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 /*
  38  * control flags for do_chunk_alloc's force field
  39  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  40  * if we really need one.
  41  *
  42  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  43  * if we have very few chunks already allocated.  This is
  44  * used as part of the clustering code to help make sure
  45  * we have a good pool of storage to cluster in, without
  46  * filling the FS with empty chunks
  47  *
  48  * CHUNK_ALLOC_FORCE means it must try to allocate one
  49  *
  50  */
  51 enum {
  52         CHUNK_ALLOC_NO_FORCE = 0,
  53         CHUNK_ALLOC_LIMITED = 1,
  54         CHUNK_ALLOC_FORCE = 2,
  55 };
  56
  57 /*
  58  * Control how reservations are dealt with.
  59  *
  60  * RESERVE_FREE - freeing a reservation.
  61  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  62  *   ENOSPC accounting
  63  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  64  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  65  */
  66 enum {
  67         RESERVE_FREE = 0,
  68         RESERVE_ALLOC = 1,
  69         RESERVE_ALLOC_NO_ACCOUNT = 2,
  70 };
  71
  72 static int update_block_group(struct btrfs_trans_handle *trans,
  73                               struct btrfs_root *root,
  74                               u64 bytenr, u64 num_bytes, int alloc);
  75 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  76                                 struct btrfs_root *root,
  77                                 u64 bytenr, u64 num_bytes, u64 parent,
  78                                 u64 root_objectid, u64 owner_objectid,
  79                                 u64 owner_offset, int refs_to_drop,
  80                                 struct btrfs_delayed_extent_op *extra_op);
  81 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  82                                     struct extent_buffer *leaf,
  83                                     struct btrfs_extent_item *ei);
  84 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  85                                       struct btrfs_root *root,
  86                                       u64 parent, u64 root_objectid,
  87                                       u64 flags, u64 owner, u64 offset,
  88                                       struct btrfs_key *ins, int ref_mod);
  89 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  90                                      struct btrfs_root *root,
  91                                      u64 parent, u64 root_objectid,
  92                                      u64 flags, struct btrfs_disk_key *key,
  93                                      int level, struct btrfs_key *ins);
  94 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  95                           struct btrfs_root *extent_root, u64 alloc_bytes,
  96                           u64 flags, int force);
  97 static int find_next_key(struct btrfs_path *path, int level,
  98                          struct btrfs_key *key);
  99 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 100                             int dump_block_groups);
 101 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 102                                        u64 num_bytes, int reserve);
 103
 104 static noinline int
 105 block_group_cache_done(struct btrfs_block_group_cache *cache)
 106 {
 107         smp_mb();
 108         return cache->cached == BTRFS_CACHE_FINISHED;
 109 }
 110
 111 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 112 {
 113         return (cache->flags & bits) == bits;
 114 }
 115
 116 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 117 {
 118         atomic_inc(&cache->count);
 119 }
 120
 121 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 122 {
 123         if (atomic_dec_and_test(&cache->count)) {
 124                 WARN_ON(cache->pinned > 0);
 125                 WARN_ON(cache->reserved > 0);
 126                 kfree(cache->free_space_ctl);
 127                 kfree(cache);
 128         }
 129 }
 130
 131 /*
 132  * this adds the block group to the fs_info rb tree for the block group
 133  * cache
 134  */
 135 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 136                                 struct btrfs_block_group_cache *block_group)
 137 {
 138         struct rb_node **p;
 139         struct rb_node *parent = NULL;
 140         struct btrfs_block_group_cache *cache;
 141
 142         spin_lock(&info->block_group_cache_lock);
 143         p = &info->block_group_cache_tree.rb_node;
 144
 145         while (*p) {
 146                 parent = *p;
 147                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 148                                  cache_node);
 149                 if (block_group->key.objectid < cache->key.objectid) {
 150                         p = &(*p)->rb_left;
 151                 } else if (block_group->key.objectid > cache->key.objectid) {
 152                         p = &(*p)->rb_right;
 153                 } else {
 154                         spin_unlock(&info->block_group_cache_lock);
 155                         return -EEXIST;
 156                 }
 157         }
 158
 159         rb_link_node(&block_group->cache_node, parent, p);
 160         rb_insert_color(&block_group->cache_node,
 161                         &info->block_group_cache_tree);
 162         spin_unlock(&info->block_group_cache_lock);
 163
 164         return 0;
 165 }
 166
 167 /*
 168  * This will return the block group at or after bytenr if contains is 0, else
 169  * it will return the block group that contains the bytenr
 170  */
 171 static struct btrfs_block_group_cache *
 172 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 173                               int contains)
 174 {
 175         struct btrfs_block_group_cache *cache, *ret = NULL;
 176         struct rb_node *n;
 177         u64 end, start;
 178
 179         spin_lock(&info->block_group_cache_lock);
 180         n = info->block_group_cache_tree.rb_node;
 181
 182         while (n) {
 183                 cache = rb_entry(n, struct btrfs_block_group_cache,
 184                                  cache_node);
 185                 end = cache->key.objectid + cache->key.offset - 1;
 186                 start = cache->key.objectid;
 187
 188                 if (bytenr < start) {
 189                         if (!contains && (!ret || start < ret->key.objectid))
 190                                 ret = cache;
 191                         n = n->rb_left;
 192                 } else if (bytenr > start) {
 193                         if (contains && bytenr <= end) {
 194                                 ret = cache;
 195                                 break;
 196                         }
 197                         n = n->rb_right;
 198                 } else {
 199                         ret = cache;
 200                         break;
 201                 }
 202         }
 203         if (ret)
 204                 btrfs_get_block_group(ret);
 205         spin_unlock(&info->block_group_cache_lock);
 206
 207         return ret;
 208 }
 209
 210 static int add_excluded_extent(struct btrfs_root *root,
 211                                u64 start, u64 num_bytes)
 212 {
 213         u64 end = start + num_bytes - 1;
 214         set_extent_bits(&root->fs_info->freed_extents[0],
 215                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 216         set_extent_bits(&root->fs_info->freed_extents[1],
 217                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 218         return 0;
 219 }
 220
 221 static void free_excluded_extents(struct btrfs_root *root,
 222                                   struct btrfs_block_group_cache *cache)
 223 {
 224         u64 start, end;
 225
 226         start = cache->key.objectid;
 227         end = start + cache->key.offset - 1;
 228
 229         clear_extent_bits(&root->fs_info->freed_extents[0],
 230                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 231         clear_extent_bits(&root->fs_info->freed_extents[1],
 232                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 233 }
 234
 235 static int exclude_super_stripes(struct btrfs_root *root,
 236                                  struct btrfs_block_group_cache *cache)
 237 {
 238         u64 bytenr;
 239         u64 *logical;
 240         int stripe_len;
 241         int i, nr, ret;
 242
 243         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 244                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 245                 cache->bytes_super += stripe_len;
 246                 ret = add_excluded_extent(root, cache->key.objectid,
 247                                           stripe_len);
 248                 BUG_ON(ret); /* -ENOMEM */
 249         }
 250
 251         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 252                 bytenr = btrfs_sb_offset(i);
 253                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 254                                        cache->key.objectid, bytenr,
 255                                        0, &logical, &nr, &stripe_len);
 256                 BUG_ON(ret); /* -ENOMEM */
 257
 258                 while (nr--) {
 259                         cache->bytes_super += stripe_len;
 260                         ret = add_excluded_extent(root, logical[nr],
 261                                                   stripe_len);
 262                         BUG_ON(ret); /* -ENOMEM */
 263                 }
 264
 265                 kfree(logical);
 266         }
 267         return 0;
 268 }
 269
 270 static struct btrfs_caching_control *
 271 get_caching_control(struct btrfs_block_group_cache *cache)
 272 {
 273         struct btrfs_caching_control *ctl;
 274
 275         spin_lock(&cache->lock);
 276         if (cache->cached != BTRFS_CACHE_STARTED) {
 277                 spin_unlock(&cache->lock);
 278                 return NULL;
 279         }
 280
 281         /* We're loading it the fast way, so we don't have a caching_ctl. */
 282         if (!cache->caching_ctl) {
 283                 spin_unlock(&cache->lock);
 284                 return NULL;
 285         }
 286
 287         ctl = cache->caching_ctl;
 288         atomic_inc(&ctl->count);
 289         spin_unlock(&cache->lock);
 290         return ctl;
 291 }
 292
 293 static void put_caching_control(struct btrfs_caching_control *ctl)
 294 {
 295         if (atomic_dec_and_test(&ctl->count))
 296                 kfree(ctl);
 297 }
 298
 299 /*
 300  * this is only called by cache_block_group, since we could have freed extents
 301  * we need to check the pinned_extents for any extents that can't be used yet
 302  * since their free space will be released as soon as the transaction commits.
 303  */
 304 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 305                               struct btrfs_fs_info *info, u64 start, u64 end)
 306 {
 307         u64 extent_start, extent_end, size, total_added = 0;
 308         int ret;
 309
 310         while (start < end) {
 311                 ret = find_first_extent_bit(info->pinned_extents, start,
 312                                             &extent_start, &extent_end,
 313                                             EXTENT_DIRTY | EXTENT_UPTODATE);
 314                 if (ret)
 315                         break;
 316
 317                 if (extent_start <= start) {
 318                         start = extent_end + 1;
 319                 } else if (extent_start > start && extent_start < end) {
 320                         size = extent_start - start;
 321                         total_added += size;
 322                         ret = btrfs_add_free_space(block_group, start,
 323                                                    size);
 324                         BUG_ON(ret); /* -ENOMEM or logic error */
 325                         start = extent_end + 1;
 326                 } else {
 327                         break;
 328                 }
 329         }
 330
 331         if (start < end) {
 332                 size = end - start;
 333                 total_added += size;
 334                 ret = btrfs_add_free_space(block_group, start, size);
 335                 BUG_ON(ret); /* -ENOMEM or logic error */
 336         }
 337
 338         return total_added;
 339 }
 340
 341 static noinline void caching_thread(struct btrfs_work *work)
 342 {
 343         struct btrfs_block_group_cache *block_group;
 344         struct btrfs_fs_info *fs_info;
 345         struct btrfs_caching_control *caching_ctl;
 346         struct btrfs_root *extent_root;
 347         struct btrfs_path *path;
 348         struct extent_buffer *leaf;
 349         struct btrfs_key key;
 350         u64 total_found = 0;
 351         u64 last = 0;
 352         u32 nritems;
 353         int ret = 0;
 354
 355         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 356         block_group = caching_ctl->block_group;
 357         fs_info = block_group->fs_info;
 358         extent_root = fs_info->extent_root;
 359
 360         path = btrfs_alloc_path();
 361         if (!path)
 362                 goto out;
 363
 364         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 365
 366         /*
 367          * We don't want to deadlock with somebody trying to allocate a new
 368          * extent for the extent root while also trying to search the extent
 369          * root to add free space.  So we skip locking and search the commit
 370          * root, since its read-only
 371          */
 372         path->skip_locking = 1;
 373         path->search_commit_root = 1;
 374         path->reada = 1;
 375
 376         key.objectid = last;
 377         key.offset = 0;
 378         key.type = BTRFS_EXTENT_ITEM_KEY;
 379 again:
 380         mutex_lock(&caching_ctl->mutex);
 381         /* need to make sure the commit_root doesn't disappear */
 382         down_read(&fs_info->extent_commit_sem);
 383
 384         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 385         if (ret < 0)
 386                 goto err;
 387
 388         leaf = path->nodes[0];
 389         nritems = btrfs_header_nritems(leaf);
 390
 391         while (1) {
 392                 if (btrfs_fs_closing(fs_info) > 1) {
 393                         last = (u64)-1;
 394                         break;
 395                 }
 396
 397                 if (path->slots[0] < nritems) {
 398                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 399                 } else {
 400                         ret = find_next_key(path, 0, &key);
 401                         if (ret)
 402                                 break;
 403
 404                         if (need_resched() ||
 405                             btrfs_next_leaf(extent_root, path)) {
 406                                 caching_ctl->progress = last;
 407                                 btrfs_release_path(path);
 408                                 up_read(&fs_info->extent_commit_sem);
 409                                 mutex_unlock(&caching_ctl->mutex);
 410                                 cond_resched();
 411                                 goto again;
 412                         }
 413                         leaf = path->nodes[0];
 414                         nritems = btrfs_header_nritems(leaf);
 415                         continue;
 416                 }
 417
 418                 if (key.objectid < block_group->key.objectid) {
 419                         path->slots[0]++;
 420                         continue;
 421                 }
 422
 423                 if (key.objectid >= block_group->key.objectid +
 424                     block_group->key.offset)
 425                         break;
 426
 427                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 428                         total_found += add_new_free_space(block_group,
 429                                                           fs_info, last,
 430                                                           key.objectid);
 431                         last = key.objectid + key.offset;
 432
 433                         if (total_found > (1024 * 1024 * 2)) {
 434                                 total_found = 0;
 435                                 wake_up(&caching_ctl->wait);
 436                         }
 437                 }
 438                 path->slots[0]++;
 439         }
 440         ret = 0;
 441
 442         total_found += add_new_free_space(block_group, fs_info, last,
 443                                           block_group->key.objectid +
 444                                           block_group->key.offset);
 445         caching_ctl->progress = (u64)-1;
 446
 447         spin_lock(&block_group->lock);
 448         block_group->caching_ctl = NULL;
 449         block_group->cached = BTRFS_CACHE_FINISHED;
 450         spin_unlock(&block_group->lock);
 451
 452 err:
 453         btrfs_free_path(path);
 454         up_read(&fs_info->extent_commit_sem);
 455
 456         free_excluded_extents(extent_root, block_group);
 457
 458         mutex_unlock(&caching_ctl->mutex);
 459 out:
 460         wake_up(&caching_ctl->wait);
 461
 462         put_caching_control(caching_ctl);
 463         btrfs_put_block_group(block_group);
 464 }
 465
 466 static int cache_block_group(struct btrfs_block_group_cache *cache,
 467                              struct btrfs_trans_handle *trans,
 468                              struct btrfs_root *root,
 469                              int load_cache_only)
 470 {
 471         DEFINE_WAIT(wait);
 472         struct btrfs_fs_info *fs_info = cache->fs_info;
 473         struct btrfs_caching_control *caching_ctl;
 474         int ret = 0;
 475
 476         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 477         if (!caching_ctl)
 478                 return -ENOMEM;
 479
 480         INIT_LIST_HEAD(&caching_ctl->list);
 481         mutex_init(&caching_ctl->mutex);
 482         init_waitqueue_head(&caching_ctl->wait);
 483         caching_ctl->block_group = cache;
 484         caching_ctl->progress = cache->key.objectid;
 485         atomic_set(&caching_ctl->count, 1);
 486         caching_ctl->work.func = caching_thread;
 487
 488         spin_lock(&cache->lock);
 489         /*
 490          * This should be a rare occasion, but this could happen I think in the
 491          * case where one thread starts to load the space cache info, and then
 492          * some other thread starts a transaction commit which tries to do an
 493          * allocation while the other thread is still loading the space cache
 494          * info.  The previous loop should have kept us from choosing this block
 495          * group, but if we've moved to the state where we will wait on caching
 496          * block groups we need to first check if we're doing a fast load here,
 497          * so we can wait for it to finish, otherwise we could end up allocating
 498          * from a block group who's cache gets evicted for one reason or
 499          * another.
 500          */
 501         while (cache->cached == BTRFS_CACHE_FAST) {
 502                 struct btrfs_caching_control *ctl;
 503
 504                 ctl = cache->caching_ctl;
 505                 atomic_inc(&ctl->count);
 506                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 507                 spin_unlock(&cache->lock);
 508
 509                 schedule();
 510
 511                 finish_wait(&ctl->wait, &wait);
 512                 put_caching_control(ctl);
 513                 spin_lock(&cache->lock);
 514         }
 515
 516         if (cache->cached != BTRFS_CACHE_NO) {
 517                 spin_unlock(&cache->lock);
 518                 kfree(caching_ctl);
 519                 return 0;
 520         }
 521         WARN_ON(cache->caching_ctl);
 522         cache->caching_ctl = caching_ctl;
 523         cache->cached = BTRFS_CACHE_FAST;
 524         spin_unlock(&cache->lock);
 525
 526         /*
 527          * We can't do the read from on-disk cache during a commit since we need
 528          * to have the normal tree locking.  Also if we are currently trying to
 529          * allocate blocks for the tree root we can't do the fast caching since
 530          * we likely hold important locks.
 531          */
 532         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 533                 ret = load_free_space_cache(fs_info, cache);
 534
 535                 spin_lock(&cache->lock);
 536                 if (ret == 1) {
 537                         cache->caching_ctl = NULL;
 538                         cache->cached = BTRFS_CACHE_FINISHED;
 539                         cache->last_byte_to_unpin = (u64)-1;
 540                 } else {
 541                         if (load_cache_only) {
 542                                 cache->caching_ctl = NULL;
 543                                 cache->cached = BTRFS_CACHE_NO;
 544                         } else {
 545                                 cache->cached = BTRFS_CACHE_STARTED;
 546                         }
 547                 }
 548                 spin_unlock(&cache->lock);
 549                 wake_up(&caching_ctl->wait);
 550                 if (ret == 1) {
 551                         put_caching_control(caching_ctl);
 552                         free_excluded_extents(fs_info->extent_root, cache);
 553                         return 0;
 554                 }
 555         } else {
 556                 /*
 557                  * We are not going to do the fast caching, set cached to the
 558                  * appropriate value and wakeup any waiters.
 559                  */
 560                 spin_lock(&cache->lock);
 561                 if (load_cache_only) {
 562                         cache->caching_ctl = NULL;
 563                         cache->cached = BTRFS_CACHE_NO;
 564                 } else {
 565                         cache->cached = BTRFS_CACHE_STARTED;
 566                 }
 567                 spin_unlock(&cache->lock);
 568                 wake_up(&caching_ctl->wait);
 569         }
 570
 571         if (load_cache_only) {
 572                 put_caching_control(caching_ctl);
 573                 return 0;
 574         }
 575
 576         down_write(&fs_info->extent_commit_sem);
 577         atomic_inc(&caching_ctl->count);
 578         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 579         up_write(&fs_info->extent_commit_sem);
 580
 581         btrfs_get_block_group(cache);
 582
 583         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 584
 585         return ret;
 586 }
 587
 588 /*
 589  * return the block group that starts at or after bytenr
 590  */
 591 static struct btrfs_block_group_cache *
 592 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 593 {
 594         struct btrfs_block_group_cache *cache;
 595
 596         cache = block_group_cache_tree_search(info, bytenr, 0);
 597
 598         return cache;
 599 }
 600
 601 /*
 602  * return the block group that contains the given bytenr
 603  */
 604 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 605                                                  struct btrfs_fs_info *info,
 606                                                  u64 bytenr)
 607 {
 608         struct btrfs_block_group_cache *cache;
 609
 610         cache = block_group_cache_tree_search(info, bytenr, 1);
 611
 612         return cache;
 613 }
 614
 615 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 616                                                   u64 flags)
 617 {
 618         struct list_head *head = &info->space_info;
 619         struct btrfs_space_info *found;
 620
 621         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 622
 623         rcu_read_lock();
 624         list_for_each_entry_rcu(found, head, list) {
 625                 if (found->flags & flags) {
 626                         rcu_read_unlock();
 627                         return found;
 628                 }
 629         }
 630         rcu_read_unlock();
 631         return NULL;
 632 }
 633
 634 /*
 635  * after adding space to the filesystem, we need to clear the full flags
 636  * on all the space infos.
 637  */
 638 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 639 {
 640         struct list_head *head = &info->space_info;
 641         struct btrfs_space_info *found;
 642
 643         rcu_read_lock();
 644         list_for_each_entry_rcu(found, head, list)
 645                 found->full = 0;
 646         rcu_read_unlock();
 647 }
 648
 649 static u64 div_factor(u64 num, int factor)
 650 {
 651         if (factor == 10)
 652                 return num;
 653         num *= factor;
 654         do_div(num, 10);
 655         return num;
 656 }
 657
 658 static u64 div_factor_fine(u64 num, int factor)
 659 {
 660         if (factor == 100)
 661                 return num;
 662         num *= factor;
 663         do_div(num, 100);
 664         return num;
 665 }
 666
 667 u64 btrfs_find_block_group(struct btrfs_root *root,
 668                            u64 search_start, u64 search_hint, int owner)
 669 {
 670         struct btrfs_block_group_cache *cache;
 671         u64 used;
 672         u64 last = max(search_hint, search_start);
 673         u64 group_start = 0;
 674         int full_search = 0;
 675         int factor = 9;
 676         int wrapped = 0;
 677 again:
 678         while (1) {
 679                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
 680                 if (!cache)
 681                         break;
 682
 683                 spin_lock(&cache->lock);
 684                 last = cache->key.objectid + cache->key.offset;
 685                 used = btrfs_block_group_used(&cache->item);
 686
 687                 if ((full_search || !cache->ro) &&
 688                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
 689                         if (used + cache->pinned + cache->reserved <
 690                             div_factor(cache->key.offset, factor)) {
 691                                 group_start = cache->key.objectid;
 692                                 spin_unlock(&cache->lock);
 693                                 btrfs_put_block_group(cache);
 694                                 goto found;
 695                         }
 696                 }
 697                 spin_unlock(&cache->lock);
 698                 btrfs_put_block_group(cache);
 699                 cond_resched();
 700         }
 701         if (!wrapped) {
 702                 last = search_start;
 703                 wrapped = 1;
 704                 goto again;
 705         }
 706         if (!full_search && factor < 10) {
 707                 last = search_start;
 708                 full_search = 1;
 709                 factor = 10;
 710                 goto again;
 711         }
 712 found:
 713         return group_start;
 714 }
 715
 716 /* simple helper to search for an existing extent at a given offset */
 717 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 718 {
 719         int ret;
 720         struct btrfs_key key;
 721         struct btrfs_path *path;
 722
 723         path = btrfs_alloc_path();
 724         if (!path)
 725                 return -ENOMEM;
 726
 727         key.objectid = start;
 728         key.offset = len;
 729         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 730         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 731                                 0, 0);
 732         btrfs_free_path(path);
 733         return ret;
 734 }
 735
 736 /*
 737  * helper function to lookup reference count and flags of extent.
 738  *
 739  * the head node for delayed ref is used to store the sum of all the
 740  * reference count modifications queued up in the rbtree. the head
 741  * node may also store the extent flags to set. This way you can check
 742  * to see what the reference count and extent flags would be if all of
 743  * the delayed refs are not processed.
 744  */
 745 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 746                              struct btrfs_root *root, u64 bytenr,
 747                              u64 num_bytes, u64 *refs, u64 *flags)
 748 {
 749         struct btrfs_delayed_ref_head *head;
 750         struct btrfs_delayed_ref_root *delayed_refs;
 751         struct btrfs_path *path;
 752         struct btrfs_extent_item *ei;
 753         struct extent_buffer *leaf;
 754         struct btrfs_key key;
 755         u32 item_size;
 756         u64 num_refs;
 757         u64 extent_flags;
 758         int ret;
 759
 760         path = btrfs_alloc_path();
 761         if (!path)
 762                 return -ENOMEM;
 763
 764         key.objectid = bytenr;
 765         key.type = BTRFS_EXTENT_ITEM_KEY;
 766         key.offset = num_bytes;
 767         if (!trans) {
 768                 path->skip_locking = 1;
 769                 path->search_commit_root = 1;
 770         }
 771 again:
 772         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 773                                 &key, path, 0, 0);
 774         if (ret < 0)
 775                 goto out_free;
 776
 777         if (ret == 0) {
 778                 leaf = path->nodes[0];
 779                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 780                 if (item_size >= sizeof(*ei)) {
 781                         ei = btrfs_item_ptr(leaf, path->slots[0],
 782                                             struct btrfs_extent_item);
 783                         num_refs = btrfs_extent_refs(leaf, ei);
 784                         extent_flags = btrfs_extent_flags(leaf, ei);
 785                 } else {
 786 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 787                         struct btrfs_extent_item_v0 *ei0;
 788                         BUG_ON(item_size != sizeof(*ei0));
 789                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 790                                              struct btrfs_extent_item_v0);
 791                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 792                         /* FIXME: this isn't correct for data */
 793                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 794 #else
 795                         BUG();
 796 #endif
 797                 }
 798                 BUG_ON(num_refs == 0);
 799         } else {
 800                 num_refs = 0;
 801                 extent_flags = 0;
 802                 ret = 0;
 803         }
 804
 805         if (!trans)
 806                 goto out;
 807
 808         delayed_refs = &trans->transaction->delayed_refs;
 809         spin_lock(&delayed_refs->lock);
 810         head = btrfs_find_delayed_ref_head(trans, bytenr);
 811         if (head) {
 812                 if (!mutex_trylock(&head->mutex)) {
 813                         atomic_inc(&head->node.refs);
 814                         spin_unlock(&delayed_refs->lock);
 815
 816                         btrfs_release_path(path);
 817
 818                         /*
 819                          * Mutex was contended, block until it's released and try
 820                          * again
 821                          */
 822                         mutex_lock(&head->mutex);
 823                         mutex_unlock(&head->mutex);
 824                         btrfs_put_delayed_ref(&head->node);
 825                         goto again;
 826                 }
 827                 if (head->extent_op && head->extent_op->update_flags)
 828                         extent_flags |= head->extent_op->flags_to_set;
 829                 else
 830                         BUG_ON(num_refs == 0);
 831
 832                 num_refs += head->node.ref_mod;
 833                 mutex_unlock(&head->mutex);
 834         }
 835         spin_unlock(&delayed_refs->lock);
 836 out:
 837         WARN_ON(num_refs == 0);
 838         if (refs)
 839                 *refs = num_refs;
 840         if (flags)
 841                 *flags = extent_flags;
 842 out_free:
 843         btrfs_free_path(path);
 844         return ret;
 845 }
 846
 847 /*
 848  * Back reference rules.  Back refs have three main goals:
 849  *
 850  * 1) differentiate between all holders of references to an extent so that
 851  *    when a reference is dropped we can make sure it was a valid reference
 852  *    before freeing the extent.
 853  *
 854  * 2) Provide enough information to quickly find the holders of an extent
 855  *    if we notice a given block is corrupted or bad.
 856  *
 857  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 858  *    maintenance.  This is actually the same as #2, but with a slightly
 859  *    different use case.
 860  *
 861  * There are two kinds of back refs. The implicit back refs is optimized
 862  * for pointers in non-shared tree blocks. For a given pointer in a block,
 863  * back refs of this kind provide information about the block's owner tree
 864  * and the pointer's key. These information allow us to find the block by
 865  * b-tree searching. The full back refs is for pointers in tree blocks not
 866  * referenced by their owner trees. The location of tree block is recorded
 867  * in the back refs. Actually the full back refs is generic, and can be
 868  * used in all cases the implicit back refs is used. The major shortcoming
 869  * of the full back refs is its overhead. Every time a tree block gets
 870  * COWed, we have to update back refs entry for all pointers in it.
 871  *
 872  * For a newly allocated tree block, we use implicit back refs for
 873  * pointers in it. This means most tree related operations only involve
 874  * implicit back refs. For a tree block created in old transaction, the
 875  * only way to drop a reference to it is COW it. So we can detect the
 876  * event that tree block loses its owner tree's reference and do the
 877  * back refs conversion.
 878  *
 879  * When a tree block is COW'd through a tree, there are four cases:
 880  *
 881  * The reference count of the block is one and the tree is the block's
 882  * owner tree. Nothing to do in this case.
 883  *
 884  * The reference count of the block is one and the tree is not the
 885  * block's owner tree. In this case, full back refs is used for pointers
 886  * in the block. Remove these full back refs, add implicit back refs for
 887  * every pointers in the new block.
 888  *
 889  * The reference count of the block is greater than one and the tree is
 890  * the block's owner tree. In this case, implicit back refs is used for
 891  * pointers in the block. Add full back refs for every pointers in the
 892  * block, increase lower level extents' reference counts. The original
 893  * implicit back refs are entailed to the new block.
 894  *
 895  * The reference count of the block is greater than one and the tree is
 896  * not the block's owner tree. Add implicit back refs for every pointer in
 897  * the new block, increase lower level extents' reference count.
 898  *
 899  * Back Reference Key composing:
 900  *
 901  * The key objectid corresponds to the first byte in the extent,
 902  * The key type is used to differentiate between types of back refs.
 903  * There are different meanings of the key offset for different types
 904  * of back refs.
 905  *
 906  * File extents can be referenced by:
 907  *
 908  * - multiple snapshots, subvolumes, or different generations in one subvol
 909  * - different files inside a single subvolume
 910  * - different offsets inside a file (bookend extents in file.c)
 911  *
 912  * The extent ref structure for the implicit back refs has fields for:
 913  *
 914  * - Objectid of the subvolume root
 915  * - objectid of the file holding the reference
 916  * - original offset in the file
 917  * - how many bookend extents
 918  *
 919  * The key offset for the implicit back refs is hash of the first
 920  * three fields.
 921  *
 922  * The extent ref structure for the full back refs has field for:
 923  *
 924  * - number of pointers in the tree leaf
 925  *
 926  * The key offset for the implicit back refs is the first byte of
 927  * the tree leaf
 928  *
 929  * When a file extent is allocated, The implicit back refs is used.
 930  * the fields are filled in:
 931  *
 932  *     (root_key.objectid, inode objectid, offset in file, 1)
 933  *
 934  * When a file extent is removed file truncation, we find the
 935  * corresponding implicit back refs and check the following fields:
 936  *
 937  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 938  *
 939  * Btree extents can be referenced by:
 940  *
 941  * - Different subvolumes
 942  *
 943  * Both the implicit back refs and the full back refs for tree blocks
 944  * only consist of key. The key offset for the implicit back refs is
 945  * objectid of block's owner tree. The key offset for the full back refs
 946  * is the first byte of parent block.
 947  *
 948  * When implicit back refs is used, information about the lowest key and
 949  * level of the tree block are required. These information are stored in
 950  * tree block info structure.
