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 #undef SCRAMBLE_DELAYED_REFS
  38
  39 /*
  40  * control flags for do_chunk_alloc's force field
  41  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  42  * if we really need one.
  43  *
  44  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  45  * if we have very few chunks already allocated.  This is
  46  * used as part of the clustering code to help make sure
  47  * we have a good pool of storage to cluster in, without
  48  * filling the FS with empty chunks
  49  *
  50  * CHUNK_ALLOC_FORCE means it must try to allocate one
  51  *
  52  */
  53 enum {
  54         CHUNK_ALLOC_NO_FORCE = 0,
  55         CHUNK_ALLOC_LIMITED = 1,
  56         CHUNK_ALLOC_FORCE = 2,
  57 };
  58
  59 /*
  60  * Control how reservations are dealt with.
  61  *
  62  * RESERVE_FREE - freeing a reservation.
  63  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  64  *   ENOSPC accounting
  65  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  66  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  67  */
  68 enum {
  69         RESERVE_FREE = 0,
  70         RESERVE_ALLOC = 1,
  71         RESERVE_ALLOC_NO_ACCOUNT = 2,
  72 };
  73
  74 static int update_block_group(struct btrfs_trans_handle *trans,
  75                               struct btrfs_root *root,
  76                               u64 bytenr, u64 num_bytes, int alloc);
  77 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  78                                 struct btrfs_root *root,
  79                                 u64 bytenr, u64 num_bytes, u64 parent,
  80                                 u64 root_objectid, u64 owner_objectid,
  81                                 u64 owner_offset, int refs_to_drop,
  82                                 struct btrfs_delayed_extent_op *extra_op);
  83 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  84                                     struct extent_buffer *leaf,
  85                                     struct btrfs_extent_item *ei);
  86 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  87                                       struct btrfs_root *root,
  88                                       u64 parent, u64 root_objectid,
  89                                       u64 flags, u64 owner, u64 offset,
  90                                       struct btrfs_key *ins, int ref_mod);
  91 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  92                                      struct btrfs_root *root,
  93                                      u64 parent, u64 root_objectid,
  94                                      u64 flags, struct btrfs_disk_key *key,
  95                                      int level, struct btrfs_key *ins);
  96 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  97                           struct btrfs_root *extent_root, u64 alloc_bytes,
  98                           u64 flags, int force);
  99 static int find_next_key(struct btrfs_path *path, int level,
 100                          struct btrfs_key *key);
 101 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 102                             int dump_block_groups);
 103 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 104                                        u64 num_bytes, int reserve);
 105
 106 static noinline int
 107 block_group_cache_done(struct btrfs_block_group_cache *cache)
 108 {
 109         smp_mb();
 110         return cache->cached == BTRFS_CACHE_FINISHED;
 111 }
 112
 113 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 114 {
 115         return (cache->flags & bits) == bits;
 116 }
 117
 118 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 119 {
 120         atomic_inc(&cache->count);
 121 }
 122
 123 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 124 {
 125         if (atomic_dec_and_test(&cache->count)) {
 126                 WARN_ON(cache->pinned > 0);
 127                 WARN_ON(cache->reserved > 0);
 128                 kfree(cache->free_space_ctl);
 129                 kfree(cache);
 130         }
 131 }
 132
 133 /*
 134  * this adds the block group to the fs_info rb tree for the block group
 135  * cache
 136  */
 137 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 138                                 struct btrfs_block_group_cache *block_group)
 139 {
 140         struct rb_node **p;
 141         struct rb_node *parent = NULL;
 142         struct btrfs_block_group_cache *cache;
 143
 144         spin_lock(&info->block_group_cache_lock);
 145         p = &info->block_group_cache_tree.rb_node;
 146
 147         while (*p) {
 148                 parent = *p;
 149                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 150                                  cache_node);
 151                 if (block_group->key.objectid < cache->key.objectid) {
 152                         p = &(*p)->rb_left;
 153                 } else if (block_group->key.objectid > cache->key.objectid) {
 154                         p = &(*p)->rb_right;
 155                 } else {
 156                         spin_unlock(&info->block_group_cache_lock);
 157                         return -EEXIST;
 158                 }
 159         }
 160
 161         rb_link_node(&block_group->cache_node, parent, p);
 162         rb_insert_color(&block_group->cache_node,
 163                         &info->block_group_cache_tree);
 164         spin_unlock(&info->block_group_cache_lock);
 165
 166         return 0;
 167 }
 168
 169 /*
 170  * This will return the block group at or after bytenr if contains is 0, else
 171  * it will return the block group that contains the bytenr
 172  */
 173 static struct btrfs_block_group_cache *
 174 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 175                               int contains)
 176 {
 177         struct btrfs_block_group_cache *cache, *ret = NULL;
 178         struct rb_node *n;
 179         u64 end, start;
 180
 181         spin_lock(&info->block_group_cache_lock);
 182         n = info->block_group_cache_tree.rb_node;
 183
 184         while (n) {
 185                 cache = rb_entry(n, struct btrfs_block_group_cache,
 186                                  cache_node);
 187                 end = cache->key.objectid + cache->key.offset - 1;
 188                 start = cache->key.objectid;
 189
 190                 if (bytenr < start) {
 191                         if (!contains && (!ret || start < ret->key.objectid))
 192                                 ret = cache;
 193                         n = n->rb_left;
 194                 } else if (bytenr > start) {
 195                         if (contains && bytenr <= end) {
 196                                 ret = cache;
 197                                 break;
 198                         }
 199                         n = n->rb_right;
 200                 } else {
 201                         ret = cache;
 202                         break;
 203                 }
 204         }
 205         if (ret)
 206                 btrfs_get_block_group(ret);
 207         spin_unlock(&info->block_group_cache_lock);
 208
 209         return ret;
 210 }
 211
 212 static int add_excluded_extent(struct btrfs_root *root,
 213                                u64 start, u64 num_bytes)
 214 {
 215         u64 end = start + num_bytes - 1;
 216         set_extent_bits(&root->fs_info->freed_extents[0],
 217                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 218         set_extent_bits(&root->fs_info->freed_extents[1],
 219                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 220         return 0;
 221 }
 222
 223 static void free_excluded_extents(struct btrfs_root *root,
 224                                   struct btrfs_block_group_cache *cache)
 225 {
 226         u64 start, end;
 227
 228         start = cache->key.objectid;
 229         end = start + cache->key.offset - 1;
 230
 231         clear_extent_bits(&root->fs_info->freed_extents[0],
 232                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 233         clear_extent_bits(&root->fs_info->freed_extents[1],
 234                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 235 }
 236
 237 static int exclude_super_stripes(struct btrfs_root *root,
 238                                  struct btrfs_block_group_cache *cache)
 239 {
 240         u64 bytenr;
 241         u64 *logical;
 242         int stripe_len;
 243         int i, nr, ret;
 244
 245         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 246                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 247                 cache->bytes_super += stripe_len;
 248                 ret = add_excluded_extent(root, cache->key.objectid,
 249                                           stripe_len);
 250                 BUG_ON(ret); /* -ENOMEM */
 251         }
 252
 253         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 254                 bytenr = btrfs_sb_offset(i);
 255                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 256                                        cache->key.objectid, bytenr,
 257                                        0, &logical, &nr, &stripe_len);
 258                 BUG_ON(ret); /* -ENOMEM */
 259
 260                 while (nr--) {
 261                         cache->bytes_super += stripe_len;
 262                         ret = add_excluded_extent(root, logical[nr],
 263                                                   stripe_len);
 264                         BUG_ON(ret); /* -ENOMEM */
 265                 }
 266
 267                 kfree(logical);
 268         }
 269         return 0;
 270 }
 271
 272 static struct btrfs_caching_control *
 273 get_caching_control(struct btrfs_block_group_cache *cache)
 274 {
 275         struct btrfs_caching_control *ctl;
 276
 277         spin_lock(&cache->lock);
 278         if (cache->cached != BTRFS_CACHE_STARTED) {
 279                 spin_unlock(&cache->lock);
 280                 return NULL;
 281         }
 282
 283         /* We're loading it the fast way, so we don't have a caching_ctl. */
 284         if (!cache->caching_ctl) {
 285                 spin_unlock(&cache->lock);
 286                 return NULL;
 287         }
 288
 289         ctl = cache->caching_ctl;
 290         atomic_inc(&ctl->count);
 291         spin_unlock(&cache->lock);
 292         return ctl;
 293 }
 294
 295 static void put_caching_control(struct btrfs_caching_control *ctl)
 296 {
 297         if (atomic_dec_and_test(&ctl->count))
 298                 kfree(ctl);
 299 }
 300
 301 /*
 302  * this is only called by cache_block_group, since we could have freed extents
 303  * we need to check the pinned_extents for any extents that can't be used yet
 304  * since their free space will be released as soon as the transaction commits.
 305  */
 306 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 307                               struct btrfs_fs_info *info, u64 start, u64 end)
 308 {
 309         u64 extent_start, extent_end, size, total_added = 0;
 310         int ret;
 311
 312         while (start < end) {
 313                 ret = find_first_extent_bit(info->pinned_extents, start,
 314                                             &extent_start, &extent_end,
 315                                             EXTENT_DIRTY | EXTENT_UPTODATE);
 316                 if (ret)
 317                         break;
 318
 319                 if (extent_start <= start) {
 320                         start = extent_end + 1;
 321                 } else if (extent_start > start && extent_start < end) {
 322                         size = extent_start - start;
 323                         total_added += size;
 324                         ret = btrfs_add_free_space(block_group, start,
 325                                                    size);
 326                         BUG_ON(ret); /* -ENOMEM or logic error */
 327                         start = extent_end + 1;
 328                 } else {
 329                         break;
 330                 }
 331         }
 332
 333         if (start < end) {
 334                 size = end - start;
 335                 total_added += size;
 336                 ret = btrfs_add_free_space(block_group, start, size);
 337                 BUG_ON(ret); /* -ENOMEM or logic error */
 338         }
 339
 340         return total_added;
 341 }
 342
 343 static noinline void caching_thread(struct btrfs_work *work)
 344 {
 345         struct btrfs_block_group_cache *block_group;
 346         struct btrfs_fs_info *fs_info;
 347         struct btrfs_caching_control *caching_ctl;
 348         struct btrfs_root *extent_root;
 349         struct btrfs_path *path;
 350         struct extent_buffer *leaf;
 351         struct btrfs_key key;
 352         u64 total_found = 0;
 353         u64 last = 0;
 354         u32 nritems;
 355         int ret = 0;
 356
 357         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 358         block_group = caching_ctl->block_group;
 359         fs_info = block_group->fs_info;
 360         extent_root = fs_info->extent_root;
 361
 362         path = btrfs_alloc_path();
 363         if (!path)
 364                 goto out;
 365
 366         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 367
 368         /*
 369          * We don't want to deadlock with somebody trying to allocate a new
 370          * extent for the extent root while also trying to search the extent
 371          * root to add free space.  So we skip locking and search the commit
 372          * root, since its read-only
 373          */
 374         path->skip_locking = 1;
 375         path->search_commit_root = 1;
 376         path->reada = 1;
 377
 378         key.objectid = last;
 379         key.offset = 0;
 380         key.type = BTRFS_EXTENT_ITEM_KEY;
 381 again:
 382         mutex_lock(&caching_ctl->mutex);
 383         /* need to make sure the commit_root doesn't disappear */
 384         down_read(&fs_info->extent_commit_sem);
 385
 386         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 387         if (ret < 0)
 388                 goto err;
 389
 390         leaf = path->nodes[0];
 391         nritems = btrfs_header_nritems(leaf);
 392
 393         while (1) {
 394                 if (btrfs_fs_closing(fs_info) > 1) {
 395                         last = (u64)-1;
 396                         break;
 397                 }
 398
 399                 if (path->slots[0] < nritems) {
 400                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 401                 } else {
 402                         ret = find_next_key(path, 0, &key);
 403                         if (ret)
 404                                 break;
 405
 406                         if (need_resched() ||
 407                             btrfs_next_leaf(extent_root, path)) {
 408                                 caching_ctl->progress = last;
 409                                 btrfs_release_path(path);
 410                                 up_read(&fs_info->extent_commit_sem);
 411                                 mutex_unlock(&caching_ctl->mutex);
 412                                 cond_resched();
 413                                 goto again;
 414                         }
 415                         leaf = path->nodes[0];
 416                         nritems = btrfs_header_nritems(leaf);
 417                         continue;
 418                 }
 419
 420                 if (key.objectid < block_group->key.objectid) {
 421                         path->slots[0]++;
 422                         continue;
 423                 }
 424
 425                 if (key.objectid >= block_group->key.objectid +
 426                     block_group->key.offset)
 427                         break;
 428
 429                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 430                         total_found += add_new_free_space(block_group,
 431                                                           fs_info, last,
 432                                                           key.objectid);
 433                         last = key.objectid + key.offset;
 434
 435                         if (total_found > (1024 * 1024 * 2)) {
 436                                 total_found = 0;
 437                                 wake_up(&caching_ctl->wait);
 438                         }
 439                 }
 440                 path->slots[0]++;
 441         }
 442         ret = 0;
 443
 444         total_found += add_new_free_space(block_group, fs_info, last,
 445                                           block_group->key.objectid +
 446                                           block_group->key.offset);
 447         caching_ctl->progress = (u64)-1;
 448
 449         spin_lock(&block_group->lock);
 450         block_group->caching_ctl = NULL;
 451         block_group->cached = BTRFS_CACHE_FINISHED;
 452         spin_unlock(&block_group->lock);
 453
 454 err:
 455         btrfs_free_path(path);
 456         up_read(&fs_info->extent_commit_sem);
 457
 458         free_excluded_extents(extent_root, block_group);
 459
 460         mutex_unlock(&caching_ctl->mutex);
 461 out:
 462         wake_up(&caching_ctl->wait);
 463
 464         put_caching_control(caching_ctl);
 465         btrfs_put_block_group(block_group);
 466 }
 467
 468 static int cache_block_group(struct btrfs_block_group_cache *cache,
 469                              struct btrfs_trans_handle *trans,
 470                              struct btrfs_root *root,
 471                              int load_cache_only)
 472 {
 473         DEFINE_WAIT(wait);
 474         struct btrfs_fs_info *fs_info = cache->fs_info;
 475         struct btrfs_caching_control *caching_ctl;
 476         int ret = 0;
 477
 478         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 479         if (!caching_ctl)
 480                 return -ENOMEM;
 481
 482         INIT_LIST_HEAD(&caching_ctl->list);
 483         mutex_init(&caching_ctl->mutex);
 484         init_waitqueue_head(&caching_ctl->wait);
 485         caching_ctl->block_group = cache;
 486         caching_ctl->progress = cache->key.objectid;
 487         atomic_set(&caching_ctl->count, 1);
 488         caching_ctl->work.func = caching_thread;
 489
 490         spin_lock(&cache->lock);
 491         /*
 492          * This should be a rare occasion, but this could happen I think in the
 493          * case where one thread starts to load the space cache info, and then
 494          * some other thread starts a transaction commit which tries to do an
 495          * allocation while the other thread is still loading the space cache
 496          * info.  The previous loop should have kept us from choosing this block
 497          * group, but if we've moved to the state where we will wait on caching
 498          * block groups we need to first check if we're doing a fast load here,
 499          * so we can wait for it to finish, otherwise we could end up allocating
 500          * from a block group who's cache gets evicted for one reason or
 501          * another.
 502          */
 503         while (cache->cached == BTRFS_CACHE_FAST) {
 504                 struct btrfs_caching_control *ctl;
 505
 506                 ctl = cache->caching_ctl;
 507                 atomic_inc(&ctl->count);
 508                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 509                 spin_unlock(&cache->lock);
 510
 511                 schedule();
 512
 513                 finish_wait(&ctl->wait, &wait);
 514                 put_caching_control(ctl);
 515                 spin_lock(&cache->lock);
 516         }
 517
 518         if (cache->cached != BTRFS_CACHE_NO) {
 519                 spin_unlock(&cache->lock);
 520                 kfree(caching_ctl);
 521                 return 0;
 522         }
 523         WARN_ON(cache->caching_ctl);
 524         cache->caching_ctl = caching_ctl;
 525         cache->cached = BTRFS_CACHE_FAST;
 526         spin_unlock(&cache->lock);
 527
 528         /*
 529          * We can't do the read from on-disk cache during a commit since we need
 530          * to have the normal tree locking.  Also if we are currently trying to
 531          * allocate blocks for the tree root we can't do the fast caching since
 532          * we likely hold important locks.
 533          */
 534         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 535                 ret = load_free_space_cache(fs_info, cache);
 536
 537                 spin_lock(&cache->lock);
 538                 if (ret == 1) {
 539                         cache->caching_ctl = NULL;
 540                         cache->cached = BTRFS_CACHE_FINISHED;
 541                         cache->last_byte_to_unpin = (u64)-1;
 542                 } else {
 543                         if (load_cache_only) {
 544                                 cache->caching_ctl = NULL;
 545                                 cache->cached = BTRFS_CACHE_NO;
 546                         } else {
 547                                 cache->cached = BTRFS_CACHE_STARTED;
 548                         }
 549                 }
 550                 spin_unlock(&cache->lock);
 551                 wake_up(&caching_ctl->wait);
 552                 if (ret == 1) {
 553                         put_caching_control(caching_ctl);
 554                         free_excluded_extents(fs_info->extent_root, cache);
 555                         return 0;
 556                 }
 557         } else {
 558                 /*
 559                  * We are not going to do the fast caching, set cached to the
 560                  * appropriate value and wakeup any waiters.
 561                  */
 562                 spin_lock(&cache->lock);
 563                 if (load_cache_only) {
 564                         cache->caching_ctl = NULL;
 565                         cache->cached = BTRFS_CACHE_NO;
 566                 } else {
 567                         cache->cached = BTRFS_CACHE_STARTED;
 568                 }
 569                 spin_unlock(&cache->lock);
 570                 wake_up(&caching_ctl->wait);
 571         }
 572
 573         if (load_cache_only) {
 574                 put_caching_control(caching_ctl);
 575                 return 0;
 576         }
 577
 578         down_write(&fs_info->extent_commit_sem);
 579         atomic_inc(&caching_ctl->count);
 580         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 581         up_write(&fs_info->extent_commit_sem);
 582
 583         btrfs_get_block_group(cache);
 584
 585         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 586
 587         return ret;
 588 }
 589
 590 /*
 591  * return the block group that starts at or after bytenr
 592  */
 593 static struct btrfs_block_group_cache *
 594 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 595 {
 596         struct btrfs_block_group_cache *cache;
 597
 598         cache = block_group_cache_tree_search(info, bytenr, 0);
 599
 600         return cache;
 601 }
 602
 603 /*
 604  * return the block group that contains the given bytenr
 605  */
 606 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 607                                                  struct btrfs_fs_info *info,
 608                                                  u64 bytenr)
 609 {
 610         struct btrfs_block_group_cache *cache;
 611
 612         cache = block_group_cache_tree_search(info, bytenr, 1);
 613
 614         return cache;
 615 }
 616
 617 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 618                                                   u64 flags)
 619 {
 620         struct list_head *head = &info->space_info;
 621         struct btrfs_space_info *found;
 622
 623         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 624
 625         rcu_read_lock();
 626         list_for_each_entry_rcu(found, head, list) {
 627                 if (found->flags & flags) {
 628                         rcu_read_unlock();
 629                         return found;
 630                 }
 631         }
 632         rcu_read_unlock();
 633         return NULL;
 634 }
 635
 636 /*
 637  * after adding space to the filesystem, we need to clear the full flags
 638  * on all the space infos.
 639  */
 640 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 641 {
 642         struct list_head *head = &info->space_info;
 643         struct btrfs_space_info *found;
 644
 645         rcu_read_lock();
 646         list_for_each_entry_rcu(found, head, list)
 647                 found->full = 0;
 648         rcu_read_unlock();
 649 }
 650
 651 static u64 div_factor(u64 num, int factor)
 652 {
 653         if (factor == 10)
 654                 return num;
 655         num *= factor;
 656         do_div(num, 10);
 657         return num;
 658 }
 659
 660 static u64 div_factor_fine(u64 num, int factor)
 661 {
 662         if (factor == 100)
 663                 return num;
 664         num *= factor;
 665         do_div(num, 100);
 666         return num;
 667 }
 668
 669 u64 btrfs_find_block_group(struct btrfs_root *root,
 670                            u64 search_start, u64 search_hint, int owner)
 671 {
 672         struct btrfs_block_group_cache *cache;
 673         u64 used;
 674         u64 last = max(search_hint, search_start);
 675         u64 group_start = 0;
 676         int full_search = 0;
 677         int factor = 9;
 678         int wrapped = 0;
 679 again:
 680         while (1) {
 681                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
 682                 if (!cache)
 683                         break;
 684
 685                 spin_lock(&cache->lock);
 686                 last = cache->key.objectid + cache->key.offset;
 687                 used = btrfs_block_group_used(&cache->item);
 688
 689                 if ((full_search || !cache->ro) &&
 690                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
 691                         if (used + cache->pinned + cache->reserved <
 692                             div_factor(cache->key.offset, factor)) {
 693                                 group_start = cache->key.objectid;
 694                                 spin_unlock(&cache->lock);
 695                                 btrfs_put_block_group(cache);
 696                                 goto found;
 697                         }
 698                 }
 699                 spin_unlock(&cache->lock);
 700                 btrfs_put_block_group(cache);
 701                 cond_resched();
 702         }
 703         if (!wrapped) {
 704                 last = search_start;
 705                 wrapped = 1;
 706                 goto again;
 707         }
 708         if (!full_search && factor < 10) {
 709                 last = search_start;
 710                 full_search = 1;
 711                 factor = 10;
 712                 goto again;
 713         }
 714 found:
 715         return group_start;
 716 }
 717
 718 /* simple helper to search for an existing extent at a given offset */
 719 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 720 {
 721         int ret;
 722         struct btrfs_key key;
 723         struct btrfs_path *path;
 724
 725         path = btrfs_alloc_path();
 726         if (!path)
 727                 return -ENOMEM;
 728
 729         key.objectid = start;
 730         key.offset = len;
 731         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 732         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 733                                 0, 0);
 734         btrfs_free_path(path);
 735         return ret;
 736 }
 737
 738 /*
 739  * helper function to lookup reference count and flags of extent.
 740  *
 741  * the head node for delayed ref is used to store the sum of all the
 742  * reference count modifications queued up in the rbtree. the head
 743  * node may also store the extent flags to set. This way you can check
 744  * to see what the reference count and extent flags would be if all of
 745  * the delayed refs are not processed.
 746  */
 747 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 748                              struct btrfs_root *root, u64 bytenr,
 749                              u64 num_bytes, u64 *refs, u64 *flags)
 750 {
 751         struct btrfs_delayed_ref_head *head;
 752         struct btrfs_delayed_ref_root *delayed_refs;
 753         struct btrfs_path *path;
 754         struct btrfs_extent_item *ei;
 755         struct extent_buffer *leaf;
 756         struct btrfs_key key;
 757         u32 item_size;
 758         u64 num_refs;
 759         u64 extent_flags;
 760         int ret;
 761
 762         path = btrfs_alloc_path();
 763         if (!path)
 764                 return -ENOMEM;
 765
 766         key.objectid = bytenr;
 767         key.type = BTRFS_EXTENT_ITEM_KEY;
 768         key.offset = num_bytes;
 769         if (!trans) {
 770                 path->skip_locking = 1;
 771                 path->search_commit_root = 1;
 772         }
 773 again:
 774         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 775                                 &key, path, 0, 0);
 776         if (ret < 0)
 777                 goto out_free;
 778
 779         if (ret == 0) {
 780                 leaf = path->nodes[0];
 781                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 782                 if (item_size >= sizeof(*ei)) {
 783                         ei = btrfs_item_ptr(leaf, path->slots[0],
 784                                             struct btrfs_extent_item);
 785                         num_refs = btrfs_extent_refs(leaf, ei);
 786                         extent_flags = btrfs_extent_flags(leaf, ei);
 787                 } else {
 788 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 789                         struct btrfs_extent_item_v0 *ei0;
 790                         BUG_ON(item_size != sizeof(*ei0));
 791                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 792                                              struct btrfs_extent_item_v0);
 793                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 794                         /* FIXME: this isn't correct for data */
 795                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 796 #else
 797                         BUG();
 798 #endif
 799                 }
 800                 BUG_ON(num_refs == 0);
 801         } else {
 802                 num_refs = 0;
 803                 extent_flags = 0;
 804                 ret = 0;
 805         }
 806
 807         if (!trans)
 808                 goto out;
 809
 810         delayed_refs = &trans->transaction->delayed_refs;
 811         spin_lock(&delayed_refs->lock);
 812         head = btrfs_find_delayed_ref_head(trans, bytenr);
 813         if (head) {
 814                 if (!mutex_trylock(&head->mutex)) {
 815                         atomic_inc(&head->node.refs);
 816                         spin_unlock(&delayed_refs->lock);
 817
 818                         btrfs_release_path(path);
 819
 820                         /*
 821                          * Mutex was contended, block until it's released and try
 822                          * again
 823                          */
 824                         mutex_lock(&head->mutex);
 825                         mutex_unlock(&head->mutex);
 826                         btrfs_put_delayed_ref(&head->node);
 827                         goto again;
 828                 }
 829                 if (head->extent_op && head->extent_op->update_flags)
 830                         extent_flags |= head->extent_op->flags_to_set;
 831                 else
 832                         BUG_ON(num_refs == 0);
 833
 834                 num_refs += head->node.ref_mod;
 835                 mutex_unlock(&head->mutex);
 836         }
 837         spin_unlock(&delayed_refs->lock);
 838 out:
 839         WARN_ON(num_refs == 0);
 840         if (refs)
 841                 *refs = num_refs;
 842         if (flags)
 843                 *flags = extent_flags;
 844 out_free:
 845         btrfs_free_path(path);
 846         return ret;
 847 }
 848
 849 /*
 850  * Back reference rules.  Back refs have three main goals:
 851  *
 852  * 1) differentiate between all holders of references to an extent so that
 853  *    when a reference is dropped we can make sure it was a valid reference
 854  *    before freeing the extent.
 855  *
 856  * 2) Provide enough information to quickly find the holders of an extent
 857  *    if we notice a given block is corrupted or bad.
 858  *
 859  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 860  *    maintenance.  This is actually the same as #2, but with a slightly
 861  *    different use case.
