again;
2719         }
2720
2721         /* We've already setup this transaction, go ahead and exit */
2722         if (block_group->cache_generation == trans->transid &&
2723             i_size_read(inode)) {
2724                 dcs = BTRFS_DC_SETUP;
2725                 goto out_put;
2726         }
2727
2728         /*
2729          * We want to set the generation to 0, that way if anything goes wrong
2730          * from here on out we know not to trust this cache when we load up next
2731          * time.
2732          */
2733         BTRFS_I(inode)->generation = 0;
2734         ret = btrfs_update_inode(trans, root, inode);
2735         WARN_ON(ret);
2736
2737         if (i_size_read(inode) > 0) {
2738                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2739                                                       inode);
2740                 if (ret)
2741                         goto out_put;
2742         }
2743
2744         spin_lock(&block_group->lock);
2745         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2746                 /* We're not cached, don't bother trying to write stuff out */
2747                 dcs = BTRFS_DC_WRITTEN;
2748                 spin_unlock(&block_group->lock);
2749                 goto out_put;
2750         }
2751         spin_unlock(&block_group->lock);
2752
2753         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2754         if (!num_pages)
2755                 num_pages = 1;
2756
2757         /*
2758          * Just to make absolutely sure we have enough space, we're going to
2759          * preallocate 12 pages worth of space for each block group.  In
2760          * practice we ought to use at most 8, but we need extra space so we can
2761          * add our header and have a terminator between the extents and the
2762          * bitmaps.
2763          */
2764         num_pages *= 16;
2765         num_pages *= PAGE_CACHE_SIZE;
2766
2767         ret = btrfs_check_data_free_space(inode, num_pages);
2768         if (ret)
2769                 goto out_put;
2770
2771         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2772                                               num_pages, num_pages,
2773                                               &alloc_hint);
2774         if (!ret)
2775                 dcs = BTRFS_DC_SETUP;
2776         btrfs_free_reserved_data_space(inode, num_pages);
2777
2778 out_put:
2779         iput(inode);
2780 out_free:
2781         btrfs_release_path(path);
2782 out:
2783         spin_lock(&block_group->lock);
2784         if (!ret)
2785                 block_group->cache_generation = trans->transid;
2786         block_group->disk_cache_state = dcs;
2787         spin_unlock(&block_group->lock);
2788
2789         return ret;
2790 }
2791
2792 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2793                                    struct btrfs_root *root)
2794 {
2795         struct btrfs_block_group_cache *cache;
2796         int err = 0;
2797         struct btrfs_path *path;
2798         u64 last = 0;
2799
2800         path = btrfs_alloc_path();
2801         if (!path)
2802                 return -ENOMEM;
2803
2804 again:
2805         while (1) {
2806                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2807                 while (cache) {
2808                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2809                                 break;
2810                         cache = next_block_group(root, cache);
2811                 }
2812                 if (!cache) {
2813                         if (last == 0)
2814                                 break;
2815                         last = 0;
2816                         continue;
2817                 }
2818                 err = cache_save_setup(cache, trans, path);
2819                 last = cache->key.objectid + cache->key.offset;
2820                 btrfs_put_block_group(cache);
2821         }
2822
2823         while (1) {
2824                 if (last == 0) {
2825                         err = btrfs_run_delayed_refs(trans, root,
2826                                                      (unsigned long)-1);
2827                         BUG_ON(err);
2828                 }
2829
2830                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2831                 while (cache) {
2832                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2833                                 btrfs_put_block_group(cache);
2834                                 goto again;
2835                         }
2836
2837                         if (cache->dirty)
2838                                 break;
2839                         cache = next_block_group(root, cache);
2840                 }
2841                 if (!cache) {
2842                         if (last == 0)
2843                                 break;
2844                         last = 0;
2845                         continue;
2846                 }
2847
2848                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
2849                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
2850                 cache->dirty = 0;
2851                 last = cache->key.objectid + cache->key.offset;
2852
2853                 err = write_one_cache_group(trans, root, path, cache);
2854                 BUG_ON(err);
2855                 btrfs_put_block_group(cache);
2856         }
2857
2858         while (1) {
2859                 /*
2860                  * I don't think this is needed since we're just marking our
2861                  * preallocated extent as written, but just in case it can't
2862                  * hurt.
2863                  */
2864                 if (last == 0) {
2865                         err = btrfs_run_delayed_refs(trans, root,
2866                                                      (unsigned long)-1);
2867                         BUG_ON(err);
2868                 }
2869
2870                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2871                 while (cache) {
2872                         /*
2873                          * Really this shouldn't happen, but it could if we
2874                          * couldn't write the entire preallocated extent and
2875                          * splitting the extent resulted in a new block.
2876                          */
2877                         if (cache->dirty) {
2878                                 btrfs_put_block_group(cache);
2879                                 goto again;
2880                         }
2881                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2882                                 break;
2883                         cache = next_block_group(root, cache);
2884                 }
2885                 if (!cache) {
2886                         if (last == 0)
2887                                 break;
2888                         last = 0;
2889                         continue;
2890                 }
2891
2892                 btrfs_write_out_cache(root, trans, cache, path);
2893
2894                 /*
2895                  * If we didn't have an error then the cache state is still
2896                  * NEED_WRITE, so we can set it to WRITTEN.
2897                  */
2898                 if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
2899                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
2900                 last = cache->key.objectid + cache->key.offset;
2901                 btrfs_put_block_group(cache);
2902         }
2903
2904         btrfs_free_path(path);
2905         return 0;
2906 }
2907
2908 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
2909 {
2910         struct btrfs_block_group_cache *block_group;
2911         int readonly = 0;
2912
2913         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
2914         if (!block_group || block_group->ro)
2915                 readonly = 1;
2916         if (block_group)
2917                 btrfs_put_block_group(block_group);
2918         return readonly;
2919 }
2920
2921 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
2922                              u64 total_bytes, u64 bytes_used,
2923                              struct btrfs_space_info **space_info)
2924 {
2925         struct btrfs_space_info *found;
2926         int i;
2927         int factor;
2928
2929         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
2930                      BTRFS_BLOCK_GROUP_RAID10))
2931                 factor = 2;
2932         else
2933                 factor = 1;
2934
2935         found = __find_space_info(info, flags);
2936         if (found) {
2937                 spin_lock(&found->lock);
2938                 found->total_bytes += total_bytes;
2939                 found->disk_total += total_bytes * factor;
2940                 found->bytes_used += bytes_used;
2941                 found->disk_used += bytes_used * factor;
2942                 found->full = 0;
2943                 spin_unlock(&found->lock);
2944                 *space_info = found;
2945                 return 0;
2946         }
2947         found = kzalloc(sizeof(*found), GFP_NOFS);
2948         if (!found)
2949                 return -ENOMEM;
2950
2951         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
2952                 INIT_LIST_HEAD(&found->block_groups[i]);
2953         init_rwsem(&found->groups_sem);
2954         spin_lock_init(&found->lock);
2955         found->flags = flags & (BTRFS_BLOCK_GROUP_DATA |
2956                                 BTRFS_BLOCK_GROUP_SYSTEM |
2957                                 BTRFS_BLOCK_GROUP_METADATA);
2958         found->total_bytes = total_bytes;
2959         found->disk_total = total_bytes * factor;
2960         found->bytes_used = bytes_used;
2961         found->disk_used = bytes_used * factor;
2962         found->bytes_pinned = 0;
2963         found->bytes_reserved = 0;
2964         found->bytes_readonly = 0;
2965         found->bytes_may_use = 0;
2966         found->full = 0;
2967         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
2968         found->chunk_alloc = 0;
2969         found->flush = 0;
2970         init_waitqueue_head(&found->wait);
2971         *space_info = found;
2972         list_add_rcu(&found->list, &info->space_info);
2973         return 0;
2974 }
2975
2976 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
2977 {
2978         u64 extra_flags = flags & (BTRFS_BLOCK_GROUP_RAID0 |
2979                                    BTRFS_BLOCK_GROUP_RAID1 |
2980                                    BTRFS_BLOCK_GROUP_RAID10 |
2981                                    BTRFS_BLOCK_GROUP_DUP);
2982         if (extra_flags) {
2983                 if (flags & BTRFS_BLOCK_GROUP_DATA)
2984                         fs_info->avail_data_alloc_bits |= extra_flags;
2985                 if (flags & BTRFS_BLOCK_GROUP_METADATA)
2986                         fs_info->avail_metadata_alloc_bits |= extra_flags;
2987                 if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
2988                         fs_info->avail_system_alloc_bits |= extra_flags;
2989         }
2990 }
2991
2992 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
2993 {
2994         /*
2995          * we add in the count of missing devices because we want
2996          * to make sure that any RAID levels on a degraded FS
2997          * continue to be honored.
2998          */
2999         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3000                 root->fs_info->fs_devices->missing_devices;
3001
3002         if (num_devices == 1)
3003                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3004         if (num_devices < 4)
3005                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3006
3007         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3008             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3009                       BTRFS_BLOCK_GROUP_RAID10))) {
3010                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3011         }
3012
3013         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3014             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3015                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3016         }
3017
3018         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3019             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3020              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3021              (flags & BTRFS_BLOCK_GROUP_DUP)))
3022                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3023         return flags;
3024 }
3025
3026 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3027 {
3028         if (flags & BTRFS_BLOCK_GROUP_DATA)
3029                 flags |= root->fs_info->avail_data_alloc_bits &
3030                          root->fs_info->data_alloc_profile;
3031         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3032                 flags |= root->fs_info->avail_system_alloc_bits &
3033                          root->fs_info->system_alloc_profile;
3034         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3035                 flags |= root->fs_info->avail_metadata_alloc_bits &
3036                          root->fs_info->metadata_alloc_profile;
3037         return btrfs_reduce_alloc_profile(root, flags);
3038 }
3039
3040 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3041 {
3042         u64 flags;
3043
3044         if (data)
3045                 flags = BTRFS_BLOCK_GROUP_DATA;
3046         else if (root == root->fs_info->chunk_root)
3047                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3048         else
3049                 flags = BTRFS_BLOCK_GROUP_METADATA;
3050
3051         return get_alloc_profile(root, flags);
3052 }
3053
3054 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3055 {
3056         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3057                                                        BTRFS_BLOCK_GROUP_DATA);
3058 }
3059
3060 /*
3061  * This will check the space that the inode allocates from to make sure we have
3062  * enough space for bytes.
3063  */
3064 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3065 {
3066         struct btrfs_space_info *data_sinfo;
3067         struct btrfs_root *root = BTRFS_I(inode)->root;
3068         u64 used;
3069         int ret = 0, committed = 0, alloc_chunk = 1;
3070
3071         /* make sure bytes are sectorsize aligned */
3072         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3073
3074         if (root == root->fs_info->tree_root ||
3075             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3076                 alloc_chunk = 0;
3077                 committed = 1;
3078         }
3079
3080         data_sinfo = BTRFS_I(inode)->space_info;
3081         if (!data_sinfo)
3082                 goto alloc;
3083
3084 again:
3085         /* make sure we have enough space to handle the data first */
3086         spin_lock(&data_sinfo->lock);
3087         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3088                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3089                 data_sinfo->bytes_may_use;
3090
3091         if (used + bytes > data_sinfo->total_bytes) {
3092                 struct btrfs_trans_handle *trans;
3093
3094                 /*
3095                  * if we don't have enough free bytes in this space then we need
3096                  * to alloc a new chunk.
3097                  */
3098                 if (!data_sinfo->full && alloc_chunk) {
3099                         u64 alloc_target;
3100
3101                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3102                         spin_unlock(&data_sinfo->lock);
3103 alloc:
3104                         alloc_target = btrfs_get_alloc_profile(root, 1);
3105                         trans = btrfs_join_transaction(root);
3106                         if (IS_ERR(trans))
3107                                 return PTR_ERR(trans);
3108
3109                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3110                                              bytes + 2 * 1024 * 1024,
3111                                              alloc_target,
3112                                              CHUNK_ALLOC_NO_FORCE);
3113                         btrfs_end_transaction(trans, root);
3114                         if (ret < 0) {
3115                                 if (ret != -ENOSPC)
3116                                         return ret;
3117                                 else
3118                                         goto commit_trans;
3119                         }
3120
3121                         if (!data_sinfo) {
3122                                 btrfs_set_inode_space_info(root, inode);
3123                                 data_sinfo = BTRFS_I(inode)->space_info;
3124                         }
3125                         goto again;
3126                 }
3127
3128                 /*
3129                  * If we have less pinned bytes than we want to allocate then
3130                  * don't bother committing the transaction, it won't help us.
3131                  */
3132                 if (data_sinfo->bytes_pinned < bytes)
3133                         committed = 1;
3134                 spin_unlock(&data_sinfo->lock);
3135
3136                 /* commit the current transaction and try again */
3137 commit_trans:
3138                 if (!committed &&
3139                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3140                         committed = 1;
3141                         trans = btrfs_join_transaction(root);
3142                         if (IS_ERR(trans))
3143                                 return PTR_ERR(trans);
3144                         ret = btrfs_commit_transaction(trans, root);
3145                         if (ret)
3146                                 return ret;
3147                         goto again;
3148                 }
3149
3150                 return -ENOSPC;
3151         }
3152         data_sinfo->bytes_may_use += bytes;
3153         spin_unlock(&data_sinfo->lock);
3154
3155         return 0;
3156 }
3157
3158 /*
3159  * Called if we need to clear a data reservation for this inode.
3160  */
3161 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3162 {
3163         struct btrfs_root *root = BTRFS_I(inode)->root;
3164         struct btrfs_space_info *data_sinfo;
3165
3166         /* make sure bytes are sectorsize aligned */
3167         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3168
3169         data_sinfo = BTRFS_I(inode)->space_info;
3170         spin_lock(&data_sinfo->lock);
3171         data_sinfo->bytes_may_use -= bytes;
3172         spin_unlock(&data_sinfo->lock);
3173 }
3174
3175 static void force_metadata_allocation(struct btrfs_fs_info *info)
3176 {
3177         struct list_head *head = &info->space_info;
3178         struct btrfs_space_info *found;
3179
3180         rcu_read_lock();
3181         list_for_each_entry_rcu(found, head, list) {
3182                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3183                         found->force_alloc = CHUNK_ALLOC_FORCE;
3184         }
3185         rcu_read_unlock();
3186 }
3187
3188 static int should_alloc_chunk(struct btrfs_root *root,
3189                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3190                               int force)
3191 {
3192         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3193         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3194         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3195         u64 thresh;
3196
3197         if (force == CHUNK_ALLOC_FORCE)
3198                 return 1;
3199
3200         /*
3201          * We need to take into account the global rsv because for all intents
3202          * and purposes it's used space.  Don't worry about locking the
3203          * global_rsv, it doesn't change except when the transaction commits.
3204          */
3205         num_allocated += global_rsv->size;
3206
3207         /*
3208          * in limited mode, we want to have some free space up to
3209          * about 1% of the FS size.
3210          */
3211         if (force == CHUNK_ALLOC_LIMITED) {
3212                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3213                 thresh = max_t(u64, 64 * 1024 * 1024,
3214                                div_factor_fine(thresh, 1));
3215
3216                 if (num_bytes - num_allocated < thresh)
3217                         return 1;
3218         }
3219
3220         /*
3221          * we have two similar checks here, one based on percentage
3222          * and once based on a hard number of 256MB.  The idea
3223          * is that if we have a good amount of free
3224          * room, don't allocate a chunk.  A good mount is
3225          * less than 80% utilized of the chunks we have allocated,
3226          * or more than 256MB free
3227          */
3228         if (num_allocated + alloc_bytes + 256 * 1024 * 1024 < num_bytes)
3229                 return 0;
3230
3231         if (num_allocated + alloc_bytes < div_factor(num_bytes, 8))
3232                 return 0;
3233
3234         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3235
3236         /* 256MB or 5% of the FS */
3237         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 5));
3238
3239         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 3))
3240                 return 0;
3241         return 1;
3242 }
3243
3244 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3245                           struct btrfs_root *extent_root, u64 alloc_bytes,
3246                           u64 flags, int force)
3247 {
3248         struct btrfs_space_info *space_info;
3249         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3250         int wait_for_alloc = 0;
3251         int ret = 0;
3252
3253         flags = btrfs_reduce_alloc_profile(extent_root, flags);
3254
3255         space_info = __find_space_info(extent_root->fs_info, flags);
3256         if (!space_info) {
3257                 ret = update_space_info(extent_root->fs_info, flags,
3258                                         0, 0, &space_info);
3259                 BUG_ON(ret);
3260         }
3261         BUG_ON(!space_info);
3262
3263 again:
3264         spin_lock(&space_info->lock);
3265         if (space_info->force_alloc)
3266                 force = space_info->force_alloc;
3267         if (space_info->full) {
3268                 spin_unlock(&space_info->lock);
3269                 return 0;
3270         }
3271
3272         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3273                 spin_unlock(&space_info->lock);
3274                 return 0;
3275         } else if (space_info->chunk_alloc) {
3276                 wait_for_alloc = 1;
3277         } else {
3278                 space_info->chunk_alloc = 1;
3279         }
3280
3281         spin_unlock(&space_info->lock);
3282
3283         mutex_lock(&fs_info->chunk_mutex);
3284
3285         /*
3286          * The chunk_mutex is held throughout the entirety of a chunk
3287          * allocation, so once we've acquired the chunk_mutex we know that the
3288          * other guy is done and we need to recheck and see if we should
3289          * allocate.
3290          */
3291         if (wait_for_alloc) {
3292                 mutex_unlock(&fs_info->chunk_mutex);
3293                 wait_for_alloc = 0;
3294                 goto again;
3295         }
3296
3297         /*
3298          * If we have mixed data/metadata chunks we want to make sure we keep
3299          * allocating mixed chunks instead of individual chunks.
3300          */
3301         if (btrfs_mixed_space_info(space_info))
3302                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3303
3304         /*
3305          * if we're doing a data chunk, go ahead and make sure that
3306          * we keep a reasonable number of metadata chunks allocated in the
3307          * FS as well.