 951  */
 952
 953 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 954 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 955                                   struct btrfs_root *root,
 956                                   struct btrfs_path *path,
 957                                   u64 owner, u32 extra_size)
 958 {
 959         struct btrfs_extent_item *item;
 960         struct btrfs_extent_item_v0 *ei0;
 961         struct btrfs_extent_ref_v0 *ref0;
 962         struct btrfs_tree_block_info *bi;
 963         struct extent_buffer *leaf;
 964         struct btrfs_key key;
 965         struct btrfs_key found_key;
 966         u32 new_size = sizeof(*item);
 967         u64 refs;
 968         int ret;
 969
 970         leaf = path->nodes[0];
 971         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 972
 973         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 974         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 975                              struct btrfs_extent_item_v0);
 976         refs = btrfs_extent_refs_v0(leaf, ei0);
 977
 978         if (owner == (u64)-1) {
 979                 while (1) {
 980                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 981                                 ret = btrfs_next_leaf(root, path);
 982                                 if (ret < 0)
 983                                         return ret;
 984                                 BUG_ON(ret > 0); /* Corruption */
 985                                 leaf = path->nodes[0];
 986                         }
 987                         btrfs_item_key_to_cpu(leaf, &found_key,
 988                                               path->slots[0]);
 989                         BUG_ON(key.objectid != found_key.objectid);
 990                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 991                                 path->slots[0]++;
 992                                 continue;
 993                         }
 994                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 995                                               struct btrfs_extent_ref_v0);
 996                         owner = btrfs_ref_objectid_v0(leaf, ref0);
 997                         break;
 998                 }
 999         }
1000         btrfs_release_path(path);
1001
1002         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1003                 new_size += sizeof(*bi);
1004
1005         new_size -= sizeof(*ei0);
1006         ret = btrfs_search_slot(trans, root, &key, path,
1007                                 new_size + extra_size, 1);
1008         if (ret < 0)
1009                 return ret;
1010         BUG_ON(ret); /* Corruption */
1011
1012         btrfs_extend_item(trans, root, path, new_size);
1013
1014         leaf = path->nodes[0];
1015         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1016         btrfs_set_extent_refs(leaf, item, refs);
1017         /* FIXME: get real generation */
1018         btrfs_set_extent_generation(leaf, item, 0);
1019         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1020                 btrfs_set_extent_flags(leaf, item,
1021                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1022                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1023                 bi = (struct btrfs_tree_block_info *)(item + 1);
1024                 /* FIXME: get first key of the block */
1025                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1026                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1027         } else {
1028                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1029         }
1030         btrfs_mark_buffer_dirty(leaf);
1031         return 0;
1032 }
1033 #endif
1034
1035 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1036 {
1037         u32 high_crc = ~(u32)0;
1038         u32 low_crc = ~(u32)0;
1039         __le64 lenum;
1040
1041         lenum = cpu_to_le64(root_objectid);
1042         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1043         lenum = cpu_to_le64(owner);
1044         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1045         lenum = cpu_to_le64(offset);
1046         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1047
1048         return ((u64)high_crc << 31) ^ (u64)low_crc;
1049 }
1050
1051 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1052                                      struct btrfs_extent_data_ref *ref)
1053 {
1054         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1055                                     btrfs_extent_data_ref_objectid(leaf, ref),
1056                                     btrfs_extent_data_ref_offset(leaf, ref));
1057 }
1058
1059 static int match_extent_data_ref(struct extent_buffer *leaf,
1060                                  struct btrfs_extent_data_ref *ref,
1061                                  u64 root_objectid, u64 owner, u64 offset)
1062 {
1063         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1064             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1065             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1066                 return 0;
1067         return 1;
1068 }
1069
1070 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1071                                            struct btrfs_root *root,
1072                                            struct btrfs_path *path,
1073                                            u64 bytenr, u64 parent,
1074                                            u64 root_objectid,
1075                                            u64 owner, u64 offset)
1076 {
1077         struct btrfs_key key;
1078         struct btrfs_extent_data_ref *ref;
1079         struct extent_buffer *leaf;
1080         u32 nritems;
1081         int ret;
1082         int recow;
1083         int err = -ENOENT;
1084
1085         key.objectid = bytenr;
1086         if (parent) {
1087                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1088                 key.offset = parent;
1089         } else {
1090                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1091                 key.offset = hash_extent_data_ref(root_objectid,
1092                                                   owner, offset);
1093         }
1094 again:
1095         recow = 0;
1096         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1097         if (ret < 0) {
1098                 err = ret;
1099                 goto fail;
1100         }
1101
1102         if (parent) {
1103                 if (!ret)
1104                         return 0;
1105 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1106                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1107                 btrfs_release_path(path);
1108                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1109                 if (ret < 0) {
1110                         err = ret;
1111                         goto fail;
1112                 }
1113                 if (!ret)
1114                         return 0;
1115 #endif
1116                 goto fail;
1117         }
1118
1119         leaf = path->nodes[0];
1120         nritems = btrfs_header_nritems(leaf);
1121         while (1) {
1122                 if (path->slots[0] >= nritems) {
1123                         ret = btrfs_next_leaf(root, path);
1124                         if (ret < 0)
1125                                 err = ret;
1126                         if (ret)
1127                                 goto fail;
1128
1129                         leaf = path->nodes[0];
1130                         nritems = btrfs_header_nritems(leaf);
1131                         recow = 1;
1132                 }
1133
1134                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1135                 if (key.objectid != bytenr ||
1136                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1137                         goto fail;
1138
1139                 ref = btrfs_item_ptr(leaf, path->slots[0],
1140                                      struct btrfs_extent_data_ref);
1141
1142                 if (match_extent_data_ref(leaf, ref, root_objectid,
1143                                           owner, offset)) {
1144                         if (recow) {
1145                                 btrfs_release_path(path);
1146                                 goto again;
1147                         }
1148                         err = 0;
1149                         break;
1150                 }
1151                 path->slots[0]++;
1152         }
1153 fail:
1154         return err;
1155 }
1156
1157 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1158                                            struct btrfs_root *root,
1159                                            struct btrfs_path *path,
1160                                            u64 bytenr, u64 parent,
1161                                            u64 root_objectid, u64 owner,
1162                                            u64 offset, int refs_to_add)
1163 {
1164         struct btrfs_key key;
1165         struct extent_buffer *leaf;
1166         u32 size;
1167         u32 num_refs;
1168         int ret;
1169
1170         key.objectid = bytenr;
1171         if (parent) {
1172                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1173                 key.offset = parent;
1174                 size = sizeof(struct btrfs_shared_data_ref);
1175         } else {
1176                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1177                 key.offset = hash_extent_data_ref(root_objectid,
1178                                                   owner, offset);
1179                 size = sizeof(struct btrfs_extent_data_ref);
1180         }
1181
1182         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1183         if (ret && ret != -EEXIST)
1184                 goto fail;
1185
1186         leaf = path->nodes[0];
1187         if (parent) {
1188                 struct btrfs_shared_data_ref *ref;
1189                 ref = btrfs_item_ptr(leaf, path->slots[0],
1190                                      struct btrfs_shared_data_ref);
1191                 if (ret == 0) {
1192                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1193                 } else {
1194                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1195                         num_refs += refs_to_add;
1196                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1197                 }
1198         } else {
1199                 struct btrfs_extent_data_ref *ref;
1200                 while (ret == -EEXIST) {
1201                         ref = btrfs_item_ptr(leaf, path->slots[0],
1202                                              struct btrfs_extent_data_ref);
1203                         if (match_extent_data_ref(leaf, ref, root_objectid,
1204                                                   owner, offset))
1205                                 break;
1206                         btrfs_release_path(path);
1207                         key.offset++;
1208                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1209                                                       size);
1210                         if (ret && ret != -EEXIST)
1211                                 goto fail;
1212
1213                         leaf = path->nodes[0];
1214                 }
1215                 ref = btrfs_item_ptr(leaf, path->slots[0],
1216                                      struct btrfs_extent_data_ref);
1217                 if (ret == 0) {
1218                         btrfs_set_extent_data_ref_root(leaf, ref,
1219                                                        root_objectid);
1220                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1221                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1222                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1223                 } else {
1224                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1225                         num_refs += refs_to_add;
1226                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1227                 }
1228         }
1229         btrfs_mark_buffer_dirty(leaf);
1230         ret = 0;
1231 fail:
1232         btrfs_release_path(path);
1233         return ret;
1234 }
1235
1236 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1237                                            struct btrfs_root *root,
1238                                            struct btrfs_path *path,
1239                                            int refs_to_drop)
1240 {
1241         struct btrfs_key key;
1242         struct btrfs_extent_data_ref *ref1 = NULL;
1243         struct btrfs_shared_data_ref *ref2 = NULL;
1244         struct extent_buffer *leaf;
1245         u32 num_refs = 0;
1246         int ret = 0;
1247
1248         leaf = path->nodes[0];
1249         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1250
1251         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1252                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1253                                       struct btrfs_extent_data_ref);
1254                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1255         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1256                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1257                                       struct btrfs_shared_data_ref);
1258                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1259 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1260         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1261                 struct btrfs_extent_ref_v0 *ref0;
1262                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1263                                       struct btrfs_extent_ref_v0);
1264                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1265 #endif
1266         } else {
1267                 BUG();
1268         }
1269
1270         BUG_ON(num_refs < refs_to_drop);
1271         num_refs -= refs_to_drop;
1272
1273         if (num_refs == 0) {
1274                 ret = btrfs_del_item(trans, root, path);
1275         } else {
1276                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1277                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1278                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1279                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1280 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1281                 else {
1282                         struct btrfs_extent_ref_v0 *ref0;
1283                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1284                                         struct btrfs_extent_ref_v0);
1285                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1286                 }
1287 #endif
1288                 btrfs_mark_buffer_dirty(leaf);
1289         }
1290         return ret;
1291 }
1292
1293 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1294                                           struct btrfs_path *path,
1295                                           struct btrfs_extent_inline_ref *iref)
1296 {
1297         struct btrfs_key key;
1298         struct extent_buffer *leaf;
1299         struct btrfs_extent_data_ref *ref1;
1300         struct btrfs_shared_data_ref *ref2;
1301         u32 num_refs = 0;
1302
1303         leaf = path->nodes[0];
1304         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1305         if (iref) {
1306                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1307                     BTRFS_EXTENT_DATA_REF_KEY) {
1308                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1309                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1310                 } else {
1311                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1312                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1313                 }
1314         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1315                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1316                                       struct btrfs_extent_data_ref);
1317                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1318         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1319                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1320                                       struct btrfs_shared_data_ref);
1321                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1322 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1323         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1324                 struct btrfs_extent_ref_v0 *ref0;
1325                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1326                                       struct btrfs_extent_ref_v0);
1327                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1328 #endif
1329         } else {
1330                 WARN_ON(1);
1331         }
1332         return num_refs;
1333 }
1334
1335 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1336                                           struct btrfs_root *root,
1337                                           struct btrfs_path *path,
1338                                           u64 bytenr, u64 parent,
1339                                           u64 root_objectid)
1340 {
1341         struct btrfs_key key;
1342         int ret;
1343
1344         key.objectid = bytenr;
1345         if (parent) {
1346                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1347                 key.offset = parent;
1348         } else {
1349                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1350                 key.offset = root_objectid;
1351         }
1352
1353         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1354         if (ret > 0)
1355                 ret = -ENOENT;
1356 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1357         if (ret == -ENOENT && parent) {
1358                 btrfs_release_path(path);
1359                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1360                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1361                 if (ret > 0)
1362                         ret = -ENOENT;
1363         }
1364 #endif
1365         return ret;
1366 }
1367
1368 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1369                                           struct btrfs_root *root,
1370                                           struct btrfs_path *path,
1371                                           u64 bytenr, u64 parent,
1372                                           u64 root_objectid)
1373 {
1374         struct btrfs_key key;
1375         int ret;
1376
1377         key.objectid = bytenr;
1378         if (parent) {
1379                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1380                 key.offset = parent;
1381         } else {
1382                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1383                 key.offset = root_objectid;
1384         }
1385
1386         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1387         btrfs_release_path(path);
1388         return ret;
1389 }
1390
1391 static inline int extent_ref_type(u64 parent, u64 owner)
1392 {
1393         int type;
1394         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1395                 if (parent > 0)
1396                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1397                 else
1398                         type = BTRFS_TREE_BLOCK_REF_KEY;
1399         } else {
1400                 if (parent > 0)
1401                         type = BTRFS_SHARED_DATA_REF_KEY;
1402                 else
1403                         type = BTRFS_EXTENT_DATA_REF_KEY;
1404         }
1405         return type;
1406 }
1407
1408 static int find_next_key(struct btrfs_path *path, int level,
1409                          struct btrfs_key *key)
1410
1411 {
1412         for (; level < BTRFS_MAX_LEVEL; level++) {
1413                 if (!path->nodes[level])
1414                         break;
1415                 if (path->slots[level] + 1 >=
1416                     btrfs_header_nritems(path->nodes[level]))
1417                         continue;
1418                 if (level == 0)
1419                         btrfs_item_key_to_cpu(path->nodes[level], key,
1420                                               path->slots[level] + 1);
1421                 else
1422                         btrfs_node_key_to_cpu(path->nodes[level], key,
1423                                               path->slots[level] + 1);
1424                 return 0;
1425         }
1426         return 1;
1427 }
1428
1429 /*
1430  * look for inline back ref. if back ref is found, *ref_ret is set
1431  * to the address of inline back ref, and 0 is returned.
1432  *
1433  * if back ref isn't found, *ref_ret is set to the address where it
1434  * should be inserted, and -ENOENT is returned.
1435  *
1436  * if insert is true and there are too many inline back refs, the path
1437  * points to the extent item, and -EAGAIN is returned.
1438  *
1439  * NOTE: inline back refs are ordered in the same way that back ref
1440  *       items in the tree are ordered.
1441  */
1442 static noinline_for_stack
1443 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1444                                  struct btrfs_root *root,
1445                                  struct btrfs_path *path,
1446                                  struct btrfs_extent_inline_ref **ref_ret,
1447                                  u64 bytenr, u64 num_bytes,
1448                                  u64 parent, u64 root_objectid,
1449                                  u64 owner, u64 offset, int insert)
1450 {
1451         struct btrfs_key key;
1452         struct extent_buffer *leaf;
1453         struct btrfs_extent_item *ei;
1454         struct btrfs_extent_inline_ref *iref;
1455         u64 flags;
1456         u64 item_size;
1457         unsigned long ptr;
1458         unsigned long end;
1459         int extra_size;
1460         int type;
1461         int want;
1462         int ret;
1463         int err = 0;
1464
1465         key.objectid = bytenr;
1466         key.type = BTRFS_EXTENT_ITEM_KEY;
1467         key.offset = num_bytes;
1468
1469         want = extent_ref_type(parent, owner);
1470         if (insert) {
1471                 extra_size = btrfs_extent_inline_ref_size(want);
1472                 path->keep_locks = 1;
1473         } else
1474                 extra_size = -1;
1475         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1476         if (ret < 0) {
1477                 err = ret;
1478                 goto out;
1479         }
1480         if (ret && !insert) {
1481                 err = -ENOENT;
1482                 goto out;
1483         }
1484         BUG_ON(ret); /* Corruption */
1485
1486         leaf = path->nodes[0];
1487         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1488 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1489         if (item_size < sizeof(*ei)) {
1490                 if (!insert) {
1491                         err = -ENOENT;
1492                         goto out;
1493                 }
1494                 ret = convert_extent_item_v0(trans, root, path, owner,
1495                                              extra_size);
1496                 if (ret < 0) {
1497                         err = ret;
1498                         goto out;
1499                 }
1500                 leaf = path->nodes[0];
1501                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1502         }
1503 #endif
1504         BUG_ON(item_size < sizeof(*ei));
1505
1506         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1507         flags = btrfs_extent_flags(leaf, ei);
1508
1509         ptr = (unsigned long)(ei + 1);
1510         end = (unsigned long)ei + item_size;
1511
1512         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1513                 ptr += sizeof(struct btrfs_tree_block_info);
1514                 BUG_ON(ptr > end);
1515         } else {
1516                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1517         }
1518
1519         err = -ENOENT;
1520         while (1) {
1521                 if (ptr >= end) {
1522                         WARN_ON(ptr > end);
1523                         break;
1524                 }
1525                 iref = (struct btrfs_extent_inline_ref *)ptr;
1526                 type = btrfs_extent_inline_ref_type(leaf, iref);
1527                 if (want < type)
1528                         break;
1529                 if (want > type) {
1530                         ptr += btrfs_extent_inline_ref_size(type);
1531                         continue;
1532                 }
1533
1534                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1535                         struct btrfs_extent_data_ref *dref;
1536                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1537                         if (match_extent_data_ref(leaf, dref, root_objectid,
1538                                                   owner, offset)) {
1539                                 err = 0;
1540                                 break;
1541                         }
1542                         if (hash_extent_data_ref_item(leaf, dref) <
1543                             hash_extent_data_ref(root_objectid, owner, offset))
1544                                 break;
1545                 } else {
1546                         u64 ref_offset;
1547                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1548                         if (parent > 0) {
1549                                 if (parent == ref_offset) {
1550                                         err = 0;
1551                                         break;
1552                                 }
1553                                 if (ref_offset < parent)
1554                                         break;
1555                         } else {
1556                                 if (root_objectid == ref_offset) {
1557                                         err = 0;
1558                                         break;
1559                                 }
1560                                 if (ref_offset < root_objectid)
1561                                         break;
1562                         }
1563                 }
1564                 ptr += btrfs_extent_inline_ref_size(type);
1565         }
1566         if (err == -ENOENT && insert) {
1567                 if (item_size + extra_size >=
1568                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1569                         err = -EAGAIN;
1570                         goto out;
1571                 }
1572                 /*
1573                  * To add new inline back ref, we have to make sure
1574                  * there is no corresponding back ref item.
1575                  * For simplicity, we just do not add new inline back
1576                  * ref if there is any kind of item for this block
1577                  */
1578                 if (find_next_key(path, 0, &key) == 0 &&
1579                     key.objectid == bytenr &&
1580                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1581                         err = -EAGAIN;
1582                         goto out;
1583                 }
1584         }
1585         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1586 out:
1587         if (insert) {
1588                 path->keep_locks = 0;
1589                 btrfs_unlock_up_safe(path, 1);
1590         }
1591         return err;
1592 }
1593
1594 /*
1595  * helper to add new inline back ref
1596  */
1597 static noinline_for_stack
1598 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1599                                  struct btrfs_root *root,
1600                                  struct btrfs_path *path,
1601                                  struct btrfs_extent_inline_ref *iref,
1602                                  u64 parent, u64 root_objectid,
1603                                  u64 owner, u64 offset, int refs_to_add,
1604                                  struct btrfs_delayed_extent_op *extent_op)
1605 {
1606         struct extent_buffer *leaf;
1607         struct btrfs_extent_item *ei;
1608         unsigned long ptr;
1609         unsigned long end;
1610         unsigned long item_offset;
1611         u64 refs;
1612         int size;
1613         int type;
1614
1615         leaf = path->nodes[0];
1616         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1617         item_offset = (unsigned long)iref - (unsigned long)ei;
1618
1619         type = extent_ref_type(parent, owner);
1620         size = btrfs_extent_inline_ref_size(type);
1621
1622         btrfs_extend_item(trans, root, path, size);
1623
1624         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1625         refs = btrfs_extent_refs(leaf, ei);
1626         refs += refs_to_add;
1627         btrfs_set_extent_refs(leaf, ei, refs);
1628         if (extent_op)
1629                 __run_delayed_extent_op(extent_op, leaf, ei);
1630
1631         ptr = (unsigned long)ei + item_offset;
1632         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1633         if (ptr < end - size)
1634                 memmove_extent_buffer(leaf, ptr + size, ptr,
1635                                       end - size - ptr);
1636
1637         iref = (struct btrfs_extent_inline_ref *)ptr;
1638         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1639         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1640                 struct btrfs_extent_data_ref *dref;
1641                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1642                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1643                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1644                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1645                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1646         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1647                 struct btrfs_shared_data_ref *sref;
1648                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1649                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1650                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1651         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1652                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1653         } else {
1654                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1655         }
1656         btrfs_mark_buffer_dirty(leaf);
1657 }
1658
1659 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1660                                  struct btrfs_root *root,
1661                                  struct btrfs_path *path,
1662                                  struct btrfs_extent_inline_ref **ref_ret,
1663                                  u64 bytenr, u64 num_bytes, u64 parent,
1664                                  u64 root_objectid, u64 owner, u64 offset)
1665 {
1666         int ret;
1667
1668         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1669                                            bytenr, num_bytes, parent,
1670                                            root_objectid, owner, offset, 0);
1671         if (ret != -ENOENT)
1672                 return ret;
1673
1674         btrfs_release_path(path);
1675         *ref_ret = NULL;
1676
1677         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1678                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1679                                             root_objectid);
1680         } else {
1681                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1682                                              root_objectid, owner, offset);
1683         }
1684         return ret;
1685 }
1686
1687 /*
1688  * helper to update/remove inline back ref
1689  */
1690 static noinline_for_stack
1691 void update_inline_extent_backref(struct btrfs_trans_handle *trans,
1692                                   struct btrfs_root *root,
1693                                   struct btrfs_path *path,
1694                                   struct btrfs_extent_inline_ref *iref,
1695                                   int refs_to_mod,
1696                                   struct btrfs_delayed_extent_op *extent_op)
1697 {
1698         struct extent_buffer *leaf;
1699         struct btrfs_extent_item *ei;
1700         struct btrfs_extent_data_ref *dref = NULL;
1701         struct btrfs_shared_data_ref *sref = NULL;
1702         unsigned long ptr;
1703         unsigned long end;
1704         u32 item_size;
1705         int size;
1706         int type;
1707         u64 refs;
1708
1709         leaf = path->nodes[0];
1710         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1711         refs = btrfs_extent_refs(leaf, ei);
1712         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1713         refs += refs_to_mod;
1714         btrfs_set_extent_refs(leaf, ei, refs);
1715         if (extent_op)
1716                 __run_delayed_extent_op(extent_op, leaf, ei);
1717
1718         type = btrfs_extent_inline_ref_type(leaf, iref);
1719
1720         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1721                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1722                 refs = btrfs_extent_data_ref_count(leaf, dref);
1723         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1724                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1725                 refs = btrfs_shared_data_ref_count(leaf, sref);
1726         } else {
1727                 refs = 1;
1728                 BUG_ON(refs_to_mod != -1);
1729         }
1730
1731         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1732         refs += refs_to_mod;
1733
1734         if (refs > 0) {
1735                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1736                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1737                 else
1738                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1739         } else {
1740                 size =  btrfs_extent_inline_ref_size(type);
1741                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1742                 ptr = (unsigned long)iref;
1743                 end = (unsigned long)ei + item_size;
1744                 if (ptr + size < end)
1745                         memmove_extent_buffer(leaf, ptr, ptr + size,
1746                                               end - ptr - size);
1747                 item_size -= size;
1748                 btrfs_truncate_item(trans, root, path, item_size, 1);
1749         }
1750         btrfs_mark_buffer_dirty(leaf);
1751 }
1752
1753 static noinline_for_stack
1754 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1755                                  struct btrfs_root *root,
1756                                  struct btrfs_path *path,
1757                                  u64 bytenr, u64 num_bytes, u64 parent,
1758                                  u64 root_objectid, u64 owner,
1759                                  u64 offset, int refs_to_add,
1760                                  struct btrfs_delayed_extent_op *extent_op)
1761 {
1762         struct btrfs_extent_inline_ref *iref;
1763         int ret;
1764
1765         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1766                                            bytenr, num_bytes, parent,
1767                                            root_objectid, owner, offset, 1);
1768         if (ret == 0) {
1769                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1770                 update_inline_extent_backref(trans, root, path, iref,
1771                                              refs_to_add, extent_op);
1772         } else if (ret == -ENOENT) {
1773                 setup_inline_extent_backref(trans, root, path, iref, parent,
1774                                             root_objectid, owner, offset,
1775                                             refs_to_add, extent_op);
1776                 ret = 0;
1777         }
1778         return ret;
1779 }
1780
1781 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1782                                  struct btrfs_root *root,
1783                                  struct btrfs_path *path,
1784                                  u64 bytenr, u64 parent, u64 root_objectid,
1785                                  u64 owner, u64 offset, int refs_to_add)
1786 {
1787         int ret;
1788         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1789                 BUG_ON(refs_to_add != 1);
1790                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1791                                             parent, root_objectid);
1792         } else {
1793                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1794                                              parent, root_objectid,
1795                                              owner, offset, refs_to_add);
1796         }
1797         return ret;
1798 }
1799
1800 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1801                                  struct btrfs_root *root,
1802                                  struct btrfs_path *path,
1803                                  struct btrfs_extent_inline_ref *iref,
1804                                  int refs_to_drop, int is_data)
1805 {
1806         int ret = 0;
1807
1808         BUG_ON(!is_data && refs_to_drop != 1);
1809         if (iref) {
1810                 update_inline_extent_backref(trans, root, path, iref,
1811                                              -refs_to_drop, NULL);
1812         } else if (is_data) {
1813                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1814         } else {
1815                 ret = btrfs_del_item(trans, root, path);
1816         }
1817         return ret;
1818 }
1819
1820 static int btrfs_issue_discard(struct block_device *bdev,
1821                                 u64 start, u64 len)
1822 {
1823         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1824 }
1825
1826 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1827                                 u64 num_bytes, u64 *actual_bytes)
1828 {
1829         int ret;
1830         u64 discarded_bytes = 0;
1831         struct btrfs_bio *bbio = NULL;
1832
1833
1834         /* Tell the block device(s) that the sectors can be discarded */
1835         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1836                               bytenr, &num_bytes, &bbio, 0);
1837         /* Error condition is -ENOMEM */
1838         if (!ret) {
1839                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1840                 int i;
1841
1842
1843                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1844                         if (!stripe->dev->can_discard)
1845                                 continue;
1846
1847                         ret = btrfs_issue_discard(stripe->dev->bdev,
1848                                                   stripe->physical,
1849                                                   stripe->length);
1850                         if (!ret)
1851                                 discarded_bytes += stripe->length;
1852                         else if (ret != -EOPNOTSUPP)
1853                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1854
1855                         /*
1856                          * Just in case we get back EOPNOTSUPP for some reason,
1857                          * just ignore the return value so we don't screw up
1858                          * people calling discard_extent.
1859                          */
1860                         ret = 0;
1861                 }
1862                 kfree(bbio);
1863         }
1864
1865         if (actual_bytes)
1866                 *actual_bytes = discarded_bytes;
1867
1868
1869         return ret;
1870 }
1871
1872 /* Can return -ENOMEM */
1873 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1874                          struct btrfs_root *root,
1875                          u64 bytenr, u64 num_bytes, u64 parent,
1876                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1877 {
1878         int ret;
1879         struct btrfs_fs_info *fs_info = root->fs_info;
1880
1881         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1882                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1883
1884         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1885                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1886                                         num_bytes,
1887                                         parent, root_objectid, (int)owner,
1888                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1889         } else {
1890                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1891                                         num_bytes,
1892                                         parent, root_objectid, owner, offset,
1893                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1894         }
1895         return ret;
1896 }
1897
1898 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1899                                   struct btrfs_root *root,
1900                                   u64 bytenr, u64 num_bytes,
1901                                   u64 parent, u64 root_objectid,
1902                                   u64 owner, u64 offset, int refs_to_add,
1903                                   struct btrfs_delayed_extent_op *extent_op)
1904 {
1905         struct btrfs_path *path;
1906         struct extent_buffer *leaf;
1907         struct btrfs_extent_item *item;
1908         u64 refs;
1909         int ret;
1910         int err = 0;
1911
1912         path = btrfs_alloc_path();
1913         if (!path)
1914                 return -ENOMEM;
1915
1916         path->reada = 1;
1917         path->leave_spinning = 1;
1918         /* this will setup the path even if it fails to insert the back ref */
1919         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1920                                            path, bytenr, num_bytes, parent,
1921                                            root_objectid, owner, offset,
1922                                            refs_to_add, extent_op);
1923         if (ret == 0)
1924                 goto out;
1925
1926         if (ret != -EAGAIN) {
1927                 err = ret;
1928                 goto out;
1929         }
1930
1931         leaf = path->nodes[0];
1932         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1933         refs = btrfs_extent_refs(leaf, item);
1934         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1935         if (extent_op)
1936                 __run_delayed_extent_op(extent_op, leaf, item);
1937
1938         btrfs_mark_buffer_dirty(leaf);
1939         btrfs_release_path(path);
1940
1941         path->reada = 1;
1942         path->leave_spinning = 1;
1943
1944         /* now insert the actual backref */
1945         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1946                                     path, bytenr, parent, root_objectid,
1947                                     owner, offset, refs_to_add);
1948         if (ret)
1949                 btrfs_abort_transaction(trans, root, ret);
1950 out:
1951         btrfs_free_path(path);
1952         return err;
1953 }
1954
1955 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1956                                 struct btrfs_root *root,
1957                                 struct btrfs_delayed_ref_node *node,
1958                                 struct btrfs_delayed_extent_op *extent_op,
1959                                 int insert_reserved)
1960 {
1961         int ret = 0;
1962         struct btrfs_delayed_data_ref *ref;
1963         struct btrfs_key ins;
1964         u64 parent = 0;
1965         u64 ref_root = 0;
1966         u64 flags = 0;
1967
1968         ins.objectid = node->bytenr;
1969         ins.offset = node->num_bytes;
1970         ins.type = BTRFS_EXTENT_ITEM_KEY;
1971
1972         ref = btrfs_delayed_node_to_data_ref(node);
1973         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1974                 parent = ref->parent;
1975         else
1976                 ref_root = ref->root;
1977
1978         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1979                 if (extent_op) {
1980                         BUG_ON(extent_op->update_key);
1981                         flags |= extent_op->flags_to_set;
1982                 }
1983                 ret = alloc_reserved_file_extent(trans, root,
1984                                                  parent, ref_root, flags,
1985                                                  ref->objectid, ref->offset,
1986                                                  &ins, node->ref_mod);
1987         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1988                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1989                                              node->num_bytes, parent,
1990                                              ref_root, ref->objectid,
1991                                              ref->offset, node->ref_mod,
1992                                              extent_op);
1993         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1994                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1995                                           node->num_bytes, parent,
1996                                           ref_root, ref->objectid,
1997                                           ref->offset, node->ref_mod,
1998                                           extent_op);
1999         } else {
2000                 BUG();
2001         }
2002         return ret;
2003 }
2004
2005 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2006                                     struct extent_buffer *leaf,
2007                                     struct btrfs_extent_item *ei)
2008 {
2009         u64 flags = btrfs_extent_flags(leaf, ei);
2010         if (extent_op->update_flags) {
2011                 flags |= extent_op->flags_to_set;
2012                 btrfs_set_extent_flags(leaf, ei, flags);
2013         }
2014
2015         if (extent_op->update_key) {
2016                 struct btrfs_tree_block_info *bi;
2017                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2018                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2019                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2020         }
2021 }
2022
2023 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2024                                  struct btrfs_root *root,
2025                                  struct btrfs_delayed_ref_node *node,
2026                                  struct btrfs_delayed_extent_op *extent_op)
2027 {
2028         struct btrfs_key key;
2029         struct btrfs_path *path;
2030         struct btrfs_extent_item *ei;
2031         struct extent_buffer *leaf;
2032         u32 item_size;
2033         int ret;
2034         int err = 0;
2035
2036         if (trans->aborted)
2037                 return 0;
2038
2039         path = btrfs_alloc_path();
2040         if (!path)
2041                 return -ENOMEM;
2042
2043         key.objectid = node->bytenr;
2044         key.type = BTRFS_EXTENT_ITEM_KEY;
2045         key.offset = node->num_bytes;
2046
2047         path->reada = 1;
2048         path->leave_spinning = 1;
2049         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2050                                 path, 0, 1);
2051         if (ret < 0) {
2052                 err = ret;
2053                 goto out;
2054         }
2055         if (ret > 0) {
2056                 err = -EIO;
2057                 goto out;
2058         }
2059
2060         leaf = path->nodes[0];
2061         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2062 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2063         if (item_size < sizeof(*ei)) {
2064                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2065                                              path, (u64)-1, 0);
2066                 if (ret < 0) {
2067                         err = ret;
2068                         goto out;
2069                 }
2070                 leaf = path->nodes[0];
2071                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2072         }
2073 #endif
2074         BUG_ON(item_size < sizeof(*ei));
2075         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2076         __run_delayed_extent_op(extent_op, leaf, ei);
2077
2078         btrfs_mark_buffer_dirty(leaf);
2079 out:
2080         btrfs_free_path(path);
2081         return err;
2082 }
2083
2084 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2085                                 struct btrfs_root *root,
2086                                 struct btrfs_delayed_ref_node *node,
2087                                 struct btrfs_delayed_extent_op *extent_op,
2088                                 int insert_reserved)
2089 {
2090         int ret = 0;
2091         struct btrfs_delayed_tree_ref *ref;
2092         struct btrfs_key ins;
2093         u64 parent = 0;
2094         u64 ref_root = 0;
2095
2096         ins.objectid = node->bytenr;
2097         ins.offset = node->num_bytes;
2098         ins.type = BTRFS_EXTENT_ITEM_KEY;
2099
2100         ref = btrfs_delayed_node_to_tree_ref(node);
2101         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2102                 parent = ref->parent;
2103         else
2104                 ref_root = ref->root;
2105
2106         BUG_ON(node->ref_mod != 1);
2107         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2108                 BUG_ON(!extent_op || !extent_op->update_flags ||
2109                        !extent_op->update_key);
2110                 ret = alloc_reserved_tree_block(trans, root,
2111                                                 parent, ref_root,
2112                                                 extent_op->flags_to_set,
2113                                                 &extent_op->key,
2114                                                 ref->level, &ins);
2115         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2116                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2117                                              node->num_bytes, parent, ref_root,
2118                                              ref->level, 0, 1, extent_op);
2119         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2120                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2121                                           node->num_bytes, parent, ref_root,
2122                                           ref->level, 0, 1, extent_op);
2123         } else {
2124                 BUG();
2125         }
2126         return ret;
2127 }
2128
2129 /* helper function to actually process a single delayed ref entry */
2130 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2131                                struct btrfs_root *root,
2132                                struct btrfs_delayed_ref_node *node,
2133                                struct btrfs_delayed_extent_op *extent_op,
2134                                int insert_reserved)
2135 {
2136         int ret = 0;
2137
2138         if (trans->aborted)
2139                 return 0;
2140
2141         if (btrfs_delayed_ref_is_head(node)) {
2142                 struct btrfs_delayed_ref_head *head;
2143                 /*
2144                  * we've hit the end of the chain and we were supposed
2145                  * to insert this extent into the tree.  But, it got
2146                  * deleted before we ever needed to insert it, so all
2147                  * we have to do is clean up the accounting
2148                  */
2149                 BUG_ON(extent_op);
2150                 head = btrfs_delayed_node_to_head(node);
2151                 if (insert_reserved) {
2152                         btrfs_pin_extent(root, node->bytenr,
2153                                          node->num_bytes, 1);
2154                         if (head->is_data) {
2155                                 ret = btrfs_del_csums(trans, root,
2156                                                       node->bytenr,
2157                                                       node->num_bytes);
2158                         }
2159                 }
2160                 mutex_unlock(&head->mutex);
2161                 return ret;
2162         }
2163
2164         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2165             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2166                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2167                                            insert_reserved);
2168         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2169                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2170                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2171                                            insert_reserved);
2172         else
2173                 BUG();
2174         return ret;
2175 }
2176
2177 static noinline struct btrfs_delayed_ref_node *
2178 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2179 {
2180         struct rb_node *node;
2181         struct btrfs_delayed_ref_node *ref;
2182         int action = BTRFS_ADD_DELAYED_REF;
2183 again:
2184         /*
2185          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2186          * this prevents ref count from going down to zero when
2187          * there still are pending delayed ref.