 862  *
 863  * There are two kinds of back refs. The implicit back refs is optimized
 864  * for pointers in non-shared tree blocks. For a given pointer in a block,
 865  * back refs of this kind provide information about the block's owner tree
 866  * and the pointer's key. These information allow us to find the block by
 867  * b-tree searching. The full back refs is for pointers in tree blocks not
 868  * referenced by their owner trees. The location of tree block is recorded
 869  * in the back refs. Actually the full back refs is generic, and can be
 870  * used in all cases the implicit back refs is used. The major shortcoming
 871  * of the full back refs is its overhead. Every time a tree block gets
 872  * COWed, we have to update back refs entry for all pointers in it.
 873  *
 874  * For a newly allocated tree block, we use implicit back refs for
 875  * pointers in it. This means most tree related operations only involve
 876  * implicit back refs. For a tree block created in old transaction, the
 877  * only way to drop a reference to it is COW it. So we can detect the
 878  * event that tree block loses its owner tree's reference and do the
 879  * back refs conversion.
 880  *
 881  * When a tree block is COW'd through a tree, there are four cases:
 882  *
 883  * The reference count of the block is one and the tree is the block's
 884  * owner tree. Nothing to do in this case.
 885  *
 886  * The reference count of the block is one and the tree is not the
 887  * block's owner tree. In this case, full back refs is used for pointers
 888  * in the block. Remove these full back refs, add implicit back refs for
 889  * every pointers in the new block.
 890  *
 891  * The reference count of the block is greater than one and the tree is
 892  * the block's owner tree. In this case, implicit back refs is used for
 893  * pointers in the block. Add full back refs for every pointers in the
 894  * block, increase lower level extents' reference counts. The original
 895  * implicit back refs are entailed to the new block.
 896  *
 897  * The reference count of the block is greater than one and the tree is
 898  * not the block's owner tree. Add implicit back refs for every pointer in
 899  * the new block, increase lower level extents' reference count.
 900  *
 901  * Back Reference Key composing:
 902  *
 903  * The key objectid corresponds to the first byte in the extent,
 904  * The key type is used to differentiate between types of back refs.
 905  * There are different meanings of the key offset for different types
 906  * of back refs.
 907  *
 908  * File extents can be referenced by:
 909  *
 910  * - multiple snapshots, subvolumes, or different generations in one subvol
 911  * - different files inside a single subvolume
 912  * - different offsets inside a file (bookend extents in file.c)
 913  *
 914  * The extent ref structure for the implicit back refs has fields for:
 915  *
 916  * - Objectid of the subvolume root
 917  * - objectid of the file holding the reference
 918  * - original offset in the file
 919  * - how many bookend extents
 920  *
 921  * The key offset for the implicit back refs is hash of the first
 922  * three fields.
 923  *
 924  * The extent ref structure for the full back refs has field for:
 925  *
 926  * - number of pointers in the tree leaf
 927  *
 928  * The key offset for the implicit back refs is the first byte of
 929  * the tree leaf
 930  *
 931  * When a file extent is allocated, The implicit back refs is used.
 932  * the fields are filled in:
 933  *
 934  *     (root_key.objectid, inode objectid, offset in file, 1)
 935  *
 936  * When a file extent is removed file truncation, we find the
 937  * corresponding implicit back refs and check the following fields:
 938  *
 939  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 940  *
 941  * Btree extents can be referenced by:
 942  *
 943  * - Different subvolumes
 944  *
 945  * Both the implicit back refs and the full back refs for tree blocks
 946  * only consist of key. The key offset for the implicit back refs is
 947  * objectid of block's owner tree. The key offset for the full back refs
 948  * is the first byte of parent block.
 949  *
 950  * When implicit back refs is used, information about the lowest key and
 951  * level of the tree block are required. These information are stored in
 952  * tree block info structure.
 953  */
 954
 955 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 956 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 957                                   struct btrfs_root *root,
 958                                   struct btrfs_path *path,
 959                                   u64 owner, u32 extra_size)
 960 {
 961         struct btrfs_extent_item *item;
 962         struct btrfs_extent_item_v0 *ei0;
 963         struct btrfs_extent_ref_v0 *ref0;
 964         struct btrfs_tree_block_info *bi;
 965         struct extent_buffer *leaf;
 966         struct btrfs_key key;
 967         struct btrfs_key found_key;
 968         u32 new_size = sizeof(*item);
 969         u64 refs;
 970         int ret;
 971
 972         leaf = path->nodes[0];
 973         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 974
 975         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 976         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 977                              struct btrfs_extent_item_v0);
 978         refs = btrfs_extent_refs_v0(leaf, ei0);
 979
 980         if (owner == (u64)-1) {
 981                 while (1) {
 982                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 983                                 ret = btrfs_next_leaf(root, path);
 984                                 if (ret < 0)
 985                                         return ret;
 986                                 BUG_ON(ret > 0); /* Corruption */
 987                                 leaf = path->nodes[0];
 988                         }
 989                         btrfs_item_key_to_cpu(leaf, &found_key,
 990                                               path->slots[0]);
 991                         BUG_ON(key.objectid != found_key.objectid);
 992                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 993                                 path->slots[0]++;
 994                                 continue;
 995                         }
 996                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 997                                               struct btrfs_extent_ref_v0);
 998                         owner = btrfs_ref_objectid_v0(leaf, ref0);
 999                         break;
1000                 }
1001         }
1002         btrfs_release_path(path);
1003
1004         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1005                 new_size += sizeof(*bi);
1006
1007         new_size -= sizeof(*ei0);
1008         ret = btrfs_search_slot(trans, root, &key, path,
1009                                 new_size + extra_size, 1);
1010         if (ret < 0)
1011                 return ret;
1012         BUG_ON(ret); /* Corruption */
1013
1014         btrfs_extend_item(trans, root, path, new_size);
1015
1016         leaf = path->nodes[0];
1017         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1018         btrfs_set_extent_refs(leaf, item, refs);
1019         /* FIXME: get real generation */
1020         btrfs_set_extent_generation(leaf, item, 0);
1021         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1022                 btrfs_set_extent_flags(leaf, item,
1023                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1024                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1025                 bi = (struct btrfs_tree_block_info *)(item + 1);
1026                 /* FIXME: get first key of the block */
1027                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1028                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1029         } else {
1030                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1031         }
1032         btrfs_mark_buffer_dirty(leaf);
1033         return 0;
1034 }
1035 #endif
1036
1037 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1038 {
1039         u32 high_crc = ~(u32)0;
1040         u32 low_crc = ~(u32)0;
1041         __le64 lenum;
1042
1043         lenum = cpu_to_le64(root_objectid);
1044         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1045         lenum = cpu_to_le64(owner);
1046         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1047         lenum = cpu_to_le64(offset);
1048         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1049
1050         return ((u64)high_crc << 31) ^ (u64)low_crc;
1051 }
1052
1053 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1054                                      struct btrfs_extent_data_ref *ref)
1055 {
1056         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1057                                     btrfs_extent_data_ref_objectid(leaf, ref),
1058                                     btrfs_extent_data_ref_offset(leaf, ref));
1059 }
1060
1061 static int match_extent_data_ref(struct extent_buffer *leaf,
1062                                  struct btrfs_extent_data_ref *ref,
1063                                  u64 root_objectid, u64 owner, u64 offset)
1064 {
1065         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1066             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1067             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1068                 return 0;
1069         return 1;
1070 }
1071
1072 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1073                                            struct btrfs_root *root,
1074                                            struct btrfs_path *path,
1075                                            u64 bytenr, u64 parent,
1076                                            u64 root_objectid,
1077                                            u64 owner, u64 offset)
1078 {
1079         struct btrfs_key key;
1080         struct btrfs_extent_data_ref *ref;
1081         struct extent_buffer *leaf;
1082         u32 nritems;
1083         int ret;
1084         int recow;
1085         int err = -ENOENT;
1086
1087         key.objectid = bytenr;
1088         if (parent) {
1089                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1090                 key.offset = parent;
1091         } else {
1092                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1093                 key.offset = hash_extent_data_ref(root_objectid,
1094                                                   owner, offset);
1095         }
1096 again:
1097         recow = 0;
1098         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1099         if (ret < 0) {
1100                 err = ret;
1101                 goto fail;
1102         }
1103
1104         if (parent) {
1105                 if (!ret)
1106                         return 0;
1107 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1108                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1109                 btrfs_release_path(path);
1110                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1111                 if (ret < 0) {
1112                         err = ret;
1113                         goto fail;
1114                 }
1115                 if (!ret)
1116                         return 0;
1117 #endif
1118                 goto fail;
1119         }
1120
1121         leaf = path->nodes[0];
1122         nritems = btrfs_header_nritems(leaf);
1123         while (1) {
1124                 if (path->slots[0] >= nritems) {
1125                         ret = btrfs_next_leaf(root, path);
1126                         if (ret < 0)
1127                                 err = ret;
1128                         if (ret)
1129                                 goto fail;
1130
1131                         leaf = path->nodes[0];
1132                         nritems = btrfs_header_nritems(leaf);
1133                         recow = 1;
1134                 }
1135
1136                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1137                 if (key.objectid != bytenr ||
1138                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1139                         goto fail;
1140
1141                 ref = btrfs_item_ptr(leaf, path->slots[0],
1142                                      struct btrfs_extent_data_ref);
1143
1144                 if (match_extent_data_ref(leaf, ref, root_objectid,
1145                                           owner, offset)) {
1146                         if (recow) {
1147                                 btrfs_release_path(path);
1148                                 goto again;
1149                         }
1150                         err = 0;
1151                         break;
1152                 }
1153                 path->slots[0]++;
1154         }
1155 fail:
1156         return err;
1157 }
1158
1159 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1160                                            struct btrfs_root *root,
1161                                            struct btrfs_path *path,
1162                                            u64 bytenr, u64 parent,
1163                                            u64 root_objectid, u64 owner,
1164                                            u64 offset, int refs_to_add)
1165 {
1166         struct btrfs_key key;
1167         struct extent_buffer *leaf;
1168         u32 size;
1169         u32 num_refs;
1170         int ret;
1171
1172         key.objectid = bytenr;
1173         if (parent) {
1174                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1175                 key.offset = parent;
1176                 size = sizeof(struct btrfs_shared_data_ref);
1177         } else {
1178                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1179                 key.offset = hash_extent_data_ref(root_objectid,
1180                                                   owner, offset);
1181                 size = sizeof(struct btrfs_extent_data_ref);
1182         }
1183
1184         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1185         if (ret && ret != -EEXIST)
1186                 goto fail;
1187
1188         leaf = path->nodes[0];
1189         if (parent) {
1190                 struct btrfs_shared_data_ref *ref;
1191                 ref = btrfs_item_ptr(leaf, path->slots[0],
1192                                      struct btrfs_shared_data_ref);
1193                 if (ret == 0) {
1194                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1195                 } else {
1196                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1197                         num_refs += refs_to_add;
1198                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1199                 }
1200         } else {
1201                 struct btrfs_extent_data_ref *ref;
1202                 while (ret == -EEXIST) {
1203                         ref = btrfs_item_ptr(leaf, path->slots[0],
1204                                              struct btrfs_extent_data_ref);
1205                         if (match_extent_data_ref(leaf, ref, root_objectid,
1206                                                   owner, offset))
1207                                 break;
1208                         btrfs_release_path(path);
1209                         key.offset++;
1210                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1211                                                       size);
1212                         if (ret && ret != -EEXIST)
1213                                 goto fail;
1214
1215                         leaf = path->nodes[0];
1216                 }
1217                 ref = btrfs_item_ptr(leaf, path->slots[0],
1218                                      struct btrfs_extent_data_ref);
1219                 if (ret == 0) {
1220                         btrfs_set_extent_data_ref_root(leaf, ref,
1221                                                        root_objectid);
1222                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1223                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1224                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1225                 } else {
1226                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1227                         num_refs += refs_to_add;
1228                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1229                 }
1230         }
1231         btrfs_mark_buffer_dirty(leaf);
1232         ret = 0;
1233 fail:
1234         btrfs_release_path(path);
1235         return ret;
1236 }
1237
1238 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1239                                            struct btrfs_root *root,
1240                                            struct btrfs_path *path,
1241                                            int refs_to_drop)
1242 {
1243         struct btrfs_key key;
1244         struct btrfs_extent_data_ref *ref1 = NULL;
1245         struct btrfs_shared_data_ref *ref2 = NULL;
1246         struct extent_buffer *leaf;
1247         u32 num_refs = 0;
1248         int ret = 0;
1249
1250         leaf = path->nodes[0];
1251         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1252
1253         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1254                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1255                                       struct btrfs_extent_data_ref);
1256                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1257         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1258                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1259                                       struct btrfs_shared_data_ref);
1260                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1261 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1262         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1263                 struct btrfs_extent_ref_v0 *ref0;
1264                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1265                                       struct btrfs_extent_ref_v0);
1266                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1267 #endif
1268         } else {
1269                 BUG();
1270         }
1271
1272         BUG_ON(num_refs < refs_to_drop);
1273         num_refs -= refs_to_drop;
1274
1275         if (num_refs == 0) {
1276                 ret = btrfs_del_item(trans, root, path);
1277         } else {
1278                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1279                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1280                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1281                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1282 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1283                 else {
1284                         struct btrfs_extent_ref_v0 *ref0;
1285                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1286                                         struct btrfs_extent_ref_v0);
1287                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1288                 }
1289 #endif
1290                 btrfs_mark_buffer_dirty(leaf);
1291         }
1292         return ret;
1293 }
1294
1295 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1296                                           struct btrfs_path *path,
1297                                           struct btrfs_extent_inline_ref *iref)
1298 {
1299         struct btrfs_key key;
1300         struct extent_buffer *leaf;
1301         struct btrfs_extent_data_ref *ref1;
1302         struct btrfs_shared_data_ref *ref2;
1303         u32 num_refs = 0;
1304
1305         leaf = path->nodes[0];
1306         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1307         if (iref) {
1308                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1309                     BTRFS_EXTENT_DATA_REF_KEY) {
1310                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1311                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1312                 } else {
1313                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1314                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1315                 }
1316         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1317                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1318                                       struct btrfs_extent_data_ref);
1319                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1320         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1321                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1322                                       struct btrfs_shared_data_ref);
1323                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1324 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1325         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1326                 struct btrfs_extent_ref_v0 *ref0;
1327                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1328                                       struct btrfs_extent_ref_v0);
1329                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1330 #endif
1331         } else {
1332                 WARN_ON(1);
1333         }
1334         return num_refs;
1335 }
1336
1337 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1338                                           struct btrfs_root *root,
1339                                           struct btrfs_path *path,
1340                                           u64 bytenr, u64 parent,
1341                                           u64 root_objectid)
1342 {
1343         struct btrfs_key key;
1344         int ret;
1345
1346         key.objectid = bytenr;
1347         if (parent) {
1348                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1349                 key.offset = parent;
1350         } else {
1351                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1352                 key.offset = root_objectid;
1353         }
1354
1355         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1356         if (ret > 0)
1357                 ret = -ENOENT;
1358 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1359         if (ret == -ENOENT && parent) {
1360                 btrfs_release_path(path);
1361                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1362                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1363                 if (ret > 0)
1364                         ret = -ENOENT;
1365         }
1366 #endif
1367         return ret;
1368 }
1369
1370 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1371                                           struct btrfs_root *root,
1372                                           struct btrfs_path *path,
1373                                           u64 bytenr, u64 parent,
1374                                           u64 root_objectid)
1375 {
1376         struct btrfs_key key;
1377         int ret;
1378
1379         key.objectid = bytenr;
1380         if (parent) {
1381                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1382                 key.offset = parent;
1383         } else {
1384                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1385                 key.offset = root_objectid;
1386         }
1387
1388         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1389         btrfs_release_path(path);
1390         return ret;
1391 }
1392
1393 static inline int extent_ref_type(u64 parent, u64 owner)
1394 {
1395         int type;
1396         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1397                 if (parent > 0)
1398                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1399                 else
1400                         type = BTRFS_TREE_BLOCK_REF_KEY;
1401         } else {
1402                 if (parent > 0)
1403                         type = BTRFS_SHARED_DATA_REF_KEY;
1404                 else
1405                         type = BTRFS_EXTENT_DATA_REF_KEY;
1406         }
1407         return type;
1408 }
1409
1410 static int find_next_key(struct btrfs_path *path, int level,
1411                          struct btrfs_key *key)
1412
1413 {
1414         for (; level < BTRFS_MAX_LEVEL; level++) {
1415                 if (!path->nodes[level])
1416                         break;
1417                 if (path->slots[level] + 1 >=
1418                     btrfs_header_nritems(path->nodes[level]))
1419                         continue;
1420                 if (level == 0)
1421                         btrfs_item_key_to_cpu(path->nodes[level], key,
1422                                               path->slots[level] + 1);
1423                 else
1424                         btrfs_node_key_to_cpu(path->nodes[level], key,
1425                                               path->slots[level] + 1);
1426                 return 0;
1427         }
1428         return 1;
1429 }
1430
1431 /*
1432  * look for inline back ref. if back ref is found, *ref_ret is set
1433  * to the address of inline back ref, and 0 is returned.
1434  *
1435  * if back ref isn't found, *ref_ret is set to the address where it
1436  * should be inserted, and -ENOENT is returned.
1437  *
1438  * if insert is true and there are too many inline back refs, the path
1439  * points to the extent item, and -EAGAIN is returned.
1440  *
1441  * NOTE: inline back refs are ordered in the same way that back ref
1442  *       items in the tree are ordered.
1443  */
1444 static noinline_for_stack
1445 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1446                                  struct btrfs_root *root,
1447                                  struct btrfs_path *path,
1448                                  struct btrfs_extent_inline_ref **ref_ret,
1449                                  u64 bytenr, u64 num_bytes,
1450                                  u64 parent, u64 root_objectid,
1451                                  u64 owner, u64 offset, int insert)
1452 {
1453         struct btrfs_key key;
1454         struct extent_buffer *leaf;
1455         struct btrfs_extent_item *ei;
1456         struct btrfs_extent_inline_ref *iref;
1457         u64 flags;
1458         u64 item_size;
1459         unsigned long ptr;
1460         unsigned long end;
1461         int extra_size;
1462         int type;
1463         int want;
1464         int ret;
1465         int err = 0;
1466
1467         key.objectid = bytenr;
1468         key.type = BTRFS_EXTENT_ITEM_KEY;
1469         key.offset = num_bytes;
1470
1471         want = extent_ref_type(parent, owner);
1472         if (insert) {
1473                 extra_size = btrfs_extent_inline_ref_size(want);
1474                 path->keep_locks = 1;
1475         } else
1476                 extra_size = -1;
1477         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1478         if (ret < 0) {
1479                 err = ret;
1480                 goto out;
1481         }
1482         if (ret && !insert) {
1483                 err = -ENOENT;
1484                 goto out;
1485         }
1486         BUG_ON(ret); /* Corruption */
1487
1488         leaf = path->nodes[0];
1489         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1490 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1491         if (item_size < sizeof(*ei)) {
1492                 if (!insert) {
1493                         err = -ENOENT;
1494                         goto out;
1495                 }
1496                 ret = convert_extent_item_v0(trans, root, path, owner,
1497                                              extra_size);
1498                 if (ret < 0) {
1499                         err = ret;
1500                         goto out;
1501                 }
1502                 leaf = path->nodes[0];
1503                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1504         }
1505 #endif
1506         BUG_ON(item_size < sizeof(*ei));
1507
1508         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1509         flags = btrfs_extent_flags(leaf, ei);
1510
1511         ptr = (unsigned long)(ei + 1);
1512         end = (unsigned long)ei + item_size;
1513
1514         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1515                 ptr += sizeof(struct btrfs_tree_block_info);
1516                 BUG_ON(ptr > end);
1517         } else {
1518                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1519         }
1520
1521         err = -ENOENT;
1522         while (1) {
1523                 if (ptr >= end) {
1524                         WARN_ON(ptr > end);
1525                         break;
1526                 }
1527                 iref = (struct btrfs_extent_inline_ref *)ptr;
1528                 type = btrfs_extent_inline_ref_type(leaf, iref);
1529                 if (want < type)
1530                         break;
1531                 if (want > type) {
1532                         ptr += btrfs_extent_inline_ref_size(type);
1533                         continue;
1534                 }
1535
1536                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1537                         struct btrfs_extent_data_ref *dref;
1538                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1539                         if (match_extent_data_ref(leaf, dref, root_objectid,
1540                                                   owner, offset)) {
1541                                 err = 0;
1542                                 break;
1543                         }
1544                         if (hash_extent_data_ref_item(leaf, dref) <
1545                             hash_extent_data_ref(root_objectid, owner, offset))
1546                                 break;
1547                 } else {
1548                         u64 ref_offset;
1549                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1550                         if (parent > 0) {
1551                                 if (parent == ref_offset) {
1552                                         err = 0;
1553                                         break;
1554                                 }
1555                                 if (ref_offset < parent)
1556                                         break;
1557                         } else {
1558                                 if (root_objectid == ref_offset) {
1559                                         err = 0;
1560                                         break;
1561                                 }
1562                                 if (ref_offset < root_objectid)
1563                                         break;
1564                         }
1565                 }
1566                 ptr += btrfs_extent_inline_ref_size(type);
1567         }
1568         if (err == -ENOENT && insert) {
1569                 if (item_size + extra_size >=
1570                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1571                         err = -EAGAIN;
1572                         goto out;
1573                 }
1574                 /*
1575                  * To add new inline back ref, we have to make sure
1576                  * there is no corresponding back ref item.
1577                  * For simplicity, we just do not add new inline back
1578                  * ref if there is any kind of item for this block
1579                  */
1580                 if (find_next_key(path, 0, &key) == 0 &&
1581                     key.objectid == bytenr &&
1582                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1583                         err = -EAGAIN;
1584                         goto out;
1585                 }
1586         }
1587         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1588 out:
1589         if (insert) {
1590                 path->keep_locks = 0;
1591                 btrfs_unlock_up_safe(path, 1);
1592         }
1593         return err;
1594 }
1595
1596 /*
1597  * helper to add new inline back ref
1598  */
1599 static noinline_for_stack
1600 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1601                                  struct btrfs_root *root,
1602                                  struct btrfs_path *path,
1603                                  struct btrfs_extent_inline_ref *iref,
1604                                  u64 parent, u64 root_objectid,
1605                                  u64 owner, u64 offset, int refs_to_add,
1606                                  struct btrfs_delayed_extent_op *extent_op)
1607 {
1608         struct extent_buffer *leaf;
1609         struct btrfs_extent_item *ei;
1610         unsigned long ptr;
1611         unsigned long end;
1612         unsigned long item_offset;
1613         u64 refs;
1614         int size;
1615         int type;
1616
1617         leaf = path->nodes[0];
1618         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1619         item_offset = (unsigned long)iref - (unsigned long)ei;
1620
1621         type = extent_ref_type(parent, owner);
1622         size = btrfs_extent_inline_ref_size(type);
1623
1624         btrfs_extend_item(trans, root, path, size);
1625
1626         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1627         refs = btrfs_extent_refs(leaf, ei);
1628         refs += refs_to_add;
1629         btrfs_set_extent_refs(leaf, ei, refs);
1630         if (extent_op)
1631                 __run_delayed_extent_op(extent_op, leaf, ei);
1632
1633         ptr = (unsigned long)ei + item_offset;
1634         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1635         if (ptr < end - size)
1636                 memmove_extent_buffer(leaf, ptr + size, ptr,
1637                                       end - size - ptr);
1638
1639         iref = (struct btrfs_extent_inline_ref *)ptr;
1640         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1641         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1642                 struct btrfs_extent_data_ref *dref;
1643                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1644                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1645                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1646                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1647                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1648         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1649                 struct btrfs_shared_data_ref *sref;
1650                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1651                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1652                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1653         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1654                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1655         } else {
1656                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1657         }
1658         btrfs_mark_buffer_dirty(leaf);
1659 }
1660
1661 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1662                                  struct btrfs_root *root,
1663                                  struct btrfs_path *path,
1664                                  struct btrfs_extent_inline_ref **ref_ret,
1665                                  u64 bytenr, u64 num_bytes, u64 parent,
1666                                  u64 root_objectid, u64 owner, u64 offset)
1667 {
1668         int ret;
1669
1670         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1671                                            bytenr, num_bytes, parent,
1672                                            root_objectid, owner, offset, 0);
1673         if (ret != -ENOENT)
1674                 return ret;
1675
1676         btrfs_release_path(path);
1677         *ref_ret = NULL;
1678
1679         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1680                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1681                                             root_objectid);
1682         } else {
1683                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1684                                              root_objectid, owner, offset);
1685         }
1686         return ret;
1687 }
1688
1689 /*
1690  * helper to update/remove inline back ref
1691  */
1692 static noinline_for_stack
1693 void update_inline_extent_backref(struct btrfs_trans_handle *trans,
1694                                   struct btrfs_root *root,
1695                                   struct btrfs_path *path,
1696                                   struct btrfs_extent_inline_ref *iref,
1697                                   int refs_to_mod,
1698                                   struct btrfs_delayed_extent_op *extent_op)
1699 {
1700         struct extent_buffer *leaf;
1701         struct btrfs_extent_item *ei;
1702         struct btrfs_extent_data_ref *dref = NULL;
1703         struct btrfs_shared_data_ref *sref = NULL;
1704         unsigned long ptr;
1705         unsigned long end;
1706         u32 item_size;
1707         int size;
1708         int type;
1709         u64 refs;
1710
1711         leaf = path->nodes[0];
1712         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1713         refs = btrfs_extent_refs(leaf, ei);
1714         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1715         refs += refs_to_mod;
1716         btrfs_set_extent_refs(leaf, ei, refs);
1717         if (extent_op)
1718                 __run_delayed_extent_op(extent_op, leaf, ei);
1719
1720         type = btrfs_extent_inline_ref_type(leaf, iref);
1721
1722         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1723                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1724                 refs = btrfs_extent_data_ref_count(leaf, dref);
1725         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1726                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1727                 refs = btrfs_shared_data_ref_count(leaf, sref);
1728         } else {
1729                 refs = 1;
1730                 BUG_ON(refs_to_mod != -1);
1731         }
1732
1733         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1734         refs += refs_to_mod;
1735
1736         if (refs > 0) {
1737                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1738                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1739                 else
1740                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1741         } else {
1742                 size =  btrfs_extent_inline_ref_size(type);
1743                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1744                 ptr = (unsigned long)iref;
1745                 end = (unsigned long)ei + item_size;
1746                 if (ptr + size < end)
1747                         memmove_extent_buffer(leaf, ptr, ptr + size,
1748                                               end - ptr - size);
1749                 item_size -= size;
1750                 btrfs_truncate_item(trans, root, path, item_size, 1);
1751         }
1752         btrfs_mark_buffer_dirty(leaf);
1753 }
1754
1755 static noinline_for_stack
1756 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1757                                  struct btrfs_root *root,
1758                                  struct btrfs_path *path,
1759                                  u64 bytenr, u64 num_bytes, u64 parent,
1760                                  u64 root_objectid, u64 owner,
1761                                  u64 offset, int refs_to_add,
1762                                  struct btrfs_delayed_extent_op *extent_op)
1763 {
1764         struct btrfs_extent_inline_ref *iref;
1765         int ret;
1766
1767         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1768                                            bytenr, num_bytes, parent,
1769                                            root_objectid, owner, offset, 1);
1770         if (ret == 0) {
1771                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1772                 update_inline_extent_backref(trans, root, path, iref,
1773                                              refs_to_add, extent_op);
1774         } else if (ret == -ENOENT) {
1775                 setup_inline_extent_backref(trans, root, path, iref, parent,
1776                                             root_objectid, owner, offset,
1777                                             refs_to_add, extent_op);
1778                 ret = 0;
1779         }
1780         return ret;
1781 }
1782
1783 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1784                                  struct btrfs_root *root,
1785                                  struct btrfs_path *path,
1786                                  u64 bytenr, u64 parent, u64 root_objectid,
1787                                  u64 owner, u64 offset, int refs_to_add)
1788 {
1789         int ret;
1790         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1791                 BUG_ON(refs_to_add != 1);
1792                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1793                                             parent, root_objectid);
1794         } else {
1795                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1796                                              parent, root_objectid,
1797                                              owner, offset, refs_to_add);
1798         }
1799         return ret;
1800 }
1801
1802 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1803                                  struct btrfs_root *root,
1804                                  struct btrfs_path *path,
1805                                  struct btrfs_extent_inline_ref *iref,
1806                                  int refs_to_drop, int is_data)
1807 {
1808         int ret = 0;
1809
1810         BUG_ON(!is_data && refs_to_drop != 1);
1811         if (iref) {
1812                 update_inline_extent_backref(trans, root, path, iref,
1813                                              -refs_to_drop, NULL);
1814         } else if (is_data) {
1815                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1816         } else {
1817                 ret = btrfs_del_item(trans, root, path);
1818         }
1819         return ret;
1820 }
1821
1822 static int btrfs_issue_discard(struct block_device *bdev,
1823                                 u64 start, u64 len)
1824 {
1825         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1826 }
1827
1828 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1829                                 u64 num_bytes, u64 *actual_bytes)
1830 {
1831         int ret;
1832         u64 discarded_bytes = 0;
1833         struct btrfs_bio *bbio = NULL;
1834
1835
1836         /* Tell the block device(s) that the sectors can be discarded */
1837         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1838                               bytenr, &num_bytes, &bbio, 0);
1839         /* Error condition is -ENOMEM */
1840         if (!ret) {
1841                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1842                 int i;
1843
1844
1845                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1846                         if (!stripe->dev->can_discard)
1847                                 continue;
1848
1849                         ret = btrfs_issue_discard(stripe->dev->bdev,
1850                                                   stripe->physical,
1851                                                   stripe->length);
1852                         if (!ret)
1853                                 discarded_bytes += stripe->length;
1854                         else if (ret != -EOPNOTSUPP)
1855                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1856
1857                         /*
1858                          * Just in case we get back EOPNOTSUPP for some reason,
1859                          * just ignore the return value so we don't screw up
1860                          * people calling discard_extent.
