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 static void wait_for_more_refs(struct btrfs_delayed_ref_root *delayed_refs,
2351                                unsigned long num_refs,
2352                                struct list_head *first_seq)
2353 {
2354         spin_unlock(&delayed_refs->lock);
2355         pr_debug("waiting for more refs (num %ld, first %p)\n",
2356                  num_refs, first_seq);
2357         wait_event(delayed_refs->seq_wait,
2358                    num_refs != delayed_refs->num_entries ||
2359                    delayed_refs->seq_head.next != first_seq);
2360         pr_debug("done waiting for more refs (num %ld, first %p)\n",
2361                  delayed_refs->num_entries, delayed_refs->seq_head.next);
2362         spin_lock(&delayed_refs->lock);
2363 }
2364
2365 /*
2366  * this starts processing the delayed reference count updates and
2367  * extent insertions we have queued up so far.  count can be
2368  * 0, which means to process everything in the tree at the start
2369  * of the run (but not newly added entries), or it can be some target
2370  * number you'd like to process.
2371  *
2372  * Returns 0 on success or if called with an aborted transaction
2373  * Returns <0 on error and aborts the transaction
2374  */
2375 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2376                            struct btrfs_root *root, unsigned long count)
2377 {
2378         struct rb_node *node;
2379         struct btrfs_delayed_ref_root *delayed_refs;
2380         struct btrfs_delayed_ref_node *ref;
2381         struct list_head cluster;
2382         struct list_head *first_seq = NULL;
2383         int ret;
2384         u64 delayed_start;
2385         int run_all = count == (unsigned long)-1;
2386         int run_most = 0;
2387         unsigned long num_refs = 0;
2388         int consider_waiting;
2389
2390         /* We'll clean this up in btrfs_cleanup_transaction */
2391         if (trans->aborted)
2392                 return 0;
2393
2394         if (root == root->fs_info->extent_root)
2395                 root = root->fs_info->tree_root;
2396
2397         do_chunk_alloc(trans, root->fs_info->extent_root,
2398                        2 * 1024 * 1024, btrfs_get_alloc_profile(root, 0),
2399                        CHUNK_ALLOC_NO_FORCE);
2400
2401         delayed_refs = &trans->transaction->delayed_refs;
2402         INIT_LIST_HEAD(&cluster);
2403 again:
2404         consider_waiting = 0;
2405         spin_lock(&delayed_refs->lock);
2406         if (count == 0) {
2407                 count = delayed_refs->num_entries * 2;
2408                 run_most = 1;
2409         }
2410         while (1) {
2411                 if (!(run_all || run_most) &&
2412                     delayed_refs->num_heads_ready < 64)
2413                         break;
2414
2415                 /*
2416                  * go find something we can process in the rbtree.  We start at
2417                  * the beginning of the tree, and then build a cluster
2418                  * of refs to process starting at the first one we are able to
2419                  * lock
2420                  */
2421                 delayed_start = delayed_refs->run_delayed_start;
2422                 ret = btrfs_find_ref_cluster(trans, &cluster,
2423                                              delayed_refs->run_delayed_start);
2424                 if (ret)
2425                         break;
2426
2427                 if (delayed_start >= delayed_refs->run_delayed_start) {
2428                         if (consider_waiting == 0) {
2429                                 /*
2430                                  * btrfs_find_ref_cluster looped. let's do one
2431                                  * more cycle. if we don't run any delayed ref
2432                                  * during that cycle (because we can't because
2433                                  * all of them are blocked) and if the number of
2434                                  * refs doesn't change, we avoid busy waiting.
2435                                  */
2436                                 consider_waiting = 1;
2437                                 num_refs = delayed_refs->num_entries;
2438                                 first_seq = root->fs_info->tree_mod_seq_list.next;
2439                         } else {
2440                                 wait_for_more_refs(delayed_refs,
2441                                                    num_refs, first_seq);
2442                                 /*
2443                                  * after waiting, things have changed. we
2444                                  * dropped the lock and someone else might have
2445                                  * run some refs, built new clusters and so on.
2446                                  * therefore, we restart staleness detection.
2447                                  */
2448                                 consider_waiting = 0;
2449                         }
2450                 }
2451
2452                 ret = run_clustered_refs(trans, root, &cluster);
2453                 if (ret < 0) {
2454                         spin_unlock(&delayed_refs->lock);
2455                         btrfs_abort_transaction(trans, root, ret);
2456                         return ret;
2457                 }
2458
2459                 count -= min_t(unsigned long, ret, count);
2460
2461                 if (count == 0)
2462                         break;
2463
2464                 if (ret || delayed_refs->run_delayed_start == 0) {
2465                         /* refs were run, let's reset staleness detection */
2466                         consider_waiting = 0;
2467                 }
2468         }
2469
2470         if (run_all) {
2471                 node = rb_first(&delayed_refs->root);
2472                 if (!node)
2473                         goto out;
2474                 count = (unsigned long)-1;
2475
2476                 while (node) {
2477                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2478                                        rb_node);
2479                         if (btrfs_delayed_ref_is_head(ref)) {
2480                                 struct btrfs_delayed_ref_head *head;
2481
2482                                 head = btrfs_delayed_node_to_head(ref);
2483                                 atomic_inc(&ref->refs);
2484
2485                                 spin_unlock(&delayed_refs->lock);
2486                                 /*
2487                                  * Mutex was contended, block until it's
2488                                  * released and try again
2489                                  */
2490                                 mutex_lock(&head->mutex);
2491                                 mutex_unlock(&head->mutex);
2492
2493                                 btrfs_put_delayed_ref(ref);
2494                                 cond_resched();
2495                                 goto again;
2496                         }
2497                         node = rb_next(node);
2498                 }
2499                 spin_unlock(&delayed_refs->lock);
2500                 schedule_timeout(1);
2501                 goto again;
2502         }
2503 out:
2504         spin_unlock(&delayed_refs->lock);
2505         return 0;
2506 }
2507
2508 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2509                                 struct btrfs_root *root,
2510                                 u64 bytenr, u64 num_bytes, u64 flags,
2511                                 int is_data)
2512 {
2513         struct btrfs_delayed_extent_op *extent_op;
2514         int ret;
2515
2516         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2517         if (!extent_op)
2518                 return -ENOMEM;
2519
2520         extent_op->flags_to_set = flags;
2521         extent_op->update_flags = 1;
2522         extent_op->update_key = 0;
2523         extent_op->is_data = is_data ? 1 : 0;
2524
2525         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2526                                           num_bytes, extent_op);
2527         if (ret)
2528                 kfree(extent_op);
2529         return ret;
2530 }
2531
2532 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2533                                       struct btrfs_root *root,
2534                                       struct btrfs_path *path,
2535                                       u64 objectid, u64 offset, u64 bytenr)
2536 {
2537         struct btrfs_delayed_ref_head *head;
2538         struct btrfs_delayed_ref_node *ref;
2539         struct btrfs_delayed_data_ref *data_ref;
2540         struct btrfs_delayed_ref_root *delayed_refs;
2541         struct rb_node *node;
2542         int ret = 0;
2543
2544         ret = -ENOENT;
2545         delayed_refs = &trans->transaction->delayed_refs;
2546         spin_lock(&delayed_refs->lock);
2547         head = btrfs_find_delayed_ref_head(trans, bytenr);
2548         if (!head)
2549                 goto out;
2550
2551         if (!mutex_trylock(&head->mutex)) {
2552                 atomic_inc(&head->node.refs);
2553                 spin_unlock(&delayed_refs->lock);
2554
2555                 btrfs_release_path(path);
2556
2557                 /*
2558                  * Mutex was contended, block until it's released and let
2559                  * caller try again
2560                  */
2561                 mutex_lock(&head->mutex);
2562                 mutex_unlock(&head->mutex);
2563                 btrfs_put_delayed_ref(&head->node);
2564                 return -EAGAIN;
2565         }
2566
2567         node = rb_prev(&head->node.rb_node);
2568         if (!node)
2569                 goto out_unlock;
2570
2571         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2572
2573         if (ref->bytenr != bytenr)
2574                 goto out_unlock;
2575
2576         ret = 1;
2577         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2578                 goto out_unlock;
2579
2580         data_ref = btrfs_delayed_node_to_data_ref(ref);
2581
2582         node = rb_prev(node);
2583         if (node) {
2584                 int seq = ref->seq;
2585
2586                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2587                 if (ref->bytenr == bytenr && ref->seq == seq)
2588                         goto out_unlock;
2589         }
2590
2591         if (data_ref->root != root->root_key.objectid ||
2592             data_ref->objectid != objectid || data_ref->offset != offset)
2593                 goto out_unlock;
2594
2595         ret = 0;
2596 out_unlock:
2597         mutex_unlock(&head->mutex);
2598 out:
2599         spin_unlock(&delayed_refs->lock);
2600         return ret;
2601 }
2602
2603 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2604                                         struct btrfs_root *root,
2605                                         struct btrfs_path *path,
2606                                         u64 objectid, u64 offset, u64 bytenr)
2607 {
2608         struct btrfs_root *extent_root = root->fs_info->extent_root;
2609         struct extent_buffer *leaf;
2610         struct btrfs_extent_data_ref *ref;
2611         struct btrfs_extent_inline_ref *iref;
2612         struct btrfs_extent_item *ei;
2613         struct btrfs_key key;
2614         u32 item_size;
2615         int ret;
2616
2617         key.objectid = bytenr;
2618         key.offset = (u64)-1;
2619         key.type = BTRFS_EXTENT_ITEM_KEY;
2620
2621         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2622         if (ret < 0)
2623                 goto out;
2624         BUG_ON(ret == 0); /* Corruption */
2625
2626         ret = -ENOENT;
2627         if (path->slots[0] == 0)
2628                 goto out;
2629
2630         path->slots[0]--;
2631         leaf = path->nodes[0];
2632         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2633
2634         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2635                 goto out;
2636
2637         ret = 1;
2638         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2639 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2640         if (item_size < sizeof(*ei)) {
2641                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2642                 goto out;
2643         }
2644 #endif
2645         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2646
2647         if (item_size != sizeof(*ei) +
2648             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2649                 goto out;
2650
2651         if (btrfs_extent_generation(leaf, ei) <=
2652             btrfs_root_last_snapshot(&root->root_item))
2653                 goto out;
2654
2655         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2656         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2657             BTRFS_EXTENT_DATA_REF_KEY)
2658                 goto out;
2659
2660         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2661         if (btrfs_extent_refs(leaf, ei) !=
2662             btrfs_extent_data_ref_count(leaf, ref) ||
2663             btrfs_extent_data_ref_root(leaf, ref) !=
2664             root->root_key.objectid ||
2665             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2666             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2667                 goto out;
2668
2669         ret = 0;
2670 out:
2671         return ret;
2672 }
2673
2674 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2675                           struct btrfs_root *root,
2676                           u64 objectid, u64 offset, u64 bytenr)
2677 {
2678         struct btrfs_path *path;
2679         int ret;
2680         int ret2;
2681
2682         path = btrfs_alloc_path();
2683         if (!path)
2684                 return -ENOENT;
2685
2686         do {
2687                 ret = check_committed_ref(trans, root, path, objectid,
2688                                           offset, bytenr);
2689                 if (ret && ret != -ENOENT)
2690                         goto out;
2691
2692                 ret2 = check_delayed_ref(trans, root, path, objectid,
2693                                          offset, bytenr);
2694         } while (ret2 == -EAGAIN);
2695
2696         if (ret2 && ret2 != -ENOENT) {
2697                 ret = ret2;
2698                 goto out;
2699         }
2700
2701         if (ret != -ENOENT || ret2 != -ENOENT)
2702                 ret = 0;
2703 out:
2704         btrfs_free_path(path);
2705         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2706                 WARN_ON(ret > 0);
2707         return ret;
2708 }
2709
2710 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2711                            struct btrfs_root *root,
2712                            struct extent_buffer *buf,
2713                            int full_backref, int inc, int for_cow)
2714 {
2715         u64 bytenr;
2716         u64 num_bytes;
2717         u64 parent;
2718         u64 ref_root;
2719         u32 nritems;
2720         struct btrfs_key key;
2721         struct btrfs_file_extent_item *fi;
2722         int i;
2723         int level;
2724         int ret = 0;
2725         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2726                             u64, u64, u64, u64, u64, u64, int);
2727
2728         ref_root = btrfs_header_owner(buf);
2729         nritems = btrfs_header_nritems(buf);
2730         level = btrfs_header_level(buf);
2731
2732         if (!root->ref_cows && level == 0)
2733                 return 0;
2734
2735         if (inc)
2736                 process_func = btrfs_inc_extent_ref;
2737         else
2738                 process_func = btrfs_free_extent;
2739
2740         if (full_backref)
2741                 parent = buf->start;
2742         else
2743                 parent = 0;
2744
2745         for (i = 0; i < nritems; i++) {
2746                 if (level == 0) {
2747                         btrfs_item_key_to_cpu(buf, &key, i);
2748                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2749                                 continue;
2750                         fi = btrfs_item_ptr(buf, i,
2751                                             struct btrfs_file_extent_item);
2752                         if (btrfs_file_extent_type(buf, fi) ==
2753                             BTRFS_FILE_EXTENT_INLINE)
2754                                 continue;
2755                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2756                         if (bytenr == 0)
2757                                 continue;
2758
2759                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2760                         key.offset -= btrfs_file_extent_offset(buf, fi);
2761                         ret = process_func(trans, root, bytenr, num_bytes,
2762                                            parent, ref_root, key.objectid,
2763                                            key.offset, for_cow);
2764                         if (ret)
2765                                 goto fail;
2766                 } else {
2767                         bytenr = btrfs_node_blockptr(buf, i);
2768                         num_bytes = btrfs_level_size(root, level - 1);
2769                         ret = process_func(trans, root, bytenr, num_bytes,
2770                                            parent, ref_root, level - 1, 0,
2771                                            for_cow);
2772                         if (ret)
2773                                 goto fail;
2774                 }
2775         }
2776         return 0;
2777 fail:
2778         return ret;
2779 }
2780
2781 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2782                   struct extent_buffer *buf, int full_backref, int for_cow)
2783 {
2784         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2785 }
2786
2787 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2788                   struct extent_buffer *buf, int full_backref, int for_cow)
2789 {
2790         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2791 }
2792
2793 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2794                                  struct btrfs_root *root,
2795                                  struct btrfs_path *path,
2796                                  struct btrfs_block_group_cache *cache)
2797 {
2798         int ret;
2799         struct btrfs_root *extent_root = root->fs_info->extent_root;
2800         unsigned long bi;
2801         struct extent_buffer *leaf;
2802
2803         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2804         if (ret < 0)
2805                 goto fail;
2806         BUG_ON(ret); /* Corruption */
2807
2808         leaf = path->nodes[0];
2809         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2810         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2811         btrfs_mark_buffer_dirty(leaf);
2812         btrfs_release_path(path);
2813 fail:
2814         if (ret) {
2815                 btrfs_abort_transaction(trans, root, ret);
2816                 return ret;
2817         }
2818         return 0;
2819
2820 }
2821
2822 static struct btrfs_block_group_cache *
2823 next_block_group(struct btrfs_root *root,
2824                  struct btrfs_block_group_cache *cache)
2825 {
2826         struct rb_node *node;
2827         spin_lock(&root->fs_info->block_group_cache_lock);
2828         node = rb_next(&cache->cache_node);
2829         btrfs_put_block_group(cache);
2830         if (node) {
2831                 cache = rb_entry(node, struct btrfs_block_group_cache,
2832                                  cache_node);
2833                 btrfs_get_block_group(cache);
2834         } else
2835                 cache = NULL;
2836         spin_unlock(&root->fs_info->block_group_cache_lock);
2837         return cache;
2838 }
2839
2840 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2841                             struct btrfs_trans_handle *trans,
2842                             struct btrfs_path *path)
2843 {
2844         struct btrfs_root *root = block_group->fs_info->tree_root;
2845         struct inode *inode = NULL;
2846         u64 alloc_hint = 0;
2847         int dcs = BTRFS_DC_ERROR;
2848         int num_pages = 0;
2849         int retries = 0;
2850         int ret = 0;
2851
2852         /*
2853          * If this block group is smaller than 100 megs don't bother caching the
2854          * block group.
2855          */
2856         if (block_group->key.offset < (100 * 1024 * 1024)) {
2857                 spin_lock(&block_group->lock);
2858                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2859                 spin_unlock(&block_group->lock);
2860                 return 0;
2861         }
2862
2863 again:
2864         inode = lookup_free_space_inode(root, block_group, path);
2865         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2866                 ret = PTR_ERR(inode);
2867                 btrfs_release_path(path);
2868                 goto out;
2869         }
2870
2871         if (IS_ERR(inode)) {
2872                 BUG_ON(retries);
2873                 retries++;
2874
2875                 if (block_group->ro)
2876                         goto out_free;
2877
2878                 ret = create_free_space_inode(root, trans, block_group, path);
2879                 if (ret)
2880                         goto out_free;
2881                 goto again;
2882         }
2883
2884         /* We've already setup this transaction, go ahead and exit */
2885         if (block_group->cache_generation == trans->transid &&
2886             i_size_read(inode)) {
2887                 dcs = BTRFS_DC_SETUP;
2888                 goto out_put;
2889         }
2890
2891         /*
2892          * We want to set the generation to 0, that way if anything goes wrong
2893          * from here on out we know not to trust this cache when we load up next
2894          * time.
2895          */
2896         BTRFS_I(inode)->generation = 0;
2897         ret = btrfs_update_inode(trans, root, inode);
2898         WARN_ON(ret);
2899
2900         if (i_size_read(inode) > 0) {
2901                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2902                                                       inode);
2903                 if (ret)
2904                         goto out_put;
2905         }
2906
2907         spin_lock(&block_group->lock);
2908         if (block_group->cached != BTRFS_CACHE_FINISHED ||
2909             !btrfs_test_opt(root, SPACE_CACHE)) {
2910                 /*
2911                  * don't bother trying to write stuff out _if_
2912                  * a) we're not cached,
2913                  * b) we're with nospace_cache mount option.
2914                  */
2915                 dcs = BTRFS_DC_WRITTEN;
2916                 spin_unlock(&block_group->lock);
2917                 goto out_put;
2918         }
2919         spin_unlock(&block_group->lock);
2920
2921         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2922         if (!num_pages)
2923                 num_pages = 1;
2924
2925         /*
2926          * Just to make absolutely sure we have enough space, we're going to
2927          * preallocate 12 pages worth of space for each block group.  In
2928          * practice we ought to use at most 8, but we need extra space so we can
2929          * add our header and have a terminator between the extents and the
2930          * bitmaps.
2931          */
2932         num_pages *= 16;
2933         num_pages *= PAGE_CACHE_SIZE;
2934
2935         ret = btrfs_check_data_free_space(inode, num_pages);
2936         if (ret)
2937                 goto out_put;
2938
2939         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2940                                               num_pages, num_pages,
2941                                               &alloc_hint);
2942         if (!ret)
2943                 dcs = BTRFS_DC_SETUP;
2944         btrfs_free_reserved_data_space(inode, num_pages);
2945
2946 out_put:
2947         iput(inode);
2948 out_free:
2949         btrfs_release_path(path);
2950 out:
2951         spin_lock(&block_group->lock);
2952         if (!ret && dcs == BTRFS_DC_SETUP)
2953                 block_group->cache_generation = trans->transid;
2954         block_group->disk_cache_state = dcs;
2955         spin_unlock(&block_group->lock);
2956
2957         return ret;
2958 }
2959
2960 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2961                                    struct btrfs_root *root)
2962 {
2963         struct btrfs_block_group_cache *cache;
2964         int err = 0;
2965         struct btrfs_path *path;
2966         u64 last = 0;
2967
2968         path = btrfs_alloc_path();
2969         if (!path)
2970                 return -ENOMEM;
2971
2972 again:
2973         while (1) {
2974                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2975                 while (cache) {
2976                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2977                                 break;
2978                         cache = next_block_group(root, cache);
2979                 }
2980                 if (!cache) {
2981                         if (last == 0)
2982                                 break;
2983                         last = 0;
2984                         continue;
2985                 }
2986                 err = cache_save_setup(cache, trans, path);
2987                 last = cache->key.objectid + cache->key.offset;
2988                 btrfs_put_block_group(cache);
2989         }
2990
2991         while (1) {
2992                 if (last == 0) {
2993                         err = btrfs_run_delayed_refs(trans, root,
2994                                                      (unsigned long)-1);
2995                         if (err) /* File system offline */
2996                                 goto out;
2997                 }
2998
2999                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3000                 while (cache) {
3001                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3002                                 btrfs_put_block_group(cache);
3003                                 goto again;
3004                         }
3005
3006                         if (cache->dirty)
3007                                 break;
3008                         cache = next_block_group(root, cache);
3009                 }
3010                 if (!cache) {
3011                         if (last == 0)
3012                                 break;
3013                         last = 0;
3014                         continue;
3015                 }
3016
3017                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3018                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3019                 cache->dirty = 0;
3020                 last = cache->key.objectid + cache->key.offset;
3021
3022                 err = write_one_cache_group(trans, root, path, cache);
3023                 if (err) /* File system offline */
3024                         goto out;
3025
3026                 btrfs_put_block_group(cache);
3027         }
3028
3029         while (1) {
3030                 /*
3031                  * I don't think this is needed since we're just marking our
3032                  * preallocated extent as written, but just in case it can't
3033                  * hurt.
3034                  */
3035                 if (last == 0) {
3036                         err = btrfs_run_delayed_refs(trans, root,
3037                                                      (unsigned long)-1);
3038                         if (err) /* File system offline */
3039                                 goto out;
3040                 }
3041
3042                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3043                 while (cache) {
3044                         /*
3045                          * Really this shouldn't happen, but it could if we
3046                          * couldn't write the entire preallocated extent and
3047                          * splitting the extent resulted in a new block.