3308          */
3309         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3310                 fs_info->data_chunk_allocations++;
3311                 if (!(fs_info->data_chunk_allocations %
3312                       fs_info->metadata_ratio))
3313                         force_metadata_allocation(fs_info);
3314         }
3315
3316         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3317         if (ret < 0 && ret != -ENOSPC)
3318                 goto out;
3319
3320         spin_lock(&space_info->lock);
3321         if (ret)
3322                 space_info->full = 1;
3323         else
3324                 ret = 1;
3325
3326         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3327         space_info->chunk_alloc = 0;
3328         spin_unlock(&space_info->lock);
3329 out:
3330         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3331         return ret;
3332 }
3333
3334 /*
3335  * shrink metadata reservation for delalloc
3336  */
3337 static int shrink_delalloc(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         struct btrfs_trans_handle *trans;
3343         u64 reserved;
3344         u64 max_reclaim;
3345         u64 reclaimed = 0;
3346         long time_left;
3347         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3348         int loops = 0;
3349         unsigned long progress;
3350
3351         trans = (struct btrfs_trans_handle *)current->journal_info;
3352         block_rsv = &root->fs_info->delalloc_block_rsv;
3353         space_info = block_rsv->space_info;
3354
3355         smp_mb();
3356         reserved = space_info->bytes_may_use;
3357         progress = space_info->reservation_progress;
3358
3359         if (reserved == 0)
3360                 return 0;
3361
3362         smp_mb();
3363         if (root->fs_info->delalloc_bytes == 0) {
3364                 if (trans)
3365                         return 0;
3366                 btrfs_wait_ordered_extents(root, 0, 0);
3367                 return 0;
3368         }
3369
3370         max_reclaim = min(reserved, to_reclaim);
3371         nr_pages = max_t(unsigned long, nr_pages,
3372                          max_reclaim >> PAGE_CACHE_SHIFT);
3373         while (loops < 1024) {
3374                 /* have the flusher threads jump in and do some IO */
3375                 smp_mb();
3376                 nr_pages = min_t(unsigned long, nr_pages,
3377                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3378                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages);
3379
3380                 spin_lock(&space_info->lock);
3381                 if (reserved > space_info->bytes_may_use)
3382                         reclaimed += reserved - space_info->bytes_may_use;
3383                 reserved = space_info->bytes_may_use;
3384                 spin_unlock(&space_info->lock);
3385
3386                 loops++;
3387
3388                 if (reserved == 0 || reclaimed >= max_reclaim)
3389                         break;
3390
3391                 if (trans && trans->transaction->blocked)
3392                         return -EAGAIN;
3393
3394                 if (wait_ordered && !trans) {
3395                         btrfs_wait_ordered_extents(root, 0, 0);
3396                 } else {
3397                         time_left = schedule_timeout_interruptible(1);
3398
3399                         /* We were interrupted, exit */
3400                         if (time_left)
3401                                 break;
3402                 }
3403
3404                 /* we've kicked the IO a few times, if anything has been freed,
3405                  * exit.  There is no sense in looping here for a long time
3406                  * when we really need to commit the transaction, or there are
3407                  * just too many writers without enough free space
3408                  */
3409
3410                 if (loops > 3) {
3411                         smp_mb();
3412                         if (progress != space_info->reservation_progress)
3413                                 break;
3414                 }
3415
3416         }
3417
3418         return reclaimed >= to_reclaim;
3419 }
3420
3421 /**
3422  * maybe_commit_transaction - possibly commit the transaction if its ok to
3423  * @root - the root we're allocating for
3424  * @bytes - the number of bytes we want to reserve
3425  * @force - force the commit
3426  *
3427  * This will check to make sure that committing the transaction will actually
3428  * get us somewhere and then commit the transaction if it does.  Otherwise it
3429  * will return -ENOSPC.
3430  */
3431 static int may_commit_transaction(struct btrfs_root *root,
3432                                   struct btrfs_space_info *space_info,
3433                                   u64 bytes, int force)
3434 {
3435         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3436         struct btrfs_trans_handle *trans;
3437
3438         trans = (struct btrfs_trans_handle *)current->journal_info;
3439         if (trans)
3440                 return -EAGAIN;
3441
3442         if (force)
3443                 goto commit;
3444
3445         /* See if there is enough pinned space to make this reservation */
3446         spin_lock(&space_info->lock);
3447         if (space_info->bytes_pinned >= bytes) {
3448                 spin_unlock(&space_info->lock);
3449                 goto commit;
3450         }
3451         spin_unlock(&space_info->lock);
3452
3453         /*
3454          * See if there is some space in the delayed insertion reservation for
3455          * this reservation.
3456          */
3457         if (space_info != delayed_rsv->space_info)
3458                 return -ENOSPC;
3459
3460         spin_lock(&delayed_rsv->lock);
3461         if (delayed_rsv->size < bytes) {
3462                 spin_unlock(&delayed_rsv->lock);
3463                 return -ENOSPC;
3464         }
3465         spin_unlock(&delayed_rsv->lock);
3466
3467 commit:
3468         trans = btrfs_join_transaction(root);
3469         if (IS_ERR(trans))
3470                 return -ENOSPC;
3471
3472         return btrfs_commit_transaction(trans, root);
3473 }
3474
3475 /**
3476  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3477  * @root - the root we're allocating for
3478  * @block_rsv - the block_rsv we're allocating for
3479  * @orig_bytes - the number of bytes we want
3480  * @flush - wether or not we can flush to make our reservation
3481  *
3482  * This will reserve orgi_bytes number of bytes from the space info associated
3483  * with the block_rsv.  If there is not enough space it will make an attempt to
3484  * flush out space to make room.  It will do this by flushing delalloc if
3485  * possible or committing the transaction.  If flush is 0 then no attempts to
3486  * regain reservations will be made and this will fail if there is not enough
3487  * space already.
3488  */
3489 static int reserve_metadata_bytes(struct btrfs_root *root,
3490                                   struct btrfs_block_rsv *block_rsv,
3491                                   u64 orig_bytes, int flush)
3492 {
3493         struct btrfs_space_info *space_info = block_rsv->space_info;
3494         u64 used;
3495         u64 num_bytes = orig_bytes;
3496         int retries = 0;
3497         int ret = 0;
3498         bool committed = false;
3499         bool flushing = false;
3500         bool wait_ordered = false;
3501
3502 again:
3503         ret = 0;
3504         spin_lock(&space_info->lock);
3505         /*
3506          * We only want to wait if somebody other than us is flushing and we are
3507          * actually alloed to flush.
3508          */
3509         while (flush && !flushing && space_info->flush) {
3510                 spin_unlock(&space_info->lock);
3511                 /*
3512                  * If we have a trans handle we can't wait because the flusher
3513                  * may have to commit the transaction, which would mean we would
3514                  * deadlock since we are waiting for the flusher to finish, but
3515                  * hold the current transaction open.
3516                  */
3517                 if (current->journal_info)
3518                         return -EAGAIN;
3519                 ret = wait_event_interruptible(space_info->wait,
3520                                                !space_info->flush);
3521                 /* Must have been interrupted, return */
3522                 if (ret)
3523                         return -EINTR;
3524
3525                 spin_lock(&space_info->lock);
3526         }
3527
3528         ret = -ENOSPC;
3529         used = space_info->bytes_used + space_info->bytes_reserved +
3530                 space_info->bytes_pinned + space_info->bytes_readonly +
3531                 space_info->bytes_may_use;
3532
3533         /*
3534          * The idea here is that we've not already over-reserved the block group
3535          * then we can go ahead and save our reservation first and then start
3536          * flushing if we need to.  Otherwise if we've already overcommitted
3537          * lets start flushing stuff first and then come back and try to make
3538          * our reservation.
3539          */
3540         if (used <= space_info->total_bytes) {
3541                 if (used + orig_bytes <= space_info->total_bytes) {
3542                         space_info->bytes_may_use += orig_bytes;
3543                         ret = 0;
3544                 } else {
3545                         /*
3546                          * Ok set num_bytes to orig_bytes since we aren't
3547                          * overocmmitted, this way we only try and reclaim what
3548                          * we need.
3549                          */
3550                         num_bytes = orig_bytes;
3551                 }
3552         } else {
3553                 /*
3554                  * Ok we're over committed, set num_bytes to the overcommitted
3555                  * amount plus the amount of bytes that we need for this
3556                  * reservation.
3557                  */
3558                 wait_ordered = true;
3559                 num_bytes = used - space_info->total_bytes +
3560                         (orig_bytes * (retries + 1));
3561         }
3562
3563         if (ret) {
3564                 u64 profile = btrfs_get_alloc_profile(root, 0);
3565                 u64 avail;
3566
3567                 /*
3568                  * If we have a lot of space that's pinned, don't bother doing
3569                  * the overcommit dance yet and just commit the transaction.
3570                  */
3571                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3572                 do_div(avail, 10);
3573                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3574                         space_info->flush = 1;
3575                         flushing = true;
3576                         spin_unlock(&space_info->lock);
3577                         ret = may_commit_transaction(root, space_info,
3578                                                      orig_bytes, 1);
3579                         if (ret)
3580                                 goto out;
3581                         committed = true;
3582                         goto again;
3583                 }
3584
3585                 spin_lock(&root->fs_info->free_chunk_lock);
3586                 avail = root->fs_info->free_chunk_space;
3587
3588                 /*
3589                  * If we have dup, raid1 or raid10 then only half of the free
3590                  * space is actually useable.
3591                  */
3592                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3593                                BTRFS_BLOCK_GROUP_RAID1 |
3594                                BTRFS_BLOCK_GROUP_RAID10))
3595                         avail >>= 1;
3596
3597                 /*
3598                  * If we aren't flushing don't let us overcommit too much, say
3599                  * 1/8th of the space.  If we can flush, let it overcommit up to
3600                  * 1/2 of the space.
3601                  */
3602                 if (flush)
3603                         avail >>= 3;
3604                 else
3605                         avail >>= 1;
3606                  spin_unlock(&root->fs_info->free_chunk_lock);
3607
3608                 if (used + num_bytes < space_info->total_bytes + avail) {
3609                         space_info->bytes_may_use += orig_bytes;
3610                         ret = 0;
3611                 } else {
3612                         wait_ordered = true;
3613                 }
3614         }
3615
3616         /*
3617          * Couldn't make our reservation, save our place so while we're trying
3618          * to reclaim space we can actually use it instead of somebody else
3619          * stealing it from us.
3620          */
3621         if (ret && flush) {
3622                 flushing = true;
3623                 space_info->flush = 1;
3624         }
3625
3626         spin_unlock(&space_info->lock);
3627
3628         if (!ret || !flush)
3629                 goto out;
3630
3631         /*
3632          * We do synchronous shrinking since we don't actually unreserve
3633          * metadata until after the IO is completed.
3634          */
3635         ret = shrink_delalloc(root, num_bytes, wait_ordered);
3636         if (ret < 0)
3637                 goto out;
3638
3639         ret = 0;
3640
3641         /*
3642          * So if we were overcommitted it's possible that somebody else flushed
3643          * out enough space and we simply didn't have enough space to reclaim,
3644          * so go back around and try again.
3645          */
3646         if (retries < 2) {
3647                 wait_ordered = true;
3648                 retries++;
3649                 goto again;
3650         }
3651
3652         ret = -ENOSPC;
3653         if (committed)
3654                 goto out;
3655
3656         ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3657         if (!ret) {
3658                 committed = true;
3659                 goto again;
3660         }
3661
3662 out:
3663         if (flushing) {
3664                 spin_lock(&space_info->lock);
3665                 space_info->flush = 0;
3666                 wake_up_all(&space_info->wait);
3667                 spin_unlock(&space_info->lock);
3668         }
3669         return ret;
3670 }
3671
3672 static struct btrfs_block_rsv *get_block_rsv(struct btrfs_trans_handle *trans,
3673                                              struct btrfs_root *root)
3674 {
3675         struct btrfs_block_rsv *block_rsv = NULL;
3676
3677         if (root->ref_cows || root == root->fs_info->csum_root)
3678                 block_rsv = trans->block_rsv;
3679
3680         if (!block_rsv)
3681                 block_rsv = root->block_rsv;
3682
3683         if (!block_rsv)
3684                 block_rsv = &root->fs_info->empty_block_rsv;
3685
3686         return block_rsv;
3687 }
3688
3689 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3690                                u64 num_bytes)
3691 {
3692         int ret = -ENOSPC;
3693         spin_lock(&block_rsv->lock);
3694         if (block_rsv->reserved >= num_bytes) {
3695                 block_rsv->reserved -= num_bytes;
3696                 if (block_rsv->reserved < block_rsv->size)
3697                         block_rsv->full = 0;
3698                 ret = 0;
3699         }
3700         spin_unlock(&block_rsv->lock);
3701         return ret;
3702 }
3703
3704 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3705                                 u64 num_bytes, int update_size)
3706 {
3707         spin_lock(&block_rsv->lock);
3708         block_rsv->reserved += num_bytes;
3709         if (update_size)
3710                 block_rsv->size += num_bytes;
3711         else if (block_rsv->reserved >= block_rsv->size)
3712                 block_rsv->full = 1;
3713         spin_unlock(&block_rsv->lock);
3714 }
3715
3716 static void block_rsv_release_bytes(struct btrfs_block_rsv *block_rsv,
3717                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3718 {
3719         struct btrfs_space_info *space_info = block_rsv->space_info;
3720
3721         spin_lock(&block_rsv->lock);
3722         if (num_bytes == (u64)-1)
3723                 num_bytes = block_rsv->size;
3724         block_rsv->size -= num_bytes;
3725         if (block_rsv->reserved >= block_rsv->size) {
3726                 num_bytes = block_rsv->reserved - block_rsv->size;
3727                 block_rsv->reserved = block_rsv->size;
3728                 block_rsv->full = 1;
3729         } else {
3730                 num_bytes = 0;
3731         }
3732         spin_unlock(&block_rsv->lock);
3733
3734         if (num_bytes > 0) {
3735                 if (dest) {
3736                         spin_lock(&dest->lock);
3737                         if (!dest->full) {
3738                                 u64 bytes_to_add;
3739
3740                                 bytes_to_add = dest->size - dest->reserved;
3741                                 bytes_to_add = min(num_bytes, bytes_to_add);
3742                                 dest->reserved += bytes_to_add;
3743                                 if (dest->reserved >= dest->size)
3744                                         dest->full = 1;
3745                                 num_bytes -= bytes_to_add;
3746                         }
3747                         spin_unlock(&dest->lock);
3748                 }
3749                 if (num_bytes) {
3750                         spin_lock(&space_info->lock);
3751                         space_info->bytes_may_use -= num_bytes;
3752                         space_info->reservation_progress++;
3753                         spin_unlock(&space_info->lock);
3754                 }
3755         }
3756 }
3757
3758 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3759                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3760 {
3761         int ret;
3762
3763         ret = block_rsv_use_bytes(src, num_bytes);
3764         if (ret)
3765                 return ret;
3766
3767         block_rsv_add_bytes(dst, num_bytes, 1);
3768         return 0;
3769 }
3770
3771 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3772 {
3773         memset(rsv, 0, sizeof(*rsv));
3774         spin_lock_init(&rsv->lock);
3775 }
3776
3777 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3778 {
3779         struct btrfs_block_rsv *block_rsv;
3780         struct btrfs_fs_info *fs_info = root->fs_info;
3781
3782         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3783         if (!block_rsv)
3784                 return NULL;
3785
3786         btrfs_init_block_rsv(block_rsv);
3787         block_rsv->space_info = __find_space_info(fs_info,
3788                                                   BTRFS_BLOCK_GROUP_METADATA);
3789         return block_rsv;
3790 }
3791
3792 void btrfs_free_block_rsv(struct btrfs_root *root,
3793                           struct btrfs_block_rsv *rsv)
3794 {
3795         btrfs_block_rsv_release(root, rsv, (u64)-1);
3796         kfree(rsv);
3797 }
3798
3799 int btrfs_block_rsv_add(struct btrfs_root *root,
3800                         struct btrfs_block_rsv *block_rsv,
3801                         u64 num_bytes)
3802 {
3803         int ret;
3804
3805         if (num_bytes == 0)
3806                 return 0;
3807
3808         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, 1);
3809         if (!ret) {
3810                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3811                 return 0;
3812         }
3813
3814         return ret;
3815 }
3816
3817 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
3818                                 struct btrfs_block_rsv *block_rsv,
3819                                 u64 num_bytes)
3820 {
3821         int ret;
3822
3823         if (num_bytes == 0)
3824                 return 0;
3825
3826         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, 0);
3827         if (!ret) {
3828                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
3829                 return 0;
3830         }
3831
3832         return ret;
3833 }
3834
3835 int btrfs_block_rsv_check(struct btrfs_root *root,
3836                           struct btrfs_block_rsv *block_rsv, int min_factor)
3837 {
3838         u64 num_bytes = 0;
3839         int ret = -ENOSPC;
3840
3841         if (!block_rsv)
3842                 return 0;
3843
3844         spin_lock(&block_rsv->lock);
3845         num_bytes = div_factor(block_rsv->size, min_factor);
3846         if (block_rsv->reserved >= num_bytes)
3847                 ret = 0;
3848         spin_unlock(&block_rsv->lock);
3849
3850         return ret;
3851 }
3852
3853 int btrfs_block_rsv_refill(struct btrfs_root *root,
3854                           struct btrfs_block_rsv *block_rsv,
3855                           u64 min_reserved)
3856 {
3857         u64 num_bytes = 0;
3858         int ret = -ENOSPC;
3859
3860         if (!block_rsv)
3861                 return 0;
3862
3863         spin_lock(&block_rsv->lock);
3864         num_bytes = min_reserved;
3865         if (block_rsv->reserved >= num_bytes)
3866                 ret = 0;
3867         else
3868                 num_bytes -= block_rsv->reserved;
3869         spin_unlock(&block_rsv->lock);
3870
3871         if (!ret)
3872                 return 0;
3873
3874         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, 1);
3875         if (!ret) {
3876                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
3877                 return 0;
3878         }
3879
3880         return ret;
3881 }
3882
3883 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
3884                             struct btrfs_block_rsv *dst_rsv,
3885                             u64 num_bytes)
3886 {
3887         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
3888 }
3889
3890 void btrfs_block_rsv_release(struct btrfs_root *root,
3891                              struct btrfs_block_rsv *block_rsv,
3892                              u64 num_bytes)
3893 {
3894         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3895         if (global_rsv->full || global_rsv == block_rsv ||
3896             block_rsv->space_info != global_rsv->space_info)
3897                 global_rsv = NULL;
3898         block_rsv_release_bytes(block_rsv, global_rsv, num_bytes);
3899 }
3900
3901 /*
3902  * helper to calculate size of global block reservation.
3903  * the desired value is sum of space used by extent tree,
3904  * checksum tree and root tree
3905  */
3906 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
3907 {
3908         struct btrfs_space_info *sinfo;
3909         u64 num_bytes;
3910         u64 meta_used;
3911         u64 data_used;
3912         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
3913
3914         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
3915         spin_lock(&sinfo->lock);
3916         data_used = sinfo->bytes_used;
3917         spin_unlock(&sinfo->lock);
3918
3919         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3920         spin_lock(&sinfo->lock);
3921         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
3922                 data_used = 0;
3923         meta_used = sinfo->bytes_used;
3924         spin_unlock(&sinfo->lock);
3925
3926         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
3927                     csum_size * 2;
3928         num_bytes += div64_u64(data_used + meta_used, 50);
3929
3930         if (num_bytes * 3 > meta_used)
3931                 num_bytes = div64_u64(meta_used, 3);
3932
3933         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
3934 }
3935
3936 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
3937 {
3938         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
3939         struct btrfs_space_info *sinfo = block_rsv->space_info;
3940         u64 num_bytes;
3941
3942         num_bytes = calc_global_metadata_size(fs_info);
3943
3944         spin_lock(&block_rsv->lock);
3945         spin_lock(&sinfo->lock);
3946
3947         block_rsv->size = num_bytes;
3948
3949         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
3950                     sinfo->bytes_reserved + sinfo->bytes_readonly +
3951                     sinfo->bytes_may_use;
3952
3953         if (sinfo->total_bytes > num_bytes) {
3954                 num_bytes = sinfo->total_bytes - num_bytes;
3955                 block_rsv->reserved += num_bytes;
3956                 sinfo->bytes_may_use += num_bytes;
3957         }
3958
3959         if (block_rsv->reserved >= block_rsv->size) {
3960                 num_bytes = block_rsv->reserved - block_rsv->size;
3961                 sinfo->bytes_may_use -= num_bytes;
3962                 sinfo->reservation_progress++;
3963                 block_rsv->reserved = block_rsv->size;
3964                 block_rsv->full = 1;
3965         }
3966
3967         spin_unlock(&sinfo->lock);
3968         spin_unlock(&block_rsv->lock);
3969 }
3970
3971 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
3972 {
3973         struct btrfs_space_info *space_info;
3974
3975         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3976         fs_info->chunk_block_rsv.space_info = space_info;
3977
3978         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
3979         fs_info->global_block_rsv.space_info = space_info;
3980         fs_info->delalloc_block_rsv.space_info = space_info;
3981         fs_info->trans_block_rsv.space_info = space_info;
3982         fs_info->empty_block_rsv.space_info = space_info;
3983         fs_info->delayed_block_rsv.space_info = space_info;
3984
3985         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
3986         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
3987         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
3988         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
3989         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
3990
3991         update_global_block_rsv(fs_info);
3992 }
3993
3994 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
3995 {
3996         block_rsv_release_bytes(&fs_info->global_block_rsv, NULL, (u64)-1);
3997         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
3998         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
3999         WARN_ON(fs_info->trans_block_rsv.size > 0);
4000         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4001         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4002         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4003         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4004         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4005 }
4006
4007 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4008                                   struct btrfs_root *root)
4009 {
4010         if (!trans->bytes_reserved)
4011                 return;
4012
4013         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4014         trans->bytes_reserved = 0;
4015 }
4016
4017 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4018                                   struct inode *inode)
4019 {
4020         struct btrfs_root *root = BTRFS_I(inode)->root;
4021         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4022         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4023
4024         /*
4025          * We need to hold space in order to delete our orphan item once we've
4026          * added it, so this takes the reservation so we can release it later
4027          * when we are truly done with the orphan item.