1861                          */
1862                         ret = 0;
1863                 }
1864                 kfree(bbio);
1865         }
1866
1867         if (actual_bytes)
1868                 *actual_bytes = discarded_bytes;
1869
1870
1871         return ret;
1872 }
1873
1874 /* Can return -ENOMEM */
1875 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1876                          struct btrfs_root *root,
1877                          u64 bytenr, u64 num_bytes, u64 parent,
1878                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1879 {
1880         int ret;
1881         struct btrfs_fs_info *fs_info = root->fs_info;
1882
1883         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1884                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1885
1886         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1887                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1888                                         num_bytes,
1889                                         parent, root_objectid, (int)owner,
1890                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1891         } else {
1892                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1893                                         num_bytes,
1894                                         parent, root_objectid, owner, offset,
1895                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1896         }
1897         return ret;
1898 }
1899
1900 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1901                                   struct btrfs_root *root,
1902                                   u64 bytenr, u64 num_bytes,
1903                                   u64 parent, u64 root_objectid,
1904                                   u64 owner, u64 offset, int refs_to_add,
1905                                   struct btrfs_delayed_extent_op *extent_op)
1906 {
1907         struct btrfs_path *path;
1908         struct extent_buffer *leaf;
1909         struct btrfs_extent_item *item;
1910         u64 refs;
1911         int ret;
1912         int err = 0;
1913
1914         path = btrfs_alloc_path();
1915         if (!path)
1916                 return -ENOMEM;
1917
1918         path->reada = 1;
1919         path->leave_spinning = 1;
1920         /* this will setup the path even if it fails to insert the back ref */
1921         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1922                                            path, bytenr, num_bytes, parent,
1923                                            root_objectid, owner, offset,
1924                                            refs_to_add, extent_op);
1925         if (ret == 0)
1926                 goto out;
1927
1928         if (ret != -EAGAIN) {
1929                 err = ret;
1930                 goto out;
1931         }
1932
1933         leaf = path->nodes[0];
1934         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1935         refs = btrfs_extent_refs(leaf, item);
1936         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1937         if (extent_op)
1938                 __run_delayed_extent_op(extent_op, leaf, item);
1939
1940         btrfs_mark_buffer_dirty(leaf);
1941         btrfs_release_path(path);
1942
1943         path->reada = 1;
1944         path->leave_spinning = 1;
1945
1946         /* now insert the actual backref */
1947         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1948                                     path, bytenr, parent, root_objectid,
1949                                     owner, offset, refs_to_add);
1950         if (ret)
1951                 btrfs_abort_transaction(trans, root, ret);
1952 out:
1953         btrfs_free_path(path);
1954         return err;
1955 }
1956
1957 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1958                                 struct btrfs_root *root,
1959                                 struct btrfs_delayed_ref_node *node,
1960                                 struct btrfs_delayed_extent_op *extent_op,
1961                                 int insert_reserved)
1962 {
1963         int ret = 0;
1964         struct btrfs_delayed_data_ref *ref;
1965         struct btrfs_key ins;
1966         u64 parent = 0;
1967         u64 ref_root = 0;
1968         u64 flags = 0;
1969
1970         ins.objectid = node->bytenr;
1971         ins.offset = node->num_bytes;
1972         ins.type = BTRFS_EXTENT_ITEM_KEY;
1973
1974         ref = btrfs_delayed_node_to_data_ref(node);
1975         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1976                 parent = ref->parent;
1977         else
1978                 ref_root = ref->root;
1979
1980         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1981                 if (extent_op) {
1982                         BUG_ON(extent_op->update_key);
1983                         flags |= extent_op->flags_to_set;
1984                 }
1985                 ret = alloc_reserved_file_extent(trans, root,
1986                                                  parent, ref_root, flags,
1987                                                  ref->objectid, ref->offset,
1988                                                  &ins, node->ref_mod);
1989         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1990                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1991                                              node->num_bytes, parent,
1992                                              ref_root, ref->objectid,
1993                                              ref->offset, node->ref_mod,
1994                                              extent_op);
1995         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1996                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1997                                           node->num_bytes, parent,
1998                                           ref_root, ref->objectid,
1999                                           ref->offset, node->ref_mod,
2000                                           extent_op);
2001         } else {
2002                 BUG();
2003         }
2004         return ret;
2005 }
2006
2007 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2008                                     struct extent_buffer *leaf,
2009                                     struct btrfs_extent_item *ei)
2010 {
2011         u64 flags = btrfs_extent_flags(leaf, ei);
2012         if (extent_op->update_flags) {
2013                 flags |= extent_op->flags_to_set;
2014                 btrfs_set_extent_flags(leaf, ei, flags);
2015         }
2016
2017         if (extent_op->update_key) {
2018                 struct btrfs_tree_block_info *bi;
2019                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2020                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2021                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2022         }
2023 }
2024
2025 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2026                                  struct btrfs_root *root,
2027                                  struct btrfs_delayed_ref_node *node,
2028                                  struct btrfs_delayed_extent_op *extent_op)
2029 {
2030         struct btrfs_key key;
2031         struct btrfs_path *path;
2032         struct btrfs_extent_item *ei;
2033         struct extent_buffer *leaf;
2034         u32 item_size;
2035         int ret;
2036         int err = 0;
2037
2038         if (trans->aborted)
2039                 return 0;
2040
2041         path = btrfs_alloc_path();
2042         if (!path)
2043                 return -ENOMEM;
2044
2045         key.objectid = node->bytenr;
2046         key.type = BTRFS_EXTENT_ITEM_KEY;
2047         key.offset = node->num_bytes;
2048
2049         path->reada = 1;
2050         path->leave_spinning = 1;
2051         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2052                                 path, 0, 1);
2053         if (ret < 0) {
2054                 err = ret;
2055                 goto out;
2056         }
2057         if (ret > 0) {
2058                 err = -EIO;
2059                 goto out;
2060         }
2061
2062         leaf = path->nodes[0];
2063         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2064 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2065         if (item_size < sizeof(*ei)) {
2066                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2067                                              path, (u64)-1, 0);
2068                 if (ret < 0) {
2069                         err = ret;
2070                         goto out;
2071                 }
2072                 leaf = path->nodes[0];
2073                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2074         }
2075 #endif
2076         BUG_ON(item_size < sizeof(*ei));
2077         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2078         __run_delayed_extent_op(extent_op, leaf, ei);
2079
2080         btrfs_mark_buffer_dirty(leaf);
2081 out:
2082         btrfs_free_path(path);
2083         return err;
2084 }
2085
2086 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2087                                 struct btrfs_root *root,
2088                                 struct btrfs_delayed_ref_node *node,
2089                                 struct btrfs_delayed_extent_op *extent_op,
2090                                 int insert_reserved)
2091 {
2092         int ret = 0;
2093         struct btrfs_delayed_tree_ref *ref;
2094         struct btrfs_key ins;
2095         u64 parent = 0;
2096         u64 ref_root = 0;
2097
2098         ins.objectid = node->bytenr;
2099         ins.offset = node->num_bytes;
2100         ins.type = BTRFS_EXTENT_ITEM_KEY;
2101
2102         ref = btrfs_delayed_node_to_tree_ref(node);
2103         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2104                 parent = ref->parent;
2105         else
2106                 ref_root = ref->root;
2107
2108         BUG_ON(node->ref_mod != 1);
2109         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2110                 BUG_ON(!extent_op || !extent_op->update_flags ||
2111                        !extent_op->update_key);
2112                 ret = alloc_reserved_tree_block(trans, root,
2113                                                 parent, ref_root,
2114                                                 extent_op->flags_to_set,
2115                                                 &extent_op->key,
2116                                                 ref->level, &ins);
2117         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2118                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2119                                              node->num_bytes, parent, ref_root,
2120                                              ref->level, 0, 1, extent_op);
2121         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2122                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2123                                           node->num_bytes, parent, ref_root,
2124                                           ref->level, 0, 1, extent_op);
2125         } else {
2126                 BUG();
2127         }
2128         return ret;
2129 }
2130
2131 /* helper function to actually process a single delayed ref entry */
2132 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2133                                struct btrfs_root *root,
2134                                struct btrfs_delayed_ref_node *node,
2135                                struct btrfs_delayed_extent_op *extent_op,
2136                                int insert_reserved)
2137 {
2138         int ret = 0;
2139
2140         if (trans->aborted)
2141                 return 0;
2142
2143         if (btrfs_delayed_ref_is_head(node)) {
2144                 struct btrfs_delayed_ref_head *head;
2145                 /*
2146                  * we've hit the end of the chain and we were supposed
2147                  * to insert this extent into the tree.  But, it got
2148                  * deleted before we ever needed to insert it, so all
2149                  * we have to do is clean up the accounting
2150                  */
2151                 BUG_ON(extent_op);
2152                 head = btrfs_delayed_node_to_head(node);
2153                 if (insert_reserved) {
2154                         btrfs_pin_extent(root, node->bytenr,
2155                                          node->num_bytes, 1);
2156                         if (head->is_data) {
2157                                 ret = btrfs_del_csums(trans, root,
2158                                                       node->bytenr,
2159                                                       node->num_bytes);
2160                         }
2161                 }
2162                 mutex_unlock(&head->mutex);
2163                 return ret;
2164         }
2165
2166         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2167             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2168                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2169                                            insert_reserved);
2170         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2171                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2172                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2173                                            insert_reserved);
2174         else
2175                 BUG();
2176         return ret;
2177 }
2178
2179 static noinline struct btrfs_delayed_ref_node *
2180 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2181 {
2182         struct rb_node *node;
2183         struct btrfs_delayed_ref_node *ref;
2184         int action = BTRFS_ADD_DELAYED_REF;
2185 again:
2186         /*
2187          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2188          * this prevents ref count from going down to zero when
2189          * there still are pending delayed ref.
2190          */
2191         node = rb_prev(&head->node.rb_node);
2192         while (1) {
2193                 if (!node)
2194                         break;
2195                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2196                                 rb_node);
2197                 if (ref->bytenr != head->node.bytenr)
2198                         break;
2199                 if (ref->action == action)
2200                         return ref;
2201                 node = rb_prev(node);
2202         }
2203         if (action == BTRFS_ADD_DELAYED_REF) {
2204                 action = BTRFS_DROP_DELAYED_REF;
2205                 goto again;
2206         }
2207         return NULL;
2208 }
2209
2210 /*
2211  * Returns 0 on success or if called with an already aborted transaction.
2212  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2213  */
2214 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2215                                        struct btrfs_root *root,
2216                                        struct list_head *cluster)
2217 {
2218         struct btrfs_delayed_ref_root *delayed_refs;
2219         struct btrfs_delayed_ref_node *ref;
2220         struct btrfs_delayed_ref_head *locked_ref = NULL;
2221         struct btrfs_delayed_extent_op *extent_op;
2222         struct btrfs_fs_info *fs_info = root->fs_info;
2223         int ret;
2224         int count = 0;
2225         int must_insert_reserved = 0;
2226
2227         delayed_refs = &trans->transaction->delayed_refs;
2228         while (1) {
2229                 if (!locked_ref) {
2230                         /* pick a new head ref from the cluster list */
2231                         if (list_empty(cluster))
2232                                 break;
2233
2234                         locked_ref = list_entry(cluster->next,
2235                                      struct btrfs_delayed_ref_head, cluster);
2236
2237                         /* grab the lock that says we are going to process
2238                          * all the refs for this head */
2239                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2240
2241                         /*
2242                          * we may have dropped the spin lock to get the head
2243                          * mutex lock, and that might have given someone else
2244                          * time to free the head.  If that's true, it has been
2245                          * removed from our list and we can move on.
2246                          */
2247                         if (ret == -EAGAIN) {
2248                                 locked_ref = NULL;
2249                                 count++;
2250                                 continue;
2251                         }
2252                 }
2253
2254                 /*
2255                  * We need to try and merge add/drops of the same ref since we
2256                  * can run into issues with relocate dropping the implicit ref
2257                  * and then it being added back again before the drop can
2258                  * finish.  If we merged anything we need to re-loop so we can
2259                  * get a good ref.
2260                  */
2261                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2262                                          locked_ref);
2263
2264                 /*
2265                  * locked_ref is the head node, so we have to go one
2266                  * node back for any delayed ref updates
2267                  */
2268                 ref = select_delayed_ref(locked_ref);
2269
2270                 if (ref && ref->seq &&
2271                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2272                         /*
2273                          * there are still refs with lower seq numbers in the
2274                          * process of being added. Don't run this ref yet.
2275                          */
2276                         list_del_init(&locked_ref->cluster);
2277                         mutex_unlock(&locked_ref->mutex);
2278                         locked_ref = NULL;
2279                         delayed_refs->num_heads_ready++;
2280                         spin_unlock(&delayed_refs->lock);
2281                         cond_resched();
2282                         spin_lock(&delayed_refs->lock);
2283                         continue;
2284                 }
2285
2286                 /*
2287                  * record the must insert reserved flag before we
2288                  * drop the spin lock.
2289                  */
2290                 must_insert_reserved = locked_ref->must_insert_reserved;
2291                 locked_ref->must_insert_reserved = 0;
2292
2293                 extent_op = locked_ref->extent_op;
2294                 locked_ref->extent_op = NULL;
2295
2296                 if (!ref) {
2297                         /* All delayed refs have been processed, Go ahead
2298                          * and send the head node to run_one_delayed_ref,
2299                          * so that any accounting fixes can happen
2300                          */
2301                         ref = &locked_ref->node;
2302
2303                         if (extent_op && must_insert_reserved) {
2304                                 kfree(extent_op);
2305                                 extent_op = NULL;
2306                         }
2307
2308                         if (extent_op) {
2309                                 spin_unlock(&delayed_refs->lock);
2310
2311                                 ret = run_delayed_extent_op(trans, root,
2312                                                             ref, extent_op);
2313                                 kfree(extent_op);
2314
2315                                 if (ret) {
2316                                         printk(KERN_DEBUG "btrfs: run_delayed_extent_op returned %d\n", ret);
2317                                         spin_lock(&delayed_refs->lock);
2318                                         return ret;
2319                                 }
2320
2321                                 goto next;
2322                         }
2323
2324                         list_del_init(&locked_ref->cluster);
2325                         locked_ref = NULL;
2326                 }
2327
2328                 ref->in_tree = 0;
2329                 rb_erase(&ref->rb_node, &delayed_refs->root);
2330                 delayed_refs->num_entries--;
2331                 if (locked_ref) {
2332                         /*
2333                          * when we play the delayed ref, also correct the
2334                          * ref_mod on head
2335                          */
2336                         switch (ref->action) {
2337                         case BTRFS_ADD_DELAYED_REF:
2338                         case BTRFS_ADD_DELAYED_EXTENT:
2339                                 locked_ref->node.ref_mod -= ref->ref_mod;
2340                                 break;
2341                         case BTRFS_DROP_DELAYED_REF:
2342                                 locked_ref->node.ref_mod += ref->ref_mod;
2343                                 break;
2344                         default:
2345                                 WARN_ON(1);
2346                         }
2347                 }
2348                 spin_unlock(&delayed_refs->lock);
2349
2350                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2351                                           must_insert_reserved);
2352
2353                 btrfs_put_delayed_ref(ref);
2354                 kfree(extent_op);
2355                 count++;
2356
2357                 if (ret) {
2358                         printk(KERN_DEBUG "btrfs: run_one_delayed_ref returned %d\n", ret);
2359                         spin_lock(&delayed_refs->lock);
2360                         return ret;
2361                 }
2362
2363 next:
2364                 do_chunk_alloc(trans, fs_info->extent_root,
2365                                2 * 1024 * 1024,
2366                                btrfs_get_alloc_profile(root, 0),
2367                                CHUNK_ALLOC_NO_FORCE);
2368                 cond_resched();
2369                 spin_lock(&delayed_refs->lock);
2370         }
2371         return count;
2372 }
2373
2374 #ifdef SCRAMBLE_DELAYED_REFS
2375 /*
2376  * Normally delayed refs get processed in ascending bytenr order. This
2377  * correlates in most cases to the order added. To expose dependencies on this
2378  * order, we start to process the tree in the middle instead of the beginning
2379  */
2380 static u64 find_middle(struct rb_root *root)
2381 {
2382         struct rb_node *n = root->rb_node;
2383         struct btrfs_delayed_ref_node *entry;
2384         int alt = 1;
2385         u64 middle;
2386         u64 first = 0, last = 0;
2387
2388         n = rb_first(root);
2389         if (n) {
2390                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2391                 first = entry->bytenr;
2392         }
2393         n = rb_last(root);
2394         if (n) {
2395                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2396                 last = entry->bytenr;
2397         }
2398         n = root->rb_node;
2399
2400         while (n) {
2401                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2402                 WARN_ON(!entry->in_tree);
2403
2404                 middle = entry->bytenr;
2405
2406                 if (alt)
2407                         n = n->rb_left;
2408                 else
2409                         n = n->rb_right;
2410
2411                 alt = 1 - alt;
2412         }
2413         return middle;
2414 }
2415 #endif
2416
2417 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2418                                          struct btrfs_fs_info *fs_info)
2419 {
2420         struct qgroup_update *qgroup_update;
2421         int ret = 0;
2422
2423         if (list_empty(&trans->qgroup_ref_list) !=
2424             !trans->delayed_ref_elem.seq) {
2425                 /* list without seq or seq without list */
2426                 printk(KERN_ERR "btrfs: qgroup accounting update error, list is%s empty, seq is %llu\n",
2427                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2428                         trans->delayed_ref_elem.seq);
2429                 BUG();
2430         }
2431
2432         if (!trans->delayed_ref_elem.seq)
2433                 return 0;
2434
2435         while (!list_empty(&trans->qgroup_ref_list)) {
2436                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2437                                                  struct qgroup_update, list);
2438                 list_del(&qgroup_update->list);
2439                 if (!ret)
2440                         ret = btrfs_qgroup_account_ref(
2441                                         trans, fs_info, qgroup_update->node,
2442                                         qgroup_update->extent_op);
2443                 kfree(qgroup_update);
2444         }
2445
2446         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2447
2448         return ret;
2449 }
2450
2451 /*
2452  * this starts processing the delayed reference count updates and
2453  * extent insertions we have queued up so far.  count can be
2454  * 0, which means to process everything in the tree at the start
2455  * of the run (but not newly added entries), or it can be some target
2456  * number you'd like to process.
2457  *
2458  * Returns 0 on success or if called with an aborted transaction
2459  * Returns <0 on error and aborts the transaction
2460  */
2461 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2462                            struct btrfs_root *root, unsigned long count)
2463 {
2464         struct rb_node *node;
2465         struct btrfs_delayed_ref_root *delayed_refs;
2466         struct btrfs_delayed_ref_node *ref;
2467         struct list_head cluster;
2468         int ret;
2469         u64 delayed_start;
2470         int run_all = count == (unsigned long)-1;
2471         int run_most = 0;
2472         int loops;
2473
2474         /* We'll clean this up in btrfs_cleanup_transaction */
2475         if (trans->aborted)
2476                 return 0;
2477
2478         if (root == root->fs_info->extent_root)
2479                 root = root->fs_info->tree_root;
2480
2481         do_chunk_alloc(trans, root->fs_info->extent_root,
2482                        2 * 1024 * 1024, btrfs_get_alloc_profile(root, 0),
2483                        CHUNK_ALLOC_NO_FORCE);
2484
2485         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2486
2487         delayed_refs = &trans->transaction->delayed_refs;
2488         INIT_LIST_HEAD(&cluster);
2489 again:
2490         loops = 0;
2491         spin_lock(&delayed_refs->lock);
2492
2493 #ifdef SCRAMBLE_DELAYED_REFS
2494         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2495 #endif
2496
2497         if (count == 0) {
2498                 count = delayed_refs->num_entries * 2;
2499                 run_most = 1;
2500         }
2501         while (1) {
2502                 if (!(run_all || run_most) &&
2503                     delayed_refs->num_heads_ready < 64)
2504                         break;
2505
2506                 /*
2507                  * go find something we can process in the rbtree.  We start at
2508                  * the beginning of the tree, and then build a cluster
2509                  * of refs to process starting at the first one we are able to
2510                  * lock
2511                  */
2512                 delayed_start = delayed_refs->run_delayed_start;
2513                 ret = btrfs_find_ref_cluster(trans, &cluster,
2514                                              delayed_refs->run_delayed_start);
2515                 if (ret)
2516                         break;
2517
2518                 ret = run_clustered_refs(trans, root, &cluster);
2519                 if (ret < 0) {
2520                         spin_unlock(&delayed_refs->lock);
2521                         btrfs_abort_transaction(trans, root, ret);
2522                         return ret;
2523                 }
2524
2525                 count -= min_t(unsigned long, ret, count);
2526
2527                 if (count == 0)
2528                         break;
2529
2530                 if (delayed_start >= delayed_refs->run_delayed_start) {
2531                         if (loops == 0) {
2532                                 /*
2533                                  * btrfs_find_ref_cluster looped. let's do one
2534                                  * more cycle. if we don't run any delayed ref
2535                                  * during that cycle (because we can't because
2536                                  * all of them are blocked), bail out.