3048                          */
3049                         if (cache->dirty) {
3050                                 btrfs_put_block_group(cache);
3051                                 goto again;
3052                         }
3053                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3054                                 break;
3055                         cache = next_block_group(root, cache);
3056                 }
3057                 if (!cache) {
3058                         if (last == 0)
3059                                 break;
3060                         last = 0;
3061                         continue;
3062                 }
3063
3064                 err = btrfs_write_out_cache(root, trans, cache, path);
3065
3066                 /*
3067                  * If we didn't have an error then the cache state is still
3068                  * NEED_WRITE, so we can set it to WRITTEN.
3069                  */
3070                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3071                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3072                 last = cache->key.objectid + cache->key.offset;
3073                 btrfs_put_block_group(cache);
3074         }
3075 out:
3076
3077         btrfs_free_path(path);
3078         return err;
3079 }
3080
3081 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3082 {
3083         struct btrfs_block_group_cache *block_group;
3084         int readonly = 0;
3085
3086         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3087         if (!block_group || block_group->ro)
3088                 readonly = 1;
3089         if (block_group)
3090                 btrfs_put_block_group(block_group);
3091         return readonly;
3092 }
3093
3094 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3095                              u64 total_bytes, u64 bytes_used,
3096                              struct btrfs_space_info **space_info)
3097 {
3098         struct btrfs_space_info *found;
3099         int i;
3100         int factor;
3101
3102         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3103                      BTRFS_BLOCK_GROUP_RAID10))
3104                 factor = 2;
3105         else
3106                 factor = 1;
3107
3108         found = __find_space_info(info, flags);
3109         if (found) {
3110                 spin_lock(&found->lock);
3111                 found->total_bytes += total_bytes;
3112                 found->disk_total += total_bytes * factor;
3113                 found->bytes_used += bytes_used;
3114                 found->disk_used += bytes_used * factor;
3115                 found->full = 0;
3116                 spin_unlock(&found->lock);
3117                 *space_info = found;
3118                 return 0;
3119         }
3120         found = kzalloc(sizeof(*found), GFP_NOFS);
3121         if (!found)
3122                 return -ENOMEM;
3123
3124         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3125                 INIT_LIST_HEAD(&found->block_groups[i]);
3126         init_rwsem(&found->groups_sem);
3127         spin_lock_init(&found->lock);
3128         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3129         found->total_bytes = total_bytes;
3130         found->disk_total = total_bytes * factor;
3131         found->bytes_used = bytes_used;
3132         found->disk_used = bytes_used * factor;
3133         found->bytes_pinned = 0;
3134         found->bytes_reserved = 0;
3135         found->bytes_readonly = 0;
3136         found->bytes_may_use = 0;
3137         found->full = 0;
3138         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3139         found->chunk_alloc = 0;
3140         found->flush = 0;
3141         init_waitqueue_head(&found->wait);
3142         *space_info = found;
3143         list_add_rcu(&found->list, &info->space_info);
3144         if (flags & BTRFS_BLOCK_GROUP_DATA)
3145                 info->data_sinfo = found;
3146         return 0;
3147 }
3148
3149 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3150 {
3151         u64 extra_flags = chunk_to_extended(flags) &
3152                                 BTRFS_EXTENDED_PROFILE_MASK;
3153
3154         if (flags & BTRFS_BLOCK_GROUP_DATA)
3155                 fs_info->avail_data_alloc_bits |= extra_flags;
3156         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3157                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3158         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3159                 fs_info->avail_system_alloc_bits |= extra_flags;
3160 }
3161
3162 /*
3163  * returns target flags in extended format or 0 if restripe for this
3164  * chunk_type is not in progress
3165  *
3166  * should be called with either volume_mutex or balance_lock held
3167  */
3168 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3169 {
3170         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3171         u64 target = 0;
3172
3173         if (!bctl)
3174                 return 0;
3175
3176         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3177             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3178                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3179         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3180                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3181                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3182         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3183                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3184                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3185         }
3186
3187         return target;
3188 }
3189
3190 /*
3191  * @flags: available profiles in extended format (see ctree.h)
3192  *
3193  * Returns reduced profile in chunk format.  If profile changing is in
3194  * progress (either running or paused) picks the target profile (if it's
3195  * already available), otherwise falls back to plain reducing.
3196  */
3197 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3198 {
3199         /*
3200          * we add in the count of missing devices because we want
3201          * to make sure that any RAID levels on a degraded FS
3202          * continue to be honored.
3203          */
3204         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3205                 root->fs_info->fs_devices->missing_devices;
3206         u64 target;
3207
3208         /*
3209          * see if restripe for this chunk_type is in progress, if so
3210          * try to reduce to the target profile
3211          */
3212         spin_lock(&root->fs_info->balance_lock);
3213         target = get_restripe_target(root->fs_info, flags);
3214         if (target) {
3215                 /* pick target profile only if it's already available */
3216                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3217                         spin_unlock(&root->fs_info->balance_lock);
3218                         return extended_to_chunk(target);
3219                 }
3220         }
3221         spin_unlock(&root->fs_info->balance_lock);
3222
3223         if (num_devices == 1)
3224                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3225         if (num_devices < 4)
3226                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3227
3228         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3229             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3230                       BTRFS_BLOCK_GROUP_RAID10))) {
3231                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3232         }
3233
3234         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3235             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3236                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3237         }
3238
3239         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3240             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3241              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3242              (flags & BTRFS_BLOCK_GROUP_DUP))) {
3243                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3244         }
3245
3246         return extended_to_chunk(flags);
3247 }
3248
3249 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3250 {
3251         if (flags & BTRFS_BLOCK_GROUP_DATA)
3252                 flags |= root->fs_info->avail_data_alloc_bits;
3253         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3254                 flags |= root->fs_info->avail_system_alloc_bits;
3255         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3256                 flags |= root->fs_info->avail_metadata_alloc_bits;
3257
3258         return btrfs_reduce_alloc_profile(root, flags);
3259 }
3260
3261 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3262 {
3263         u64 flags;
3264
3265         if (data)
3266                 flags = BTRFS_BLOCK_GROUP_DATA;
3267         else if (root == root->fs_info->chunk_root)
3268                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3269         else
3270                 flags = BTRFS_BLOCK_GROUP_METADATA;
3271
3272         return get_alloc_profile(root, flags);
3273 }
3274
3275 /*
3276  * This will check the space that the inode allocates from to make sure we have
3277  * enough space for bytes.
3278  */
3279 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3280 {
3281         struct btrfs_space_info *data_sinfo;
3282         struct btrfs_root *root = BTRFS_I(inode)->root;
3283         struct btrfs_fs_info *fs_info = root->fs_info;
3284         u64 used;
3285         int ret = 0, committed = 0, alloc_chunk = 1;
3286
3287         /* make sure bytes are sectorsize aligned */
3288         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3289
3290         if (root == root->fs_info->tree_root ||
3291             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3292                 alloc_chunk = 0;
3293                 committed = 1;
3294         }
3295
3296         data_sinfo = fs_info->data_sinfo;
3297         if (!data_sinfo)
3298                 goto alloc;
3299
3300 again:
3301         /* make sure we have enough space to handle the data first */
3302         spin_lock(&data_sinfo->lock);
3303         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3304                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3305                 data_sinfo->bytes_may_use;
3306
3307         if (used + bytes > data_sinfo->total_bytes) {
3308                 struct btrfs_trans_handle *trans;
3309
3310                 /*
3311                  * if we don't have enough free bytes in this space then we need
3312                  * to alloc a new chunk.
3313                  */
3314                 if (!data_sinfo->full && alloc_chunk) {
3315                         u64 alloc_target;
3316
3317                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3318                         spin_unlock(&data_sinfo->lock);
3319 alloc:
3320                         alloc_target = btrfs_get_alloc_profile(root, 1);
3321                         trans = btrfs_join_transaction(root);
3322                         if (IS_ERR(trans))
3323                                 return PTR_ERR(trans);
3324
3325                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3326                                              bytes + 2 * 1024 * 1024,
3327                                              alloc_target,
3328                                              CHUNK_ALLOC_NO_FORCE);
3329                         btrfs_end_transaction(trans, root);
3330                         if (ret < 0) {
3331                                 if (ret != -ENOSPC)
3332                                         return ret;
3333                                 else
3334                                         goto commit_trans;
3335                         }
3336
3337                         if (!data_sinfo)
3338                                 data_sinfo = fs_info->data_sinfo;
3339
3340                         goto again;
3341                 }
3342
3343                 /*
3344                  * If we have less pinned bytes than we want to allocate then
3345                  * don't bother committing the transaction, it won't help us.
3346                  */
3347                 if (data_sinfo->bytes_pinned < bytes)
3348                         committed = 1;
3349                 spin_unlock(&data_sinfo->lock);
3350
3351                 /* commit the current transaction and try again */
3352 commit_trans:
3353                 if (!committed &&
3354                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3355                         committed = 1;
3356                         trans = btrfs_join_transaction(root);
3357                         if (IS_ERR(trans))
3358                                 return PTR_ERR(trans);
3359                         ret = btrfs_commit_transaction(trans, root);
3360                         if (ret)
3361                                 return ret;
3362                         goto again;
3363                 }
3364
3365                 return -ENOSPC;
3366         }
3367         data_sinfo->bytes_may_use += bytes;
3368         trace_btrfs_space_reservation(root->fs_info, "space_info",
3369                                       data_sinfo->flags, bytes, 1);
3370         spin_unlock(&data_sinfo->lock);
3371
3372         return 0;
3373 }
3374
3375 /*
3376  * Called if we need to clear a data reservation for this inode.
3377  */
3378 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3379 {
3380         struct btrfs_root *root = BTRFS_I(inode)->root;
3381         struct btrfs_space_info *data_sinfo;
3382
3383         /* make sure bytes are sectorsize aligned */
3384         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3385
3386         data_sinfo = root->fs_info->data_sinfo;
3387         spin_lock(&data_sinfo->lock);
3388         data_sinfo->bytes_may_use -= bytes;
3389         trace_btrfs_space_reservation(root->fs_info, "space_info",
3390                                       data_sinfo->flags, bytes, 0);
3391         spin_unlock(&data_sinfo->lock);
3392 }
3393
3394 static void force_metadata_allocation(struct btrfs_fs_info *info)
3395 {
3396         struct list_head *head = &info->space_info;
3397         struct btrfs_space_info *found;
3398
3399         rcu_read_lock();
3400         list_for_each_entry_rcu(found, head, list) {
3401                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3402                         found->force_alloc = CHUNK_ALLOC_FORCE;
3403         }
3404         rcu_read_unlock();
3405 }
3406
3407 static int should_alloc_chunk(struct btrfs_root *root,
3408                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3409                               int force)
3410 {
3411         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3412         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3413         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3414         u64 thresh;
3415
3416         if (force == CHUNK_ALLOC_FORCE)
3417                 return 1;
3418
3419         /*
3420          * We need to take into account the global rsv because for all intents
3421          * and purposes it's used space.  Don't worry about locking the
3422          * global_rsv, it doesn't change except when the transaction commits.
3423          */
3424         num_allocated += global_rsv->size;
3425
3426         /*
3427          * in limited mode, we want to have some free space up to
3428          * about 1% of the FS size.
3429          */
3430         if (force == CHUNK_ALLOC_LIMITED) {
3431                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3432                 thresh = max_t(u64, 64 * 1024 * 1024,
3433                                div_factor_fine(thresh, 1));
3434
3435                 if (num_bytes - num_allocated < thresh)
3436                         return 1;
3437         }
3438         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3439
3440         /* 256MB or 2% of the FS */
3441         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 2));
3442         /* system chunks need a much small threshold */
3443         if (sinfo->flags & BTRFS_BLOCK_GROUP_SYSTEM)
3444                 thresh = 32 * 1024 * 1024;
3445
3446         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 8))
3447                 return 0;
3448         return 1;
3449 }
3450
3451 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3452 {
3453         u64 num_dev;
3454
3455         if (type & BTRFS_BLOCK_GROUP_RAID10 ||
3456             type & BTRFS_BLOCK_GROUP_RAID0)
3457                 num_dev = root->fs_info->fs_devices->rw_devices;
3458         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3459                 num_dev = 2;
3460         else
3461                 num_dev = 1;    /* DUP or single */
3462
3463         /* metadata for updaing devices and chunk tree */
3464         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3465 }
3466
3467 static void check_system_chunk(struct btrfs_trans_handle *trans,
3468                                struct btrfs_root *root, u64 type)
3469 {
3470         struct btrfs_space_info *info;
3471         u64 left;
3472         u64 thresh;
3473
3474         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3475         spin_lock(&info->lock);
3476         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3477                 info->bytes_reserved - info->bytes_readonly;
3478         spin_unlock(&info->lock);
3479
3480         thresh = get_system_chunk_thresh(root, type);
3481         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3482                 printk(KERN_INFO "left=%llu, need=%llu, flags=%llu\n",
3483                        left, thresh, type);
3484                 dump_space_info(info, 0, 0);
3485         }
3486
3487         if (left < thresh) {
3488                 u64 flags;
3489
3490                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3491                 btrfs_alloc_chunk(trans, root, flags);
3492         }
3493 }
3494
3495 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3496                           struct btrfs_root *extent_root, u64 alloc_bytes,
3497                           u64 flags, int force)
3498 {
3499         struct btrfs_space_info *space_info;
3500         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3501         int wait_for_alloc = 0;
3502         int ret = 0;
3503
3504         space_info = __find_space_info(extent_root->fs_info, flags);
3505         if (!space_info) {
3506                 ret = update_space_info(extent_root->fs_info, flags,
3507                                         0, 0, &space_info);
3508                 BUG_ON(ret); /* -ENOMEM */
3509         }
3510         BUG_ON(!space_info); /* Logic error */
3511
3512 again:
3513         spin_lock(&space_info->lock);
3514         if (force < space_info->force_alloc)
3515                 force = space_info->force_alloc;
3516         if (space_info->full) {
3517                 spin_unlock(&space_info->lock);
3518                 return 0;
3519         }
3520
3521         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3522                 spin_unlock(&space_info->lock);
3523                 return 0;
3524         } else if (space_info->chunk_alloc) {
3525                 wait_for_alloc = 1;
3526         } else {
3527                 space_info->chunk_alloc = 1;
3528         }
3529
3530         spin_unlock(&space_info->lock);
3531
3532         mutex_lock(&fs_info->chunk_mutex);
3533
3534         /*
3535          * The chunk_mutex is held throughout the entirety of a chunk
3536          * allocation, so once we've acquired the chunk_mutex we know that the
3537          * other guy is done and we need to recheck and see if we should
3538          * allocate.
3539          */
3540         if (wait_for_alloc) {
3541                 mutex_unlock(&fs_info->chunk_mutex);
3542                 wait_for_alloc = 0;
3543                 goto again;
3544         }
3545
3546         /*
3547          * If we have mixed data/metadata chunks we want to make sure we keep
3548          * allocating mixed chunks instead of individual chunks.
3549          */
3550         if (btrfs_mixed_space_info(space_info))
3551                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3552
3553         /*
3554          * if we're doing a data chunk, go ahead and make sure that
3555          * we keep a reasonable number of metadata chunks allocated in the
3556          * FS as well.
3557          */
3558         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3559                 fs_info->data_chunk_allocations++;
3560                 if (!(fs_info->data_chunk_allocations %
3561                       fs_info->metadata_ratio))
3562                         force_metadata_allocation(fs_info);
3563         }
3564
3565         /*
3566          * Check if we have enough space in SYSTEM chunk because we may need
3567          * to update devices.
3568          */
3569         check_system_chunk(trans, extent_root, flags);
3570
3571         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3572         if (ret < 0 && ret != -ENOSPC)
3573                 goto out;
3574
3575         spin_lock(&space_info->lock);
3576         if (ret)
3577                 space_info->full = 1;
3578         else
3579                 ret = 1;
3580
3581         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3582         space_info->chunk_alloc = 0;
3583         spin_unlock(&space_info->lock);
3584 out:
3585         mutex_unlock(&fs_info->chunk_mutex);
3586         return ret;
3587 }
3588
3589 /*
3590  * shrink metadata reservation for delalloc
3591  */
3592 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3593                             bool wait_ordered)
3594 {
3595         struct btrfs_block_rsv *block_rsv;
3596         struct btrfs_space_info *space_info;
3597         struct btrfs_trans_handle *trans;
3598         u64 delalloc_bytes;
3599         u64 max_reclaim;
3600         long time_left;
3601         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3602         int loops = 0;
3603
3604         trans = (struct btrfs_trans_handle *)current->journal_info;
3605         block_rsv = &root->fs_info->delalloc_block_rsv;
3606         space_info = block_rsv->space_info;
3607
3608         smp_mb();
3609         delalloc_bytes = root->fs_info->delalloc_bytes;
3610         if (delalloc_bytes == 0) {
3611                 if (trans)
3612                         return;
3613                 btrfs_wait_ordered_extents(root, 0, 0);
3614                 return;
3615         }
3616
3617         while (delalloc_bytes && loops < 3) {
3618                 max_reclaim = min(delalloc_bytes, to_reclaim);
3619                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3620                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3621                                                WB_REASON_FS_FREE_SPACE);
3622
3623                 spin_lock(&space_info->lock);
3624                 if (space_info->bytes_used + space_info->bytes_reserved +
3625                     space_info->bytes_pinned + space_info->bytes_readonly +
3626                     space_info->bytes_may_use + orig <=
3627                     space_info->total_bytes) {
3628                         spin_unlock(&space_info->lock);
3629                         break;
3630                 }
3631                 spin_unlock(&space_info->lock);
3632
3633                 loops++;
3634                 if (wait_ordered && !trans) {
3635                         btrfs_wait_ordered_extents(root, 0, 0);
3636                 } else {
3637                         time_left = schedule_timeout_killable(1);
3638                         if (time_left)
3639                                 break;
3640                 }
3641                 smp_mb();
3642                 delalloc_bytes = root->fs_info->delalloc_bytes;
3643         }
3644 }
3645
3646 /**
3647  * maybe_commit_transaction - possibly commit the transaction if its ok to
3648  * @root - the root we're allocating for
3649  * @bytes - the number of bytes we want to reserve
3650  * @force - force the commit
3651  *
3652  * This will check to make sure that committing the transaction will actually
3653  * get us somewhere and then commit the transaction if it does.  Otherwise it
3654  * will return -ENOSPC.
3655  */
3656 static int may_commit_transaction(struct btrfs_root *root,
3657                                   struct btrfs_space_info *space_info,
3658                                   u64 bytes, int force)
3659 {
3660         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3661         struct btrfs_trans_handle *trans;
3662
3663         trans = (struct btrfs_trans_handle *)current->journal_info;
3664         if (trans)
3665                 return -EAGAIN;
3666
3667         if (force)
3668                 goto commit;
3669
3670         /* See if there is enough pinned space to make this reservation */
3671         spin_lock(&space_info->lock);
3672         if (space_info->bytes_pinned >= bytes) {
3673                 spin_unlock(&space_info->lock);
3674                 goto commit;
3675         }
3676         spin_unlock(&space_info->lock);
3677
3678         /*
3679          * See if there is some space in the delayed insertion reservation for
3680          * this reservation.
3681          */
3682         if (space_info != delayed_rsv->space_info)
3683                 return -ENOSPC;
3684
3685         spin_lock(&space_info->lock);
3686         spin_lock(&delayed_rsv->lock);
3687         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3688                 spin_unlock(&delayed_rsv->lock);
3689                 spin_unlock(&space_info->lock);
3690                 return -ENOSPC;
3691         }
3692         spin_unlock(&delayed_rsv->lock);
3693         spin_unlock(&space_info->lock);
3694
3695 commit:
3696         trans = btrfs_join_transaction(root);
3697         if (IS_ERR(trans))
3698                 return -ENOSPC;
3699
3700         return btrfs_commit_transaction(trans, root);
3701 }
3702
3703 enum flush_state {
3704         FLUSH_DELALLOC          =       1,
3705         FLUSH_DELALLOC_WAIT     =       2,
3706         FLUSH_DELAYED_ITEMS_NR  =       3,
3707         FLUSH_DELAYED_ITEMS     =       4,
3708         COMMIT_TRANS            =       5,
3709 };
3710
3711 static int flush_space(struct btrfs_root *root,
3712                        struct btrfs_space_info *space_info, u64 num_bytes,
3713                        u64 orig_bytes, int state)
3714 {
3715         struct btrfs_trans_handle *trans;
3716         int nr;
3717         int ret = 0;
3718
3719         switch (state) {
3720         case FLUSH_DELALLOC:
3721         case FLUSH_DELALLOC_WAIT:
3722                 shrink_delalloc(root, num_bytes, orig_bytes,
3723                                 state == FLUSH_DELALLOC_WAIT);
3724                 break;
3725         case FLUSH_DELAYED_ITEMS_NR:
3726         case FLUSH_DELAYED_ITEMS:
3727                 if (state == FLUSH_DELAYED_ITEMS_NR) {
3728                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
3729
3730                         nr = (int)div64_u64(num_bytes, bytes);
3731                         if (!nr)
3732                                 nr = 1;
3733                         nr *= 2;
3734                 } else {
3735                         nr = -1;
3736                 }
3737                 trans = btrfs_join_transaction(root);
3738                 if (IS_ERR(trans)) {
3739                         ret = PTR_ERR(trans);
3740                         break;
3741                 }
3742                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
3743                 btrfs_end_transaction(trans, root);
3744                 break;
3745         case COMMIT_TRANS:
3746                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3747                 break;
3748         default:
3749                 ret = -ENOSPC;
3750                 break;
3751         }
3752
3753         return ret;
3754 }
3755 /**
3756  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3757  * @root - the root we're allocating for
3758  * @block_rsv - the block_rsv we're allocating for
3759  * @orig_bytes - the number of bytes we want
3760  * @flush - wether or not we can flush to make our reservation
3761  *
3762  * This will reserve orgi_bytes number of bytes from the space info associated
3763  * with the block_rsv.  If there is not enough space it will make an attempt to
3764  * flush out space to make room.  It will do this by flushing delalloc if
3765  * possible or committing the transaction.  If flush is 0 then no attempts to
3766  * regain reservations will be made and this will fail if there is not enough
3767  * space already.