2188          */
2189         node = rb_prev(&head->node.rb_node);
2190         while (1) {
2191                 if (!node)
2192                         break;
2193                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2194                                 rb_node);
2195                 if (ref->bytenr != head->node.bytenr)
2196                         break;
2197                 if (ref->action == action)
2198                         return ref;
2199                 node = rb_prev(node);
2200         }
2201         if (action == BTRFS_ADD_DELAYED_REF) {
2202                 action = BTRFS_DROP_DELAYED_REF;
2203                 goto again;
2204         }
2205         return NULL;
2206 }
2207
2208 /*
2209  * Returns 0 on success or if called with an already aborted transaction.
2210  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2211  */
2212 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2213                                        struct btrfs_root *root,
2214                                        struct list_head *cluster)
2215 {
2216         struct btrfs_delayed_ref_root *delayed_refs;
2217         struct btrfs_delayed_ref_node *ref;
2218         struct btrfs_delayed_ref_head *locked_ref = NULL;
2219         struct btrfs_delayed_extent_op *extent_op;
2220         int ret;
2221         int count = 0;
2222         int must_insert_reserved = 0;
2223
2224         delayed_refs = &trans->transaction->delayed_refs;
2225         while (1) {
2226                 if (!locked_ref) {
2227                         /* pick a new head ref from the cluster list */
2228                         if (list_empty(cluster))
2229                                 break;
2230
2231                         locked_ref = list_entry(cluster->next,
2232                                      struct btrfs_delayed_ref_head, cluster);
2233
2234                         /* grab the lock that says we are going to process
2235                          * all the refs for this head */
2236                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2237
2238                         /*
2239                          * we may have dropped the spin lock to get the head
2240                          * mutex lock, and that might have given someone else
2241                          * time to free the head.  If that's true, it has been
2242                          * removed from our list and we can move on.
2243                          */
2244                         if (ret == -EAGAIN) {
2245                                 locked_ref = NULL;
2246                                 count++;
2247                                 continue;
2248                         }
2249                 }
2250
2251                 /*
2252                  * locked_ref is the head node, so we have to go one
2253                  * node back for any delayed ref updates
2254                  */
2255                 ref = select_delayed_ref(locked_ref);
2256
2257                 if (ref && ref->seq &&
2258                     btrfs_check_delayed_seq(delayed_refs, ref->seq)) {
2259                         /*
2260                          * there are still refs with lower seq numbers in the
2261                          * process of being added. Don't run this ref yet.
2262                          */
2263                         list_del_init(&locked_ref->cluster);
2264                         mutex_unlock(&locked_ref->mutex);
2265                         locked_ref = NULL;
2266                         delayed_refs->num_heads_ready++;
2267                         spin_unlock(&delayed_refs->lock);
2268                         cond_resched();
2269                         spin_lock(&delayed_refs->lock);
2270                         continue;
2271                 }
2272
2273                 /*
2274                  * record the must insert reserved flag before we
2275                  * drop the spin lock.
2276                  */
2277                 must_insert_reserved = locked_ref->must_insert_reserved;
2278                 locked_ref->must_insert_reserved = 0;
2279
2280                 extent_op = locked_ref->extent_op;
2281                 locked_ref->extent_op = NULL;
2282
2283                 if (!ref) {
2284                         /* All delayed refs have been processed, Go ahead
2285                          * and send the head node to run_one_delayed_ref,
2286                          * so that any accounting fixes can happen
2287                          */
2288                         ref = &locked_ref->node;
2289
2290                         if (extent_op && must_insert_reserved) {
2291                                 kfree(extent_op);
2292                                 extent_op = NULL;
2293                         }
2294
2295                         if (extent_op) {
2296                                 spin_unlock(&delayed_refs->lock);
2297
2298                                 ret = run_delayed_extent_op(trans, root,
2299                                                             ref, extent_op);
2300                                 kfree(extent_op);
2301
2302                                 if (ret) {
2303                                         printk(KERN_DEBUG "btrfs: run_delayed_extent_op returned %d\n", ret);
2304                                         spin_lock(&delayed_refs->lock);
2305                                         return ret;
2306                                 }
2307
2308                                 goto next;
2309                         }
2310
2311                         list_del_init(&locked_ref->cluster);
2312                         locked_ref = NULL;
2313                 }
2314
2315                 ref->in_tree = 0;
2316                 rb_erase(&ref->rb_node, &delayed_refs->root);
2317                 delayed_refs->num_entries--;
2318                 /*
2319                  * we modified num_entries, but as we're currently running
2320                  * delayed refs, skip
2321                  *     wake_up(&delayed_refs->seq_wait);
2322                  * here.
2323                  */
2324                 spin_unlock(&delayed_refs->lock);
2325
2326                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2327                                           must_insert_reserved);
2328
2329                 btrfs_put_delayed_ref(ref);
2330                 kfree(extent_op);
2331                 count++;
2332
2333                 if (ret) {
2334                         printk(KERN_DEBUG "btrfs: run_one_delayed_ref returned %d\n", ret);
2335                         spin_lock(&delayed_refs->lock);
2336                         return ret;
2337                 }
2338
2339 next:
2340                 do_chunk_alloc(trans, root->fs_info->extent_root,
2341                                2 * 1024 * 1024,
2342                                btrfs_get_alloc_profile(root, 0),
2343                                CHUNK_ALLOC_NO_FORCE);
2344                 cond_resched();
2345                 spin_lock(&delayed_refs->lock);
2346         }
2347         return count;
2348 }
2349
2350
2351 static void wait_for_more_refs(struct btrfs_delayed_ref_root *delayed_refs,
2352                         unsigned long num_refs)
2353 {
2354         struct list_head *first_seq = delayed_refs->seq_head.next;
2355
2356         spin_unlock(&delayed_refs->lock);
2357         pr_debug("waiting for more refs (num %ld, first %p)\n",
2358                  num_refs, first_seq);
2359         wait_event(delayed_refs->seq_wait,
2360                    num_refs != delayed_refs->num_entries ||
2361                    delayed_refs->seq_head.next != first_seq);
2362         pr_debug("done waiting for more refs (num %ld, first %p)\n",
2363                  delayed_refs->num_entries, delayed_refs->seq_head.next);
2364         spin_lock(&delayed_refs->lock);
2365 }
2366
2367 /*
2368  * this starts processing the delayed reference count updates and
2369  * extent insertions we have queued up so far.  count can be
2370  * 0, which means to process everything in the tree at the start
2371  * of the run (but not newly added entries), or it can be some target
2372  * number you'd like to process.
2373  *
2374  * Returns 0 on success or if called with an aborted transaction
2375  * Returns <0 on error and aborts the transaction
2376  */
2377 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2378                            struct btrfs_root *root, unsigned long count)
2379 {
2380         struct rb_node *node;
2381         struct btrfs_delayed_ref_root *delayed_refs;
2382         struct btrfs_delayed_ref_node *ref;
2383         struct list_head cluster;
2384         int ret;
2385         u64 delayed_start;
2386         int run_all = count == (unsigned long)-1;
2387         int run_most = 0;
2388         unsigned long num_refs = 0;
2389         int consider_waiting;
2390
2391         /* We'll clean this up in btrfs_cleanup_transaction */
2392         if (trans->aborted)
2393                 return 0;
2394
2395         if (root == root->fs_info->extent_root)
2396                 root = root->fs_info->tree_root;
2397
2398         do_chunk_alloc(trans, root->fs_info->extent_root,
2399                        2 * 1024 * 1024, btrfs_get_alloc_profile(root, 0),
2400                        CHUNK_ALLOC_NO_FORCE);
2401
2402         delayed_refs = &trans->transaction->delayed_refs;
2403         INIT_LIST_HEAD(&cluster);
2404 again:
2405         consider_waiting = 0;
2406         spin_lock(&delayed_refs->lock);
2407         if (count == 0) {
2408                 count = delayed_refs->num_entries * 2;
2409                 run_most = 1;
2410         }
2411         while (1) {
2412                 if (!(run_all || run_most) &&
2413                     delayed_refs->num_heads_ready < 64)
2414                         break;
2415
2416                 /*
2417                  * go find something we can process in the rbtree.  We start at
2418                  * the beginning of the tree, and then build a cluster
2419                  * of refs to process starting at the first one we are able to
2420                  * lock
2421                  */
2422                 delayed_start = delayed_refs->run_delayed_start;
2423                 ret = btrfs_find_ref_cluster(trans, &cluster,
2424                                              delayed_refs->run_delayed_start);
2425                 if (ret)
2426                         break;
2427
2428                 if (delayed_start >= delayed_refs->run_delayed_start) {
2429                         if (consider_waiting == 0) {
2430                                 /*
2431                                  * btrfs_find_ref_cluster looped. let's do one
2432                                  * more cycle. if we don't run any delayed ref
2433                                  * during that cycle (because we can't because
2434                                  * all of them are blocked) and if the number of
2435                                  * refs doesn't change, we avoid busy waiting.
2436                                  */
2437                                 consider_waiting = 1;
2438                                 num_refs = delayed_refs->num_entries;
2439                         } else {
2440                                 wait_for_more_refs(delayed_refs, num_refs);
2441                                 /*
2442                                  * after waiting, things have changed. we
2443                                  * dropped the lock and someone else might have
2444                                  * run some refs, built new clusters and so on.
2445                                  * therefore, we restart staleness detection.
2446                                  */
2447                                 consider_waiting = 0;
2448                         }
2449                 }
2450
2451                 ret = run_clustered_refs(trans, root, &cluster);
2452                 if (ret < 0) {
2453                         spin_unlock(&delayed_refs->lock);
2454                         btrfs_abort_transaction(trans, root, ret);
2455                         return ret;
2456                 }
2457
2458                 count -= min_t(unsigned long, ret, count);
2459
2460                 if (count == 0)
2461                         break;
2462
2463                 if (ret || delayed_refs->run_delayed_start == 0) {
2464                         /* refs were run, let's reset staleness detection */
2465                         consider_waiting = 0;
2466                 }
2467         }
2468
2469         if (run_all) {
2470                 node = rb_first(&delayed_refs->root);
2471                 if (!node)
2472                         goto out;
2473                 count = (unsigned long)-1;
2474
2475                 while (node) {
2476                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2477                                        rb_node);
2478                         if (btrfs_delayed_ref_is_head(ref)) {
2479                                 struct btrfs_delayed_ref_head *head;
2480
2481                                 head = btrfs_delayed_node_to_head(ref);
2482                                 atomic_inc(&ref->refs);
2483
2484                                 spin_unlock(&delayed_refs->lock);
2485                                 /*
2486                                  * Mutex was contended, block until it's
2487                                  * released and try again
2488                                  */
2489                                 mutex_lock(&head->mutex);
2490                                 mutex_unlock(&head->mutex);
2491
2492                                 btrfs_put_delayed_ref(ref);
2493                                 cond_resched();
2494                                 goto again;
2495                         }
2496                         node = rb_next(node);
2497                 }
2498                 spin_unlock(&delayed_refs->lock);
2499                 schedule_timeout(1);
2500                 goto again;
2501         }
2502 out:
2503         spin_unlock(&delayed_refs->lock);
2504         return 0;
2505 }
2506
2507 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2508                                 struct btrfs_root *root,
2509                                 u64 bytenr, u64 num_bytes, u64 flags,
2510                                 int is_data)
2511 {
2512         struct btrfs_delayed_extent_op *extent_op;
2513         int ret;
2514
2515         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2516         if (!extent_op)
2517                 return -ENOMEM;
2518
2519         extent_op->flags_to_set = flags;
2520         extent_op->update_flags = 1;
2521         extent_op->update_key = 0;
2522         extent_op->is_data = is_data ? 1 : 0;
2523
2524         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2525                                           num_bytes, extent_op);
2526         if (ret)
2527                 kfree(extent_op);
2528         return ret;
2529 }
2530
2531 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2532                                       struct btrfs_root *root,
2533                                       struct btrfs_path *path,
2534                                       u64 objectid, u64 offset, u64 bytenr)
2535 {
2536         struct btrfs_delayed_ref_head *head;
2537         struct btrfs_delayed_ref_node *ref;
2538         struct btrfs_delayed_data_ref *data_ref;
2539         struct btrfs_delayed_ref_root *delayed_refs;
2540         struct rb_node *node;
2541         int ret = 0;
2542
2543         ret = -ENOENT;
2544         delayed_refs = &trans->transaction->delayed_refs;
2545         spin_lock(&delayed_refs->lock);
2546         head = btrfs_find_delayed_ref_head(trans, bytenr);
2547         if (!head)
2548                 goto out;
2549
2550         if (!mutex_trylock(&head->mutex)) {
2551                 atomic_inc(&head->node.refs);
2552                 spin_unlock(&delayed_refs->lock);
2553
2554                 btrfs_release_path(path);
2555
2556                 /*
2557                  * Mutex was contended, block until it's released and let
2558                  * caller try again
2559                  */
2560                 mutex_lock(&head->mutex);
2561                 mutex_unlock(&head->mutex);
2562                 btrfs_put_delayed_ref(&head->node);
2563                 return -EAGAIN;
2564         }
2565
2566         node = rb_prev(&head->node.rb_node);
2567         if (!node)
2568                 goto out_unlock;
2569
2570         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2571
2572         if (ref->bytenr != bytenr)
2573                 goto out_unlock;
2574
2575         ret = 1;
2576         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2577                 goto out_unlock;
2578
2579         data_ref = btrfs_delayed_node_to_data_ref(ref);
2580
2581         node = rb_prev(node);
2582         if (node) {
2583                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2584                 if (ref->bytenr == bytenr)
2585                         goto out_unlock;
2586         }
2587
2588         if (data_ref->root != root->root_key.objectid ||
2589             data_ref->objectid != objectid || data_ref->offset != offset)
2590                 goto out_unlock;
2591
2592         ret = 0;
2593 out_unlock:
2594         mutex_unlock(&head->mutex);
2595 out:
2596         spin_unlock(&delayed_refs->lock);
2597         return ret;
2598 }
2599
2600 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2601                                         struct btrfs_root *root,
2602                                         struct btrfs_path *path,
2603                                         u64 objectid, u64 offset, u64 bytenr)
2604 {
2605         struct btrfs_root *extent_root = root->fs_info->extent_root;
2606         struct extent_buffer *leaf;
2607         struct btrfs_extent_data_ref *ref;
2608         struct btrfs_extent_inline_ref *iref;
2609         struct btrfs_extent_item *ei;
2610         struct btrfs_key key;
2611         u32 item_size;
2612         int ret;
2613
2614         key.objectid = bytenr;
2615         key.offset = (u64)-1;
2616         key.type = BTRFS_EXTENT_ITEM_KEY;
2617
2618         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2619         if (ret < 0)
2620                 goto out;
2621         BUG_ON(ret == 0); /* Corruption */
2622
2623         ret = -ENOENT;
2624         if (path->slots[0] == 0)
2625                 goto out;
2626
2627         path->slots[0]--;
2628         leaf = path->nodes[0];
2629         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2630
2631         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2632                 goto out;
2633
2634         ret = 1;
2635         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2636 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2637         if (item_size < sizeof(*ei)) {
2638                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2639                 goto out;
2640         }
2641 #endif
2642         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2643
2644         if (item_size != sizeof(*ei) +
2645             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2646                 goto out;
2647
2648         if (btrfs_extent_generation(leaf, ei) <=
2649             btrfs_root_last_snapshot(&root->root_item))
2650                 goto out;
2651
2652         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2653         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2654             BTRFS_EXTENT_DATA_REF_KEY)
2655                 goto out;
2656
2657         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2658         if (btrfs_extent_refs(leaf, ei) !=
2659             btrfs_extent_data_ref_count(leaf, ref) ||
2660             btrfs_extent_data_ref_root(leaf, ref) !=
2661             root->root_key.objectid ||
2662             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2663             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2664                 goto out;
2665
2666         ret = 0;
2667 out:
2668         return ret;
2669 }
2670
2671 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2672                           struct btrfs_root *root,
2673                           u64 objectid, u64 offset, u64 bytenr)
2674 {
2675         struct btrfs_path *path;
2676         int ret;
2677         int ret2;
2678
2679         path = btrfs_alloc_path();
2680         if (!path)
2681                 return -ENOENT;
2682
2683         do {
2684                 ret = check_committed_ref(trans, root, path, objectid,
2685                                           offset, bytenr);
2686                 if (ret && ret != -ENOENT)
2687                         goto out;
2688
2689                 ret2 = check_delayed_ref(trans, root, path, objectid,
2690                                          offset, bytenr);
2691         } while (ret2 == -EAGAIN);
2692
2693         if (ret2 && ret2 != -ENOENT) {
2694                 ret = ret2;
2695                 goto out;
2696         }
2697
2698         if (ret != -ENOENT || ret2 != -ENOENT)
2699                 ret = 0;
2700 out:
2701         btrfs_free_path(path);
2702         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2703                 WARN_ON(ret > 0);
2704         return ret;
2705 }
2706
2707 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2708                            struct btrfs_root *root,
2709                            struct extent_buffer *buf,
2710                            int full_backref, int inc, int for_cow)
2711 {
2712         u64 bytenr;
2713         u64 num_bytes;
2714         u64 parent;
2715         u64 ref_root;
2716         u32 nritems;
2717         struct btrfs_key key;
2718         struct btrfs_file_extent_item *fi;
2719         int i;
2720         int level;
2721         int ret = 0;
2722         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2723                             u64, u64, u64, u64, u64, u64, int);
2724
2725         ref_root = btrfs_header_owner(buf);
2726         nritems = btrfs_header_nritems(buf);
2727         level = btrfs_header_level(buf);
2728
2729         if (!root->ref_cows && level == 0)
2730                 return 0;
2731
2732         if (inc)
2733                 process_func = btrfs_inc_extent_ref;
2734         else
2735                 process_func = btrfs_free_extent;
2736
2737         if (full_backref)
2738                 parent = buf->start;
2739         else
2740                 parent = 0;
2741
2742         for (i = 0; i < nritems; i++) {
2743                 if (level == 0) {
2744                         btrfs_item_key_to_cpu(buf, &key, i);
2745                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2746                                 continue;
2747                         fi = btrfs_item_ptr(buf, i,
2748                                             struct btrfs_file_extent_item);
2749                         if (btrfs_file_extent_type(buf, fi) ==
2750                             BTRFS_FILE_EXTENT_INLINE)
2751                                 continue;
2752                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2753                         if (bytenr == 0)
2754                                 continue;
2755
2756                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2757                         key.offset -= btrfs_file_extent_offset(buf, fi);
2758                         ret = process_func(trans, root, bytenr, num_bytes,
2759                                            parent, ref_root, key.objectid,
2760                                            key.offset, for_cow);
2761                         if (ret)
2762                                 goto fail;
2763                 } else {
2764                         bytenr = btrfs_node_blockptr(buf, i);
2765                         num_bytes = btrfs_level_size(root, level - 1);
2766                         ret = process_func(trans, root, bytenr, num_bytes,
2767                                            parent, ref_root, level - 1, 0,
2768                                            for_cow);
2769                         if (ret)
2770                                 goto fail;
2771                 }
2772         }
2773         return 0;
2774 fail:
2775         return ret;
2776 }
2777
2778 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2779                   struct extent_buffer *buf, int full_backref, int for_cow)
2780 {
2781         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2782 }
2783
2784 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2785                   struct extent_buffer *buf, int full_backref, int for_cow)
2786 {
2787         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2788 }
2789
2790 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2791                                  struct btrfs_root *root,
2792                                  struct btrfs_path *path,
2793                                  struct btrfs_block_group_cache *cache)
2794 {
2795         int ret;
2796         struct btrfs_root *extent_root = root->fs_info->extent_root;
2797         unsigned long bi;
2798         struct extent_buffer *leaf;
2799
2800         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2801         if (ret < 0)
2802                 goto fail;
2803         BUG_ON(ret); /* Corruption */
2804
2805         leaf = path->nodes[0];
2806         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2807         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2808         btrfs_mark_buffer_dirty(leaf);
2809         btrfs_release_path(path);
2810 fail:
2811         if (ret) {
2812                 btrfs_abort_transaction(trans, root, ret);
2813                 return ret;
2814         }
2815         return 0;
2816
2817 }
2818
2819 static struct btrfs_block_group_cache *
2820 next_block_group(struct btrfs_root *root,
2821                  struct btrfs_block_group_cache *cache)
2822 {
2823         struct rb_node *node;
2824         spin_lock(&root->fs_info->block_group_cache_lock);
2825         node = rb_next(&cache->cache_node);
2826         btrfs_put_block_group(cache);
2827         if (node) {
2828                 cache = rb_entry(node, struct btrfs_block_group_cache,
2829                                  cache_node);
2830                 btrfs_get_block_group(cache);
2831         } else
2832                 cache = NULL;
2833         spin_unlock(&root->fs_info->block_group_cache_lock);
2834         return cache;
2835 }
2836
2837 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2838                             struct btrfs_trans_handle *trans,
2839                             struct btrfs_path *path)
2840 {
2841         struct btrfs_root *root = block_group->fs_info->tree_root;
2842         struct inode *inode = NULL;
2843         u64 alloc_hint = 0;
2844         int dcs = BTRFS_DC_ERROR;
2845         int num_pages = 0;
2846         int retries = 0;
2847         int ret = 0;
2848
2849         /*
2850          * If this block group is smaller than 100 megs don't bother caching the
2851          * block group.
2852          */
2853         if (block_group->key.offset < (100 * 1024 * 1024)) {
2854                 spin_lock(&block_group->lock);
2855                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2856                 spin_unlock(&block_group->lock);
2857                 return 0;
2858         }
2859
2860 again:
2861         inode = lookup_free_space_inode(root, block_group, path);
2862         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2863                 ret = PTR_ERR(inode);
2864                 btrfs_release_path(path);
2865                 goto out;
2866         }
2867
2868         if (IS_ERR(inode)) {
2869                 BUG_ON(retries);
2870                 retries++;
2871
2872                 if (block_group->ro)
2873                         goto out_free;
2874
2875                 ret = create_free_space_inode(root, trans, block_group, path);
2876                 if (ret)
2877                         goto out_free;
2878                 goto again;
2879         }
2880
2881         /* We've already setup this transaction, go ahead and exit */
2882         if (block_group->cache_generation == trans->transid &&
2883             i_size_read(inode)) {
2884                 dcs = BTRFS_DC_SETUP;
2885                 goto out_put;
2886         }
2887
2888         /*
2889          * We want to set the generation to 0, that way if anything goes wrong
2890          * from here on out we know not to trust this cache when we load up next
2891          * time.
2892          */
2893         BTRFS_I(inode)->generation = 0;
2894         ret = btrfs_update_inode(trans, root, inode);
2895         WARN_ON(ret);
2896
2897         if (i_size_read(inode) > 0) {
2898                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2899                                                       inode);
2900                 if (ret)
2901                         goto out_put;
2902         }
2903
2904         spin_lock(&block_group->lock);
2905         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2906                 /* We're not cached, don't bother trying to write stuff out */
2907                 dcs = BTRFS_DC_WRITTEN;
2908                 spin_unlock(&block_group->lock);
2909                 goto out_put;
2910         }
2911         spin_unlock(&block_group->lock);
2912
2913         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2914         if (!num_pages)
2915                 num_pages = 1;
2916
2917         /*
2918          * Just to make absolutely sure we have enough space, we're going to
2919          * preallocate 12 pages worth of space for each block group.  In
2920          * practice we ought to use at most 8, but we need extra space so we can
2921          * add our header and have a terminator between the extents and the
2922          * bitmaps.
2923          */
2924         num_pages *= 16;
2925         num_pages *= PAGE_CACHE_SIZE;
2926
2927         ret = btrfs_check_data_free_space(inode, num_pages);
2928         if (ret)
2929                 goto out_put;
2930
2931         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2932                                               num_pages, num_pages,
2933                                               &alloc_hint);
2934         if (!ret)
2935                 dcs = BTRFS_DC_SETUP;
2936         btrfs_free_reserved_data_space(inode, num_pages);
2937
2938 out_put:
2939         iput(inode);
2940 out_free:
2941         btrfs_release_path(path);
2942 out:
2943         spin_lock(&block_group->lock);
2944         if (!ret && dcs == BTRFS_DC_SETUP)
2945                 block_group->cache_generation = trans->transid;
2946         block_group->disk_cache_state = dcs;
2947         spin_unlock(&block_group->lock);
2948
2949         return ret;
2950 }
2951
2952 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2953                                    struct btrfs_root *root)
2954 {
2955         struct btrfs_block_group_cache *cache;
2956         int err = 0;
2957         struct btrfs_path *path;
2958         u64 last = 0;
2959
2960         path = btrfs_alloc_path();
2961         if (!path)
2962                 return -ENOMEM;
2963
2964 again:
2965         while (1) {
2966                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2967                 while (cache) {
2968                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2969                                 break;
2970                         cache = next_block_group(root, cache);
2971                 }
2972                 if (!cache) {
2973                         if (last == 0)
2974                                 break;
2975                         last = 0;
2976                         continue;
2977                 }
2978                 err = cache_save_setup(cache, trans, path);
2979                 last = cache->key.objectid + cache->key.offset;
2980                 btrfs_put_block_group(cache);
2981         }
2982
2983         while (1) {
2984                 if (last == 0) {
2985                         err = btrfs_run_delayed_refs(trans, root,
2986                                                      (unsigned long)-1);
2987                         if (err) /* File system offline */
2988                                 goto out;
2989                 }
2990
2991                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2992                 while (cache) {
2993                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2994                                 btrfs_put_block_group(cache);
2995                                 goto again;
2996                         }
2997
2998                         if (cache->dirty)
2999                                 break;
3000                         cache = next_block_group(root, cache);
3001                 }
3002                 if (!cache) {
3003                         if (last == 0)
3004                                 break;
3005                         last = 0;
3006                         continue;
3007                 }
3008
3009                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3010                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3011                 cache->dirty = 0;
3012                 last = cache->key.objectid + cache->key.offset;
3013
3014                 err = write_one_cache_group(trans, root, path, cache);
3015                 if (err) /* File system offline */
3016                         goto out;
3017
3018                 btrfs_put_block_group(cache);
3019         }
3020
3021         while (1) {
3022                 /*
3023                  * I don't think this is needed since we're just marking our
3024                  * preallocated extent as written, but just in case it can't
3025                  * hurt.
3026                  */
3027                 if (last == 0) {
3028                         err = btrfs_run_delayed_refs(trans, root,
3029                                                      (unsigned long)-1);
3030                         if (err) /* File system offline */
3031                                 goto out;
3032                 }
3033
3034                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3035                 while (cache) {
3036                         /*
3037                          * Really this shouldn't happen, but it could if we
3038                          * couldn't write the entire preallocated extent and
3039                          * splitting the extent resulted in a new block.
3040                          */
3041                         if (cache->dirty) {
3042                                 btrfs_put_block_group(cache);
3043                                 goto again;
3044                         }
3045                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3046                                 break;
3047                         cache = next_block_group(root, cache);
3048                 }
3049                 if (!cache) {
3050                         if (last == 0)
3051                                 break;
3052                         last = 0;
3053                         continue;
3054                 }
3055
3056                 err = btrfs_write_out_cache(root, trans, cache, path);
3057
3058                 /*
3059                  * If we didn't have an error then the cache state is still
3060                  * NEED_WRITE, so we can set it to WRITTEN.