4028          */
4029         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4030         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4031 }
4032
4033 void btrfs_orphan_release_metadata(struct inode *inode)
4034 {
4035         struct btrfs_root *root = BTRFS_I(inode)->root;
4036         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4037         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4038 }
4039
4040 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4041                                 struct btrfs_pending_snapshot *pending)
4042 {
4043         struct btrfs_root *root = pending->root;
4044         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4045         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4046         /*
4047          * two for root back/forward refs, two for directory entries
4048          * and one for root of the snapshot.
4049          */
4050         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4051         dst_rsv->space_info = src_rsv->space_info;
4052         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4053 }
4054
4055 /**
4056  * drop_outstanding_extent - drop an outstanding extent
4057  * @inode: the inode we're dropping the extent for
4058  *
4059  * This is called when we are freeing up an outstanding extent, either called
4060  * after an error or after an extent is written.  This will return the number of
4061  * reserved extents that need to be freed.  This must be called with
4062  * BTRFS_I(inode)->lock held.
4063  */
4064 static unsigned drop_outstanding_extent(struct inode *inode)
4065 {
4066         unsigned drop_inode_space = 0;
4067         unsigned dropped_extents = 0;
4068
4069         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4070         BTRFS_I(inode)->outstanding_extents--;
4071
4072         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4073             BTRFS_I(inode)->delalloc_meta_reserved) {
4074                 drop_inode_space = 1;
4075                 BTRFS_I(inode)->delalloc_meta_reserved = 0;
4076         }
4077
4078         /*
4079          * If we have more or the same amount of outsanding extents than we have
4080          * reserved then we need to leave the reserved extents count alone.
4081          */
4082         if (BTRFS_I(inode)->outstanding_extents >=
4083             BTRFS_I(inode)->reserved_extents)
4084                 return drop_inode_space;
4085
4086         dropped_extents = BTRFS_I(inode)->reserved_extents -
4087                 BTRFS_I(inode)->outstanding_extents;
4088         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4089         return dropped_extents + drop_inode_space;
4090 }
4091
4092 /**
4093  * calc_csum_metadata_size - return the amount of metada space that must be
4094  *      reserved/free'd for the given bytes.
4095  * @inode: the inode we're manipulating
4096  * @num_bytes: the number of bytes in question
4097  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4098  *
4099  * This adjusts the number of csum_bytes in the inode and then returns the
4100  * correct amount of metadata that must either be reserved or freed.  We
4101  * calculate how many checksums we can fit into one leaf and then divide the
4102  * number of bytes that will need to be checksumed by this value to figure out
4103  * how many checksums will be required.  If we are adding bytes then the number
4104  * may go up and we will return the number of additional bytes that must be
4105  * reserved.  If it is going down we will return the number of bytes that must
4106  * be freed.
4107  *
4108  * This must be called with BTRFS_I(inode)->lock held.
4109  */
4110 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4111                                    int reserve)
4112 {
4113         struct btrfs_root *root = BTRFS_I(inode)->root;
4114         u64 csum_size;
4115         int num_csums_per_leaf;
4116         int num_csums;
4117         int old_csums;
4118
4119         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4120             BTRFS_I(inode)->csum_bytes == 0)
4121                 return 0;
4122
4123         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4124         if (reserve)
4125                 BTRFS_I(inode)->csum_bytes += num_bytes;
4126         else
4127                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4128         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4129         num_csums_per_leaf = (int)div64_u64(csum_size,
4130                                             sizeof(struct btrfs_csum_item) +
4131                                             sizeof(struct btrfs_disk_key));
4132         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4133         num_csums = num_csums + num_csums_per_leaf - 1;
4134         num_csums = num_csums / num_csums_per_leaf;
4135
4136         old_csums = old_csums + num_csums_per_leaf - 1;
4137         old_csums = old_csums / num_csums_per_leaf;
4138
4139         /* No change, no need to reserve more */
4140         if (old_csums == num_csums)
4141                 return 0;
4142
4143         if (reserve)
4144                 return btrfs_calc_trans_metadata_size(root,
4145                                                       num_csums - old_csums);
4146
4147         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4148 }
4149
4150 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4151 {
4152         struct btrfs_root *root = BTRFS_I(inode)->root;
4153         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4154         u64 to_reserve = 0;
4155         unsigned nr_extents = 0;
4156         int flush = 1;
4157         int ret;
4158
4159         if (btrfs_is_free_space_inode(root, inode))
4160                 flush = 0;
4161
4162         if (flush && btrfs_transaction_in_commit(root->fs_info))
4163                 schedule_timeout(1);
4164
4165         num_bytes = ALIGN(num_bytes, root->sectorsize);
4166
4167         spin_lock(&BTRFS_I(inode)->lock);
4168         BTRFS_I(inode)->outstanding_extents++;
4169
4170         if (BTRFS_I(inode)->outstanding_extents >
4171             BTRFS_I(inode)->reserved_extents) {
4172                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4173                         BTRFS_I(inode)->reserved_extents;
4174                 BTRFS_I(inode)->reserved_extents += nr_extents;
4175         }
4176
4177         /*
4178          * Add an item to reserve for updating the inode when we complete the
4179          * delalloc io.
4180          */
4181         if (!BTRFS_I(inode)->delalloc_meta_reserved) {
4182                 nr_extents++;
4183                 BTRFS_I(inode)->delalloc_meta_reserved = 1;
4184         }
4185
4186         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4187         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4188         spin_unlock(&BTRFS_I(inode)->lock);
4189
4190         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4191         if (ret) {
4192                 u64 to_free = 0;
4193                 unsigned dropped;
4194
4195                 spin_lock(&BTRFS_I(inode)->lock);
4196                 dropped = drop_outstanding_extent(inode);
4197                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4198                 spin_unlock(&BTRFS_I(inode)->lock);
4199                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4200
4201                 /*
4202                  * Somebody could have come in and twiddled with the
4203                  * reservation, so if we have to free more than we would have
4204                  * reserved from this reservation go ahead and release those
4205                  * bytes.
4206                  */
4207                 to_free -= to_reserve;
4208                 if (to_free)
4209                         btrfs_block_rsv_release(root, block_rsv, to_free);
4210                 return ret;
4211         }
4212
4213         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4214
4215         return 0;
4216 }
4217
4218 /**
4219  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4220  * @inode: the inode to release the reservation for
4221  * @num_bytes: the number of bytes we're releasing
4222  *
4223  * This will release the metadata reservation for an inode.  This can be called
4224  * once we complete IO for a given set of bytes to release their metadata
4225  * reservations.
4226  */
4227 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4228 {
4229         struct btrfs_root *root = BTRFS_I(inode)->root;
4230         u64 to_free = 0;
4231         unsigned dropped;
4232
4233         num_bytes = ALIGN(num_bytes, root->sectorsize);
4234         spin_lock(&BTRFS_I(inode)->lock);
4235         dropped = drop_outstanding_extent(inode);
4236
4237         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4238         spin_unlock(&BTRFS_I(inode)->lock);
4239         if (dropped > 0)
4240                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4241
4242         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4243                                 to_free);
4244 }
4245
4246 /**
4247  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4248  * @inode: inode we're writing to
4249  * @num_bytes: the number of bytes we want to allocate
4250  *
4251  * This will do the following things
4252  *
4253  * o reserve space in the data space info for num_bytes
4254  * o reserve space in the metadata space info based on number of outstanding
4255  *   extents and how much csums will be needed
4256  * o add to the inodes ->delalloc_bytes
4257  * o add it to the fs_info's delalloc inodes list.
4258  *
4259  * This will return 0 for success and -ENOSPC if there is no space left.
4260  */
4261 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4262 {
4263         int ret;
4264
4265         ret = btrfs_check_data_free_space(inode, num_bytes);
4266         if (ret)
4267                 return ret;
4268
4269         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4270         if (ret) {
4271                 btrfs_free_reserved_data_space(inode, num_bytes);
4272                 return ret;
4273         }
4274
4275         return 0;
4276 }
4277
4278 /**
4279  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4280  * @inode: inode we're releasing space for
4281  * @num_bytes: the number of bytes we want to free up
4282  *
4283  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4284  * called in the case that we don't need the metadata AND data reservations
4285  * anymore.  So if there is an error or we insert an inline extent.
4286  *
4287  * This function will release the metadata space that was not used and will
4288  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4289  * list if there are no delalloc bytes left.
4290  */
4291 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4292 {
4293         btrfs_delalloc_release_metadata(inode, num_bytes);
4294         btrfs_free_reserved_data_space(inode, num_bytes);
4295 }
4296
4297 static int update_block_group(struct btrfs_trans_handle *trans,
4298                               struct btrfs_root *root,
4299                               u64 bytenr, u64 num_bytes, int alloc)
4300 {
4301         struct btrfs_block_group_cache *cache = NULL;
4302         struct btrfs_fs_info *info = root->fs_info;
4303         u64 total = num_bytes;
4304         u64 old_val;
4305         u64 byte_in_group;
4306         int factor;
4307
4308         /* block accounting for super block */
4309         spin_lock(&info->delalloc_lock);
4310         old_val = btrfs_super_bytes_used(info->super_copy);
4311         if (alloc)
4312                 old_val += num_bytes;
4313         else
4314                 old_val -= num_bytes;
4315         btrfs_set_super_bytes_used(info->super_copy, old_val);
4316         spin_unlock(&info->delalloc_lock);
4317
4318         while (total) {
4319                 cache = btrfs_lookup_block_group(info, bytenr);
4320                 if (!cache)
4321                         return -1;
4322                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4323                                     BTRFS_BLOCK_GROUP_RAID1 |
4324                                     BTRFS_BLOCK_GROUP_RAID10))
4325                         factor = 2;
4326                 else
4327                         factor = 1;
4328                 /*
4329                  * If this block group has free space cache written out, we
4330                  * need to make sure to load it if we are removing space.  This
4331                  * is because we need the unpinning stage to actually add the
4332                  * space back to the block group, otherwise we will leak space.
4333                  */
4334                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4335                         cache_block_group(cache, trans, NULL, 1);
4336
4337                 byte_in_group = bytenr - cache->key.objectid;
4338                 WARN_ON(byte_in_group > cache->key.offset);
4339
4340                 spin_lock(&cache->space_info->lock);
4341                 spin_lock(&cache->lock);
4342
4343                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4344                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4345                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4346
4347                 cache->dirty = 1;
4348                 old_val = btrfs_block_group_used(&cache->item);
4349                 num_bytes = min(total, cache->key.offset - byte_in_group);
4350                 if (alloc) {
4351                         old_val += num_bytes;
4352                         btrfs_set_block_group_used(&cache->item, old_val);
4353                         cache->reserved -= num_bytes;
4354                         cache->space_info->bytes_reserved -= num_bytes;
4355                         cache->space_info->bytes_used += num_bytes;
4356                         cache->space_info->disk_used += num_bytes * factor;
4357                         spin_unlock(&cache->lock);
4358                         spin_unlock(&cache->space_info->lock);
4359                 } else {
4360                         old_val -= num_bytes;
4361                         btrfs_set_block_group_used(&cache->item, old_val);
4362                         cache->pinned += num_bytes;
4363                         cache->space_info->bytes_pinned += num_bytes;
4364                         cache->space_info->bytes_used -= num_bytes;
4365                         cache->space_info->disk_used -= num_bytes * factor;
4366                         spin_unlock(&cache->lock);
4367                         spin_unlock(&cache->space_info->lock);
4368
4369                         set_extent_dirty(info->pinned_extents,
4370                                          bytenr, bytenr + num_bytes - 1,
4371                                          GFP_NOFS | __GFP_NOFAIL);
4372                 }
4373                 btrfs_put_block_group(cache);
4374                 total -= num_bytes;
4375                 bytenr += num_bytes;
4376         }
4377         return 0;
4378 }
4379
4380 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4381 {
4382         struct btrfs_block_group_cache *cache;
4383         u64 bytenr;
4384
4385         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4386         if (!cache)
4387                 return 0;
4388
4389         bytenr = cache->key.objectid;
4390         btrfs_put_block_group(cache);
4391
4392         return bytenr;
4393 }
4394
4395 static int pin_down_extent(struct btrfs_root *root,
4396                            struct btrfs_block_group_cache *cache,
4397                            u64 bytenr, u64 num_bytes, int reserved)
4398 {
4399         spin_lock(&cache->space_info->lock);
4400         spin_lock(&cache->lock);
4401         cache->pinned += num_bytes;
4402         cache->space_info->bytes_pinned += num_bytes;
4403         if (reserved) {
4404                 cache->reserved -= num_bytes;
4405                 cache->space_info->bytes_reserved -= num_bytes;
4406         }
4407         spin_unlock(&cache->lock);
4408         spin_unlock(&cache->space_info->lock);
4409
4410         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4411                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4412         return 0;
4413 }
4414
4415 /*
4416  * this function must be called within transaction
4417  */
4418 int btrfs_pin_extent(struct btrfs_root *root,
4419                      u64 bytenr, u64 num_bytes, int reserved)
4420 {
4421         struct btrfs_block_group_cache *cache;
4422
4423         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4424         BUG_ON(!cache);
4425
4426         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4427
4428         btrfs_put_block_group(cache);
4429         return 0;
4430 }
4431
4432 /*
4433  * this function must be called within transaction
4434  */
4435 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4436                                     struct btrfs_root *root,
4437                                     u64 bytenr, u64 num_bytes)
4438 {
4439         struct btrfs_block_group_cache *cache;
4440
4441         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4442         BUG_ON(!cache);
4443
4444         /*
4445          * pull in the free space cache (if any) so that our pin
4446          * removes the free space from the cache.  We have load_only set
4447          * to one because the slow code to read in the free extents does check
4448          * the pinned extents.
4449          */
4450         cache_block_group(cache, trans, root, 1);
4451
4452         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4453
4454         /* remove us from the free space cache (if we're there at all) */
4455         btrfs_remove_free_space(cache, bytenr, num_bytes);
4456         btrfs_put_block_group(cache);
4457         return 0;
4458 }
4459
4460 /**
4461  * btrfs_update_reserved_bytes - update the block_group and space info counters
4462  * @cache:      The cache we are manipulating
4463  * @num_bytes:  The number of bytes in question
4464  * @reserve:    One of the reservation enums
4465  *
4466  * This is called by the allocator when it reserves space, or by somebody who is
4467  * freeing space that was never actually used on disk.  For example if you
4468  * reserve some space for a new leaf in transaction A and before transaction A
4469  * commits you free that leaf, you call this with reserve set to 0 in order to
4470  * clear the reservation.
4471  *
4472  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4473  * ENOSPC accounting.  For data we handle the reservation through clearing the
4474  * delalloc bits in the io_tree.  We have to do this since we could end up
4475  * allocating less disk space for the amount of data we have reserved in the
4476  * case of compression.
4477  *
4478  * If this is a reservation and the block group has become read only we cannot
4479  * make the reservation and return -EAGAIN, otherwise this function always
4480  * succeeds.