3768  */
3769 static int reserve_metadata_bytes(struct btrfs_root *root,
3770                                   struct btrfs_block_rsv *block_rsv,
3771                                   u64 orig_bytes, int flush)
3772 {
3773         struct btrfs_space_info *space_info = block_rsv->space_info;
3774         u64 used;
3775         u64 num_bytes = orig_bytes;
3776         int flush_state = FLUSH_DELALLOC;
3777         int ret = 0;
3778         bool flushing = false;
3779         bool committed = false;
3780
3781 again:
3782         ret = 0;
3783         spin_lock(&space_info->lock);
3784         /*
3785          * We only want to wait if somebody other than us is flushing and we are
3786          * actually alloed to flush.
3787          */
3788         while (flush && !flushing && space_info->flush) {
3789                 spin_unlock(&space_info->lock);
3790                 /*
3791                  * If we have a trans handle we can't wait because the flusher
3792                  * may have to commit the transaction, which would mean we would
3793                  * deadlock since we are waiting for the flusher to finish, but
3794                  * hold the current transaction open.
3795                  */
3796                 if (current->journal_info)
3797                         return -EAGAIN;
3798                 ret = wait_event_killable(space_info->wait, !space_info->flush);
3799                 /* Must have been killed, return */
3800                 if (ret)
3801                         return -EINTR;
3802
3803                 spin_lock(&space_info->lock);
3804         }
3805
3806         ret = -ENOSPC;
3807         used = space_info->bytes_used + space_info->bytes_reserved +
3808                 space_info->bytes_pinned + space_info->bytes_readonly +
3809                 space_info->bytes_may_use;
3810
3811         /*
3812          * The idea here is that we've not already over-reserved the block group
3813          * then we can go ahead and save our reservation first and then start
3814          * flushing if we need to.  Otherwise if we've already overcommitted
3815          * lets start flushing stuff first and then come back and try to make
3816          * our reservation.
3817          */
3818         if (used <= space_info->total_bytes) {
3819                 if (used + orig_bytes <= space_info->total_bytes) {
3820                         space_info->bytes_may_use += orig_bytes;
3821                         trace_btrfs_space_reservation(root->fs_info,
3822                                 "space_info", space_info->flags, orig_bytes, 1);
3823                         ret = 0;
3824                 } else {
3825                         /*
3826                          * Ok set num_bytes to orig_bytes since we aren't
3827                          * overocmmitted, this way we only try and reclaim what
3828                          * we need.
3829                          */
3830                         num_bytes = orig_bytes;
3831                 }
3832         } else {
3833                 /*
3834                  * Ok we're over committed, set num_bytes to the overcommitted
3835                  * amount plus the amount of bytes that we need for this
3836                  * reservation.
3837                  */
3838                 num_bytes = used - space_info->total_bytes +
3839                         (orig_bytes * 2);
3840         }
3841
3842         if (ret) {
3843                 u64 profile = btrfs_get_alloc_profile(root, 0);
3844                 u64 avail;
3845
3846                 /*
3847                  * If we have a lot of space that's pinned, don't bother doing
3848                  * the overcommit dance yet and just commit the transaction.
3849                  */
3850                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3851                 do_div(avail, 10);
3852                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3853                         space_info->flush = 1;
3854                         flushing = true;
3855                         spin_unlock(&space_info->lock);
3856                         ret = may_commit_transaction(root, space_info,
3857                                                      orig_bytes, 1);
3858                         if (ret)
3859                                 goto out;
3860                         committed = true;
3861                         goto again;
3862                 }
3863
3864                 spin_lock(&root->fs_info->free_chunk_lock);
3865                 avail = root->fs_info->free_chunk_space;
3866
3867                 /*
3868                  * If we have dup, raid1 or raid10 then only half of the free
3869                  * space is actually useable.
3870                  */
3871                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3872                                BTRFS_BLOCK_GROUP_RAID1 |
3873                                BTRFS_BLOCK_GROUP_RAID10))
3874                         avail >>= 1;
3875
3876                 /*
3877                  * If we aren't flushing don't let us overcommit too much, say
3878                  * 1/8th of the space.  If we can flush, let it overcommit up to
3879                  * 1/2 of the space.
3880                  */
3881                 if (flush)
3882                         avail >>= 3;
3883                 else
3884                         avail >>= 1;
3885                  spin_unlock(&root->fs_info->free_chunk_lock);
3886
3887                 if (used + num_bytes < space_info->total_bytes + avail) {
3888                         space_info->bytes_may_use += orig_bytes;
3889                         trace_btrfs_space_reservation(root->fs_info,
3890                                 "space_info", space_info->flags, orig_bytes, 1);
3891                         ret = 0;
3892                 }
3893         }
3894
3895         /*
3896          * Couldn't make our reservation, save our place so while we're trying
3897          * to reclaim space we can actually use it instead of somebody else
3898          * stealing it from us.
3899          */
3900         if (ret && flush) {
3901                 flushing = true;
3902                 space_info->flush = 1;
3903         }
3904
3905         spin_unlock(&space_info->lock);
3906
3907         if (!ret || !flush)
3908                 goto out;
3909
3910         ret = flush_space(root, space_info, num_bytes, orig_bytes,
3911                           flush_state);
3912         flush_state++;
3913         if (!ret)
3914                 goto again;
3915         else if (flush_state <= COMMIT_TRANS)
3916                 goto again;
3917
3918 out:
3919         if (flushing) {
3920                 spin_lock(&space_info->lock);
3921                 space_info->flush = 0;
3922                 wake_up_all(&space_info->wait);
3923                 spin_unlock(&space_info->lock);
3924         }
3925         return ret;
3926 }
3927
3928 static struct btrfs_block_rsv *get_block_rsv(
3929                                         const struct btrfs_trans_handle *trans,
3930                                         const struct btrfs_root *root)
3931 {
3932         struct btrfs_block_rsv *block_rsv = NULL;
3933
3934         if (root->ref_cows)
3935                 block_rsv = trans->block_rsv;
3936
3937         if (root == root->fs_info->csum_root && trans->adding_csums)
3938                 block_rsv = trans->block_rsv;
3939
3940         if (!block_rsv)
3941                 block_rsv = root->block_rsv;
3942
3943         if (!block_rsv)
3944                 block_rsv = &root->fs_info->empty_block_rsv;
3945
3946         return block_rsv;
3947 }
3948
3949 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3950                                u64 num_bytes)
3951 {
3952         int ret = -ENOSPC;
3953         spin_lock(&block_rsv->lock);
3954         if (block_rsv->reserved >= num_bytes) {
3955                 block_rsv->reserved -= num_bytes;
3956                 if (block_rsv->reserved < block_rsv->size)
3957                         block_rsv->full = 0;
3958                 ret = 0;
3959         }
3960         spin_unlock(&block_rsv->lock);
3961         return ret;
3962 }
3963
3964 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3965                                 u64 num_bytes, int update_size)
3966 {
3967         spin_lock(&block_rsv->lock);
3968         block_rsv->reserved += num_bytes;
3969         if (update_size)
3970                 block_rsv->size += num_bytes;
3971         else if (block_rsv->reserved >= block_rsv->size)
3972                 block_rsv->full = 1;
3973         spin_unlock(&block_rsv->lock);
3974 }
3975
3976 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
3977                                     struct btrfs_block_rsv *block_rsv,
3978                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3979 {
3980         struct btrfs_space_info *space_info = block_rsv->space_info;
3981
3982         spin_lock(&block_rsv->lock);
3983         if (num_bytes == (u64)-1)
3984                 num_bytes = block_rsv->size;
3985         block_rsv->size -= num_bytes;
3986         if (block_rsv->reserved >= block_rsv->size) {
3987                 num_bytes = block_rsv->reserved - block_rsv->size;
3988                 block_rsv->reserved = block_rsv->size;
3989                 block_rsv->full = 1;
3990         } else {
3991                 num_bytes = 0;
3992         }
3993         spin_unlock(&block_rsv->lock);
3994
3995         if (num_bytes > 0) {
3996                 if (dest) {
3997                         spin_lock(&dest->lock);
3998                         if (!dest->full) {
3999                                 u64 bytes_to_add;
4000
4001                                 bytes_to_add = dest->size - dest->reserved;
4002                                 bytes_to_add = min(num_bytes, bytes_to_add);
4003                                 dest->reserved += bytes_to_add;
4004                                 if (dest->reserved >= dest->size)
4005                                         dest->full = 1;
4006                                 num_bytes -= bytes_to_add;
4007                         }
4008                         spin_unlock(&dest->lock);
4009                 }
4010                 if (num_bytes) {
4011                         spin_lock(&space_info->lock);
4012                         space_info->bytes_may_use -= num_bytes;
4013                         trace_btrfs_space_reservation(fs_info, "space_info",
4014                                         space_info->flags, num_bytes, 0);
4015                         space_info->reservation_progress++;
4016                         spin_unlock(&space_info->lock);
4017                 }
4018         }
4019 }
4020
4021 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4022                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4023 {
4024         int ret;
4025
4026         ret = block_rsv_use_bytes(src, num_bytes);
4027         if (ret)
4028                 return ret;
4029
4030         block_rsv_add_bytes(dst, num_bytes, 1);
4031         return 0;
4032 }
4033
4034 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
4035 {
4036         memset(rsv, 0, sizeof(*rsv));
4037         spin_lock_init(&rsv->lock);
4038 }
4039
4040 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
4041 {
4042         struct btrfs_block_rsv *block_rsv;
4043         struct btrfs_fs_info *fs_info = root->fs_info;
4044
4045         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4046         if (!block_rsv)
4047                 return NULL;
4048
4049         btrfs_init_block_rsv(block_rsv);
4050         block_rsv->space_info = __find_space_info(fs_info,
4051                                                   BTRFS_BLOCK_GROUP_METADATA);
4052         return block_rsv;
4053 }
4054
4055 void btrfs_free_block_rsv(struct btrfs_root *root,
4056                           struct btrfs_block_rsv *rsv)
4057 {
4058         btrfs_block_rsv_release(root, rsv, (u64)-1);
4059         kfree(rsv);
4060 }
4061
4062 static inline int __block_rsv_add(struct btrfs_root *root,
4063                                   struct btrfs_block_rsv *block_rsv,
4064                                   u64 num_bytes, int flush)
4065 {
4066         int ret;
4067
4068         if (num_bytes == 0)
4069                 return 0;
4070
4071         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4072         if (!ret) {
4073                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4074                 return 0;
4075         }
4076
4077         return ret;
4078 }
4079
4080 int btrfs_block_rsv_add(struct btrfs_root *root,
4081                         struct btrfs_block_rsv *block_rsv,
4082                         u64 num_bytes)
4083 {
4084         return __block_rsv_add(root, block_rsv, num_bytes, 1);
4085 }
4086
4087 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
4088                                 struct btrfs_block_rsv *block_rsv,
4089                                 u64 num_bytes)
4090 {
4091         return __block_rsv_add(root, block_rsv, num_bytes, 0);
4092 }
4093
4094 int btrfs_block_rsv_check(struct btrfs_root *root,
4095                           struct btrfs_block_rsv *block_rsv, int min_factor)
4096 {
4097         u64 num_bytes = 0;
4098         int ret = -ENOSPC;
4099
4100         if (!block_rsv)
4101                 return 0;
4102
4103         spin_lock(&block_rsv->lock);
4104         num_bytes = div_factor(block_rsv->size, min_factor);
4105         if (block_rsv->reserved >= num_bytes)
4106                 ret = 0;
4107         spin_unlock(&block_rsv->lock);
4108
4109         return ret;
4110 }
4111
4112 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
4113                                            struct btrfs_block_rsv *block_rsv,
4114                                            u64 min_reserved, int flush)
4115 {
4116         u64 num_bytes = 0;
4117         int ret = -ENOSPC;
4118
4119         if (!block_rsv)
4120                 return 0;
4121
4122         spin_lock(&block_rsv->lock);
4123         num_bytes = min_reserved;
4124         if (block_rsv->reserved >= num_bytes)
4125                 ret = 0;
4126         else
4127                 num_bytes -= block_rsv->reserved;
4128         spin_unlock(&block_rsv->lock);
4129
4130         if (!ret)
4131                 return 0;
4132
4133         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4134         if (!ret) {
4135                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4136                 return 0;
4137         }
4138
4139         return ret;
4140 }
4141
4142 int btrfs_block_rsv_refill(struct btrfs_root *root,
4143                            struct btrfs_block_rsv *block_rsv,
4144                            u64 min_reserved)
4145 {
4146         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
4147 }
4148
4149 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
4150                                    struct btrfs_block_rsv *block_rsv,
4151                                    u64 min_reserved)
4152 {
4153         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
4154 }
4155
4156 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4157                             struct btrfs_block_rsv *dst_rsv,
4158                             u64 num_bytes)
4159 {
4160         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4161 }
4162
4163 void btrfs_block_rsv_release(struct btrfs_root *root,
4164                              struct btrfs_block_rsv *block_rsv,
4165                              u64 num_bytes)
4166 {
4167         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4168         if (global_rsv->full || global_rsv == block_rsv ||
4169             block_rsv->space_info != global_rsv->space_info)
4170                 global_rsv = NULL;
4171         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4172                                 num_bytes);
4173 }
4174
4175 /*
4176  * helper to calculate size of global block reservation.
4177  * the desired value is sum of space used by extent tree,
4178  * checksum tree and root tree
4179  */
4180 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4181 {
4182         struct btrfs_space_info *sinfo;
4183         u64 num_bytes;
4184         u64 meta_used;
4185         u64 data_used;
4186         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4187
4188         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4189         spin_lock(&sinfo->lock);
4190         data_used = sinfo->bytes_used;
4191         spin_unlock(&sinfo->lock);
4192
4193         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4194         spin_lock(&sinfo->lock);
4195         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4196                 data_used = 0;
4197         meta_used = sinfo->bytes_used;
4198         spin_unlock(&sinfo->lock);
4199
4200         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4201                     csum_size * 2;
4202         num_bytes += div64_u64(data_used + meta_used, 50);
4203
4204         if (num_bytes * 3 > meta_used)
4205                 num_bytes = div64_u64(meta_used, 3);
4206
4207         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4208 }
4209
4210 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4211 {
4212         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4213         struct btrfs_space_info *sinfo = block_rsv->space_info;
4214         u64 num_bytes;
4215
4216         num_bytes = calc_global_metadata_size(fs_info);
4217
4218         spin_lock(&sinfo->lock);
4219         spin_lock(&block_rsv->lock);
4220
4221         block_rsv->size = num_bytes;
4222
4223         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4224                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4225                     sinfo->bytes_may_use;
4226
4227         if (sinfo->total_bytes > num_bytes) {
4228                 num_bytes = sinfo->total_bytes - num_bytes;
4229                 block_rsv->reserved += num_bytes;
4230                 sinfo->bytes_may_use += num_bytes;
4231                 trace_btrfs_space_reservation(fs_info, "space_info",
4232                                       sinfo->flags, num_bytes, 1);
4233         }
4234
4235         if (block_rsv->reserved >= block_rsv->size) {
4236                 num_bytes = block_rsv->reserved - block_rsv->size;
4237                 sinfo->bytes_may_use -= num_bytes;
4238                 trace_btrfs_space_reservation(fs_info, "space_info",
4239                                       sinfo->flags, num_bytes, 0);
4240                 sinfo->reservation_progress++;
4241                 block_rsv->reserved = block_rsv->size;
4242                 block_rsv->full = 1;
4243         }
4244
4245         spin_unlock(&block_rsv->lock);
4246         spin_unlock(&sinfo->lock);
4247 }
4248
4249 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4250 {
4251         struct btrfs_space_info *space_info;
4252
4253         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4254         fs_info->chunk_block_rsv.space_info = space_info;
4255
4256         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4257         fs_info->global_block_rsv.space_info = space_info;
4258         fs_info->delalloc_block_rsv.space_info = space_info;
4259         fs_info->trans_block_rsv.space_info = space_info;
4260         fs_info->empty_block_rsv.space_info = space_info;
4261         fs_info->delayed_block_rsv.space_info = space_info;
4262
4263         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4264         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4265         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4266         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4267         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4268
4269         update_global_block_rsv(fs_info);
4270 }
4271
4272 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4273 {
4274         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4275                                 (u64)-1);
4276         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4277         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4278         WARN_ON(fs_info->trans_block_rsv.size > 0);
4279         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4280         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4281         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4282         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4283         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4284 }
4285
4286 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4287                                   struct btrfs_root *root)
4288 {
4289         if (!trans->block_rsv)
4290                 return;
4291
4292         if (!trans->bytes_reserved)
4293                 return;
4294
4295         trace_btrfs_space_reservation(root->fs_info, "transaction",
4296                                       trans->transid, trans->bytes_reserved, 0);
4297         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4298         trans->bytes_reserved = 0;
4299 }
4300
4301 /* Can only return 0 or -ENOSPC */
4302 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4303                                   struct inode *inode)
4304 {
4305         struct btrfs_root *root = BTRFS_I(inode)->root;
4306         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4307         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4308
4309         /*
4310          * We need to hold space in order to delete our orphan item once we've
4311          * added it, so this takes the reservation so we can release it later
4312          * when we are truly done with the orphan item.
4313          */
4314         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4315         trace_btrfs_space_reservation(root->fs_info, "orphan",
4316                                       btrfs_ino(inode), num_bytes, 1);
4317         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4318 }
4319
4320 void btrfs_orphan_release_metadata(struct inode *inode)
4321 {
4322         struct btrfs_root *root = BTRFS_I(inode)->root;
4323         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4324         trace_btrfs_space_reservation(root->fs_info, "orphan",
4325                                       btrfs_ino(inode), num_bytes, 0);
4326         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4327 }
4328
4329 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4330                                 struct btrfs_pending_snapshot *pending)
4331 {
4332         struct btrfs_root *root = pending->root;
4333         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4334         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4335         /*
4336          * two for root back/forward refs, two for directory entries
4337          * and one for root of the snapshot.
4338          */
4339         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4340         dst_rsv->space_info = src_rsv->space_info;
4341         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4342 }
4343
4344 /**
4345  * drop_outstanding_extent - drop an outstanding extent
4346  * @inode: the inode we're dropping the extent for
4347  *
4348  * This is called when we are freeing up an outstanding extent, either called
4349  * after an error or after an extent is written.  This will return the number of
4350  * reserved extents that need to be freed.  This must be called with
4351  * BTRFS_I(inode)->lock held.
4352  */
4353 static unsigned drop_outstanding_extent(struct inode *inode)
4354 {
4355         unsigned drop_inode_space = 0;
4356         unsigned dropped_extents = 0;
4357
4358         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4359         BTRFS_I(inode)->outstanding_extents--;
4360
4361         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4362             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4363                                &BTRFS_I(inode)->runtime_flags))
4364                 drop_inode_space = 1;
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(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 (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4472                       &BTRFS_I(inode)->runtime_flags)) {
4473                 nr_extents++;
4474                 extra_reserve = 1;
4475         }
4476
4477         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4478         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4479         csum_bytes = BTRFS_I(inode)->csum_bytes;
4480         spin_unlock(&BTRFS_I(inode)->lock);
4481
4482         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4483         if (ret) {
4484                 u64 to_free = 0;
4485                 unsigned dropped;
4486
4487                 spin_lock(&BTRFS_I(inode)->lock);
4488                 dropped = drop_outstanding_extent(inode);
4489                 /*
4490                  * If the inodes csum_bytes is the same as the original
4491                  * csum_bytes then we know we haven't raced with any free()ers
4492                  * so we can just reduce our inodes csum bytes and carry on.
4493                  * Otherwise we have to do the normal free thing to account for
4494                  * the case that the free side didn't free up its reserve
4495                  * because of this outstanding reservation.
4496                  */
4497                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4498                         calc_csum_metadata_size(inode, num_bytes, 0);
4499                 else
4500                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4501                 spin_unlock(&BTRFS_I(inode)->lock);
4502                 if (dropped)
4503                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4504
4505                 if (to_free) {
4506                         btrfs_block_rsv_release(root, block_rsv, to_free);
4507                         trace_btrfs_space_reservation(root->fs_info,
4508                                                       "delalloc",
4509                                                       btrfs_ino(inode),
4510                                                       to_free, 0);
4511                 }
4512                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4513                 return ret;
4514         }
4515
4516         spin_lock(&BTRFS_I(inode)->lock);
4517         if (extra_reserve) {
4518                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4519                         &BTRFS_I(inode)->runtime_flags);
4520                 nr_extents--;
4521         }
4522         BTRFS_I(inode)->reserved_extents += nr_extents;
4523         spin_unlock(&BTRFS_I(inode)->lock);
4524         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4525
4526         if (to_reserve)
4527                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4528                                               btrfs_ino(inode), to_reserve, 1);
4529         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4530
4531         return 0;
4532 }
4533
4534 /**
4535  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4536  * @inode: the inode to release the reservation for
4537  * @num_bytes: the number of bytes we're releasing
4538  *
4539  * This will release the metadata reservation for an inode.  This can be called
4540  * once we complete IO for a given set of bytes to release their metadata
4541  * reservations.
4542  */
4543 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4544 {
4545         struct btrfs_root *root = BTRFS_I(inode)->root;
4546         u64 to_free = 0;
4547         unsigned dropped;
4548
4549         num_bytes = ALIGN(num_bytes, root->sectorsize);
4550         spin_lock(&BTRFS_I(inode)->lock);
4551         dropped = drop_outstanding_extent(inode);
4552
4553         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4554         spin_unlock(&BTRFS_I(inode)->lock);
4555         if (dropped > 0)
4556                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4557
4558         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4559                                       btrfs_ino(inode), to_free, 0);
4560         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4561                                 to_free);
4562 }
4563
4564 /**
4565  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4566  * @inode: inode we're writing to
4567  * @num_bytes: the number of bytes we want to allocate
4568  *
4569  * This will do the following things
4570  *
4571  * o reserve space in the data space info for num_bytes
4572  * o reserve space in the metadata space info based on number of outstanding
4573  *   extents and how much csums will be needed
4574  * o add to the inodes ->delalloc_bytes
4575  * o add it to the fs_info's delalloc inodes list.