2537                                  */
2538                                 loops = 1;
2539                         } else {
2540                                 /*
2541                                  * no runnable refs left, stop trying
2542                                  */
2543                                 BUG_ON(run_all);
2544                                 break;
2545                         }
2546                 }
2547                 if (ret) {
2548                         /* refs were run, let's reset staleness detection */
2549                         loops = 0;
2550                 }
2551         }
2552
2553         if (run_all) {
2554                 node = rb_first(&delayed_refs->root);
2555                 if (!node)
2556                         goto out;
2557                 count = (unsigned long)-1;
2558
2559                 while (node) {
2560                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2561                                        rb_node);
2562                         if (btrfs_delayed_ref_is_head(ref)) {
2563                                 struct btrfs_delayed_ref_head *head;
2564
2565                                 head = btrfs_delayed_node_to_head(ref);
2566                                 atomic_inc(&ref->refs);
2567
2568                                 spin_unlock(&delayed_refs->lock);
2569                                 /*
2570                                  * Mutex was contended, block until it's
2571                                  * released and try again
2572                                  */
2573                                 mutex_lock(&head->mutex);
2574                                 mutex_unlock(&head->mutex);
2575
2576                                 btrfs_put_delayed_ref(ref);
2577                                 cond_resched();
2578                                 goto again;
2579                         }
2580                         node = rb_next(node);
2581                 }
2582                 spin_unlock(&delayed_refs->lock);
2583                 schedule_timeout(1);
2584                 goto again;
2585         }
2586 out:
2587         spin_unlock(&delayed_refs->lock);
2588         assert_qgroups_uptodate(trans);
2589         return 0;
2590 }
2591
2592 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2593                                 struct btrfs_root *root,
2594                                 u64 bytenr, u64 num_bytes, u64 flags,
2595                                 int is_data)
2596 {
2597         struct btrfs_delayed_extent_op *extent_op;
2598         int ret;
2599
2600         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2601         if (!extent_op)
2602                 return -ENOMEM;
2603
2604         extent_op->flags_to_set = flags;
2605         extent_op->update_flags = 1;
2606         extent_op->update_key = 0;
2607         extent_op->is_data = is_data ? 1 : 0;
2608
2609         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2610                                           num_bytes, extent_op);
2611         if (ret)
2612                 kfree(extent_op);
2613         return ret;
2614 }
2615
2616 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2617                                       struct btrfs_root *root,
2618                                       struct btrfs_path *path,
2619                                       u64 objectid, u64 offset, u64 bytenr)
2620 {
2621         struct btrfs_delayed_ref_head *head;
2622         struct btrfs_delayed_ref_node *ref;
2623         struct btrfs_delayed_data_ref *data_ref;
2624         struct btrfs_delayed_ref_root *delayed_refs;
2625         struct rb_node *node;
2626         int ret = 0;
2627
2628         ret = -ENOENT;
2629         delayed_refs = &trans->transaction->delayed_refs;
2630         spin_lock(&delayed_refs->lock);
2631         head = btrfs_find_delayed_ref_head(trans, bytenr);
2632         if (!head)
2633                 goto out;
2634
2635         if (!mutex_trylock(&head->mutex)) {
2636                 atomic_inc(&head->node.refs);
2637                 spin_unlock(&delayed_refs->lock);
2638
2639                 btrfs_release_path(path);
2640
2641                 /*
2642                  * Mutex was contended, block until it's released and let
2643                  * caller try again
2644                  */
2645                 mutex_lock(&head->mutex);
2646                 mutex_unlock(&head->mutex);
2647                 btrfs_put_delayed_ref(&head->node);
2648                 return -EAGAIN;
2649         }
2650
2651         node = rb_prev(&head->node.rb_node);
2652         if (!node)
2653                 goto out_unlock;
2654
2655         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2656
2657         if (ref->bytenr != bytenr)
2658                 goto out_unlock;
2659
2660         ret = 1;
2661         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2662                 goto out_unlock;
2663
2664         data_ref = btrfs_delayed_node_to_data_ref(ref);
2665
2666         node = rb_prev(node);
2667         if (node) {
2668                 int seq = ref->seq;
2669
2670                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2671                 if (ref->bytenr == bytenr && ref->seq == seq)
2672                         goto out_unlock;
2673         }
2674
2675         if (data_ref->root != root->root_key.objectid ||
2676             data_ref->objectid != objectid || data_ref->offset != offset)
2677                 goto out_unlock;
2678
2679         ret = 0;
2680 out_unlock:
2681         mutex_unlock(&head->mutex);
2682 out:
2683         spin_unlock(&delayed_refs->lock);
2684         return ret;
2685 }
2686
2687 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2688                                         struct btrfs_root *root,
2689                                         struct btrfs_path *path,
2690                                         u64 objectid, u64 offset, u64 bytenr)
2691 {
2692         struct btrfs_root *extent_root = root->fs_info->extent_root;
2693         struct extent_buffer *leaf;
2694         struct btrfs_extent_data_ref *ref;
2695         struct btrfs_extent_inline_ref *iref;
2696         struct btrfs_extent_item *ei;
2697         struct btrfs_key key;
2698         u32 item_size;
2699         int ret;
2700
2701         key.objectid = bytenr;
2702         key.offset = (u64)-1;
2703         key.type = BTRFS_EXTENT_ITEM_KEY;
2704
2705         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2706         if (ret < 0)
2707                 goto out;
2708         BUG_ON(ret == 0); /* Corruption */
2709
2710         ret = -ENOENT;
2711         if (path->slots[0] == 0)
2712                 goto out;
2713
2714         path->slots[0]--;
2715         leaf = path->nodes[0];
2716         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2717
2718         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2719                 goto out;
2720
2721         ret = 1;
2722         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2723 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2724         if (item_size < sizeof(*ei)) {
2725                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2726                 goto out;
2727         }
2728 #endif
2729         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2730
2731         if (item_size != sizeof(*ei) +
2732             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2733                 goto out;
2734
2735         if (btrfs_extent_generation(leaf, ei) <=
2736             btrfs_root_last_snapshot(&root->root_item))
2737                 goto out;
2738
2739         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2740         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2741             BTRFS_EXTENT_DATA_REF_KEY)
2742                 goto out;
2743
2744         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2745         if (btrfs_extent_refs(leaf, ei) !=
2746             btrfs_extent_data_ref_count(leaf, ref) ||
2747             btrfs_extent_data_ref_root(leaf, ref) !=
2748             root->root_key.objectid ||
2749             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2750             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2751                 goto out;
2752
2753         ret = 0;
2754 out:
2755         return ret;
2756 }
2757
2758 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2759                           struct btrfs_root *root,
2760                           u64 objectid, u64 offset, u64 bytenr)
2761 {
2762         struct btrfs_path *path;
2763         int ret;
2764         int ret2;
2765
2766         path = btrfs_alloc_path();
2767         if (!path)
2768                 return -ENOENT;
2769
2770         do {
2771                 ret = check_committed_ref(trans, root, path, objectid,
2772                                           offset, bytenr);
2773                 if (ret && ret != -ENOENT)
2774                         goto out;
2775
2776                 ret2 = check_delayed_ref(trans, root, path, objectid,
2777                                          offset, bytenr);
2778         } while (ret2 == -EAGAIN);
2779
2780         if (ret2 && ret2 != -ENOENT) {
2781                 ret = ret2;
2782                 goto out;
2783         }
2784
2785         if (ret != -ENOENT || ret2 != -ENOENT)
2786                 ret = 0;
2787 out:
2788         btrfs_free_path(path);
2789         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2790                 WARN_ON(ret > 0);
2791         return ret;
2792 }
2793
2794 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2795                            struct btrfs_root *root,
2796                            struct extent_buffer *buf,
2797                            int full_backref, int inc, int for_cow)
2798 {
2799         u64 bytenr;
2800         u64 num_bytes;
2801         u64 parent;
2802         u64 ref_root;
2803         u32 nritems;
2804         struct btrfs_key key;
2805         struct btrfs_file_extent_item *fi;
2806         int i;
2807         int level;
2808         int ret = 0;
2809         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2810                             u64, u64, u64, u64, u64, u64, int);
2811
2812         ref_root = btrfs_header_owner(buf);
2813         nritems = btrfs_header_nritems(buf);
2814         level = btrfs_header_level(buf);
2815
2816         if (!root->ref_cows && level == 0)
2817                 return 0;
2818
2819         if (inc)
2820                 process_func = btrfs_inc_extent_ref;
2821         else
2822                 process_func = btrfs_free_extent;
2823
2824         if (full_backref)
2825                 parent = buf->start;
2826         else
2827                 parent = 0;
2828
2829         for (i = 0; i < nritems; i++) {
2830                 if (level == 0) {
2831                         btrfs_item_key_to_cpu(buf, &key, i);
2832                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2833                                 continue;
2834                         fi = btrfs_item_ptr(buf, i,
2835                                             struct btrfs_file_extent_item);
2836                         if (btrfs_file_extent_type(buf, fi) ==
2837                             BTRFS_FILE_EXTENT_INLINE)
2838                                 continue;
2839                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2840                         if (bytenr == 0)
2841                                 continue;
2842
2843                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2844                         key.offset -= btrfs_file_extent_offset(buf, fi);
2845                         ret = process_func(trans, root, bytenr, num_bytes,
2846                                            parent, ref_root, key.objectid,
2847                                            key.offset, for_cow);
2848                         if (ret)
2849                                 goto fail;
2850                 } else {
2851                         bytenr = btrfs_node_blockptr(buf, i);
2852                         num_bytes = btrfs_level_size(root, level - 1);
2853                         ret = process_func(trans, root, bytenr, num_bytes,
2854                                            parent, ref_root, level - 1, 0,
2855                                            for_cow);
2856                         if (ret)
2857                                 goto fail;
2858                 }
2859         }
2860         return 0;
2861 fail:
2862         return ret;
2863 }
2864
2865 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2866                   struct extent_buffer *buf, int full_backref, int for_cow)
2867 {
2868         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2869 }
2870
2871 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2872                   struct extent_buffer *buf, int full_backref, int for_cow)
2873 {
2874         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2875 }
2876
2877 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2878                                  struct btrfs_root *root,
2879                                  struct btrfs_path *path,
2880                                  struct btrfs_block_group_cache *cache)
2881 {
2882         int ret;
2883         struct btrfs_root *extent_root = root->fs_info->extent_root;
2884         unsigned long bi;
2885         struct extent_buffer *leaf;
2886
2887         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2888         if (ret < 0)
2889                 goto fail;
2890         BUG_ON(ret); /* Corruption */
2891
2892         leaf = path->nodes[0];
2893         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2894         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2895         btrfs_mark_buffer_dirty(leaf);
2896         btrfs_release_path(path);
2897 fail:
2898         if (ret) {
2899                 btrfs_abort_transaction(trans, root, ret);
2900                 return ret;
2901         }
2902         return 0;
2903
2904 }
2905
2906 static struct btrfs_block_group_cache *
2907 next_block_group(struct btrfs_root *root,
2908                  struct btrfs_block_group_cache *cache)
2909 {
2910         struct rb_node *node;
2911         spin_lock(&root->fs_info->block_group_cache_lock);
2912         node = rb_next(&cache->cache_node);
2913         btrfs_put_block_group(cache);
2914         if (node) {
2915                 cache = rb_entry(node, struct btrfs_block_group_cache,
2916                                  cache_node);
2917                 btrfs_get_block_group(cache);
2918         } else
2919                 cache = NULL;
2920         spin_unlock(&root->fs_info->block_group_cache_lock);
2921         return cache;
2922 }
2923
2924 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2925                             struct btrfs_trans_handle *trans,
2926                             struct btrfs_path *path)
2927 {
2928         struct btrfs_root *root = block_group->fs_info->tree_root;
2929         struct inode *inode = NULL;
2930         u64 alloc_hint = 0;
2931         int dcs = BTRFS_DC_ERROR;
2932         int num_pages = 0;
2933         int retries = 0;
2934         int ret = 0;
2935
2936         /*
2937          * If this block group is smaller than 100 megs don't bother caching the
2938          * block group.
2939          */
2940         if (block_group->key.offset < (100 * 1024 * 1024)) {
2941                 spin_lock(&block_group->lock);
2942                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2943                 spin_unlock(&block_group->lock);
2944                 return 0;
2945         }
2946
2947 again:
2948         inode = lookup_free_space_inode(root, block_group, path);
2949         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2950                 ret = PTR_ERR(inode);
2951                 btrfs_release_path(path);
2952                 goto out;
2953         }
2954
2955         if (IS_ERR(inode)) {
2956                 BUG_ON(retries);
2957                 retries++;
2958
2959                 if (block_group->ro)
2960                         goto out_free;
2961
2962                 ret = create_free_space_inode(root, trans, block_group, path);
2963                 if (ret)
2964                         goto out_free;
2965                 goto again;
2966         }
2967
2968         /* We've already setup this transaction, go ahead and exit */
2969         if (block_group->cache_generation == trans->transid &&
2970             i_size_read(inode)) {
2971                 dcs = BTRFS_DC_SETUP;
2972                 goto out_put;
2973         }
2974
2975         /*
2976          * We want to set the generation to 0, that way if anything goes wrong
2977          * from here on out we know not to trust this cache when we load up next
2978          * time.
2979          */
2980         BTRFS_I(inode)->generation = 0;
2981         ret = btrfs_update_inode(trans, root, inode);
2982         WARN_ON(ret);
2983
2984         if (i_size_read(inode) > 0) {
2985                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2986                                                       inode);
2987                 if (ret)
2988                         goto out_put;
2989         }
2990
2991         spin_lock(&block_group->lock);
2992         if (block_group->cached != BTRFS_CACHE_FINISHED ||
2993             !btrfs_test_opt(root, SPACE_CACHE)) {
2994                 /*
2995                  * don't bother trying to write stuff out _if_
2996                  * a) we're not cached,
2997                  * b) we're with nospace_cache mount option.
2998                  */
2999                 dcs = BTRFS_DC_WRITTEN;
3000                 spin_unlock(&block_group->lock);
3001                 goto out_put;
3002         }
3003         spin_unlock(&block_group->lock);
3004
3005         /*
3006          * Try to preallocate enough space based on how big the block group is.
3007          * Keep in mind this has to include any pinned space which could end up
3008          * taking up quite a bit since it's not folded into the other space
3009          * cache.
3010          */
3011         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3012         if (!num_pages)
3013                 num_pages = 1;
3014
3015         num_pages *= 16;
3016         num_pages *= PAGE_CACHE_SIZE;
3017
3018         ret = btrfs_check_data_free_space(inode, num_pages);
3019         if (ret)
3020                 goto out_put;
3021
3022         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3023                                               num_pages, num_pages,
3024                                               &alloc_hint);
3025         if (!ret)
3026                 dcs = BTRFS_DC_SETUP;
3027         btrfs_free_reserved_data_space(inode, num_pages);
3028
3029 out_put:
3030         iput(inode);
3031 out_free:
3032         btrfs_release_path(path);
3033 out:
3034         spin_lock(&block_group->lock);
3035         if (!ret && dcs == BTRFS_DC_SETUP)
3036                 block_group->cache_generation = trans->transid;
3037         block_group->disk_cache_state = dcs;
3038         spin_unlock(&block_group->lock);
3039
3040         return ret;
3041 }
3042
3043 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3044                                    struct btrfs_root *root)
3045 {
3046         struct btrfs_block_group_cache *cache;
3047         int err = 0;
3048         struct btrfs_path *path;
3049         u64 last = 0;
3050
3051         path = btrfs_alloc_path();
3052         if (!path)
3053                 return -ENOMEM;
3054
3055 again:
3056         while (1) {
3057                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3058                 while (cache) {
3059                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3060                                 break;
3061                         cache = next_block_group(root, cache);
3062                 }
3063                 if (!cache) {
3064                         if (last == 0)
3065                                 break;
3066                         last = 0;
3067                         continue;
3068                 }
3069                 err = cache_save_setup(cache, trans, path);
3070                 last = cache->key.objectid + cache->key.offset;
3071                 btrfs_put_block_group(cache);
3072         }
3073
3074         while (1) {
3075                 if (last == 0) {
3076                         err = btrfs_run_delayed_refs(trans, root,
3077                                                      (unsigned long)-1);
3078                         if (err) /* File system offline */
3079                                 goto out;
3080                 }
3081
3082                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3083                 while (cache) {
3084                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3085                                 btrfs_put_block_group(cache);
3086                                 goto again;
3087                         }
3088
3089                         if (cache->dirty)
3090                                 break;
3091                         cache = next_block_group(root, cache);
3092                 }
3093                 if (!cache) {
3094                         if (last == 0)
3095                                 break;
3096                         last = 0;
3097                         continue;
3098                 }
3099
3100                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3101                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3102                 cache->dirty = 0;
3103                 last = cache->key.objectid + cache->key.offset;
3104
3105                 err = write_one_cache_group(trans, root, path, cache);
3106                 if (err) /* File system offline */
3107                         goto out;
3108
3109                 btrfs_put_block_group(cache);
3110         }
3111
3112         while (1) {
3113                 /*
3114                  * I don't think this is needed since we're just marking our
3115                  * preallocated extent as written, but just in case it can't
3116                  * hurt.
3117                  */
3118                 if (last == 0) {
3119                         err = btrfs_run_delayed_refs(trans, root,
3120                                                      (unsigned long)-1);
3121                         if (err) /* File system offline */
3122                                 goto out;
3123                 }
3124
3125                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3126                 while (cache) {
3127                         /*
3128                          * Really this shouldn't happen, but it could if we
3129                          * couldn't write the entire preallocated extent and
3130                          * splitting the extent resulted in a new block.
3131                          */
3132                         if (cache->dirty) {
3133                                 btrfs_put_block_group(cache);
3134                                 goto again;
3135                         }
3136                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3137                                 break;
3138                         cache = next_block_group(root, cache);
3139                 }
3140                 if (!cache) {
3141                         if (last == 0)
3142                                 break;
3143                         last = 0;
3144                         continue;
3145                 }
3146
3147                 err = btrfs_write_out_cache(root, trans, cache, path);
3148
3149                 /*
3150                  * If we didn't have an error then the cache state is still
3151                  * NEED_WRITE, so we can set it to WRITTEN.
3152                  */
3153                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3154                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3155                 last = cache->key.objectid + cache->key.offset;
3156                 btrfs_put_block_group(cache);
3157         }
3158 out:
3159
3160         btrfs_free_path(path);
3161         return err;
3162 }
3163
3164 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3165 {
3166         struct btrfs_block_group_cache *block_group;
3167         int readonly = 0;
3168
3169         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3170         if (!block_group || block_group->ro)
3171                 readonly = 1;
3172         if (block_group)
3173                 btrfs_put_block_group(block_group);
3174         return readonly;
3175 }
3176
3177 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3178                              u64 total_bytes, u64 bytes_used,
3179                              struct btrfs_space_info **space_info)
3180 {
3181         struct btrfs_space_info *found;
3182         int i;
3183         int factor;
3184
3185         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3186                      BTRFS_BLOCK_GROUP_RAID10))
3187                 factor = 2;
3188         else
3189                 factor = 1;
3190
3191         found = __find_space_info(info, flags);
3192         if (found) {
3193                 spin_lock(&found->lock);
3194                 found->total_bytes += total_bytes;
3195                 found->disk_total += total_bytes * factor;
3196                 found->bytes_used += bytes_used;
3197                 found->disk_used += bytes_used * factor;
3198                 found->full = 0;
3199                 spin_unlock(&found->lock);
3200                 *space_info = found;
3201                 return 0;
3202         }
3203         found = kzalloc(sizeof(*found), GFP_NOFS);
3204         if (!found)
3205                 return -ENOMEM;
3206
3207         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3208                 INIT_LIST_HEAD(&found->block_groups[i]);
3209         init_rwsem(&found->groups_sem);
3210         spin_lock_init(&found->lock);
3211         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3212         found->total_bytes = total_bytes;
3213         found->disk_total = total_bytes * factor;
3214         found->bytes_used = bytes_used;
3215         found->disk_used = bytes_used * factor;
3216         found->bytes_pinned = 0;
3217         found->bytes_reserved = 0;
3218         found->bytes_readonly = 0;
3219         found->bytes_may_use = 0;
3220         found->full = 0;
3221         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3222         found->chunk_alloc = 0;
3223         found->flush = 0;
3224         init_waitqueue_head(&found->wait);
3225         *space_info = found;
3226         list_add_rcu(&found->list, &info->space_info);
3227         if (flags & BTRFS_BLOCK_GROUP_DATA)
3228                 info->data_sinfo = found;
3229         return 0;
3230 }
3231
3232 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3233 {
3234         u64 extra_flags = chunk_to_extended(flags) &
3235                                 BTRFS_EXTENDED_PROFILE_MASK;
3236
3237         if (flags & BTRFS_BLOCK_GROUP_DATA)
3238                 fs_info->avail_data_alloc_bits |= extra_flags;
3239         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3240                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3241         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3242                 fs_info->avail_system_alloc_bits |= extra_flags;
3243 }
3244
3245 /*
3246  * returns target flags in extended format or 0 if restripe for this
3247  * chunk_type is not in progress
3248  *
3249  * should be called with either volume_mutex or balance_lock held
3250  */
3251 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3252 {
3253         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3254         u64 target = 0;
3255
3256         if (!bctl)
3257                 return 0;
3258
3259         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3260             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3261                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3262         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3263                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3264                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3265         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3266                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3267                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3268         }
3269
3270         return target;
3271 }
3272
3273 /*
3274  * @flags: available profiles in extended format (see ctree.h)
3275  *
3276  * Returns reduced profile in chunk format.  If profile changing is in
3277  * progress (either running or paused) picks the target profile (if it's
3278  * already available), otherwise falls back to plain reducing.
3279  */
3280 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3281 {
3282         /*
3283          * we add in the count of missing devices because we want
3284          * to make sure that any RAID levels on a degraded FS
3285          * continue to be honored.
3286          */
3287         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3288                 root->fs_info->fs_devices->missing_devices;
3289         u64 target;
3290
3291         /*
3292          * see if restripe for this chunk_type is in progress, if so
3293          * try to reduce to the target profile
3294          */
3295         spin_lock(&root->fs_info->balance_lock);
3296         target = get_restripe_target(root->fs_info, flags);
3297         if (target) {
3298                 /* pick target profile only if it's already available */
3299                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3300                         spin_unlock(&root->fs_info->balance_lock);
3301                         return extended_to_chunk(target);
3302                 }
3303         }
3304         spin_unlock(&root->fs_info->balance_lock);
3305
3306         if (num_devices == 1)
3307                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3308         if (num_devices < 4)
3309                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3310
3311         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3312             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3313                       BTRFS_BLOCK_GROUP_RAID10))) {
3314                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3315         }
3316
3317         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3318             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3319                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3320         }
3321
3322         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3323             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3324              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3325              (flags & BTRFS_BLOCK_GROUP_DUP))) {
3326                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3327         }
3328
3329         return extended_to_chunk(flags);
3330 }
3331
3332 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3333 {
3334         if (flags & BTRFS_BLOCK_GROUP_DATA)
3335                 flags |= root->fs_info->avail_data_alloc_bits;
3336         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3337                 flags |= root->fs_info->avail_system_alloc_bits;
3338         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3339                 flags |= root->fs_info->avail_metadata_alloc_bits;
3340
3341         return btrfs_reduce_alloc_profile(root, flags);
3342 }
3343
3344 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3345 {
3346         u64 flags;
3347
3348         if (data)
3349                 flags = BTRFS_BLOCK_GROUP_DATA;
3350         else if (root == root->fs_info->chunk_root)
3351                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3352         else
3353                 flags = BTRFS_BLOCK_GROUP_METADATA;
3354
3355         return get_alloc_profile(root, flags);
3356 }
3357
3358 /*
3359  * This will check the space that the inode allocates from to make sure we have
3360  * enough space for bytes.
3361  */
3362 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3363 {
3364         struct btrfs_space_info *data_sinfo;
3365         struct btrfs_root *root = BTRFS_I(inode)->root;
3366         struct btrfs_fs_info *fs_info = root->fs_info;
3367         u64 used;
3368         int ret = 0, committed = 0, alloc_chunk = 1;
3369
3370         /* make sure bytes are sectorsize aligned */
3371         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3372
3373         if (root == root->fs_info->tree_root ||
3374             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3375                 alloc_chunk = 0;
3376                 committed = 1;
3377         }
3378
3379         data_sinfo = fs_info->data_sinfo;
3380         if (!data_sinfo)
3381                 goto alloc;
3382
3383 again:
3384         /* make sure we have enough space to handle the data first */
3385         spin_lock(&data_sinfo->lock);
3386         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3387                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3388                 data_sinfo->bytes_may_use;
3389
3390         if (used + bytes > data_sinfo->total_bytes) {
3391                 struct btrfs_trans_handle *trans;
3392
3393                 /*
3394                  * if we don't have enough free bytes in this space then we need
3395                  * to alloc a new chunk.
3396                  */
3397                 if (!data_sinfo->full && alloc_chunk) {
3398                         u64 alloc_target;
3399
3400                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3401                         spin_unlock(&data_sinfo->lock);
3402 alloc:
3403                         alloc_target = btrfs_get_alloc_profile(root, 1);
3404                         trans = btrfs_join_transaction(root);
3405                         if (IS_ERR(trans))
3406                                 return PTR_ERR(trans);
3407
3408                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3409                                              bytes + 2 * 1024 * 1024,
3410                                              alloc_target,
3411                                              CHUNK_ALLOC_NO_FORCE);
3412                         btrfs_end_transaction(trans, root);
3413                         if (ret < 0) {
3414                                 if (ret != -ENOSPC)
3415                                         return ret;
3416                                 else
3417                                         goto commit_trans;
3418                         }
3419
3420                         if (!data_sinfo)
3421                                 data_sinfo = fs_info->data_sinfo;
3422
3423                         goto again;
3424                 }
3425
3426                 /*
3427                  * If we have less pinned bytes than we want to allocate then
3428                  * don't bother committing the transaction, it won't help us.
3429                  */
3430                 if (data_sinfo->bytes_pinned < bytes)
3431                         committed = 1;
3432                 spin_unlock(&data_sinfo->lock);
3433
3434                 /* commit the current transaction and try again */
3435 commit_trans:
3436                 if (!committed &&
3437                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3438                         committed = 1;
3439                         trans = btrfs_join_transaction(root);
3440                         if (IS_ERR(trans))
3441                                 return PTR_ERR(trans);
3442                         ret = btrfs_commit_transaction(trans, root);
3443                         if (ret)
3444                                 return ret;
3445                         goto again;
3446                 }
3447
3448                 return -ENOSPC;
3449         }
3450         data_sinfo->bytes_may_use += bytes;
3451         trace_btrfs_space_reservation(root->fs_info, "space_info",
3452                                       data_sinfo->flags, bytes, 1);
3453         spin_unlock(&data_sinfo->lock);
3454
3455         return 0;
3456 }
3457
3458 /*
3459  * Called if we need to clear a data reservation for this inode.
3460  */
3461 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3462 {
3463         struct btrfs_root *root = BTRFS_I(inode)->root;
3464         struct btrfs_space_info *data_sinfo;
3465
3466         /* make sure bytes are sectorsize aligned */
3467         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3468
3469         data_sinfo = root->fs_info->data_sinfo;
3470         spin_lock(&data_sinfo->lock);
3471         data_sinfo->bytes_may_use -= bytes;
3472         trace_btrfs_space_reservation(root->fs_info, "space_info",
3473                                       data_sinfo->flags, bytes, 0);
3474         spin_unlock(&data_sinfo->lock);
3475 }
3476
3477 static void force_metadata_allocation(struct btrfs_fs_info *info)
3478 {
3479         struct list_head *head = &info->space_info;
3480         struct btrfs_space_info *found;
3481
3482         rcu_read_lock();
3483         list_for_each_entry_rcu(found, head, list) {
3484                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3485                         found->force_alloc = CHUNK_ALLOC_FORCE;
3486         }
3487         rcu_read_unlock();
3488 }
3489
3490 static int should_alloc_chunk(struct btrfs_root *root,
3491                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3492                               int force)
3493 {
3494         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3495         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3496         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3497         u64 thresh;
3498
3499         if (force == CHUNK_ALLOC_FORCE)
3500                 return 1;
3501
3502         /*
3503          * We need to take into account the global rsv because for all intents
3504          * and purposes it's used space.  Don't worry about locking the
3505          * global_rsv, it doesn't change except when the transaction commits.
3506          */
3507         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3508                 num_allocated += global_rsv->size;
3509
3510         /*
3511          * in limited mode, we want to have some free space up to
3512          * about 1% of the FS size.