4481  */
4482 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4483                                        u64 num_bytes, int reserve)
4484 {
4485         struct btrfs_space_info *space_info = cache->space_info;
4486         int ret = 0;
4487         spin_lock(&space_info->lock);
4488         spin_lock(&cache->lock);
4489         if (reserve != RESERVE_FREE) {
4490                 if (cache->ro) {
4491                         ret = -EAGAIN;
4492                 } else {
4493                         cache->reserved += num_bytes;
4494                         space_info->bytes_reserved += num_bytes;
4495                         if (reserve == RESERVE_ALLOC) {
4496                                 BUG_ON(space_info->bytes_may_use < num_bytes);
4497                                 space_info->bytes_may_use -= num_bytes;
4498                         }
4499                 }
4500         } else {
4501                 if (cache->ro)
4502                         space_info->bytes_readonly += num_bytes;
4503                 cache->reserved -= num_bytes;
4504                 space_info->bytes_reserved -= num_bytes;
4505                 space_info->reservation_progress++;
4506         }
4507         spin_unlock(&cache->lock);
4508         spin_unlock(&space_info->lock);
4509         return ret;
4510 }
4511
4512 int btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4513                                 struct btrfs_root *root)
4514 {
4515         struct btrfs_fs_info *fs_info = root->fs_info;
4516         struct btrfs_caching_control *next;
4517         struct btrfs_caching_control *caching_ctl;
4518         struct btrfs_block_group_cache *cache;
4519
4520         down_write(&fs_info->extent_commit_sem);
4521
4522         list_for_each_entry_safe(caching_ctl, next,
4523                                  &fs_info->caching_block_groups, list) {
4524                 cache = caching_ctl->block_group;
4525                 if (block_group_cache_done(cache)) {
4526                         cache->last_byte_to_unpin = (u64)-1;
4527                         list_del_init(&caching_ctl->list);
4528                         put_caching_control(caching_ctl);
4529                 } else {
4530                         cache->last_byte_to_unpin = caching_ctl->progress;
4531                 }
4532         }
4533
4534         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4535                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4536         else
4537                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4538
4539         up_write(&fs_info->extent_commit_sem);
4540
4541         update_global_block_rsv(fs_info);
4542         return 0;
4543 }
4544
4545 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4546 {
4547         struct btrfs_fs_info *fs_info = root->fs_info;
4548         struct btrfs_block_group_cache *cache = NULL;
4549         u64 len;
4550
4551         while (start <= end) {
4552                 if (!cache ||
4553                     start >= cache->key.objectid + cache->key.offset) {
4554                         if (cache)
4555                                 btrfs_put_block_group(cache);
4556                         cache = btrfs_lookup_block_group(fs_info, start);
4557                         BUG_ON(!cache);
4558                 }
4559
4560                 len = cache->key.objectid + cache->key.offset - start;
4561                 len = min(len, end + 1 - start);
4562
4563                 if (start < cache->last_byte_to_unpin) {
4564                         len = min(len, cache->last_byte_to_unpin - start);
4565                         btrfs_add_free_space(cache, start, len);
4566                 }
4567
4568                 start += len;
4569
4570                 spin_lock(&cache->space_info->lock);
4571                 spin_lock(&cache->lock);
4572                 cache->pinned -= len;
4573                 cache->space_info->bytes_pinned -= len;
4574                 if (cache->ro)
4575                         cache->space_info->bytes_readonly += len;
4576                 spin_unlock(&cache->lock);
4577                 spin_unlock(&cache->space_info->lock);
4578         }
4579
4580         if (cache)
4581                 btrfs_put_block_group(cache);
4582         return 0;
4583 }
4584
4585 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4586                                struct btrfs_root *root)
4587 {
4588         struct btrfs_fs_info *fs_info = root->fs_info;
4589         struct extent_io_tree *unpin;
4590         u64 start;
4591         u64 end;
4592         int ret;
4593
4594         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4595                 unpin = &fs_info->freed_extents[1];
4596         else
4597                 unpin = &fs_info->freed_extents[0];
4598
4599         while (1) {
4600                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4601                                             EXTENT_DIRTY);
4602                 if (ret)
4603                         break;
4604
4605                 if (btrfs_test_opt(root, DISCARD))
4606                         ret = btrfs_discard_extent(root, start,
4607                                                    end + 1 - start, NULL);
4608
4609                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4610                 unpin_extent_range(root, start, end);
4611                 cond_resched();
4612         }
4613
4614         return 0;
4615 }
4616
4617 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4618                                 struct btrfs_root *root,
4619                                 u64 bytenr, u64 num_bytes, u64 parent,
4620                                 u64 root_objectid, u64 owner_objectid,
4621                                 u64 owner_offset, int refs_to_drop,
4622                                 struct btrfs_delayed_extent_op *extent_op)
4623 {
4624         struct btrfs_key key;
4625         struct btrfs_path *path;
4626         struct btrfs_fs_info *info = root->fs_info;
4627         struct btrfs_root *extent_root = info->extent_root;
4628         struct extent_buffer *leaf;
4629         struct btrfs_extent_item *ei;
4630         struct btrfs_extent_inline_ref *iref;
4631         int ret;
4632         int is_data;
4633         int extent_slot = 0;
4634         int found_extent = 0;
4635         int num_to_del = 1;
4636         u32 item_size;
4637         u64 refs;
4638
4639         path = btrfs_alloc_path();
4640         if (!path)
4641                 return -ENOMEM;
4642
4643         path->reada = 1;
4644         path->leave_spinning = 1;
4645
4646         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4647         BUG_ON(!is_data && refs_to_drop != 1);
4648
4649         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4650                                     bytenr, num_bytes, parent,
4651                                     root_objectid, owner_objectid,
4652                                     owner_offset);
4653         if (ret == 0) {
4654                 extent_slot = path->slots[0];
4655                 while (extent_slot >= 0) {
4656                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4657                                               extent_slot);
4658                         if (key.objectid != bytenr)
4659                                 break;
4660                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4661                             key.offset == num_bytes) {
4662                                 found_extent = 1;
4663                                 break;
4664                         }
4665                         if (path->slots[0] - extent_slot > 5)
4666                                 break;
4667                         extent_slot--;
4668                 }
4669 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4670                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4671                 if (found_extent && item_size < sizeof(*ei))
4672                         found_extent = 0;
4673 #endif
4674                 if (!found_extent) {
4675                         BUG_ON(iref);
4676                         ret = remove_extent_backref(trans, extent_root, path,
4677                                                     NULL, refs_to_drop,
4678                                                     is_data);
4679                         BUG_ON(ret);
4680                         btrfs_release_path(path);
4681                         path->leave_spinning = 1;
4682
4683                         key.objectid = bytenr;
4684                         key.type = BTRFS_EXTENT_ITEM_KEY;
4685                         key.offset = num_bytes;
4686
4687                         ret = btrfs_search_slot(trans, extent_root,
4688                                                 &key, path, -1, 1);
4689                         if (ret) {
4690                                 printk(KERN_ERR "umm, got %d back from search"
4691                                        ", was looking for %llu\n", ret,
4692                                        (unsigned long long)bytenr);
4693                                 if (ret > 0)
4694                                         btrfs_print_leaf(extent_root,
4695                                                          path->nodes[0]);
4696                         }
4697                         BUG_ON(ret);
4698                         extent_slot = path->slots[0];
4699                 }
4700         } else {
4701                 btrfs_print_leaf(extent_root, path->nodes[0]);
4702                 WARN_ON(1);
4703                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4704                        "parent %llu root %llu  owner %llu offset %llu\n",
4705                        (unsigned long long)bytenr,
4706                        (unsigned long long)parent,
4707                        (unsigned long long)root_objectid,
4708                        (unsigned long long)owner_objectid,
4709                        (unsigned long long)owner_offset);
4710         }
4711
4712         leaf = path->nodes[0];
4713         item_size = btrfs_item_size_nr(leaf, extent_slot);
4714 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4715         if (item_size < sizeof(*ei)) {
4716                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4717                 ret = convert_extent_item_v0(trans, extent_root, path,
4718                                              owner_objectid, 0);
4719                 BUG_ON(ret < 0);
4720
4721                 btrfs_release_path(path);
4722                 path->leave_spinning = 1;
4723
4724                 key.objectid = bytenr;
4725                 key.type = BTRFS_EXTENT_ITEM_KEY;
4726                 key.offset = num_bytes;
4727
4728                 ret = btrfs_search_slot(trans, extent_root, &key, path,
4729                                         -1, 1);
4730                 if (ret) {
4731                         printk(KERN_ERR "umm, got %d back from search"
4732                                ", was looking for %llu\n", ret,
4733                                (unsigned long long)bytenr);
4734                         btrfs_print_leaf(extent_root, path->nodes[0]);
4735                 }
4736                 BUG_ON(ret);
4737                 extent_slot = path->slots[0];
4738                 leaf = path->nodes[0];
4739                 item_size = btrfs_item_size_nr(leaf, extent_slot);
4740         }
4741 #endif
4742         BUG_ON(item_size < sizeof(*ei));
4743         ei = btrfs_item_ptr(leaf, extent_slot,
4744                             struct btrfs_extent_item);
4745         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
4746                 struct btrfs_tree_block_info *bi;
4747                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
4748                 bi = (struct btrfs_tree_block_info *)(ei + 1);
4749                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
4750         }
4751
4752         refs = btrfs_extent_refs(leaf, ei);
4753         BUG_ON(refs < refs_to_drop);
4754         refs -= refs_to_drop;
4755
4756         if (refs > 0) {
4757                 if (extent_op)
4758                         __run_delayed_extent_op(extent_op, leaf, ei);
4759                 /*
4760                  * In the case of inline back ref, reference count will
4761                  * be updated by remove_extent_backref
4762                  */
4763                 if (iref) {
4764                         BUG_ON(!found_extent);
4765                 } else {
4766                         btrfs_set_extent_refs(leaf, ei, refs);
4767                         btrfs_mark_buffer_dirty(leaf);
4768                 }
4769                 if (found_extent) {
4770                         ret = remove_extent_backref(trans, extent_root, path,
4771                                                     iref, refs_to_drop,
4772                                                     is_data);
4773                         BUG_ON(ret);
4774                 }
4775         } else {
4776                 if (found_extent) {
4777                         BUG_ON(is_data && refs_to_drop !=
4778                                extent_data_ref_count(root, path, iref));
4779                         if (iref) {
4780                                 BUG_ON(path->slots[0] != extent_slot);
4781                         } else {
4782                                 BUG_ON(path->slots[0] != extent_slot + 1);
4783                                 path->slots[0] = extent_slot;
4784                                 num_to_del = 2;
4785                         }
4786                 }
4787
4788                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
4789                                       num_to_del);
4790                 BUG_ON(ret);
4791                 btrfs_release_path(path);
4792
4793                 if (is_data) {
4794                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
4795                         BUG_ON(ret);
4796                 } else {
4797                         invalidate_mapping_pages(info->btree_inode->i_mapping,
4798                              bytenr >> PAGE_CACHE_SHIFT,
4799                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
4800                 }
4801
4802                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
4803                 BUG_ON(ret);
4804         }
4805         btrfs_free_path(path);
4806         return ret;
4807 }
4808
4809 /*
4810  * when we free an block, it is possible (and likely) that we free the last
4811  * delayed ref for that extent as well.  This searches the delayed ref tree for
4812  * a given extent, and if there are no other delayed refs to be processed, it
4813  * removes it from the tree.
4814  */
4815 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
4816                                       struct btrfs_root *root, u64 bytenr)
4817 {
4818         struct btrfs_delayed_ref_head *head;
4819         struct btrfs_delayed_ref_root *delayed_refs;
4820         struct btrfs_delayed_ref_node *ref;
4821         struct rb_node *node;
4822         int ret = 0;
4823
4824         delayed_refs = &trans->transaction->delayed_refs;
4825         spin_lock(&delayed_refs->lock);
4826         head = btrfs_find_delayed_ref_head(trans, bytenr);
4827         if (!head)
4828                 goto out;
4829
4830         node = rb_prev(&head->node.rb_node);
4831         if (!node)
4832                 goto out;
4833
4834         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
4835
4836         /* there are still entries for this ref, we can't drop it */
4837         if (ref->bytenr == bytenr)
4838                 goto out;
4839
4840         if (head->extent_op) {
4841                 if (!head->must_insert_reserved)
4842                         goto out;
4843                 kfree(head->extent_op);
4844                 head->extent_op = NULL;
4845         }
4846
4847         /*
4848          * waiting for the lock here would deadlock.  If someone else has it
4849          * locked they are already in the process of dropping it anyway
4850          */
4851         if (!mutex_trylock(&head->mutex))
4852                 goto out;
4853
4854         /*
4855          * at this point we have a head with no other entries.  Go
4856          * ahead and process it.
4857          */
4858         head->node.in_tree = 0;
4859         rb_erase(&head->node.rb_node, &delayed_refs->root);
4860
4861         delayed_refs->num_entries--;
4862
4863         /*
4864          * we don't take a ref on the node because we're removing it from the
4865          * tree, so we just steal the ref the tree was holding.
4866          */
4867         delayed_refs->num_heads--;
4868         if (list_empty(&head->cluster))
4869                 delayed_refs->num_heads_ready--;
4870
4871         list_del_init(&head->cluster);
4872         spin_unlock(&delayed_refs->lock);
4873
4874         BUG_ON(head->extent_op);
4875         if (head->must_insert_reserved)
4876                 ret = 1;
4877
4878         mutex_unlock(&head->mutex);
4879         btrfs_put_delayed_ref(&head->node);
4880         return ret;
4881 out:
4882         spin_unlock(&delayed_refs->lock);
4883         return 0;
4884 }
4885
4886 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
4887                            struct btrfs_root *root,
4888                            struct extent_buffer *buf,
4889                            u64 parent, int last_ref)
4890 {
4891         struct btrfs_block_group_cache *cache = NULL;
4892         int ret;
4893
4894         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4895                 ret = btrfs_add_delayed_tree_ref(trans, buf->start, buf->len,
4896                                                 parent, root->root_key.objectid,
4897                                                 btrfs_header_level(buf),
4898                                                 BTRFS_DROP_DELAYED_REF, NULL);
4899                 BUG_ON(ret);
4900         }
4901
4902         if (!last_ref)
4903                 return;
4904
4905         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
4906
4907         if (btrfs_header_generation(buf) == trans->transid) {
4908                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
4909                         ret = check_ref_cleanup(trans, root, buf->start);
4910                         if (!ret)
4911                                 goto out;
4912                 }
4913
4914                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
4915                         pin_down_extent(root, cache, buf->start, buf->len, 1);
4916                         goto out;
4917                 }
4918
4919                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
4920
4921                 btrfs_add_free_space(cache, buf->start, buf->len);
4922                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
4923         }
4924 out:
4925         /*
4926          * Deleting the buffer, clear the corrupt flag since it doesn't matter
4927          * anymore.
4928          */
4929         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
4930         btrfs_put_block_group(cache);
4931 }
4932
4933 int btrfs_free_extent(struct btrfs_trans_handle *trans,
4934                       struct btrfs_root *root,
4935                       u64 bytenr, u64 num_bytes, u64 parent,
4936                       u64 root_objectid, u64 owner, u64 offset)
4937 {
4938         int ret;
4939
4940         /*
4941          * tree log blocks never actually go into the extent allocation
4942          * tree, just update pinning info and exit early.
4943          */
4944         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
4945                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
4946                 /* unlocks the pinned mutex */
4947                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
4948                 ret = 0;
4949         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
4950                 ret = btrfs_add_delayed_tree_ref(trans, bytenr, num_bytes,
4951                                         parent, root_objectid, (int)owner,
4952                                         BTRFS_DROP_DELAYED_REF, NULL);
4953                 BUG_ON(ret);
4954         } else {
4955                 ret = btrfs_add_delayed_data_ref(trans, bytenr, num_bytes,
4956                                         parent, root_objectid, owner,
4957                                         offset, BTRFS_DROP_DELAYED_REF, NULL);
4958                 BUG_ON(ret);
4959         }
4960         return ret;
4961 }
4962
4963 static u64 stripe_align(struct btrfs_root *root, u64 val)
4964 {
4965         u64 mask = ((u64)root->stripesize - 1);
4966         u64 ret = (val + mask) & ~mask;
4967         return ret;
4968 }
4969
4970 /*
4971  * when we wait for progress in the block group caching, its because
4972  * our allocation attempt failed at least once.  So, we must sleep
4973  * and let some progress happen before we try again.
4974  *
4975  * This function will sleep at least once waiting for new free space to
4976  * show up, and then it will check the block group free space numbers
4977  * for our min num_bytes.  Another option is to have it go ahead
4978  * and look in the rbtree for a free extent of a given size, but this
4979  * is a good start.
4980  */
4981 static noinline int
4982 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
4983                                 u64 num_bytes)
4984 {
4985         struct btrfs_caching_control *caching_ctl;
4986         DEFINE_WAIT(wait);
4987
4988         caching_ctl = get_caching_control(cache);
4989         if (!caching_ctl)
4990                 return 0;
4991
4992         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
4993                    (cache->free_space_ctl->free_space >= num_bytes));
4994
4995         put_caching_control(caching_ctl);
4996         return 0;
4997 }
4998
4999 static noinline int
5000 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5001 {
5002         struct btrfs_caching_control *caching_ctl;
5003         DEFINE_WAIT(wait);
5004
5005         caching_ctl = get_caching_control(cache);
5006         if (!caching_ctl)
5007                 return 0;
5008
5009         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5010
5011         put_caching_control(caching_ctl);
5012         return 0;
5013 }
5014
5015 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5016 {
5017         int index;
5018         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
5019                 index = 0;
5020         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
5021                 index = 1;
5022         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
5023                 index = 2;
5024         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
5025                 index = 3;
5026         else
5027                 index = 4;
5028         return index;
5029 }
5030
5031 enum btrfs_loop_type {
5032         LOOP_FIND_IDEAL = 0,
5033         LOOP_CACHING_NOWAIT = 1,
5034         LOOP_CACHING_WAIT = 2,
5035         LOOP_ALLOC_CHUNK = 3,
5036         LOOP_NO_EMPTY_SIZE = 4,
5037 };
5038
5039 /*
5040  * walks the btree of allocated extents and find a hole of a given size.
5041  * The key ins is changed to record the hole:
5042  * ins->objectid == block start
5043  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5044  * ins->offset == number of blocks
5045  * Any available blocks before search_start are skipped.
5046  */
5047 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5048                                      struct btrfs_root *orig_root,
5049                                      u64 num_bytes, u64 empty_size,
5050                                      u64 search_start, u64 search_end,
5051                                      u64 hint_byte, struct btrfs_key *ins,
5052                                      u64 data)
5053 {
5054         int ret = 0;
5055         struct btrfs_root *root = orig_root->fs_info->extent_root;
5056         struct btrfs_free_cluster *last_ptr = NULL;
5057         struct btrfs_block_group_cache *block_group = NULL;
5058         int empty_cluster = 2 * 1024 * 1024;
5059         int allowed_chunk_alloc = 0;
5060         int done_chunk_alloc = 0;
5061         struct btrfs_space_info *space_info;
5062         int last_ptr_loop = 0;
5063         int loop = 0;
5064         int index = 0;
5065         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5066                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5067         bool found_uncached_bg = false;
5068         bool failed_cluster_refill = false;
5069         bool failed_alloc = false;
5070         bool use_cluster = true;
5071         bool have_caching_bg = false;
5072         u64 ideal_cache_percent = 0;
5073         u64 ideal_cache_offset = 0;
5074
5075         WARN_ON(num_bytes < root->sectorsize);
5076         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5077         ins->objectid = 0;
5078         ins->offset = 0;
5079
5080         space_info = __find_space_info(root->fs_info, data);
5081         if (!space_info) {
5082                 printk(KERN_ERR "No space info for %llu\n", data);
5083                 return -ENOSPC;
5084         }
5085
5086         /*
5087          * If the space info is for both data and metadata it means we have a
5088          * small filesystem and we can't use the clustering stuff.
5089          */
5090         if (btrfs_mixed_space_info(space_info))
5091                 use_cluster = false;
5092
5093         if (orig_root->ref_cows || empty_size)
5094                 allowed_chunk_alloc = 1;
5095
5096         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5097                 last_ptr = &root->fs_info->meta_alloc_cluster;
5098                 if (!btrfs_test_opt(root, SSD))
5099                         empty_cluster = 64 * 1024;
5100         }
5101
5102         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5103             btrfs_test_opt(root, SSD)) {
5104                 last_ptr = &root->fs_info->data_alloc_cluster;
5105         }
5106
5107         if (last_ptr) {
5108                 spin_lock(&last_ptr->lock);
5109                 if (last_ptr->block_group)
5110                         hint_byte = last_ptr->window_start;
5111                 spin_unlock(&last_ptr->lock);
5112         }
5113
5114         search_start = max(search_start, first_logical_byte(root, 0));
5115         search_start = max(search_start, hint_byte);
5116
5117         if (!last_ptr)
5118                 empty_cluster = 0;
5119
5120         if (search_start == hint_byte) {
5121 ideal_cache:
5122                 block_group = btrfs_lookup_block_group(root->fs_info,
5123                                                        search_start);
5124                 /*
5125                  * we don't want to use the block group if it doesn't match our
5126                  * allocation bits, or if its not cached.
5127                  *
5128                  * However if we are re-searching with an ideal block group
5129                  * picked out then we don't care that the block group is cached.
5130                  */
5131                 if (block_group && block_group_bits(block_group, data) &&
5132                     (block_group->cached != BTRFS_CACHE_NO ||
5133                      search_start == ideal_cache_offset)) {
5134                         down_read(&space_info->groups_sem);
5135                         if (list_empty(&block_group->list) ||
5136                             block_group->ro) {
5137                                 /*
5138                                  * someone is removing this block group,
5139                                  * we can't jump into the have_block_group
5140                                  * target because our list pointers are not
5141                                  * valid
5142                                  */
5143                                 btrfs_put_block_group(block_group);
5144                                 up_read(&space_info->groups_sem);
5145                         } else {
5146                                 index = get_block_group_index(block_group);
5147                                 goto have_block_group;
5148                         }
5149                 } else if (block_group) {
5150                         btrfs_put_block_group(block_group);
5151                 }
5152         }
5153 search:
5154         have_caching_bg = false;
5155         down_read(&space_info->groups_sem);
5156         list_for_each_entry(block_group, &space_info->block_groups[index],
5157                             list) {
5158                 u64 offset;
5159                 int cached;
5160
5161                 btrfs_get_block_group(block_group);
5162                 search_start = block_group->key.objectid;
5163
5164                 /*
5165                  * this can happen if we end up cycling through all the
5166                  * raid types, but we want to make sure we only allocate
5167                  * for the proper type.
5168                  */
5169                 if (!block_group_bits(block_group, data)) {
5170                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5171                                 BTRFS_BLOCK_GROUP_RAID1 |
5172                                 BTRFS_BLOCK_GROUP_RAID10;
5173
5174                         /*
5175                          * if they asked for extra copies and this block group
5176                          * doesn't provide them, bail.  This does allow us to
5177                          * fill raid0 from raid1.