4576  *
4577  * This will return 0 for success and -ENOSPC if there is no space left.
4578  */
4579 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4580 {
4581         int ret;
4582
4583         ret = btrfs_check_data_free_space(inode, num_bytes);
4584         if (ret)
4585                 return ret;
4586
4587         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4588         if (ret) {
4589                 btrfs_free_reserved_data_space(inode, num_bytes);
4590                 return ret;
4591         }
4592
4593         return 0;
4594 }
4595
4596 /**
4597  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4598  * @inode: inode we're releasing space for
4599  * @num_bytes: the number of bytes we want to free up
4600  *
4601  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4602  * called in the case that we don't need the metadata AND data reservations
4603  * anymore.  So if there is an error or we insert an inline extent.
4604  *
4605  * This function will release the metadata space that was not used and will
4606  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4607  * list if there are no delalloc bytes left.
4608  */
4609 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4610 {
4611         btrfs_delalloc_release_metadata(inode, num_bytes);
4612         btrfs_free_reserved_data_space(inode, num_bytes);
4613 }
4614
4615 static int update_block_group(struct btrfs_trans_handle *trans,
4616                               struct btrfs_root *root,
4617                               u64 bytenr, u64 num_bytes, int alloc)
4618 {
4619         struct btrfs_block_group_cache *cache = NULL;
4620         struct btrfs_fs_info *info = root->fs_info;
4621         u64 total = num_bytes;
4622         u64 old_val;
4623         u64 byte_in_group;
4624         int factor;
4625
4626         /* block accounting for super block */
4627         spin_lock(&info->delalloc_lock);
4628         old_val = btrfs_super_bytes_used(info->super_copy);
4629         if (alloc)
4630                 old_val += num_bytes;
4631         else
4632                 old_val -= num_bytes;
4633         btrfs_set_super_bytes_used(info->super_copy, old_val);
4634         spin_unlock(&info->delalloc_lock);
4635
4636         while (total) {
4637                 cache = btrfs_lookup_block_group(info, bytenr);
4638                 if (!cache)
4639                         return -ENOENT;
4640                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4641                                     BTRFS_BLOCK_GROUP_RAID1 |
4642                                     BTRFS_BLOCK_GROUP_RAID10))
4643                         factor = 2;
4644                 else
4645                         factor = 1;
4646                 /*
4647                  * If this block group has free space cache written out, we
4648                  * need to make sure to load it if we are removing space.  This
4649                  * is because we need the unpinning stage to actually add the
4650                  * space back to the block group, otherwise we will leak space.
4651                  */
4652                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4653                         cache_block_group(cache, trans, NULL, 1);
4654
4655                 byte_in_group = bytenr - cache->key.objectid;
4656                 WARN_ON(byte_in_group > cache->key.offset);
4657
4658                 spin_lock(&cache->space_info->lock);
4659                 spin_lock(&cache->lock);
4660
4661                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4662                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4663                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4664
4665                 cache->dirty = 1;
4666                 old_val = btrfs_block_group_used(&cache->item);
4667                 num_bytes = min(total, cache->key.offset - byte_in_group);
4668                 if (alloc) {
4669                         old_val += num_bytes;
4670                         btrfs_set_block_group_used(&cache->item, old_val);
4671                         cache->reserved -= num_bytes;
4672                         cache->space_info->bytes_reserved -= num_bytes;
4673                         cache->space_info->bytes_used += num_bytes;
4674                         cache->space_info->disk_used += num_bytes * factor;
4675                         spin_unlock(&cache->lock);
4676                         spin_unlock(&cache->space_info->lock);
4677                 } else {
4678                         old_val -= num_bytes;
4679                         btrfs_set_block_group_used(&cache->item, old_val);
4680                         cache->pinned += num_bytes;
4681                         cache->space_info->bytes_pinned += num_bytes;
4682                         cache->space_info->bytes_used -= num_bytes;
4683                         cache->space_info->disk_used -= num_bytes * factor;
4684                         spin_unlock(&cache->lock);
4685                         spin_unlock(&cache->space_info->lock);
4686
4687                         set_extent_dirty(info->pinned_extents,
4688                                          bytenr, bytenr + num_bytes - 1,
4689                                          GFP_NOFS | __GFP_NOFAIL);
4690                 }
4691                 btrfs_put_block_group(cache);
4692                 total -= num_bytes;
4693                 bytenr += num_bytes;
4694         }
4695         return 0;
4696 }
4697
4698 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4699 {
4700         struct btrfs_block_group_cache *cache;
4701         u64 bytenr;
4702
4703         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4704         if (!cache)
4705                 return 0;
4706
4707         bytenr = cache->key.objectid;
4708         btrfs_put_block_group(cache);
4709
4710         return bytenr;
4711 }
4712
4713 static int pin_down_extent(struct btrfs_root *root,
4714                            struct btrfs_block_group_cache *cache,
4715                            u64 bytenr, u64 num_bytes, int reserved)
4716 {
4717         spin_lock(&cache->space_info->lock);
4718         spin_lock(&cache->lock);
4719         cache->pinned += num_bytes;
4720         cache->space_info->bytes_pinned += num_bytes;
4721         if (reserved) {
4722                 cache->reserved -= num_bytes;
4723                 cache->space_info->bytes_reserved -= num_bytes;
4724         }
4725         spin_unlock(&cache->lock);
4726         spin_unlock(&cache->space_info->lock);
4727
4728         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4729                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4730         return 0;
4731 }
4732
4733 /*
4734  * this function must be called within transaction
4735  */
4736 int btrfs_pin_extent(struct btrfs_root *root,
4737                      u64 bytenr, u64 num_bytes, int reserved)
4738 {
4739         struct btrfs_block_group_cache *cache;
4740
4741         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4742         BUG_ON(!cache); /* Logic error */
4743
4744         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4745
4746         btrfs_put_block_group(cache);
4747         return 0;
4748 }
4749
4750 /*
4751  * this function must be called within transaction
4752  */
4753 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4754                                     struct btrfs_root *root,
4755                                     u64 bytenr, u64 num_bytes)
4756 {
4757         struct btrfs_block_group_cache *cache;
4758
4759         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4760         BUG_ON(!cache); /* Logic error */
4761
4762         /*
4763          * pull in the free space cache (if any) so that our pin
4764          * removes the free space from the cache.  We have load_only set
4765          * to one because the slow code to read in the free extents does check
4766          * the pinned extents.
4767          */
4768         cache_block_group(cache, trans, root, 1);
4769
4770         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4771
4772         /* remove us from the free space cache (if we're there at all) */
4773         btrfs_remove_free_space(cache, bytenr, num_bytes);
4774         btrfs_put_block_group(cache);
4775         return 0;
4776 }
4777
4778 /**
4779  * btrfs_update_reserved_bytes - update the block_group and space info counters
4780  * @cache:      The cache we are manipulating
4781  * @num_bytes:  The number of bytes in question
4782  * @reserve:    One of the reservation enums
4783  *
4784  * This is called by the allocator when it reserves space, or by somebody who is
4785  * freeing space that was never actually used on disk.  For example if you
4786  * reserve some space for a new leaf in transaction A and before transaction A
4787  * commits you free that leaf, you call this with reserve set to 0 in order to
4788  * clear the reservation.
4789  *
4790  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4791  * ENOSPC accounting.  For data we handle the reservation through clearing the
4792  * delalloc bits in the io_tree.  We have to do this since we could end up
4793  * allocating less disk space for the amount of data we have reserved in the
4794  * case of compression.
4795  *
4796  * If this is a reservation and the block group has become read only we cannot
4797  * make the reservation and return -EAGAIN, otherwise this function always
4798  * succeeds.
4799  */
4800 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4801                                        u64 num_bytes, int reserve)
4802 {
4803         struct btrfs_space_info *space_info = cache->space_info;
4804         int ret = 0;
4805
4806         spin_lock(&space_info->lock);
4807         spin_lock(&cache->lock);
4808         if (reserve != RESERVE_FREE) {
4809                 if (cache->ro) {
4810                         ret = -EAGAIN;
4811                 } else {
4812                         cache->reserved += num_bytes;
4813                         space_info->bytes_reserved += num_bytes;
4814                         if (reserve == RESERVE_ALLOC) {
4815                                 trace_btrfs_space_reservation(cache->fs_info,
4816                                                 "space_info", space_info->flags,
4817                                                 num_bytes, 0);
4818                                 space_info->bytes_may_use -= num_bytes;
4819                         }
4820                 }
4821         } else {
4822                 if (cache->ro)
4823                         space_info->bytes_readonly += num_bytes;
4824                 cache->reserved -= num_bytes;
4825                 space_info->bytes_reserved -= num_bytes;
4826                 space_info->reservation_progress++;
4827         }
4828         spin_unlock(&cache->lock);
4829         spin_unlock(&space_info->lock);
4830         return ret;
4831 }
4832
4833 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4834                                 struct btrfs_root *root)
4835 {
4836         struct btrfs_fs_info *fs_info = root->fs_info;
4837         struct btrfs_caching_control *next;
4838         struct btrfs_caching_control *caching_ctl;
4839         struct btrfs_block_group_cache *cache;
4840
4841         down_write(&fs_info->extent_commit_sem);
4842
4843         list_for_each_entry_safe(caching_ctl, next,
4844                                  &fs_info->caching_block_groups, list) {
4845                 cache = caching_ctl->block_group;
4846                 if (block_group_cache_done(cache)) {
4847                         cache->last_byte_to_unpin = (u64)-1;
4848                         list_del_init(&caching_ctl->list);
4849                         put_caching_control(caching_ctl);
4850                 } else {
4851                         cache->last_byte_to_unpin = caching_ctl->progress;
4852                 }
4853         }
4854
4855         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4856                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4857         else
4858                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4859
4860         up_write(&fs_info->extent_commit_sem);
4861
4862         update_global_block_rsv(fs_info);
4863 }
4864
4865 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4866 {
4867         struct btrfs_fs_info *fs_info = root->fs_info;
4868         struct btrfs_block_group_cache *cache = NULL;
4869         u64 len;
4870
4871         while (start <= end) {
4872                 if (!cache ||
4873                     start >= cache->key.objectid + cache->key.offset) {
4874                         if (cache)
4875                                 btrfs_put_block_group(cache);
4876                         cache = btrfs_lookup_block_group(fs_info, start);
4877                         BUG_ON(!cache); /* Logic error */
4878                 }
4879
4880                 len = cache->key.objectid + cache->key.offset - start;
4881                 len = min(len, end + 1 - start);
4882
4883                 if (start < cache->last_byte_to_unpin) {
4884                         len = min(len, cache->last_byte_to_unpin - start);
4885                         btrfs_add_free_space(cache, start, len);
4886                 }
4887
4888                 start += len;
4889
4890                 spin_lock(&cache->space_info->lock);
4891                 spin_lock(&cache->lock);
4892                 cache->pinned -= len;
4893                 cache->space_info->bytes_pinned -= len;
4894                 if (cache->ro)
4895                         cache->space_info->bytes_readonly += len;
4896                 spin_unlock(&cache->lock);
4897                 spin_unlock(&cache->space_info->lock);
4898         }
4899
4900         if (cache)
4901                 btrfs_put_block_group(cache);
4902         return 0;
4903 }
4904
4905 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4906                                struct btrfs_root *root)
4907 {
4908         struct btrfs_fs_info *fs_info = root->fs_info;
4909         struct extent_io_tree *unpin;
4910         u64 start;
4911         u64 end;
4912         int ret;
4913
4914         if (trans->aborted)
4915                 return 0;
4916
4917         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4918                 unpin = &fs_info->freed_extents[1];
4919         else
4920                 unpin = &fs_info->freed_extents[0];
4921
4922         while (1) {
4923                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4924                                             EXTENT_DIRTY);
4925                 if (ret)
4926                         break;
4927
4928                 if (btrfs_test_opt(root, DISCARD))
4929                         ret = btrfs_discard_extent(root, start,
4930                                                    end + 1 - start, NULL);
4931
4932                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4933                 unpin_extent_range(root, start, end);
4934                 cond_resched();
4935         }
4936
4937         return 0;
4938 }
4939
4940 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4941                                 struct btrfs_root *root,
4942                                 u64 bytenr, u64 num_bytes, u64 parent,
4943                                 u64 root_objectid, u64 owner_objectid,
4944                                 u64 owner_offset, int refs_to_drop,
4945                                 struct btrfs_delayed_extent_op *extent_op)
4946 {
4947         struct btrfs_key key;
4948         struct btrfs_path *path;
4949         struct btrfs_fs_info *info = root->fs_info;
4950         struct btrfs_root *extent_root = info->extent_root;
4951         struct extent_buffer *leaf;
4952         struct btrfs_extent_item *ei;
4953         struct btrfs_extent_inline_ref *iref;
4954         int ret;
4955         int is_data;
4956         int extent_slot = 0;
4957         int found_extent = 0;
4958         int num_to_del = 1;
4959         u32 item_size;
4960         u64 refs;
4961
4962         path = btrfs_alloc_path();
4963         if (!path)
4964                 return -ENOMEM;
4965
4966         path->reada = 1;
4967         path->leave_spinning = 1;
4968
4969         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4970         BUG_ON(!is_data && refs_to_drop != 1);
4971
4972         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4973                                     bytenr, num_bytes, parent,
4974                                     root_objectid, owner_objectid,
4975                                     owner_offset);
4976         if (ret == 0) {
4977                 extent_slot = path->slots[0];
4978                 while (extent_slot >= 0) {
4979                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4980                                               extent_slot);
4981                         if (key.objectid != bytenr)
4982                                 break;
4983                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4984                             key.offset == num_bytes) {
4985                                 found_extent = 1;
4986                                 break;
4987                         }
4988                         if (path->slots[0] - extent_slot > 5)
4989                                 break;
4990                         extent_slot--;
4991                 }
4992 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4993                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4994                 if (found_extent && item_size < sizeof(*ei))
4995                         found_extent = 0;
4996 #endif
4997                 if (!found_extent) {
4998                         BUG_ON(iref);
4999                         ret = remove_extent_backref(trans, extent_root, path,
5000                                                     NULL, refs_to_drop,
5001                                                     is_data);
5002                         if (ret)
5003                                 goto abort;
5004                         btrfs_release_path(path);
5005                         path->leave_spinning = 1;
5006
5007                         key.objectid = bytenr;
5008                         key.type = BTRFS_EXTENT_ITEM_KEY;
5009                         key.offset = num_bytes;
5010
5011                         ret = btrfs_search_slot(trans, extent_root,
5012                                                 &key, path, -1, 1);
5013                         if (ret) {
5014                                 printk(KERN_ERR "umm, got %d back from search"
5015                                        ", was looking for %llu\n", ret,
5016                                        (unsigned long long)bytenr);
5017                                 if (ret > 0)
5018                                         btrfs_print_leaf(extent_root,
5019                                                          path->nodes[0]);
5020                         }
5021                         if (ret < 0)
5022                                 goto abort;
5023                         extent_slot = path->slots[0];
5024                 }
5025         } else if (ret == -ENOENT) {
5026                 btrfs_print_leaf(extent_root, path->nodes[0]);
5027                 WARN_ON(1);
5028                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
5029                        "parent %llu root %llu  owner %llu offset %llu\n",
5030                        (unsigned long long)bytenr,
5031                        (unsigned long long)parent,
5032                        (unsigned long long)root_objectid,
5033                        (unsigned long long)owner_objectid,
5034                        (unsigned long long)owner_offset);
5035         } else {
5036                 goto abort;
5037         }
5038
5039         leaf = path->nodes[0];
5040         item_size = btrfs_item_size_nr(leaf, extent_slot);
5041 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5042         if (item_size < sizeof(*ei)) {
5043                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5044                 ret = convert_extent_item_v0(trans, extent_root, path,
5045                                              owner_objectid, 0);
5046                 if (ret < 0)
5047                         goto abort;
5048
5049                 btrfs_release_path(path);
5050                 path->leave_spinning = 1;
5051
5052                 key.objectid = bytenr;
5053                 key.type = BTRFS_EXTENT_ITEM_KEY;
5054                 key.offset = num_bytes;
5055
5056                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5057                                         -1, 1);
5058                 if (ret) {
5059                         printk(KERN_ERR "umm, got %d back from search"
5060                                ", was looking for %llu\n", ret,
5061                                (unsigned long long)bytenr);
5062                         btrfs_print_leaf(extent_root, path->nodes[0]);
5063                 }
5064                 if (ret < 0)
5065                         goto abort;
5066                 extent_slot = path->slots[0];
5067                 leaf = path->nodes[0];
5068                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5069         }
5070 #endif
5071         BUG_ON(item_size < sizeof(*ei));
5072         ei = btrfs_item_ptr(leaf, extent_slot,
5073                             struct btrfs_extent_item);
5074         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5075                 struct btrfs_tree_block_info *bi;
5076                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5077                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5078                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5079         }
5080
5081         refs = btrfs_extent_refs(leaf, ei);
5082         BUG_ON(refs < refs_to_drop);
5083         refs -= refs_to_drop;
5084
5085         if (refs > 0) {
5086                 if (extent_op)
5087                         __run_delayed_extent_op(extent_op, leaf, ei);
5088                 /*
5089                  * In the case of inline back ref, reference count will
5090                  * be updated by remove_extent_backref
5091                  */
5092                 if (iref) {
5093                         BUG_ON(!found_extent);
5094                 } else {
5095                         btrfs_set_extent_refs(leaf, ei, refs);
5096                         btrfs_mark_buffer_dirty(leaf);
5097                 }
5098                 if (found_extent) {
5099                         ret = remove_extent_backref(trans, extent_root, path,
5100                                                     iref, refs_to_drop,
5101                                                     is_data);
5102                         if (ret)
5103                                 goto abort;
5104                 }
5105         } else {
5106                 if (found_extent) {
5107                         BUG_ON(is_data && refs_to_drop !=
5108                                extent_data_ref_count(root, path, iref));
5109                         if (iref) {
5110                                 BUG_ON(path->slots[0] != extent_slot);
5111                         } else {
5112                                 BUG_ON(path->slots[0] != extent_slot + 1);
5113                                 path->slots[0] = extent_slot;
5114                                 num_to_del = 2;
5115                         }
5116                 }
5117
5118                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5119                                       num_to_del);
5120                 if (ret)
5121                         goto abort;
5122                 btrfs_release_path(path);
5123
5124                 if (is_data) {
5125                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5126                         if (ret)
5127                                 goto abort;
5128                 }
5129
5130                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
5131                 if (ret)
5132                         goto abort;
5133         }
5134 out:
5135         btrfs_free_path(path);
5136         return ret;
5137
5138 abort:
5139         btrfs_abort_transaction(trans, extent_root, ret);
5140         goto out;
5141 }
5142
5143 /*
5144  * when we free an block, it is possible (and likely) that we free the last
5145  * delayed ref for that extent as well.  This searches the delayed ref tree for
5146  * a given extent, and if there are no other delayed refs to be processed, it
5147  * removes it from the tree.
5148  */
5149 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5150                                       struct btrfs_root *root, u64 bytenr)
5151 {
5152         struct btrfs_delayed_ref_head *head;
5153         struct btrfs_delayed_ref_root *delayed_refs;
5154         struct btrfs_delayed_ref_node *ref;
5155         struct rb_node *node;
5156         int ret = 0;
5157
5158         delayed_refs = &trans->transaction->delayed_refs;
5159         spin_lock(&delayed_refs->lock);
5160         head = btrfs_find_delayed_ref_head(trans, bytenr);
5161         if (!head)
5162                 goto out;
5163
5164         node = rb_prev(&head->node.rb_node);
5165         if (!node)
5166                 goto out;
5167
5168         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5169
5170         /* there are still entries for this ref, we can't drop it */
5171         if (ref->bytenr == bytenr)
5172                 goto out;
5173
5174         if (head->extent_op) {
5175                 if (!head->must_insert_reserved)
5176                         goto out;
5177                 kfree(head->extent_op);
5178                 head->extent_op = NULL;
5179         }
5180
5181         /*
5182          * waiting for the lock here would deadlock.  If someone else has it
5183          * locked they are already in the process of dropping it anyway
5184          */
5185         if (!mutex_trylock(&head->mutex))
5186                 goto out;
5187
5188         /*
5189          * at this point we have a head with no other entries.  Go
5190          * ahead and process it.
5191          */
5192         head->node.in_tree = 0;
5193         rb_erase(&head->node.rb_node, &delayed_refs->root);
5194
5195         delayed_refs->num_entries--;
5196         if (waitqueue_active(&delayed_refs->seq_wait))
5197                 wake_up(&delayed_refs->seq_wait);
5198
5199         /*
5200          * we don't take a ref on the node because we're removing it from the
5201          * tree, so we just steal the ref the tree was holding.
5202          */
5203         delayed_refs->num_heads--;
5204         if (list_empty(&head->cluster))
5205                 delayed_refs->num_heads_ready--;
5206
5207         list_del_init(&head->cluster);
5208         spin_unlock(&delayed_refs->lock);
5209
5210         BUG_ON(head->extent_op);
5211         if (head->must_insert_reserved)
5212                 ret = 1;
5213
5214         mutex_unlock(&head->mutex);
5215         btrfs_put_delayed_ref(&head->node);
5216         return ret;
5217 out:
5218         spin_unlock(&delayed_refs->lock);
5219         return 0;
5220 }
5221
5222 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5223                            struct btrfs_root *root,
5224                            struct extent_buffer *buf,
5225                            u64 parent, int last_ref)
5226 {
5227         struct btrfs_block_group_cache *cache = NULL;
5228         int ret;
5229
5230         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5231                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5232                                         buf->start, buf->len,
5233                                         parent, root->root_key.objectid,
5234                                         btrfs_header_level(buf),
5235                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5236                 BUG_ON(ret); /* -ENOMEM */
5237         }
5238
5239         if (!last_ref)
5240                 return;
5241
5242         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5243
5244         if (btrfs_header_generation(buf) == trans->transid) {
5245                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5246                         ret = check_ref_cleanup(trans, root, buf->start);
5247                         if (!ret)
5248                                 goto out;
5249                 }
5250
5251                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5252                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5253                         goto out;
5254                 }
5255
5256                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5257
5258                 btrfs_add_free_space(cache, buf->start, buf->len);
5259                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5260         }
5261 out:
5262         /*
5263          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5264          * anymore.