3061                  */
3062                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3063                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3064                 last = cache->key.objectid + cache->key.offset;
3065                 btrfs_put_block_group(cache);
3066         }
3067 out:
3068
3069         btrfs_free_path(path);
3070         return err;
3071 }
3072
3073 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3074 {
3075         struct btrfs_block_group_cache *block_group;
3076         int readonly = 0;
3077
3078         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3079         if (!block_group || block_group->ro)
3080                 readonly = 1;
3081         if (block_group)
3082                 btrfs_put_block_group(block_group);
3083         return readonly;
3084 }
3085
3086 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3087                              u64 total_bytes, u64 bytes_used,
3088                              struct btrfs_space_info **space_info)
3089 {
3090         struct btrfs_space_info *found;
3091         int i;
3092         int factor;
3093
3094         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3095                      BTRFS_BLOCK_GROUP_RAID10))
3096                 factor = 2;
3097         else
3098                 factor = 1;
3099
3100         found = __find_space_info(info, flags);
3101         if (found) {
3102                 spin_lock(&found->lock);
3103                 found->total_bytes += total_bytes;
3104                 found->disk_total += total_bytes * factor;
3105                 found->bytes_used += bytes_used;
3106                 found->disk_used += bytes_used * factor;
3107                 found->full = 0;
3108                 spin_unlock(&found->lock);
3109                 *space_info = found;
3110                 return 0;
3111         }
3112         found = kzalloc(sizeof(*found), GFP_NOFS);
3113         if (!found)
3114                 return -ENOMEM;
3115
3116         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3117                 INIT_LIST_HEAD(&found->block_groups[i]);
3118         init_rwsem(&found->groups_sem);
3119         spin_lock_init(&found->lock);
3120         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3121         found->total_bytes = total_bytes;
3122         found->disk_total = total_bytes * factor;
3123         found->bytes_used = bytes_used;
3124         found->disk_used = bytes_used * factor;
3125         found->bytes_pinned = 0;
3126         found->bytes_reserved = 0;
3127         found->bytes_readonly = 0;
3128         found->bytes_may_use = 0;
3129         found->full = 0;
3130         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3131         found->chunk_alloc = 0;
3132         found->flush = 0;
3133         init_waitqueue_head(&found->wait);
3134         *space_info = found;
3135         list_add_rcu(&found->list, &info->space_info);
3136         return 0;
3137 }
3138
3139 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3140 {
3141         u64 extra_flags = chunk_to_extended(flags) &
3142                                 BTRFS_EXTENDED_PROFILE_MASK;
3143
3144         if (flags & BTRFS_BLOCK_GROUP_DATA)
3145                 fs_info->avail_data_alloc_bits |= extra_flags;
3146         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3147                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3148         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3149                 fs_info->avail_system_alloc_bits |= extra_flags;
3150 }
3151
3152 /*
3153  * returns target flags in extended format or 0 if restripe for this
3154  * chunk_type is not in progress
3155  *
3156  * should be called with either volume_mutex or balance_lock held
3157  */
3158 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3159 {
3160         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3161         u64 target = 0;
3162
3163         if (!bctl)
3164                 return 0;
3165
3166         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3167             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3168                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3169         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3170                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3171                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3172         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3173                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3174                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3175         }
3176
3177         return target;
3178 }
3179
3180 /*
3181  * @flags: available profiles in extended format (see ctree.h)
3182  *
3183  * Returns reduced profile in chunk format.  If profile changing is in
3184  * progress (either running or paused) picks the target profile (if it's
3185  * already available), otherwise falls back to plain reducing.
3186  */
3187 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3188 {
3189         /*
3190          * we add in the count of missing devices because we want
3191          * to make sure that any RAID levels on a degraded FS
3192          * continue to be honored.
3193          */
3194         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3195                 root->fs_info->fs_devices->missing_devices;
3196         u64 target;
3197
3198         /*
3199          * see if restripe for this chunk_type is in progress, if so
3200          * try to reduce to the target profile
3201          */
3202         spin_lock(&root->fs_info->balance_lock);
3203         target = get_restripe_target(root->fs_info, flags);
3204         if (target) {
3205                 /* pick target profile only if it's already available */
3206                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3207                         spin_unlock(&root->fs_info->balance_lock);
3208                         return extended_to_chunk(target);
3209                 }
3210         }
3211         spin_unlock(&root->fs_info->balance_lock);
3212
3213         if (num_devices == 1)
3214                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3215         if (num_devices < 4)
3216                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3217
3218         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3219             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3220                       BTRFS_BLOCK_GROUP_RAID10))) {
3221                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3222         }
3223
3224         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3225             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3226                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3227         }
3228
3229         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3230             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3231              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3232              (flags & BTRFS_BLOCK_GROUP_DUP))) {
3233                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3234         }
3235
3236         return extended_to_chunk(flags);
3237 }
3238
3239 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3240 {
3241         if (flags & BTRFS_BLOCK_GROUP_DATA)
3242                 flags |= root->fs_info->avail_data_alloc_bits;
3243         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3244                 flags |= root->fs_info->avail_system_alloc_bits;
3245         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3246                 flags |= root->fs_info->avail_metadata_alloc_bits;
3247
3248         return btrfs_reduce_alloc_profile(root, flags);
3249 }
3250
3251 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3252 {
3253         u64 flags;
3254
3255         if (data)
3256                 flags = BTRFS_BLOCK_GROUP_DATA;
3257         else if (root == root->fs_info->chunk_root)
3258                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3259         else
3260                 flags = BTRFS_BLOCK_GROUP_METADATA;
3261
3262         return get_alloc_profile(root, flags);
3263 }
3264
3265 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3266 {
3267         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3268                                                        BTRFS_BLOCK_GROUP_DATA);
3269 }
3270
3271 /*
3272  * This will check the space that the inode allocates from to make sure we have
3273  * enough space for bytes.
3274  */
3275 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3276 {
3277         struct btrfs_space_info *data_sinfo;
3278         struct btrfs_root *root = BTRFS_I(inode)->root;
3279         u64 used;
3280         int ret = 0, committed = 0, alloc_chunk = 1;
3281
3282         /* make sure bytes are sectorsize aligned */
3283         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3284
3285         if (root == root->fs_info->tree_root ||
3286             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3287                 alloc_chunk = 0;
3288                 committed = 1;
3289         }
3290
3291         data_sinfo = BTRFS_I(inode)->space_info;
3292         if (!data_sinfo)
3293                 goto alloc;
3294
3295 again:
3296         /* make sure we have enough space to handle the data first */
3297         spin_lock(&data_sinfo->lock);
3298         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3299                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3300                 data_sinfo->bytes_may_use;
3301
3302         if (used + bytes > data_sinfo->total_bytes) {
3303                 struct btrfs_trans_handle *trans;
3304
3305                 /*
3306                  * if we don't have enough free bytes in this space then we need
3307                  * to alloc a new chunk.
3308                  */
3309                 if (!data_sinfo->full && alloc_chunk) {
3310                         u64 alloc_target;
3311
3312                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3313                         spin_unlock(&data_sinfo->lock);
3314 alloc:
3315                         alloc_target = btrfs_get_alloc_profile(root, 1);
3316                         trans = btrfs_join_transaction(root);
3317                         if (IS_ERR(trans))
3318                                 return PTR_ERR(trans);
3319
3320                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3321                                              bytes + 2 * 1024 * 1024,
3322                                              alloc_target,
3323                                              CHUNK_ALLOC_NO_FORCE);
3324                         btrfs_end_transaction(trans, root);
3325                         if (ret < 0) {
3326                                 if (ret != -ENOSPC)
3327                                         return ret;
3328                                 else
3329                                         goto commit_trans;
3330                         }
3331
3332                         if (!data_sinfo) {
3333                                 btrfs_set_inode_space_info(root, inode);
3334                                 data_sinfo = BTRFS_I(inode)->space_info;
3335                         }
3336                         goto again;
3337                 }
3338
3339                 /*
3340                  * If we have less pinned bytes than we want to allocate then
3341                  * don't bother committing the transaction, it won't help us.
3342                  */
3343                 if (data_sinfo->bytes_pinned < bytes)
3344                         committed = 1;
3345                 spin_unlock(&data_sinfo->lock);
3346
3347                 /* commit the current transaction and try again */
3348 commit_trans:
3349                 if (!committed &&
3350                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3351                         committed = 1;
3352                         trans = btrfs_join_transaction(root);
3353                         if (IS_ERR(trans))
3354                                 return PTR_ERR(trans);
3355                         ret = btrfs_commit_transaction(trans, root);
3356                         if (ret)
3357                                 return ret;
3358                         goto again;
3359                 }
3360
3361                 return -ENOSPC;
3362         }
3363         data_sinfo->bytes_may_use += bytes;
3364         trace_btrfs_space_reservation(root->fs_info, "space_info",
3365                                       data_sinfo->flags, bytes, 1);
3366         spin_unlock(&data_sinfo->lock);
3367
3368         return 0;
3369 }
3370
3371 /*
3372  * Called if we need to clear a data reservation for this inode.
3373  */
3374 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3375 {
3376         struct btrfs_root *root = BTRFS_I(inode)->root;
3377         struct btrfs_space_info *data_sinfo;
3378
3379         /* make sure bytes are sectorsize aligned */
3380         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3381
3382         data_sinfo = BTRFS_I(inode)->space_info;
3383         spin_lock(&data_sinfo->lock);
3384         data_sinfo->bytes_may_use -= bytes;
3385         trace_btrfs_space_reservation(root->fs_info, "space_info",
3386                                       data_sinfo->flags, bytes, 0);
3387         spin_unlock(&data_sinfo->lock);
3388 }
3389
3390 static void force_metadata_allocation(struct btrfs_fs_info *info)
3391 {
3392         struct list_head *head = &info->space_info;
3393         struct btrfs_space_info *found;
3394
3395         rcu_read_lock();
3396         list_for_each_entry_rcu(found, head, list) {
3397                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3398                         found->force_alloc = CHUNK_ALLOC_FORCE;
3399         }
3400         rcu_read_unlock();
3401 }
3402
3403 static int should_alloc_chunk(struct btrfs_root *root,
3404                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3405                               int force)
3406 {
3407         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3408         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3409         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3410         u64 thresh;
3411
3412         if (force == CHUNK_ALLOC_FORCE)
3413                 return 1;
3414
3415         /*
3416          * We need to take into account the global rsv because for all intents
3417          * and purposes it's used space.  Don't worry about locking the
3418          * global_rsv, it doesn't change except when the transaction commits.
3419          */
3420         num_allocated += global_rsv->size;
3421
3422         /*
3423          * in limited mode, we want to have some free space up to
3424          * about 1% of the FS size.
3425          */
3426         if (force == CHUNK_ALLOC_LIMITED) {
3427                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3428                 thresh = max_t(u64, 64 * 1024 * 1024,
3429                                div_factor_fine(thresh, 1));
3430
3431                 if (num_bytes - num_allocated < thresh)
3432                         return 1;
3433         }
3434         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3435
3436         /* 256MB or 2% of the FS */
3437         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 2));
3438         /* system chunks need a much small threshold */
3439         if (sinfo->flags & BTRFS_BLOCK_GROUP_SYSTEM)
3440                 thresh = 32 * 1024 * 1024;
3441
3442         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 8))
3443                 return 0;
3444         return 1;
3445 }
3446
3447 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3448 {
3449         u64 num_dev;
3450
3451         if (type & BTRFS_BLOCK_GROUP_RAID10 ||
3452             type & BTRFS_BLOCK_GROUP_RAID0)
3453                 num_dev = root->fs_info->fs_devices->rw_devices;
3454         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3455                 num_dev = 2;
3456         else
3457                 num_dev = 1;    /* DUP or single */
3458
3459         /* metadata for updaing devices and chunk tree */
3460         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3461 }
3462
3463 static void check_system_chunk(struct btrfs_trans_handle *trans,
3464                                struct btrfs_root *root, u64 type)
3465 {
3466         struct btrfs_space_info *info;
3467         u64 left;
3468         u64 thresh;
3469
3470         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3471         spin_lock(&info->lock);
3472         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3473                 info->bytes_reserved - info->bytes_readonly;
3474         spin_unlock(&info->lock);
3475
3476         thresh = get_system_chunk_thresh(root, type);
3477         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3478                 printk(KERN_INFO "left=%llu, need=%llu, flags=%llu\n",
3479                        left, thresh, type);
3480                 dump_space_info(info, 0, 0);
3481         }
3482
3483         if (left < thresh) {
3484                 u64 flags;
3485
3486                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3487                 btrfs_alloc_chunk(trans, root, flags);
3488         }
3489 }
3490
3491 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3492                           struct btrfs_root *extent_root, u64 alloc_bytes,
3493                           u64 flags, int force)
3494 {
3495         struct btrfs_space_info *space_info;
3496         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3497         int wait_for_alloc = 0;
3498         int ret = 0;
3499
3500         space_info = __find_space_info(extent_root->fs_info, flags);
3501         if (!space_info) {
3502                 ret = update_space_info(extent_root->fs_info, flags,
3503                                         0, 0, &space_info);
3504                 BUG_ON(ret); /* -ENOMEM */
3505         }
3506         BUG_ON(!space_info); /* Logic error */
3507
3508 again:
3509         spin_lock(&space_info->lock);
3510         if (force < space_info->force_alloc)
3511                 force = space_info->force_alloc;
3512         if (space_info->full) {
3513                 spin_unlock(&space_info->lock);
3514                 return 0;
3515         }
3516
3517         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3518                 spin_unlock(&space_info->lock);
3519                 return 0;
3520         } else if (space_info->chunk_alloc) {
3521                 wait_for_alloc = 1;
3522         } else {
3523                 space_info->chunk_alloc = 1;
3524         }
3525
3526         spin_unlock(&space_info->lock);
3527
3528         mutex_lock(&fs_info->chunk_mutex);
3529
3530         /*
3531          * The chunk_mutex is held throughout the entirety of a chunk
3532          * allocation, so once we've acquired the chunk_mutex we know that the
3533          * other guy is done and we need to recheck and see if we should
3534          * allocate.
3535          */
3536         if (wait_for_alloc) {
3537                 mutex_unlock(&fs_info->chunk_mutex);
3538                 wait_for_alloc = 0;
3539                 goto again;
3540         }
3541
3542         /*
3543          * If we have mixed data/metadata chunks we want to make sure we keep
3544          * allocating mixed chunks instead of individual chunks.
3545          */
3546         if (btrfs_mixed_space_info(space_info))
3547                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3548
3549         /*
3550          * if we're doing a data chunk, go ahead and make sure that
3551          * we keep a reasonable number of metadata chunks allocated in the
3552          * FS as well.
3553          */
3554         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3555                 fs_info->data_chunk_allocations++;
3556                 if (!(fs_info->data_chunk_allocations %
3557                       fs_info->metadata_ratio))
3558                         force_metadata_allocation(fs_info);
3559         }
3560
3561         /*
3562          * Check if we have enough space in SYSTEM chunk because we may need
3563          * to update devices.
3564          */
3565         check_system_chunk(trans, extent_root, flags);
3566
3567         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3568         if (ret < 0 && ret != -ENOSPC)
3569                 goto out;
3570
3571         spin_lock(&space_info->lock);
3572         if (ret)
3573                 space_info->full = 1;
3574         else
3575                 ret = 1;
3576
3577         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3578         space_info->chunk_alloc = 0;
3579         spin_unlock(&space_info->lock);
3580 out:
3581         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3582         return ret;
3583 }
3584
3585 /*
3586  * shrink metadata reservation for delalloc
3587  */
3588 static int shrink_delalloc(struct btrfs_root *root, u64 to_reclaim,
3589                            bool wait_ordered)
3590 {
3591         struct btrfs_block_rsv *block_rsv;
3592         struct btrfs_space_info *space_info;
3593         struct btrfs_trans_handle *trans;
3594         u64 reserved;
3595         u64 max_reclaim;
3596         u64 reclaimed = 0;
3597         long time_left;
3598         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3599         int loops = 0;
3600         unsigned long progress;
3601
3602         trans = (struct btrfs_trans_handle *)current->journal_info;
3603         block_rsv = &root->fs_info->delalloc_block_rsv;
3604         space_info = block_rsv->space_info;
3605
3606         smp_mb();
3607         reserved = space_info->bytes_may_use;
3608         progress = space_info->reservation_progress;
3609
3610         if (reserved == 0)
3611                 return 0;
3612
3613         smp_mb();
3614         if (root->fs_info->delalloc_bytes == 0) {
3615                 if (trans)
3616                         return 0;
3617                 btrfs_wait_ordered_extents(root, 0, 0);
3618                 return 0;
3619         }
3620
3621         max_reclaim = min(reserved, to_reclaim);
3622         nr_pages = max_t(unsigned long, nr_pages,
3623                          max_reclaim >> PAGE_CACHE_SHIFT);
3624         while (loops < 1024) {
3625                 /* have the flusher threads jump in and do some IO */
3626                 smp_mb();
3627                 nr_pages = min_t(unsigned long, nr_pages,
3628                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3629                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3630                                                 WB_REASON_FS_FREE_SPACE);
3631
3632                 spin_lock(&space_info->lock);
3633                 if (reserved > space_info->bytes_may_use)
3634                         reclaimed += reserved - space_info->bytes_may_use;
3635                 reserved = space_info->bytes_may_use;
3636                 spin_unlock(&space_info->lock);
3637
3638                 loops++;
3639
3640                 if (reserved == 0 || reclaimed >= max_reclaim)
3641                         break;
3642
3643                 if (trans && trans->transaction->blocked)
3644                         return -EAGAIN;
3645
3646                 if (wait_ordered && !trans) {
3647                         btrfs_wait_ordered_extents(root, 0, 0);
3648                 } else {
3649                         time_left = schedule_timeout_interruptible(1);
3650
3651                         /* We were interrupted, exit */
3652                         if (time_left)
3653                                 break;
3654                 }
3655
3656                 /* we've kicked the IO a few times, if anything has been freed,
3657                  * exit.  There is no sense in looping here for a long time
3658                  * when we really need to commit the transaction, or there are
3659                  * just too many writers without enough free space
3660                  */
3661
3662                 if (loops > 3) {
3663                         smp_mb();
3664                         if (progress != space_info->reservation_progress)
3665                                 break;
3666                 }
3667
3668         }
3669
3670         return reclaimed >= to_reclaim;
3671 }
3672
3673 /**
3674  * maybe_commit_transaction - possibly commit the transaction if its ok to
3675  * @root - the root we're allocating for
3676  * @bytes - the number of bytes we want to reserve
3677  * @force - force the commit
3678  *
3679  * This will check to make sure that committing the transaction will actually
3680  * get us somewhere and then commit the transaction if it does.  Otherwise it
3681  * will return -ENOSPC.
3682  */
3683 static int may_commit_transaction(struct btrfs_root *root,
3684                                   struct btrfs_space_info *space_info,
3685                                   u64 bytes, int force)
3686 {
3687         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3688         struct btrfs_trans_handle *trans;
3689
3690         trans = (struct btrfs_trans_handle *)current->journal_info;
3691         if (trans)
3692                 return -EAGAIN;
3693
3694         if (force)
3695                 goto commit;
3696
3697         /* See if there is enough pinned space to make this reservation */
3698         spin_lock(&space_info->lock);
3699         if (space_info->bytes_pinned >= bytes) {
3700                 spin_unlock(&space_info->lock);
3701                 goto commit;
3702         }
3703         spin_unlock(&space_info->lock);
3704
3705         /*
3706          * See if there is some space in the delayed insertion reservation for
3707          * this reservation.
3708          */
3709         if (space_info != delayed_rsv->space_info)
3710                 return -ENOSPC;
3711
3712         spin_lock(&space_info->lock);
3713         spin_lock(&delayed_rsv->lock);
3714         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3715                 spin_unlock(&delayed_rsv->lock);
3716                 spin_unlock(&space_info->lock);
3717                 return -ENOSPC;
3718         }
3719         spin_unlock(&delayed_rsv->lock);
3720         spin_unlock(&space_info->lock);
3721
3722 commit:
3723         trans = btrfs_join_transaction(root);
3724         if (IS_ERR(trans))
3725                 return -ENOSPC;
3726
3727         return btrfs_commit_transaction(trans, root);
3728 }
3729
3730 /**
3731  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3732  * @root - the root we're allocating for
3733  * @block_rsv - the block_rsv we're allocating for
3734  * @orig_bytes - the number of bytes we want
3735  * @flush - wether or not we can flush to make our reservation
3736  *
3737  * This will reserve orgi_bytes number of bytes from the space info associated
3738  * with the block_rsv.  If there is not enough space it will make an attempt to
3739  * flush out space to make room.  It will do this by flushing delalloc if
3740  * possible or committing the transaction.  If flush is 0 then no attempts to
3741  * regain reservations will be made and this will fail if there is not enough
3742  * space already.
3743  */
3744 static int reserve_metadata_bytes(struct btrfs_root *root,
3745                                   struct btrfs_block_rsv *block_rsv,
3746                                   u64 orig_bytes, int flush)
3747 {
3748         struct btrfs_space_info *space_info = block_rsv->space_info;
3749         u64 used;
3750         u64 num_bytes = orig_bytes;
3751         int retries = 0;
3752         int ret = 0;
3753         bool committed = false;
3754         bool flushing = false;
3755         bool wait_ordered = false;
3756
3757 again:
3758         ret = 0;
3759         spin_lock(&space_info->lock);
3760         /*
3761          * We only want to wait if somebody other than us is flushing and we are
3762          * actually alloed to flush.
3763          */
3764         while (flush && !flushing && space_info->flush) {
3765                 spin_unlock(&space_info->lock);
3766                 /*
3767                  * If we have a trans handle we can't wait because the flusher
3768                  * may have to commit the transaction, which would mean we would
3769                  * deadlock since we are waiting for the flusher to finish, but
3770                  * hold the current transaction open.
3771                  */
3772                 if (current->journal_info)
3773                         return -EAGAIN;
3774                 ret = wait_event_interruptible(space_info->wait,
3775                                                !space_info->flush);
3776                 /* Must have been interrupted, return */
3777                 if (ret) {
3778                         printk(KERN_DEBUG "btrfs: %s returning -EINTR\n", __func__);
3779                         return -EINTR;
3780                 }
3781
3782                 spin_lock(&space_info->lock);
3783         }
3784
3785         ret = -ENOSPC;
3786         used = space_info->bytes_used + space_info->bytes_reserved +
3787                 space_info->bytes_pinned + space_info->bytes_readonly +
3788                 space_info->bytes_may_use;
3789
3790         /*
3791          * The idea here is that we've not already over-reserved the block group
3792          * then we can go ahead and save our reservation first and then start
3793          * flushing if we need to.  Otherwise if we've already overcommitted
3794          * lets start flushing stuff first and then come back and try to make
3795          * our reservation.
3796          */
3797         if (used <= space_info->total_bytes) {
3798                 if (used + orig_bytes <= space_info->total_bytes) {
3799                         space_info->bytes_may_use += orig_bytes;
3800                         trace_btrfs_space_reservation(root->fs_info,
3801                                 "space_info", space_info->flags, orig_bytes, 1);
3802                         ret = 0;
3803                 } else {
3804                         /*
3805                          * Ok set num_bytes to orig_bytes since we aren't
3806                          * overocmmitted, this way we only try and reclaim what
3807                          * we need.
3808                          */
3809                         num_bytes = orig_bytes;
3810                 }
3811         } else {
3812                 /*
3813                  * Ok we're over committed, set num_bytes to the overcommitted
3814                  * amount plus the amount of bytes that we need for this
3815                  * reservation.
3816                  */
3817                 wait_ordered = true;
3818                 num_bytes = used - space_info->total_bytes +
3819                         (orig_bytes * (retries + 1));
3820         }
3821
3822         if (ret) {
3823                 u64 profile = btrfs_get_alloc_profile(root, 0);
3824                 u64 avail;
3825
3826                 /*
3827                  * If we have a lot of space that's pinned, don't bother doing
3828                  * the overcommit dance yet and just commit the transaction.
3829                  */
3830                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3831                 do_div(avail, 10);
3832                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3833                         space_info->flush = 1;
3834                         flushing = true;
3835                         spin_unlock(&space_info->lock);
3836                         ret = may_commit_transaction(root, space_info,
3837                                                      orig_bytes, 1);
3838                         if (ret)
3839                                 goto out;
3840                         committed = true;
3841                         goto again;
3842                 }
3843
3844                 spin_lock(&root->fs_info->free_chunk_lock);
3845                 avail = root->fs_info->free_chunk_space;
3846
3847                 /*
3848                  * If we have dup, raid1 or raid10 then only half of the free
3849                  * space is actually useable.
3850                  */
3851                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3852                                BTRFS_BLOCK_GROUP_RAID1 |
3853                                BTRFS_BLOCK_GROUP_RAID10))
3854                         avail >>= 1;
3855
3856                 /*
3857                  * If we aren't flushing don't let us overcommit too much, say
3858                  * 1/8th of the space.  If we can flush, let it overcommit up to
3859                  * 1/2 of the space.
3860                  */
3861                 if (flush)
3862                         avail >>= 3;
3863                 else
3864                         avail >>= 1;
3865                  spin_unlock(&root->fs_info->free_chunk_lock);
3866
3867                 if (used + num_bytes < space_info->total_bytes + avail) {
3868                         space_info->bytes_may_use += orig_bytes;
3869                         trace_btrfs_space_reservation(root->fs_info,
3870                                 "space_info", space_info->flags, orig_bytes, 1);
3871                         ret = 0;
3872                 } else {
3873                         wait_ordered = true;
3874                 }
3875         }
3876
3877         /*
3878          * Couldn't make our reservation, save our place so while we're trying
3879          * to reclaim space we can actually use it instead of somebody else
3880          * stealing it from us.
3881          */
3882         if (ret && flush) {
3883                 flushing = true;
3884                 space_info->flush = 1;
3885         }
3886
3887         spin_unlock(&space_info->lock);
3888
3889         if (!ret || !flush)
3890                 goto out;
3891
3892         /*
3893          * We do synchronous shrinking since we don't actually unreserve
3894          * metadata until after the IO is completed.
3895          */
3896         ret = shrink_delalloc(root, num_bytes, wait_ordered);
3897         if (ret < 0)
3898                 goto out;
3899
3900         ret = 0;
3901
3902         /*
3903          * So if we were overcommitted it's possible that somebody else flushed
3904          * out enough space and we simply didn't have enough space to reclaim,
3905          * so go back around and try again.
3906          */
3907         if (retries < 2) {
3908                 wait_ordered = true;
3909                 retries++;
3910                 goto again;
3911         }
3912
3913         ret = -ENOSPC;
3914         if (committed)
3915                 goto out;
3916
3917         ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3918         if (!ret) {
3919                 committed = true;
3920                 goto again;
3921         }
3922
3923 out:
3924         if (flushing) {
3925                 spin_lock(&space_info->lock);
3926                 space_info->flush = 0;
3927                 wake_up_all(&space_info->wait);
3928                 spin_unlock(&space_info->lock);
3929         }
3930         return ret;
3931 }
3932
3933 static struct btrfs_block_rsv *get_block_rsv(
3934                                         const struct btrfs_trans_handle *trans,
3935                                         const struct btrfs_root *root)
3936 {
3937         struct btrfs_block_rsv *block_rsv = NULL;
3938
3939         if (root->ref_cows || root == root->fs_info->csum_root)
3940                 block_rsv = trans->block_rsv;
3941
3942         if (!block_rsv)
3943                 block_rsv = root->block_rsv;
3944
3945         if (!block_rsv)
3946                 block_rsv = &root->fs_info->empty_block_rsv;
3947
3948         return block_rsv;
3949 }
3950
3951 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3952                                u64 num_bytes)
3953 {
3954         int ret = -ENOSPC;
3955         spin_lock(&block_rsv->lock);
3956         if (block_rsv->reserved >= num_bytes) {
3957                 block_rsv->reserved -= num_bytes;
3958                 if (block_rsv->reserved < block_rsv->size)
3959                         block_rsv->full = 0;
3960                 ret = 0;
3961         }
3962         spin_unlock(&block_rsv->lock);
3963         return ret;
3964 }
3965
3966 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3967                                 u64 num_bytes, int update_size)
3968 {
3969         spin_lock(&block_rsv->lock);
3970         block_rsv->reserved += num_bytes;
3971         if (update_size)
3972                 block_rsv->size += num_bytes;
3973         else if (block_rsv->reserved >= block_rsv->size)
3974                 block_rsv->full = 1;
3975         spin_unlock(&block_rsv->lock);
3976 }
3977
3978 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
3979                                     struct btrfs_block_rsv *block_rsv,
3980                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3981 {
3982         struct btrfs_space_info *space_info = block_rsv->space_info;
3983
3984         spin_lock(&block_rsv->lock);
3985         if (num_bytes == (u64)-1)
3986                 num_bytes = block_rsv->size;
3987         block_rsv->size -= num_bytes;
3988         if (block_rsv->reserved >= block_rsv->size) {
3989                 num_bytes = block_rsv->reserved - block_rsv->size;
3990                 block_rsv->reserved = block_rsv->size;
3991                 block_rsv->full = 1;
3992         } else {
3993                 num_bytes = 0;
3994         }
3995         spin_unlock(&block_rsv->lock);
3996
3997         if (num_bytes > 0) {
3998                 if (dest) {
3999                         spin_lock(&dest->lock);
4000                         if (!dest->full) {
4001                                 u64 bytes_to_add;
4002
4003                                 bytes_to_add = dest->size - dest->reserved;
4004                                 bytes_to_add = min(num_bytes, bytes_to_add);
4005                                 dest->reserved += bytes_to_add;
4006                                 if (dest->reserved >= dest->size)
4007                                         dest->full = 1;
4008                                 num_bytes -= bytes_to_add;
4009                         }
4010                         spin_unlock(&dest->lock);
4011                 }
4012                 if (num_bytes) {
4013                         spin_lock(&space_info->lock);
4014                         space_info->bytes_may_use -= num_bytes;
4015                         trace_btrfs_space_reservation(fs_info, "space_info",
4016                                         space_info->flags, num_bytes, 0);
4017                         space_info->reservation_progress++;
4018                         spin_unlock(&space_info->lock);
4019                 }
4020         }
4021 }
4022
4023 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4024                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4025 {
4026         int ret;
4027
4028         ret = block_rsv_use_bytes(src, num_bytes);
4029         if (ret)
4030                 return ret;
4031
4032         block_rsv_add_bytes(dst, num_bytes, 1);
4033         return 0;
4034 }
4035
4036 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
4037 {
4038         memset(rsv, 0, sizeof(*rsv));
4039         spin_lock_init(&rsv->lock);
4040 }
4041
4042 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
4043 {
4044         struct btrfs_block_rsv *block_rsv;
4045         struct btrfs_fs_info *fs_info = root->fs_info;
4046
4047         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4048         if (!block_rsv)
4049                 return NULL;
4050
4051         btrfs_init_block_rsv(block_rsv);
4052         block_rsv->space_info = __find_space_info(fs_info,
4053                                                   BTRFS_BLOCK_GROUP_METADATA);
4054         return block_rsv;
4055 }
4056
4057 void btrfs_free_block_rsv(struct btrfs_root *root,
4058                           struct btrfs_block_rsv *rsv)
4059 {
4060         btrfs_block_rsv_release(root, rsv, (u64)-1);
4061         kfree(rsv);
4062 }
4063
4064 static inline int __block_rsv_add(struct btrfs_root *root,
4065                                   struct btrfs_block_rsv *block_rsv,
4066                                   u64 num_bytes, int flush)
4067 {
4068         int ret;
4069
4070         if (num_bytes == 0)
4071                 return 0;
4072
4073         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4074         if (!ret) {
4075                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4076                 return 0;
4077         }
4078
4079         return ret;
4080 }
4081
4082 int btrfs_block_rsv_add(struct btrfs_root *root,
4083                         struct btrfs_block_rsv *block_rsv,
4084                         u64 num_bytes)
4085 {
4086         return __block_rsv_add(root, block_rsv, num_bytes, 1);
4087 }
4088
4089 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
4090                                 struct btrfs_block_rsv *block_rsv,
4091                                 u64 num_bytes)
4092 {
4093         return __block_rsv_add(root, block_rsv, num_bytes, 0);
4094 }
4095
4096 int btrfs_block_rsv_check(struct btrfs_root *root,
4097                           struct btrfs_block_rsv *block_rsv, int min_factor)
4098 {
4099         u64 num_bytes = 0;
4100         int ret = -ENOSPC;
4101
4102         if (!block_rsv)
4103                 return 0;
4104
4105         spin_lock(&block_rsv->lock);
4106         num_bytes = div_factor(block_rsv->size, min_factor);
4107         if (block_rsv->reserved >= num_bytes)
4108                 ret = 0;
4109         spin_unlock(&block_rsv->lock);
4110
4111         return ret;
4112 }
4113
4114 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
4115                                            struct btrfs_block_rsv *block_rsv,
4116                                            u64 min_reserved, int flush)
4117 {
4118         u64 num_bytes = 0;
4119         int ret = -ENOSPC;
4120
4121         if (!block_rsv)
4122                 return 0;
4123
4124         spin_lock(&block_rsv->lock);
4125         num_bytes = min_reserved;
4126         if (block_rsv->reserved >= num_bytes)
4127                 ret = 0;
4128         else
4129                 num_bytes -= block_rsv->reserved;
4130         spin_unlock(&block_rsv->lock);
4131
4132         if (!ret)
4133                 return 0;
4134
4135         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4136         if (!ret) {
4137                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4138                 return 0;
4139         }
4140
4141         return ret;
4142 }
4143
4144 int btrfs_block_rsv_refill(struct btrfs_root *root,
4145                            struct btrfs_block_rsv *block_rsv,
4146                            u64 min_reserved)
4147 {
4148         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
4149 }
4150
4151 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
4152                                    struct btrfs_block_rsv *block_rsv,
4153                                    u64 min_reserved)
4154 {
4155         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
4156 }
4157
4158 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4159                             struct btrfs_block_rsv *dst_rsv,
4160                             u64 num_bytes)
4161 {
4162         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4163 }
4164
4165 void btrfs_block_rsv_release(struct btrfs_root *root,
4166                              struct btrfs_block_rsv *block_rsv,
4167                              u64 num_bytes)
4168 {
4169         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4170         if (global_rsv->full || global_rsv == block_rsv ||
4171             block_rsv->space_info != global_rsv->space_info)
4172                 global_rsv = NULL;
4173         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4174                                 num_bytes);
4175 }
4176
4177 /*
4178  * helper to calculate size of global block reservation.