5178                          */
5179                         if ((data & extra) && !(block_group->flags & extra))
5180                                 goto loop;
5181                 }
5182
5183 have_block_group:
5184                 if (unlikely(block_group->cached == BTRFS_CACHE_NO)) {
5185                         u64 free_percent;
5186
5187                         ret = cache_block_group(block_group, trans,
5188                                                 orig_root, 1);
5189                         if (block_group->cached == BTRFS_CACHE_FINISHED)
5190                                 goto have_block_group;
5191
5192                         free_percent = btrfs_block_group_used(&block_group->item);
5193                         free_percent *= 100;
5194                         free_percent = div64_u64(free_percent,
5195                                                  block_group->key.offset);
5196                         free_percent = 100 - free_percent;
5197                         if (free_percent > ideal_cache_percent &&
5198                             likely(!block_group->ro)) {
5199                                 ideal_cache_offset = block_group->key.objectid;
5200                                 ideal_cache_percent = free_percent;
5201                         }
5202
5203                         /*
5204                          * The caching workers are limited to 2 threads, so we
5205                          * can queue as much work as we care to.
5206                          */
5207                         if (loop > LOOP_FIND_IDEAL) {
5208                                 ret = cache_block_group(block_group, trans,
5209                                                         orig_root, 0);
5210                                 BUG_ON(ret);
5211                         }
5212                         found_uncached_bg = true;
5213
5214                         /*
5215                          * If loop is set for cached only, try the next block
5216                          * group.
5217                          */
5218                         if (loop == LOOP_FIND_IDEAL)
5219                                 goto loop;
5220                 }
5221
5222                 cached = block_group_cache_done(block_group);
5223                 if (unlikely(!cached))
5224                         found_uncached_bg = true;
5225
5226                 if (unlikely(block_group->ro))
5227                         goto loop;
5228
5229                 spin_lock(&block_group->free_space_ctl->tree_lock);
5230                 if (cached &&
5231                     block_group->free_space_ctl->free_space <
5232                     num_bytes + empty_size) {
5233                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5234                         goto loop;
5235                 }
5236                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5237
5238                 /*
5239                  * Ok we want to try and use the cluster allocator, so lets look
5240                  * there, unless we are on LOOP_NO_EMPTY_SIZE, since we will
5241                  * have tried the cluster allocator plenty of times at this
5242                  * point and not have found anything, so we are likely way too
5243                  * fragmented for the clustering stuff to find anything, so lets
5244                  * just skip it and let the allocator find whatever block it can
5245                  * find
5246                  */
5247                 if (last_ptr && loop < LOOP_NO_EMPTY_SIZE) {
5248                         /*
5249                          * the refill lock keeps out other
5250                          * people trying to start a new cluster
5251                          */
5252                         spin_lock(&last_ptr->refill_lock);
5253                         if (last_ptr->block_group &&
5254                             (last_ptr->block_group->ro ||
5255                             !block_group_bits(last_ptr->block_group, data))) {
5256                                 offset = 0;
5257                                 goto refill_cluster;
5258                         }
5259
5260                         offset = btrfs_alloc_from_cluster(block_group, last_ptr,
5261                                                  num_bytes, search_start);
5262                         if (offset) {
5263                                 /* we have a block, we're done */
5264                                 spin_unlock(&last_ptr->refill_lock);
5265                                 goto checks;
5266                         }
5267
5268                         spin_lock(&last_ptr->lock);
5269                         /*
5270                          * whoops, this cluster doesn't actually point to
5271                          * this block group.  Get a ref on the block
5272                          * group is does point to and try again
5273                          */
5274                         if (!last_ptr_loop && last_ptr->block_group &&
5275                             last_ptr->block_group != block_group &&
5276                             index <=
5277                                  get_block_group_index(last_ptr->block_group)) {
5278
5279                                 btrfs_put_block_group(block_group);
5280                                 block_group = last_ptr->block_group;
5281                                 btrfs_get_block_group(block_group);
5282                                 spin_unlock(&last_ptr->lock);
5283                                 spin_unlock(&last_ptr->refill_lock);
5284
5285                                 last_ptr_loop = 1;
5286                                 search_start = block_group->key.objectid;
5287                                 /*
5288                                  * we know this block group is properly
5289                                  * in the list because
5290                                  * btrfs_remove_block_group, drops the
5291                                  * cluster before it removes the block
5292                                  * group from the list
5293                                  */
5294                                 goto have_block_group;
5295                         }
5296                         spin_unlock(&last_ptr->lock);
5297 refill_cluster:
5298                         /*
5299                          * this cluster didn't work out, free it and
5300                          * start over
5301                          */
5302                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5303
5304                         last_ptr_loop = 0;
5305
5306                         /* allocate a cluster in this block group */
5307                         ret = btrfs_find_space_cluster(trans, root,
5308                                                block_group, last_ptr,
5309                                                offset, num_bytes,
5310                                                empty_cluster + empty_size);
5311                         if (ret == 0) {
5312                                 /*
5313                                  * now pull our allocation out of this
5314                                  * cluster
5315                                  */
5316                                 offset = btrfs_alloc_from_cluster(block_group,
5317                                                   last_ptr, num_bytes,
5318                                                   search_start);
5319                                 if (offset) {
5320                                         /* we found one, proceed */
5321                                         spin_unlock(&last_ptr->refill_lock);
5322                                         goto checks;
5323                                 }
5324                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5325                                    && !failed_cluster_refill) {
5326                                 spin_unlock(&last_ptr->refill_lock);
5327
5328                                 failed_cluster_refill = true;
5329                                 wait_block_group_cache_progress(block_group,
5330                                        num_bytes + empty_cluster + empty_size);
5331                                 goto have_block_group;
5332                         }
5333
5334                         /*
5335                          * at this point we either didn't find a cluster
5336                          * or we weren't able to allocate a block from our
5337                          * cluster.  Free the cluster we've been trying
5338                          * to use, and go to the next block group
5339                          */
5340                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5341                         spin_unlock(&last_ptr->refill_lock);
5342                         goto loop;
5343                 }
5344
5345                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5346                                                     num_bytes, empty_size);
5347                 /*
5348                  * If we didn't find a chunk, and we haven't failed on this
5349                  * block group before, and this block group is in the middle of
5350                  * caching and we are ok with waiting, then go ahead and wait
5351                  * for progress to be made, and set failed_alloc to true.
5352                  *
5353                  * If failed_alloc is true then we've already waited on this
5354                  * block group once and should move on to the next block group.
5355                  */
5356                 if (!offset && !failed_alloc && !cached &&
5357                     loop > LOOP_CACHING_NOWAIT) {
5358                         wait_block_group_cache_progress(block_group,
5359                                                 num_bytes + empty_size);
5360                         failed_alloc = true;
5361                         goto have_block_group;
5362                 } else if (!offset) {
5363                         if (!cached)
5364                                 have_caching_bg = true;
5365                         goto loop;
5366                 }
5367 checks:
5368                 search_start = stripe_align(root, offset);
5369                 /* move on to the next group */
5370                 if (search_start + num_bytes >= search_end) {
5371                         btrfs_add_free_space(block_group, offset, num_bytes);
5372                         goto loop;
5373                 }
5374
5375                 /* move on to the next group */
5376                 if (search_start + num_bytes >
5377                     block_group->key.objectid + block_group->key.offset) {
5378                         btrfs_add_free_space(block_group, offset, num_bytes);
5379                         goto loop;
5380                 }
5381
5382                 ins->objectid = search_start;
5383                 ins->offset = num_bytes;
5384
5385                 if (offset < search_start)
5386                         btrfs_add_free_space(block_group, offset,
5387                                              search_start - offset);
5388                 BUG_ON(offset > search_start);
5389
5390                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
5391                                                   alloc_type);
5392                 if (ret == -EAGAIN) {
5393                         btrfs_add_free_space(block_group, offset, num_bytes);
5394                         goto loop;
5395                 }
5396
5397                 /* we are all good, lets return */
5398                 ins->objectid = search_start;
5399                 ins->offset = num_bytes;
5400
5401                 if (offset < search_start)
5402                         btrfs_add_free_space(block_group, offset,
5403                                              search_start - offset);
5404                 BUG_ON(offset > search_start);
5405                 btrfs_put_block_group(block_group);
5406                 break;
5407 loop:
5408                 failed_cluster_refill = false;
5409                 failed_alloc = false;
5410                 BUG_ON(index != get_block_group_index(block_group));
5411                 btrfs_put_block_group(block_group);
5412         }
5413         up_read(&space_info->groups_sem);
5414
5415         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
5416                 goto search;
5417
5418         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5419                 goto search;
5420
5421         /* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
5422          *                      for them to make caching progress.  Also
5423          *                      determine the best possible bg to cache
5424          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5425          *                      caching kthreads as we move along
5426          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5427          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5428          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5429          *                      again
5430          */
5431         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5432                 index = 0;
5433                 if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
5434                         found_uncached_bg = false;
5435                         loop++;
5436                         if (!ideal_cache_percent)
5437                                 goto search;
5438
5439                         /*
5440                          * 1 of the following 2 things have happened so far
5441                          *
5442                          * 1) We found an ideal block group for caching that
5443                          * is mostly full and will cache quickly, so we might
5444                          * as well wait for it.
5445                          *
5446                          * 2) We searched for cached only and we didn't find
5447                          * anything, and we didn't start any caching kthreads
5448                          * either, so chances are we will loop through and
5449                          * start a couple caching kthreads, and then come back
5450                          * around and just wait for them.  This will be slower
5451                          * because we will have 2 caching kthreads reading at
5452                          * the same time when we could have just started one
5453                          * and waited for it to get far enough to give us an
5454                          * allocation, so go ahead and go to the wait caching
5455                          * loop.
5456                          */
5457                         loop = LOOP_CACHING_WAIT;
5458                         search_start = ideal_cache_offset;
5459                         ideal_cache_percent = 0;
5460                         goto ideal_cache;
5461                 } else if (loop == LOOP_FIND_IDEAL) {
5462                         /*
5463                          * Didn't find a uncached bg, wait on anything we find
5464                          * next.
5465                          */
5466                         loop = LOOP_CACHING_WAIT;
5467                         goto search;
5468                 }
5469
5470                 loop++;
5471
5472                 if (loop == LOOP_ALLOC_CHUNK) {
5473                        if (allowed_chunk_alloc) {
5474                                 ret = do_chunk_alloc(trans, root, num_bytes +
5475                                                      2 * 1024 * 1024, data,
5476                                                      CHUNK_ALLOC_LIMITED);
5477                                 allowed_chunk_alloc = 0;
5478                                 if (ret == 1)
5479                                         done_chunk_alloc = 1;
5480                         } else if (!done_chunk_alloc &&
5481                                    space_info->force_alloc ==
5482                                    CHUNK_ALLOC_NO_FORCE) {
5483                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5484                         }
5485
5486                        /*
5487                         * We didn't allocate a chunk, go ahead and drop the
5488                         * empty size and loop again.
5489                         */
5490                        if (!done_chunk_alloc)
5491                                loop = LOOP_NO_EMPTY_SIZE;
5492                 }
5493
5494                 if (loop == LOOP_NO_EMPTY_SIZE) {
5495                         empty_size = 0;
5496                         empty_cluster = 0;
5497                 }
5498
5499                 goto search;
5500         } else if (!ins->objectid) {
5501                 ret = -ENOSPC;
5502         } else if (ins->objectid) {
5503                 ret = 0;
5504         }
5505
5506         return ret;
5507 }
5508
5509 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5510                             int dump_block_groups)
5511 {
5512         struct btrfs_block_group_cache *cache;
5513         int index = 0;
5514
5515         spin_lock(&info->lock);
5516         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5517                (unsigned long long)info->flags,
5518                (unsigned long long)(info->total_bytes - info->bytes_used -
5519                                     info->bytes_pinned - info->bytes_reserved -
5520                                     info->bytes_readonly),
5521                (info->full) ? "" : "not ");
5522         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5523                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5524                (unsigned long long)info->total_bytes,
5525                (unsigned long long)info->bytes_used,
5526                (unsigned long long)info->bytes_pinned,
5527                (unsigned long long)info->bytes_reserved,
5528                (unsigned long long)info->bytes_may_use,
5529                (unsigned long long)info->bytes_readonly);
5530         spin_unlock(&info->lock);
5531
5532         if (!dump_block_groups)
5533                 return;
5534
5535         down_read(&info->groups_sem);
5536 again:
5537         list_for_each_entry(cache, &info->block_groups[index], list) {
5538                 spin_lock(&cache->lock);
5539                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5540                        "%llu pinned %llu reserved\n",
5541                        (unsigned long long)cache->key.objectid,
5542                        (unsigned long long)cache->key.offset,
5543                        (unsigned long long)btrfs_block_group_used(&cache->item),
5544                        (unsigned long long)cache->pinned,
5545                        (unsigned long long)cache->reserved);
5546                 btrfs_dump_free_space(cache, bytes);
5547                 spin_unlock(&cache->lock);
5548         }
5549         if (++index < BTRFS_NR_RAID_TYPES)
5550                 goto again;
5551         up_read(&info->groups_sem);
5552 }
5553
5554 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5555                          struct btrfs_root *root,
5556                          u64 num_bytes, u64 min_alloc_size,
5557                          u64 empty_size, u64 hint_byte,
5558                          u64 search_end, struct btrfs_key *ins,
5559                          u64 data)
5560 {
5561         int ret;
5562         u64 search_start = 0;
5563
5564         data = btrfs_get_alloc_profile(root, data);
5565 again:
5566         /*
5567          * the only place that sets empty_size is btrfs_realloc_node, which
5568          * is not called recursively on allocations
5569          */
5570         if (empty_size || root->ref_cows)
5571                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5572                                      num_bytes + 2 * 1024 * 1024, data,
5573                                      CHUNK_ALLOC_NO_FORCE);
5574
5575         WARN_ON(num_bytes < root->sectorsize);
5576         ret = find_free_extent(trans, root, num_bytes, empty_size,
5577                                search_start, search_end, hint_byte,
5578                                ins, data);
5579
5580         if (ret == -ENOSPC && num_bytes > min_alloc_size) {
5581                 num_bytes = num_bytes >> 1;
5582                 num_bytes = num_bytes & ~(root->sectorsize - 1);
5583                 num_bytes = max(num_bytes, min_alloc_size);
5584                 do_chunk_alloc(trans, root->fs_info->extent_root,
5585                                num_bytes, data, CHUNK_ALLOC_FORCE);
5586                 goto again;
5587         }
5588         if (ret == -ENOSPC && btrfs_test_opt(root, ENOSPC_DEBUG)) {
5589                 struct btrfs_space_info *sinfo;
5590
5591                 sinfo = __find_space_info(root->fs_info, data);
5592                 printk(KERN_ERR "btrfs allocation failed flags %llu, "
5593                        "wanted %llu\n", (unsigned long long)data,
5594                        (unsigned long long)num_bytes);
5595                 dump_space_info(sinfo, num_bytes, 1);
5596         }
5597
5598         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
5599
5600         return ret;
5601 }
5602
5603 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
5604                                         u64 start, u64 len, int pin)
5605 {
5606         struct btrfs_block_group_cache *cache;
5607         int ret = 0;
5608
5609         cache = btrfs_lookup_block_group(root->fs_info, start);
5610         if (!cache) {
5611                 printk(KERN_ERR "Unable to find block group for %llu\n",
5612                        (unsigned long long)start);
5613                 return -ENOSPC;
5614         }
5615
5616         if (btrfs_test_opt(root, DISCARD))
5617                 ret = btrfs_discard_extent(root, start, len, NULL);
5618
5619         if (pin)
5620                 pin_down_extent(root, cache, start, len, 1);
5621         else {
5622                 btrfs_add_free_space(cache, start, len);
5623                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
5624         }
5625         btrfs_put_block_group(cache);
5626
5627         trace_btrfs_reserved_extent_free(root, start, len);
5628
5629         return ret;
5630 }
5631
5632 int btrfs_free_reserved_extent(struct btrfs_root *root,
5633                                         u64 start, u64 len)
5634 {
5635         return __btrfs_free_reserved_extent(root, start, len, 0);
5636 }
5637
5638 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
5639                                        u64 start, u64 len)
5640 {
5641         return __btrfs_free_reserved_extent(root, start, len, 1);
5642 }
5643
5644 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5645                                       struct btrfs_root *root,
5646                                       u64 parent, u64 root_objectid,
5647                                       u64 flags, u64 owner, u64 offset,
5648                                       struct btrfs_key *ins, int ref_mod)
5649 {
5650         int ret;
5651         struct btrfs_fs_info *fs_info = root->fs_info;
5652         struct btrfs_extent_item *extent_item;
5653         struct btrfs_extent_inline_ref *iref;
5654         struct btrfs_path *path;
5655         struct extent_buffer *leaf;
5656         int type;
5657         u32 size;
5658
5659         if (parent > 0)
5660                 type = BTRFS_SHARED_DATA_REF_KEY;
5661         else
5662                 type = BTRFS_EXTENT_DATA_REF_KEY;
5663
5664         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5665
5666         path = btrfs_alloc_path();
5667         if (!path)
5668                 return -ENOMEM;
5669
5670         path->leave_spinning = 1;
5671         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5672                                       ins, size);
5673         BUG_ON(ret);
5674
5675         leaf = path->nodes[0];
5676         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5677                                      struct btrfs_extent_item);
5678         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
5679         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5680         btrfs_set_extent_flags(leaf, extent_item,
5681                                flags | BTRFS_EXTENT_FLAG_DATA);
5682
5683         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
5684         btrfs_set_extent_inline_ref_type(leaf, iref, type);
5685         if (parent > 0) {
5686                 struct btrfs_shared_data_ref *ref;
5687                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
5688                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5689                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
5690         } else {
5691                 struct btrfs_extent_data_ref *ref;
5692                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
5693                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
5694                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
5695                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
5696                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
5697         }
5698
5699         btrfs_mark_buffer_dirty(path->nodes[0]);
5700         btrfs_free_path(path);
5701
5702         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5703         if (ret) {
5704                 printk(KERN_ERR "btrfs update block group failed for %llu "
5705                        "%llu\n", (unsigned long long)ins->objectid,
5706                        (unsigned long long)ins->offset);
5707                 BUG();
5708         }
5709         return ret;
5710 }
5711
5712 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
5713                                      struct btrfs_root *root,
5714                                      u64 parent, u64 root_objectid,
5715                                      u64 flags, struct btrfs_disk_key *key,
5716                                      int level, struct btrfs_key *ins)
5717 {
5718         int ret;
5719         struct btrfs_fs_info *fs_info = root->fs_info;
5720         struct btrfs_extent_item *extent_item;
5721         struct btrfs_tree_block_info *block_info;
5722         struct btrfs_extent_inline_ref *iref;
5723         struct btrfs_path *path;
5724         struct extent_buffer *leaf;
5725         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
5726
5727         path = btrfs_alloc_path();
5728         if (!path)
5729                 return -ENOMEM;
5730
5731         path->leave_spinning = 1;
5732         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5733                                       ins, size);
5734         BUG_ON(ret);
5735
5736         leaf = path->nodes[0];
5737         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5738                                      struct btrfs_extent_item);
5739         btrfs_set_extent_refs(leaf, extent_item, 1);
5740         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5741         btrfs_set_extent_flags(leaf, extent_item,
5742                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
5743         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
5744
5745         btrfs_set_tree_block_key(leaf, block_info, key);
5746         btrfs_set_tree_block_level(leaf, block_info, level);
5747
5748         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
5749         if (parent > 0) {
5750                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
5751                 btrfs_set_extent_inline_ref_type(leaf, iref,
5752                                                  BTRFS_SHARED_BLOCK_REF_KEY);
5753                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5754         } else {
5755                 btrfs_set_extent_inline_ref_type(leaf, iref,
5756                                                  BTRFS_TREE_BLOCK_REF_KEY);
5757                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
5758         }
5759
5760         btrfs_mark_buffer_dirty(leaf);
5761         btrfs_free_path(path);
5762
5763         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5764         if (ret) {
5765                 printk(KERN_ERR "btrfs update block group failed for %llu "
5766                        "%llu\n", (unsigned long long)ins->objectid,
5767                        (unsigned long long)ins->offset);
5768                 BUG();
5769         }
5770         return ret;
5771 }
5772
5773 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5774                                      struct btrfs_root *root,
5775                                      u64 root_objectid, u64 owner,
5776                                      u64 offset, struct btrfs_key *ins)
5777 {
5778         int ret;
5779
5780         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
5781
5782         ret = btrfs_add_delayed_data_ref(trans, ins->objectid, ins->offset,
5783                                          0, root_objectid, owner, offset,
5784                                          BTRFS_ADD_DELAYED_EXTENT, NULL);
5785         return ret;
5786 }
5787
5788 /*
5789  * this is used by the tree logging recovery code.  It records that
5790  * an extent has been allocated and makes sure to clear the free
5791  * space cache bits as well
5792  */
5793 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
5794                                    struct btrfs_root *root,
5795                                    u64 root_objectid, u64 owner, u64 offset,
5796                                    struct btrfs_key *ins)
5797 {
5798         int ret;
5799         struct btrfs_block_group_cache *block_group;
5800         struct btrfs_caching_control *caching_ctl;
5801         u64 start = ins->objectid;
5802         u64 num_bytes = ins->offset;
5803
5804         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
5805         cache_block_group(block_group, trans, NULL, 0);
5806         caching_ctl = get_caching_control(block_group);
5807
5808         if (!caching_ctl) {
5809                 BUG_ON(!block_group_cache_done(block_group));
5810                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5811                 BUG_ON(ret);
5812         } else {
5813                 mutex_lock(&caching_ctl->mutex);
5814
5815                 if (start >= caching_ctl->progress) {
5816                         ret = add_excluded_extent(root, start, num_bytes);
5817                         BUG_ON(ret);
5818                 } else if (start + num_bytes <= caching_ctl->progress) {
5819                         ret = btrfs_remove_free_space(block_group,
5820                                                       start, num_bytes);
5821                         BUG_ON(ret);
5822                 } else {
5823                         num_bytes = caching_ctl->progress - start;
5824                         ret = btrfs_remove_free_space(block_group,
5825                                                       start, num_bytes);
5826                         BUG_ON(ret);
5827
5828                         start = caching_ctl->progress;
5829                         num_bytes = ins->objectid + ins->offset -
5830                                     caching_ctl->progress;
5831                         ret = add_excluded_extent(root, start, num_bytes);
5832                         BUG_ON(ret);
5833                 }
5834
5835                 mutex_unlock(&caching_ctl->mutex);
5836                 put_caching_control(caching_ctl);
5837         }
5838
5839         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
5840                                           RESERVE_ALLOC_NO_ACCOUNT);
5841         BUG_ON(ret);
5842         btrfs_put_block_group(block_group);
5843         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
5844                                          0, owner, offset, ins, 1);
5845         return ret;
5846 }
5847
5848 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
5849                                             struct btrfs_root *root,
5850                                             u64 bytenr, u32 blocksize,
5851                                             int level)
5852 {
5853         struct extent_buffer *buf;
5854
5855         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
5856         if (!buf)
5857                 return ERR_PTR(-ENOMEM);
5858         btrfs_set_header_generation(buf, trans->transid);
5859         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
5860         btrfs_tree_lock(buf);
5861         clean_tree_block(trans, root, buf);
5862
5863         btrfs_set_lock_blocking(buf);
5864         btrfs_set_buffer_uptodate(buf);
5865
5866         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
5867                 /*
5868                  * we allow two log transactions at a time, use different
5869                  * EXENT bit to differentiate dirty pages.