5265          */
5266         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5267         btrfs_put_block_group(cache);
5268 }
5269
5270 /* Can return -ENOMEM */
5271 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5272                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5273                       u64 owner, u64 offset, int for_cow)
5274 {
5275         int ret;
5276         struct btrfs_fs_info *fs_info = root->fs_info;
5277
5278         /*
5279          * tree log blocks never actually go into the extent allocation
5280          * tree, just update pinning info and exit early.
5281          */
5282         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5283                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5284                 /* unlocks the pinned mutex */
5285                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5286                 ret = 0;
5287         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5288                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5289                                         num_bytes,
5290                                         parent, root_objectid, (int)owner,
5291                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5292         } else {
5293                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5294                                                 num_bytes,
5295                                                 parent, root_objectid, owner,
5296                                                 offset, BTRFS_DROP_DELAYED_REF,
5297                                                 NULL, for_cow);
5298         }
5299         return ret;
5300 }
5301
5302 static u64 stripe_align(struct btrfs_root *root, u64 val)
5303 {
5304         u64 mask = ((u64)root->stripesize - 1);
5305         u64 ret = (val + mask) & ~mask;
5306         return ret;
5307 }
5308
5309 /*
5310  * when we wait for progress in the block group caching, its because
5311  * our allocation attempt failed at least once.  So, we must sleep
5312  * and let some progress happen before we try again.
5313  *
5314  * This function will sleep at least once waiting for new free space to
5315  * show up, and then it will check the block group free space numbers
5316  * for our min num_bytes.  Another option is to have it go ahead
5317  * and look in the rbtree for a free extent of a given size, but this
5318  * is a good start.
5319  */
5320 static noinline int
5321 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5322                                 u64 num_bytes)
5323 {
5324         struct btrfs_caching_control *caching_ctl;
5325         DEFINE_WAIT(wait);
5326
5327         caching_ctl = get_caching_control(cache);
5328         if (!caching_ctl)
5329                 return 0;
5330
5331         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5332                    (cache->free_space_ctl->free_space >= num_bytes));
5333
5334         put_caching_control(caching_ctl);
5335         return 0;
5336 }
5337
5338 static noinline int
5339 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5340 {
5341         struct btrfs_caching_control *caching_ctl;
5342         DEFINE_WAIT(wait);
5343
5344         caching_ctl = get_caching_control(cache);
5345         if (!caching_ctl)
5346                 return 0;
5347
5348         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5349
5350         put_caching_control(caching_ctl);
5351         return 0;
5352 }
5353
5354 static int __get_block_group_index(u64 flags)
5355 {
5356         int index;
5357
5358         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5359                 index = 0;
5360         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5361                 index = 1;
5362         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5363                 index = 2;
5364         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5365                 index = 3;
5366         else
5367                 index = 4;
5368
5369         return index;
5370 }
5371
5372 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5373 {
5374         return __get_block_group_index(cache->flags);
5375 }
5376
5377 enum btrfs_loop_type {
5378         LOOP_CACHING_NOWAIT = 0,
5379         LOOP_CACHING_WAIT = 1,
5380         LOOP_ALLOC_CHUNK = 2,
5381         LOOP_NO_EMPTY_SIZE = 3,
5382 };
5383
5384 /*
5385  * walks the btree of allocated extents and find a hole of a given size.
5386  * The key ins is changed to record the hole:
5387  * ins->objectid == block start
5388  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5389  * ins->offset == number of blocks
5390  * Any available blocks before search_start are skipped.
5391  */
5392 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5393                                      struct btrfs_root *orig_root,
5394                                      u64 num_bytes, u64 empty_size,
5395                                      u64 hint_byte, struct btrfs_key *ins,
5396                                      u64 data)
5397 {
5398         int ret = 0;
5399         struct btrfs_root *root = orig_root->fs_info->extent_root;
5400         struct btrfs_free_cluster *last_ptr = NULL;
5401         struct btrfs_block_group_cache *block_group = NULL;
5402         struct btrfs_block_group_cache *used_block_group;
5403         u64 search_start = 0;
5404         int empty_cluster = 2 * 1024 * 1024;
5405         int allowed_chunk_alloc = 0;
5406         int done_chunk_alloc = 0;
5407         struct btrfs_space_info *space_info;
5408         int loop = 0;
5409         int index = 0;
5410         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5411                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5412         bool found_uncached_bg = false;
5413         bool failed_cluster_refill = false;
5414         bool failed_alloc = false;
5415         bool use_cluster = true;
5416         bool have_caching_bg = false;
5417
5418         WARN_ON(num_bytes < root->sectorsize);
5419         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5420         ins->objectid = 0;
5421         ins->offset = 0;
5422
5423         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5424
5425         space_info = __find_space_info(root->fs_info, data);
5426         if (!space_info) {
5427                 printk(KERN_ERR "No space info for %llu\n", data);
5428                 return -ENOSPC;
5429         }
5430
5431         /*
5432          * If the space info is for both data and metadata it means we have a
5433          * small filesystem and we can't use the clustering stuff.
5434          */
5435         if (btrfs_mixed_space_info(space_info))
5436                 use_cluster = false;
5437
5438         if (orig_root->ref_cows || empty_size)
5439                 allowed_chunk_alloc = 1;
5440
5441         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5442                 last_ptr = &root->fs_info->meta_alloc_cluster;
5443                 if (!btrfs_test_opt(root, SSD))
5444                         empty_cluster = 64 * 1024;
5445         }
5446
5447         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5448             btrfs_test_opt(root, SSD)) {
5449                 last_ptr = &root->fs_info->data_alloc_cluster;
5450         }
5451
5452         if (last_ptr) {
5453                 spin_lock(&last_ptr->lock);
5454                 if (last_ptr->block_group)
5455                         hint_byte = last_ptr->window_start;
5456                 spin_unlock(&last_ptr->lock);
5457         }
5458
5459         search_start = max(search_start, first_logical_byte(root, 0));
5460         search_start = max(search_start, hint_byte);
5461
5462         if (!last_ptr)
5463                 empty_cluster = 0;
5464
5465         if (search_start == hint_byte) {
5466                 block_group = btrfs_lookup_block_group(root->fs_info,
5467                                                        search_start);
5468                 used_block_group = block_group;
5469                 /*
5470                  * we don't want to use the block group if it doesn't match our
5471                  * allocation bits, or if its not cached.
5472                  *
5473                  * However if we are re-searching with an ideal block group
5474                  * picked out then we don't care that the block group is cached.
5475                  */
5476                 if (block_group && block_group_bits(block_group, data) &&
5477                     block_group->cached != BTRFS_CACHE_NO) {
5478                         down_read(&space_info->groups_sem);
5479                         if (list_empty(&block_group->list) ||
5480                             block_group->ro) {
5481                                 /*
5482                                  * someone is removing this block group,
5483                                  * we can't jump into the have_block_group
5484                                  * target because our list pointers are not
5485                                  * valid
5486                                  */
5487                                 btrfs_put_block_group(block_group);
5488                                 up_read(&space_info->groups_sem);
5489                         } else {
5490                                 index = get_block_group_index(block_group);
5491                                 goto have_block_group;
5492                         }
5493                 } else if (block_group) {
5494                         btrfs_put_block_group(block_group);
5495                 }
5496         }
5497 search:
5498         have_caching_bg = false;
5499         down_read(&space_info->groups_sem);
5500         list_for_each_entry(block_group, &space_info->block_groups[index],
5501                             list) {
5502                 u64 offset;
5503                 int cached;
5504
5505                 used_block_group = block_group;
5506                 btrfs_get_block_group(block_group);
5507                 search_start = block_group->key.objectid;
5508
5509                 /*
5510                  * this can happen if we end up cycling through all the
5511                  * raid types, but we want to make sure we only allocate
5512                  * for the proper type.
5513                  */
5514                 if (!block_group_bits(block_group, data)) {
5515                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5516                                 BTRFS_BLOCK_GROUP_RAID1 |
5517                                 BTRFS_BLOCK_GROUP_RAID10;
5518
5519                         /*
5520                          * if they asked for extra copies and this block group
5521                          * doesn't provide them, bail.  This does allow us to
5522                          * fill raid0 from raid1.
5523                          */
5524                         if ((data & extra) && !(block_group->flags & extra))
5525                                 goto loop;
5526                 }
5527
5528 have_block_group:
5529                 cached = block_group_cache_done(block_group);
5530                 if (unlikely(!cached)) {
5531                         found_uncached_bg = true;
5532                         ret = cache_block_group(block_group, trans,
5533                                                 orig_root, 0);
5534                         BUG_ON(ret < 0);
5535                         ret = 0;
5536                 }
5537
5538                 if (unlikely(block_group->ro))
5539                         goto loop;
5540
5541                 /*
5542                  * Ok we want to try and use the cluster allocator, so
5543                  * lets look there
5544                  */
5545                 if (last_ptr) {
5546                         /*
5547                          * the refill lock keeps out other
5548                          * people trying to start a new cluster
5549                          */
5550                         spin_lock(&last_ptr->refill_lock);
5551                         used_block_group = last_ptr->block_group;
5552                         if (used_block_group != block_group &&
5553                             (!used_block_group ||
5554                              used_block_group->ro ||
5555                              !block_group_bits(used_block_group, data))) {
5556                                 used_block_group = block_group;
5557                                 goto refill_cluster;
5558                         }
5559
5560                         if (used_block_group != block_group)
5561                                 btrfs_get_block_group(used_block_group);
5562
5563                         offset = btrfs_alloc_from_cluster(used_block_group,
5564                           last_ptr, num_bytes, used_block_group->key.objectid);
5565                         if (offset) {
5566                                 /* we have a block, we're done */
5567                                 spin_unlock(&last_ptr->refill_lock);
5568                                 trace_btrfs_reserve_extent_cluster(root,
5569                                         block_group, search_start, num_bytes);
5570                                 goto checks;
5571                         }
5572
5573                         WARN_ON(last_ptr->block_group != used_block_group);
5574                         if (used_block_group != block_group) {
5575                                 btrfs_put_block_group(used_block_group);
5576                                 used_block_group = block_group;
5577                         }
5578 refill_cluster:
5579                         BUG_ON(used_block_group != block_group);
5580                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5581                          * set up a new clusters, so lets just skip it
5582                          * and let the allocator find whatever block
5583                          * it can find.  If we reach this point, we
5584                          * will have tried the cluster allocator
5585                          * plenty of times and not have found
5586                          * anything, so we are likely way too
5587                          * fragmented for the clustering stuff to find
5588                          * anything.
5589                          *
5590                          * However, if the cluster is taken from the
5591                          * current block group, release the cluster
5592                          * first, so that we stand a better chance of
5593                          * succeeding in the unclustered
5594                          * allocation.  */
5595                         if (loop >= LOOP_NO_EMPTY_SIZE &&
5596                             last_ptr->block_group != block_group) {
5597                                 spin_unlock(&last_ptr->refill_lock);
5598                                 goto unclustered_alloc;
5599                         }
5600
5601                         /*
5602                          * this cluster didn't work out, free it and
5603                          * start over
5604                          */
5605                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5606
5607                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5608                                 spin_unlock(&last_ptr->refill_lock);
5609                                 goto unclustered_alloc;
5610                         }
5611
5612                         /* allocate a cluster in this block group */
5613                         ret = btrfs_find_space_cluster(trans, root,
5614                                                block_group, last_ptr,
5615                                                search_start, num_bytes,
5616                                                empty_cluster + empty_size);
5617                         if (ret == 0) {
5618                                 /*
5619                                  * now pull our allocation out of this
5620                                  * cluster
5621                                  */
5622                                 offset = btrfs_alloc_from_cluster(block_group,
5623                                                   last_ptr, num_bytes,
5624                                                   search_start);
5625                                 if (offset) {
5626                                         /* we found one, proceed */
5627                                         spin_unlock(&last_ptr->refill_lock);
5628                                         trace_btrfs_reserve_extent_cluster(root,
5629                                                 block_group, search_start,
5630                                                 num_bytes);
5631                                         goto checks;
5632                                 }
5633                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5634                                    && !failed_cluster_refill) {
5635                                 spin_unlock(&last_ptr->refill_lock);
5636
5637                                 failed_cluster_refill = true;
5638                                 wait_block_group_cache_progress(block_group,
5639                                        num_bytes + empty_cluster + empty_size);
5640                                 goto have_block_group;
5641                         }
5642
5643                         /*
5644                          * at this point we either didn't find a cluster
5645                          * or we weren't able to allocate a block from our
5646                          * cluster.  Free the cluster we've been trying
5647                          * to use, and go to the next block group
5648                          */
5649                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5650                         spin_unlock(&last_ptr->refill_lock);
5651                         goto loop;
5652                 }
5653
5654 unclustered_alloc:
5655                 spin_lock(&block_group->free_space_ctl->tree_lock);
5656                 if (cached &&
5657                     block_group->free_space_ctl->free_space <
5658                     num_bytes + empty_cluster + empty_size) {
5659                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5660                         goto loop;
5661                 }
5662                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5663
5664                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5665                                                     num_bytes, empty_size);
5666                 /*
5667                  * If we didn't find a chunk, and we haven't failed on this
5668                  * block group before, and this block group is in the middle of
5669                  * caching and we are ok with waiting, then go ahead and wait
5670                  * for progress to be made, and set failed_alloc to true.
5671                  *
5672                  * If failed_alloc is true then we've already waited on this
5673                  * block group once and should move on to the next block group.
5674                  */
5675                 if (!offset && !failed_alloc && !cached &&
5676                     loop > LOOP_CACHING_NOWAIT) {
5677                         wait_block_group_cache_progress(block_group,
5678                                                 num_bytes + empty_size);
5679                         failed_alloc = true;
5680                         goto have_block_group;
5681                 } else if (!offset) {
5682                         if (!cached)
5683                                 have_caching_bg = true;
5684                         goto loop;
5685                 }
5686 checks:
5687                 search_start = stripe_align(root, offset);
5688
5689                 /* move on to the next group */
5690                 if (search_start + num_bytes >
5691                     used_block_group->key.objectid + used_block_group->key.offset) {
5692                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5693                         goto loop;
5694                 }
5695
5696                 if (offset < search_start)
5697                         btrfs_add_free_space(used_block_group, offset,
5698                                              search_start - offset);
5699                 BUG_ON(offset > search_start);
5700
5701                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
5702                                                   alloc_type);
5703                 if (ret == -EAGAIN) {
5704                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5705                         goto loop;
5706                 }
5707
5708                 /* we are all good, lets return */
5709                 ins->objectid = search_start;
5710                 ins->offset = num_bytes;
5711
5712                 trace_btrfs_reserve_extent(orig_root, block_group,
5713                                            search_start, num_bytes);
5714                 if (offset < search_start)
5715                         btrfs_add_free_space(used_block_group, offset,
5716                                              search_start - offset);
5717                 BUG_ON(offset > search_start);
5718                 if (used_block_group != block_group)
5719                         btrfs_put_block_group(used_block_group);
5720                 btrfs_put_block_group(block_group);
5721                 break;
5722 loop:
5723                 failed_cluster_refill = false;
5724                 failed_alloc = false;
5725                 BUG_ON(index != get_block_group_index(block_group));
5726                 if (used_block_group != block_group)
5727                         btrfs_put_block_group(used_block_group);
5728                 btrfs_put_block_group(block_group);
5729         }
5730         up_read(&space_info->groups_sem);
5731
5732         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
5733                 goto search;
5734
5735         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5736                 goto search;
5737
5738         /*
5739          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5740          *                      caching kthreads as we move along
5741          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5742          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5743          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5744          *                      again
5745          */
5746         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5747                 index = 0;
5748                 loop++;
5749                 if (loop == LOOP_ALLOC_CHUNK) {
5750                        if (allowed_chunk_alloc) {
5751                                 ret = do_chunk_alloc(trans, root, num_bytes +
5752                                                      2 * 1024 * 1024, data,
5753                                                      CHUNK_ALLOC_LIMITED);
5754                                 /*
5755                                  * Do not bail out on ENOSPC since we
5756                                  * can do more things.
5757                                  */
5758                                 if (ret < 0 && ret != -ENOSPC) {
5759                                         btrfs_abort_transaction(trans,
5760                                                                 root, ret);
5761                                         goto out;
5762                                 }
5763                                 allowed_chunk_alloc = 0;
5764                                 if (ret == 1)
5765                                         done_chunk_alloc = 1;
5766                         } else if (!done_chunk_alloc &&
5767                                    space_info->force_alloc ==
5768                                    CHUNK_ALLOC_NO_FORCE) {
5769                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5770                         }
5771
5772                        /*
5773                         * We didn't allocate a chunk, go ahead and drop the
5774                         * empty size and loop again.