3513          */
3514         if (force == CHUNK_ALLOC_LIMITED) {
3515                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3516                 thresh = max_t(u64, 64 * 1024 * 1024,
3517                                div_factor_fine(thresh, 1));
3518
3519                 if (num_bytes - num_allocated < thresh)
3520                         return 1;
3521         }
3522
3523         if (num_allocated + alloc_bytes < div_factor(num_bytes, 8))
3524                 return 0;
3525         return 1;
3526 }
3527
3528 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3529 {
3530         u64 num_dev;
3531
3532         if (type & BTRFS_BLOCK_GROUP_RAID10 ||
3533             type & BTRFS_BLOCK_GROUP_RAID0)
3534                 num_dev = root->fs_info->fs_devices->rw_devices;
3535         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3536                 num_dev = 2;
3537         else
3538                 num_dev = 1;    /* DUP or single */
3539
3540         /* metadata for updaing devices and chunk tree */
3541         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3542 }
3543
3544 static void check_system_chunk(struct btrfs_trans_handle *trans,
3545                                struct btrfs_root *root, u64 type)
3546 {
3547         struct btrfs_space_info *info;
3548         u64 left;
3549         u64 thresh;
3550
3551         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3552         spin_lock(&info->lock);
3553         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3554                 info->bytes_reserved - info->bytes_readonly;
3555         spin_unlock(&info->lock);
3556
3557         thresh = get_system_chunk_thresh(root, type);
3558         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3559                 printk(KERN_INFO "left=%llu, need=%llu, flags=%llu\n",
3560                        left, thresh, type);
3561                 dump_space_info(info, 0, 0);
3562         }
3563
3564         if (left < thresh) {
3565                 u64 flags;
3566
3567                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3568                 btrfs_alloc_chunk(trans, root, flags);
3569         }
3570 }
3571
3572 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3573                           struct btrfs_root *extent_root, u64 alloc_bytes,
3574                           u64 flags, int force)
3575 {
3576         struct btrfs_space_info *space_info;
3577         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3578         int wait_for_alloc = 0;
3579         int ret = 0;
3580
3581         space_info = __find_space_info(extent_root->fs_info, flags);
3582         if (!space_info) {
3583                 ret = update_space_info(extent_root->fs_info, flags,
3584                                         0, 0, &space_info);
3585                 BUG_ON(ret); /* -ENOMEM */
3586         }
3587         BUG_ON(!space_info); /* Logic error */
3588
3589 again:
3590         spin_lock(&space_info->lock);
3591         if (force < space_info->force_alloc)
3592                 force = space_info->force_alloc;
3593         if (space_info->full) {
3594                 spin_unlock(&space_info->lock);
3595                 return 0;
3596         }
3597
3598         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3599                 spin_unlock(&space_info->lock);
3600                 return 0;
3601         } else if (space_info->chunk_alloc) {
3602                 wait_for_alloc = 1;
3603         } else {
3604                 space_info->chunk_alloc = 1;
3605         }
3606
3607         spin_unlock(&space_info->lock);
3608
3609         mutex_lock(&fs_info->chunk_mutex);
3610
3611         /*
3612          * The chunk_mutex is held throughout the entirety of a chunk
3613          * allocation, so once we've acquired the chunk_mutex we know that the
3614          * other guy is done and we need to recheck and see if we should
3615          * allocate.
3616          */
3617         if (wait_for_alloc) {
3618                 mutex_unlock(&fs_info->chunk_mutex);
3619                 wait_for_alloc = 0;
3620                 goto again;
3621         }
3622
3623         /*
3624          * If we have mixed data/metadata chunks we want to make sure we keep
3625          * allocating mixed chunks instead of individual chunks.
3626          */
3627         if (btrfs_mixed_space_info(space_info))
3628                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3629
3630         /*
3631          * if we're doing a data chunk, go ahead and make sure that
3632          * we keep a reasonable number of metadata chunks allocated in the
3633          * FS as well.
3634          */
3635         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3636                 fs_info->data_chunk_allocations++;
3637                 if (!(fs_info->data_chunk_allocations %
3638                       fs_info->metadata_ratio))
3639                         force_metadata_allocation(fs_info);
3640         }
3641
3642         /*
3643          * Check if we have enough space in SYSTEM chunk because we may need
3644          * to update devices.
3645          */
3646         check_system_chunk(trans, extent_root, flags);
3647
3648         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3649         if (ret < 0 && ret != -ENOSPC)
3650                 goto out;
3651
3652         spin_lock(&space_info->lock);
3653         if (ret)
3654                 space_info->full = 1;
3655         else
3656                 ret = 1;
3657
3658         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3659         space_info->chunk_alloc = 0;
3660         spin_unlock(&space_info->lock);
3661 out:
3662         mutex_unlock(&fs_info->chunk_mutex);
3663         return ret;
3664 }
3665
3666 static int can_overcommit(struct btrfs_root *root,
3667                           struct btrfs_space_info *space_info, u64 bytes,
3668                           int flush)
3669 {
3670         u64 profile = btrfs_get_alloc_profile(root, 0);
3671         u64 avail;
3672         u64 used;
3673
3674         used = space_info->bytes_used + space_info->bytes_reserved +
3675                 space_info->bytes_pinned + space_info->bytes_readonly +
3676                 space_info->bytes_may_use;
3677
3678         spin_lock(&root->fs_info->free_chunk_lock);
3679         avail = root->fs_info->free_chunk_space;
3680         spin_unlock(&root->fs_info->free_chunk_lock);
3681
3682         /*
3683          * If we have dup, raid1 or raid10 then only half of the free
3684          * space is actually useable.
3685          */
3686         if (profile & (BTRFS_BLOCK_GROUP_DUP |
3687                        BTRFS_BLOCK_GROUP_RAID1 |
3688                        BTRFS_BLOCK_GROUP_RAID10))
3689                 avail >>= 1;
3690
3691         /*
3692          * If we aren't flushing don't let us overcommit too much, say
3693          * 1/8th of the space.  If we can flush, let it overcommit up to
3694          * 1/2 of the space.
3695          */
3696         if (flush)
3697                 avail >>= 3;
3698         else
3699                 avail >>= 1;
3700
3701         if (used + bytes < space_info->total_bytes + avail)
3702                 return 1;
3703         return 0;
3704 }
3705
3706 /*
3707  * shrink metadata reservation for delalloc
3708  */
3709 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3710                             bool wait_ordered)
3711 {
3712         struct btrfs_block_rsv *block_rsv;
3713         struct btrfs_space_info *space_info;
3714         struct btrfs_trans_handle *trans;
3715         u64 delalloc_bytes;
3716         u64 max_reclaim;
3717         long time_left;
3718         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3719         int loops = 0;
3720
3721         trans = (struct btrfs_trans_handle *)current->journal_info;
3722         block_rsv = &root->fs_info->delalloc_block_rsv;
3723         space_info = block_rsv->space_info;
3724
3725         smp_mb();
3726         delalloc_bytes = root->fs_info->delalloc_bytes;
3727         if (delalloc_bytes == 0) {
3728                 if (trans)
3729                         return;
3730                 btrfs_wait_ordered_extents(root, 0, 0);
3731                 return;
3732         }
3733
3734         while (delalloc_bytes && loops < 3) {
3735                 max_reclaim = min(delalloc_bytes, to_reclaim);
3736                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3737                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3738                                                WB_REASON_FS_FREE_SPACE);
3739
3740                 /*
3741                  * We need to wait for the async pages to actually start before
3742                  * we do anything.
3743                  */
3744                 wait_event(root->fs_info->async_submit_wait,
3745                            !atomic_read(&root->fs_info->async_delalloc_pages));
3746
3747                 spin_lock(&space_info->lock);
3748                 if (can_overcommit(root, space_info, orig, !trans)) {
3749                         spin_unlock(&space_info->lock);
3750                         break;
3751                 }
3752                 spin_unlock(&space_info->lock);
3753
3754                 loops++;
3755                 if (wait_ordered && !trans) {
3756                         btrfs_wait_ordered_extents(root, 0, 0);
3757                 } else {
3758                         time_left = schedule_timeout_killable(1);
3759                         if (time_left)
3760                                 break;
3761                 }
3762                 smp_mb();
3763                 delalloc_bytes = root->fs_info->delalloc_bytes;
3764         }
3765 }
3766
3767 /**
3768  * maybe_commit_transaction - possibly commit the transaction if its ok to
3769  * @root - the root we're allocating for
3770  * @bytes - the number of bytes we want to reserve
3771  * @force - force the commit
3772  *
3773  * This will check to make sure that committing the transaction will actually
3774  * get us somewhere and then commit the transaction if it does.  Otherwise it
3775  * will return -ENOSPC.
3776  */
3777 static int may_commit_transaction(struct btrfs_root *root,
3778                                   struct btrfs_space_info *space_info,
3779                                   u64 bytes, int force)
3780 {
3781         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3782         struct btrfs_trans_handle *trans;
3783
3784         trans = (struct btrfs_trans_handle *)current->journal_info;
3785         if (trans)
3786                 return -EAGAIN;
3787
3788         if (force)
3789                 goto commit;
3790
3791         /* See if there is enough pinned space to make this reservation */
3792         spin_lock(&space_info->lock);
3793         if (space_info->bytes_pinned >= bytes) {
3794                 spin_unlock(&space_info->lock);
3795                 goto commit;
3796         }
3797         spin_unlock(&space_info->lock);
3798
3799         /*
3800          * See if there is some space in the delayed insertion reservation for
3801          * this reservation.
3802          */
3803         if (space_info != delayed_rsv->space_info)
3804                 return -ENOSPC;
3805
3806         spin_lock(&space_info->lock);
3807         spin_lock(&delayed_rsv->lock);
3808         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3809                 spin_unlock(&delayed_rsv->lock);
3810                 spin_unlock(&space_info->lock);
3811                 return -ENOSPC;
3812         }
3813         spin_unlock(&delayed_rsv->lock);
3814         spin_unlock(&space_info->lock);
3815
3816 commit:
3817         trans = btrfs_join_transaction(root);
3818         if (IS_ERR(trans))
3819                 return -ENOSPC;
3820
3821         return btrfs_commit_transaction(trans, root);
3822 }
3823
3824 enum flush_state {
3825         FLUSH_DELALLOC          =       1,
3826         FLUSH_DELALLOC_WAIT     =       2,
3827         FLUSH_DELAYED_ITEMS_NR  =       3,
3828         FLUSH_DELAYED_ITEMS     =       4,
3829         COMMIT_TRANS            =       5,
3830 };
3831
3832 static int flush_space(struct btrfs_root *root,
3833                        struct btrfs_space_info *space_info, u64 num_bytes,
3834                        u64 orig_bytes, int state)
3835 {
3836         struct btrfs_trans_handle *trans;
3837         int nr;
3838         int ret = 0;
3839
3840         switch (state) {
3841         case FLUSH_DELALLOC:
3842         case FLUSH_DELALLOC_WAIT:
3843                 shrink_delalloc(root, num_bytes, orig_bytes,
3844                                 state == FLUSH_DELALLOC_WAIT);
3845                 break;
3846         case FLUSH_DELAYED_ITEMS_NR:
3847         case FLUSH_DELAYED_ITEMS:
3848                 if (state == FLUSH_DELAYED_ITEMS_NR) {
3849                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
3850
3851                         nr = (int)div64_u64(num_bytes, bytes);
3852                         if (!nr)
3853                                 nr = 1;
3854                         nr *= 2;
3855                 } else {
3856                         nr = -1;
3857                 }
3858                 trans = btrfs_join_transaction(root);
3859                 if (IS_ERR(trans)) {
3860                         ret = PTR_ERR(trans);
3861                         break;
3862                 }
3863                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
3864                 btrfs_end_transaction(trans, root);
3865                 break;
3866         case COMMIT_TRANS:
3867                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3868                 break;
3869         default:
3870                 ret = -ENOSPC;
3871                 break;
3872         }
3873
3874         return ret;
3875 }
3876 /**
3877  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3878  * @root - the root we're allocating for
3879  * @block_rsv - the block_rsv we're allocating for
3880  * @orig_bytes - the number of bytes we want
3881  * @flush - wether or not we can flush to make our reservation
3882  *
3883  * This will reserve orgi_bytes number of bytes from the space info associated
3884  * with the block_rsv.  If there is not enough space it will make an attempt to
3885  * flush out space to make room.  It will do this by flushing delalloc if
3886  * possible or committing the transaction.  If flush is 0 then no attempts to
3887  * regain reservations will be made and this will fail if there is not enough
3888  * space already.
3889  */
3890 static int reserve_metadata_bytes(struct btrfs_root *root,
3891                                   struct btrfs_block_rsv *block_rsv,
3892                                   u64 orig_bytes, int flush)
3893 {
3894         struct btrfs_space_info *space_info = block_rsv->space_info;
3895         u64 used;
3896         u64 num_bytes = orig_bytes;
3897         int flush_state = FLUSH_DELALLOC;
3898         int ret = 0;
3899         bool flushing = false;
3900         bool committed = false;
3901
3902 again:
3903         ret = 0;
3904         spin_lock(&space_info->lock);
3905         /*
3906          * We only want to wait if somebody other than us is flushing and we are
3907          * actually alloed to flush.
3908          */
3909         while (flush && !flushing && space_info->flush) {
3910                 spin_unlock(&space_info->lock);
3911                 /*
3912                  * If we have a trans handle we can't wait because the flusher
3913                  * may have to commit the transaction, which would mean we would
3914                  * deadlock since we are waiting for the flusher to finish, but
3915                  * hold the current transaction open.
3916                  */
3917                 if (current->journal_info)
3918                         return -EAGAIN;
3919                 ret = wait_event_killable(space_info->wait, !space_info->flush);
3920                 /* Must have been killed, return */
3921                 if (ret)
3922                         return -EINTR;
3923
3924                 spin_lock(&space_info->lock);
3925         }
3926
3927         ret = -ENOSPC;
3928         used = space_info->bytes_used + space_info->bytes_reserved +
3929                 space_info->bytes_pinned + space_info->bytes_readonly +
3930                 space_info->bytes_may_use;
3931
3932         /*
3933          * The idea here is that we've not already over-reserved the block group
3934          * then we can go ahead and save our reservation first and then start
3935          * flushing if we need to.  Otherwise if we've already overcommitted
3936          * lets start flushing stuff first and then come back and try to make
3937          * our reservation.
3938          */
3939         if (used <= space_info->total_bytes) {
3940                 if (used + orig_bytes <= space_info->total_bytes) {
3941                         space_info->bytes_may_use += orig_bytes;
3942                         trace_btrfs_space_reservation(root->fs_info,
3943                                 "space_info", space_info->flags, orig_bytes, 1);
3944                         ret = 0;
3945                 } else {
3946                         /*
3947                          * Ok set num_bytes to orig_bytes since we aren't
3948                          * overocmmitted, this way we only try and reclaim what
3949                          * we need.
3950                          */
3951                         num_bytes = orig_bytes;
3952                 }
3953         } else {
3954                 /*
3955                  * Ok we're over committed, set num_bytes to the overcommitted
3956                  * amount plus the amount of bytes that we need for this
3957                  * reservation.
3958                  */
3959                 num_bytes = used - space_info->total_bytes +
3960                         (orig_bytes * 2);
3961         }
3962
3963         if (ret) {
3964                 u64 avail;
3965
3966                 /*
3967                  * If we have a lot of space that's pinned, don't bother doing
3968                  * the overcommit dance yet and just commit the transaction.
3969                  */
3970                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3971                 do_div(avail, 10);
3972                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3973                         space_info->flush = 1;
3974                         flushing = true;
3975                         spin_unlock(&space_info->lock);
3976                         ret = may_commit_transaction(root, space_info,
3977                                                      orig_bytes, 1);
3978                         if (ret)
3979                                 goto out;
3980                         committed = true;
3981                         goto again;
3982                 }
3983
3984                 if (can_overcommit(root, space_info, orig_bytes, flush)) {
3985                         space_info->bytes_may_use += orig_bytes;
3986                         trace_btrfs_space_reservation(root->fs_info,
3987                                 "space_info", space_info->flags, orig_bytes, 1);
3988                         ret = 0;
3989                 }
3990         }
3991
3992         /*
3993          * Couldn't make our reservation, save our place so while we're trying
3994          * to reclaim space we can actually use it instead of somebody else
3995          * stealing it from us.
3996          */
3997         if (ret && flush) {
3998                 flushing = true;
3999                 space_info->flush = 1;
4000         }
4001
4002         spin_unlock(&space_info->lock);
4003
4004         if (!ret || !flush)
4005                 goto out;
4006
4007         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4008                           flush_state);
4009         flush_state++;
4010         if (!ret)
4011                 goto again;
4012         else if (flush_state <= COMMIT_TRANS)
4013                 goto again;
4014
4015 out:
4016         if (flushing) {
4017                 spin_lock(&space_info->lock);
4018                 space_info->flush = 0;
4019                 wake_up_all(&space_info->wait);
4020                 spin_unlock(&space_info->lock);
4021         }
4022         return ret;
4023 }
4024
4025 static struct btrfs_block_rsv *get_block_rsv(
4026                                         const struct btrfs_trans_handle *trans,
4027                                         const struct btrfs_root *root)
4028 {
4029         struct btrfs_block_rsv *block_rsv = NULL;
4030
4031         if (root->ref_cows)
4032                 block_rsv = trans->block_rsv;
4033
4034         if (root == root->fs_info->csum_root && trans->adding_csums)
4035                 block_rsv = trans->block_rsv;
4036
4037         if (!block_rsv)
4038                 block_rsv = root->block_rsv;
4039
4040         if (!block_rsv)
4041                 block_rsv = &root->fs_info->empty_block_rsv;
4042
4043         return block_rsv;
4044 }
4045
4046 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4047                                u64 num_bytes)
4048 {
4049         int ret = -ENOSPC;
4050         spin_lock(&block_rsv->lock);
4051         if (block_rsv->reserved >= num_bytes) {
4052                 block_rsv->reserved -= num_bytes;
4053                 if (block_rsv->reserved < block_rsv->size)
4054                         block_rsv->full = 0;
4055                 ret = 0;
4056         }
4057         spin_unlock(&block_rsv->lock);
4058         return ret;
4059 }
4060
4061 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4062                                 u64 num_bytes, int update_size)
4063 {
4064         spin_lock(&block_rsv->lock);
4065         block_rsv->reserved += num_bytes;
4066         if (update_size)
4067                 block_rsv->size += num_bytes;
4068         else if (block_rsv->reserved >= block_rsv->size)
4069                 block_rsv->full = 1;
4070         spin_unlock(&block_rsv->lock);
4071 }
4072
4073 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4074                                     struct btrfs_block_rsv *block_rsv,
4075                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4076 {
4077         struct btrfs_space_info *space_info = block_rsv->space_info;
4078
4079         spin_lock(&block_rsv->lock);
4080         if (num_bytes == (u64)-1)
4081                 num_bytes = block_rsv->size;
4082         block_rsv->size -= num_bytes;
4083         if (block_rsv->reserved >= block_rsv->size) {
4084                 num_bytes = block_rsv->reserved - block_rsv->size;
4085                 block_rsv->reserved = block_rsv->size;
4086                 block_rsv->full = 1;
4087         } else {
4088                 num_bytes = 0;
4089         }
4090         spin_unlock(&block_rsv->lock);
4091
4092         if (num_bytes > 0) {
4093                 if (dest) {
4094                         spin_lock(&dest->lock);
4095                         if (!dest->full) {
4096                                 u64 bytes_to_add;
4097
4098                                 bytes_to_add = dest->size - dest->reserved;
4099                                 bytes_to_add = min(num_bytes, bytes_to_add);
4100                                 dest->reserved += bytes_to_add;
4101                                 if (dest->reserved >= dest->size)
4102                                         dest->full = 1;
4103                                 num_bytes -= bytes_to_add;
4104                         }
4105                         spin_unlock(&dest->lock);
4106                 }
4107                 if (num_bytes) {
4108                         spin_lock(&space_info->lock);
4109                         space_info->bytes_may_use -= num_bytes;
4110                         trace_btrfs_space_reservation(fs_info, "space_info",
4111                                         space_info->flags, num_bytes, 0);
4112                         space_info->reservation_progress++;
4113                         spin_unlock(&space_info->lock);
4114                 }
4115         }
4116 }
4117
4118 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4119                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4120 {
4121         int ret;
4122
4123         ret = block_rsv_use_bytes(src, num_bytes);
4124         if (ret)
4125                 return ret;
4126
4127         block_rsv_add_bytes(dst, num_bytes, 1);
4128         return 0;
4129 }
4130
4131 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4132 {
4133         memset(rsv, 0, sizeof(*rsv));
4134         spin_lock_init(&rsv->lock);
4135         rsv->type = type;
4136 }
4137
4138 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4139                                               unsigned short type)
4140 {
4141         struct btrfs_block_rsv *block_rsv;
4142         struct btrfs_fs_info *fs_info = root->fs_info;
4143
4144         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4145         if (!block_rsv)
4146                 return NULL;
4147
4148         btrfs_init_block_rsv(block_rsv, type);
4149         block_rsv->space_info = __find_space_info(fs_info,
4150                                                   BTRFS_BLOCK_GROUP_METADATA);
4151         return block_rsv;
4152 }
4153
4154 void btrfs_free_block_rsv(struct btrfs_root *root,
4155                           struct btrfs_block_rsv *rsv)
4156 {
4157         if (!rsv)
4158                 return;
4159         btrfs_block_rsv_release(root, rsv, (u64)-1);
4160         kfree(rsv);
4161 }
4162
4163 static inline int __block_rsv_add(struct btrfs_root *root,
4164                                   struct btrfs_block_rsv *block_rsv,
4165                                   u64 num_bytes, int flush)
4166 {
4167         int ret;
4168
4169         if (num_bytes == 0)
4170                 return 0;
4171
4172         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4173         if (!ret) {
4174                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4175                 return 0;
4176         }
4177
4178         return ret;
4179 }
4180
4181 int btrfs_block_rsv_add(struct btrfs_root *root,
4182                         struct btrfs_block_rsv *block_rsv,
4183                         u64 num_bytes)
4184 {
4185         return __block_rsv_add(root, block_rsv, num_bytes, 1);
4186 }
4187
4188 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
4189                                 struct btrfs_block_rsv *block_rsv,
4190                                 u64 num_bytes)
4191 {
4192         return __block_rsv_add(root, block_rsv, num_bytes, 0);
4193 }
4194
4195 int btrfs_block_rsv_check(struct btrfs_root *root,
4196                           struct btrfs_block_rsv *block_rsv, int min_factor)
4197 {
4198         u64 num_bytes = 0;
4199         int ret = -ENOSPC;
4200
4201         if (!block_rsv)
4202                 return 0;
4203
4204         spin_lock(&block_rsv->lock);
4205         num_bytes = div_factor(block_rsv->size, min_factor);
4206         if (block_rsv->reserved >= num_bytes)
4207                 ret = 0;
4208         spin_unlock(&block_rsv->lock);
4209
4210         return ret;
4211 }
4212
4213 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
4214                                            struct btrfs_block_rsv *block_rsv,
4215                                            u64 min_reserved, int flush)
4216 {
4217         u64 num_bytes = 0;
4218         int ret = -ENOSPC;
4219
4220         if (!block_rsv)
4221                 return 0;
4222
4223         spin_lock(&block_rsv->lock);
4224         num_bytes = min_reserved;
4225         if (block_rsv->reserved >= num_bytes)
4226                 ret = 0;
4227         else
4228                 num_bytes -= block_rsv->reserved;
4229         spin_unlock(&block_rsv->lock);
4230
4231         if (!ret)
4232                 return 0;
4233
4234         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4235         if (!ret) {
4236                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4237                 return 0;
4238         }
4239
4240         return ret;
4241 }
4242
4243 int btrfs_block_rsv_refill(struct btrfs_root *root,
4244                            struct btrfs_block_rsv *block_rsv,
4245                            u64 min_reserved)
4246 {
4247         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
4248 }
4249
4250 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
4251                                    struct btrfs_block_rsv *block_rsv,
4252                                    u64 min_reserved)
4253 {
4254         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
4255 }
4256
4257 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4258                             struct btrfs_block_rsv *dst_rsv,
4259                             u64 num_bytes)
4260 {
4261         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4262 }
4263
4264 void btrfs_block_rsv_release(struct btrfs_root *root,
4265                              struct btrfs_block_rsv *block_rsv,
4266                              u64 num_bytes)
4267 {
4268         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4269         if (global_rsv->full || global_rsv == block_rsv ||
4270             block_rsv->space_info != global_rsv->space_info)
4271                 global_rsv = NULL;
4272         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4273                                 num_bytes);
4274 }
4275
4276 /*
4277  * helper to calculate size of global block reservation.