5870                  */
5871                 if (root->log_transid % 2 == 0)
5872                         set_extent_dirty(&root->dirty_log_pages, buf->start,
5873                                         buf->start + buf->len - 1, GFP_NOFS);
5874                 else
5875                         set_extent_new(&root->dirty_log_pages, buf->start,
5876                                         buf->start + buf->len - 1, GFP_NOFS);
5877         } else {
5878                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
5879                          buf->start + buf->len - 1, GFP_NOFS);
5880         }
5881         trans->blocks_used++;
5882         /* this returns a buffer locked for blocking */
5883         return buf;
5884 }
5885
5886 static struct btrfs_block_rsv *
5887 use_block_rsv(struct btrfs_trans_handle *trans,
5888               struct btrfs_root *root, u32 blocksize)
5889 {
5890         struct btrfs_block_rsv *block_rsv;
5891         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5892         int ret;
5893
5894         block_rsv = get_block_rsv(trans, root);
5895
5896         if (block_rsv->size == 0) {
5897                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
5898                 /*
5899                  * If we couldn't reserve metadata bytes try and use some from
5900                  * the global reserve.
5901                  */
5902                 if (ret && block_rsv != global_rsv) {
5903                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5904                         if (!ret)
5905                                 return global_rsv;
5906                         return ERR_PTR(ret);
5907                 } else if (ret) {
5908                         return ERR_PTR(ret);
5909                 }
5910                 return block_rsv;
5911         }
5912
5913         ret = block_rsv_use_bytes(block_rsv, blocksize);
5914         if (!ret)
5915                 return block_rsv;
5916         if (ret) {
5917                 static DEFINE_RATELIMIT_STATE(_rs,
5918                                 DEFAULT_RATELIMIT_INTERVAL,
5919                                 /*DEFAULT_RATELIMIT_BURST*/ 2);
5920                 if (__ratelimit(&_rs)) {
5921                         printk(KERN_DEBUG "btrfs: block rsv returned %d\n", ret);
5922                         WARN_ON(1);
5923                 }
5924                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
5925                 if (!ret) {
5926                         return block_rsv;
5927                 } else if (ret && block_rsv != global_rsv) {
5928                         ret = block_rsv_use_bytes(global_rsv, blocksize);
5929                         if (!ret)
5930                                 return global_rsv;
5931                 }
5932         }
5933
5934         return ERR_PTR(-ENOSPC);
5935 }
5936
5937 static void unuse_block_rsv(struct btrfs_block_rsv *block_rsv, u32 blocksize)
5938 {
5939         block_rsv_add_bytes(block_rsv, blocksize, 0);
5940         block_rsv_release_bytes(block_rsv, NULL, 0);
5941 }
5942
5943 /*
5944  * finds a free extent and does all the dirty work required for allocation
5945  * returns the key for the extent through ins, and a tree buffer for
5946  * the first block of the extent through buf.
5947  *
5948  * returns the tree buffer or NULL.
5949  */
5950 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
5951                                         struct btrfs_root *root, u32 blocksize,
5952                                         u64 parent, u64 root_objectid,
5953                                         struct btrfs_disk_key *key, int level,
5954                                         u64 hint, u64 empty_size)
5955 {
5956         struct btrfs_key ins;
5957         struct btrfs_block_rsv *block_rsv;
5958         struct extent_buffer *buf;
5959         u64 flags = 0;
5960         int ret;
5961
5962
5963         block_rsv = use_block_rsv(trans, root, blocksize);
5964         if (IS_ERR(block_rsv))
5965                 return ERR_CAST(block_rsv);
5966
5967         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
5968                                    empty_size, hint, (u64)-1, &ins, 0);
5969         if (ret) {
5970                 unuse_block_rsv(block_rsv, blocksize);
5971                 return ERR_PTR(ret);
5972         }
5973
5974         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
5975                                     blocksize, level);
5976         BUG_ON(IS_ERR(buf));
5977
5978         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
5979                 if (parent == 0)
5980                         parent = ins.objectid;
5981                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
5982         } else
5983                 BUG_ON(parent > 0);
5984
5985         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
5986                 struct btrfs_delayed_extent_op *extent_op;
5987                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
5988                 BUG_ON(!extent_op);
5989                 if (key)
5990                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
5991                 else
5992                         memset(&extent_op->key, 0, sizeof(extent_op->key));
5993                 extent_op->flags_to_set = flags;
5994                 extent_op->update_key = 1;
5995                 extent_op->update_flags = 1;
5996                 extent_op->is_data = 0;
5997
5998                 ret = btrfs_add_delayed_tree_ref(trans, ins.objectid,
5999                                         ins.offset, parent, root_objectid,
6000                                         level, BTRFS_ADD_DELAYED_EXTENT,
6001                                         extent_op);
6002                 BUG_ON(ret);
6003         }
6004         return buf;
6005 }
6006
6007 struct walk_control {
6008         u64 refs[BTRFS_MAX_LEVEL];
6009         u64 flags[BTRFS_MAX_LEVEL];
6010         struct btrfs_key update_progress;
6011         int stage;
6012         int level;
6013         int shared_level;
6014         int update_ref;
6015         int keep_locks;
6016         int reada_slot;
6017         int reada_count;
6018 };
6019
6020 #define DROP_REFERENCE  1
6021 #define UPDATE_BACKREF  2
6022
6023 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6024                                      struct btrfs_root *root,
6025                                      struct walk_control *wc,
6026                                      struct btrfs_path *path)
6027 {
6028         u64 bytenr;
6029         u64 generation;
6030         u64 refs;
6031         u64 flags;
6032         u32 nritems;
6033         u32 blocksize;
6034         struct btrfs_key key;
6035         struct extent_buffer *eb;
6036         int ret;
6037         int slot;
6038         int nread = 0;
6039
6040         if (path->slots[wc->level] < wc->reada_slot) {
6041                 wc->reada_count = wc->reada_count * 2 / 3;
6042                 wc->reada_count = max(wc->reada_count, 2);
6043         } else {
6044                 wc->reada_count = wc->reada_count * 3 / 2;
6045                 wc->reada_count = min_t(int, wc->reada_count,
6046                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6047         }
6048
6049         eb = path->nodes[wc->level];
6050         nritems = btrfs_header_nritems(eb);
6051         blocksize = btrfs_level_size(root, wc->level - 1);
6052
6053         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6054                 if (nread >= wc->reada_count)
6055                         break;
6056
6057                 cond_resched();
6058                 bytenr = btrfs_node_blockptr(eb, slot);
6059                 generation = btrfs_node_ptr_generation(eb, slot);
6060
6061                 if (slot == path->slots[wc->level])
6062                         goto reada;
6063
6064                 if (wc->stage == UPDATE_BACKREF &&
6065                     generation <= root->root_key.offset)
6066                         continue;
6067
6068                 /* We don't lock the tree block, it's OK to be racy here */
6069                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6070                                                &refs, &flags);
6071                 BUG_ON(ret);
6072                 BUG_ON(refs == 0);
6073
6074                 if (wc->stage == DROP_REFERENCE) {
6075                         if (refs == 1)
6076                                 goto reada;
6077
6078                         if (wc->level == 1 &&
6079                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6080                                 continue;
6081                         if (!wc->update_ref ||
6082                             generation <= root->root_key.offset)
6083                                 continue;
6084                         btrfs_node_key_to_cpu(eb, &key, slot);
6085                         ret = btrfs_comp_cpu_keys(&key,
6086                                                   &wc->update_progress);
6087                         if (ret < 0)
6088                                 continue;
6089                 } else {
6090                         if (wc->level == 1 &&
6091                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6092                                 continue;
6093                 }
6094 reada:
6095                 ret = readahead_tree_block(root, bytenr, blocksize,
6096                                            generation);
6097                 if (ret)
6098                         break;
6099                 nread++;
6100         }
6101         wc->reada_slot = slot;
6102 }
6103
6104 /*
6105  * hepler to process tree block while walking down the tree.
6106  *
6107  * when wc->stage == UPDATE_BACKREF, this function updates
6108  * back refs for pointers in the block.
6109  *
6110  * NOTE: return value 1 means we should stop walking down.
6111  */
6112 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6113                                    struct btrfs_root *root,
6114                                    struct btrfs_path *path,
6115                                    struct walk_control *wc, int lookup_info)
6116 {
6117         int level = wc->level;
6118         struct extent_buffer *eb = path->nodes[level];
6119         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6120         int ret;
6121
6122         if (wc->stage == UPDATE_BACKREF &&
6123             btrfs_header_owner(eb) != root->root_key.objectid)
6124                 return 1;
6125
6126         /*
6127          * when reference count of tree block is 1, it won't increase
6128          * again. once full backref flag is set, we never clear it.
6129          */
6130         if (lookup_info &&
6131             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6132              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6133                 BUG_ON(!path->locks[level]);
6134                 ret = btrfs_lookup_extent_info(trans, root,
6135                                                eb->start, eb->len,
6136                                                &wc->refs[level],
6137                                                &wc->flags[level]);
6138                 BUG_ON(ret);
6139                 BUG_ON(wc->refs[level] == 0);
6140         }
6141
6142         if (wc->stage == DROP_REFERENCE) {
6143                 if (wc->refs[level] > 1)
6144                         return 1;
6145
6146                 if (path->locks[level] && !wc->keep_locks) {
6147                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6148                         path->locks[level] = 0;
6149                 }
6150                 return 0;
6151         }
6152
6153         /* wc->stage == UPDATE_BACKREF */
6154         if (!(wc->flags[level] & flag)) {
6155                 BUG_ON(!path->locks[level]);
6156                 ret = btrfs_inc_ref(trans, root, eb, 1);
6157                 BUG_ON(ret);
6158                 ret = btrfs_dec_ref(trans, root, eb, 0);
6159                 BUG_ON(ret);
6160                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6161                                                   eb->len, flag, 0);
6162                 BUG_ON(ret);
6163                 wc->flags[level] |= flag;
6164         }
6165
6166         /*
6167          * the block is shared by multiple trees, so it's not good to
6168          * keep the tree lock
6169          */
6170         if (path->locks[level] && level > 0) {
6171                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6172                 path->locks[level] = 0;
6173         }
6174         return 0;
6175 }
6176
6177 /*
6178  * hepler to process tree block pointer.
6179  *
6180  * when wc->stage == DROP_REFERENCE, this function checks
6181  * reference count of the block pointed to. if the block
6182  * is shared and we need update back refs for the subtree
6183  * rooted at the block, this function changes wc->stage to
6184  * UPDATE_BACKREF. if the block is shared and there is no
6185  * need to update back, this function drops the reference
6186  * to the block.
6187  *
6188  * NOTE: return value 1 means we should stop walking down.
6189  */
6190 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6191                                  struct btrfs_root *root,
6192                                  struct btrfs_path *path,
6193                                  struct walk_control *wc, int *lookup_info)
6194 {
6195         u64 bytenr;
6196         u64 generation;
6197         u64 parent;
6198         u32 blocksize;
6199         struct btrfs_key key;
6200         struct extent_buffer *next;
6201         int level = wc->level;
6202         int reada = 0;
6203         int ret = 0;
6204
6205         generation = btrfs_node_ptr_generation(path->nodes[level],
6206                                                path->slots[level]);
6207         /*
6208          * if the lower level block was created before the snapshot
6209          * was created, we know there is no need to update back refs
6210          * for the subtree
6211          */
6212         if (wc->stage == UPDATE_BACKREF &&
6213             generation <= root->root_key.offset) {
6214                 *lookup_info = 1;
6215                 return 1;
6216         }
6217
6218         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6219         blocksize = btrfs_level_size(root, level - 1);
6220
6221         next = btrfs_find_tree_block(root, bytenr, blocksize);
6222         if (!next) {
6223                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6224                 if (!next)
6225                         return -ENOMEM;
6226                 reada = 1;
6227         }
6228         btrfs_tree_lock(next);
6229         btrfs_set_lock_blocking(next);
6230
6231         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6232                                        &wc->refs[level - 1],
6233                                        &wc->flags[level - 1]);
6234         BUG_ON(ret);
6235         BUG_ON(wc->refs[level - 1] == 0);
6236         *lookup_info = 0;
6237
6238         if (wc->stage == DROP_REFERENCE) {
6239                 if (wc->refs[level - 1] > 1) {
6240                         if (level == 1 &&
6241                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6242                                 goto skip;
6243
6244                         if (!wc->update_ref ||
6245                             generation <= root->root_key.offset)
6246                                 goto skip;
6247
6248                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6249                                               path->slots[level]);
6250                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6251                         if (ret < 0)
6252                                 goto skip;
6253
6254                         wc->stage = UPDATE_BACKREF;
6255                         wc->shared_level = level - 1;
6256                 }
6257         } else {
6258                 if (level == 1 &&
6259                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6260                         goto skip;
6261         }
6262
6263         if (!btrfs_buffer_uptodate(next, generation)) {
6264                 btrfs_tree_unlock(next);
6265                 free_extent_buffer(next);
6266                 next = NULL;
6267                 *lookup_info = 1;
6268         }
6269
6270         if (!next) {
6271                 if (reada && level == 1)
6272                         reada_walk_down(trans, root, wc, path);
6273                 next = read_tree_block(root, bytenr, blocksize, generation);
6274                 if (!next)
6275                         return -EIO;
6276                 btrfs_tree_lock(next);
6277                 btrfs_set_lock_blocking(next);
6278         }
6279
6280         level--;
6281         BUG_ON(level != btrfs_header_level(next));
6282         path->nodes[level] = next;
6283         path->slots[level] = 0;
6284         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6285         wc->level = level;
6286         if (wc->level == 1)
6287                 wc->reada_slot = 0;
6288         return 0;
6289 skip:
6290         wc->refs[level - 1] = 0;
6291         wc->flags[level - 1] = 0;
6292         if (wc->stage == DROP_REFERENCE) {
6293                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6294                         parent = path->nodes[level]->start;
6295                 } else {
6296                         BUG_ON(root->root_key.objectid !=
6297                                btrfs_header_owner(path->nodes[level]));
6298                         parent = 0;
6299                 }
6300
6301                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6302                                         root->root_key.objectid, level - 1, 0);
6303                 BUG_ON(ret);
6304         }
6305         btrfs_tree_unlock(next);
6306         free_extent_buffer(next);
6307         *lookup_info = 1;
6308         return 1;
6309 }
6310
6311 /*
6312  * hepler to process tree block while walking up the tree.
6313  *
6314  * when wc->stage == DROP_REFERENCE, this function drops
6315  * reference count on the block.
6316  *
6317  * when wc->stage == UPDATE_BACKREF, this function changes
6318  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6319  * to UPDATE_BACKREF previously while processing the block.
6320  *
6321  * NOTE: return value 1 means we should stop walking up.
6322  */
6323 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6324                                  struct btrfs_root *root,
6325                                  struct btrfs_path *path,
6326                                  struct walk_control *wc)
6327 {
6328         int ret;
6329         int level = wc->level;
6330         struct extent_buffer *eb = path->nodes[level];
6331         u64 parent = 0;
6332
6333         if (wc->stage == UPDATE_BACKREF) {
6334                 BUG_ON(wc->shared_level < level);
6335                 if (level < wc->shared_level)
6336                         goto out;
6337
6338                 ret = find_next_key(path, level + 1, &wc->update_progress);
6339                 if (ret > 0)
6340                         wc->update_ref = 0;
6341
6342                 wc->stage = DROP_REFERENCE;
6343                 wc->shared_level = -1;
6344                 path->slots[level] = 0;
6345
6346                 /*
6347                  * check reference count again if the block isn't locked.
6348                  * we should start walking down the tree again if reference
6349                  * count is one.