4179  * the desired value is sum of space used by extent tree,
4180  * checksum tree and root tree
4181  */
4182 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4183 {
4184         struct btrfs_space_info *sinfo;
4185         u64 num_bytes;
4186         u64 meta_used;
4187         u64 data_used;
4188         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4189
4190         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4191         spin_lock(&sinfo->lock);
4192         data_used = sinfo->bytes_used;
4193         spin_unlock(&sinfo->lock);
4194
4195         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4196         spin_lock(&sinfo->lock);
4197         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4198                 data_used = 0;
4199         meta_used = sinfo->bytes_used;
4200         spin_unlock(&sinfo->lock);
4201
4202         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4203                     csum_size * 2;
4204         num_bytes += div64_u64(data_used + meta_used, 50);
4205
4206         if (num_bytes * 3 > meta_used)
4207                 num_bytes = div64_u64(meta_used, 3);
4208
4209         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4210 }
4211
4212 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4213 {
4214         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4215         struct btrfs_space_info *sinfo = block_rsv->space_info;
4216         u64 num_bytes;
4217
4218         num_bytes = calc_global_metadata_size(fs_info);
4219
4220         spin_lock(&block_rsv->lock);
4221         spin_lock(&sinfo->lock);
4222
4223         block_rsv->size = num_bytes;
4224
4225         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4226                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4227                     sinfo->bytes_may_use;
4228
4229         if (sinfo->total_bytes > num_bytes) {
4230                 num_bytes = sinfo->total_bytes - num_bytes;
4231                 block_rsv->reserved += num_bytes;
4232                 sinfo->bytes_may_use += num_bytes;
4233                 trace_btrfs_space_reservation(fs_info, "space_info",
4234                                       sinfo->flags, num_bytes, 1);
4235         }
4236
4237         if (block_rsv->reserved >= block_rsv->size) {
4238                 num_bytes = block_rsv->reserved - block_rsv->size;
4239                 sinfo->bytes_may_use -= num_bytes;
4240                 trace_btrfs_space_reservation(fs_info, "space_info",
4241                                       sinfo->flags, num_bytes, 0);
4242                 sinfo->reservation_progress++;
4243                 block_rsv->reserved = block_rsv->size;
4244                 block_rsv->full = 1;
4245         }
4246
4247         spin_unlock(&sinfo->lock);
4248         spin_unlock(&block_rsv->lock);
4249 }
4250
4251 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4252 {
4253         struct btrfs_space_info *space_info;
4254
4255         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4256         fs_info->chunk_block_rsv.space_info = space_info;
4257
4258         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4259         fs_info->global_block_rsv.space_info = space_info;
4260         fs_info->delalloc_block_rsv.space_info = space_info;
4261         fs_info->trans_block_rsv.space_info = space_info;
4262         fs_info->empty_block_rsv.space_info = space_info;
4263         fs_info->delayed_block_rsv.space_info = space_info;
4264
4265         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4266         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4267         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4268         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4269         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4270
4271         update_global_block_rsv(fs_info);
4272 }
4273
4274 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4275 {
4276         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4277                                 (u64)-1);
4278         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4279         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4280         WARN_ON(fs_info->trans_block_rsv.size > 0);
4281         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4282         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4283         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4284         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4285         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4286 }
4287
4288 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4289                                   struct btrfs_root *root)
4290 {
4291         if (!trans->bytes_reserved)
4292                 return;
4293
4294         trace_btrfs_space_reservation(root->fs_info, "transaction",
4295                                       trans->transid, trans->bytes_reserved, 0);
4296         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4297         trans->bytes_reserved = 0;
4298 }
4299
4300 /* Can only return 0 or -ENOSPC */
4301 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4302                                   struct inode *inode)
4303 {
4304         struct btrfs_root *root = BTRFS_I(inode)->root;
4305         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4306         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4307
4308         /*
4309          * We need to hold space in order to delete our orphan item once we've
4310          * added it, so this takes the reservation so we can release it later
4311          * when we are truly done with the orphan item.
4312          */
4313         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4314         trace_btrfs_space_reservation(root->fs_info, "orphan",
4315                                       btrfs_ino(inode), num_bytes, 1);
4316         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4317 }
4318
4319 void btrfs_orphan_release_metadata(struct inode *inode)
4320 {
4321         struct btrfs_root *root = BTRFS_I(inode)->root;
4322         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4323         trace_btrfs_space_reservation(root->fs_info, "orphan",
4324                                       btrfs_ino(inode), num_bytes, 0);
4325         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4326 }
4327
4328 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4329                                 struct btrfs_pending_snapshot *pending)
4330 {
4331         struct btrfs_root *root = pending->root;
4332         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4333         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4334         /*
4335          * two for root back/forward refs, two for directory entries
4336          * and one for root of the snapshot.
4337          */
4338         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4339         dst_rsv->space_info = src_rsv->space_info;
4340         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4341 }
4342
4343 /**
4344  * drop_outstanding_extent - drop an outstanding extent
4345  * @inode: the inode we're dropping the extent for
4346  *
4347  * This is called when we are freeing up an outstanding extent, either called
4348  * after an error or after an extent is written.  This will return the number of
4349  * reserved extents that need to be freed.  This must be called with
4350  * BTRFS_I(inode)->lock held.
4351  */
4352 static unsigned drop_outstanding_extent(struct inode *inode)
4353 {
4354         unsigned drop_inode_space = 0;
4355         unsigned dropped_extents = 0;
4356
4357         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4358         BTRFS_I(inode)->outstanding_extents--;
4359
4360         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4361             BTRFS_I(inode)->delalloc_meta_reserved) {
4362                 drop_inode_space = 1;
4363                 BTRFS_I(inode)->delalloc_meta_reserved = 0;
4364         }
4365
4366         /*
4367          * If we have more or the same amount of outsanding extents than we have
4368          * reserved then we need to leave the reserved extents count alone.
4369          */
4370         if (BTRFS_I(inode)->outstanding_extents >=
4371             BTRFS_I(inode)->reserved_extents)
4372                 return drop_inode_space;
4373
4374         dropped_extents = BTRFS_I(inode)->reserved_extents -
4375                 BTRFS_I(inode)->outstanding_extents;
4376         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4377         return dropped_extents + drop_inode_space;
4378 }
4379
4380 /**
4381  * calc_csum_metadata_size - return the amount of metada space that must be
4382  *      reserved/free'd for the given bytes.
4383  * @inode: the inode we're manipulating
4384  * @num_bytes: the number of bytes in question
4385  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4386  *
4387  * This adjusts the number of csum_bytes in the inode and then returns the
4388  * correct amount of metadata that must either be reserved or freed.  We
4389  * calculate how many checksums we can fit into one leaf and then divide the
4390  * number of bytes that will need to be checksumed by this value to figure out
4391  * how many checksums will be required.  If we are adding bytes then the number
4392  * may go up and we will return the number of additional bytes that must be
4393  * reserved.  If it is going down we will return the number of bytes that must
4394  * be freed.
4395  *
4396  * This must be called with BTRFS_I(inode)->lock held.
4397  */
4398 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4399                                    int reserve)
4400 {
4401         struct btrfs_root *root = BTRFS_I(inode)->root;
4402         u64 csum_size;
4403         int num_csums_per_leaf;
4404         int num_csums;
4405         int old_csums;
4406
4407         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4408             BTRFS_I(inode)->csum_bytes == 0)
4409                 return 0;
4410
4411         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4412         if (reserve)
4413                 BTRFS_I(inode)->csum_bytes += num_bytes;
4414         else
4415                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4416         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4417         num_csums_per_leaf = (int)div64_u64(csum_size,
4418                                             sizeof(struct btrfs_csum_item) +
4419                                             sizeof(struct btrfs_disk_key));
4420         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4421         num_csums = num_csums + num_csums_per_leaf - 1;
4422         num_csums = num_csums / num_csums_per_leaf;
4423
4424         old_csums = old_csums + num_csums_per_leaf - 1;
4425         old_csums = old_csums / num_csums_per_leaf;
4426
4427         /* No change, no need to reserve more */
4428         if (old_csums == num_csums)
4429                 return 0;
4430
4431         if (reserve)
4432                 return btrfs_calc_trans_metadata_size(root,
4433                                                       num_csums - old_csums);
4434
4435         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4436 }
4437
4438 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4439 {
4440         struct btrfs_root *root = BTRFS_I(inode)->root;
4441         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4442         u64 to_reserve = 0;
4443         u64 csum_bytes;
4444         unsigned nr_extents = 0;
4445         int extra_reserve = 0;
4446         int flush = 1;
4447         int ret;
4448
4449         /* Need to be holding the i_mutex here if we aren't free space cache */
4450         if (btrfs_is_free_space_inode(root, inode))
4451                 flush = 0;
4452
4453         if (flush && btrfs_transaction_in_commit(root->fs_info))
4454                 schedule_timeout(1);
4455
4456         mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4457         num_bytes = ALIGN(num_bytes, root->sectorsize);
4458
4459         spin_lock(&BTRFS_I(inode)->lock);
4460         BTRFS_I(inode)->outstanding_extents++;
4461
4462         if (BTRFS_I(inode)->outstanding_extents >
4463             BTRFS_I(inode)->reserved_extents)
4464                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4465                         BTRFS_I(inode)->reserved_extents;
4466
4467         /*
4468          * Add an item to reserve for updating the inode when we complete the
4469          * delalloc io.
4470          */
4471         if (!BTRFS_I(inode)->delalloc_meta_reserved) {
4472                 nr_extents++;
4473                 extra_reserve = 1;
4474         }
4475
4476         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4477         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4478         csum_bytes = BTRFS_I(inode)->csum_bytes;
4479         spin_unlock(&BTRFS_I(inode)->lock);
4480
4481         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4482         if (ret) {
4483                 u64 to_free = 0;
4484                 unsigned dropped;
4485
4486                 spin_lock(&BTRFS_I(inode)->lock);
4487                 dropped = drop_outstanding_extent(inode);
4488                 /*
4489                  * If the inodes csum_bytes is the same as the original
4490                  * csum_bytes then we know we haven't raced with any free()ers
4491                  * so we can just reduce our inodes csum bytes and carry on.
4492                  * Otherwise we have to do the normal free thing to account for
4493                  * the case that the free side didn't free up its reserve
4494                  * because of this outstanding reservation.
4495                  */
4496                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4497                         calc_csum_metadata_size(inode, num_bytes, 0);
4498                 else
4499                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4500                 spin_unlock(&BTRFS_I(inode)->lock);
4501                 if (dropped)
4502                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4503
4504                 if (to_free) {
4505                         btrfs_block_rsv_release(root, block_rsv, to_free);
4506                         trace_btrfs_space_reservation(root->fs_info,
4507                                                       "delalloc",
4508                                                       btrfs_ino(inode),
4509                                                       to_free, 0);
4510                 }
4511                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4512                 return ret;
4513         }
4514
4515         spin_lock(&BTRFS_I(inode)->lock);
4516         if (extra_reserve) {
4517                 BTRFS_I(inode)->delalloc_meta_reserved = 1;
4518                 nr_extents--;
4519         }
4520         BTRFS_I(inode)->reserved_extents += nr_extents;
4521         spin_unlock(&BTRFS_I(inode)->lock);
4522         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4523
4524         if (to_reserve)
4525                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4526                                               btrfs_ino(inode), to_reserve, 1);
4527         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4528
4529         return 0;
4530 }
4531
4532 /**
4533  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4534  * @inode: the inode to release the reservation for
4535  * @num_bytes: the number of bytes we're releasing
4536  *
4537  * This will release the metadata reservation for an inode.  This can be called
4538  * once we complete IO for a given set of bytes to release their metadata
4539  * reservations.
4540  */
4541 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4542 {
4543         struct btrfs_root *root = BTRFS_I(inode)->root;
4544         u64 to_free = 0;
4545         unsigned dropped;
4546
4547         num_bytes = ALIGN(num_bytes, root->sectorsize);
4548         spin_lock(&BTRFS_I(inode)->lock);
4549         dropped = drop_outstanding_extent(inode);
4550
4551         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4552         spin_unlock(&BTRFS_I(inode)->lock);
4553         if (dropped > 0)
4554                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4555
4556         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4557                                       btrfs_ino(inode), to_free, 0);
4558         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4559                                 to_free);
4560 }
4561
4562 /**
4563  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4564  * @inode: inode we're writing to
4565  * @num_bytes: the number of bytes we want to allocate
4566  *
4567  * This will do the following things
4568  *
4569  * o reserve space in the data space info for num_bytes
4570  * o reserve space in the metadata space info based on number of outstanding
4571  *   extents and how much csums will be needed
4572  * o add to the inodes ->delalloc_bytes
4573  * o add it to the fs_info's delalloc inodes list.
4574  *
4575  * This will return 0 for success and -ENOSPC if there is no space left.
4576  */
4577 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4578 {
4579         int ret;
4580
4581         ret = btrfs_check_data_free_space(inode, num_bytes);
4582         if (ret)
4583                 return ret;
4584
4585         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4586         if (ret) {
4587                 btrfs_free_reserved_data_space(inode, num_bytes);
4588                 return ret;
4589         }
4590
4591         return 0;
4592 }
4593
4594 /**
4595  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4596  * @inode: inode we're releasing space for
4597  * @num_bytes: the number of bytes we want to free up
4598  *
4599  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4600  * called in the case that we don't need the metadata AND data reservations
4601  * anymore.  So if there is an error or we insert an inline extent.
4602  *
4603  * This function will release the metadata space that was not used and will
4604  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4605  * list if there are no delalloc bytes left.
4606  */
4607 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4608 {
4609         btrfs_delalloc_release_metadata(inode, num_bytes);
4610         btrfs_free_reserved_data_space(inode, num_bytes);
4611 }
4612
4613 static int update_block_group(struct btrfs_trans_handle *trans,
4614                               struct btrfs_root *root,
4615                               u64 bytenr, u64 num_bytes, int alloc)
4616 {
4617         struct btrfs_block_group_cache *cache = NULL;
4618         struct btrfs_fs_info *info = root->fs_info;
4619         u64 total = num_bytes;
4620         u64 old_val;
4621         u64 byte_in_group;
4622         int factor;
4623
4624         /* block accounting for super block */
4625         spin_lock(&info->delalloc_lock);
4626         old_val = btrfs_super_bytes_used(info->super_copy);
4627         if (alloc)
4628                 old_val += num_bytes;
4629         else
4630                 old_val -= num_bytes;
4631         btrfs_set_super_bytes_used(info->super_copy, old_val);
4632         spin_unlock(&info->delalloc_lock);
4633
4634         while (total) {
4635                 cache = btrfs_lookup_block_group(info, bytenr);
4636                 if (!cache)
4637                         return -ENOENT;
4638                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4639                                     BTRFS_BLOCK_GROUP_RAID1 |
4640                                     BTRFS_BLOCK_GROUP_RAID10))
4641                         factor = 2;
4642                 else
4643                         factor = 1;
4644                 /*
4645                  * If this block group has free space cache written out, we
4646                  * need to make sure to load it if we are removing space.  This
4647                  * is because we need the unpinning stage to actually add the
4648                  * space back to the block group, otherwise we will leak space.
4649                  */
4650                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4651                         cache_block_group(cache, trans, NULL, 1);
4652
4653                 byte_in_group = bytenr - cache->key.objectid;
4654                 WARN_ON(byte_in_group > cache->key.offset);
4655
4656                 spin_lock(&cache->space_info->lock);
4657                 spin_lock(&cache->lock);
4658
4659                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4660                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4661                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4662
4663                 cache->dirty = 1;
4664                 old_val = btrfs_block_group_used(&cache->item);
4665                 num_bytes = min(total, cache->key.offset - byte_in_group);
4666                 if (alloc) {
4667                         old_val += num_bytes;
4668                         btrfs_set_block_group_used(&cache->item, old_val);
4669                         cache->reserved -= num_bytes;
4670                         cache->space_info->bytes_reserved -= num_bytes;
4671                         cache->space_info->bytes_used += num_bytes;
4672                         cache->space_info->disk_used += num_bytes * factor;
4673                         spin_unlock(&cache->lock);
4674                         spin_unlock(&cache->space_info->lock);
4675                 } else {
4676                         old_val -= num_bytes;
4677                         btrfs_set_block_group_used(&cache->item, old_val);
4678                         cache->pinned += num_bytes;
4679                         cache->space_info->bytes_pinned += num_bytes;
4680                         cache->space_info->bytes_used -= num_bytes;
4681                         cache->space_info->disk_used -= num_bytes * factor;
4682                         spin_unlock(&cache->lock);
4683                         spin_unlock(&cache->space_info->lock);
4684
4685                         set_extent_dirty(info->pinned_extents,
4686                                          bytenr, bytenr + num_bytes - 1,
4687                                          GFP_NOFS | __GFP_NOFAIL);
4688                 }
4689                 btrfs_put_block_group(cache);
4690                 total -= num_bytes;
4691                 bytenr += num_bytes;
4692         }
4693         return 0;
4694 }
4695
4696 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4697 {
4698         struct btrfs_block_group_cache *cache;
4699         u64 bytenr;
4700
4701         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4702         if (!cache)
4703                 return 0;
4704
4705         bytenr = cache->key.objectid;
4706         btrfs_put_block_group(cache);
4707
4708         return bytenr;
4709 }
4710
4711 static int pin_down_extent(struct btrfs_root *root,
4712                            struct btrfs_block_group_cache *cache,
4713                            u64 bytenr, u64 num_bytes, int reserved)
4714 {
4715         spin_lock(&cache->space_info->lock);
4716         spin_lock(&cache->lock);
4717         cache->pinned += num_bytes;
4718         cache->space_info->bytes_pinned += num_bytes;
4719         if (reserved) {
4720                 cache->reserved -= num_bytes;
4721                 cache->space_info->bytes_reserved -= num_bytes;
4722         }
4723         spin_unlock(&cache->lock);
4724         spin_unlock(&cache->space_info->lock);
4725
4726         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4727                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4728         return 0;
4729 }
4730
4731 /*
4732  * this function must be called within transaction
4733  */
4734 int btrfs_pin_extent(struct btrfs_root *root,
4735                      u64 bytenr, u64 num_bytes, int reserved)
4736 {
4737         struct btrfs_block_group_cache *cache;
4738
4739         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4740         BUG_ON(!cache); /* Logic error */
4741
4742         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4743
4744         btrfs_put_block_group(cache);
4745         return 0;
4746 }
4747
4748 /*
4749  * this function must be called within transaction
4750  */
4751 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4752                                     struct btrfs_root *root,
4753                                     u64 bytenr, u64 num_bytes)
4754 {
4755         struct btrfs_block_group_cache *cache;
4756
4757         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4758         BUG_ON(!cache); /* Logic error */
4759
4760         /*
4761          * pull in the free space cache (if any) so that our pin
4762          * removes the free space from the cache.  We have load_only set
4763          * to one because the slow code to read in the free extents does check
4764          * the pinned extents.
4765          */
4766         cache_block_group(cache, trans, root, 1);
4767
4768         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4769
4770         /* remove us from the free space cache (if we're there at all) */
4771         btrfs_remove_free_space(cache, bytenr, num_bytes);
4772         btrfs_put_block_group(cache);
4773         return 0;
4774 }
4775
4776 /**
4777  * btrfs_update_reserved_bytes - update the block_group and space info counters
4778  * @cache:      The cache we are manipulating
4779  * @num_bytes:  The number of bytes in question
4780  * @reserve:    One of the reservation enums
4781  *
4782  * This is called by the allocator when it reserves space, or by somebody who is
4783  * freeing space that was never actually used on disk.  For example if you
4784  * reserve some space for a new leaf in transaction A and before transaction A
4785  * commits you free that leaf, you call this with reserve set to 0 in order to
4786  * clear the reservation.
4787  *
4788  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4789  * ENOSPC accounting.  For data we handle the reservation through clearing the
4790  * delalloc bits in the io_tree.  We have to do this since we could end up
4791  * allocating less disk space for the amount of data we have reserved in the
4792  * case of compression.
4793  *
4794  * If this is a reservation and the block group has become read only we cannot
4795  * make the reservation and return -EAGAIN, otherwise this function always
4796  * succeeds.
4797  */
4798 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4799                                        u64 num_bytes, int reserve)
4800 {
4801         struct btrfs_space_info *space_info = cache->space_info;
4802         int ret = 0;
4803
4804         spin_lock(&space_info->lock);
4805         spin_lock(&cache->lock);
4806         if (reserve != RESERVE_FREE) {
4807                 if (cache->ro) {
4808                         ret = -EAGAIN;
4809                 } else {
4810                         cache->reserved += num_bytes;
4811                         space_info->bytes_reserved += num_bytes;
4812                         if (reserve == RESERVE_ALLOC) {
4813                                 trace_btrfs_space_reservation(cache->fs_info,
4814                                                 "space_info", space_info->flags,
4815                                                 num_bytes, 0);
4816                                 space_info->bytes_may_use -= num_bytes;
4817                         }
4818                 }
4819         } else {
4820                 if (cache->ro)
4821                         space_info->bytes_readonly += num_bytes;
4822                 cache->reserved -= num_bytes;
4823                 space_info->bytes_reserved -= num_bytes;
4824                 space_info->reservation_progress++;
4825         }
4826         spin_unlock(&cache->lock);
4827         spin_unlock(&space_info->lock);
4828         return ret;
4829 }
4830
4831 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4832                                 struct btrfs_root *root)
4833 {
4834         struct btrfs_fs_info *fs_info = root->fs_info;
4835         struct btrfs_caching_control *next;
4836         struct btrfs_caching_control *caching_ctl;
4837         struct btrfs_block_group_cache *cache;
4838
4839         down_write(&fs_info->extent_commit_sem);
4840
4841         list_for_each_entry_safe(caching_ctl, next,
4842                                  &fs_info->caching_block_groups, list) {
4843                 cache = caching_ctl->block_group;
4844                 if (block_group_cache_done(cache)) {
4845                         cache->last_byte_to_unpin = (u64)-1;
4846                         list_del_init(&caching_ctl->list);
4847                         put_caching_control(caching_ctl);
4848                 } else {
4849                         cache->last_byte_to_unpin = caching_ctl->progress;
4850                 }
4851         }
4852
4853         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4854                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4855         else
4856                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4857
4858         up_write(&fs_info->extent_commit_sem);
4859
4860         update_global_block_rsv(fs_info);
4861 }
4862
4863 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4864 {
4865         struct btrfs_fs_info *fs_info = root->fs_info;
4866         struct btrfs_block_group_cache *cache = NULL;
4867         u64 len;
4868
4869         while (start <= end) {
4870                 if (!cache ||
4871                     start >= cache->key.objectid + cache->key.offset) {
4872                         if (cache)
4873                                 btrfs_put_block_group(cache);
4874                         cache = btrfs_lookup_block_group(fs_info, start);
4875                         BUG_ON(!cache); /* Logic error */
4876                 }
4877
4878                 len = cache->key.objectid + cache->key.offset - start;
4879                 len = min(len, end + 1 - start);
4880
4881                 if (start < cache->last_byte_to_unpin) {
4882                         len = min(len, cache->last_byte_to_unpin - start);
4883                         btrfs_add_free_space(cache, start, len);
4884                 }
4885
4886                 start += len;
4887
4888                 spin_lock(&cache->space_info->lock);
4889                 spin_lock(&cache->lock);
4890                 cache->pinned -= len;
4891                 cache->space_info->bytes_pinned -= len;
4892                 if (cache->ro)
4893                         cache->space_info->bytes_readonly += len;
4894                 spin_unlock(&cache->lock);
4895                 spin_unlock(&cache->space_info->lock);
4896         }
4897
4898         if (cache)
4899                 btrfs_put_block_group(cache);
4900         return 0;
4901 }
4902
4903 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4904                                struct btrfs_root *root)
4905 {
4906         struct btrfs_fs_info *fs_info = root->fs_info;
4907         struct extent_io_tree *unpin;
4908         u64 start;
4909         u64 end;
4910         int ret;
4911
4912         if (trans->aborted)
4913                 return 0;
4914
4915         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4916                 unpin = &fs_info->freed_extents[1];
4917         else
4918                 unpin = &fs_info->freed_extents[0];
4919
4920         while (1) {
4921                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4922                                             EXTENT_DIRTY);
4923                 if (ret)
4924                         break;
4925
4926                 if (btrfs_test_opt(root, DISCARD))
4927                         ret = btrfs_discard_extent(root, start,
4928                                                    end + 1 - start, NULL);
4929
4930                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4931                 unpin_extent_range(root, start, end);
4932                 cond_resched();
4933         }
4934
4935         return 0;
4936 }
4937
4938 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4939                                 struct btrfs_root *root,
4940                                 u64 bytenr, u64 num_bytes, u64 parent,
4941                                 u64 root_objectid, u64 owner_objectid,
4942                                 u64 owner_offset, int refs_to_drop,
4943                                 struct btrfs_delayed_extent_op *extent_op)
4944 {
4945         struct btrfs_key key;
4946         struct btrfs_path *path;
4947         struct btrfs_fs_info *info = root->fs_info;
4948         struct btrfs_root *extent_root = info->extent_root;
4949         struct extent_buffer *leaf;
4950         struct btrfs_extent_item *ei;
4951         struct btrfs_extent_inline_ref *iref;
4952         int ret;
4953         int is_data;
4954         int extent_slot = 0;
4955         int found_extent = 0;
4956         int num_to_del = 1;
4957         u32 item_size;
4958         u64 refs;
4959
4960         path = btrfs_alloc_path();
4961         if (!path)
4962                 return -ENOMEM;
4963
4964         path->reada = 1;
4965         path->leave_spinning = 1;
4966
4967         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4968         BUG_ON(!is_data && refs_to_drop != 1);
4969
4970         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4971                                     bytenr, num_bytes, parent,
4972                                     root_objectid, owner_objectid,
4973                                     owner_offset);
4974         if (ret == 0) {
4975                 extent_slot = path->slots[0];
4976                 while (extent_slot >= 0) {
4977                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4978                                               extent_slot);
4979                         if (key.objectid != bytenr)
4980                                 break;
4981                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4982                             key.offset == num_bytes) {
4983                                 found_extent = 1;
4984                                 break;
4985                         }
4986                         if (path->slots[0] - extent_slot > 5)
4987                                 break;
4988                         extent_slot--;
4989                 }
4990 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4991                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4992                 if (found_extent && item_size < sizeof(*ei))
4993                         found_extent = 0;
4994 #endif
4995                 if (!found_extent) {
4996                         BUG_ON(iref);
4997                         ret = remove_extent_backref(trans, extent_root, path,
4998                                                     NULL, refs_to_drop,
4999                                                     is_data);
5000                         if (ret)
5001                                 goto abort;
5002                         btrfs_release_path(path);
5003                         path->leave_spinning = 1;
5004
5005                         key.objectid = bytenr;
5006                         key.type = BTRFS_EXTENT_ITEM_KEY;
5007                         key.offset = num_bytes;
5008
5009                         ret = btrfs_search_slot(trans, extent_root,
5010                                                 &key, path, -1, 1);
5011                         if (ret) {
5012                                 printk(KERN_ERR "umm, got %d back from search"
5013                                        ", was looking for %llu\n", ret,
5014                                        (unsigned long long)bytenr);
5015                                 if (ret > 0)
5016                                         btrfs_print_leaf(extent_root,
5017                                                          path->nodes[0]);
5018                         }
5019                         if (ret < 0)
5020                                 goto abort;
5021                         extent_slot = path->slots[0];
5022                 }
5023         } else if (ret == -ENOENT) {
5024                 btrfs_print_leaf(extent_root, path->nodes[0]);
5025                 WARN_ON(1);
5026                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
5027                        "parent %llu root %llu  owner %llu offset %llu\n",
5028                        (unsigned long long)bytenr,
5029                        (unsigned long long)parent,
5030                        (unsigned long long)root_objectid,
5031                        (unsigned long long)owner_objectid,
5032                        (unsigned long long)owner_offset);
5033         } else {
5034                 goto abort;
5035         }
5036
5037         leaf = path->nodes[0];
5038         item_size = btrfs_item_size_nr(leaf, extent_slot);
5039 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5040         if (item_size < sizeof(*ei)) {
5041                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5042                 ret = convert_extent_item_v0(trans, extent_root, path,
5043                                              owner_objectid, 0);
5044                 if (ret < 0)
5045                         goto abort;
5046
5047                 btrfs_release_path(path);
5048                 path->leave_spinning = 1;
5049
5050                 key.objectid = bytenr;
5051                 key.type = BTRFS_EXTENT_ITEM_KEY;
5052                 key.offset = num_bytes;
5053
5054                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5055                                         -1, 1);
5056                 if (ret) {
5057                         printk(KERN_ERR "umm, got %d back from search"
5058                                ", was looking for %llu\n", ret,
5059                                (unsigned long long)bytenr);
5060                         btrfs_print_leaf(extent_root, path->nodes[0]);
5061                 }
5062                 if (ret < 0)
5063                         goto abort;
5064                 extent_slot = path->slots[0];
5065                 leaf = path->nodes[0];
5066                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5067         }
5068 #endif
5069         BUG_ON(item_size < sizeof(*ei));
5070         ei = btrfs_item_ptr(leaf, extent_slot,
5071                             struct btrfs_extent_item);
5072         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5073                 struct btrfs_tree_block_info *bi;
5074                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5075                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5076                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5077         }
5078
5079         refs = btrfs_extent_refs(leaf, ei);
5080         BUG_ON(refs < refs_to_drop);
5081         refs -= refs_to_drop;
5082
5083         if (refs > 0) {
5084                 if (extent_op)
5085                         __run_delayed_extent_op(extent_op, leaf, ei);
5086                 /*
5087                  * In the case of inline back ref, reference count will
5088                  * be updated by remove_extent_backref
5089                  */
5090                 if (iref) {
5091                         BUG_ON(!found_extent);
5092                 } else {
5093                         btrfs_set_extent_refs(leaf, ei, refs);
5094                         btrfs_mark_buffer_dirty(leaf);
5095                 }
5096                 if (found_extent) {
5097                         ret = remove_extent_backref(trans, extent_root, path,
5098                                                     iref, refs_to_drop,
5099                                                     is_data);
5100                         if (ret)
5101                                 goto abort;
5102                 }
5103         } else {
5104                 if (found_extent) {
5105                         BUG_ON(is_data && refs_to_drop !=
5106                                extent_data_ref_count(root, path, iref));
5107                         if (iref) {
5108                                 BUG_ON(path->slots[0] != extent_slot);
5109                         } else {
5110                                 BUG_ON(path->slots[0] != extent_slot + 1);
5111                                 path->slots[0] = extent_slot;
5112                                 num_to_del = 2;
5113                         }
5114                 }
5115
5116                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5117                                       num_to_del);
5118                 if (ret)
5119                         goto abort;
5120                 btrfs_release_path(path);
5121
5122                 if (is_data) {
5123                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5124                         if (ret)
5125                                 goto abort;
5126                 }
5127
5128                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
5129                 if (ret)
5130                         goto abort;
5131         }
5132 out:
5133         btrfs_free_path(path);
5134         return ret;
5135
5136 abort:
5137         btrfs_abort_transaction(trans, extent_root, ret);
5138         goto out;
5139 }
5140
5141 /*
5142  * when we free an block, it is possible (and likely) that we free the last
5143  * delayed ref for that extent as well.  This searches the delayed ref tree for
5144  * a given extent, and if there are no other delayed refs to be processed, it
5145  * removes it from the tree.