4278  * the desired value is sum of space used by extent tree,
4279  * checksum tree and root tree
4280  */
4281 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4282 {
4283         struct btrfs_space_info *sinfo;
4284         u64 num_bytes;
4285         u64 meta_used;
4286         u64 data_used;
4287         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4288
4289         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4290         spin_lock(&sinfo->lock);
4291         data_used = sinfo->bytes_used;
4292         spin_unlock(&sinfo->lock);
4293
4294         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4295         spin_lock(&sinfo->lock);
4296         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4297                 data_used = 0;
4298         meta_used = sinfo->bytes_used;
4299         spin_unlock(&sinfo->lock);
4300
4301         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4302                     csum_size * 2;
4303         num_bytes += div64_u64(data_used + meta_used, 50);
4304
4305         if (num_bytes * 3 > meta_used)
4306                 num_bytes = div64_u64(meta_used, 3);
4307
4308         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4309 }
4310
4311 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4312 {
4313         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4314         struct btrfs_space_info *sinfo = block_rsv->space_info;
4315         u64 num_bytes;
4316
4317         num_bytes = calc_global_metadata_size(fs_info);
4318
4319         spin_lock(&sinfo->lock);
4320         spin_lock(&block_rsv->lock);
4321
4322         block_rsv->size = num_bytes;
4323
4324         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4325                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4326                     sinfo->bytes_may_use;
4327
4328         if (sinfo->total_bytes > num_bytes) {
4329                 num_bytes = sinfo->total_bytes - num_bytes;
4330                 block_rsv->reserved += num_bytes;
4331                 sinfo->bytes_may_use += num_bytes;
4332                 trace_btrfs_space_reservation(fs_info, "space_info",
4333                                       sinfo->flags, num_bytes, 1);
4334         }
4335
4336         if (block_rsv->reserved >= block_rsv->size) {
4337                 num_bytes = block_rsv->reserved - block_rsv->size;
4338                 sinfo->bytes_may_use -= num_bytes;
4339                 trace_btrfs_space_reservation(fs_info, "space_info",
4340                                       sinfo->flags, num_bytes, 0);
4341                 sinfo->reservation_progress++;
4342                 block_rsv->reserved = block_rsv->size;
4343                 block_rsv->full = 1;
4344         }
4345
4346         spin_unlock(&block_rsv->lock);
4347         spin_unlock(&sinfo->lock);
4348 }
4349
4350 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4351 {
4352         struct btrfs_space_info *space_info;
4353
4354         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4355         fs_info->chunk_block_rsv.space_info = space_info;
4356
4357         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4358         fs_info->global_block_rsv.space_info = space_info;
4359         fs_info->delalloc_block_rsv.space_info = space_info;
4360         fs_info->trans_block_rsv.space_info = space_info;
4361         fs_info->empty_block_rsv.space_info = space_info;
4362         fs_info->delayed_block_rsv.space_info = space_info;
4363
4364         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4365         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4366         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4367         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4368         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4369
4370         update_global_block_rsv(fs_info);
4371 }
4372
4373 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4374 {
4375         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4376                                 (u64)-1);
4377         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4378         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4379         WARN_ON(fs_info->trans_block_rsv.size > 0);
4380         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4381         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4382         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4383         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4384         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4385 }
4386
4387 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4388                                   struct btrfs_root *root)
4389 {
4390         if (!trans->block_rsv)
4391                 return;
4392
4393         if (!trans->bytes_reserved)
4394                 return;
4395
4396         trace_btrfs_space_reservation(root->fs_info, "transaction",
4397                                       trans->transid, trans->bytes_reserved, 0);
4398         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4399         trans->bytes_reserved = 0;
4400 }
4401
4402 /* Can only return 0 or -ENOSPC */
4403 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4404                                   struct inode *inode)
4405 {
4406         struct btrfs_root *root = BTRFS_I(inode)->root;
4407         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4408         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4409
4410         /*
4411          * We need to hold space in order to delete our orphan item once we've
4412          * added it, so this takes the reservation so we can release it later
4413          * when we are truly done with the orphan item.
4414          */
4415         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4416         trace_btrfs_space_reservation(root->fs_info, "orphan",
4417                                       btrfs_ino(inode), num_bytes, 1);
4418         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4419 }
4420
4421 void btrfs_orphan_release_metadata(struct inode *inode)
4422 {
4423         struct btrfs_root *root = BTRFS_I(inode)->root;
4424         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4425         trace_btrfs_space_reservation(root->fs_info, "orphan",
4426                                       btrfs_ino(inode), num_bytes, 0);
4427         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4428 }
4429
4430 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4431                                 struct btrfs_pending_snapshot *pending)
4432 {
4433         struct btrfs_root *root = pending->root;
4434         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4435         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4436         /*
4437          * two for root back/forward refs, two for directory entries,
4438          * one for root of the snapshot and one for parent inode.
4439          */
4440         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 6);
4441         dst_rsv->space_info = src_rsv->space_info;
4442         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4443 }
4444
4445 /**
4446  * drop_outstanding_extent - drop an outstanding extent
4447  * @inode: the inode we're dropping the extent for
4448  *
4449  * This is called when we are freeing up an outstanding extent, either called
4450  * after an error or after an extent is written.  This will return the number of
4451  * reserved extents that need to be freed.  This must be called with
4452  * BTRFS_I(inode)->lock held.
4453  */
4454 static unsigned drop_outstanding_extent(struct inode *inode)
4455 {
4456         unsigned drop_inode_space = 0;
4457         unsigned dropped_extents = 0;
4458
4459         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4460         BTRFS_I(inode)->outstanding_extents--;
4461
4462         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4463             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4464                                &BTRFS_I(inode)->runtime_flags))
4465                 drop_inode_space = 1;
4466
4467         /*
4468          * If we have more or the same amount of outsanding extents than we have
4469          * reserved then we need to leave the reserved extents count alone.
4470          */
4471         if (BTRFS_I(inode)->outstanding_extents >=
4472             BTRFS_I(inode)->reserved_extents)
4473                 return drop_inode_space;
4474
4475         dropped_extents = BTRFS_I(inode)->reserved_extents -
4476                 BTRFS_I(inode)->outstanding_extents;
4477         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4478         return dropped_extents + drop_inode_space;
4479 }
4480
4481 /**
4482  * calc_csum_metadata_size - return the amount of metada space that must be
4483  *      reserved/free'd for the given bytes.
4484  * @inode: the inode we're manipulating
4485  * @num_bytes: the number of bytes in question
4486  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4487  *
4488  * This adjusts the number of csum_bytes in the inode and then returns the
4489  * correct amount of metadata that must either be reserved or freed.  We
4490  * calculate how many checksums we can fit into one leaf and then divide the
4491  * number of bytes that will need to be checksumed by this value to figure out
4492  * how many checksums will be required.  If we are adding bytes then the number
4493  * may go up and we will return the number of additional bytes that must be
4494  * reserved.  If it is going down we will return the number of bytes that must
4495  * be freed.
4496  *
4497  * This must be called with BTRFS_I(inode)->lock held.
4498  */
4499 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4500                                    int reserve)
4501 {
4502         struct btrfs_root *root = BTRFS_I(inode)->root;
4503         u64 csum_size;
4504         int num_csums_per_leaf;
4505         int num_csums;
4506         int old_csums;
4507
4508         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4509             BTRFS_I(inode)->csum_bytes == 0)
4510                 return 0;
4511
4512         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4513         if (reserve)
4514                 BTRFS_I(inode)->csum_bytes += num_bytes;
4515         else
4516                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4517         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4518         num_csums_per_leaf = (int)div64_u64(csum_size,
4519                                             sizeof(struct btrfs_csum_item) +
4520                                             sizeof(struct btrfs_disk_key));
4521         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4522         num_csums = num_csums + num_csums_per_leaf - 1;
4523         num_csums = num_csums / num_csums_per_leaf;
4524
4525         old_csums = old_csums + num_csums_per_leaf - 1;
4526         old_csums = old_csums / num_csums_per_leaf;
4527
4528         /* No change, no need to reserve more */
4529         if (old_csums == num_csums)
4530                 return 0;
4531
4532         if (reserve)
4533                 return btrfs_calc_trans_metadata_size(root,
4534                                                       num_csums - old_csums);
4535
4536         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4537 }
4538
4539 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4540 {
4541         struct btrfs_root *root = BTRFS_I(inode)->root;
4542         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4543         u64 to_reserve = 0;
4544         u64 csum_bytes;
4545         unsigned nr_extents = 0;
4546         int extra_reserve = 0;
4547         int flush = 1;
4548         int ret;
4549
4550         /* Need to be holding the i_mutex here if we aren't free space cache */
4551         if (btrfs_is_free_space_inode(inode))
4552                 flush = 0;
4553
4554         if (flush && btrfs_transaction_in_commit(root->fs_info))
4555                 schedule_timeout(1);
4556
4557         mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4558         num_bytes = ALIGN(num_bytes, root->sectorsize);
4559
4560         spin_lock(&BTRFS_I(inode)->lock);
4561         BTRFS_I(inode)->outstanding_extents++;
4562
4563         if (BTRFS_I(inode)->outstanding_extents >
4564             BTRFS_I(inode)->reserved_extents)
4565                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4566                         BTRFS_I(inode)->reserved_extents;
4567
4568         /*
4569          * Add an item to reserve for updating the inode when we complete the
4570          * delalloc io.
4571          */
4572         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4573                       &BTRFS_I(inode)->runtime_flags)) {
4574                 nr_extents++;
4575                 extra_reserve = 1;
4576         }
4577
4578         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4579         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4580         csum_bytes = BTRFS_I(inode)->csum_bytes;
4581         spin_unlock(&BTRFS_I(inode)->lock);
4582
4583         if (root->fs_info->quota_enabled) {
4584                 ret = btrfs_qgroup_reserve(root, num_bytes +
4585                                            nr_extents * root->leafsize);
4586                 if (ret) {
4587                         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4588                         return ret;
4589                 }
4590         }
4591
4592         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4593         if (ret) {
4594                 u64 to_free = 0;
4595                 unsigned dropped;
4596
4597                 spin_lock(&BTRFS_I(inode)->lock);
4598                 dropped = drop_outstanding_extent(inode);
4599                 /*
4600                  * If the inodes csum_bytes is the same as the original
4601                  * csum_bytes then we know we haven't raced with any free()ers
4602                  * so we can just reduce our inodes csum bytes and carry on.
4603                  * Otherwise we have to do the normal free thing to account for
4604                  * the case that the free side didn't free up its reserve
4605                  * because of this outstanding reservation.
4606                  */
4607                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4608                         calc_csum_metadata_size(inode, num_bytes, 0);
4609                 else
4610                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4611                 spin_unlock(&BTRFS_I(inode)->lock);
4612                 if (dropped)
4613                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4614
4615                 if (to_free) {
4616                         btrfs_block_rsv_release(root, block_rsv, to_free);
4617                         trace_btrfs_space_reservation(root->fs_info,
4618                                                       "delalloc",
4619                                                       btrfs_ino(inode),
4620                                                       to_free, 0);
4621                 }
4622                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4623                 return ret;
4624         }
4625
4626         spin_lock(&BTRFS_I(inode)->lock);
4627         if (extra_reserve) {
4628                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4629                         &BTRFS_I(inode)->runtime_flags);
4630                 nr_extents--;
4631         }
4632         BTRFS_I(inode)->reserved_extents += nr_extents;
4633         spin_unlock(&BTRFS_I(inode)->lock);
4634         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4635
4636         if (to_reserve)
4637                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4638                                               btrfs_ino(inode), to_reserve, 1);
4639         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4640
4641         return 0;
4642 }
4643
4644 /**
4645  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4646  * @inode: the inode to release the reservation for
4647  * @num_bytes: the number of bytes we're releasing
4648  *
4649  * This will release the metadata reservation for an inode.  This can be called
4650  * once we complete IO for a given set of bytes to release their metadata
4651  * reservations.
4652  */
4653 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4654 {
4655         struct btrfs_root *root = BTRFS_I(inode)->root;
4656         u64 to_free = 0;
4657         unsigned dropped;
4658
4659         num_bytes = ALIGN(num_bytes, root->sectorsize);
4660         spin_lock(&BTRFS_I(inode)->lock);
4661         dropped = drop_outstanding_extent(inode);
4662
4663         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4664         spin_unlock(&BTRFS_I(inode)->lock);
4665         if (dropped > 0)
4666                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4667
4668         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4669                                       btrfs_ino(inode), to_free, 0);
4670         if (root->fs_info->quota_enabled) {
4671                 btrfs_qgroup_free(root, num_bytes +
4672                                         dropped * root->leafsize);
4673         }
4674
4675         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4676                                 to_free);
4677 }
4678
4679 /**
4680  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4681  * @inode: inode we're writing to
4682  * @num_bytes: the number of bytes we want to allocate
4683  *
4684  * This will do the following things
4685  *
4686  * o reserve space in the data space info for num_bytes
4687  * o reserve space in the metadata space info based on number of outstanding
4688  *   extents and how much csums will be needed
4689  * o add to the inodes ->delalloc_bytes
4690  * o add it to the fs_info's delalloc inodes list.
4691  *
4692  * This will return 0 for success and -ENOSPC if there is no space left.
4693  */
4694 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4695 {
4696         int ret;
4697
4698         ret = btrfs_check_data_free_space(inode, num_bytes);
4699         if (ret)
4700                 return ret;
4701
4702         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4703         if (ret) {
4704                 btrfs_free_reserved_data_space(inode, num_bytes);
4705                 return ret;
4706         }
4707
4708         return 0;
4709 }
4710
4711 /**
4712  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4713  * @inode: inode we're releasing space for
4714  * @num_bytes: the number of bytes we want to free up
4715  *
4716  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4717  * called in the case that we don't need the metadata AND data reservations
4718  * anymore.  So if there is an error or we insert an inline extent.
4719  *
4720  * This function will release the metadata space that was not used and will
4721  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4722  * list if there are no delalloc bytes left.
4723  */
4724 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4725 {
4726         btrfs_delalloc_release_metadata(inode, num_bytes);
4727         btrfs_free_reserved_data_space(inode, num_bytes);
4728 }
4729
4730 static int update_block_group(struct btrfs_trans_handle *trans,
4731                               struct btrfs_root *root,
4732                               u64 bytenr, u64 num_bytes, int alloc)
4733 {
4734         struct btrfs_block_group_cache *cache = NULL;
4735         struct btrfs_fs_info *info = root->fs_info;
4736         u64 total = num_bytes;
4737         u64 old_val;
4738         u64 byte_in_group;
4739         int factor;
4740
4741         /* block accounting for super block */
4742         spin_lock(&info->delalloc_lock);
4743         old_val = btrfs_super_bytes_used(info->super_copy);
4744         if (alloc)
4745                 old_val += num_bytes;
4746         else
4747                 old_val -= num_bytes;
4748         btrfs_set_super_bytes_used(info->super_copy, old_val);
4749         spin_unlock(&info->delalloc_lock);
4750
4751         while (total) {
4752                 cache = btrfs_lookup_block_group(info, bytenr);
4753                 if (!cache)
4754                         return -ENOENT;
4755                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4756                                     BTRFS_BLOCK_GROUP_RAID1 |
4757                                     BTRFS_BLOCK_GROUP_RAID10))
4758                         factor = 2;
4759                 else
4760                         factor = 1;
4761                 /*
4762                  * If this block group has free space cache written out, we
4763                  * need to make sure to load it if we are removing space.  This
4764                  * is because we need the unpinning stage to actually add the
4765                  * space back to the block group, otherwise we will leak space.
4766                  */
4767                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4768                         cache_block_group(cache, trans, NULL, 1);
4769
4770                 byte_in_group = bytenr - cache->key.objectid;
4771                 WARN_ON(byte_in_group > cache->key.offset);
4772
4773                 spin_lock(&cache->space_info->lock);
4774                 spin_lock(&cache->lock);
4775
4776                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4777                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4778                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4779
4780                 cache->dirty = 1;
4781                 old_val = btrfs_block_group_used(&cache->item);
4782                 num_bytes = min(total, cache->key.offset - byte_in_group);
4783                 if (alloc) {
4784                         old_val += num_bytes;
4785                         btrfs_set_block_group_used(&cache->item, old_val);
4786                         cache->reserved -= num_bytes;
4787                         cache->space_info->bytes_reserved -= num_bytes;
4788                         cache->space_info->bytes_used += num_bytes;
4789                         cache->space_info->disk_used += num_bytes * factor;
4790                         spin_unlock(&cache->lock);
4791                         spin_unlock(&cache->space_info->lock);
4792                 } else {
4793                         old_val -= num_bytes;
4794                         btrfs_set_block_group_used(&cache->item, old_val);
4795                         cache->pinned += num_bytes;
4796                         cache->space_info->bytes_pinned += num_bytes;
4797                         cache->space_info->bytes_used -= num_bytes;
4798                         cache->space_info->disk_used -= num_bytes * factor;
4799                         spin_unlock(&cache->lock);
4800                         spin_unlock(&cache->space_info->lock);
4801
4802                         set_extent_dirty(info->pinned_extents,
4803                                          bytenr, bytenr + num_bytes - 1,
4804                                          GFP_NOFS | __GFP_NOFAIL);
4805                 }
4806                 btrfs_put_block_group(cache);
4807                 total -= num_bytes;
4808                 bytenr += num_bytes;
4809         }
4810         return 0;
4811 }
4812
4813 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4814 {
4815         struct btrfs_block_group_cache *cache;
4816         u64 bytenr;
4817
4818         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4819         if (!cache)
4820                 return 0;
4821
4822         bytenr = cache->key.objectid;
4823         btrfs_put_block_group(cache);
4824
4825         return bytenr;
4826 }
4827
4828 static int pin_down_extent(struct btrfs_root *root,
4829                            struct btrfs_block_group_cache *cache,
4830                            u64 bytenr, u64 num_bytes, int reserved)
4831 {
4832         spin_lock(&cache->space_info->lock);
4833         spin_lock(&cache->lock);
4834         cache->pinned += num_bytes;
4835         cache->space_info->bytes_pinned += num_bytes;
4836         if (reserved) {
4837                 cache->reserved -= num_bytes;
4838                 cache->space_info->bytes_reserved -= num_bytes;
4839         }
4840         spin_unlock(&cache->lock);
4841         spin_unlock(&cache->space_info->lock);
4842
4843         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4844                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4845         return 0;
4846 }
4847
4848 /*
4849  * this function must be called within transaction
4850  */
4851 int btrfs_pin_extent(struct btrfs_root *root,
4852                      u64 bytenr, u64 num_bytes, int reserved)
4853 {
4854         struct btrfs_block_group_cache *cache;
4855
4856         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4857         BUG_ON(!cache); /* Logic error */
4858
4859         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4860
4861         btrfs_put_block_group(cache);
4862         return 0;
4863 }
4864
4865 /*
4866  * this function must be called within transaction
4867  */
4868 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4869                                     struct btrfs_root *root,
4870                                     u64 bytenr, u64 num_bytes)
4871 {
4872         struct btrfs_block_group_cache *cache;
4873
4874         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4875         BUG_ON(!cache); /* Logic error */
4876
4877         /*
4878          * pull in the free space cache (if any) so that our pin
4879          * removes the free space from the cache.  We have load_only set
4880          * to one because the slow code to read in the free extents does check
4881          * the pinned extents.
4882          */
4883         cache_block_group(cache, trans, root, 1);
4884
4885         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4886
4887         /* remove us from the free space cache (if we're there at all) */
4888         btrfs_remove_free_space(cache, bytenr, num_bytes);
4889         btrfs_put_block_group(cache);
4890         return 0;
4891 }
4892
4893 /**
4894  * btrfs_update_reserved_bytes - update the block_group and space info counters
4895  * @cache:      The cache we are manipulating
4896  * @num_bytes:  The number of bytes in question
4897  * @reserve:    One of the reservation enums
4898  *
4899  * This is called by the allocator when it reserves space, or by somebody who is
4900  * freeing space that was never actually used on disk.  For example if you
4901  * reserve some space for a new leaf in transaction A and before transaction A
4902  * commits you free that leaf, you call this with reserve set to 0 in order to
4903  * clear the reservation.
4904  *
4905  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4906  * ENOSPC accounting.  For data we handle the reservation through clearing the
4907  * delalloc bits in the io_tree.  We have to do this since we could end up
4908  * allocating less disk space for the amount of data we have reserved in the
4909  * case of compression.
4910  *
4911  * If this is a reservation and the block group has become read only we cannot
4912  * make the reservation and return -EAGAIN, otherwise this function always
4913  * succeeds.
4914  */
4915 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4916                                        u64 num_bytes, int reserve)
4917 {
4918         struct btrfs_space_info *space_info = cache->space_info;
4919         int ret = 0;
4920
4921         spin_lock(&space_info->lock);
4922         spin_lock(&cache->lock);
4923         if (reserve != RESERVE_FREE) {
4924                 if (cache->ro) {
4925                         ret = -EAGAIN;
4926                 } else {
4927                         cache->reserved += num_bytes;
4928                         space_info->bytes_reserved += num_bytes;
4929                         if (reserve == RESERVE_ALLOC) {
4930                                 trace_btrfs_space_reservation(cache->fs_info,
4931                                                 "space_info", space_info->flags,
4932                                                 num_bytes, 0);
4933                                 space_info->bytes_may_use -= num_bytes;
4934                         }
4935                 }
4936         } else {
4937                 if (cache->ro)
4938                         space_info->bytes_readonly += num_bytes;
4939                 cache->reserved -= num_bytes;
4940                 space_info->bytes_reserved -= num_bytes;
4941                 space_info->reservation_progress++;
4942         }
4943         spin_unlock(&cache->lock);
4944         spin_unlock(&space_info->lock);
4945         return ret;
4946 }
4947
4948 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4949                                 struct btrfs_root *root)
4950 {
4951         struct btrfs_fs_info *fs_info = root->fs_info;
4952         struct btrfs_caching_control *next;
4953         struct btrfs_caching_control *caching_ctl;
4954         struct btrfs_block_group_cache *cache;
4955
4956         down_write(&fs_info->extent_commit_sem);
4957
4958         list_for_each_entry_safe(caching_ctl, next,
4959                                  &fs_info->caching_block_groups, list) {
4960                 cache = caching_ctl->block_group;
4961                 if (block_group_cache_done(cache)) {
4962                         cache->last_byte_to_unpin = (u64)-1;
4963                         list_del_init(&caching_ctl->list);
4964                         put_caching_control(caching_ctl);
4965                 } else {
4966                         cache->last_byte_to_unpin = caching_ctl->progress;
4967                 }
4968         }
4969
4970         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4971                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4972         else
4973                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4974
4975         up_write(&fs_info->extent_commit_sem);
4976
4977         update_global_block_rsv(fs_info);
4978 }
4979
4980 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4981 {
4982         struct btrfs_fs_info *fs_info = root->fs_info;
4983         struct btrfs_block_group_cache *cache = NULL;
4984         u64 len;
4985
4986         while (start <= end) {
4987                 if (!cache ||
4988                     start >= cache->key.objectid + cache->key.offset) {
4989                         if (cache)
4990                                 btrfs_put_block_group(cache);
4991                         cache = btrfs_lookup_block_group(fs_info, start);
4992                         BUG_ON(!cache); /* Logic error */
4993                 }
4994
4995                 len = cache->key.objectid + cache->key.offset - start;
4996                 len = min(len, end + 1 - start);
4997
4998                 if (start < cache->last_byte_to_unpin) {
4999                         len = min(len, cache->last_byte_to_unpin - start);
5000                         btrfs_add_free_space(cache, start, len);
5001                 }
5002
5003                 start += len;
5004
5005                 spin_lock(&cache->space_info->lock);
5006                 spin_lock(&cache->lock);
5007                 cache->pinned -= len;
5008                 cache->space_info->bytes_pinned -= len;
5009                 if (cache->ro)
5010                         cache->space_info->bytes_readonly += len;
5011                 spin_unlock(&cache->lock);
5012                 spin_unlock(&cache->space_info->lock);
5013         }
5014
5015         if (cache)
5016                 btrfs_put_block_group(cache);
5017         return 0;
5018 }
5019
5020 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5021                                struct btrfs_root *root)
5022 {
5023         struct btrfs_fs_info *fs_info = root->fs_info;
5024         struct extent_io_tree *unpin;
5025         u64 start;
5026         u64 end;
5027         int ret;
5028
5029         if (trans->aborted)
5030                 return 0;
5031
5032         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5033                 unpin = &fs_info->freed_extents[1];
5034         else
5035                 unpin = &fs_info->freed_extents[0];
5036
5037         while (1) {
5038                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5039                                             EXTENT_DIRTY);
5040                 if (ret)
5041                         break;
5042
5043                 if (btrfs_test_opt(root, DISCARD))
5044                         ret = btrfs_discard_extent(root, start,
5045                                                    end + 1 - start, NULL);
5046
5047                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5048                 unpin_extent_range(root, start, end);
5049                 cond_resched();
5050         }
5051
5052         return 0;
5053 }
5054
5055 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5056                                 struct btrfs_root *root,
5057                                 u64 bytenr, u64 num_bytes, u64 parent,
5058                                 u64 root_objectid, u64 owner_objectid,
5059                                 u64 owner_offset, int refs_to_drop,
5060                                 struct btrfs_delayed_extent_op *extent_op)
5061 {
5062         struct btrfs_key key;
5063         struct btrfs_path *path;
5064         struct btrfs_fs_info *info = root->fs_info;
5065         struct btrfs_root *extent_root = info->extent_root;
5066         struct extent_buffer *leaf;
5067         struct btrfs_extent_item *ei;
5068         struct btrfs_extent_inline_ref *iref;
5069         int ret;
5070         int is_data;
5071         int extent_slot = 0;
5072         int found_extent = 0;
5073         int num_to_del = 1;
5074         u32 item_size;
5075         u64 refs;
5076
5077         path = btrfs_alloc_path();
5078         if (!path)
5079                 return -ENOMEM;
5080
5081         path->reada = 1;
5082         path->leave_spinning = 1;
5083
5084         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5085         BUG_ON(!is_data && refs_to_drop != 1);
5086
5087         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5088                                     bytenr, num_bytes, parent,
5089                                     root_objectid, owner_objectid,
5090                                     owner_offset);
5091         if (ret == 0) {
5092                 extent_slot = path->slots[0];
5093                 while (extent_slot >= 0) {
5094                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5095                                               extent_slot);
5096                         if (key.objectid != bytenr)
5097                                 break;
5098                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5099                             key.offset == num_bytes) {
5100                                 found_extent = 1;
5101                                 break;
5102                         }
5103                         if (path->slots[0] - extent_slot > 5)
5104                                 break;
5105                         extent_slot--;
5106                 }
5107 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5108                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5109                 if (found_extent && item_size < sizeof(*ei))
5110                         found_extent = 0;
5111 #endif
5112                 if (!found_extent) {
5113                         BUG_ON(iref);
5114                         ret = remove_extent_backref(trans, extent_root, path,
5115                                                     NULL, refs_to_drop,
5116                                                     is_data);
5117                         if (ret)
5118                                 goto abort;
5119                         btrfs_release_path(path);
5120                         path->leave_spinning = 1;
5121
5122                         key.objectid = bytenr;
5123                         key.type = BTRFS_EXTENT_ITEM_KEY;
5124                         key.offset = num_bytes;
5125
5126                         ret = btrfs_search_slot(trans, extent_root,
5127                                                 &key, path, -1, 1);
5128                         if (ret) {
5129                                 printk(KERN_ERR "umm, got %d back from search"
5130                                        ", was looking for %llu\n", ret,
5131                                        (unsigned long long)bytenr);
5132                                 if (ret > 0)
5133                                         btrfs_print_leaf(extent_root,
5134                                                          path->nodes[0]);
5135                         }
5136                         if (ret < 0)
5137                                 goto abort;
5138                         extent_slot = path->slots[0];
5139                 }
5140         } else if (ret == -ENOENT) {
5141                 btrfs_print_leaf(extent_root, path->nodes[0]);
5142                 WARN_ON(1);
5143                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
5144                        "parent %llu root %llu  owner %llu offset %llu\n",
5145                        (unsigned long long)bytenr,
5146                        (unsigned long long)parent,
5147                        (unsigned long long)root_objectid,
5148                        (unsigned long long)owner_objectid,
5149                        (unsigned long long)owner_offset);
5150         } else {
5151                 goto abort;
5152         }
5153
5154         leaf = path->nodes[0];
5155         item_size = btrfs_item_size_nr(leaf, extent_slot);
5156 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5157         if (item_size < sizeof(*ei)) {
5158                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5159                 ret = convert_extent_item_v0(trans, extent_root, path,
5160                                              owner_objectid, 0);
5161                 if (ret < 0)
5162                         goto abort;
5163
5164                 btrfs_release_path(path);
5165                 path->leave_spinning = 1;
5166
5167                 key.objectid = bytenr;
5168                 key.type = BTRFS_EXTENT_ITEM_KEY;
5169                 key.offset = num_bytes;
5170
5171                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5172                                         -1, 1);
5173                 if (ret) {
5174                         printk(KERN_ERR "umm, got %d back from search"
5175                                ", was looking for %llu\n", ret,
5176                                (unsigned long long)bytenr);
5177                         btrfs_print_leaf(extent_root, path->nodes[0]);
5178                 }
5179                 if (ret < 0)
5180                         goto abort;
5181                 extent_slot = path->slots[0];
5182                 leaf = path->nodes[0];
5183                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5184         }
5185 #endif
5186         BUG_ON(item_size < sizeof(*ei));
5187         ei = btrfs_item_ptr(leaf, extent_slot,
5188                             struct btrfs_extent_item);
5189         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5190                 struct btrfs_tree_block_info *bi;
5191                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5192                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5193                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5194         }
5195
5196         refs = btrfs_extent_refs(leaf, ei);
5197         BUG_ON(refs < refs_to_drop);
5198         refs -= refs_to_drop;
5199
5200         if (refs > 0) {
5201                 if (extent_op)
5202                         __run_delayed_extent_op(extent_op, leaf, ei);
5203                 /*
5204                  * In the case of inline back ref, reference count will
5205                  * be updated by remove_extent_backref
5206                  */
5207                 if (iref) {
5208                         BUG_ON(!found_extent);
5209                 } else {
5210                         btrfs_set_extent_refs(leaf, ei, refs);
5211                         btrfs_mark_buffer_dirty(leaf);
5212                 }
5213                 if (found_extent) {
5214                         ret = remove_extent_backref(trans, extent_root, path,
5215                                                     iref, refs_to_drop,
5216                                                     is_data);
5217                         if (ret)
5218                                 goto abort;
5219                 }
5220         } else {
5221                 if (found_extent) {
5222                         BUG_ON(is_data && refs_to_drop !=
5223                                extent_data_ref_count(root, path, iref));
5224                         if (iref) {
5225                                 BUG_ON(path->slots[0] != extent_slot);
5226                         } else {
5227                                 BUG_ON(path->slots[0] != extent_slot + 1);
5228                                 path->slots[0] = extent_slot;
5229                                 num_to_del = 2;
5230                         }
5231                 }
5232
5233                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5234                                       num_to_del);
5235                 if (ret)
5236                         goto abort;
5237                 btrfs_release_path(path);
5238
5239                 if (is_data) {
5240                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5241                         if (ret)
5242                                 goto abort;
5243                 }
5244
5245                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
5246                 if (ret)
5247                         goto abort;
5248         }
5249 out:
5250         btrfs_free_path(path);
5251         return ret;
5252
5253 abort:
5254         btrfs_abort_transaction(trans, extent_root, ret);
5255         goto out;
5256 }
5257
5258 /*
5259  * when we free an block, it is possible (and likely) that we free the last
5260  * delayed ref for that extent as well.  This searches the delayed ref tree for
5261  * a given extent, and if there are no other delayed refs to be processed, it
5262  * removes it from the tree.