5775                         */
5776                        if (!done_chunk_alloc)
5777                                loop = LOOP_NO_EMPTY_SIZE;
5778                 }
5779
5780                 if (loop == LOOP_NO_EMPTY_SIZE) {
5781                         empty_size = 0;
5782                         empty_cluster = 0;
5783                 }
5784
5785                 goto search;
5786         } else if (!ins->objectid) {
5787                 ret = -ENOSPC;
5788         } else if (ins->objectid) {
5789                 ret = 0;
5790         }
5791 out:
5792
5793         return ret;
5794 }
5795
5796 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5797                             int dump_block_groups)
5798 {
5799         struct btrfs_block_group_cache *cache;
5800         int index = 0;
5801
5802         spin_lock(&info->lock);
5803         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5804                (unsigned long long)info->flags,
5805                (unsigned long long)(info->total_bytes - info->bytes_used -
5806                                     info->bytes_pinned - info->bytes_reserved -
5807                                     info->bytes_readonly),
5808                (info->full) ? "" : "not ");
5809         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5810                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5811                (unsigned long long)info->total_bytes,
5812                (unsigned long long)info->bytes_used,
5813                (unsigned long long)info->bytes_pinned,
5814                (unsigned long long)info->bytes_reserved,
5815                (unsigned long long)info->bytes_may_use,
5816                (unsigned long long)info->bytes_readonly);
5817         spin_unlock(&info->lock);
5818
5819         if (!dump_block_groups)
5820                 return;
5821
5822         down_read(&info->groups_sem);
5823 again:
5824         list_for_each_entry(cache, &info->block_groups[index], list) {
5825                 spin_lock(&cache->lock);
5826                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s\n",
5827                        (unsigned long long)cache->key.objectid,
5828                        (unsigned long long)cache->key.offset,
5829                        (unsigned long long)btrfs_block_group_used(&cache->item),
5830                        (unsigned long long)cache->pinned,
5831                        (unsigned long long)cache->reserved,
5832                        cache->ro ? "[readonly]" : "");
5833                 btrfs_dump_free_space(cache, bytes);
5834                 spin_unlock(&cache->lock);
5835         }
5836         if (++index < BTRFS_NR_RAID_TYPES)
5837                 goto again;
5838         up_read(&info->groups_sem);
5839 }
5840
5841 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5842                          struct btrfs_root *root,
5843                          u64 num_bytes, u64 min_alloc_size,
5844                          u64 empty_size, u64 hint_byte,
5845                          struct btrfs_key *ins, u64 data)
5846 {
5847         bool final_tried = false;
5848         int ret;
5849
5850         data = btrfs_get_alloc_profile(root, data);
5851 again:
5852         /*
5853          * the only place that sets empty_size is btrfs_realloc_node, which
5854          * is not called recursively on allocations
5855          */
5856         if (empty_size || root->ref_cows) {
5857                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5858                                      num_bytes + 2 * 1024 * 1024, data,
5859                                      CHUNK_ALLOC_NO_FORCE);
5860                 if (ret < 0 && ret != -ENOSPC) {
5861                         btrfs_abort_transaction(trans, root, ret);
5862                         return ret;
5863                 }
5864         }
5865
5866         WARN_ON(num_bytes < root->sectorsize);
5867         ret = find_free_extent(trans, root, num_bytes, empty_size,
5868                                hint_byte, ins, data);
5869
5870         if (ret == -ENOSPC) {
5871                 if (!final_tried) {
5872                         num_bytes = num_bytes >> 1;
5873                         num_bytes = num_bytes & ~(root->sectorsize - 1);
5874                         num_bytes = max(num_bytes, min_alloc_size);
5875                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5876                                        num_bytes, data, CHUNK_ALLOC_FORCE);
5877                         if (ret < 0 && ret != -ENOSPC) {
5878                                 btrfs_abort_transaction(trans, root, ret);
5879                                 return ret;
5880                         }
5881                         if (num_bytes == min_alloc_size)
5882                                 final_tried = true;
5883                         goto again;
5884                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
5885                         struct btrfs_space_info *sinfo;
5886
5887                         sinfo = __find_space_info(root->fs_info, data);
5888                         printk(KERN_ERR "btrfs allocation failed flags %llu, "
5889                                "wanted %llu\n", (unsigned long long)data,
5890                                (unsigned long long)num_bytes);
5891                         if (sinfo)
5892                                 dump_space_info(sinfo, num_bytes, 1);
5893                 }
5894         }
5895
5896         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
5897
5898         return ret;
5899 }
5900
5901 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
5902                                         u64 start, u64 len, int pin)
5903 {
5904         struct btrfs_block_group_cache *cache;
5905         int ret = 0;
5906
5907         cache = btrfs_lookup_block_group(root->fs_info, start);
5908         if (!cache) {
5909                 printk(KERN_ERR "Unable to find block group for %llu\n",
5910                        (unsigned long long)start);
5911                 return -ENOSPC;
5912         }
5913
5914         if (btrfs_test_opt(root, DISCARD))
5915                 ret = btrfs_discard_extent(root, start, len, NULL);
5916
5917         if (pin)
5918                 pin_down_extent(root, cache, start, len, 1);
5919         else {
5920                 btrfs_add_free_space(cache, start, len);
5921                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
5922         }
5923         btrfs_put_block_group(cache);
5924
5925         trace_btrfs_reserved_extent_free(root, start, len);
5926
5927         return ret;
5928 }
5929
5930 int btrfs_free_reserved_extent(struct btrfs_root *root,
5931                                         u64 start, u64 len)
5932 {
5933         return __btrfs_free_reserved_extent(root, start, len, 0);
5934 }
5935
5936 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
5937                                        u64 start, u64 len)
5938 {
5939         return __btrfs_free_reserved_extent(root, start, len, 1);
5940 }
5941
5942 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5943                                       struct btrfs_root *root,
5944                                       u64 parent, u64 root_objectid,
5945                                       u64 flags, u64 owner, u64 offset,
5946                                       struct btrfs_key *ins, int ref_mod)
5947 {
5948         int ret;
5949         struct btrfs_fs_info *fs_info = root->fs_info;
5950         struct btrfs_extent_item *extent_item;
5951         struct btrfs_extent_inline_ref *iref;
5952         struct btrfs_path *path;
5953         struct extent_buffer *leaf;
5954         int type;
5955         u32 size;
5956
5957         if (parent > 0)
5958                 type = BTRFS_SHARED_DATA_REF_KEY;
5959         else
5960                 type = BTRFS_EXTENT_DATA_REF_KEY;
5961
5962         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5963
5964         path = btrfs_alloc_path();
5965         if (!path)
5966                 return -ENOMEM;
5967
5968         path->leave_spinning = 1;
5969         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5970                                       ins, size);
5971         if (ret) {
5972                 btrfs_free_path(path);
5973                 return ret;
5974         }
5975
5976         leaf = path->nodes[0];
5977         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5978                                      struct btrfs_extent_item);
5979         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
5980         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5981         btrfs_set_extent_flags(leaf, extent_item,
5982                                flags | BTRFS_EXTENT_FLAG_DATA);
5983
5984         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
5985         btrfs_set_extent_inline_ref_type(leaf, iref, type);
5986         if (parent > 0) {
5987                 struct btrfs_shared_data_ref *ref;
5988                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
5989                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5990                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
5991         } else {
5992                 struct btrfs_extent_data_ref *ref;
5993                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
5994                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
5995                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
5996                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
5997                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
5998         }
5999
6000         btrfs_mark_buffer_dirty(path->nodes[0]);
6001         btrfs_free_path(path);
6002
6003         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
6004         if (ret) { /* -ENOENT, logic error */
6005                 printk(KERN_ERR "btrfs update block group failed for %llu "
6006                        "%llu\n", (unsigned long long)ins->objectid,
6007                        (unsigned long long)ins->offset);
6008                 BUG();
6009         }
6010         return ret;
6011 }
6012
6013 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6014                                      struct btrfs_root *root,
6015                                      u64 parent, u64 root_objectid,
6016                                      u64 flags, struct btrfs_disk_key *key,
6017                                      int level, struct btrfs_key *ins)
6018 {
6019         int ret;
6020         struct btrfs_fs_info *fs_info = root->fs_info;
6021         struct btrfs_extent_item *extent_item;
6022         struct btrfs_tree_block_info *block_info;
6023         struct btrfs_extent_inline_ref *iref;
6024         struct btrfs_path *path;
6025         struct extent_buffer *leaf;
6026         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
6027
6028         path = btrfs_alloc_path();
6029         if (!path)
6030                 return -ENOMEM;
6031
6032         path->leave_spinning = 1;
6033         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6034                                       ins, size);
6035         if (ret) {
6036                 btrfs_free_path(path);
6037                 return ret;
6038         }
6039
6040         leaf = path->nodes[0];
6041         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6042                                      struct btrfs_extent_item);
6043         btrfs_set_extent_refs(leaf, extent_item, 1);
6044         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6045         btrfs_set_extent_flags(leaf, extent_item,
6046                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6047         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6048
6049         btrfs_set_tree_block_key(leaf, block_info, key);
6050         btrfs_set_tree_block_level(leaf, block_info, level);
6051
6052         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6053         if (parent > 0) {
6054                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6055                 btrfs_set_extent_inline_ref_type(leaf, iref,
6056                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6057                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6058         } else {
6059                 btrfs_set_extent_inline_ref_type(leaf, iref,
6060                                                  BTRFS_TREE_BLOCK_REF_KEY);
6061                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6062         }
6063
6064         btrfs_mark_buffer_dirty(leaf);
6065         btrfs_free_path(path);
6066
6067         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
6068         if (ret) { /* -ENOENT, logic error */
6069                 printk(KERN_ERR "btrfs update block group failed for %llu "
6070                        "%llu\n", (unsigned long long)ins->objectid,
6071                        (unsigned long long)ins->offset);
6072                 BUG();
6073         }
6074         return ret;
6075 }
6076
6077 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6078                                      struct btrfs_root *root,
6079                                      u64 root_objectid, u64 owner,
6080                                      u64 offset, struct btrfs_key *ins)
6081 {
6082         int ret;
6083
6084         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6085
6086         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6087                                          ins->offset, 0,
6088                                          root_objectid, owner, offset,
6089                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6090         return ret;
6091 }
6092
6093 /*
6094  * this is used by the tree logging recovery code.  It records that
6095  * an extent has been allocated and makes sure to clear the free
6096  * space cache bits as well
6097  */
6098 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6099                                    struct btrfs_root *root,
6100                                    u64 root_objectid, u64 owner, u64 offset,
6101                                    struct btrfs_key *ins)
6102 {
6103         int ret;
6104         struct btrfs_block_group_cache *block_group;
6105         struct btrfs_caching_control *caching_ctl;
6106         u64 start = ins->objectid;
6107         u64 num_bytes = ins->offset;
6108
6109         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6110         cache_block_group(block_group, trans, NULL, 0);
6111         caching_ctl = get_caching_control(block_group);
6112
6113         if (!caching_ctl) {
6114                 BUG_ON(!block_group_cache_done(block_group));
6115                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6116                 BUG_ON(ret); /* -ENOMEM */
6117         } else {
6118                 mutex_lock(&caching_ctl->mutex);
6119
6120                 if (start >= caching_ctl->progress) {
6121                         ret = add_excluded_extent(root, start, num_bytes);
6122                         BUG_ON(ret); /* -ENOMEM */
6123                 } else if (start + num_bytes <= caching_ctl->progress) {
6124                         ret = btrfs_remove_free_space(block_group,
6125                                                       start, num_bytes);
6126                         BUG_ON(ret); /* -ENOMEM */
6127                 } else {
6128                         num_bytes = caching_ctl->progress - start;
6129                         ret = btrfs_remove_free_space(block_group,
6130                                                       start, num_bytes);
6131                         BUG_ON(ret); /* -ENOMEM */
6132
6133                         start = caching_ctl->progress;
6134                         num_bytes = ins->objectid + ins->offset -
6135                                     caching_ctl->progress;
6136                         ret = add_excluded_extent(root, start, num_bytes);
6137                         BUG_ON(ret); /* -ENOMEM */
6138                 }
6139
6140                 mutex_unlock(&caching_ctl->mutex);
6141                 put_caching_control(caching_ctl);
6142         }
6143
6144         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6145                                           RESERVE_ALLOC_NO_ACCOUNT);
6146         BUG_ON(ret); /* logic error */
6147         btrfs_put_block_group(block_group);
6148         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6149                                          0, owner, offset, ins, 1);
6150         return ret;
6151 }
6152
6153 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
6154                                             struct btrfs_root *root,
6155                                             u64 bytenr, u32 blocksize,
6156                                             int level)
6157 {
6158         struct extent_buffer *buf;
6159
6160         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6161         if (!buf)
6162                 return ERR_PTR(-ENOMEM);
6163         btrfs_set_header_generation(buf, trans->transid);
6164         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6165         btrfs_tree_lock(buf);
6166         clean_tree_block(trans, root, buf);
6167         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6168
6169         btrfs_set_lock_blocking(buf);
6170         btrfs_set_buffer_uptodate(buf);
6171
6172         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6173                 /*
6174                  * we allow two log transactions at a time, use different
6175                  * EXENT bit to differentiate dirty pages.
6176                  */
6177                 if (root->log_transid % 2 == 0)
6178                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6179                                         buf->start + buf->len - 1, GFP_NOFS);
6180                 else
6181                         set_extent_new(&root->dirty_log_pages, buf->start,
6182                                         buf->start + buf->len - 1, GFP_NOFS);
6183         } else {
6184                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6185                          buf->start + buf->len - 1, GFP_NOFS);
6186         }
6187         trans->blocks_used++;
6188         /* this returns a buffer locked for blocking */
6189         return buf;
6190 }
6191
6192 static struct btrfs_block_rsv *
6193 use_block_rsv(struct btrfs_trans_handle *trans,
6194               struct btrfs_root *root, u32 blocksize)
6195 {
6196         struct btrfs_block_rsv *block_rsv;
6197         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6198         int ret;
6199
6200         block_rsv = get_block_rsv(trans, root);
6201
6202         if (block_rsv->size == 0) {
6203                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
6204                 /*
6205                  * If we couldn't reserve metadata bytes try and use some from
6206                  * the global reserve.
6207                  */
6208                 if (ret && block_rsv != global_rsv) {
6209                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6210                         if (!ret)
6211                                 return global_rsv;
6212                         return ERR_PTR(ret);
6213                 } else if (ret) {
6214                         return ERR_PTR(ret);
6215                 }
6216                 return block_rsv;
6217         }
6218
6219         ret = block_rsv_use_bytes(block_rsv, blocksize);
6220         if (!ret)
6221                 return block_rsv;
6222         if (ret) {
6223                 static DEFINE_RATELIMIT_STATE(_rs,
6224                                 DEFAULT_RATELIMIT_INTERVAL,
6225                                 /*DEFAULT_RATELIMIT_BURST*/ 2);
6226                 if (__ratelimit(&_rs)) {
6227                         printk(KERN_DEBUG "btrfs: block rsv returned %d\n", ret);
6228                         WARN_ON(1);
6229                 }
6230                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
6231                 if (!ret) {
6232                         return block_rsv;
6233                 } else if (ret && block_rsv != global_rsv) {
6234                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6235                         if (!ret)
6236                                 return global_rsv;
6237                 }
6238         }
6239
6240         return ERR_PTR(-ENOSPC);
6241 }
6242
6243 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6244                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6245 {
6246         block_rsv_add_bytes(block_rsv, blocksize, 0);
6247         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6248 }
6249
6250 /*
6251  * finds a free extent and does all the dirty work required for allocation
6252  * returns the key for the extent through ins, and a tree buffer for
6253  * the first block of the extent through buf.
6254  *
6255  * returns the tree buffer or NULL.
6256  */
6257 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6258                                         struct btrfs_root *root, u32 blocksize,
6259                                         u64 parent, u64 root_objectid,
6260                                         struct btrfs_disk_key *key, int level,
6261                                         u64 hint, u64 empty_size)
6262 {
6263         struct btrfs_key ins;
6264         struct btrfs_block_rsv *block_rsv;
6265         struct extent_buffer *buf;
6266         u64 flags = 0;
6267         int ret;
6268
6269
6270         block_rsv = use_block_rsv(trans, root, blocksize);
6271         if (IS_ERR(block_rsv))
6272                 return ERR_CAST(block_rsv);
6273
6274         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6275                                    empty_size, hint, &ins, 0);
6276         if (ret) {
6277                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6278                 return ERR_PTR(ret);
6279         }
6280
6281         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6282                                     blocksize, level);
6283         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6284
6285         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6286                 if (parent == 0)
6287                         parent = ins.objectid;
6288                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6289         } else
6290                 BUG_ON(parent > 0);
6291
6292         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6293                 struct btrfs_delayed_extent_op *extent_op;
6294                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
6295                 BUG_ON(!extent_op); /* -ENOMEM */
6296                 if (key)
6297                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6298                 else
6299                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6300                 extent_op->flags_to_set = flags;
6301                 extent_op->update_key = 1;
6302                 extent_op->update_flags = 1;
6303                 extent_op->is_data = 0;
6304
6305                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6306                                         ins.objectid,
6307                                         ins.offset, parent, root_objectid,
6308                                         level, BTRFS_ADD_DELAYED_EXTENT,
6309                                         extent_op, 0);
6310                 BUG_ON(ret); /* -ENOMEM */
6311         }
6312         return buf;
6313 }
6314
6315 struct walk_control {
6316         u64 refs[BTRFS_MAX_LEVEL];
6317         u64 flags[BTRFS_MAX_LEVEL];
6318         struct btrfs_key update_progress;
6319         int stage;
6320         int level;
6321         int shared_level;
6322         int update_ref;
6323         int keep_locks;
6324         int reada_slot;
6325         int reada_count;
6326         int for_reloc;
6327 };
6328
6329 #define DROP_REFERENCE  1
6330 #define UPDATE_BACKREF  2
6331
6332 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6333                                      struct btrfs_root *root,
6334                                      struct walk_control *wc,
6335                                      struct btrfs_path *path)
6336 {
6337         u64 bytenr;
6338         u64 generation;
6339         u64 refs;
6340         u64 flags;
6341         u32 nritems;
6342         u32 blocksize;
6343         struct btrfs_key key;
6344         struct extent_buffer *eb;
6345         int ret;
6346         int slot;
6347         int nread = 0;
6348
6349         if (path->slots[wc->level] < wc->reada_slot) {
6350                 wc->reada_count = wc->reada_count * 2 / 3;
6351                 wc->reada_count = max(wc->reada_count, 2);
6352         } else {
6353                 wc->reada_count = wc->reada_count * 3 / 2;
6354                 wc->reada_count = min_t(int, wc->reada_count,
6355                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6356         }
6357
6358         eb = path->nodes[wc->level];
6359         nritems = btrfs_header_nritems(eb);
6360         blocksize = btrfs_level_size(root, wc->level - 1);
6361
6362         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6363                 if (nread >= wc->reada_count)
6364                         break;
6365
6366                 cond_resched();
6367                 bytenr = btrfs_node_blockptr(eb, slot);
6368                 generation = btrfs_node_ptr_generation(eb, slot);
6369
6370                 if (slot == path->slots[wc->level])
6371                         goto reada;
6372
6373                 if (wc->stage == UPDATE_BACKREF &&
6374                     generation <= root->root_key.offset)
6375                         continue;
6376
6377                 /* We don't lock the tree block, it's OK to be racy here */
6378                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6379                                                &refs, &flags);
6380                 /* We don't care about errors in readahead. */
6381                 if (ret < 0)
6382                         continue;
6383                 BUG_ON(refs == 0);
6384
6385                 if (wc->stage == DROP_REFERENCE) {
6386                         if (refs == 1)
6387                                 goto reada;
6388
6389                         if (wc->level == 1 &&
6390                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6391                                 continue;
6392                         if (!wc->update_ref ||
6393                             generation <= root->root_key.offset)
6394                                 continue;
6395                         btrfs_node_key_to_cpu(eb, &key, slot);
6396                         ret = btrfs_comp_cpu_keys(&key,
6397                                                   &wc->update_progress);
6398                         if (ret < 0)
6399                                 continue;
6400                 } else {
6401                         if (wc->level == 1 &&
6402                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6403                                 continue;
6404                 }
6405 reada:
6406                 ret = readahead_tree_block(root, bytenr, blocksize,
6407                                            generation);
6408                 if (ret)
6409                         break;
6410                 nread++;
6411         }
6412         wc->reada_slot = slot;
6413 }
6414
6415 /*
6416  * hepler to process tree block while walking down the tree.
6417  *
6418  * when wc->stage == UPDATE_BACKREF, this function updates
6419  * back refs for pointers in the block.
6420  *
6421  * NOTE: return value 1 means we should stop walking down.
6422  */
6423 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6424                                    struct btrfs_root *root,
6425                                    struct btrfs_path *path,
6426                                    struct walk_control *wc, int lookup_info)
6427 {
6428         int level = wc->level;
6429         struct extent_buffer *eb = path->nodes[level];
6430         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6431         int ret;
6432
6433         if (wc->stage == UPDATE_BACKREF &&
6434             btrfs_header_owner(eb) != root->root_key.objectid)
6435                 return 1;
6436
6437         /*
6438          * when reference count of tree block is 1, it won't increase
6439          * again. once full backref flag is set, we never clear it.
6440          */
6441         if (lookup_info &&
6442             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6443              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6444                 BUG_ON(!path->locks[level]);
6445                 ret = btrfs_lookup_extent_info(trans, root,
6446                                                eb->start, eb->len,
6447                                                &wc->refs[level],
6448                                                &wc->flags[level]);
6449                 BUG_ON(ret == -ENOMEM);
6450                 if (ret)
6451                         return ret;
6452                 BUG_ON(wc->refs[level] == 0);
6453         }
6454
6455         if (wc->stage == DROP_REFERENCE) {
6456                 if (wc->refs[level] > 1)
6457                         return 1;
6458
6459                 if (path->locks[level] && !wc->keep_locks) {
6460                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6461                         path->locks[level] = 0;
6462                 }
6463                 return 0;
6464         }
6465
6466         /* wc->stage == UPDATE_BACKREF */
6467         if (!(wc->flags[level] & flag)) {
6468                 BUG_ON(!path->locks[level]);
6469                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
6470                 BUG_ON(ret); /* -ENOMEM */
6471                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
6472                 BUG_ON(ret); /* -ENOMEM */
6473                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6474                                                   eb->len, flag, 0);
6475                 BUG_ON(ret); /* -ENOMEM */
6476                 wc->flags[level] |= flag;
6477         }
6478
6479         /*
6480          * the block is shared by multiple trees, so it's not good to
6481          * keep the tree lock
6482          */
6483         if (path->locks[level] && level > 0) {
6484                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6485                 path->locks[level] = 0;
6486         }
6487         return 0;
6488 }
6489
6490 /*
6491  * hepler to process tree block pointer.
6492  *
6493  * when wc->stage == DROP_REFERENCE, this function checks
6494  * reference count of the block pointed to. if the block
6495  * is shared and we need update back refs for the subtree
6496  * rooted at the block, this function changes wc->stage to
6497  * UPDATE_BACKREF. if the block is shared and there is no
6498  * need to update back, this function drops the reference
6499  * to the block.
6500  *
6501  * NOTE: return value 1 means we should stop walking down.
6502  */
6503 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6504                                  struct btrfs_root *root,
6505                                  struct btrfs_path *path,
6506                                  struct walk_control *wc, int *lookup_info)
6507 {
6508         u64 bytenr;
6509         u64 generation;
6510         u64 parent;
6511         u32 blocksize;
6512         struct btrfs_key key;
6513         struct extent_buffer *next;
6514         int level = wc->level;
6515         int reada = 0;
6516         int ret = 0;
6517
6518         generation = btrfs_node_ptr_generation(path->nodes[level],
6519                                                path->slots[level]);
6520         /*
6521          * if the lower level block was created before the snapshot
6522          * was created, we know there is no need to update back refs
6523          * for the subtree
6524          */
6525         if (wc->stage == UPDATE_BACKREF &&
6526             generation <= root->root_key.offset) {
6527                 *lookup_info = 1;
6528                 return 1;
6529         }
6530
6531         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6532         blocksize = btrfs_level_size(root, level - 1);
6533
6534         next = btrfs_find_tree_block(root, bytenr, blocksize);
6535         if (!next) {
6536                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6537                 if (!next)
6538                         return -ENOMEM;
6539                 reada = 1;
6540         }
6541         btrfs_tree_lock(next);
6542         btrfs_set_lock_blocking(next);
6543
6544         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6545                                        &wc->refs[level - 1],
6546                                        &wc->flags[level - 1]);
6547         if (ret < 0) {
6548                 btrfs_tree_unlock(next);
6549                 return ret;
6550         }
6551
6552         BUG_ON(wc->refs[level - 1] == 0);
6553         *lookup_info = 0;
6554
6555         if (wc->stage == DROP_REFERENCE) {
6556                 if (wc->refs[level - 1] > 1) {
6557                         if (level == 1 &&
6558                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6559                                 goto skip;
6560
6561                         if (!wc->update_ref ||
6562                             generation <= root->root_key.offset)
6563                                 goto skip;
6564
6565                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6566                                               path->slots[level]);
6567                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6568                         if (ret < 0)
6569                                 goto skip;
6570
6571                         wc->stage = UPDATE_BACKREF;
6572                         wc->shared_level = level - 1;
6573                 }
6574         } else {
6575                 if (level == 1 &&
6576                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6577                         goto skip;
6578         }
6579
6580         if (!btrfs_buffer_uptodate(next, generation, 0)) {
6581                 btrfs_tree_unlock(next);
6582                 free_extent_buffer(next);
6583                 next = NULL;
6584                 *lookup_info = 1;
6585         }
6586
6587         if (!next) {
6588                 if (reada && level == 1)
6589                         reada_walk_down(trans, root, wc, path);
6590                 next = read_tree_block(root, bytenr, blocksize, generation);
6591                 if (!next)
6592                         return -EIO;
6593                 btrfs_tree_lock(next);
6594                 btrfs_set_lock_blocking(next);
6595         }
6596
6597         level--;
6598         BUG_ON(level != btrfs_header_level(next));
6599         path->nodes[level] = next;
6600         path->slots[level] = 0;
6601         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6602         wc->level = level;
6603         if (wc->level == 1)
6604                 wc->reada_slot = 0;
6605         return 0;
6606 skip:
6607         wc->refs[level - 1] = 0;
6608         wc->flags[level - 1] = 0;
6609         if (wc->stage == DROP_REFERENCE) {
6610                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6611                         parent = path->nodes[level]->start;
6612                 } else {
6613                         BUG_ON(root->root_key.objectid !=
6614                                btrfs_header_owner(path->nodes[level]));
6615                         parent = 0;
6616                 }
6617
6618                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6619                                 root->root_key.objectid, level - 1, 0, 0);
6620                 BUG_ON(ret); /* -ENOMEM */
6621         }
6622         btrfs_tree_unlock(next);
6623         free_extent_buffer(next);
6624         *lookup_info = 1;
6625         return 1;
6626 }
6627
6628 /*
6629  * hepler to process tree block while walking up the tree.