5146  */
5147 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5148                                       struct btrfs_root *root, u64 bytenr)
5149 {
5150         struct btrfs_delayed_ref_head *head;
5151         struct btrfs_delayed_ref_root *delayed_refs;
5152         struct btrfs_delayed_ref_node *ref;
5153         struct rb_node *node;
5154         int ret = 0;
5155
5156         delayed_refs = &trans->transaction->delayed_refs;
5157         spin_lock(&delayed_refs->lock);
5158         head = btrfs_find_delayed_ref_head(trans, bytenr);
5159         if (!head)
5160                 goto out;
5161
5162         node = rb_prev(&head->node.rb_node);
5163         if (!node)
5164                 goto out;
5165
5166         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5167
5168         /* there are still entries for this ref, we can't drop it */
5169         if (ref->bytenr == bytenr)
5170                 goto out;
5171
5172         if (head->extent_op) {
5173                 if (!head->must_insert_reserved)
5174                         goto out;
5175                 kfree(head->extent_op);
5176                 head->extent_op = NULL;
5177         }
5178
5179         /*
5180          * waiting for the lock here would deadlock.  If someone else has it
5181          * locked they are already in the process of dropping it anyway
5182          */
5183         if (!mutex_trylock(&head->mutex))
5184                 goto out;
5185
5186         /*
5187          * at this point we have a head with no other entries.  Go
5188          * ahead and process it.
5189          */
5190         head->node.in_tree = 0;
5191         rb_erase(&head->node.rb_node, &delayed_refs->root);
5192
5193         delayed_refs->num_entries--;
5194         if (waitqueue_active(&delayed_refs->seq_wait))
5195                 wake_up(&delayed_refs->seq_wait);
5196
5197         /*
5198          * we don't take a ref on the node because we're removing it from the
5199          * tree, so we just steal the ref the tree was holding.
5200          */
5201         delayed_refs->num_heads--;
5202         if (list_empty(&head->cluster))
5203                 delayed_refs->num_heads_ready--;
5204
5205         list_del_init(&head->cluster);
5206         spin_unlock(&delayed_refs->lock);
5207
5208         BUG_ON(head->extent_op);
5209         if (head->must_insert_reserved)
5210                 ret = 1;
5211
5212         mutex_unlock(&head->mutex);
5213         btrfs_put_delayed_ref(&head->node);
5214         return ret;
5215 out:
5216         spin_unlock(&delayed_refs->lock);
5217         return 0;
5218 }
5219
5220 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5221                            struct btrfs_root *root,
5222                            struct extent_buffer *buf,
5223                            u64 parent, int last_ref, int for_cow)
5224 {
5225         struct btrfs_block_group_cache *cache = NULL;
5226         int ret;
5227
5228         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5229                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5230                                         buf->start, buf->len,
5231                                         parent, root->root_key.objectid,
5232                                         btrfs_header_level(buf),
5233                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5234                 BUG_ON(ret); /* -ENOMEM */
5235         }
5236
5237         if (!last_ref)
5238                 return;
5239
5240         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5241
5242         if (btrfs_header_generation(buf) == trans->transid) {
5243                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5244                         ret = check_ref_cleanup(trans, root, buf->start);
5245                         if (!ret)
5246                                 goto out;
5247                 }
5248
5249                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5250                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5251                         goto out;
5252                 }
5253
5254                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5255
5256                 btrfs_add_free_space(cache, buf->start, buf->len);
5257                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5258         }
5259 out:
5260         /*
5261          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5262          * anymore.
5263          */
5264         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5265         btrfs_put_block_group(cache);
5266 }
5267
5268 /* Can return -ENOMEM */
5269 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5270                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5271                       u64 owner, u64 offset, int for_cow)
5272 {
5273         int ret;
5274         struct btrfs_fs_info *fs_info = root->fs_info;
5275
5276         /*
5277          * tree log blocks never actually go into the extent allocation
5278          * tree, just update pinning info and exit early.
5279          */
5280         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5281                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5282                 /* unlocks the pinned mutex */
5283                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5284                 ret = 0;
5285         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5286                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5287                                         num_bytes,
5288                                         parent, root_objectid, (int)owner,
5289                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5290         } else {
5291                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5292                                                 num_bytes,
5293                                                 parent, root_objectid, owner,
5294                                                 offset, BTRFS_DROP_DELAYED_REF,
5295                                                 NULL, for_cow);
5296         }
5297         return ret;
5298 }
5299
5300 static u64 stripe_align(struct btrfs_root *root, u64 val)
5301 {
5302         u64 mask = ((u64)root->stripesize - 1);
5303         u64 ret = (val + mask) & ~mask;
5304         return ret;
5305 }
5306
5307 /*
5308  * when we wait for progress in the block group caching, its because
5309  * our allocation attempt failed at least once.  So, we must sleep
5310  * and let some progress happen before we try again.
5311  *
5312  * This function will sleep at least once waiting for new free space to
5313  * show up, and then it will check the block group free space numbers
5314  * for our min num_bytes.  Another option is to have it go ahead
5315  * and look in the rbtree for a free extent of a given size, but this
5316  * is a good start.
5317  */
5318 static noinline int
5319 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5320                                 u64 num_bytes)
5321 {
5322         struct btrfs_caching_control *caching_ctl;
5323         DEFINE_WAIT(wait);
5324
5325         caching_ctl = get_caching_control(cache);
5326         if (!caching_ctl)
5327                 return 0;
5328
5329         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5330                    (cache->free_space_ctl->free_space >= num_bytes));
5331
5332         put_caching_control(caching_ctl);
5333         return 0;
5334 }
5335
5336 static noinline int
5337 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5338 {
5339         struct btrfs_caching_control *caching_ctl;
5340         DEFINE_WAIT(wait);
5341
5342         caching_ctl = get_caching_control(cache);
5343         if (!caching_ctl)
5344                 return 0;
5345
5346         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5347
5348         put_caching_control(caching_ctl);
5349         return 0;
5350 }
5351
5352 static int __get_block_group_index(u64 flags)
5353 {
5354         int index;
5355
5356         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5357                 index = 0;
5358         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5359                 index = 1;
5360         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5361                 index = 2;
5362         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5363                 index = 3;
5364         else
5365                 index = 4;
5366
5367         return index;
5368 }
5369
5370 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5371 {
5372         return __get_block_group_index(cache->flags);
5373 }
5374
5375 enum btrfs_loop_type {
5376         LOOP_CACHING_NOWAIT = 0,
5377         LOOP_CACHING_WAIT = 1,
5378         LOOP_ALLOC_CHUNK = 2,
5379         LOOP_NO_EMPTY_SIZE = 3,
5380 };
5381
5382 /*
5383  * walks the btree of allocated extents and find a hole of a given size.
5384  * The key ins is changed to record the hole:
5385  * ins->objectid == block start
5386  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5387  * ins->offset == number of blocks
5388  * Any available blocks before search_start are skipped.
5389  */
5390 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5391                                      struct btrfs_root *orig_root,
5392                                      u64 num_bytes, u64 empty_size,
5393                                      u64 hint_byte, struct btrfs_key *ins,
5394                                      u64 data)
5395 {
5396         int ret = 0;
5397         struct btrfs_root *root = orig_root->fs_info->extent_root;
5398         struct btrfs_free_cluster *last_ptr = NULL;
5399         struct btrfs_block_group_cache *block_group = NULL;
5400         struct btrfs_block_group_cache *used_block_group;
5401         u64 search_start = 0;
5402         int empty_cluster = 2 * 1024 * 1024;
5403         int allowed_chunk_alloc = 0;
5404         int done_chunk_alloc = 0;
5405         struct btrfs_space_info *space_info;
5406         int loop = 0;
5407         int index = 0;
5408         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5409                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5410         bool found_uncached_bg = false;
5411         bool failed_cluster_refill = false;
5412         bool failed_alloc = false;
5413         bool use_cluster = true;
5414         bool have_caching_bg = false;
5415
5416         WARN_ON(num_bytes < root->sectorsize);
5417         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5418         ins->objectid = 0;
5419         ins->offset = 0;
5420
5421         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5422
5423         space_info = __find_space_info(root->fs_info, data);
5424         if (!space_info) {
5425                 printk(KERN_ERR "No space info for %llu\n", data);
5426                 return -ENOSPC;
5427         }
5428
5429         /*
5430          * If the space info is for both data and metadata it means we have a
5431          * small filesystem and we can't use the clustering stuff.
5432          */
5433         if (btrfs_mixed_space_info(space_info))
5434                 use_cluster = false;
5435
5436         if (orig_root->ref_cows || empty_size)
5437                 allowed_chunk_alloc = 1;
5438
5439         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5440                 last_ptr = &root->fs_info->meta_alloc_cluster;
5441                 if (!btrfs_test_opt(root, SSD))
5442                         empty_cluster = 64 * 1024;
5443         }
5444
5445         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5446             btrfs_test_opt(root, SSD)) {
5447                 last_ptr = &root->fs_info->data_alloc_cluster;
5448         }
5449
5450         if (last_ptr) {
5451                 spin_lock(&last_ptr->lock);
5452                 if (last_ptr->block_group)
5453                         hint_byte = last_ptr->window_start;
5454                 spin_unlock(&last_ptr->lock);
5455         }
5456
5457         search_start = max(search_start, first_logical_byte(root, 0));
5458         search_start = max(search_start, hint_byte);
5459
5460         if (!last_ptr)
5461                 empty_cluster = 0;
5462
5463         if (search_start == hint_byte) {
5464                 block_group = btrfs_lookup_block_group(root->fs_info,
5465                                                        search_start);
5466                 used_block_group = block_group;
5467                 /*
5468                  * we don't want to use the block group if it doesn't match our
5469                  * allocation bits, or if its not cached.
5470                  *
5471                  * However if we are re-searching with an ideal block group
5472                  * picked out then we don't care that the block group is cached.
5473                  */
5474                 if (block_group && block_group_bits(block_group, data) &&
5475                     block_group->cached != BTRFS_CACHE_NO) {
5476                         down_read(&space_info->groups_sem);
5477                         if (list_empty(&block_group->list) ||
5478                             block_group->ro) {
5479                                 /*
5480                                  * someone is removing this block group,
5481                                  * we can't jump into the have_block_group
5482                                  * target because our list pointers are not
5483                                  * valid
5484                                  */
5485                                 btrfs_put_block_group(block_group);
5486                                 up_read(&space_info->groups_sem);
5487                         } else {
5488                                 index = get_block_group_index(block_group);
5489                                 goto have_block_group;
5490                         }
5491                 } else if (block_group) {
5492                         btrfs_put_block_group(block_group);
5493                 }
5494         }
5495 search:
5496         have_caching_bg = false;
5497         down_read(&space_info->groups_sem);
5498         list_for_each_entry(block_group, &space_info->block_groups[index],
5499                             list) {
5500                 u64 offset;
5501                 int cached;
5502
5503                 used_block_group = block_group;
5504                 btrfs_get_block_group(block_group);
5505                 search_start = block_group->key.objectid;
5506
5507                 /*
5508                  * this can happen if we end up cycling through all the
5509                  * raid types, but we want to make sure we only allocate
5510                  * for the proper type.
5511                  */
5512                 if (!block_group_bits(block_group, data)) {
5513                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5514                                 BTRFS_BLOCK_GROUP_RAID1 |
5515                                 BTRFS_BLOCK_GROUP_RAID10;
5516
5517                         /*
5518                          * if they asked for extra copies and this block group
5519                          * doesn't provide them, bail.  This does allow us to
5520                          * fill raid0 from raid1.
5521                          */
5522                         if ((data & extra) && !(block_group->flags & extra))
5523                                 goto loop;
5524                 }
5525
5526 have_block_group:
5527                 cached = block_group_cache_done(block_group);
5528                 if (unlikely(!cached)) {
5529                         found_uncached_bg = true;
5530                         ret = cache_block_group(block_group, trans,
5531                                                 orig_root, 0);
5532                         BUG_ON(ret < 0);
5533                         ret = 0;
5534                 }
5535
5536                 if (unlikely(block_group->ro))
5537                         goto loop;
5538
5539                 /*
5540                  * Ok we want to try and use the cluster allocator, so
5541                  * lets look there
5542                  */
5543                 if (last_ptr) {
5544                         /*
5545                          * the refill lock keeps out other
5546                          * people trying to start a new cluster
5547                          */
5548                         spin_lock(&last_ptr->refill_lock);
5549                         used_block_group = last_ptr->block_group;
5550                         if (used_block_group != block_group &&
5551                             (!used_block_group ||
5552                              used_block_group->ro ||
5553                              !block_group_bits(used_block_group, data))) {
5554                                 used_block_group = block_group;
5555                                 goto refill_cluster;
5556                         }
5557
5558                         if (used_block_group != block_group)
5559                                 btrfs_get_block_group(used_block_group);
5560
5561                         offset = btrfs_alloc_from_cluster(used_block_group,
5562                           last_ptr, num_bytes, used_block_group->key.objectid);
5563                         if (offset) {
5564                                 /* we have a block, we're done */
5565                                 spin_unlock(&last_ptr->refill_lock);
5566                                 trace_btrfs_reserve_extent_cluster(root,
5567                                         block_group, search_start, num_bytes);
5568                                 goto checks;
5569                         }
5570
5571                         WARN_ON(last_ptr->block_group != used_block_group);
5572                         if (used_block_group != block_group) {
5573                                 btrfs_put_block_group(used_block_group);
5574                                 used_block_group = block_group;
5575                         }
5576 refill_cluster:
5577                         BUG_ON(used_block_group != block_group);
5578                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5579                          * set up a new clusters, so lets just skip it
5580                          * and let the allocator find whatever block
5581                          * it can find.  If we reach this point, we
5582                          * will have tried the cluster allocator
5583                          * plenty of times and not have found
5584                          * anything, so we are likely way too
5585                          * fragmented for the clustering stuff to find
5586                          * anything.
5587                          *
5588                          * However, if the cluster is taken from the
5589                          * current block group, release the cluster
5590                          * first, so that we stand a better chance of
5591                          * succeeding in the unclustered
5592                          * allocation.  */
5593                         if (loop >= LOOP_NO_EMPTY_SIZE &&
5594                             last_ptr->block_group != block_group) {
5595                                 spin_unlock(&last_ptr->refill_lock);
5596                                 goto unclustered_alloc;
5597                         }
5598
5599                         /*
5600                          * this cluster didn't work out, free it and
5601                          * start over
5602                          */
5603                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5604
5605                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5606                                 spin_unlock(&last_ptr->refill_lock);
5607                                 goto unclustered_alloc;
5608                         }
5609
5610                         /* allocate a cluster in this block group */
5611                         ret = btrfs_find_space_cluster(trans, root,
5612                                                block_group, last_ptr,
5613                                                search_start, num_bytes,
5614                                                empty_cluster + empty_size);
5615                         if (ret == 0) {
5616                                 /*
5617                                  * now pull our allocation out of this
5618                                  * cluster
5619                                  */
5620                                 offset = btrfs_alloc_from_cluster(block_group,
5621                                                   last_ptr, num_bytes,
5622                                                   search_start);
5623                                 if (offset) {
5624                                         /* we found one, proceed */
5625                                         spin_unlock(&last_ptr->refill_lock);
5626                                         trace_btrfs_reserve_extent_cluster(root,
5627                                                 block_group, search_start,
5628                                                 num_bytes);
5629                                         goto checks;
5630                                 }
5631                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5632                                    && !failed_cluster_refill) {
5633                                 spin_unlock(&last_ptr->refill_lock);
5634
5635                                 failed_cluster_refill = true;
5636                                 wait_block_group_cache_progress(block_group,
5637                                        num_bytes + empty_cluster + empty_size);
5638                                 goto have_block_group;
5639                         }
5640
5641                         /*
5642                          * at this point we either didn't find a cluster
5643                          * or we weren't able to allocate a block from our
5644                          * cluster.  Free the cluster we've been trying
5645                          * to use, and go to the next block group
5646                          */
5647                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5648                         spin_unlock(&last_ptr->refill_lock);
5649                         goto loop;
5650                 }
5651
5652 unclustered_alloc:
5653                 spin_lock(&block_group->free_space_ctl->tree_lock);
5654                 if (cached &&
5655                     block_group->free_space_ctl->free_space <
5656                     num_bytes + empty_cluster + empty_size) {
5657                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5658                         goto loop;
5659                 }
5660                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5661
5662                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5663                                                     num_bytes, empty_size);
5664                 /*
5665                  * If we didn't find a chunk, and we haven't failed on this
5666                  * block group before, and this block group is in the middle of
5667                  * caching and we are ok with waiting, then go ahead and wait
5668                  * for progress to be made, and set failed_alloc to true.
5669                  *
5670                  * If failed_alloc is true then we've already waited on this
5671                  * block group once and should move on to the next block group.
5672                  */
5673                 if (!offset && !failed_alloc && !cached &&
5674                     loop > LOOP_CACHING_NOWAIT) {
5675                         wait_block_group_cache_progress(block_group,
5676                                                 num_bytes + empty_size);
5677                         failed_alloc = true;
5678                         goto have_block_group;
5679                 } else if (!offset) {
5680                         if (!cached)
5681                                 have_caching_bg = true;
5682                         goto loop;
5683                 }
5684 checks:
5685                 search_start = stripe_align(root, offset);
5686
5687                 /* move on to the next group */
5688                 if (search_start + num_bytes >
5689                     used_block_group->key.objectid + used_block_group->key.offset) {
5690                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5691                         goto loop;
5692                 }
5693
5694                 if (offset < search_start)
5695                         btrfs_add_free_space(used_block_group, offset,
5696                                              search_start - offset);
5697                 BUG_ON(offset > search_start);
5698
5699                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
5700                                                   alloc_type);
5701                 if (ret == -EAGAIN) {
5702                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5703                         goto loop;
5704                 }
5705
5706                 /* we are all good, lets return */
5707                 ins->objectid = search_start;
5708                 ins->offset = num_bytes;
5709
5710                 trace_btrfs_reserve_extent(orig_root, block_group,
5711                                            search_start, num_bytes);
5712                 if (offset < search_start)
5713                         btrfs_add_free_space(used_block_group, offset,
5714                                              search_start - offset);
5715                 BUG_ON(offset > search_start);
5716                 if (used_block_group != block_group)
5717                         btrfs_put_block_group(used_block_group);
5718                 btrfs_put_block_group(block_group);
5719                 break;
5720 loop:
5721                 failed_cluster_refill = false;
5722                 failed_alloc = false;
5723                 BUG_ON(index != get_block_group_index(block_group));
5724                 if (used_block_group != block_group)
5725                         btrfs_put_block_group(used_block_group);
5726                 btrfs_put_block_group(block_group);
5727         }
5728         up_read(&space_info->groups_sem);
5729
5730         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
5731                 goto search;
5732
5733         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5734                 goto search;
5735
5736         /*
5737          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5738          *                      caching kthreads as we move along
5739          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5740          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5741          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5742          *                      again
5743          */
5744         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5745                 index = 0;
5746                 loop++;
5747                 if (loop == LOOP_ALLOC_CHUNK) {
5748                        if (allowed_chunk_alloc) {
5749                                 ret = do_chunk_alloc(trans, root, num_bytes +
5750                                                      2 * 1024 * 1024, data,
5751                                                      CHUNK_ALLOC_LIMITED);
5752                                 if (ret < 0) {
5753                                         btrfs_abort_transaction(trans,
5754                                                                 root, ret);
5755                                         goto out;
5756                                 }
5757                                 allowed_chunk_alloc = 0;
5758                                 if (ret == 1)
5759                                         done_chunk_alloc = 1;
5760                         } else if (!done_chunk_alloc &&
5761                                    space_info->force_alloc ==
5762                                    CHUNK_ALLOC_NO_FORCE) {
5763                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5764                         }
5765
5766                        /*
5767                         * We didn't allocate a chunk, go ahead and drop the
5768                         * empty size and loop again.
5769                         */
5770                        if (!done_chunk_alloc)
5771                                loop = LOOP_NO_EMPTY_SIZE;
5772                 }
5773
5774                 if (loop == LOOP_NO_EMPTY_SIZE) {
5775                         empty_size = 0;
5776                         empty_cluster = 0;
5777                 }
5778
5779                 goto search;
5780         } else if (!ins->objectid) {
5781                 ret = -ENOSPC;
5782         } else if (ins->objectid) {
5783                 ret = 0;
5784         }
5785 out:
5786
5787         return ret;
5788 }
5789
5790 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5791                             int dump_block_groups)
5792 {
5793         struct btrfs_block_group_cache *cache;
5794         int index = 0;
5795
5796         spin_lock(&info->lock);
5797         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5798                (unsigned long long)info->flags,
5799                (unsigned long long)(info->total_bytes - info->bytes_used -
5800                                     info->bytes_pinned - info->bytes_reserved -
5801                                     info->bytes_readonly),
5802                (info->full) ? "" : "not ");
5803         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5804                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5805                (unsigned long long)info->total_bytes,
5806                (unsigned long long)info->bytes_used,
5807                (unsigned long long)info->bytes_pinned,
5808                (unsigned long long)info->bytes_reserved,
5809                (unsigned long long)info->bytes_may_use,
5810                (unsigned long long)info->bytes_readonly);
5811         spin_unlock(&info->lock);
5812
5813         if (!dump_block_groups)
5814                 return;
5815
5816         down_read(&info->groups_sem);
5817 again:
5818         list_for_each_entry(cache, &info->block_groups[index], list) {
5819                 spin_lock(&cache->lock);
5820                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5821                        "%llu pinned %llu reserved\n",
5822                        (unsigned long long)cache->key.objectid,
5823                        (unsigned long long)cache->key.offset,
5824                        (unsigned long long)btrfs_block_group_used(&cache->item),
5825                        (unsigned long long)cache->pinned,
5826                        (unsigned long long)cache->reserved);
5827                 btrfs_dump_free_space(cache, bytes);
5828                 spin_unlock(&cache->lock);
5829         }
5830         if (++index < BTRFS_NR_RAID_TYPES)
5831                 goto again;
5832         up_read(&info->groups_sem);
5833 }
5834
5835 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5836                          struct btrfs_root *root,
5837                          u64 num_bytes, u64 min_alloc_size,
5838                          u64 empty_size, u64 hint_byte,
5839                          struct btrfs_key *ins, u64 data)
5840 {
5841         bool final_tried = false;
5842         int ret;
5843
5844         data = btrfs_get_alloc_profile(root, data);
5845 again:
5846         /*
5847          * the only place that sets empty_size is btrfs_realloc_node, which
5848          * is not called recursively on allocations
5849          */
5850         if (empty_size || root->ref_cows) {
5851                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5852                                      num_bytes + 2 * 1024 * 1024, data,
5853                                      CHUNK_ALLOC_NO_FORCE);
5854                 if (ret < 0 && ret != -ENOSPC) {
5855                         btrfs_abort_transaction(trans, root, ret);
5856                         return ret;
5857                 }
5858         }
5859
5860         WARN_ON(num_bytes < root->sectorsize);
5861         ret = find_free_extent(trans, root, num_bytes, empty_size,
5862                                hint_byte, ins, data);
5863
5864         if (ret == -ENOSPC) {
5865                 if (!final_tried) {
5866                         num_bytes = num_bytes >> 1;
5867                         num_bytes = num_bytes & ~(root->sectorsize - 1);
5868                         num_bytes = max(num_bytes, min_alloc_size);
5869                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5870                                        num_bytes, data, CHUNK_ALLOC_FORCE);
5871                         if (ret < 0 && ret != -ENOSPC) {
5872                                 btrfs_abort_transaction(trans, root, ret);
5873                                 return ret;
5874                         }
5875                         if (num_bytes == min_alloc_size)
5876                                 final_tried = true;
5877                         goto again;
5878                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
5879                         struct btrfs_space_info *sinfo;
5880
5881                         sinfo = __find_space_info(root->fs_info, data);
5882                         printk(KERN_ERR "btrfs allocation failed flags %llu, "
5883                                "wanted %llu\n", (unsigned long long)data,
5884                                (unsigned long long)num_bytes);
5885                         if (sinfo)
5886                                 dump_space_info(sinfo, num_bytes, 1);
5887                 }
5888         }
5889
5890         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
5891
5892         return ret;
5893 }
5894
5895 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
5896                                         u64 start, u64 len, int pin)
5897 {
5898         struct btrfs_block_group_cache *cache;
5899         int ret = 0;
5900
5901         cache = btrfs_lookup_block_group(root->fs_info, start);
5902         if (!cache) {
5903                 printk(KERN_ERR "Unable to find block group for %llu\n",
5904                        (unsigned long long)start);
5905                 return -ENOSPC;
5906         }
5907
5908         if (btrfs_test_opt(root, DISCARD))
5909                 ret = btrfs_discard_extent(root, start, len, NULL);
5910
5911         if (pin)
5912                 pin_down_extent(root, cache, start, len, 1);
5913         else {
5914                 btrfs_add_free_space(cache, start, len);
5915                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
5916         }
5917         btrfs_put_block_group(cache);
5918
5919         trace_btrfs_reserved_extent_free(root, start, len);
5920
5921         return ret;
5922 }
5923
5924 int btrfs_free_reserved_extent(struct btrfs_root *root,
5925                                         u64 start, u64 len)
5926 {
5927         return __btrfs_free_reserved_extent(root, start, len, 0);
5928 }
5929
5930 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
5931                                        u64 start, u64 len)
5932 {
5933         return __btrfs_free_reserved_extent(root, start, len, 1);
5934 }
5935
5936 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5937                                       struct btrfs_root *root,
5938                                       u64 parent, u64 root_objectid,
5939                                       u64 flags, u64 owner, u64 offset,
5940                                       struct btrfs_key *ins, int ref_mod)
5941 {
5942         int ret;
5943         struct btrfs_fs_info *fs_info = root->fs_info;
5944         struct btrfs_extent_item *extent_item;
5945         struct btrfs_extent_inline_ref *iref;
5946         struct btrfs_path *path;
5947         struct extent_buffer *leaf;
5948         int type;
5949         u32 size;
5950
5951         if (parent > 0)
5952                 type = BTRFS_SHARED_DATA_REF_KEY;
5953         else
5954                 type = BTRFS_EXTENT_DATA_REF_KEY;
5955
5956         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5957
5958         path = btrfs_alloc_path();
5959         if (!path)
5960                 return -ENOMEM;
5961
5962         path->leave_spinning = 1;
5963         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5964                                       ins, size);
5965         if (ret) {
5966                 btrfs_free_path(path);
5967                 return ret;
5968         }
5969
5970         leaf = path->nodes[0];
5971         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5972                                      struct btrfs_extent_item);
5973         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
5974         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5975         btrfs_set_extent_flags(leaf, extent_item,
5976                                flags | BTRFS_EXTENT_FLAG_DATA);
5977
5978         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
5979         btrfs_set_extent_inline_ref_type(leaf, iref, type);
5980         if (parent > 0) {
5981                 struct btrfs_shared_data_ref *ref;
5982                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
5983                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5984                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
5985         } else {
5986                 struct btrfs_extent_data_ref *ref;
5987                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
5988                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
5989                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
5990                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
5991                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
5992         }
5993
5994         btrfs_mark_buffer_dirty(path->nodes[0]);
5995         btrfs_free_path(path);
5996
5997         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
5998         if (ret) { /* -ENOENT, logic error */
5999                 printk(KERN_ERR "btrfs update block group failed for %llu "
6000                        "%llu\n", (unsigned long long)ins->objectid,
6001                        (unsigned long long)ins->offset);
6002                 BUG();
6003         }
6004         return ret;
6005 }
6006
6007 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6008                                      struct btrfs_root *root,
6009                                      u64 parent, u64 root_objectid,
6010                                      u64 flags, struct btrfs_disk_key *key,
6011                                      int level, struct btrfs_key *ins)
6012 {
6013         int ret;
6014         struct btrfs_fs_info *fs_info = root->fs_info;
6015         struct btrfs_extent_item *extent_item;
6016         struct btrfs_tree_block_info *block_info;
6017         struct btrfs_extent_inline_ref *iref;
6018         struct btrfs_path *path;
6019         struct extent_buffer *leaf;
6020         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
6021
6022         path = btrfs_alloc_path();
6023         if (!path)
6024                 return -ENOMEM;
6025
6026         path->leave_spinning = 1;
6027         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6028                                       ins, size);
6029         if (ret) {
6030                 btrfs_free_path(path);
6031                 return ret;
6032         }
6033
6034         leaf = path->nodes[0];
6035         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6036                                      struct btrfs_extent_item);
6037         btrfs_set_extent_refs(leaf, extent_item, 1);
6038         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6039         btrfs_set_extent_flags(leaf, extent_item,
6040                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6041         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6042
6043         btrfs_set_tree_block_key(leaf, block_info, key);
6044         btrfs_set_tree_block_level(leaf, block_info, level);
6045
6046         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6047         if (parent > 0) {
6048                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6049                 btrfs_set_extent_inline_ref_type(leaf, iref,
6050                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6051                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6052         } else {
6053                 btrfs_set_extent_inline_ref_type(leaf, iref,
6054                                                  BTRFS_TREE_BLOCK_REF_KEY);
6055                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6056         }
6057
6058         btrfs_mark_buffer_dirty(leaf);
6059         btrfs_free_path(path);
6060
6061         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
6062         if (ret) { /* -ENOENT, logic error */
6063                 printk(KERN_ERR "btrfs update block group failed for %llu "
6064                        "%llu\n", (unsigned long long)ins->objectid,
6065                        (unsigned long long)ins->offset);
6066                 BUG();
6067         }
6068         return ret;
6069 }
6070
6071 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6072                                      struct btrfs_root *root,
6073                                      u64 root_objectid, u64 owner,
6074                                      u64 offset, struct btrfs_key *ins)
6075 {
6076         int ret;
6077
6078         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6079
6080         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6081                                          ins->offset, 0,
6082                                          root_objectid, owner, offset,
6083                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6084         return ret;
6085 }
6086
6087 /*
6088  * this is used by the tree logging recovery code.  It records that
6089  * an extent has been allocated and makes sure to clear the free
6090  * space cache bits as well
6091  */
6092 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6093                                    struct btrfs_root *root,
6094                                    u64 root_objectid, u64 owner, u64 offset,
6095                                    struct btrfs_key *ins)
6096 {
6097         int ret;
6098         struct btrfs_block_group_cache *block_group;
6099         struct btrfs_caching_control *caching_ctl;
6100         u64 start = ins->objectid;
6101         u64 num_bytes = ins->offset;
6102
6103         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6104         cache_block_group(block_group, trans, NULL, 0);
6105         caching_ctl = get_caching_control(block_group);
6106
6107         if (!caching_ctl) {
6108                 BUG_ON(!block_group_cache_done(block_group));
6109                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6110                 BUG_ON(ret); /* -ENOMEM */
6111         } else {
6112                 mutex_lock(&caching_ctl->mutex);
6113
6114                 if (start >= caching_ctl->progress) {
6115                         ret = add_excluded_extent(root, start, num_bytes);
6116                         BUG_ON(ret); /* -ENOMEM */
6117                 } else if (start + num_bytes <= caching_ctl->progress) {
6118                         ret = btrfs_remove_free_space(block_group,
6119                                                       start, num_bytes);
6120                         BUG_ON(ret); /* -ENOMEM */
6121                 } else {
6122                         num_bytes = caching_ctl->progress - start;
6123                         ret = btrfs_remove_free_space(block_group,
6124                                                       start, num_bytes);
6125                         BUG_ON(ret); /* -ENOMEM */
6126
6127                         start = caching_ctl->progress;
6128                         num_bytes = ins->objectid + ins->offset -
6129                                     caching_ctl->progress;
6130                         ret = add_excluded_extent(root, start, num_bytes);
6131                         BUG_ON(ret); /* -ENOMEM */
6132                 }
6133
6134                 mutex_unlock(&caching_ctl->mutex);
6135                 put_caching_control(caching_ctl);
6136         }
6137
6138         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6139                                           RESERVE_ALLOC_NO_ACCOUNT);
6140         BUG_ON(ret); /* logic error */
6141         btrfs_put_block_group(block_group);
6142         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6143                                          0, owner, offset, ins, 1);
6144         return ret;
6145 }
6146
6147 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
6148                                             struct btrfs_root *root,
6149                                             u64 bytenr, u32 blocksize,
6150                                             int level)
6151 {
6152         struct extent_buffer *buf;
6153
6154         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6155         if (!buf)
6156                 return ERR_PTR(-ENOMEM);
6157         btrfs_set_header_generation(buf, trans->transid);
6158         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6159         btrfs_tree_lock(buf);
6160         clean_tree_block(trans, root, buf);
6161         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6162
6163         btrfs_set_lock_blocking(buf);
6164         btrfs_set_buffer_uptodate(buf);
6165
6166         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6167                 /*
6168                  * we allow two log transactions at a time, use different
6169                  * EXENT bit to differentiate dirty pages.