6350                  */
6351                 if (!path->locks[level]) {
6352                         BUG_ON(level == 0);
6353                         btrfs_tree_lock(eb);
6354                         btrfs_set_lock_blocking(eb);
6355                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6356
6357                         ret = btrfs_lookup_extent_info(trans, root,
6358                                                        eb->start, eb->len,
6359                                                        &wc->refs[level],
6360                                                        &wc->flags[level]);
6361                         BUG_ON(ret);
6362                         BUG_ON(wc->refs[level] == 0);
6363                         if (wc->refs[level] == 1) {
6364                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6365                                 return 1;
6366                         }
6367                 }
6368         }
6369
6370         /* wc->stage == DROP_REFERENCE */
6371         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6372
6373         if (wc->refs[level] == 1) {
6374                 if (level == 0) {
6375                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6376                                 ret = btrfs_dec_ref(trans, root, eb, 1);
6377                         else
6378                                 ret = btrfs_dec_ref(trans, root, eb, 0);
6379                         BUG_ON(ret);
6380                 }
6381                 /* make block locked assertion in clean_tree_block happy */
6382                 if (!path->locks[level] &&
6383                     btrfs_header_generation(eb) == trans->transid) {
6384                         btrfs_tree_lock(eb);
6385                         btrfs_set_lock_blocking(eb);
6386                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6387                 }
6388                 clean_tree_block(trans, root, eb);
6389         }
6390
6391         if (eb == root->node) {
6392                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6393                         parent = eb->start;
6394                 else
6395                         BUG_ON(root->root_key.objectid !=
6396                                btrfs_header_owner(eb));
6397         } else {
6398                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6399                         parent = path->nodes[level + 1]->start;
6400                 else
6401                         BUG_ON(root->root_key.objectid !=
6402                                btrfs_header_owner(path->nodes[level + 1]));
6403         }
6404
6405         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6406 out:
6407         wc->refs[level] = 0;
6408         wc->flags[level] = 0;
6409         return 0;
6410 }
6411
6412 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6413                                    struct btrfs_root *root,
6414                                    struct btrfs_path *path,
6415                                    struct walk_control *wc)
6416 {
6417         int level = wc->level;
6418         int lookup_info = 1;
6419         int ret;
6420
6421         while (level >= 0) {
6422                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6423                 if (ret > 0)
6424                         break;
6425
6426                 if (level == 0)
6427                         break;
6428
6429                 if (path->slots[level] >=
6430                     btrfs_header_nritems(path->nodes[level]))
6431                         break;
6432
6433                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6434                 if (ret > 0) {
6435                         path->slots[level]++;
6436                         continue;
6437                 } else if (ret < 0)
6438                         return ret;
6439                 level = wc->level;
6440         }
6441         return 0;
6442 }
6443
6444 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6445                                  struct btrfs_root *root,
6446                                  struct btrfs_path *path,
6447                                  struct walk_control *wc, int max_level)
6448 {
6449         int level = wc->level;
6450         int ret;
6451
6452         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6453         while (level < max_level && path->nodes[level]) {
6454                 wc->level = level;
6455                 if (path->slots[level] + 1 <
6456                     btrfs_header_nritems(path->nodes[level])) {
6457                         path->slots[level]++;
6458                         return 0;
6459                 } else {
6460                         ret = walk_up_proc(trans, root, path, wc);
6461                         if (ret > 0)
6462                                 return 0;
6463
6464                         if (path->locks[level]) {
6465                                 btrfs_tree_unlock_rw(path->nodes[level],
6466                                                      path->locks[level]);
6467                                 path->locks[level] = 0;
6468                         }
6469                         free_extent_buffer(path->nodes[level]);
6470                         path->nodes[level] = NULL;
6471                         level++;
6472                 }
6473         }
6474         return 1;
6475 }
6476
6477 /*
6478  * drop a subvolume tree.
6479  *
6480  * this function traverses the tree freeing any blocks that only
6481  * referenced by the tree.
6482  *
6483  * when a shared tree block is found. this function decreases its
6484  * reference count by one. if update_ref is true, this function
6485  * also make sure backrefs for the shared block and all lower level
6486  * blocks are properly updated.
6487  */
6488 void btrfs_drop_snapshot(struct btrfs_root *root,
6489                          struct btrfs_block_rsv *block_rsv, int update_ref)
6490 {
6491         struct btrfs_path *path;
6492         struct btrfs_trans_handle *trans;
6493         struct btrfs_root *tree_root = root->fs_info->tree_root;
6494         struct btrfs_root_item *root_item = &root->root_item;
6495         struct walk_control *wc;
6496         struct btrfs_key key;
6497         int err = 0;
6498         int ret;
6499         int level;
6500
6501         path = btrfs_alloc_path();
6502         if (!path) {
6503                 err = -ENOMEM;
6504                 goto out;
6505         }
6506
6507         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6508         if (!wc) {
6509                 btrfs_free_path(path);
6510                 err = -ENOMEM;
6511                 goto out;
6512         }
6513
6514         trans = btrfs_start_transaction(tree_root, 0);
6515         BUG_ON(IS_ERR(trans));
6516
6517         if (block_rsv)
6518                 trans->block_rsv = block_rsv;
6519
6520         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6521                 level = btrfs_header_level(root->node);
6522                 path->nodes[level] = btrfs_lock_root_node(root);
6523                 btrfs_set_lock_blocking(path->nodes[level]);
6524                 path->slots[level] = 0;
6525                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6526                 memset(&wc->update_progress, 0,
6527                        sizeof(wc->update_progress));
6528         } else {
6529                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6530                 memcpy(&wc->update_progress, &key,
6531                        sizeof(wc->update_progress));
6532
6533                 level = root_item->drop_level;
6534                 BUG_ON(level == 0);
6535                 path->lowest_level = level;
6536                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6537                 path->lowest_level = 0;
6538                 if (ret < 0) {
6539                         err = ret;
6540                         goto out_free;
6541                 }
6542                 WARN_ON(ret > 0);
6543
6544                 /*
6545                  * unlock our path, this is safe because only this
6546                  * function is allowed to delete this snapshot
6547                  */
6548                 btrfs_unlock_up_safe(path, 0);
6549
6550                 level = btrfs_header_level(root->node);
6551                 while (1) {
6552                         btrfs_tree_lock(path->nodes[level]);
6553                         btrfs_set_lock_blocking(path->nodes[level]);
6554
6555                         ret = btrfs_lookup_extent_info(trans, root,
6556                                                 path->nodes[level]->start,
6557                                                 path->nodes[level]->len,
6558                                                 &wc->refs[level],
6559                                                 &wc->flags[level]);
6560                         BUG_ON(ret);
6561                         BUG_ON(wc->refs[level] == 0);
6562
6563                         if (level == root_item->drop_level)
6564                                 break;
6565
6566                         btrfs_tree_unlock(path->nodes[level]);
6567                         WARN_ON(wc->refs[level] != 1);
6568                         level--;
6569                 }
6570         }
6571
6572         wc->level = level;
6573         wc->shared_level = -1;
6574         wc->stage = DROP_REFERENCE;
6575         wc->update_ref = update_ref;
6576         wc->keep_locks = 0;
6577         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6578
6579         while (1) {
6580                 ret = walk_down_tree(trans, root, path, wc);
6581                 if (ret < 0) {
6582                         err = ret;
6583                         break;
6584                 }
6585
6586                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
6587                 if (ret < 0) {
6588                         err = ret;
6589                         break;
6590                 }
6591
6592                 if (ret > 0) {
6593                         BUG_ON(wc->stage != DROP_REFERENCE);
6594                         break;
6595                 }
6596
6597                 if (wc->stage == DROP_REFERENCE) {
6598                         level = wc->level;
6599                         btrfs_node_key(path->nodes[level],
6600                                        &root_item->drop_progress,
6601                                        path->slots[level]);
6602                         root_item->drop_level = level;
6603                 }
6604
6605                 BUG_ON(wc->level == 0);
6606                 if (btrfs_should_end_transaction(trans, tree_root)) {
6607                         ret = btrfs_update_root(trans, tree_root,
6608                                                 &root->root_key,
6609                                                 root_item);
6610                         BUG_ON(ret);
6611
6612                         btrfs_end_transaction_throttle(trans, tree_root);
6613                         trans = btrfs_start_transaction(tree_root, 0);
6614                         BUG_ON(IS_ERR(trans));
6615                         if (block_rsv)
6616                                 trans->block_rsv = block_rsv;
6617                 }
6618         }
6619         btrfs_release_path(path);
6620         BUG_ON(err);
6621
6622         ret = btrfs_del_root(trans, tree_root, &root->root_key);
6623         BUG_ON(ret);
6624
6625         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
6626                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
6627                                            NULL, NULL);
6628                 BUG_ON(ret < 0);
6629                 if (ret > 0) {
6630                         /* if we fail to delete the orphan item this time
6631                          * around, it'll get picked up the next time.
6632                          *
6633                          * The most common failure here is just -ENOENT.
6634                          */
6635                         btrfs_del_orphan_item(trans, tree_root,
6636                                               root->root_key.objectid);
6637                 }
6638         }
6639
6640         if (root->in_radix) {
6641                 btrfs_free_fs_root(tree_root->fs_info, root);
6642         } else {
6643                 free_extent_buffer(root->node);
6644                 free_extent_buffer(root->commit_root);
6645                 kfree(root);
6646         }
6647 out_free:
6648         btrfs_end_transaction_throttle(trans, tree_root);
6649         kfree(wc);
6650         btrfs_free_path(path);
6651 out:
6652         if (err)
6653                 btrfs_std_error(root->fs_info, err);
6654         return;
6655 }
6656
6657 /*
6658  * drop subtree rooted at tree block 'node'.
6659  *
6660  * NOTE: this function will unlock and release tree block 'node'
6661  */
6662 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
6663                         struct btrfs_root *root,
6664                         struct extent_buffer *node,
6665                         struct extent_buffer *parent)
6666 {
6667         struct btrfs_path *path;
6668         struct walk_control *wc;
6669         int level;
6670         int parent_level;
6671         int ret = 0;
6672         int wret;
6673
6674         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
6675
6676         path = btrfs_alloc_path();
6677         if (!path)
6678                 return -ENOMEM;
6679
6680         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6681         if (!wc) {
6682                 btrfs_free_path(path);
6683                 return -ENOMEM;
6684         }
6685
6686         btrfs_assert_tree_locked(parent);
6687         parent_level = btrfs_header_level(parent);
6688         extent_buffer_get(parent);
6689         path->nodes[parent_level] = parent;
6690         path->slots[parent_level] = btrfs_header_nritems(parent);
6691
6692         btrfs_assert_tree_locked(node);
6693         level = btrfs_header_level(node);
6694         path->nodes[level] = node;
6695         path->slots[level] = 0;
6696         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6697
6698         wc->refs[parent_level] = 1;
6699         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6700         wc->level = level;
6701         wc->shared_level = -1;
6702         wc->stage = DROP_REFERENCE;
6703         wc->update_ref = 0;
6704         wc->keep_locks = 1;
6705         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6706
6707         while (1) {
6708                 wret = walk_down_tree(trans, root, path, wc);
6709                 if (wret < 0) {
6710                         ret = wret;
6711                         break;
6712                 }
6713
6714                 wret = walk_up_tree(trans, root, path, wc, parent_level);
6715                 if (wret < 0)
6716                         ret = wret;
6717                 if (wret != 0)
6718                         break;
6719         }
6720
6721         kfree(wc);
6722         btrfs_free_path(path);
6723         return ret;
6724 }
6725
6726 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
6727 {
6728         u64 num_devices;
6729         u64 stripped = BTRFS_BLOCK_GROUP_RAID0 |
6730                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
6731
6732         /*
6733          * we add in the count of missing devices because we want
6734          * to make sure that any RAID levels on a degraded FS
6735          * continue to be honored.
6736          */
6737         num_devices = root->fs_info->fs_devices->rw_devices +
6738                 root->fs_info->fs_devices->missing_devices;
6739
6740         if (num_devices == 1) {
6741                 stripped |= BTRFS_BLOCK_GROUP_DUP;
6742                 stripped = flags & ~stripped;
6743
6744                 /* turn raid0 into single device chunks */
6745                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
6746                         return stripped;
6747
6748                 /* turn mirroring into duplication */
6749                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
6750                              BTRFS_BLOCK_GROUP_RAID10))
6751                         return stripped | BTRFS_BLOCK_GROUP_DUP;
6752                 return flags;
6753         } else {
6754                 /* they already had raid on here, just return */
6755                 if (flags & stripped)
6756                         return flags;
6757
6758                 stripped |= BTRFS_BLOCK_GROUP_DUP;
6759                 stripped = flags & ~stripped;
6760
6761                 /* switch duplicated blocks with raid1 */
6762                 if (flags & BTRFS_BLOCK_GROUP_DUP)
6763                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
6764
6765                 /* turn single device chunks into raid0 */
6766                 return stripped | BTRFS_BLOCK_GROUP_RAID0;
6767         }
6768         return flags;
6769 }
6770
6771 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
6772 {
6773         struct btrfs_space_info *sinfo = cache->space_info;
6774         u64 num_bytes;
6775         u64 min_allocable_bytes;
6776         int ret = -ENOSPC;
6777
6778
6779         /*
6780          * We need some metadata space and system metadata space for
6781          * allocating chunks in some corner cases until we force to set
6782          * it to be readonly.
6783          */
6784         if ((sinfo->flags &
6785              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
6786             !force)
6787                 min_allocable_bytes = 1 * 1024 * 1024;
6788         else
6789                 min_allocable_bytes = 0;
6790
6791         spin_lock(&sinfo->lock);
6792         spin_lock(&cache->lock);
6793
6794         if (cache->ro) {
6795                 ret = 0;
6796                 goto out;
6797         }
6798
6799         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
6800                     cache->bytes_super - btrfs_block_group_used(&cache->item);
6801
6802         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
6803             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
6804             min_allocable_bytes <= sinfo->total_bytes) {
6805                 sinfo->bytes_readonly += num_bytes;
6806                 cache->ro = 1;
6807                 ret = 0;
6808         }
6809 out:
6810         spin_unlock(&cache->lock);
6811         spin_unlock(&sinfo->lock);
6812         return ret;
6813 }
6814
6815 int btrfs_set_block_group_ro(struct btrfs_root *root,
6816                              struct btrfs_block_group_cache *cache)
6817
6818 {
6819         struct btrfs_trans_handle *trans;
6820         u64 alloc_flags;
6821         int ret;
6822
6823         BUG_ON(cache->ro);
6824
6825         trans = btrfs_join_transaction(root);
6826         BUG_ON(IS_ERR(trans));
6827
6828         alloc_flags = update_block_group_flags(root, cache->flags);
6829         if (alloc_flags != cache->flags)
6830                 do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6831                                CHUNK_ALLOC_FORCE);
6832
6833         ret = set_block_group_ro(cache, 0);
6834         if (!ret)
6835                 goto out;
6836         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
6837         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6838                              CHUNK_ALLOC_FORCE);
6839         if (ret < 0)
6840                 goto out;
6841         ret = set_block_group_ro(cache, 0);
6842 out:
6843         btrfs_end_transaction(trans, root);
6844         return ret;
6845 }
6846
6847 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
6848                             struct btrfs_root *root, u64 type)
6849 {
6850         u64 alloc_flags = get_alloc_profile(root, type);
6851         return do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
6852                               CHUNK_ALLOC_FORCE);
6853 }
6854
6855 /*
6856  * helper to account the unused space of all the readonly block group in the
6857  * list. takes mirrors into account.
6858  */
6859 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
6860 {
6861         struct btrfs_block_group_cache *block_group;
6862         u64 free_bytes = 0;
6863         int factor;
6864
6865         list_for_each_entry(block_group, groups_list, list) {
6866                 spin_lock(&block_group->lock);
6867
6868                 if (!block_group->ro) {
6869                         spin_unlock(&block_group->lock);
6870                         continue;
6871                 }
6872
6873                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
6874                                           BTRFS_BLOCK_GROUP_RAID10 |
6875                                           BTRFS_BLOCK_GROUP_DUP))
6876                         factor = 2;
6877                 else
6878                         factor = 1;
6879
6880                 free_bytes += (block_group->key.offset -
6881                                btrfs_block_group_used(&block_group->item)) *
6882                                factor;
6883
6884                 spin_unlock(&block_group->lock);
6885         }
6886
6887         return free_bytes;
6888 }
6889
6890 /*
6891  * helper to account the unused space of all the readonly block group in the
6892  * space_info. takes mirrors into account.
6893  */
6894 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
6895 {
6896         int i;
6897         u64 free_bytes = 0;
6898
6899         spin_lock(&sinfo->lock);
6900
6901         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
6902                 if (!list_empty(&sinfo->block_groups[i]))
6903                         free_bytes += __btrfs_get_ro_block_group_free_space(
6904                                                 &sinfo->block_groups[i]);
6905
6906         spin_unlock(&sinfo->lock);
6907
6908         return free_bytes;
6909 }
6910
6911 int btrfs_set_block_group_rw(struct btrfs_root *root,
6912                               struct btrfs_block_group_cache *cache)
6913 {
6914         struct btrfs_space_info *sinfo = cache->space_info;
6915         u64 num_bytes;
6916
6917         BUG_ON(!cache->ro);
6918
6919         spin_lock(&sinfo->lock);
6920         spin_lock(&cache->lock);
6921         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
6922                     cache->bytes_super - btrfs_block_group_used(&cache->item);
6923         sinfo->bytes_readonly -= num_bytes;
6924         cache->ro = 0;
6925         spin_unlock(&cache->lock);
6926         spin_unlock(&sinfo->lock);
6927         return 0;
6928 }
6929
6930 /*
6931  * checks to see if its even possible to relocate this block group.
6932  *
6933  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
6934  * ok to go ahead and try.
6935  */
6936 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
6937 {
6938         struct btrfs_block_group_cache *block_group;
6939         struct btrfs_space_info *space_info;
6940         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
6941         struct btrfs_device *device;
6942         u64 min_free;
6943         u64 dev_min = 1;
6944         u64 dev_nr = 0;
6945         int index;
6946         int full = 0;
6947         int ret = 0;
6948
6949         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
6950
6951         /* odd, couldn't find the block group, leave it alone */
6952         if (!block_group)
6953                 return -1;
6954
6955         min_free = btrfs_block_group_used(&block_group->item);
6956
6957         /* no bytes used, we're good */
6958         if (!min_free)
6959                 goto out;
6960
6961         space_info = block_group->space_info;
6962         spin_lock(&space_info->lock);
6963
6964         full = space_info->full;
6965
6966         /*
6967          * if this is the last block group we have in this space, we can't
6968          * relocate it unless we're able to allocate a new chunk below.
6969          *
6970          * Otherwise, we need to make sure we have room in the space to handle
6971          * all of the extents from this block group.  If we can, we're good
6972          */
6973         if ((space_info->total_bytes != block_group->key.offset) &&
6974             (space_info->bytes_used + space_info->bytes_reserved +
6975              space_info->bytes_pinned + space_info->bytes_readonly +
6976              min_free < space_info->total_bytes)) {
6977                 spin_unlock(&space_info->lock);
6978                 goto out;
6979         }
6980         spin_unlock(&space_info->lock);
6981
6982         /*
6983          * ok we don't have enough space, but maybe we have free space on our
6984          * devices to allocate new chunks for relocation, so loop through our
6985          * alloc devices and guess if we have enough space.  However, if we
6986          * were marked as full, then we know there aren't enough chunks, and we
6987          * can just return.
6988          */
6989         ret = -1;
6990         if (full)
6991                 goto out;
6992
6993         /*
6994          * index:
6995          *      0: raid10
6996          *      1: raid1
6997          *      2: dup
6998          *      3: raid0
6999          *      4: single
7000          */
7001         index = get_block_group_index(block_group);
7002         if (index == 0) {
7003                 dev_min = 4;
7004                 /* Divide by 2 */
7005                 min_free >>= 1;
7006         } else if (index == 1) {
7007                 dev_min = 2;
7008         } else if (index == 2) {
7009                 /* Multiply by 2 */
7010                 min_free <<= 1;
7011         } else if (index == 3) {
7012                 dev_min = fs_devices->rw_devices;
7013                 do_div(min_free, dev_min);
7014         }
7015
7016         mutex_lock(&root->fs_info->chunk_mutex);
7017         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7018                 u64 dev_offset;
7019
7020                 /*
7021                  * check to make sure we can actually find a chunk with enough
7022                  * space to fit our block group in.