6630  *
6631  * when wc->stage == DROP_REFERENCE, this function drops
6632  * reference count on the block.
6633  *
6634  * when wc->stage == UPDATE_BACKREF, this function changes
6635  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6636  * to UPDATE_BACKREF previously while processing the block.
6637  *
6638  * NOTE: return value 1 means we should stop walking up.
6639  */
6640 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6641                                  struct btrfs_root *root,
6642                                  struct btrfs_path *path,
6643                                  struct walk_control *wc)
6644 {
6645         int ret;
6646         int level = wc->level;
6647         struct extent_buffer *eb = path->nodes[level];
6648         u64 parent = 0;
6649
6650         if (wc->stage == UPDATE_BACKREF) {
6651                 BUG_ON(wc->shared_level < level);
6652                 if (level < wc->shared_level)
6653                         goto out;
6654
6655                 ret = find_next_key(path, level + 1, &wc->update_progress);
6656                 if (ret > 0)
6657                         wc->update_ref = 0;
6658
6659                 wc->stage = DROP_REFERENCE;
6660                 wc->shared_level = -1;
6661                 path->slots[level] = 0;
6662
6663                 /*
6664                  * check reference count again if the block isn't locked.
6665                  * we should start walking down the tree again if reference
6666                  * count is one.
6667                  */
6668                 if (!path->locks[level]) {
6669                         BUG_ON(level == 0);
6670                         btrfs_tree_lock(eb);
6671                         btrfs_set_lock_blocking(eb);
6672                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6673
6674                         ret = btrfs_lookup_extent_info(trans, root,
6675                                                        eb->start, eb->len,
6676                                                        &wc->refs[level],
6677                                                        &wc->flags[level]);
6678                         if (ret < 0) {
6679                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6680                                 return ret;
6681                         }
6682                         BUG_ON(wc->refs[level] == 0);
6683                         if (wc->refs[level] == 1) {
6684                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6685                                 return 1;
6686                         }
6687                 }
6688         }
6689
6690         /* wc->stage == DROP_REFERENCE */
6691         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6692
6693         if (wc->refs[level] == 1) {
6694                 if (level == 0) {
6695                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6696                                 ret = btrfs_dec_ref(trans, root, eb, 1,
6697                                                     wc->for_reloc);
6698                         else
6699                                 ret = btrfs_dec_ref(trans, root, eb, 0,
6700                                                     wc->for_reloc);
6701                         BUG_ON(ret); /* -ENOMEM */
6702                 }
6703                 /* make block locked assertion in clean_tree_block happy */
6704                 if (!path->locks[level] &&
6705                     btrfs_header_generation(eb) == trans->transid) {
6706                         btrfs_tree_lock(eb);
6707                         btrfs_set_lock_blocking(eb);
6708                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6709                 }
6710                 clean_tree_block(trans, root, eb);
6711         }
6712
6713         if (eb == root->node) {
6714                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6715                         parent = eb->start;
6716                 else
6717                         BUG_ON(root->root_key.objectid !=
6718                                btrfs_header_owner(eb));
6719         } else {
6720                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6721                         parent = path->nodes[level + 1]->start;
6722                 else
6723                         BUG_ON(root->root_key.objectid !=
6724                                btrfs_header_owner(path->nodes[level + 1]));
6725         }
6726
6727         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6728 out:
6729         wc->refs[level] = 0;
6730         wc->flags[level] = 0;
6731         return 0;
6732 }
6733
6734 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6735                                    struct btrfs_root *root,
6736                                    struct btrfs_path *path,
6737                                    struct walk_control *wc)
6738 {
6739         int level = wc->level;
6740         int lookup_info = 1;
6741         int ret;
6742
6743         while (level >= 0) {
6744                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6745                 if (ret > 0)
6746                         break;
6747
6748                 if (level == 0)
6749                         break;
6750
6751                 if (path->slots[level] >=
6752                     btrfs_header_nritems(path->nodes[level]))
6753                         break;
6754
6755                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6756                 if (ret > 0) {
6757                         path->slots[level]++;
6758                         continue;
6759                 } else if (ret < 0)
6760                         return ret;
6761                 level = wc->level;
6762         }
6763         return 0;
6764 }
6765
6766 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6767                                  struct btrfs_root *root,
6768                                  struct btrfs_path *path,
6769                                  struct walk_control *wc, int max_level)
6770 {
6771         int level = wc->level;
6772         int ret;
6773
6774         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6775         while (level < max_level && path->nodes[level]) {
6776                 wc->level = level;
6777                 if (path->slots[level] + 1 <
6778                     btrfs_header_nritems(path->nodes[level])) {
6779                         path->slots[level]++;
6780                         return 0;
6781                 } else {
6782                         ret = walk_up_proc(trans, root, path, wc);
6783                         if (ret > 0)
6784                                 return 0;
6785
6786                         if (path->locks[level]) {
6787                                 btrfs_tree_unlock_rw(path->nodes[level],
6788                                                      path->locks[level]);
6789                                 path->locks[level] = 0;
6790                         }
6791                         free_extent_buffer(path->nodes[level]);
6792                         path->nodes[level] = NULL;
6793                         level++;
6794                 }
6795         }
6796         return 1;
6797 }
6798
6799 /*
6800  * drop a subvolume tree.
6801  *
6802  * this function traverses the tree freeing any blocks that only
6803  * referenced by the tree.
6804  *
6805  * when a shared tree block is found. this function decreases its
6806  * reference count by one. if update_ref is true, this function
6807  * also make sure backrefs for the shared block and all lower level
6808  * blocks are properly updated.
6809  */
6810 int btrfs_drop_snapshot(struct btrfs_root *root,
6811                          struct btrfs_block_rsv *block_rsv, int update_ref,
6812                          int for_reloc)
6813 {
6814         struct btrfs_path *path;
6815         struct btrfs_trans_handle *trans;
6816         struct btrfs_root *tree_root = root->fs_info->tree_root;
6817         struct btrfs_root_item *root_item = &root->root_item;
6818         struct walk_control *wc;
6819         struct btrfs_key key;
6820         int err = 0;
6821         int ret;
6822         int level;
6823
6824         path = btrfs_alloc_path();
6825         if (!path) {
6826                 err = -ENOMEM;
6827                 goto out;
6828         }
6829
6830         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6831         if (!wc) {
6832                 btrfs_free_path(path);
6833                 err = -ENOMEM;
6834                 goto out;
6835         }
6836
6837         trans = btrfs_start_transaction(tree_root, 0);
6838         if (IS_ERR(trans)) {
6839                 err = PTR_ERR(trans);
6840                 goto out_free;
6841         }
6842
6843         if (block_rsv)
6844                 trans->block_rsv = block_rsv;
6845
6846         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6847                 level = btrfs_header_level(root->node);
6848                 path->nodes[level] = btrfs_lock_root_node(root);
6849                 btrfs_set_lock_blocking(path->nodes[level]);
6850                 path->slots[level] = 0;
6851                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6852                 memset(&wc->update_progress, 0,
6853                        sizeof(wc->update_progress));
6854         } else {
6855                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6856                 memcpy(&wc->update_progress, &key,
6857                        sizeof(wc->update_progress));
6858
6859                 level = root_item->drop_level;
6860                 BUG_ON(level == 0);
6861                 path->lowest_level = level;
6862                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6863                 path->lowest_level = 0;
6864                 if (ret < 0) {
6865                         err = ret;
6866                         goto out_end_trans;
6867                 }
6868                 WARN_ON(ret > 0);
6869
6870                 /*
6871                  * unlock our path, this is safe because only this
6872                  * function is allowed to delete this snapshot
6873                  */
6874                 btrfs_unlock_up_safe(path, 0);
6875
6876                 level = btrfs_header_level(root->node);
6877                 while (1) {
6878                         btrfs_tree_lock(path->nodes[level]);
6879                         btrfs_set_lock_blocking(path->nodes[level]);
6880
6881                         ret = btrfs_lookup_extent_info(trans, root,
6882                                                 path->nodes[level]->start,
6883                                                 path->nodes[level]->len,
6884                                                 &wc->refs[level],
6885                                                 &wc->flags[level]);
6886                         if (ret < 0) {
6887                                 err = ret;
6888                                 goto out_end_trans;
6889                         }
6890                         BUG_ON(wc->refs[level] == 0);
6891
6892                         if (level == root_item->drop_level)
6893                                 break;
6894
6895                         btrfs_tree_unlock(path->nodes[level]);
6896                         WARN_ON(wc->refs[level] != 1);
6897                         level--;
6898                 }
6899         }
6900
6901         wc->level = level;
6902         wc->shared_level = -1;
6903         wc->stage = DROP_REFERENCE;
6904         wc->update_ref = update_ref;
6905         wc->keep_locks = 0;
6906         wc->for_reloc = for_reloc;
6907         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6908
6909         while (1) {
6910                 ret = walk_down_tree(trans, root, path, wc);
6911                 if (ret < 0) {
6912                         err = ret;
6913                         break;
6914                 }
6915
6916                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
6917                 if (ret < 0) {
6918                         err = ret;
6919                         break;
6920                 }
6921
6922                 if (ret > 0) {
6923                         BUG_ON(wc->stage != DROP_REFERENCE);
6924                         break;
6925                 }
6926
6927                 if (wc->stage == DROP_REFERENCE) {
6928                         level = wc->level;
6929                         btrfs_node_key(path->nodes[level],
6930                                        &root_item->drop_progress,
6931                                        path->slots[level]);
6932                         root_item->drop_level = level;
6933                 }
6934
6935                 BUG_ON(wc->level == 0);
6936                 if (btrfs_should_end_transaction(trans, tree_root)) {
6937                         ret = btrfs_update_root(trans, tree_root,
6938                                                 &root->root_key,
6939                                                 root_item);
6940                         if (ret) {
6941                                 btrfs_abort_transaction(trans, tree_root, ret);
6942                                 err = ret;
6943                                 goto out_end_trans;
6944                         }
6945
6946                         btrfs_end_transaction_throttle(trans, tree_root);
6947                         trans = btrfs_start_transaction(tree_root, 0);
6948                         if (IS_ERR(trans)) {
6949                                 err = PTR_ERR(trans);
6950                                 goto out_free;
6951                         }
6952                         if (block_rsv)
6953                                 trans->block_rsv = block_rsv;
6954                 }
6955         }
6956         btrfs_release_path(path);
6957         if (err)
6958                 goto out_end_trans;
6959
6960         ret = btrfs_del_root(trans, tree_root, &root->root_key);
6961         if (ret) {
6962                 btrfs_abort_transaction(trans, tree_root, ret);
6963                 goto out_end_trans;
6964         }
6965
6966         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
6967                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
6968                                            NULL, NULL);
6969                 if (ret < 0) {
6970                         btrfs_abort_transaction(trans, tree_root, ret);
6971                         err = ret;
6972                         goto out_end_trans;
6973                 } else if (ret > 0) {
6974                         /* if we fail to delete the orphan item this time
6975                          * around, it'll get picked up the next time.
6976                          *
6977                          * The most common failure here is just -ENOENT.
6978                          */
6979                         btrfs_del_orphan_item(trans, tree_root,
6980                                               root->root_key.objectid);
6981                 }
6982         }
6983
6984         if (root->in_radix) {
6985                 btrfs_free_fs_root(tree_root->fs_info, root);
6986         } else {
6987                 free_extent_buffer(root->node);
6988                 free_extent_buffer(root->commit_root);
6989                 kfree(root);
6990         }
6991 out_end_trans:
6992         btrfs_end_transaction_throttle(trans, tree_root);
6993 out_free:
6994         kfree(wc);
6995         btrfs_free_path(path);
6996 out:
6997         if (err)
6998                 btrfs_std_error(root->fs_info, err);
6999         return err;
7000 }
7001
7002 /*
7003  * drop subtree rooted at tree block 'node'.
7004  *
7005  * NOTE: this function will unlock and release tree block 'node'
7006  * only used by relocation code
7007  */
7008 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7009                         struct btrfs_root *root,
7010                         struct extent_buffer *node,
7011                         struct extent_buffer *parent)
7012 {
7013         struct btrfs_path *path;
7014         struct walk_control *wc;
7015         int level;
7016         int parent_level;
7017         int ret = 0;
7018         int wret;
7019
7020         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7021
7022         path = btrfs_alloc_path();
7023         if (!path)
7024                 return -ENOMEM;
7025
7026         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7027         if (!wc) {
7028                 btrfs_free_path(path);
7029                 return -ENOMEM;
7030         }
7031
7032         btrfs_assert_tree_locked(parent);
7033         parent_level = btrfs_header_level(parent);
7034         extent_buffer_get(parent);
7035         path->nodes[parent_level] = parent;
7036         path->slots[parent_level] = btrfs_header_nritems(parent);
7037
7038         btrfs_assert_tree_locked(node);
7039         level = btrfs_header_level(node);
7040         path->nodes[level] = node;
7041         path->slots[level] = 0;
7042         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7043
7044         wc->refs[parent_level] = 1;
7045         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7046         wc->level = level;
7047         wc->shared_level = -1;
7048         wc->stage = DROP_REFERENCE;
7049         wc->update_ref = 0;
7050         wc->keep_locks = 1;
7051         wc->for_reloc = 1;
7052         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7053
7054         while (1) {
7055                 wret = walk_down_tree(trans, root, path, wc);
7056                 if (wret < 0) {
7057                         ret = wret;
7058                         break;
7059                 }
7060
7061                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7062                 if (wret < 0)
7063                         ret = wret;
7064                 if (wret != 0)
7065                         break;
7066         }
7067
7068         kfree(wc);
7069         btrfs_free_path(path);
7070         return ret;
7071 }
7072
7073 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7074 {
7075         u64 num_devices;
7076         u64 stripped;
7077
7078         /*
7079          * if restripe for this chunk_type is on pick target profile and
7080          * return, otherwise do the usual balance
7081          */
7082         stripped = get_restripe_target(root->fs_info, flags);
7083         if (stripped)
7084                 return extended_to_chunk(stripped);
7085
7086         /*
7087          * we add in the count of missing devices because we want
7088          * to make sure that any RAID levels on a degraded FS
7089          * continue to be honored.
7090          */
7091         num_devices = root->fs_info->fs_devices->rw_devices +
7092                 root->fs_info->fs_devices->missing_devices;
7093
7094         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7095                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7096
7097         if (num_devices == 1) {
7098                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7099                 stripped = flags & ~stripped;
7100
7101                 /* turn raid0 into single device chunks */
7102                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7103                         return stripped;
7104
7105                 /* turn mirroring into duplication */
7106                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7107                              BTRFS_BLOCK_GROUP_RAID10))
7108                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7109         } else {
7110                 /* they already had raid on here, just return */
7111                 if (flags & stripped)
7112                         return flags;
7113
7114                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7115                 stripped = flags & ~stripped;
7116
7117                 /* switch duplicated blocks with raid1 */
7118                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7119                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7120
7121                 /* this is drive concat, leave it alone */
7122         }
7123
7124         return flags;
7125 }
7126
7127 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7128 {
7129         struct btrfs_space_info *sinfo = cache->space_info;
7130         u64 num_bytes;
7131         u64 min_allocable_bytes;
7132         int ret = -ENOSPC;
7133
7134
7135         /*
7136          * We need some metadata space and system metadata space for
7137          * allocating chunks in some corner cases until we force to set
7138          * it to be readonly.
7139          */
7140         if ((sinfo->flags &
7141              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7142             !force)
7143                 min_allocable_bytes = 1 * 1024 * 1024;
7144         else
7145                 min_allocable_bytes = 0;
7146
7147         spin_lock(&sinfo->lock);
7148         spin_lock(&cache->lock);
7149
7150         if (cache->ro) {
7151                 ret = 0;
7152                 goto out;
7153         }
7154
7155         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7156                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7157
7158         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7159             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7160             min_allocable_bytes <= sinfo->total_bytes) {
7161                 sinfo->bytes_readonly += num_bytes;
7162                 cache->ro = 1;
7163                 ret = 0;
7164         }
7165 out:
7166         spin_unlock(&cache->lock);
7167         spin_unlock(&sinfo->lock);
7168         return ret;
7169 }
7170
7171 int btrfs_set_block_group_ro(struct btrfs_root *root,
7172                              struct btrfs_block_group_cache *cache)
7173
7174 {
7175         struct btrfs_trans_handle *trans;
7176         u64 alloc_flags;
7177         int ret;
7178
7179         BUG_ON(cache->ro);
7180
7181         trans = btrfs_join_transaction(root);
7182         if (IS_ERR(trans))
7183                 return PTR_ERR(trans);
7184
7185         alloc_flags = update_block_group_flags(root, cache->flags);
7186         if (alloc_flags != cache->flags) {
7187                 ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7188                                      CHUNK_ALLOC_FORCE);
7189                 if (ret < 0)
7190                         goto out;
7191         }
7192
7193         ret = set_block_group_ro(cache, 0);
7194         if (!ret)
7195                 goto out;
7196         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7197         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7198                              CHUNK_ALLOC_FORCE);
7199         if (ret < 0)
7200                 goto out;
7201         ret = set_block_group_ro(cache, 0);
7202 out:
7203         btrfs_end_transaction(trans, root);
7204         return ret;
7205 }
7206
7207 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7208                             struct btrfs_root *root, u64 type)
7209 {
7210         u64 alloc_flags = get_alloc_profile(root, type);
7211         return do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7212                               CHUNK_ALLOC_FORCE);
7213 }
7214
7215 /*
7216  * helper to account the unused space of all the readonly block group in the
7217  * list. takes mirrors into account.
7218  */
7219 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7220 {
7221         struct btrfs_block_group_cache *block_group;
7222         u64 free_bytes = 0;
7223         int factor;
7224
7225         list_for_each_entry(block_group, groups_list, list) {
7226                 spin_lock(&block_group->lock);
7227
7228                 if (!block_group->ro) {
7229                         spin_unlock(&block_group->lock);
7230                         continue;
7231                 }
7232
7233                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7234                                           BTRFS_BLOCK_GROUP_RAID10 |
7235                                           BTRFS_BLOCK_GROUP_DUP))
7236                         factor = 2;
7237                 else
7238                         factor = 1;
7239
7240                 free_bytes += (block_group->key.offset -
7241                                btrfs_block_group_used(&block_group->item)) *
7242                                factor;
7243
7244                 spin_unlock(&block_group->lock);
7245         }
7246
7247         return free_bytes;
7248 }
7249
7250 /*
7251  * helper to account the unused space of all the readonly block group in the
7252  * space_info. takes mirrors into account.
7253  */
7254 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7255 {
7256         int i;
7257         u64 free_bytes = 0;
7258
7259         spin_lock(&sinfo->lock);
7260
7261         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7262                 if (!list_empty(&sinfo->block_groups[i]))
7263                         free_bytes += __btrfs_get_ro_block_group_free_space(
7264                                                 &sinfo->block_groups[i]);
7265
7266         spin_unlock(&sinfo->lock);
7267
7268         return free_bytes;
7269 }
7270
7271 void btrfs_set_block_group_rw(struct btrfs_root *root,
7272                               struct btrfs_block_group_cache *cache)
7273 {
7274         struct btrfs_space_info *sinfo = cache->space_info;
7275         u64 num_bytes;
7276
7277         BUG_ON(!cache->ro);
7278
7279         spin_lock(&sinfo->lock);
7280         spin_lock(&cache->lock);
7281         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7282                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7283         sinfo->bytes_readonly -= num_bytes;
7284         cache->ro = 0;
7285         spin_unlock(&cache->lock);
7286         spin_unlock(&sinfo->lock);
7287 }
7288
7289 /*
7290  * checks to see if its even possible to relocate this block group.
7291  *
7292  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7293  * ok to go ahead and try.
7294  */
7295 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7296 {
7297         struct btrfs_block_group_cache *block_group;
7298         struct btrfs_space_info *space_info;
7299         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7300         struct btrfs_device *device;
7301         u64 min_free;
7302         u64 dev_min = 1;
7303         u64 dev_nr = 0;
7304         u64 target;
7305         int index;
7306         int full = 0;
7307         int ret = 0;
7308
7309         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7310
7311         /* odd, couldn't find the block group, leave it alone */
7312         if (!block_group)
7313                 return -1;
7314
7315         min_free = btrfs_block_group_used(&block_group->item);
7316
7317         /* no bytes used, we're good */
7318         if (!min_free)
7319                 goto out;
7320
7321         space_info = block_group->space_info;
7322         spin_lock(&space_info->lock);
7323
7324         full = space_info->full;
7325
7326         /*
7327          * if this is the last block group we have in this space, we can't
7328          * relocate it unless we're able to allocate a new chunk below.
7329          *
7330          * Otherwise, we need to make sure we have room in the space to handle
7331          * all of the extents from this block group.  If we can, we're good
7332          */
7333         if ((space_info->total_bytes != block_group->key.offset) &&
7334             (space_info->bytes_used + space_info->bytes_reserved +
7335              space_info->bytes_pinned + space_info->bytes_readonly +
7336              min_free < space_info->total_bytes)) {
7337                 spin_unlock(&space_info->lock);
7338                 goto out;
7339         }
7340         spin_unlock(&space_info->lock);
7341
7342         /*
7343          * ok we don't have enough space, but maybe we have free space on our
7344          * devices to allocate new chunks for relocation, so loop through our
7345          * alloc devices and guess if we have enough space.  if this block
7346          * group is going to be restriped, run checks against the target
7347          * profile instead of the current one.