6170                  */
6171                 if (root->log_transid % 2 == 0)
6172                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6173                                         buf->start + buf->len - 1, GFP_NOFS);
6174                 else
6175                         set_extent_new(&root->dirty_log_pages, buf->start,
6176                                         buf->start + buf->len - 1, GFP_NOFS);
6177         } else {
6178                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6179                          buf->start + buf->len - 1, GFP_NOFS);
6180         }
6181         trans->blocks_used++;
6182         /* this returns a buffer locked for blocking */
6183         return buf;
6184 }
6185
6186 static struct btrfs_block_rsv *
6187 use_block_rsv(struct btrfs_trans_handle *trans,
6188               struct btrfs_root *root, u32 blocksize)
6189 {
6190         struct btrfs_block_rsv *block_rsv;
6191         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6192         int ret;
6193
6194         block_rsv = get_block_rsv(trans, root);
6195
6196         if (block_rsv->size == 0) {
6197                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
6198                 /*
6199                  * If we couldn't reserve metadata bytes try and use some from
6200                  * the global reserve.
6201                  */
6202                 if (ret && block_rsv != global_rsv) {
6203                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6204                         if (!ret)
6205                                 return global_rsv;
6206                         return ERR_PTR(ret);
6207                 } else if (ret) {
6208                         return ERR_PTR(ret);
6209                 }
6210                 return block_rsv;
6211         }
6212
6213         ret = block_rsv_use_bytes(block_rsv, blocksize);
6214         if (!ret)
6215                 return block_rsv;
6216         if (ret) {
6217                 static DEFINE_RATELIMIT_STATE(_rs,
6218                                 DEFAULT_RATELIMIT_INTERVAL,
6219                                 /*DEFAULT_RATELIMIT_BURST*/ 2);
6220                 if (__ratelimit(&_rs)) {
6221                         printk(KERN_DEBUG "btrfs: block rsv returned %d\n", ret);
6222                         WARN_ON(1);
6223                 }
6224                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
6225                 if (!ret) {
6226                         return block_rsv;
6227                 } else if (ret && block_rsv != global_rsv) {
6228                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6229                         if (!ret)
6230                                 return global_rsv;
6231                 }
6232         }
6233
6234         return ERR_PTR(-ENOSPC);
6235 }
6236
6237 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6238                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6239 {
6240         block_rsv_add_bytes(block_rsv, blocksize, 0);
6241         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6242 }
6243
6244 /*
6245  * finds a free extent and does all the dirty work required for allocation
6246  * returns the key for the extent through ins, and a tree buffer for
6247  * the first block of the extent through buf.
6248  *
6249  * returns the tree buffer or NULL.
6250  */
6251 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6252                                         struct btrfs_root *root, u32 blocksize,
6253                                         u64 parent, u64 root_objectid,
6254                                         struct btrfs_disk_key *key, int level,
6255                                         u64 hint, u64 empty_size, int for_cow)
6256 {
6257         struct btrfs_key ins;
6258         struct btrfs_block_rsv *block_rsv;
6259         struct extent_buffer *buf;
6260         u64 flags = 0;
6261         int ret;
6262
6263
6264         block_rsv = use_block_rsv(trans, root, blocksize);
6265         if (IS_ERR(block_rsv))
6266                 return ERR_CAST(block_rsv);
6267
6268         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6269                                    empty_size, hint, &ins, 0);
6270         if (ret) {
6271                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6272                 return ERR_PTR(ret);
6273         }
6274
6275         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6276                                     blocksize, level);
6277         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6278
6279         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6280                 if (parent == 0)
6281                         parent = ins.objectid;
6282                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6283         } else
6284                 BUG_ON(parent > 0);
6285
6286         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6287                 struct btrfs_delayed_extent_op *extent_op;
6288                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
6289                 BUG_ON(!extent_op); /* -ENOMEM */
6290                 if (key)
6291                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6292                 else
6293                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6294                 extent_op->flags_to_set = flags;
6295                 extent_op->update_key = 1;
6296                 extent_op->update_flags = 1;
6297                 extent_op->is_data = 0;
6298
6299                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6300                                         ins.objectid,
6301                                         ins.offset, parent, root_objectid,
6302                                         level, BTRFS_ADD_DELAYED_EXTENT,
6303                                         extent_op, for_cow);
6304                 BUG_ON(ret); /* -ENOMEM */
6305         }
6306         return buf;
6307 }
6308
6309 struct walk_control {
6310         u64 refs[BTRFS_MAX_LEVEL];
6311         u64 flags[BTRFS_MAX_LEVEL];
6312         struct btrfs_key update_progress;
6313         int stage;
6314         int level;
6315         int shared_level;
6316         int update_ref;
6317         int keep_locks;
6318         int reada_slot;
6319         int reada_count;
6320         int for_reloc;
6321 };
6322
6323 #define DROP_REFERENCE  1
6324 #define UPDATE_BACKREF  2
6325
6326 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6327                                      struct btrfs_root *root,
6328                                      struct walk_control *wc,
6329                                      struct btrfs_path *path)
6330 {
6331         u64 bytenr;
6332         u64 generation;
6333         u64 refs;
6334         u64 flags;
6335         u32 nritems;
6336         u32 blocksize;
6337         struct btrfs_key key;
6338         struct extent_buffer *eb;
6339         int ret;
6340         int slot;
6341         int nread = 0;
6342
6343         if (path->slots[wc->level] < wc->reada_slot) {
6344                 wc->reada_count = wc->reada_count * 2 / 3;
6345                 wc->reada_count = max(wc->reada_count, 2);
6346         } else {
6347                 wc->reada_count = wc->reada_count * 3 / 2;
6348                 wc->reada_count = min_t(int, wc->reada_count,
6349                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6350         }
6351
6352         eb = path->nodes[wc->level];
6353         nritems = btrfs_header_nritems(eb);
6354         blocksize = btrfs_level_size(root, wc->level - 1);
6355
6356         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6357                 if (nread >= wc->reada_count)
6358                         break;
6359
6360                 cond_resched();
6361                 bytenr = btrfs_node_blockptr(eb, slot);
6362                 generation = btrfs_node_ptr_generation(eb, slot);
6363
6364                 if (slot == path->slots[wc->level])
6365                         goto reada;
6366
6367                 if (wc->stage == UPDATE_BACKREF &&
6368                     generation <= root->root_key.offset)
6369                         continue;
6370
6371                 /* We don't lock the tree block, it's OK to be racy here */
6372                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6373                                                &refs, &flags);
6374                 /* We don't care about errors in readahead. */
6375                 if (ret < 0)
6376                         continue;
6377                 BUG_ON(refs == 0);
6378
6379                 if (wc->stage == DROP_REFERENCE) {
6380                         if (refs == 1)
6381                                 goto reada;
6382
6383                         if (wc->level == 1 &&
6384                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6385                                 continue;
6386                         if (!wc->update_ref ||
6387                             generation <= root->root_key.offset)
6388                                 continue;
6389                         btrfs_node_key_to_cpu(eb, &key, slot);
6390                         ret = btrfs_comp_cpu_keys(&key,
6391                                                   &wc->update_progress);
6392                         if (ret < 0)
6393                                 continue;
6394                 } else {
6395                         if (wc->level == 1 &&
6396                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6397                                 continue;
6398                 }
6399 reada:
6400                 ret = readahead_tree_block(root, bytenr, blocksize,
6401                                            generation);
6402                 if (ret)
6403                         break;
6404                 nread++;
6405         }
6406         wc->reada_slot = slot;
6407 }
6408
6409 /*
6410  * hepler to process tree block while walking down the tree.
6411  *
6412  * when wc->stage == UPDATE_BACKREF, this function updates
6413  * back refs for pointers in the block.
6414  *
6415  * NOTE: return value 1 means we should stop walking down.
6416  */
6417 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6418                                    struct btrfs_root *root,
6419                                    struct btrfs_path *path,
6420                                    struct walk_control *wc, int lookup_info)
6421 {
6422         int level = wc->level;
6423         struct extent_buffer *eb = path->nodes[level];
6424         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6425         int ret;
6426
6427         if (wc->stage == UPDATE_BACKREF &&
6428             btrfs_header_owner(eb) != root->root_key.objectid)
6429                 return 1;
6430
6431         /*
6432          * when reference count of tree block is 1, it won't increase
6433          * again. once full backref flag is set, we never clear it.
6434          */
6435         if (lookup_info &&
6436             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6437              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6438                 BUG_ON(!path->locks[level]);
6439                 ret = btrfs_lookup_extent_info(trans, root,
6440                                                eb->start, eb->len,
6441                                                &wc->refs[level],
6442                                                &wc->flags[level]);
6443                 BUG_ON(ret == -ENOMEM);
6444                 if (ret)
6445                         return ret;
6446                 BUG_ON(wc->refs[level] == 0);
6447         }
6448
6449         if (wc->stage == DROP_REFERENCE) {
6450                 if (wc->refs[level] > 1)
6451                         return 1;
6452
6453                 if (path->locks[level] && !wc->keep_locks) {
6454                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6455                         path->locks[level] = 0;
6456                 }
6457                 return 0;
6458         }
6459
6460         /* wc->stage == UPDATE_BACKREF */
6461         if (!(wc->flags[level] & flag)) {
6462                 BUG_ON(!path->locks[level]);
6463                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
6464                 BUG_ON(ret); /* -ENOMEM */
6465                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
6466                 BUG_ON(ret); /* -ENOMEM */
6467                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6468                                                   eb->len, flag, 0);
6469                 BUG_ON(ret); /* -ENOMEM */
6470                 wc->flags[level] |= flag;
6471         }
6472
6473         /*
6474          * the block is shared by multiple trees, so it's not good to
6475          * keep the tree lock
6476          */
6477         if (path->locks[level] && level > 0) {
6478                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6479                 path->locks[level] = 0;
6480         }
6481         return 0;
6482 }
6483
6484 /*
6485  * hepler to process tree block pointer.
6486  *
6487  * when wc->stage == DROP_REFERENCE, this function checks
6488  * reference count of the block pointed to. if the block
6489  * is shared and we need update back refs for the subtree
6490  * rooted at the block, this function changes wc->stage to
6491  * UPDATE_BACKREF. if the block is shared and there is no
6492  * need to update back, this function drops the reference
6493  * to the block.
6494  *
6495  * NOTE: return value 1 means we should stop walking down.
6496  */
6497 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6498                                  struct btrfs_root *root,
6499                                  struct btrfs_path *path,
6500                                  struct walk_control *wc, int *lookup_info)
6501 {
6502         u64 bytenr;
6503         u64 generation;
6504         u64 parent;
6505         u32 blocksize;
6506         struct btrfs_key key;
6507         struct extent_buffer *next;
6508         int level = wc->level;
6509         int reada = 0;
6510         int ret = 0;
6511
6512         generation = btrfs_node_ptr_generation(path->nodes[level],
6513                                                path->slots[level]);
6514         /*
6515          * if the lower level block was created before the snapshot
6516          * was created, we know there is no need to update back refs
6517          * for the subtree
6518          */
6519         if (wc->stage == UPDATE_BACKREF &&
6520             generation <= root->root_key.offset) {
6521                 *lookup_info = 1;
6522                 return 1;
6523         }
6524
6525         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6526         blocksize = btrfs_level_size(root, level - 1);
6527
6528         next = btrfs_find_tree_block(root, bytenr, blocksize);
6529         if (!next) {
6530                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6531                 if (!next)
6532                         return -ENOMEM;
6533                 reada = 1;
6534         }
6535         btrfs_tree_lock(next);
6536         btrfs_set_lock_blocking(next);
6537
6538         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6539                                        &wc->refs[level - 1],
6540                                        &wc->flags[level - 1]);
6541         if (ret < 0) {
6542                 btrfs_tree_unlock(next);
6543                 return ret;
6544         }
6545
6546         BUG_ON(wc->refs[level - 1] == 0);
6547         *lookup_info = 0;
6548
6549         if (wc->stage == DROP_REFERENCE) {
6550                 if (wc->refs[level - 1] > 1) {
6551                         if (level == 1 &&
6552                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6553                                 goto skip;
6554
6555                         if (!wc->update_ref ||
6556                             generation <= root->root_key.offset)
6557                                 goto skip;
6558
6559                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6560                                               path->slots[level]);
6561                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6562                         if (ret < 0)
6563                                 goto skip;
6564
6565                         wc->stage = UPDATE_BACKREF;
6566                         wc->shared_level = level - 1;
6567                 }
6568         } else {
6569                 if (level == 1 &&
6570                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6571                         goto skip;
6572         }
6573
6574         if (!btrfs_buffer_uptodate(next, generation)) {
6575                 btrfs_tree_unlock(next);
6576                 free_extent_buffer(next);
6577                 next = NULL;
6578                 *lookup_info = 1;
6579         }
6580
6581         if (!next) {
6582                 if (reada && level == 1)
6583                         reada_walk_down(trans, root, wc, path);
6584                 next = read_tree_block(root, bytenr, blocksize, generation);
6585                 if (!next)
6586                         return -EIO;
6587                 btrfs_tree_lock(next);
6588                 btrfs_set_lock_blocking(next);
6589         }
6590
6591         level--;
6592         BUG_ON(level != btrfs_header_level(next));
6593         path->nodes[level] = next;
6594         path->slots[level] = 0;
6595         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6596         wc->level = level;
6597         if (wc->level == 1)
6598                 wc->reada_slot = 0;
6599         return 0;
6600 skip:
6601         wc->refs[level - 1] = 0;
6602         wc->flags[level - 1] = 0;
6603         if (wc->stage == DROP_REFERENCE) {
6604                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6605                         parent = path->nodes[level]->start;
6606                 } else {
6607                         BUG_ON(root->root_key.objectid !=
6608                                btrfs_header_owner(path->nodes[level]));
6609                         parent = 0;
6610                 }
6611
6612                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6613                                 root->root_key.objectid, level - 1, 0, 0);
6614                 BUG_ON(ret); /* -ENOMEM */
6615         }
6616         btrfs_tree_unlock(next);
6617         free_extent_buffer(next);
6618         *lookup_info = 1;
6619         return 1;
6620 }
6621
6622 /*
6623  * hepler to process tree block while walking up the tree.
6624  *
6625  * when wc->stage == DROP_REFERENCE, this function drops
6626  * reference count on the block.
6627  *
6628  * when wc->stage == UPDATE_BACKREF, this function changes
6629  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6630  * to UPDATE_BACKREF previously while processing the block.
6631  *
6632  * NOTE: return value 1 means we should stop walking up.
6633  */
6634 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6635                                  struct btrfs_root *root,
6636                                  struct btrfs_path *path,
6637                                  struct walk_control *wc)
6638 {
6639         int ret;
6640         int level = wc->level;
6641         struct extent_buffer *eb = path->nodes[level];
6642         u64 parent = 0;
6643
6644         if (wc->stage == UPDATE_BACKREF) {
6645                 BUG_ON(wc->shared_level < level);
6646                 if (level < wc->shared_level)
6647                         goto out;
6648
6649                 ret = find_next_key(path, level + 1, &wc->update_progress);
6650                 if (ret > 0)
6651                         wc->update_ref = 0;
6652
6653                 wc->stage = DROP_REFERENCE;
6654                 wc->shared_level = -1;
6655                 path->slots[level] = 0;
6656
6657                 /*
6658                  * check reference count again if the block isn't locked.
6659                  * we should start walking down the tree again if reference
6660                  * count is one.
6661                  */
6662                 if (!path->locks[level]) {
6663                         BUG_ON(level == 0);
6664                         btrfs_tree_lock(eb);
6665                         btrfs_set_lock_blocking(eb);
6666                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6667
6668                         ret = btrfs_lookup_extent_info(trans, root,
6669                                                        eb->start, eb->len,
6670                                                        &wc->refs[level],
6671                                                        &wc->flags[level]);
6672                         if (ret < 0) {
6673                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6674                                 return ret;
6675                         }
6676                         BUG_ON(wc->refs[level] == 0);
6677                         if (wc->refs[level] == 1) {
6678                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6679                                 return 1;
6680                         }
6681                 }
6682         }
6683
6684         /* wc->stage == DROP_REFERENCE */
6685         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6686
6687         if (wc->refs[level] == 1) {
6688                 if (level == 0) {
6689                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6690                                 ret = btrfs_dec_ref(trans, root, eb, 1,
6691                                                     wc->for_reloc);
6692                         else
6693                                 ret = btrfs_dec_ref(trans, root, eb, 0,
6694                                                     wc->for_reloc);
6695                         BUG_ON(ret); /* -ENOMEM */
6696                 }
6697                 /* make block locked assertion in clean_tree_block happy */
6698                 if (!path->locks[level] &&
6699                     btrfs_header_generation(eb) == trans->transid) {
6700                         btrfs_tree_lock(eb);
6701                         btrfs_set_lock_blocking(eb);
6702                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6703                 }
6704                 clean_tree_block(trans, root, eb);
6705         }
6706
6707         if (eb == root->node) {
6708                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6709                         parent = eb->start;
6710                 else
6711                         BUG_ON(root->root_key.objectid !=
6712                                btrfs_header_owner(eb));
6713         } else {
6714                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6715                         parent = path->nodes[level + 1]->start;
6716                 else
6717                         BUG_ON(root->root_key.objectid !=
6718                                btrfs_header_owner(path->nodes[level + 1]));
6719         }
6720
6721         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1, 0);
6722 out:
6723         wc->refs[level] = 0;
6724         wc->flags[level] = 0;
6725         return 0;
6726 }
6727
6728 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6729                                    struct btrfs_root *root,
6730                                    struct btrfs_path *path,
6731                                    struct walk_control *wc)
6732 {
6733         int level = wc->level;
6734         int lookup_info = 1;
6735         int ret;
6736
6737         while (level >= 0) {
6738                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6739                 if (ret > 0)
6740                         break;
6741
6742                 if (level == 0)
6743                         break;
6744
6745                 if (path->slots[level] >=
6746                     btrfs_header_nritems(path->nodes[level]))
6747                         break;
6748
6749                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6750                 if (ret > 0) {
6751                         path->slots[level]++;
6752                         continue;
6753                 } else if (ret < 0)
6754                         return ret;
6755                 level = wc->level;
6756         }
6757         return 0;
6758 }
6759
6760 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6761                                  struct btrfs_root *root,
6762                                  struct btrfs_path *path,
6763                                  struct walk_control *wc, int max_level)
6764 {
6765         int level = wc->level;
6766         int ret;
6767
6768         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6769         while (level < max_level && path->nodes[level]) {
6770                 wc->level = level;
6771                 if (path->slots[level] + 1 <
6772                     btrfs_header_nritems(path->nodes[level])) {
6773                         path->slots[level]++;
6774                         return 0;
6775                 } else {
6776                         ret = walk_up_proc(trans, root, path, wc);
6777                         if (ret > 0)
6778                                 return 0;
6779
6780                         if (path->locks[level]) {
6781                                 btrfs_tree_unlock_rw(path->nodes[level],
6782                                                      path->locks[level]);
6783                                 path->locks[level] = 0;
6784                         }
6785                         free_extent_buffer(path->nodes[level]);
6786                         path->nodes[level] = NULL;
6787                         level++;
6788                 }
6789         }
6790         return 1;
6791 }
6792
6793 /*
6794  * drop a subvolume tree.
6795  *
6796  * this function traverses the tree freeing any blocks that only
6797  * referenced by the tree.
6798  *
6799  * when a shared tree block is found. this function decreases its
6800  * reference count by one. if update_ref is true, this function
6801  * also make sure backrefs for the shared block and all lower level
6802  * blocks are properly updated.
6803  */
6804 int btrfs_drop_snapshot(struct btrfs_root *root,
6805                          struct btrfs_block_rsv *block_rsv, int update_ref,
6806                          int for_reloc)
6807 {
6808         struct btrfs_path *path;
6809         struct btrfs_trans_handle *trans;
6810         struct btrfs_root *tree_root = root->fs_info->tree_root;
6811         struct btrfs_root_item *root_item = &root->root_item;
6812         struct walk_control *wc;
6813         struct btrfs_key key;
6814         int err = 0;
6815         int ret;
6816         int level;
6817
6818         path = btrfs_alloc_path();
6819         if (!path) {
6820                 err = -ENOMEM;
6821                 goto out;
6822         }
6823
6824         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6825         if (!wc) {
6826                 btrfs_free_path(path);
6827                 err = -ENOMEM;
6828                 goto out;
6829         }
6830
6831         trans = btrfs_start_transaction(tree_root, 0);
6832         if (IS_ERR(trans)) {
6833                 err = PTR_ERR(trans);
6834                 goto out_free;
6835         }
6836
6837         if (block_rsv)
6838                 trans->block_rsv = block_rsv;
6839
6840         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6841                 level = btrfs_header_level(root->node);
6842                 path->nodes[level] = btrfs_lock_root_node(root);
6843                 btrfs_set_lock_blocking(path->nodes[level]);
6844                 path->slots[level] = 0;
6845                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6846                 memset(&wc->update_progress, 0,
6847                        sizeof(wc->update_progress));
6848         } else {
6849                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6850                 memcpy(&wc->update_progress, &key,
6851                        sizeof(wc->update_progress));
6852
6853                 level = root_item->drop_level;
6854                 BUG_ON(level == 0);
6855                 path->lowest_level = level;
6856                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6857                 path->lowest_level = 0;
6858                 if (ret < 0) {
6859                         err = ret;
6860                         goto out_end_trans;
6861                 }
6862                 WARN_ON(ret > 0);
6863
6864                 /*
6865                  * unlock our path, this is safe because only this
6866                  * function is allowed to delete this snapshot
6867                  */
6868                 btrfs_unlock_up_safe(path, 0);
6869
6870                 level = btrfs_header_level(root->node);
6871                 while (1) {
6872                         btrfs_tree_lock(path->nodes[level]);
6873                         btrfs_set_lock_blocking(path->nodes[level]);
6874
6875                         ret = btrfs_lookup_extent_info(trans, root,
6876                                                 path->nodes[level]->start,
6877                                                 path->nodes[level]->len,
6878                                                 &wc->refs[level],
6879                                                 &wc->flags[level]);
6880                         if (ret < 0) {
6881                                 err = ret;
6882                                 goto out_end_trans;
6883                         }
6884                         BUG_ON(wc->refs[level] == 0);
6885
6886                         if (level == root_item->drop_level)
6887                                 break;
6888
6889                         btrfs_tree_unlock(path->nodes[level]);
6890                         WARN_ON(wc->refs[level] != 1);
6891                         level--;
6892                 }
6893         }
6894
6895         wc->level = level;
6896         wc->shared_level = -1;
6897         wc->stage = DROP_REFERENCE;
6898         wc->update_ref = update_ref;
6899         wc->keep_locks = 0;
6900         wc->for_reloc = for_reloc;
6901         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6902
6903         while (1) {
6904                 ret = walk_down_tree(trans, root, path, wc);
6905                 if (ret < 0) {
6906                         err = ret;
6907                         break;
6908                 }
6909
6910                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
6911                 if (ret < 0) {
6912                         err = ret;
6913                         break;
6914                 }
6915
6916                 if (ret > 0) {
6917                         BUG_ON(wc->stage != DROP_REFERENCE);
6918                         break;
6919                 }
6920
6921                 if (wc->stage == DROP_REFERENCE) {
6922                         level = wc->level;
6923                         btrfs_node_key(path->nodes[level],
6924                                        &root_item->drop_progress,
6925                                        path->slots[level]);
6926                         root_item->drop_level = level;
6927                 }
6928
6929                 BUG_ON(wc->level == 0);
6930                 if (btrfs_should_end_transaction(trans, tree_root)) {
6931                         ret = btrfs_update_root(trans, tree_root,
6932                                                 &root->root_key,
6933                                                 root_item);
6934                         if (ret) {
6935                                 btrfs_abort_transaction(trans, tree_root, ret);
6936                                 err = ret;
6937                                 goto out_end_trans;
6938                         }
6939
6940                         btrfs_end_transaction_throttle(trans, tree_root);
6941                         trans = btrfs_start_transaction(tree_root, 0);
6942                         if (IS_ERR(trans)) {
6943                                 err = PTR_ERR(trans);
6944                                 goto out_free;
6945                         }
6946                         if (block_rsv)
6947                                 trans->block_rsv = block_rsv;
6948                 }
6949         }
6950         btrfs_release_path(path);
6951         if (err)
6952                 goto out_end_trans;
6953
6954         ret = btrfs_del_root(trans, tree_root, &root->root_key);
6955         if (ret) {
6956                 btrfs_abort_transaction(trans, tree_root, ret);
6957                 goto out_end_trans;
6958         }
6959
6960         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
6961                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
6962                                            NULL, NULL);
6963                 if (ret < 0) {
6964                         btrfs_abort_transaction(trans, tree_root, ret);
6965                         err = ret;
6966                         goto out_end_trans;
6967                 } else if (ret > 0) {
6968                         /* if we fail to delete the orphan item this time
6969                          * around, it'll get picked up the next time.
6970                          *
6971                          * The most common failure here is just -ENOENT.
6972                          */
6973                         btrfs_del_orphan_item(trans, tree_root,
6974                                               root->root_key.objectid);
6975                 }
6976         }
6977
6978         if (root->in_radix) {
6979                 btrfs_free_fs_root(tree_root->fs_info, root);
6980         } else {
6981                 free_extent_buffer(root->node);
6982                 free_extent_buffer(root->commit_root);
6983                 kfree(root);
6984         }
6985 out_end_trans:
6986         btrfs_end_transaction_throttle(trans, tree_root);
6987 out_free:
6988         kfree(wc);
6989         btrfs_free_path(path);
6990 out:
6991         if (err)
6992                 btrfs_std_error(root->fs_info, err);
6993         return err;
6994 }
6995
6996 /*
6997  * drop subtree rooted at tree block 'node'.