7023                  */
7024                 if (device->total_bytes > device->bytes_used + min_free) {
7025                         ret = find_free_dev_extent(NULL, device, min_free,
7026                                                    &dev_offset, NULL);
7027                         if (!ret)
7028                                 dev_nr++;
7029
7030                         if (dev_nr >= dev_min)
7031                                 break;
7032
7033                         ret = -1;
7034                 }
7035         }
7036         mutex_unlock(&root->fs_info->chunk_mutex);
7037 out:
7038         btrfs_put_block_group(block_group);
7039         return ret;
7040 }
7041
7042 static int find_first_block_group(struct btrfs_root *root,
7043                 struct btrfs_path *path, struct btrfs_key *key)
7044 {
7045         int ret = 0;
7046         struct btrfs_key found_key;
7047         struct extent_buffer *leaf;
7048         int slot;
7049
7050         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7051         if (ret < 0)
7052                 goto out;
7053
7054         while (1) {
7055                 slot = path->slots[0];
7056                 leaf = path->nodes[0];
7057                 if (slot >= btrfs_header_nritems(leaf)) {
7058                         ret = btrfs_next_leaf(root, path);
7059                         if (ret == 0)
7060                                 continue;
7061                         if (ret < 0)
7062                                 goto out;
7063                         break;
7064                 }
7065                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7066
7067                 if (found_key.objectid >= key->objectid &&
7068                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7069                         ret = 0;
7070                         goto out;
7071                 }
7072                 path->slots[0]++;
7073         }
7074 out:
7075         return ret;
7076 }
7077
7078 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7079 {
7080         struct btrfs_block_group_cache *block_group;
7081         u64 last = 0;
7082
7083         while (1) {
7084                 struct inode *inode;
7085
7086                 block_group = btrfs_lookup_first_block_group(info, last);
7087                 while (block_group) {
7088                         spin_lock(&block_group->lock);
7089                         if (block_group->iref)
7090                                 break;
7091                         spin_unlock(&block_group->lock);
7092                         block_group = next_block_group(info->tree_root,
7093                                                        block_group);
7094                 }
7095                 if (!block_group) {
7096                         if (last == 0)
7097                                 break;
7098                         last = 0;
7099                         continue;
7100                 }
7101
7102                 inode = block_group->inode;
7103                 block_group->iref = 0;
7104                 block_group->inode = NULL;
7105                 spin_unlock(&block_group->lock);
7106                 iput(inode);
7107                 last = block_group->key.objectid + block_group->key.offset;
7108                 btrfs_put_block_group(block_group);
7109         }
7110 }
7111
7112 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7113 {
7114         struct btrfs_block_group_cache *block_group;
7115         struct btrfs_space_info *space_info;
7116         struct btrfs_caching_control *caching_ctl;
7117         struct rb_node *n;
7118
7119         down_write(&info->extent_commit_sem);
7120         while (!list_empty(&info->caching_block_groups)) {
7121                 caching_ctl = list_entry(info->caching_block_groups.next,
7122                                          struct btrfs_caching_control, list);
7123                 list_del(&caching_ctl->list);
7124                 put_caching_control(caching_ctl);
7125         }
7126         up_write(&info->extent_commit_sem);
7127
7128         spin_lock(&info->block_group_cache_lock);
7129         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7130                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7131                                        cache_node);
7132                 rb_erase(&block_group->cache_node,
7133                          &info->block_group_cache_tree);
7134                 spin_unlock(&info->block_group_cache_lock);
7135
7136                 down_write(&block_group->space_info->groups_sem);
7137                 list_del(&block_group->list);
7138                 up_write(&block_group->space_info->groups_sem);
7139
7140                 if (block_group->cached == BTRFS_CACHE_STARTED)
7141                         wait_block_group_cache_done(block_group);
7142
7143                 /*
7144                  * We haven't cached this block group, which means we could
7145                  * possibly have excluded extents on this block group.
7146                  */
7147                 if (block_group->cached == BTRFS_CACHE_NO)
7148                         free_excluded_extents(info->extent_root, block_group);
7149
7150                 btrfs_remove_free_space_cache(block_group);
7151                 btrfs_put_block_group(block_group);
7152
7153                 spin_lock(&info->block_group_cache_lock);
7154         }
7155         spin_unlock(&info->block_group_cache_lock);
7156
7157         /* now that all the block groups are freed, go through and
7158          * free all the space_info structs.  This is only called during
7159          * the final stages of unmount, and so we know nobody is
7160          * using them.  We call synchronize_rcu() once before we start,
7161          * just to be on the safe side.
7162          */
7163         synchronize_rcu();
7164
7165         release_global_block_rsv(info);
7166
7167         while(!list_empty(&info->space_info)) {
7168                 space_info = list_entry(info->space_info.next,
7169                                         struct btrfs_space_info,
7170                                         list);
7171                 if (space_info->bytes_pinned > 0 ||
7172                     space_info->bytes_reserved > 0 ||
7173                     space_info->bytes_may_use > 0) {
7174                         WARN_ON(1);
7175                         dump_space_info(space_info, 0, 0);
7176                 }
7177                 list_del(&space_info->list);
7178                 kfree(space_info);
7179         }
7180         return 0;
7181 }
7182
7183 static void __link_block_group(struct btrfs_space_info *space_info,
7184                                struct btrfs_block_group_cache *cache)
7185 {
7186         int index = get_block_group_index(cache);
7187
7188         down_write(&space_info->groups_sem);
7189         list_add_tail(&cache->list, &space_info->block_groups[index]);
7190         up_write(&space_info->groups_sem);
7191 }
7192
7193 int btrfs_read_block_groups(struct btrfs_root *root)
7194 {
7195         struct btrfs_path *path;
7196         int ret;
7197         struct btrfs_block_group_cache *cache;
7198         struct btrfs_fs_info *info = root->fs_info;
7199         struct btrfs_space_info *space_info;
7200         struct btrfs_key key;
7201         struct btrfs_key found_key;
7202         struct extent_buffer *leaf;
7203         int need_clear = 0;
7204         u64 cache_gen;
7205
7206         root = info->extent_root;
7207         key.objectid = 0;
7208         key.offset = 0;
7209         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7210         path = btrfs_alloc_path();
7211         if (!path)
7212                 return -ENOMEM;
7213         path->reada = 1;
7214
7215         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7216         if (btrfs_test_opt(root, SPACE_CACHE) &&
7217             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7218                 need_clear = 1;
7219         if (btrfs_test_opt(root, CLEAR_CACHE))
7220                 need_clear = 1;
7221
7222         while (1) {
7223                 ret = find_first_block_group(root, path, &key);
7224                 if (ret > 0)
7225                         break;
7226                 if (ret != 0)
7227                         goto error;
7228                 leaf = path->nodes[0];
7229                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7230                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7231                 if (!cache) {
7232                         ret = -ENOMEM;
7233                         goto error;
7234                 }
7235                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7236                                                 GFP_NOFS);
7237                 if (!cache->free_space_ctl) {
7238                         kfree(cache);
7239                         ret = -ENOMEM;
7240                         goto error;
7241                 }
7242
7243                 atomic_set(&cache->count, 1);
7244                 spin_lock_init(&cache->lock);
7245                 cache->fs_info = info;
7246                 INIT_LIST_HEAD(&cache->list);
7247                 INIT_LIST_HEAD(&cache->cluster_list);
7248
7249                 if (need_clear)
7250                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7251
7252                 read_extent_buffer(leaf, &cache->item,
7253                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7254                                    sizeof(cache->item));
7255                 memcpy(&cache->key, &found_key, sizeof(found_key));
7256
7257                 key.objectid = found_key.objectid + found_key.offset;
7258                 btrfs_release_path(path);
7259                 cache->flags = btrfs_block_group_flags(&cache->item);
7260                 cache->sectorsize = root->sectorsize;
7261
7262                 btrfs_init_free_space_ctl(cache);
7263
7264                 /*
7265                  * We need to exclude the super stripes now so that the space
7266                  * info has super bytes accounted for, otherwise we'll think
7267                  * we have more space than we actually do.
7268                  */
7269                 exclude_super_stripes(root, cache);
7270
7271                 /*
7272                  * check for two cases, either we are full, and therefore
7273                  * don't need to bother with the caching work since we won't
7274                  * find any space, or we are empty, and we can just add all
7275                  * the space in and be done with it.  This saves us _alot_ of
7276                  * time, particularly in the full case.
7277                  */
7278                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7279                         cache->last_byte_to_unpin = (u64)-1;
7280                         cache->cached = BTRFS_CACHE_FINISHED;
7281                         free_excluded_extents(root, cache);
7282                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7283                         cache->last_byte_to_unpin = (u64)-1;
7284                         cache->cached = BTRFS_CACHE_FINISHED;
7285                         add_new_free_space(cache, root->fs_info,
7286                                            found_key.objectid,
7287                                            found_key.objectid +
7288                                            found_key.offset);
7289                         free_excluded_extents(root, cache);
7290                 }
7291
7292                 ret = update_space_info(info, cache->flags, found_key.offset,
7293                                         btrfs_block_group_used(&cache->item),
7294                                         &space_info);
7295                 BUG_ON(ret);
7296                 cache->space_info = space_info;
7297                 spin_lock(&cache->space_info->lock);
7298                 cache->space_info->bytes_readonly += cache->bytes_super;
7299                 spin_unlock(&cache->space_info->lock);
7300
7301                 __link_block_group(space_info, cache);
7302
7303                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7304                 BUG_ON(ret);
7305
7306                 set_avail_alloc_bits(root->fs_info, cache->flags);
7307                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7308                         set_block_group_ro(cache, 1);
7309         }
7310
7311         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7312                 if (!(get_alloc_profile(root, space_info->flags) &
7313                       (BTRFS_BLOCK_GROUP_RAID10 |
7314                        BTRFS_BLOCK_GROUP_RAID1 |
7315                        BTRFS_BLOCK_GROUP_DUP)))
7316                         continue;
7317                 /*
7318                  * avoid allocating from un-mirrored block group if there are
7319                  * mirrored block groups.
7320                  */
7321                 list_for_each_entry(cache, &space_info->block_groups[3], list)
7322                         set_block_group_ro(cache, 1);
7323                 list_for_each_entry(cache, &space_info->block_groups[4], list)
7324                         set_block_group_ro(cache, 1);
7325         }
7326
7327         init_global_block_rsv(info);
7328         ret = 0;
7329 error:
7330         btrfs_free_path(path);
7331         return ret;
7332 }
7333
7334 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7335                            struct btrfs_root *root, u64 bytes_used,
7336                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7337                            u64 size)
7338 {
7339         int ret;
7340         struct btrfs_root *extent_root;
7341         struct btrfs_block_group_cache *cache;
7342
7343         extent_root = root->fs_info->extent_root;
7344
7345         root->fs_info->last_trans_log_full_commit = trans->transid;
7346
7347         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7348         if (!cache)
7349                 return -ENOMEM;
7350         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7351                                         GFP_NOFS);
7352         if (!cache->free_space_ctl) {
7353                 kfree(cache);
7354                 return -ENOMEM;
7355         }
7356
7357         cache->key.objectid = chunk_offset;
7358         cache->key.offset = size;
7359         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7360         cache->sectorsize = root->sectorsize;
7361         cache->fs_info = root->fs_info;
7362
7363         atomic_set(&cache->count, 1);
7364         spin_lock_init(&cache->lock);
7365         INIT_LIST_HEAD(&cache->list);
7366         INIT_LIST_HEAD(&cache->cluster_list);
7367
7368         btrfs_init_free_space_ctl(cache);
7369
7370         btrfs_set_block_group_used(&cache->item, bytes_used);
7371         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7372         cache->flags = type;
7373         btrfs_set_block_group_flags(&cache->item, type);
7374
7375         cache->last_byte_to_unpin = (u64)-1;
7376         cache->cached = BTRFS_CACHE_FINISHED;
7377         exclude_super_stripes(root, cache);
7378
7379         add_new_free_space(cache, root->fs_info, chunk_offset,
7380                            chunk_offset + size);
7381
7382         free_excluded_extents(root, cache);
7383
7384         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7385                                 &cache->space_info);
7386         BUG_ON(ret);
7387
7388         spin_lock(&cache->space_info->lock);
7389         cache->space_info->bytes_readonly += cache->bytes_super;
7390         spin_unlock(&cache->space_info->lock);
7391
7392         __link_block_group(cache->space_info, cache);
7393
7394         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7395         BUG_ON(ret);
7396
7397         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
7398                                 sizeof(cache->item));
7399         BUG_ON(ret);
7400
7401         set_avail_alloc_bits(extent_root->fs_info, type);
7402
7403         return 0;
7404 }
7405
7406 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7407                              struct btrfs_root *root, u64 group_start)
7408 {
7409         struct btrfs_path *path;
7410         struct btrfs_block_group_cache *block_group;
7411         struct btrfs_free_cluster *cluster;
7412         struct btrfs_root *tree_root = root->fs_info->tree_root;
7413         struct btrfs_key key;
7414         struct inode *inode;
7415         int ret;
7416         int factor;
7417
7418         root = root->fs_info->extent_root;
7419
7420         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7421         BUG_ON(!block_group);
7422         BUG_ON(!block_group->ro);
7423
7424         /*
7425          * Free the reserved super bytes from this block group before
7426          * remove it.
7427          */
7428         free_excluded_extents(root, block_group);
7429
7430         memcpy(&key, &block_group->key, sizeof(key));
7431         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7432                                   BTRFS_BLOCK_GROUP_RAID1 |
7433                                   BTRFS_BLOCK_GROUP_RAID10))
7434                 factor = 2;
7435         else
7436                 factor = 1;
7437
7438         /* make sure this block group isn't part of an allocation cluster */
7439         cluster = &root->fs_info->data_alloc_cluster;
7440         spin_lock(&cluster->refill_lock);
7441         btrfs_return_cluster_to_free_space(block_group, cluster);
7442         spin_unlock(&cluster->refill_lock);
7443
7444         /*
7445          * make sure this block group isn't part of a metadata
7446          * allocation cluster
7447          */
7448         cluster = &root->fs_info->meta_alloc_cluster;
7449         spin_lock(&cluster->refill_lock);
7450         btrfs_return_cluster_to_free_space(block_group, cluster);
7451         spin_unlock(&cluster->refill_lock);
7452
7453         path = btrfs_alloc_path();
7454         if (!path) {
7455                 ret = -ENOMEM;
7456                 goto out;
7457         }
7458
7459         inode = lookup_free_space_inode(tree_root, block_group, path);
7460         if (!IS_ERR(inode)) {
7461                 ret = btrfs_orphan_add(trans, inode);
7462                 BUG_ON(ret);
7463                 clear_nlink(inode);
7464                 /* One for the block groups ref */
7465                 spin_lock(&block_group->lock);
7466                 if (block_group->iref) {
7467                         block_group->iref = 0;
7468                         block_group->inode = NULL;
7469                         spin_unlock(&block_group->lock);
7470                         iput(inode);
7471                 } else {
7472                         spin_unlock(&block_group->lock);
7473                 }
7474                 /* One for our lookup ref */
7475                 btrfs_add_delayed_iput(inode);
7476         }
7477
7478         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
7479         key.offset = block_group->key.objectid;
7480         key.type = 0;
7481
7482         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
7483         if (ret < 0)
7484                 goto out;
7485         if (ret > 0)
7486                 btrfs_release_path(path);
7487         if (ret == 0) {
7488                 ret = btrfs_del_item(trans, tree_root, path);
7489                 if (ret)
7490                         goto out;
7491                 btrfs_release_path(path);
7492         }
7493
7494         spin_lock(&root->fs_info->block_group_cache_lock);
7495         rb_erase(&block_group->cache_node,
7496                  &root->fs_info->block_group_cache_tree);
7497         spin_unlock(&root->fs_info->block_group_cache_lock);
7498
7499         down_write(&block_group->space_info->groups_sem);
7500         /*
7501          * we must use list_del_init so people can check to see if they
7502          * are still on the list after taking the semaphore
7503          */
7504         list_del_init(&block_group->list);
7505         up_write(&block_group->space_info->groups_sem);
7506
7507         if (block_group->cached == BTRFS_CACHE_STARTED)
7508                 wait_block_group_cache_done(block_group);
7509
7510         btrfs_remove_free_space_cache(block_group);
7511
7512         spin_lock(&block_group->space_info->lock);
7513         block_group->space_info->total_bytes -= block_group->key.offset;
7514         block_group->space_info->bytes_readonly -= block_group->key.offset;
7515         block_group->space_info->disk_total -= block_group->key.offset * factor;
7516         spin_unlock(&block_group->space_info->lock);
7517
7518         memcpy(&key, &block_group->key, sizeof(key));
7519
7520         btrfs_clear_space_info_full(root->fs_info);
7521
7522         btrfs_put_block_group(block_group);
7523         btrfs_put_block_group(block_group);
7524
7525         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
7526         if (ret > 0)
7527                 ret = -EIO;
7528         if (ret < 0)
7529                 goto out;
7530
7531         ret = btrfs_del_item(trans, root, path);
7532 out:
7533         btrfs_free_path(path);
7534         return ret;
7535 }
7536
7537 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
7538 {
7539         struct btrfs_space_info *space_info;
7540         struct btrfs_super_block *disk_super;
7541         u64 features;
7542         u64 flags;
7543         int mixed = 0;
7544         int ret;
7545
7546         disk_super = fs_info->super_copy;
7547         if (!btrfs_super_root(disk_super))
7548                 return 1;
7549
7550         features = btrfs_super_incompat_flags(disk_super);
7551         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
7552                 mixed = 1;
7553
7554         flags = BTRFS_BLOCK_GROUP_SYSTEM;
7555         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7556         if (ret)
7557                 goto out;
7558
7559         if (mixed) {
7560                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
7561                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7562         } else {
7563                 flags = BTRFS_BLOCK_GROUP_METADATA;
7564                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7565                 if (ret)
7566                         goto out;
7567
7568                 flags = BTRFS_BLOCK_GROUP_DATA;
7569                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7570         }
7571 out:
7572         return ret;
7573 }
7574
7575 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
7576 {
7577         return unpin_extent_range(root, start, end);
7578 }
7579
7580 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
7581                                u64 num_bytes, u64 *actual_bytes)
7582 {
7583         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
7584 }
7585
7586 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
7587 {
7588         struct btrfs_fs_info *fs_info = root->fs_info;
7589         struct btrfs_block_group_cache *cache = NULL;
7590         u64 group_trimmed;
7591         u64 start;
7592         u64 end;
7593         u64 trimmed = 0;
7594         int ret = 0;
7595
7596         cache = btrfs_lookup_block_group(fs_info, range->start);
7597
7598         while (cache) {
7599                 if (cache->key.objectid >= (range->start + range->len)) {
7600                         btrfs_put_block_group(cache);
7601                         break;
7602                 }
7603
7604                 start = max(range->start, cache->key.objectid);
7605                 end = min(range->start + range->len,
7606                                 cache->key.objectid + cache->key.offset);
7607
7608                 if (end - start >= range->minlen) {
7609                         if (!block_group_cache_done(cache)) {
7610                                 ret = cache_block_group(cache, NULL, root, 0);
7611                                 if (!ret)
7612                                         wait_block_group_cache_done(cache);
7613                         }
7614                         ret = btrfs_trim_block_group(cache,
7615                                                      &group_trimmed,
7616                                                      start,
7617                                                      end,
7618                                                      range->minlen);
7619
7620                         trimmed += group_trimmed;
7621                         if (ret) {
7622                                 btrfs_put_block_group(cache);
7623                                 break;
7624                         }
7625                 }
7626
7627                 cache = next_block_group(fs_info->tree_root, cache);
7628         }
7629
7630         range->len = trimmed;
7631         return ret;
7632 }