5263  */
5264 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5265                                       struct btrfs_root *root, u64 bytenr)
5266 {
5267         struct btrfs_delayed_ref_head *head;
5268         struct btrfs_delayed_ref_root *delayed_refs;
5269         struct btrfs_delayed_ref_node *ref;
5270         struct rb_node *node;
5271         int ret = 0;
5272
5273         delayed_refs = &trans->transaction->delayed_refs;
5274         spin_lock(&delayed_refs->lock);
5275         head = btrfs_find_delayed_ref_head(trans, bytenr);
5276         if (!head)
5277                 goto out;
5278
5279         node = rb_prev(&head->node.rb_node);
5280         if (!node)
5281                 goto out;
5282
5283         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5284
5285         /* there are still entries for this ref, we can't drop it */
5286         if (ref->bytenr == bytenr)
5287                 goto out;
5288
5289         if (head->extent_op) {
5290                 if (!head->must_insert_reserved)
5291                         goto out;
5292                 kfree(head->extent_op);
5293                 head->extent_op = NULL;
5294         }
5295
5296         /*
5297          * waiting for the lock here would deadlock.  If someone else has it
5298          * locked they are already in the process of dropping it anyway
5299          */
5300         if (!mutex_trylock(&head->mutex))
5301                 goto out;
5302
5303         /*
5304          * at this point we have a head with no other entries.  Go
5305          * ahead and process it.
5306          */
5307         head->node.in_tree = 0;
5308         rb_erase(&head->node.rb_node, &delayed_refs->root);
5309
5310         delayed_refs->num_entries--;
5311
5312         /*
5313          * we don't take a ref on the node because we're removing it from the
5314          * tree, so we just steal the ref the tree was holding.
5315          */
5316         delayed_refs->num_heads--;
5317         if (list_empty(&head->cluster))
5318                 delayed_refs->num_heads_ready--;
5319
5320         list_del_init(&head->cluster);
5321         spin_unlock(&delayed_refs->lock);
5322
5323         BUG_ON(head->extent_op);
5324         if (head->must_insert_reserved)
5325                 ret = 1;
5326
5327         mutex_unlock(&head->mutex);
5328         btrfs_put_delayed_ref(&head->node);
5329         return ret;
5330 out:
5331         spin_unlock(&delayed_refs->lock);
5332         return 0;
5333 }
5334
5335 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5336                            struct btrfs_root *root,
5337                            struct extent_buffer *buf,
5338                            u64 parent, int last_ref)
5339 {
5340         struct btrfs_block_group_cache *cache = NULL;
5341         int ret;
5342
5343         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5344                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5345                                         buf->start, buf->len,
5346                                         parent, root->root_key.objectid,
5347                                         btrfs_header_level(buf),
5348                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5349                 BUG_ON(ret); /* -ENOMEM */
5350         }
5351
5352         if (!last_ref)
5353                 return;
5354
5355         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5356
5357         if (btrfs_header_generation(buf) == trans->transid) {
5358                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5359                         ret = check_ref_cleanup(trans, root, buf->start);
5360                         if (!ret)
5361                                 goto out;
5362                 }
5363
5364                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5365                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5366                         goto out;
5367                 }
5368
5369                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5370
5371                 btrfs_add_free_space(cache, buf->start, buf->len);
5372                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5373         }
5374 out:
5375         /*
5376          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5377          * anymore.
5378          */
5379         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5380         btrfs_put_block_group(cache);
5381 }
5382
5383 /* Can return -ENOMEM */
5384 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5385                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5386                       u64 owner, u64 offset, int for_cow)
5387 {
5388         int ret;
5389         struct btrfs_fs_info *fs_info = root->fs_info;
5390
5391         /*
5392          * tree log blocks never actually go into the extent allocation
5393          * tree, just update pinning info and exit early.
5394          */
5395         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5396                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5397                 /* unlocks the pinned mutex */
5398                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5399                 ret = 0;
5400         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5401                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5402                                         num_bytes,
5403                                         parent, root_objectid, (int)owner,
5404                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5405         } else {
5406                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5407                                                 num_bytes,
5408                                                 parent, root_objectid, owner,
5409                                                 offset, BTRFS_DROP_DELAYED_REF,
5410                                                 NULL, for_cow);
5411         }
5412         return ret;
5413 }
5414
5415 static u64 stripe_align(struct btrfs_root *root, u64 val)
5416 {
5417         u64 mask = ((u64)root->stripesize - 1);
5418         u64 ret = (val + mask) & ~mask;
5419         return ret;
5420 }
5421
5422 /*
5423  * when we wait for progress in the block group caching, its because
5424  * our allocation attempt failed at least once.  So, we must sleep
5425  * and let some progress happen before we try again.
5426  *
5427  * This function will sleep at least once waiting for new free space to
5428  * show up, and then it will check the block group free space numbers
5429  * for our min num_bytes.  Another option is to have it go ahead
5430  * and look in the rbtree for a free extent of a given size, but this
5431  * is a good start.
5432  */
5433 static noinline int
5434 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5435                                 u64 num_bytes)
5436 {
5437         struct btrfs_caching_control *caching_ctl;
5438         DEFINE_WAIT(wait);
5439
5440         caching_ctl = get_caching_control(cache);
5441         if (!caching_ctl)
5442                 return 0;
5443
5444         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5445                    (cache->free_space_ctl->free_space >= num_bytes));
5446
5447         put_caching_control(caching_ctl);
5448         return 0;
5449 }
5450
5451 static noinline int
5452 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5453 {
5454         struct btrfs_caching_control *caching_ctl;
5455         DEFINE_WAIT(wait);
5456
5457         caching_ctl = get_caching_control(cache);
5458         if (!caching_ctl)
5459                 return 0;
5460
5461         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5462
5463         put_caching_control(caching_ctl);
5464         return 0;
5465 }
5466
5467 static int __get_block_group_index(u64 flags)
5468 {
5469         int index;
5470
5471         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5472                 index = 0;
5473         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5474                 index = 1;
5475         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5476                 index = 2;
5477         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5478                 index = 3;
5479         else
5480                 index = 4;
5481
5482         return index;
5483 }
5484
5485 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5486 {
5487         return __get_block_group_index(cache->flags);
5488 }
5489
5490 enum btrfs_loop_type {
5491         LOOP_CACHING_NOWAIT = 0,
5492         LOOP_CACHING_WAIT = 1,
5493         LOOP_ALLOC_CHUNK = 2,
5494         LOOP_NO_EMPTY_SIZE = 3,
5495 };
5496
5497 /*
5498  * walks the btree of allocated extents and find a hole of a given size.
5499  * The key ins is changed to record the hole:
5500  * ins->objectid == block start
5501  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5502  * ins->offset == number of blocks
5503  * Any available blocks before search_start are skipped.
5504  */
5505 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5506                                      struct btrfs_root *orig_root,
5507                                      u64 num_bytes, u64 empty_size,
5508                                      u64 hint_byte, struct btrfs_key *ins,
5509                                      u64 data)
5510 {
5511         int ret = 0;
5512         struct btrfs_root *root = orig_root->fs_info->extent_root;
5513         struct btrfs_free_cluster *last_ptr = NULL;
5514         struct btrfs_block_group_cache *block_group = NULL;
5515         struct btrfs_block_group_cache *used_block_group;
5516         u64 search_start = 0;
5517         int empty_cluster = 2 * 1024 * 1024;
5518         int allowed_chunk_alloc = 0;
5519         int done_chunk_alloc = 0;
5520         struct btrfs_space_info *space_info;
5521         int loop = 0;
5522         int index = 0;
5523         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5524                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5525         bool found_uncached_bg = false;
5526         bool failed_cluster_refill = false;
5527         bool failed_alloc = false;
5528         bool use_cluster = true;
5529         bool have_caching_bg = false;
5530
5531         WARN_ON(num_bytes < root->sectorsize);
5532         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5533         ins->objectid = 0;
5534         ins->offset = 0;
5535
5536         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5537
5538         space_info = __find_space_info(root->fs_info, data);
5539         if (!space_info) {
5540                 printk(KERN_ERR "No space info for %llu\n", data);
5541                 return -ENOSPC;
5542         }
5543
5544         /*
5545          * If the space info is for both data and metadata it means we have a
5546          * small filesystem and we can't use the clustering stuff.
5547          */
5548         if (btrfs_mixed_space_info(space_info))
5549                 use_cluster = false;
5550
5551         if (orig_root->ref_cows || empty_size)
5552                 allowed_chunk_alloc = 1;
5553
5554         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5555                 last_ptr = &root->fs_info->meta_alloc_cluster;
5556                 if (!btrfs_test_opt(root, SSD))
5557                         empty_cluster = 64 * 1024;
5558         }
5559
5560         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5561             btrfs_test_opt(root, SSD)) {
5562                 last_ptr = &root->fs_info->data_alloc_cluster;
5563         }
5564
5565         if (last_ptr) {
5566                 spin_lock(&last_ptr->lock);
5567                 if (last_ptr->block_group)
5568                         hint_byte = last_ptr->window_start;
5569                 spin_unlock(&last_ptr->lock);
5570         }
5571
5572         search_start = max(search_start, first_logical_byte(root, 0));
5573         search_start = max(search_start, hint_byte);
5574
5575         if (!last_ptr)
5576                 empty_cluster = 0;
5577
5578         if (search_start == hint_byte) {
5579                 block_group = btrfs_lookup_block_group(root->fs_info,
5580                                                        search_start);
5581                 used_block_group = block_group;
5582                 /*
5583                  * we don't want to use the block group if it doesn't match our
5584                  * allocation bits, or if its not cached.
5585                  *
5586                  * However if we are re-searching with an ideal block group
5587                  * picked out then we don't care that the block group is cached.
5588                  */
5589                 if (block_group && block_group_bits(block_group, data) &&
5590                     block_group->cached != BTRFS_CACHE_NO) {
5591                         down_read(&space_info->groups_sem);
5592                         if (list_empty(&block_group->list) ||
5593                             block_group->ro) {
5594                                 /*
5595                                  * someone is removing this block group,
5596                                  * we can't jump into the have_block_group
5597                                  * target because our list pointers are not
5598                                  * valid
5599                                  */
5600                                 btrfs_put_block_group(block_group);
5601                                 up_read(&space_info->groups_sem);
5602                         } else {
5603                                 index = get_block_group_index(block_group);
5604                                 goto have_block_group;
5605                         }
5606                 } else if (block_group) {
5607                         btrfs_put_block_group(block_group);
5608                 }
5609         }
5610 search:
5611         have_caching_bg = false;
5612         down_read(&space_info->groups_sem);
5613         list_for_each_entry(block_group, &space_info->block_groups[index],
5614                             list) {
5615                 u64 offset;
5616                 int cached;
5617
5618                 used_block_group = block_group;
5619                 btrfs_get_block_group(block_group);
5620                 search_start = block_group->key.objectid;
5621
5622                 /*
5623                  * this can happen if we end up cycling through all the
5624                  * raid types, but we want to make sure we only allocate
5625                  * for the proper type.
5626                  */
5627                 if (!block_group_bits(block_group, data)) {
5628                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5629                                 BTRFS_BLOCK_GROUP_RAID1 |
5630                                 BTRFS_BLOCK_GROUP_RAID10;
5631
5632                         /*
5633                          * if they asked for extra copies and this block group
5634                          * doesn't provide them, bail.  This does allow us to
5635                          * fill raid0 from raid1.
5636                          */
5637                         if ((data & extra) && !(block_group->flags & extra))
5638                                 goto loop;
5639                 }
5640
5641 have_block_group:
5642                 cached = block_group_cache_done(block_group);
5643                 if (unlikely(!cached)) {
5644                         found_uncached_bg = true;
5645                         ret = cache_block_group(block_group, trans,
5646                                                 orig_root, 0);
5647                         BUG_ON(ret < 0);
5648                         ret = 0;
5649                 }
5650
5651                 if (unlikely(block_group->ro))
5652                         goto loop;
5653
5654                 /*
5655                  * Ok we want to try and use the cluster allocator, so
5656                  * lets look there
5657                  */
5658                 if (last_ptr) {
5659                         /*
5660                          * the refill lock keeps out other
5661                          * people trying to start a new cluster
5662                          */
5663                         spin_lock(&last_ptr->refill_lock);
5664                         used_block_group = last_ptr->block_group;
5665                         if (used_block_group != block_group &&
5666                             (!used_block_group ||
5667                              used_block_group->ro ||
5668                              !block_group_bits(used_block_group, data))) {
5669                                 used_block_group = block_group;
5670                                 goto refill_cluster;
5671                         }
5672
5673                         if (used_block_group != block_group)
5674                                 btrfs_get_block_group(used_block_group);
5675
5676                         offset = btrfs_alloc_from_cluster(used_block_group,
5677                           last_ptr, num_bytes, used_block_group->key.objectid);
5678                         if (offset) {
5679                                 /* we have a block, we're done */
5680                                 spin_unlock(&last_ptr->refill_lock);
5681                                 trace_btrfs_reserve_extent_cluster(root,
5682                                         block_group, search_start, num_bytes);
5683                                 goto checks;
5684                         }
5685
5686                         WARN_ON(last_ptr->block_group != used_block_group);
5687                         if (used_block_group != block_group) {
5688                                 btrfs_put_block_group(used_block_group);
5689                                 used_block_group = block_group;
5690                         }
5691 refill_cluster:
5692                         BUG_ON(used_block_group != block_group);
5693                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5694                          * set up a new clusters, so lets just skip it
5695                          * and let the allocator find whatever block
5696                          * it can find.  If we reach this point, we
5697                          * will have tried the cluster allocator
5698                          * plenty of times and not have found
5699                          * anything, so we are likely way too
5700                          * fragmented for the clustering stuff to find
5701                          * anything.
5702                          *
5703                          * However, if the cluster is taken from the
5704                          * current block group, release the cluster
5705                          * first, so that we stand a better chance of
5706                          * succeeding in the unclustered
5707                          * allocation.  */
5708                         if (loop >= LOOP_NO_EMPTY_SIZE &&
5709                             last_ptr->block_group != block_group) {
5710                                 spin_unlock(&last_ptr->refill_lock);
5711                                 goto unclustered_alloc;
5712                         }
5713
5714                         /*
5715                          * this cluster didn't work out, free it and
5716                          * start over
5717                          */
5718                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5719
5720                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5721                                 spin_unlock(&last_ptr->refill_lock);
5722                                 goto unclustered_alloc;
5723                         }
5724
5725                         /* allocate a cluster in this block group */
5726                         ret = btrfs_find_space_cluster(trans, root,
5727                                                block_group, last_ptr,
5728                                                search_start, num_bytes,
5729                                                empty_cluster + empty_size);
5730                         if (ret == 0) {
5731                                 /*
5732                                  * now pull our allocation out of this
5733                                  * cluster
5734                                  */
5735                                 offset = btrfs_alloc_from_cluster(block_group,
5736                                                   last_ptr, num_bytes,
5737                                                   search_start);
5738                                 if (offset) {
5739                                         /* we found one, proceed */
5740                                         spin_unlock(&last_ptr->refill_lock);
5741                                         trace_btrfs_reserve_extent_cluster(root,
5742                                                 block_group, search_start,
5743                                                 num_bytes);
5744                                         goto checks;
5745                                 }
5746                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5747                                    && !failed_cluster_refill) {
5748                                 spin_unlock(&last_ptr->refill_lock);
5749
5750                                 failed_cluster_refill = true;
5751                                 wait_block_group_cache_progress(block_group,
5752                                        num_bytes + empty_cluster + empty_size);
5753                                 goto have_block_group;
5754                         }
5755
5756                         /*
5757                          * at this point we either didn't find a cluster
5758                          * or we weren't able to allocate a block from our
5759                          * cluster.  Free the cluster we've been trying
5760                          * to use, and go to the next block group
5761                          */
5762                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5763                         spin_unlock(&last_ptr->refill_lock);
5764                         goto loop;
5765                 }
5766
5767 unclustered_alloc:
5768                 spin_lock(&block_group->free_space_ctl->tree_lock);
5769                 if (cached &&
5770                     block_group->free_space_ctl->free_space <
5771                     num_bytes + empty_cluster + empty_size) {
5772                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5773                         goto loop;
5774                 }
5775                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5776
5777                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5778                                                     num_bytes, empty_size);
5779                 /*
5780                  * If we didn't find a chunk, and we haven't failed on this
5781                  * block group before, and this block group is in the middle of
5782                  * caching and we are ok with waiting, then go ahead and wait
5783                  * for progress to be made, and set failed_alloc to true.
5784                  *
5785                  * If failed_alloc is true then we've already waited on this
5786                  * block group once and should move on to the next block group.
5787                  */
5788                 if (!offset && !failed_alloc && !cached &&
5789                     loop > LOOP_CACHING_NOWAIT) {
5790                         wait_block_group_cache_progress(block_group,
5791                                                 num_bytes + empty_size);
5792                         failed_alloc = true;
5793                         goto have_block_group;
5794                 } else if (!offset) {
5795                         if (!cached)
5796                                 have_caching_bg = true;
5797                         goto loop;
5798                 }
5799 checks:
5800                 search_start = stripe_align(root, offset);
5801
5802                 /* move on to the next group */
5803                 if (search_start + num_bytes >
5804                     used_block_group->key.objectid + used_block_group->key.offset) {
5805                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5806                         goto loop;
5807                 }
5808
5809                 if (offset < search_start)
5810                         btrfs_add_free_space(used_block_group, offset,
5811                                              search_start - offset);
5812                 BUG_ON(offset > search_start);
5813
5814                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
5815                                                   alloc_type);
5816                 if (ret == -EAGAIN) {
5817                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5818                         goto loop;
5819                 }
5820
5821                 /* we are all good, lets return */
5822                 ins->objectid = search_start;
5823                 ins->offset = num_bytes;
5824
5825                 trace_btrfs_reserve_extent(orig_root, block_group,
5826                                            search_start, num_bytes);
5827                 if (offset < search_start)
5828                         btrfs_add_free_space(used_block_group, offset,
5829                                              search_start - offset);
5830                 BUG_ON(offset > search_start);
5831                 if (used_block_group != block_group)
5832                         btrfs_put_block_group(used_block_group);
5833                 btrfs_put_block_group(block_group);
5834                 break;
5835 loop:
5836                 failed_cluster_refill = false;
5837                 failed_alloc = false;
5838                 BUG_ON(index != get_block_group_index(block_group));
5839                 if (used_block_group != block_group)
5840                         btrfs_put_block_group(used_block_group);
5841                 btrfs_put_block_group(block_group);
5842         }
5843         up_read(&space_info->groups_sem);
5844
5845         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
5846                 goto search;
5847
5848         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5849                 goto search;
5850
5851         /*
5852          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5853          *                      caching kthreads as we move along
5854          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5855          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5856          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5857          *                      again
5858          */
5859         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5860                 index = 0;
5861                 loop++;
5862                 if (loop == LOOP_ALLOC_CHUNK) {
5863                        if (allowed_chunk_alloc) {
5864                                 ret = do_chunk_alloc(trans, root, num_bytes +
5865                                                      2 * 1024 * 1024, data,
5866                                                      CHUNK_ALLOC_LIMITED);
5867                                 /*
5868                                  * Do not bail out on ENOSPC since we
5869                                  * can do more things.
5870                                  */
5871                                 if (ret < 0 && ret != -ENOSPC) {
5872                                         btrfs_abort_transaction(trans,
5873                                                                 root, ret);
5874                                         goto out;
5875                                 }
5876                                 allowed_chunk_alloc = 0;
5877                                 if (ret == 1)
5878                                         done_chunk_alloc = 1;
5879                         } else if (!done_chunk_alloc &&
5880                                    space_info->force_alloc ==
5881                                    CHUNK_ALLOC_NO_FORCE) {
5882                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5883                         }
5884
5885                        /*
5886                         * We didn't allocate a chunk, go ahead and drop the
5887                         * empty size and loop again.