6998  *
6999  * NOTE: this function will unlock and release tree block 'node'
7000  * only used by relocation code
7001  */
7002 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7003                         struct btrfs_root *root,
7004                         struct extent_buffer *node,
7005                         struct extent_buffer *parent)
7006 {
7007         struct btrfs_path *path;
7008         struct walk_control *wc;
7009         int level;
7010         int parent_level;
7011         int ret = 0;
7012         int wret;
7013
7014         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7015
7016         path = btrfs_alloc_path();
7017         if (!path)
7018                 return -ENOMEM;
7019
7020         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7021         if (!wc) {
7022                 btrfs_free_path(path);
7023                 return -ENOMEM;
7024         }
7025
7026         btrfs_assert_tree_locked(parent);
7027         parent_level = btrfs_header_level(parent);
7028         extent_buffer_get(parent);
7029         path->nodes[parent_level] = parent;
7030         path->slots[parent_level] = btrfs_header_nritems(parent);
7031
7032         btrfs_assert_tree_locked(node);
7033         level = btrfs_header_level(node);
7034         path->nodes[level] = node;
7035         path->slots[level] = 0;
7036         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7037
7038         wc->refs[parent_level] = 1;
7039         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7040         wc->level = level;
7041         wc->shared_level = -1;
7042         wc->stage = DROP_REFERENCE;
7043         wc->update_ref = 0;
7044         wc->keep_locks = 1;
7045         wc->for_reloc = 1;
7046         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7047
7048         while (1) {
7049                 wret = walk_down_tree(trans, root, path, wc);
7050                 if (wret < 0) {
7051                         ret = wret;
7052                         break;
7053                 }
7054
7055                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7056                 if (wret < 0)
7057                         ret = wret;
7058                 if (wret != 0)
7059                         break;
7060         }
7061
7062         kfree(wc);
7063         btrfs_free_path(path);
7064         return ret;
7065 }
7066
7067 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7068 {
7069         u64 num_devices;
7070         u64 stripped;
7071
7072         /*
7073          * if restripe for this chunk_type is on pick target profile and
7074          * return, otherwise do the usual balance
7075          */
7076         stripped = get_restripe_target(root->fs_info, flags);
7077         if (stripped)
7078                 return extended_to_chunk(stripped);
7079
7080         /*
7081          * we add in the count of missing devices because we want
7082          * to make sure that any RAID levels on a degraded FS
7083          * continue to be honored.
7084          */
7085         num_devices = root->fs_info->fs_devices->rw_devices +
7086                 root->fs_info->fs_devices->missing_devices;
7087
7088         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7089                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7090
7091         if (num_devices == 1) {
7092                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7093                 stripped = flags & ~stripped;
7094
7095                 /* turn raid0 into single device chunks */
7096                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7097                         return stripped;
7098
7099                 /* turn mirroring into duplication */
7100                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7101                              BTRFS_BLOCK_GROUP_RAID10))
7102                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7103         } else {
7104                 /* they already had raid on here, just return */
7105                 if (flags & stripped)
7106                         return flags;
7107
7108                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7109                 stripped = flags & ~stripped;
7110
7111                 /* switch duplicated blocks with raid1 */
7112                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7113                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7114
7115                 /* this is drive concat, leave it alone */
7116         }
7117
7118         return flags;
7119 }
7120
7121 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7122 {
7123         struct btrfs_space_info *sinfo = cache->space_info;
7124         u64 num_bytes;
7125         u64 min_allocable_bytes;
7126         int ret = -ENOSPC;
7127
7128
7129         /*
7130          * We need some metadata space and system metadata space for
7131          * allocating chunks in some corner cases until we force to set
7132          * it to be readonly.
7133          */
7134         if ((sinfo->flags &
7135              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7136             !force)
7137                 min_allocable_bytes = 1 * 1024 * 1024;
7138         else
7139                 min_allocable_bytes = 0;
7140
7141         spin_lock(&sinfo->lock);
7142         spin_lock(&cache->lock);
7143
7144         if (cache->ro) {
7145                 ret = 0;
7146                 goto out;
7147         }
7148
7149         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7150                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7151
7152         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7153             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7154             min_allocable_bytes <= sinfo->total_bytes) {
7155                 sinfo->bytes_readonly += num_bytes;
7156                 cache->ro = 1;
7157                 ret = 0;
7158         }
7159 out:
7160         spin_unlock(&cache->lock);
7161         spin_unlock(&sinfo->lock);
7162         return ret;
7163 }
7164
7165 int btrfs_set_block_group_ro(struct btrfs_root *root,
7166                              struct btrfs_block_group_cache *cache)
7167
7168 {
7169         struct btrfs_trans_handle *trans;
7170         u64 alloc_flags;
7171         int ret;
7172
7173         BUG_ON(cache->ro);
7174
7175         trans = btrfs_join_transaction(root);
7176         if (IS_ERR(trans))
7177                 return PTR_ERR(trans);
7178
7179         alloc_flags = update_block_group_flags(root, cache->flags);
7180         if (alloc_flags != cache->flags) {
7181                 ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7182                                      CHUNK_ALLOC_FORCE);
7183                 if (ret < 0)
7184                         goto out;
7185         }
7186
7187         ret = set_block_group_ro(cache, 0);
7188         if (!ret)
7189                 goto out;
7190         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7191         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7192                              CHUNK_ALLOC_FORCE);
7193         if (ret < 0)
7194                 goto out;
7195         ret = set_block_group_ro(cache, 0);
7196 out:
7197         btrfs_end_transaction(trans, root);
7198         return ret;
7199 }
7200
7201 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7202                             struct btrfs_root *root, u64 type)
7203 {
7204         u64 alloc_flags = get_alloc_profile(root, type);
7205         return do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7206                               CHUNK_ALLOC_FORCE);
7207 }
7208
7209 /*
7210  * helper to account the unused space of all the readonly block group in the
7211  * list. takes mirrors into account.
7212  */
7213 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7214 {
7215         struct btrfs_block_group_cache *block_group;
7216         u64 free_bytes = 0;
7217         int factor;
7218
7219         list_for_each_entry(block_group, groups_list, list) {
7220                 spin_lock(&block_group->lock);
7221
7222                 if (!block_group->ro) {
7223                         spin_unlock(&block_group->lock);
7224                         continue;
7225                 }
7226
7227                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7228                                           BTRFS_BLOCK_GROUP_RAID10 |
7229                                           BTRFS_BLOCK_GROUP_DUP))
7230                         factor = 2;
7231                 else
7232                         factor = 1;
7233
7234                 free_bytes += (block_group->key.offset -
7235                                btrfs_block_group_used(&block_group->item)) *
7236                                factor;
7237
7238                 spin_unlock(&block_group->lock);
7239         }
7240
7241         return free_bytes;
7242 }
7243
7244 /*
7245  * helper to account the unused space of all the readonly block group in the
7246  * space_info. takes mirrors into account.
7247  */
7248 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7249 {
7250         int i;
7251         u64 free_bytes = 0;
7252
7253         spin_lock(&sinfo->lock);
7254
7255         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7256                 if (!list_empty(&sinfo->block_groups[i]))
7257                         free_bytes += __btrfs_get_ro_block_group_free_space(
7258                                                 &sinfo->block_groups[i]);
7259
7260         spin_unlock(&sinfo->lock);
7261
7262         return free_bytes;
7263 }
7264
7265 void btrfs_set_block_group_rw(struct btrfs_root *root,
7266                               struct btrfs_block_group_cache *cache)
7267 {
7268         struct btrfs_space_info *sinfo = cache->space_info;
7269         u64 num_bytes;
7270
7271         BUG_ON(!cache->ro);
7272
7273         spin_lock(&sinfo->lock);
7274         spin_lock(&cache->lock);
7275         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7276                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7277         sinfo->bytes_readonly -= num_bytes;
7278         cache->ro = 0;
7279         spin_unlock(&cache->lock);
7280         spin_unlock(&sinfo->lock);
7281 }
7282
7283 /*
7284  * checks to see if its even possible to relocate this block group.
7285  *
7286  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7287  * ok to go ahead and try.
7288  */
7289 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7290 {
7291         struct btrfs_block_group_cache *block_group;
7292         struct btrfs_space_info *space_info;
7293         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7294         struct btrfs_device *device;
7295         u64 min_free;
7296         u64 dev_min = 1;
7297         u64 dev_nr = 0;
7298         u64 target;
7299         int index;
7300         int full = 0;
7301         int ret = 0;
7302
7303         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7304
7305         /* odd, couldn't find the block group, leave it alone */
7306         if (!block_group)
7307                 return -1;
7308
7309         min_free = btrfs_block_group_used(&block_group->item);
7310
7311         /* no bytes used, we're good */
7312         if (!min_free)
7313                 goto out;
7314
7315         space_info = block_group->space_info;
7316         spin_lock(&space_info->lock);
7317
7318         full = space_info->full;
7319
7320         /*
7321          * if this is the last block group we have in this space, we can't
7322          * relocate it unless we're able to allocate a new chunk below.
7323          *
7324          * Otherwise, we need to make sure we have room in the space to handle
7325          * all of the extents from this block group.  If we can, we're good
7326          */
7327         if ((space_info->total_bytes != block_group->key.offset) &&
7328             (space_info->bytes_used + space_info->bytes_reserved +
7329              space_info->bytes_pinned + space_info->bytes_readonly +
7330              min_free < space_info->total_bytes)) {
7331                 spin_unlock(&space_info->lock);
7332                 goto out;
7333         }
7334         spin_unlock(&space_info->lock);
7335
7336         /*
7337          * ok we don't have enough space, but maybe we have free space on our
7338          * devices to allocate new chunks for relocation, so loop through our
7339          * alloc devices and guess if we have enough space.  if this block
7340          * group is going to be restriped, run checks against the target
7341          * profile instead of the current one.
7342          */
7343         ret = -1;
7344
7345         /*
7346          * index:
7347          *      0: raid10
7348          *      1: raid1
7349          *      2: dup
7350          *      3: raid0
7351          *      4: single
7352          */
7353         target = get_restripe_target(root->fs_info, block_group->flags);
7354         if (target) {
7355                 index = __get_block_group_index(extended_to_chunk(target));
7356         } else {
7357                 /*
7358                  * this is just a balance, so if we were marked as full
7359                  * we know there is no space for a new chunk
7360                  */
7361                 if (full)
7362                         goto out;
7363
7364                 index = get_block_group_index(block_group);
7365         }
7366
7367         if (index == 0) {
7368                 dev_min = 4;
7369                 /* Divide by 2 */
7370                 min_free >>= 1;
7371         } else if (index == 1) {
7372                 dev_min = 2;
7373         } else if (index == 2) {
7374                 /* Multiply by 2 */
7375                 min_free <<= 1;
7376         } else if (index == 3) {
7377                 dev_min = fs_devices->rw_devices;
7378                 do_div(min_free, dev_min);
7379         }
7380
7381         mutex_lock(&root->fs_info->chunk_mutex);
7382         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7383                 u64 dev_offset;
7384
7385                 /*
7386                  * check to make sure we can actually find a chunk with enough
7387                  * space to fit our block group in.
7388                  */
7389                 if (device->total_bytes > device->bytes_used + min_free) {
7390                         ret = find_free_dev_extent(device, min_free,
7391                                                    &dev_offset, NULL);
7392                         if (!ret)
7393                                 dev_nr++;
7394
7395                         if (dev_nr >= dev_min)
7396                                 break;
7397
7398                         ret = -1;
7399                 }
7400         }
7401         mutex_unlock(&root->fs_info->chunk_mutex);
7402 out:
7403         btrfs_put_block_group(block_group);
7404         return ret;
7405 }
7406
7407 static int find_first_block_group(struct btrfs_root *root,
7408                 struct btrfs_path *path, struct btrfs_key *key)
7409 {
7410         int ret = 0;
7411         struct btrfs_key found_key;
7412         struct extent_buffer *leaf;
7413         int slot;
7414
7415         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7416         if (ret < 0)
7417                 goto out;
7418
7419         while (1) {
7420                 slot = path->slots[0];
7421                 leaf = path->nodes[0];
7422                 if (slot >= btrfs_header_nritems(leaf)) {
7423                         ret = btrfs_next_leaf(root, path);
7424                         if (ret == 0)
7425                                 continue;
7426                         if (ret < 0)
7427                                 goto out;
7428                         break;
7429                 }
7430                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7431
7432                 if (found_key.objectid >= key->objectid &&
7433                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7434                         ret = 0;
7435                         goto out;
7436                 }
7437                 path->slots[0]++;
7438         }
7439 out:
7440         return ret;
7441 }
7442
7443 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7444 {
7445         struct btrfs_block_group_cache *block_group;
7446         u64 last = 0;
7447
7448         while (1) {
7449                 struct inode *inode;
7450
7451                 block_group = btrfs_lookup_first_block_group(info, last);
7452                 while (block_group) {
7453                         spin_lock(&block_group->lock);
7454                         if (block_group->iref)
7455                                 break;
7456                         spin_unlock(&block_group->lock);
7457                         block_group = next_block_group(info->tree_root,
7458                                                        block_group);
7459                 }
7460                 if (!block_group) {
7461                         if (last == 0)
7462                                 break;
7463                         last = 0;
7464                         continue;
7465                 }
7466
7467                 inode = block_group->inode;
7468                 block_group->iref = 0;
7469                 block_group->inode = NULL;
7470                 spin_unlock(&block_group->lock);
7471                 iput(inode);
7472                 last = block_group->key.objectid + block_group->key.offset;
7473                 btrfs_put_block_group(block_group);
7474         }
7475 }
7476
7477 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7478 {
7479         struct btrfs_block_group_cache *block_group;
7480         struct btrfs_space_info *space_info;
7481         struct btrfs_caching_control *caching_ctl;
7482         struct rb_node *n;
7483
7484         down_write(&info->extent_commit_sem);
7485         while (!list_empty(&info->caching_block_groups)) {
7486                 caching_ctl = list_entry(info->caching_block_groups.next,
7487                                          struct btrfs_caching_control, list);
7488                 list_del(&caching_ctl->list);
7489                 put_caching_control(caching_ctl);
7490         }
7491         up_write(&info->extent_commit_sem);
7492
7493         spin_lock(&info->block_group_cache_lock);
7494         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7495                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7496                                        cache_node);
7497                 rb_erase(&block_group->cache_node,
7498                          &info->block_group_cache_tree);
7499                 spin_unlock(&info->block_group_cache_lock);
7500
7501                 down_write(&block_group->space_info->groups_sem);
7502                 list_del(&block_group->list);
7503                 up_write(&block_group->space_info->groups_sem);
7504
7505                 if (block_group->cached == BTRFS_CACHE_STARTED)
7506                         wait_block_group_cache_done(block_group);
7507
7508                 /*
7509                  * We haven't cached this block group, which means we could
7510                  * possibly have excluded extents on this block group.
7511                  */
7512                 if (block_group->cached == BTRFS_CACHE_NO)
7513                         free_excluded_extents(info->extent_root, block_group);
7514
7515                 btrfs_remove_free_space_cache(block_group);
7516                 btrfs_put_block_group(block_group);
7517
7518                 spin_lock(&info->block_group_cache_lock);
7519         }
7520         spin_unlock(&info->block_group_cache_lock);
7521
7522         /* now that all the block groups are freed, go through and
7523          * free all the space_info structs.  This is only called during
7524          * the final stages of unmount, and so we know nobody is
7525          * using them.  We call synchronize_rcu() once before we start,
7526          * just to be on the safe side.
7527          */
7528         synchronize_rcu();
7529
7530         release_global_block_rsv(info);
7531
7532         while(!list_empty(&info->space_info)) {
7533                 space_info = list_entry(info->space_info.next,
7534                                         struct btrfs_space_info,
7535                                         list);
7536                 if (space_info->bytes_pinned > 0 ||
7537                     space_info->bytes_reserved > 0 ||
7538                     space_info->bytes_may_use > 0) {
7539                         WARN_ON(1);
7540                         dump_space_info(space_info, 0, 0);
7541                 }
7542                 list_del(&space_info->list);
7543                 kfree(space_info);
7544         }
7545         return 0;
7546 }
7547
7548 static void __link_block_group(struct btrfs_space_info *space_info,
7549                                struct btrfs_block_group_cache *cache)
7550 {
7551         int index = get_block_group_index(cache);
7552
7553         down_write(&space_info->groups_sem);
7554         list_add_tail(&cache->list, &space_info->block_groups[index]);
7555         up_write(&space_info->groups_sem);
7556 }
7557
7558 int btrfs_read_block_groups(struct btrfs_root *root)
7559 {
7560         struct btrfs_path *path;
7561         int ret;
7562         struct btrfs_block_group_cache *cache;
7563         struct btrfs_fs_info *info = root->fs_info;
7564         struct btrfs_space_info *space_info;
7565         struct btrfs_key key;
7566         struct btrfs_key found_key;
7567         struct extent_buffer *leaf;
7568         int need_clear = 0;
7569         u64 cache_gen;
7570
7571         root = info->extent_root;
7572         key.objectid = 0;
7573         key.offset = 0;
7574         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7575         path = btrfs_alloc_path();
7576         if (!path)
7577                 return -ENOMEM;
7578         path->reada = 1;
7579
7580         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7581         if (btrfs_test_opt(root, SPACE_CACHE) &&
7582             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7583                 need_clear = 1;
7584         if (btrfs_test_opt(root, CLEAR_CACHE))
7585                 need_clear = 1;
7586
7587         while (1) {
7588                 ret = find_first_block_group(root, path, &key);
7589                 if (ret > 0)
7590                         break;
7591                 if (ret != 0)
7592                         goto error;
7593                 leaf = path->nodes[0];
7594                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7595                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7596                 if (!cache) {
7597                         ret = -ENOMEM;
7598                         goto error;
7599                 }
7600                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7601                                                 GFP_NOFS);
7602                 if (!cache->free_space_ctl) {
7603                         kfree(cache);
7604                         ret = -ENOMEM;
7605                         goto error;
7606                 }
7607
7608                 atomic_set(&cache->count, 1);
7609                 spin_lock_init(&cache->lock);
7610                 cache->fs_info = info;
7611                 INIT_LIST_HEAD(&cache->list);
7612                 INIT_LIST_HEAD(&cache->cluster_list);
7613
7614                 if (need_clear)
7615                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7616
7617                 read_extent_buffer(leaf, &cache->item,
7618                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7619                                    sizeof(cache->item));
7620                 memcpy(&cache->key, &found_key, sizeof(found_key));
7621
7622                 key.objectid = found_key.objectid + found_key.offset;
7623                 btrfs_release_path(path);
7624                 cache->flags = btrfs_block_group_flags(&cache->item);
7625                 cache->sectorsize = root->sectorsize;
7626
7627                 btrfs_init_free_space_ctl(cache);
7628
7629                 /*
7630                  * We need to exclude the super stripes now so that the space
7631                  * info has super bytes accounted for, otherwise we'll think
7632                  * we have more space than we actually do.
7633                  */
7634                 exclude_super_stripes(root, cache);
7635
7636                 /*
7637                  * check for two cases, either we are full, and therefore
7638                  * don't need to bother with the caching work since we won't
7639                  * find any space, or we are empty, and we can just add all
7640                  * the space in and be done with it.  This saves us _alot_ of
7641                  * time, particularly in the full case.
7642                  */
7643                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7644                         cache->last_byte_to_unpin = (u64)-1;
7645                         cache->cached = BTRFS_CACHE_FINISHED;
7646                         free_excluded_extents(root, cache);
7647                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7648                         cache->last_byte_to_unpin = (u64)-1;
7649                         cache->cached = BTRFS_CACHE_FINISHED;
7650                         add_new_free_space(cache, root->fs_info,
7651                                            found_key.objectid,
7652                                            found_key.objectid +
7653                                            found_key.offset);
7654                         free_excluded_extents(root, cache);
7655                 }
7656
7657                 ret = update_space_info(info, cache->flags, found_key.offset,
7658                                         btrfs_block_group_used(&cache->item),
7659                                         &space_info);
7660                 BUG_ON(ret); /* -ENOMEM */
7661                 cache->space_info = space_info;
7662                 spin_lock(&cache->space_info->lock);
7663                 cache->space_info->bytes_readonly += cache->bytes_super;
7664                 spin_unlock(&cache->space_info->lock);
7665
7666                 __link_block_group(space_info, cache);
7667
7668                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7669                 BUG_ON(ret); /* Logic error */
7670
7671                 set_avail_alloc_bits(root->fs_info, cache->flags);
7672                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7673                         set_block_group_ro(cache, 1);
7674         }
7675
7676         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7677                 if (!(get_alloc_profile(root, space_info->flags) &
7678                       (BTRFS_BLOCK_GROUP_RAID10 |
7679                        BTRFS_BLOCK_GROUP_RAID1 |
7680                        BTRFS_BLOCK_GROUP_DUP)))
7681                         continue;
7682                 /*
7683                  * avoid allocating from un-mirrored block group if there are
7684                  * mirrored block groups.
7685                  */
7686                 list_for_each_entry(cache, &space_info->block_groups[3], list)
7687                         set_block_group_ro(cache, 1);
7688                 list_for_each_entry(cache, &space_info->block_groups[4], list)
7689                         set_block_group_ro(cache, 1);
7690         }
7691
7692         init_global_block_rsv(info);
7693         ret = 0;
7694 error:
7695         btrfs_free_path(path);
7696         return ret;
7697 }
7698
7699 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7700                            struct btrfs_root *root, u64 bytes_used,
7701                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7702                            u64 size)
7703 {
7704         int ret;
7705         struct btrfs_root *extent_root;
7706         struct btrfs_block_group_cache *cache;
7707
7708         extent_root = root->fs_info->extent_root;
7709
7710         root->fs_info->last_trans_log_full_commit = trans->transid;
7711
7712         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7713         if (!cache)
7714                 return -ENOMEM;
7715         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7716                                         GFP_NOFS);
7717         if (!cache->free_space_ctl) {
7718                 kfree(cache);
7719                 return -ENOMEM;
7720         }
7721
7722         cache->key.objectid = chunk_offset;
7723         cache->key.offset = size;
7724         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7725         cache->sectorsize = root->sectorsize;
7726         cache->fs_info = root->fs_info;
7727
7728         atomic_set(&cache->count, 1);
7729         spin_lock_init(&cache->lock);
7730         INIT_LIST_HEAD(&cache->list);
7731         INIT_LIST_HEAD(&cache->cluster_list);
7732
7733         btrfs_init_free_space_ctl(cache);
7734
7735         btrfs_set_block_group_used(&cache->item, bytes_used);
7736         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7737         cache->flags = type;
7738         btrfs_set_block_group_flags(&cache->item, type);
7739
7740         cache->last_byte_to_unpin = (u64)-1;
7741         cache->cached = BTRFS_CACHE_FINISHED;
7742         exclude_super_stripes(root, cache);
7743
7744         add_new_free_space(cache, root->fs_info, chunk_offset,
7745                            chunk_offset + size);
7746
7747         free_excluded_extents(root, cache);
7748
7749         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7750                                 &cache->space_info);
7751         BUG_ON(ret); /* -ENOMEM */
7752         update_global_block_rsv(root->fs_info);
7753
7754         spin_lock(&cache->space_info->lock);
7755         cache->space_info->bytes_readonly += cache->bytes_super;
7756         spin_unlock(&cache->space_info->lock);
7757
7758         __link_block_group(cache->space_info, cache);
7759
7760         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7761         BUG_ON(ret); /* Logic error */
7762
7763         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
7764                                 sizeof(cache->item));
7765         if (ret) {
7766                 btrfs_abort_transaction(trans, extent_root, ret);
7767                 return ret;
7768         }
7769
7770         set_avail_alloc_bits(extent_root->fs_info, type);
7771
7772         return 0;
7773 }
7774
7775 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
7776 {
7777         u64 extra_flags = chunk_to_extended(flags) &
7778                                 BTRFS_EXTENDED_PROFILE_MASK;
7779
7780         if (flags & BTRFS_BLOCK_GROUP_DATA)
7781                 fs_info->avail_data_alloc_bits &= ~extra_flags;
7782         if (flags & BTRFS_BLOCK_GROUP_METADATA)
7783                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
7784         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
7785                 fs_info->avail_system_alloc_bits &= ~extra_flags;
7786 }
7787
7788 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7789                              struct btrfs_root *root, u64 group_start)
7790 {
7791         struct btrfs_path *path;
7792         struct btrfs_block_group_cache *block_group;
7793         struct btrfs_free_cluster *cluster;
7794         struct btrfs_root *tree_root = root->fs_info->tree_root;
7795         struct btrfs_key key;
7796         struct inode *inode;
7797         int ret;
7798         int index;
7799         int factor;
7800
7801         root = root->fs_info->extent_root;
7802
7803         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7804         BUG_ON(!block_group);
7805         BUG_ON(!block_group->ro);
7806
7807         /*
7808          * Free the reserved super bytes from this block group before
7809          * remove it.
7810          */
7811         free_excluded_extents(root, block_group);
7812
7813         memcpy(&key, &block_group->key, sizeof(key));
7814         index = get_block_group_index(block_group);
7815         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7816                                   BTRFS_BLOCK_GROUP_RAID1 |
7817                                   BTRFS_BLOCK_GROUP_RAID10))
7818                 factor = 2;
7819         else
7820                 factor = 1;
7821
7822         /* make sure this block group isn't part of an allocation cluster */
7823         cluster = &root->fs_info->data_alloc_cluster;
7824         spin_lock(&cluster->refill_lock);
7825         btrfs_return_cluster_to_free_space(block_group, cluster);
7826         spin_unlock(&cluster->refill_lock);
7827
7828         /*
7829          * make sure this block group isn't part of a metadata
7830          * allocation cluster
7831          */
7832         cluster = &root->fs_info->meta_alloc_cluster;
7833         spin_lock(&cluster->refill_lock);
7834         btrfs_return_cluster_to_free_space(block_group, cluster);
7835         spin_unlock(&cluster->refill_lock);
7836
7837         path = btrfs_alloc_path();
7838         if (!path) {
7839                 ret = -ENOMEM;
7840                 goto out;
7841         }
7842
7843         inode = lookup_free_space_inode(tree_root, block_group, path);
7844         if (!IS_ERR(inode)) {
7845                 ret = btrfs_orphan_add(trans, inode);
7846                 if (ret) {
7847                         btrfs_add_delayed_iput(inode);
7848                         goto out;
7849                 }
7850                 clear_nlink(inode);
7851                 /* One for the block groups ref */
7852                 spin_lock(&block_group->lock);
7853                 if (block_group->iref) {
7854                         block_group->iref = 0;
7855                         block_group->inode = NULL;
7856                         spin_unlock(&block_group->lock);
7857                         iput(inode);
7858                 } else {
7859                         spin_unlock(&block_group->lock);
7860                 }
7861                 /* One for our lookup ref */
7862                 btrfs_add_delayed_iput(inode);
7863         }
7864
7865         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
7866         key.offset = block_group->key.objectid;
7867         key.type = 0;
7868
7869         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
7870         if (ret < 0)
7871                 goto out;
7872         if (ret > 0)
7873                 btrfs_release_path(path);
7874         if (ret == 0) {
7875                 ret = btrfs_del_item(trans, tree_root, path);
7876                 if (ret)
7877                         goto out;
7878                 btrfs_release_path(path);
7879         }
7880
7881         spin_lock(&root->fs_info->block_group_cache_lock);
7882         rb_erase(&block_group->cache_node,
7883                  &root->fs_info->block_group_cache_tree);
7884         spin_unlock(&root->fs_info->block_group_cache_lock);
7885
7886         down_write(&block_group->space_info->groups_sem);
7887         /*
7888          * we must use list_del_init so people can check to see if they
7889          * are still on the list after taking the semaphore
7890          */
7891         list_del_init(&block_group->list);
7892         if (list_empty(&block_group->space_info->block_groups[index]))
7893                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
7894         up_write(&block_group->space_info->groups_sem);
7895
7896         if (block_group->cached == BTRFS_CACHE_STARTED)
7897                 wait_block_group_cache_done(block_group);
7898
7899         btrfs_remove_free_space_cache(block_group);
7900
7901         spin_lock(&block_group->space_info->lock);
7902         block_group->space_info->total_bytes -= block_group->key.offset;
7903         block_group->space_info->bytes_readonly -= block_group->key.offset;
7904         block_group->space_info->disk_total -= block_group->key.offset * factor;
7905         spin_unlock(&block_group->space_info->lock);
7906
7907         memcpy(&key, &block_group->key, sizeof(key));
7908
7909         btrfs_clear_space_info_full(root->fs_info);
7910
7911         btrfs_put_block_group(block_group);
7912         btrfs_put_block_group(block_group);
7913
7914         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
7915         if (ret > 0)
7916                 ret = -EIO;
7917         if (ret < 0)
7918                 goto out;
7919
7920         ret = btrfs_del_item(trans, root, path);
7921 out:
7922         btrfs_free_path(path);
7923         return ret;
7924 }
7925
7926 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
7927 {
7928         struct btrfs_space_info *space_info;
7929         struct btrfs_super_block *disk_super;
7930         u64 features;
7931         u64 flags;
7932         int mixed = 0;
7933         int ret;
7934
7935         disk_super = fs_info->super_copy;
7936         if (!btrfs_super_root(disk_super))
7937                 return 1;
7938
7939         features = btrfs_super_incompat_flags(disk_super);
7940         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
7941                 mixed = 1;
7942
7943         flags = BTRFS_BLOCK_GROUP_SYSTEM;
7944         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7945         if (ret)
7946                 goto out;
7947
7948         if (mixed) {
7949                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
7950                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7951         } else {
7952                 flags = BTRFS_BLOCK_GROUP_METADATA;
7953                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7954                 if (ret)
7955                         goto out;
7956
7957                 flags = BTRFS_BLOCK_GROUP_DATA;
7958                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7959         }
7960 out:
7961         return ret;
7962 }
7963
7964 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
7965 {
7966         return unpin_extent_range(root, start, end);
7967 }
7968
7969 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
7970                                u64 num_bytes, u64 *actual_bytes)
7971 {
7972         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
7973 }
7974
7975 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
7976 {
7977         struct btrfs_fs_info *fs_info = root->fs_info;
7978         struct btrfs_block_group_cache *cache = NULL;
7979         u64 group_trimmed;
7980         u64 start;
7981         u64 end;
7982         u64 trimmed = 0;
7983         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
7984         int ret = 0;
7985
7986         /*
7987          * try to trim all FS space, our block group may start from non-zero.
7988          */
7989         if (range->len == total_bytes)
7990                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
7991         else
7992                 cache = btrfs_lookup_block_group(fs_info, range->start);
7993
7994         while (cache) {
7995                 if (cache->key.objectid >= (range->start + range->len)) {
7996                         btrfs_put_block_group(cache);
7997                         break;
7998                 }
7999
8000                 start = max(range->start, cache->key.objectid);
8001                 end = min(range->start + range->len,
8002                                 cache->key.objectid + cache->key.offset);
8003
8004                 if (end - start >= range->minlen) {
8005                         if (!block_group_cache_done(cache)) {
8006                                 ret = cache_block_group(cache, NULL, root, 0);
8007                                 if (!ret)
8008                                         wait_block_group_cache_done(cache);
8009                         }
8010                         ret = btrfs_trim_block_group(cache,
8011                                                      &group_trimmed,
8012                                                      start,
8013                                                      end,
8014                                                      range->minlen);
8015
8016                         trimmed += group_trimmed;
8017                         if (ret) {
8018                                 btrfs_put_block_group(cache);
8019                                 break;
8020                         }
8021                 }
8022
8023                 cache = next_block_group(fs_info->tree_root, cache);
8024         }
8025
8026         range->len = trimmed;
8027         return ret;
8028 }