7348          */
7349         ret = -1;
7350
7351         /*
7352          * index:
7353          *      0: raid10
7354          *      1: raid1
7355          *      2: dup
7356          *      3: raid0
7357          *      4: single
7358          */
7359         target = get_restripe_target(root->fs_info, block_group->flags);
7360         if (target) {
7361                 index = __get_block_group_index(extended_to_chunk(target));
7362         } else {
7363                 /*
7364                  * this is just a balance, so if we were marked as full
7365                  * we know there is no space for a new chunk
7366                  */
7367                 if (full)
7368                         goto out;
7369
7370                 index = get_block_group_index(block_group);
7371         }
7372
7373         if (index == 0) {
7374                 dev_min = 4;
7375                 /* Divide by 2 */
7376                 min_free >>= 1;
7377         } else if (index == 1) {
7378                 dev_min = 2;
7379         } else if (index == 2) {
7380                 /* Multiply by 2 */
7381                 min_free <<= 1;
7382         } else if (index == 3) {
7383                 dev_min = fs_devices->rw_devices;
7384                 do_div(min_free, dev_min);
7385         }
7386
7387         mutex_lock(&root->fs_info->chunk_mutex);
7388         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7389                 u64 dev_offset;
7390
7391                 /*
7392                  * check to make sure we can actually find a chunk with enough
7393                  * space to fit our block group in.
7394                  */
7395                 if (device->total_bytes > device->bytes_used + min_free) {
7396                         ret = find_free_dev_extent(device, min_free,
7397                                                    &dev_offset, NULL);
7398                         if (!ret)
7399                                 dev_nr++;
7400
7401                         if (dev_nr >= dev_min)
7402                                 break;
7403
7404                         ret = -1;
7405                 }
7406         }
7407         mutex_unlock(&root->fs_info->chunk_mutex);
7408 out:
7409         btrfs_put_block_group(block_group);
7410         return ret;
7411 }
7412
7413 static int find_first_block_group(struct btrfs_root *root,
7414                 struct btrfs_path *path, struct btrfs_key *key)
7415 {
7416         int ret = 0;
7417         struct btrfs_key found_key;
7418         struct extent_buffer *leaf;
7419         int slot;
7420
7421         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7422         if (ret < 0)
7423                 goto out;
7424
7425         while (1) {
7426                 slot = path->slots[0];
7427                 leaf = path->nodes[0];
7428                 if (slot >= btrfs_header_nritems(leaf)) {
7429                         ret = btrfs_next_leaf(root, path);
7430                         if (ret == 0)
7431                                 continue;
7432                         if (ret < 0)
7433                                 goto out;
7434                         break;
7435                 }
7436                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7437
7438                 if (found_key.objectid >= key->objectid &&
7439                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7440                         ret = 0;
7441                         goto out;
7442                 }
7443                 path->slots[0]++;
7444         }
7445 out:
7446         return ret;
7447 }
7448
7449 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7450 {
7451         struct btrfs_block_group_cache *block_group;
7452         u64 last = 0;
7453
7454         while (1) {
7455                 struct inode *inode;
7456
7457                 block_group = btrfs_lookup_first_block_group(info, last);
7458                 while (block_group) {
7459                         spin_lock(&block_group->lock);
7460                         if (block_group->iref)
7461                                 break;
7462                         spin_unlock(&block_group->lock);
7463                         block_group = next_block_group(info->tree_root,
7464                                                        block_group);
7465                 }
7466                 if (!block_group) {
7467                         if (last == 0)
7468                                 break;
7469                         last = 0;
7470                         continue;
7471                 }
7472
7473                 inode = block_group->inode;
7474                 block_group->iref = 0;
7475                 block_group->inode = NULL;
7476                 spin_unlock(&block_group->lock);
7477                 iput(inode);
7478                 last = block_group->key.objectid + block_group->key.offset;
7479                 btrfs_put_block_group(block_group);
7480         }
7481 }
7482
7483 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7484 {
7485         struct btrfs_block_group_cache *block_group;
7486         struct btrfs_space_info *space_info;
7487         struct btrfs_caching_control *caching_ctl;
7488         struct rb_node *n;
7489
7490         down_write(&info->extent_commit_sem);
7491         while (!list_empty(&info->caching_block_groups)) {
7492                 caching_ctl = list_entry(info->caching_block_groups.next,
7493                                          struct btrfs_caching_control, list);
7494                 list_del(&caching_ctl->list);
7495                 put_caching_control(caching_ctl);
7496         }
7497         up_write(&info->extent_commit_sem);
7498
7499         spin_lock(&info->block_group_cache_lock);
7500         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7501                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7502                                        cache_node);
7503                 rb_erase(&block_group->cache_node,
7504                          &info->block_group_cache_tree);
7505                 spin_unlock(&info->block_group_cache_lock);
7506
7507                 down_write(&block_group->space_info->groups_sem);
7508                 list_del(&block_group->list);
7509                 up_write(&block_group->space_info->groups_sem);
7510
7511                 if (block_group->cached == BTRFS_CACHE_STARTED)
7512                         wait_block_group_cache_done(block_group);
7513
7514                 /*
7515                  * We haven't cached this block group, which means we could
7516                  * possibly have excluded extents on this block group.
7517                  */
7518                 if (block_group->cached == BTRFS_CACHE_NO)
7519                         free_excluded_extents(info->extent_root, block_group);
7520
7521                 btrfs_remove_free_space_cache(block_group);
7522                 btrfs_put_block_group(block_group);
7523
7524                 spin_lock(&info->block_group_cache_lock);
7525         }
7526         spin_unlock(&info->block_group_cache_lock);
7527
7528         /* now that all the block groups are freed, go through and
7529          * free all the space_info structs.  This is only called during
7530          * the final stages of unmount, and so we know nobody is
7531          * using them.  We call synchronize_rcu() once before we start,
7532          * just to be on the safe side.
7533          */
7534         synchronize_rcu();
7535
7536         release_global_block_rsv(info);
7537
7538         while(!list_empty(&info->space_info)) {
7539                 space_info = list_entry(info->space_info.next,
7540                                         struct btrfs_space_info,
7541                                         list);
7542                 if (space_info->bytes_pinned > 0 ||
7543                     space_info->bytes_reserved > 0 ||
7544                     space_info->bytes_may_use > 0) {
7545                         WARN_ON(1);
7546                         dump_space_info(space_info, 0, 0);
7547                 }
7548                 list_del(&space_info->list);
7549                 kfree(space_info);
7550         }
7551         return 0;
7552 }
7553
7554 static void __link_block_group(struct btrfs_space_info *space_info,
7555                                struct btrfs_block_group_cache *cache)
7556 {
7557         int index = get_block_group_index(cache);
7558
7559         down_write(&space_info->groups_sem);
7560         list_add_tail(&cache->list, &space_info->block_groups[index]);
7561         up_write(&space_info->groups_sem);
7562 }
7563
7564 int btrfs_read_block_groups(struct btrfs_root *root)
7565 {
7566         struct btrfs_path *path;
7567         int ret;
7568         struct btrfs_block_group_cache *cache;
7569         struct btrfs_fs_info *info = root->fs_info;
7570         struct btrfs_space_info *space_info;
7571         struct btrfs_key key;
7572         struct btrfs_key found_key;
7573         struct extent_buffer *leaf;
7574         int need_clear = 0;
7575         u64 cache_gen;
7576
7577         root = info->extent_root;
7578         key.objectid = 0;
7579         key.offset = 0;
7580         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7581         path = btrfs_alloc_path();
7582         if (!path)
7583                 return -ENOMEM;
7584         path->reada = 1;
7585
7586         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7587         if (btrfs_test_opt(root, SPACE_CACHE) &&
7588             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7589                 need_clear = 1;
7590         if (btrfs_test_opt(root, CLEAR_CACHE))
7591                 need_clear = 1;
7592
7593         while (1) {
7594                 ret = find_first_block_group(root, path, &key);
7595                 if (ret > 0)
7596                         break;
7597                 if (ret != 0)
7598                         goto error;
7599                 leaf = path->nodes[0];
7600                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7601                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7602                 if (!cache) {
7603                         ret = -ENOMEM;
7604                         goto error;
7605                 }
7606                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7607                                                 GFP_NOFS);
7608                 if (!cache->free_space_ctl) {
7609                         kfree(cache);
7610                         ret = -ENOMEM;
7611                         goto error;
7612                 }
7613
7614                 atomic_set(&cache->count, 1);
7615                 spin_lock_init(&cache->lock);
7616                 cache->fs_info = info;
7617                 INIT_LIST_HEAD(&cache->list);
7618                 INIT_LIST_HEAD(&cache->cluster_list);
7619
7620                 if (need_clear) {
7621                         /*
7622                          * When we mount with old space cache, we need to
7623                          * set BTRFS_DC_CLEAR and set dirty flag.
7624                          *
7625                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
7626                          *    truncate the old free space cache inode and
7627                          *    setup a new one.
7628                          * b) Setting 'dirty flag' makes sure that we flush
7629                          *    the new space cache info onto disk.
7630                          */
7631                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7632                         if (btrfs_test_opt(root, SPACE_CACHE))
7633                                 cache->dirty = 1;
7634                 }
7635
7636                 read_extent_buffer(leaf, &cache->item,
7637                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7638                                    sizeof(cache->item));
7639                 memcpy(&cache->key, &found_key, sizeof(found_key));
7640
7641                 key.objectid = found_key.objectid + found_key.offset;
7642                 btrfs_release_path(path);
7643                 cache->flags = btrfs_block_group_flags(&cache->item);
7644                 cache->sectorsize = root->sectorsize;
7645
7646                 btrfs_init_free_space_ctl(cache);
7647
7648                 /*
7649                  * We need to exclude the super stripes now so that the space
7650                  * info has super bytes accounted for, otherwise we'll think
7651                  * we have more space than we actually do.
7652                  */
7653                 exclude_super_stripes(root, cache);
7654
7655                 /*
7656                  * check for two cases, either we are full, and therefore
7657                  * don't need to bother with the caching work since we won't
7658                  * find any space, or we are empty, and we can just add all
7659                  * the space in and be done with it.  This saves us _alot_ of
7660                  * time, particularly in the full case.
7661                  */
7662                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7663                         cache->last_byte_to_unpin = (u64)-1;
7664                         cache->cached = BTRFS_CACHE_FINISHED;
7665                         free_excluded_extents(root, cache);
7666                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7667                         cache->last_byte_to_unpin = (u64)-1;
7668                         cache->cached = BTRFS_CACHE_FINISHED;
7669                         add_new_free_space(cache, root->fs_info,
7670                                            found_key.objectid,
7671                                            found_key.objectid +
7672                                            found_key.offset);
7673                         free_excluded_extents(root, cache);
7674                 }
7675
7676                 ret = update_space_info(info, cache->flags, found_key.offset,
7677                                         btrfs_block_group_used(&cache->item),
7678                                         &space_info);
7679                 BUG_ON(ret); /* -ENOMEM */
7680                 cache->space_info = space_info;
7681                 spin_lock(&cache->space_info->lock);
7682                 cache->space_info->bytes_readonly += cache->bytes_super;
7683                 spin_unlock(&cache->space_info->lock);
7684
7685                 __link_block_group(space_info, cache);
7686
7687                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7688                 BUG_ON(ret); /* Logic error */
7689
7690                 set_avail_alloc_bits(root->fs_info, cache->flags);
7691                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7692                         set_block_group_ro(cache, 1);
7693         }
7694
7695         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7696                 if (!(get_alloc_profile(root, space_info->flags) &
7697                       (BTRFS_BLOCK_GROUP_RAID10 |
7698                        BTRFS_BLOCK_GROUP_RAID1 |
7699                        BTRFS_BLOCK_GROUP_DUP)))
7700                         continue;
7701                 /*
7702                  * avoid allocating from un-mirrored block group if there are
7703                  * mirrored block groups.
7704                  */
7705                 list_for_each_entry(cache, &space_info->block_groups[3], list)
7706                         set_block_group_ro(cache, 1);
7707                 list_for_each_entry(cache, &space_info->block_groups[4], list)
7708                         set_block_group_ro(cache, 1);
7709         }
7710
7711         init_global_block_rsv(info);
7712         ret = 0;
7713 error:
7714         btrfs_free_path(path);
7715         return ret;
7716 }
7717
7718 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7719                            struct btrfs_root *root, u64 bytes_used,
7720                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7721                            u64 size)
7722 {
7723         int ret;
7724         struct btrfs_root *extent_root;
7725         struct btrfs_block_group_cache *cache;
7726
7727         extent_root = root->fs_info->extent_root;
7728
7729         root->fs_info->last_trans_log_full_commit = trans->transid;
7730
7731         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7732         if (!cache)
7733                 return -ENOMEM;
7734         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7735                                         GFP_NOFS);
7736         if (!cache->free_space_ctl) {
7737                 kfree(cache);
7738                 return -ENOMEM;
7739         }
7740
7741         cache->key.objectid = chunk_offset;
7742         cache->key.offset = size;
7743         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7744         cache->sectorsize = root->sectorsize;
7745         cache->fs_info = root->fs_info;
7746
7747         atomic_set(&cache->count, 1);
7748         spin_lock_init(&cache->lock);
7749         INIT_LIST_HEAD(&cache->list);
7750         INIT_LIST_HEAD(&cache->cluster_list);
7751
7752         btrfs_init_free_space_ctl(cache);
7753
7754         btrfs_set_block_group_used(&cache->item, bytes_used);
7755         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7756         cache->flags = type;
7757         btrfs_set_block_group_flags(&cache->item, type);
7758
7759         cache->last_byte_to_unpin = (u64)-1;
7760         cache->cached = BTRFS_CACHE_FINISHED;
7761         exclude_super_stripes(root, cache);
7762
7763         add_new_free_space(cache, root->fs_info, chunk_offset,
7764                            chunk_offset + size);
7765
7766         free_excluded_extents(root, cache);
7767
7768         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7769                                 &cache->space_info);
7770         BUG_ON(ret); /* -ENOMEM */
7771         update_global_block_rsv(root->fs_info);
7772
7773         spin_lock(&cache->space_info->lock);
7774         cache->space_info->bytes_readonly += cache->bytes_super;
7775         spin_unlock(&cache->space_info->lock);
7776
7777         __link_block_group(cache->space_info, cache);
7778
7779         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7780         BUG_ON(ret); /* Logic error */
7781
7782         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
7783                                 sizeof(cache->item));
7784         if (ret) {
7785                 btrfs_abort_transaction(trans, extent_root, ret);
7786                 return ret;
7787         }
7788
7789         set_avail_alloc_bits(extent_root->fs_info, type);
7790
7791         return 0;
7792 }
7793
7794 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
7795 {
7796         u64 extra_flags = chunk_to_extended(flags) &
7797                                 BTRFS_EXTENDED_PROFILE_MASK;
7798
7799         if (flags & BTRFS_BLOCK_GROUP_DATA)
7800                 fs_info->avail_data_alloc_bits &= ~extra_flags;
7801         if (flags & BTRFS_BLOCK_GROUP_METADATA)
7802                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
7803         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
7804                 fs_info->avail_system_alloc_bits &= ~extra_flags;
7805 }
7806
7807 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7808                              struct btrfs_root *root, u64 group_start)
7809 {
7810         struct btrfs_path *path;
7811         struct btrfs_block_group_cache *block_group;
7812         struct btrfs_free_cluster *cluster;
7813         struct btrfs_root *tree_root = root->fs_info->tree_root;
7814         struct btrfs_key key;
7815         struct inode *inode;
7816         int ret;
7817         int index;
7818         int factor;
7819
7820         root = root->fs_info->extent_root;
7821
7822         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7823         BUG_ON(!block_group);
7824         BUG_ON(!block_group->ro);
7825
7826         /*
7827          * Free the reserved super bytes from this block group before
7828          * remove it.
7829          */
7830         free_excluded_extents(root, block_group);
7831
7832         memcpy(&key, &block_group->key, sizeof(key));
7833         index = get_block_group_index(block_group);
7834         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7835                                   BTRFS_BLOCK_GROUP_RAID1 |
7836                                   BTRFS_BLOCK_GROUP_RAID10))
7837                 factor = 2;
7838         else
7839                 factor = 1;
7840
7841         /* make sure this block group isn't part of an allocation cluster */
7842         cluster = &root->fs_info->data_alloc_cluster;
7843         spin_lock(&cluster->refill_lock);
7844         btrfs_return_cluster_to_free_space(block_group, cluster);
7845         spin_unlock(&cluster->refill_lock);
7846
7847         /*
7848          * make sure this block group isn't part of a metadata
7849          * allocation cluster
7850          */
7851         cluster = &root->fs_info->meta_alloc_cluster;
7852         spin_lock(&cluster->refill_lock);
7853         btrfs_return_cluster_to_free_space(block_group, cluster);
7854         spin_unlock(&cluster->refill_lock);
7855
7856         path = btrfs_alloc_path();
7857         if (!path) {
7858                 ret = -ENOMEM;
7859                 goto out;
7860         }
7861
7862         inode = lookup_free_space_inode(tree_root, block_group, path);
7863         if (!IS_ERR(inode)) {
7864                 ret = btrfs_orphan_add(trans, inode);
7865                 if (ret) {
7866                         btrfs_add_delayed_iput(inode);
7867                         goto out;
7868                 }
7869                 clear_nlink(inode);
7870                 /* One for the block groups ref */
7871                 spin_lock(&block_group->lock);
7872                 if (block_group->iref) {
7873                         block_group->iref = 0;
7874                         block_group->inode = NULL;
7875                         spin_unlock(&block_group->lock);
7876                         iput(inode);
7877                 } else {
7878                         spin_unlock(&block_group->lock);
7879                 }
7880                 /* One for our lookup ref */
7881                 btrfs_add_delayed_iput(inode);
7882         }
7883
7884         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
7885         key.offset = block_group->key.objectid;
7886         key.type = 0;
7887
7888         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
7889         if (ret < 0)
7890                 goto out;
7891         if (ret > 0)
7892                 btrfs_release_path(path);
7893         if (ret == 0) {
7894                 ret = btrfs_del_item(trans, tree_root, path);
7895                 if (ret)
7896                         goto out;
7897                 btrfs_release_path(path);
7898         }
7899
7900         spin_lock(&root->fs_info->block_group_cache_lock);
7901         rb_erase(&block_group->cache_node,
7902                  &root->fs_info->block_group_cache_tree);
7903         spin_unlock(&root->fs_info->block_group_cache_lock);
7904
7905         down_write(&block_group->space_info->groups_sem);
7906         /*
7907          * we must use list_del_init so people can check to see if they
7908          * are still on the list after taking the semaphore
7909          */
7910         list_del_init(&block_group->list);
7911         if (list_empty(&block_group->space_info->block_groups[index]))
7912                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
7913         up_write(&block_group->space_info->groups_sem);
7914
7915         if (block_group->cached == BTRFS_CACHE_STARTED)
7916                 wait_block_group_cache_done(block_group);
7917
7918         btrfs_remove_free_space_cache(block_group);
7919
7920         spin_lock(&block_group->space_info->lock);
7921         block_group->space_info->total_bytes -= block_group->key.offset;
7922         block_group->space_info->bytes_readonly -= block_group->key.offset;
7923         block_group->space_info->disk_total -= block_group->key.offset * factor;
7924         spin_unlock(&block_group->space_info->lock);
7925
7926         memcpy(&key, &block_group->key, sizeof(key));
7927
7928         btrfs_clear_space_info_full(root->fs_info);
7929
7930         btrfs_put_block_group(block_group);
7931         btrfs_put_block_group(block_group);
7932
7933         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
7934         if (ret > 0)
7935                 ret = -EIO;
7936         if (ret < 0)
7937                 goto out;
7938
7939         ret = btrfs_del_item(trans, root, path);
7940 out:
7941         btrfs_free_path(path);
7942         return ret;
7943 }
7944
7945 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
7946 {
7947         struct btrfs_space_info *space_info;
7948         struct btrfs_super_block *disk_super;
7949         u64 features;
7950         u64 flags;
7951         int mixed = 0;
7952         int ret;
7953
7954         disk_super = fs_info->super_copy;
7955         if (!btrfs_super_root(disk_super))
7956                 return 1;
7957
7958         features = btrfs_super_incompat_flags(disk_super);
7959         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
7960                 mixed = 1;
7961
7962         flags = BTRFS_BLOCK_GROUP_SYSTEM;
7963         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7964         if (ret)
7965                 goto out;
7966
7967         if (mixed) {
7968                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
7969                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7970         } else {
7971                 flags = BTRFS_BLOCK_GROUP_METADATA;
7972                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7973                 if (ret)
7974                         goto out;
7975
7976                 flags = BTRFS_BLOCK_GROUP_DATA;
7977                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7978         }
7979 out:
7980         return ret;
7981 }
7982
7983 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
7984 {
7985         return unpin_extent_range(root, start, end);
7986 }
7987
7988 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
7989                                u64 num_bytes, u64 *actual_bytes)
7990 {
7991         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
7992 }
7993
7994 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
7995 {
7996         struct btrfs_fs_info *fs_info = root->fs_info;
7997         struct btrfs_block_group_cache *cache = NULL;
7998         u64 group_trimmed;
7999         u64 start;
8000         u64 end;
8001         u64 trimmed = 0;
8002         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8003         int ret = 0;
8004
8005         /*
8006          * try to trim all FS space, our block group may start from non-zero.
8007          */
8008         if (range->len == total_bytes)
8009                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8010         else
8011                 cache = btrfs_lookup_block_group(fs_info, range->start);
8012
8013         while (cache) {
8014                 if (cache->key.objectid >= (range->start + range->len)) {
8015                         btrfs_put_block_group(cache);
8016                         break;
8017                 }
8018
8019                 start = max(range->start, cache->key.objectid);
8020                 end = min(range->start + range->len,
8021                                 cache->key.objectid + cache->key.offset);
8022
8023                 if (end - start >= range->minlen) {
8024                         if (!block_group_cache_done(cache)) {
8025                                 ret = cache_block_group(cache, NULL, root, 0);
8026                                 if (!ret)
8027                                         wait_block_group_cache_done(cache);
8028                         }
8029                         ret = btrfs_trim_block_group(cache,
8030                                                      &group_trimmed,
8031                                                      start,
8032                                                      end,
8033                                                      range->minlen);
8034
8035                         trimmed += group_trimmed;
8036                         if (ret) {
8037                                 btrfs_put_block_group(cache);
8038                                 break;
8039                         }
8040                 }
8041
8042                 cache = next_block_group(fs_info->tree_root, cache);
8043         }
8044
8045         range->len = trimmed;
8046         return ret;
8047 }