5888                         */
5889                        if (!done_chunk_alloc)
5890                                loop = LOOP_NO_EMPTY_SIZE;
5891                 }
5892
5893                 if (loop == LOOP_NO_EMPTY_SIZE) {
5894                         empty_size = 0;
5895                         empty_cluster = 0;
5896                 }
5897
5898                 goto search;
5899         } else if (!ins->objectid) {
5900                 ret = -ENOSPC;
5901         } else if (ins->objectid) {
5902                 ret = 0;
5903         }
5904 out:
5905
5906         return ret;
5907 }
5908
5909 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5910                             int dump_block_groups)
5911 {
5912         struct btrfs_block_group_cache *cache;
5913         int index = 0;
5914
5915         spin_lock(&info->lock);
5916         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5917                (unsigned long long)info->flags,
5918                (unsigned long long)(info->total_bytes - info->bytes_used -
5919                                     info->bytes_pinned - info->bytes_reserved -
5920                                     info->bytes_readonly),
5921                (info->full) ? "" : "not ");
5922         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5923                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5924                (unsigned long long)info->total_bytes,
5925                (unsigned long long)info->bytes_used,
5926                (unsigned long long)info->bytes_pinned,
5927                (unsigned long long)info->bytes_reserved,
5928                (unsigned long long)info->bytes_may_use,
5929                (unsigned long long)info->bytes_readonly);
5930         spin_unlock(&info->lock);
5931
5932         if (!dump_block_groups)
5933                 return;
5934
5935         down_read(&info->groups_sem);
5936 again:
5937         list_for_each_entry(cache, &info->block_groups[index], list) {
5938                 spin_lock(&cache->lock);
5939                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s\n",
5940                        (unsigned long long)cache->key.objectid,
5941                        (unsigned long long)cache->key.offset,
5942                        (unsigned long long)btrfs_block_group_used(&cache->item),
5943                        (unsigned long long)cache->pinned,
5944                        (unsigned long long)cache->reserved,
5945                        cache->ro ? "[readonly]" : "");
5946                 btrfs_dump_free_space(cache, bytes);
5947                 spin_unlock(&cache->lock);
5948         }
5949         if (++index < BTRFS_NR_RAID_TYPES)
5950                 goto again;
5951         up_read(&info->groups_sem);
5952 }
5953
5954 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5955                          struct btrfs_root *root,
5956                          u64 num_bytes, u64 min_alloc_size,
5957                          u64 empty_size, u64 hint_byte,
5958                          struct btrfs_key *ins, u64 data)
5959 {
5960         bool final_tried = false;
5961         int ret;
5962
5963         data = btrfs_get_alloc_profile(root, data);
5964 again:
5965         /*
5966          * the only place that sets empty_size is btrfs_realloc_node, which
5967          * is not called recursively on allocations
5968          */
5969         if (empty_size || root->ref_cows) {
5970                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5971                                      num_bytes + 2 * 1024 * 1024, data,
5972                                      CHUNK_ALLOC_NO_FORCE);
5973                 if (ret < 0 && ret != -ENOSPC) {
5974                         btrfs_abort_transaction(trans, root, ret);
5975                         return ret;
5976                 }
5977         }
5978
5979         WARN_ON(num_bytes < root->sectorsize);
5980         ret = find_free_extent(trans, root, num_bytes, empty_size,
5981                                hint_byte, ins, data);
5982
5983         if (ret == -ENOSPC) {
5984                 if (!final_tried) {
5985                         num_bytes = num_bytes >> 1;
5986                         num_bytes = num_bytes & ~(root->sectorsize - 1);
5987                         num_bytes = max(num_bytes, min_alloc_size);
5988                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5989                                        num_bytes, data, CHUNK_ALLOC_FORCE);
5990                         if (ret < 0 && ret != -ENOSPC) {
5991                                 btrfs_abort_transaction(trans, root, ret);
5992                                 return ret;
5993                         }
5994                         if (num_bytes == min_alloc_size)
5995                                 final_tried = true;
5996                         goto again;
5997                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
5998                         struct btrfs_space_info *sinfo;
5999
6000                         sinfo = __find_space_info(root->fs_info, data);
6001                         printk(KERN_ERR "btrfs allocation failed flags %llu, "
6002                                "wanted %llu\n", (unsigned long long)data,
6003                                (unsigned long long)num_bytes);
6004                         if (sinfo)
6005                                 dump_space_info(sinfo, num_bytes, 1);
6006                 }
6007         }
6008
6009         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
6010
6011         return ret;
6012 }
6013
6014 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6015                                         u64 start, u64 len, int pin)
6016 {
6017         struct btrfs_block_group_cache *cache;
6018         int ret = 0;
6019
6020         cache = btrfs_lookup_block_group(root->fs_info, start);
6021         if (!cache) {
6022                 printk(KERN_ERR "Unable to find block group for %llu\n",
6023                        (unsigned long long)start);
6024                 return -ENOSPC;
6025         }
6026
6027         if (btrfs_test_opt(root, DISCARD))
6028                 ret = btrfs_discard_extent(root, start, len, NULL);
6029
6030         if (pin)
6031                 pin_down_extent(root, cache, start, len, 1);
6032         else {
6033                 btrfs_add_free_space(cache, start, len);
6034                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6035         }
6036         btrfs_put_block_group(cache);
6037
6038         trace_btrfs_reserved_extent_free(root, start, len);
6039
6040         return ret;
6041 }
6042
6043 int btrfs_free_reserved_extent(struct btrfs_root *root,
6044                                         u64 start, u64 len)
6045 {
6046         return __btrfs_free_reserved_extent(root, start, len, 0);
6047 }
6048
6049 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6050                                        u64 start, u64 len)
6051 {
6052         return __btrfs_free_reserved_extent(root, start, len, 1);
6053 }
6054
6055 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6056                                       struct btrfs_root *root,
6057                                       u64 parent, u64 root_objectid,
6058                                       u64 flags, u64 owner, u64 offset,
6059                                       struct btrfs_key *ins, int ref_mod)
6060 {
6061         int ret;
6062         struct btrfs_fs_info *fs_info = root->fs_info;
6063         struct btrfs_extent_item *extent_item;
6064         struct btrfs_extent_inline_ref *iref;
6065         struct btrfs_path *path;
6066         struct extent_buffer *leaf;
6067         int type;
6068         u32 size;
6069
6070         if (parent > 0)
6071                 type = BTRFS_SHARED_DATA_REF_KEY;
6072         else
6073                 type = BTRFS_EXTENT_DATA_REF_KEY;
6074
6075         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6076
6077         path = btrfs_alloc_path();
6078         if (!path)
6079                 return -ENOMEM;
6080
6081         path->leave_spinning = 1;
6082         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6083                                       ins, size);
6084         if (ret) {
6085                 btrfs_free_path(path);
6086                 return ret;
6087         }
6088
6089         leaf = path->nodes[0];
6090         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6091                                      struct btrfs_extent_item);
6092         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6093         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6094         btrfs_set_extent_flags(leaf, extent_item,
6095                                flags | BTRFS_EXTENT_FLAG_DATA);
6096
6097         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6098         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6099         if (parent > 0) {
6100                 struct btrfs_shared_data_ref *ref;
6101                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6102                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6103                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6104         } else {
6105                 struct btrfs_extent_data_ref *ref;
6106                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6107                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6108                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6109                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6110                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6111         }
6112
6113         btrfs_mark_buffer_dirty(path->nodes[0]);
6114         btrfs_free_path(path);
6115
6116         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
6117         if (ret) { /* -ENOENT, logic error */
6118                 printk(KERN_ERR "btrfs update block group failed for %llu "
6119                        "%llu\n", (unsigned long long)ins->objectid,
6120                        (unsigned long long)ins->offset);
6121                 BUG();
6122         }
6123         return ret;
6124 }
6125
6126 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6127                                      struct btrfs_root *root,
6128                                      u64 parent, u64 root_objectid,
6129                                      u64 flags, struct btrfs_disk_key *key,
6130                                      int level, struct btrfs_key *ins)
6131 {
6132         int ret;
6133         struct btrfs_fs_info *fs_info = root->fs_info;
6134         struct btrfs_extent_item *extent_item;
6135         struct btrfs_tree_block_info *block_info;
6136         struct btrfs_extent_inline_ref *iref;
6137         struct btrfs_path *path;
6138         struct extent_buffer *leaf;
6139         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
6140
6141         path = btrfs_alloc_path();
6142         if (!path)
6143                 return -ENOMEM;
6144
6145         path->leave_spinning = 1;
6146         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6147                                       ins, size);
6148         if (ret) {
6149                 btrfs_free_path(path);
6150                 return ret;
6151         }
6152
6153         leaf = path->nodes[0];
6154         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6155                                      struct btrfs_extent_item);
6156         btrfs_set_extent_refs(leaf, extent_item, 1);
6157         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6158         btrfs_set_extent_flags(leaf, extent_item,
6159                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6160         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6161
6162         btrfs_set_tree_block_key(leaf, block_info, key);
6163         btrfs_set_tree_block_level(leaf, block_info, level);
6164
6165         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6166         if (parent > 0) {
6167                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6168                 btrfs_set_extent_inline_ref_type(leaf, iref,
6169                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6170                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6171         } else {
6172                 btrfs_set_extent_inline_ref_type(leaf, iref,
6173                                                  BTRFS_TREE_BLOCK_REF_KEY);
6174                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6175         }
6176
6177         btrfs_mark_buffer_dirty(leaf);
6178         btrfs_free_path(path);
6179
6180         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
6181         if (ret) { /* -ENOENT, logic error */
6182                 printk(KERN_ERR "btrfs update block group failed for %llu "
6183                        "%llu\n", (unsigned long long)ins->objectid,
6184                        (unsigned long long)ins->offset);
6185                 BUG();
6186         }
6187         return ret;
6188 }
6189
6190 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6191                                      struct btrfs_root *root,
6192                                      u64 root_objectid, u64 owner,
6193                                      u64 offset, struct btrfs_key *ins)
6194 {
6195         int ret;
6196
6197         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6198
6199         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6200                                          ins->offset, 0,
6201                                          root_objectid, owner, offset,
6202                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6203         return ret;
6204 }
6205
6206 /*
6207  * this is used by the tree logging recovery code.  It records that
6208  * an extent has been allocated and makes sure to clear the free
6209  * space cache bits as well
6210  */
6211 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6212                                    struct btrfs_root *root,
6213                                    u64 root_objectid, u64 owner, u64 offset,
6214                                    struct btrfs_key *ins)
6215 {
6216         int ret;
6217         struct btrfs_block_group_cache *block_group;
6218         struct btrfs_caching_control *caching_ctl;
6219         u64 start = ins->objectid;
6220         u64 num_bytes = ins->offset;
6221
6222         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6223         cache_block_group(block_group, trans, NULL, 0);
6224         caching_ctl = get_caching_control(block_group);
6225
6226         if (!caching_ctl) {
6227                 BUG_ON(!block_group_cache_done(block_group));
6228                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6229                 BUG_ON(ret); /* -ENOMEM */
6230         } else {
6231                 mutex_lock(&caching_ctl->mutex);
6232
6233                 if (start >= caching_ctl->progress) {
6234                         ret = add_excluded_extent(root, start, num_bytes);
6235                         BUG_ON(ret); /* -ENOMEM */
6236                 } else if (start + num_bytes <= caching_ctl->progress) {
6237                         ret = btrfs_remove_free_space(block_group,
6238                                                       start, num_bytes);
6239                         BUG_ON(ret); /* -ENOMEM */
6240                 } else {
6241                         num_bytes = caching_ctl->progress - start;
6242                         ret = btrfs_remove_free_space(block_group,
6243                                                       start, num_bytes);
6244                         BUG_ON(ret); /* -ENOMEM */
6245
6246                         start = caching_ctl->progress;
6247                         num_bytes = ins->objectid + ins->offset -
6248                                     caching_ctl->progress;
6249                         ret = add_excluded_extent(root, start, num_bytes);
6250                         BUG_ON(ret); /* -ENOMEM */
6251                 }
6252
6253                 mutex_unlock(&caching_ctl->mutex);
6254                 put_caching_control(caching_ctl);
6255         }
6256
6257         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6258                                           RESERVE_ALLOC_NO_ACCOUNT);
6259         BUG_ON(ret); /* logic error */
6260         btrfs_put_block_group(block_group);
6261         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6262                                          0, owner, offset, ins, 1);
6263         return ret;
6264 }
6265
6266 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
6267                                             struct btrfs_root *root,
6268                                             u64 bytenr, u32 blocksize,
6269                                             int level)
6270 {
6271         struct extent_buffer *buf;
6272
6273         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6274         if (!buf)
6275                 return ERR_PTR(-ENOMEM);
6276         btrfs_set_header_generation(buf, trans->transid);
6277         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6278         btrfs_tree_lock(buf);
6279         clean_tree_block(trans, root, buf);
6280         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6281
6282         btrfs_set_lock_blocking(buf);
6283         btrfs_set_buffer_uptodate(buf);
6284
6285         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6286                 /*
6287                  * we allow two log transactions at a time, use different
6288                  * EXENT bit to differentiate dirty pages.
6289                  */
6290                 if (root->log_transid % 2 == 0)
6291                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6292                                         buf->start + buf->len - 1, GFP_NOFS);
6293                 else
6294                         set_extent_new(&root->dirty_log_pages, buf->start,
6295                                         buf->start + buf->len - 1, GFP_NOFS);
6296         } else {
6297                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6298                          buf->start + buf->len - 1, GFP_NOFS);
6299         }
6300         trans->blocks_used++;
6301         /* this returns a buffer locked for blocking */
6302         return buf;
6303 }
6304
6305 static struct btrfs_block_rsv *
6306 use_block_rsv(struct btrfs_trans_handle *trans,
6307               struct btrfs_root *root, u32 blocksize)
6308 {
6309         struct btrfs_block_rsv *block_rsv;
6310         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6311         int ret;
6312
6313         block_rsv = get_block_rsv(trans, root);
6314
6315         if (block_rsv->size == 0) {
6316                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
6317                 /*
6318                  * If we couldn't reserve metadata bytes try and use some from
6319                  * the global reserve.
6320                  */
6321                 if (ret && block_rsv != global_rsv) {
6322                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6323                         if (!ret)
6324                                 return global_rsv;
6325                         return ERR_PTR(ret);
6326                 } else if (ret) {
6327                         return ERR_PTR(ret);
6328                 }
6329                 return block_rsv;
6330         }
6331
6332         ret = block_rsv_use_bytes(block_rsv, blocksize);
6333         if (!ret)
6334                 return block_rsv;
6335         if (ret && !block_rsv->failfast) {
6336                 static DEFINE_RATELIMIT_STATE(_rs,
6337                                 DEFAULT_RATELIMIT_INTERVAL,
6338                                 /*DEFAULT_RATELIMIT_BURST*/ 2);
6339                 if (__ratelimit(&_rs)) {
6340                         printk(KERN_DEBUG "btrfs: block rsv returned %d\n", ret);
6341                         WARN_ON(1);
6342                 }
6343                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
6344                 if (!ret) {
6345                         return block_rsv;
6346                 } else if (ret && block_rsv != global_rsv) {
6347                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6348                         if (!ret)
6349                                 return global_rsv;
6350                 }
6351         }
6352
6353         return ERR_PTR(-ENOSPC);
6354 }
6355
6356 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6357                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6358 {
6359         block_rsv_add_bytes(block_rsv, blocksize, 0);
6360         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6361 }
6362
6363 /*
6364  * finds a free extent and does all the dirty work required for allocation
6365  * returns the key for the extent through ins, and a tree buffer for
6366  * the first block of the extent through buf.
6367  *
6368  * returns the tree buffer or NULL.
6369  */
6370 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6371                                         struct btrfs_root *root, u32 blocksize,
6372                                         u64 parent, u64 root_objectid,
6373                                         struct btrfs_disk_key *key, int level,
6374                                         u64 hint, u64 empty_size)
6375 {
6376         struct btrfs_key ins;
6377         struct btrfs_block_rsv *block_rsv;
6378         struct extent_buffer *buf;
6379         u64 flags = 0;
6380         int ret;
6381
6382
6383         block_rsv = use_block_rsv(trans, root, blocksize);
6384         if (IS_ERR(block_rsv))
6385                 return ERR_CAST(block_rsv);
6386
6387         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6388                                    empty_size, hint, &ins, 0);
6389         if (ret) {
6390                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6391                 return ERR_PTR(ret);
6392         }
6393
6394         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6395                                     blocksize, level);
6396         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6397
6398         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6399                 if (parent == 0)
6400                         parent = ins.objectid;
6401                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6402         } else
6403                 BUG_ON(parent > 0);
6404
6405         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6406                 struct btrfs_delayed_extent_op *extent_op;
6407                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
6408                 BUG_ON(!extent_op); /* -ENOMEM */
6409                 if (key)
6410                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6411                 else
6412                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6413                 extent_op->flags_to_set = flags;
6414                 extent_op->update_key = 1;
6415                 extent_op->update_flags = 1;
6416                 extent_op->is_data = 0;
6417
6418                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6419                                         ins.objectid,
6420                                         ins.offset, parent, root_objectid,
6421                                         level, BTRFS_ADD_DELAYED_EXTENT,
6422                                         extent_op, 0);
6423                 BUG_ON(ret); /* -ENOMEM */
6424         }
6425         return buf;
6426 }
6427
6428 struct walk_control {
6429         u64 refs[BTRFS_MAX_LEVEL];
6430         u64 flags[BTRFS_MAX_LEVEL];
6431         struct btrfs_key update_progress;
6432         int stage;
6433         int level;
6434         int shared_level;
6435         int update_ref;
6436         int keep_locks;
6437         int reada_slot;
6438         int reada_count;
6439         int for_reloc;
6440 };
6441
6442 #define DROP_REFERENCE  1
6443 #define UPDATE_BACKREF  2
6444
6445 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6446                                      struct btrfs_root *root,
6447                                      struct walk_control *wc,
6448                                      struct btrfs_path *path)
6449 {
6450         u64 bytenr;
6451         u64 generation;
6452         u64 refs;
6453         u64 flags;
6454         u32 nritems;
6455         u32 blocksize;
6456         struct btrfs_key key;
6457         struct extent_buffer *eb;
6458         int ret;
6459         int slot;
6460         int nread = 0;
6461
6462         if (path->slots[wc->level] < wc->reada_slot) {
6463                 wc->reada_count = wc->reada_count * 2 / 3;
6464                 wc->reada_count = max(wc->reada_count, 2);
6465         } else {
6466                 wc->reada_count = wc->reada_count * 3 / 2;
6467                 wc->reada_count = min_t(int, wc->reada_count,
6468                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6469         }
6470
6471         eb = path->nodes[wc->level];
6472         nritems = btrfs_header_nritems(eb);
6473         blocksize = btrfs_level_size(root, wc->level - 1);
6474
6475         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6476                 if (nread >= wc->reada_count)
6477                         break;
6478
6479                 cond_resched();
6480                 bytenr = btrfs_node_blockptr(eb, slot);
6481                 generation = btrfs_node_ptr_generation(eb, slot);
6482
6483                 if (slot == path->slots[wc->level])
6484                         goto reada;
6485
6486                 if (wc->stage == UPDATE_BACKREF &&
6487                     generation <= root->root_key.offset)
6488                         continue;
6489
6490                 /* We don't lock the tree block, it's OK to be racy here */
6491                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6492                                                &refs, &flags);
6493                 /* We don't care about errors in readahead. */
6494                 if (ret < 0)
6495                         continue;
6496                 BUG_ON(refs == 0);
6497
6498                 if (wc->stage == DROP_REFERENCE) {
6499                         if (refs == 1)
6500                                 goto reada;
6501
6502                         if (wc->level == 1 &&
6503                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6504                                 continue;
6505                         if (!wc->update_ref ||
6506                             generation <= root->root_key.offset)
6507                                 continue;
6508                         btrfs_node_key_to_cpu(eb, &key, slot);
6509                         ret = btrfs_comp_cpu_keys(&key,
6510                                                   &wc->update_progress);
6511                         if (ret < 0)
6512                                 continue;
6513                 } else {
6514                         if (wc->level == 1 &&
6515                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6516                                 continue;
6517                 }
6518 reada:
6519                 ret = readahead_tree_block(root, bytenr, blocksize,
6520                                            generation);
6521                 if (ret)
6522                         break;
6523                 nread++;
6524         }
6525         wc->reada_slot = slot;
6526 }
6527
6528 /*
6529  * hepler to process tree block while walking down the tree.
6530  *
6531  * when wc->stage == UPDATE_BACKREF, this function updates
6532  * back refs for pointers in the block.
6533  *
6534  * NOTE: return value 1 means we should stop walking down.
6535  */
6536 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6537                                    struct btrfs_root *root,
6538                                    struct btrfs_path *path,
6539                                    struct walk_control *wc, int lookup_info)
6540 {
6541         int level = wc->level;
6542         struct extent_buffer *eb = path->nodes[level];
6543         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6544         int ret;
6545
6546         if (wc->stage == UPDATE_BACKREF &&
6547             btrfs_header_owner(eb) != root->root_key.objectid)
6548                 return 1;
6549
6550         /*
6551          * when reference count of tree block is 1, it won't increase
6552          * again. once full backref flag is set, we never clear it.
6553          */
6554         if (lookup_info &&
6555             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6556              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6557                 BUG_ON(!path->locks[level]);
6558                 ret = btrfs_lookup_extent_info(trans, root,
6559                                                eb->start, eb->len,
6560                                                &wc->refs[level],
6561                                                &wc->flags[level]);
6562                 BUG_ON(ret == -ENOMEM);
6563                 if (ret)
6564                         return ret;
6565                 BUG_ON(wc->refs[level] == 0);
6566         }
6567
6568         if (wc->stage == DROP_REFERENCE) {
6569                 if (wc->refs[level] > 1)
6570                         return 1;
6571
6572                 if (path->locks[level] && !wc->keep_locks) {
6573                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6574                         path->locks[level] = 0;
6575                 }
6576                 return 0;
6577         }
6578
6579         /* wc->stage == UPDATE_BACKREF */
6580         if (!(wc->flags[level] & flag)) {
6581                 BUG_ON(!path->locks[level]);
6582                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
6583                 BUG_ON(ret); /* -ENOMEM */
6584                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
6585                 BUG_ON(ret); /* -ENOMEM */
6586                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6587                                                   eb->len, flag, 0);
6588                 BUG_ON(ret); /* -ENOMEM */
6589                 wc->flags[level] |= flag;
6590         }
6591
6592         /*
6593          * the block is shared by multiple trees, so it's not good to
6594          * keep the tree lock
6595          */
6596         if (path->locks[level] && level > 0) {
6597                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6598                 path->locks[level] = 0;
6599         }
6600         return 0;
6601 }
6602
6603 /*
6604  * hepler to process tree block pointer.
6605  *
6606  * when wc->stage == DROP_REFERENCE, this function checks
6607  * reference count of the block pointed to. if the block
6608  * is shared and we need update back refs for the subtree
6609  * rooted at the block, this function changes wc->stage to
6610  * UPDATE_BACKREF. if the block is shared and there is no
6611  * need to update back, this function drops the reference
6612  * to the block.
6613  *
6614  * NOTE: return value 1 means we should stop walking down.
6615  */
6616 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6617                                  struct btrfs_root *root,
6618                                  struct btrfs_path *path,
6619                                  struct walk_control *wc, int *lookup_info)
6620 {
6621         u64 bytenr;
6622         u64 generation;
6623         u64 parent;
6624         u32 blocksize;
6625         struct btrfs_key key;
6626         struct extent_buffer *next;
6627         int level = wc->level;
6628         int reada = 0;
6629         int ret = 0;
6630
6631         generation = btrfs_node_ptr_generation(path->nodes[level],
6632                                                path->slots[level]);
6633         /*
6634          * if the lower level block was created before the snapshot
6635          * was created, we know there is no need to update back refs
6636          * for the subtree
6637          */
6638         if (wc->stage == UPDATE_BACKREF &&
6639             generation <= root->root_key.offset) {
6640                 *lookup_info = 1;
6641                 return 1;
6642         }
6643
6644         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6645         blocksize = btrfs_level_size(root, level - 1);
6646
6647         next = btrfs_find_tree_block(root, bytenr, blocksize);
6648         if (!next) {
6649                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6650                 if (!next)
6651                         return -ENOMEM;
6652                 reada = 1;
6653         }
6654         btrfs_tree_lock(next);
6655         btrfs_set_lock_blocking(next);
6656
6657         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6658                                        &wc->refs[level - 1],
6659                                        &wc->flags[level - 1]);
6660         if (ret < 0) {
6661                 btrfs_tree_unlock(next);
6662                 return ret;
6663         }
6664
6665         BUG_ON(wc->refs[level - 1] == 0);
6666         *lookup_info = 0;
6667
6668         if (wc->stage == DROP_REFERENCE) {
6669                 if (wc->refs[level - 1] > 1) {
6670                         if (level == 1 &&
6671                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6672                                 goto skip;
6673
6674                         if (!wc->update_ref ||
6675                             generation <= root->root_key.offset)
6676                                 goto skip;
6677
6678                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6679                                               path->slots[level]);
6680                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6681                         if (ret < 0)
6682                                 goto skip;
6683
6684                         wc->stage = UPDATE_BACKREF;
6685                         wc->shared_level = level - 1;
6686                 }
6687         } else {
6688                 if (level == 1 &&
6689                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6690                         goto skip;
6691         }
6692
6693         if (!btrfs_buffer_uptodate(next, generation, 0)) {
6694                 btrfs_tree_unlock(next);
6695                 free_extent_buffer(next);
6696                 next = NULL;
6697                 *lookup_info = 1;
6698         }
6699
6700         if (!next) {
6701                 if (reada && level == 1)
6702                         reada_walk_down(trans, root, wc, path);
6703                 next = read_tree_block(root, bytenr, blocksize, generation);
6704                 if (!next)
6705                         return -EIO;
6706                 btrfs_tree_lock(next);
6707                 btrfs_set_lock_blocking(next);
6708         }
6709
6710         level--;
6711         BUG_ON(level != btrfs_header_level(next));
6712         path->nodes[level] = next;
6713         path->slots[level] = 0;
6714         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6715         wc->level = level;
6716         if (wc->level == 1)
6717                 wc->reada_slot = 0;
6718         return 0;
6719 skip:
6720         wc->refs[level - 1] = 0;
6721         wc->flags[level - 1] = 0;
6722         if (wc->stage == DROP_REFERENCE) {
6723                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6724                         parent = path->nodes[level]->start;
6725                 } else {
6726                         BUG_ON(root->root_key.objectid !=
6727                                btrfs_header_owner(path->nodes[level]));
6728                         parent = 0;
6729                 }
6730
6731                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6732                                 root->root_key.objectid, level - 1, 0, 0);
6733                 BUG_ON(ret); /* -ENOMEM */
6734         }
6735         btrfs_tree_unlock(next);
6736         free_extent_buffer(next);
6737         *lookup_info = 1;
6738         return 1;
6739 }
6740
6741 /*
6742  * hepler to process tree block while walking up the tree.
6743  *
6744  * when wc->stage == DROP_REFERENCE, this function drops
6745  * reference count on the block.
6746  *
6747  * when wc->stage == UPDATE_BACKREF, this function changes
6748  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6749  * to UPDATE_BACKREF previously while processing the block.
6750  *
6751  * NOTE: return value 1 means we should stop walking up.
6752  */
6753 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6754                                  struct btrfs_root *root,
6755                                  struct btrfs_path *path,
6756                                  struct walk_control *wc)
6757 {
6758         int ret;
6759         int level = wc->level;
6760         struct extent_buffer *eb = path->nodes[level];
6761         u64 parent = 0;
6762
6763         if (wc->stage == UPDATE_BACKREF) {
6764                 BUG_ON(wc->shared_level < level);
6765                 if (level < wc->shared_level)
6766                         goto out;
6767
6768                 ret = find_next_key(path, level + 1, &wc->update_progress);
6769                 if (ret > 0)
6770                         wc->update_ref = 0;
6771
6772                 wc->stage = DROP_REFERENCE;
6773                 wc->shared_level = -1;
6774                 path->slots[level] = 0;
6775
6776                 /*
6777                  * check reference count again if the block isn't locked.
6778                  * we should start walking down the tree again if reference
6779                  * count is one.
6780                  */
6781                 if (!path->locks[level]) {
6782                         BUG_ON(level == 0);
6783                         btrfs_tree_lock(eb);
6784                         btrfs_set_lock_blocking(eb);
6785                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6786
6787                         ret = btrfs_lookup_extent_info(trans, root,
6788                                                        eb->start, eb->len,
6789                                                        &wc->refs[level],
6790                                                        &wc->flags[level]);
6791                         if (ret < 0) {
6792                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6793                                 return ret;
6794                         }
6795                         BUG_ON(wc->refs[level] == 0);
6796                         if (wc->refs[level] == 1) {
6797                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6798                                 return 1;
6799                         }
6800                 }
6801         }
6802
6803         /* wc->stage == DROP_REFERENCE */
6804         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6805
6806         if (wc->refs[level] == 1) {
6807                 if (level == 0) {
6808                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6809                                 ret = btrfs_dec_ref(trans, root, eb, 1,
6810                                                     wc->for_reloc);
6811                         else
6812                                 ret = btrfs_dec_ref(trans, root, eb, 0,
6813                                                     wc->for_reloc);
6814                         BUG_ON(ret); /* -ENOMEM */
6815                 }
6816                 /* make block locked assertion in clean_tree_block happy */
6817                 if (!path->locks[level] &&
6818                     btrfs_header_generation(eb) == trans->transid) {
6819                         btrfs_tree_lock(eb);
6820                         btrfs_set_lock_blocking(eb);
6821                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6822                 }
6823                 clean_tree_block(trans, root, eb);
6824         }
6825
6826         if (eb == root->node) {
6827                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6828                         parent = eb->start;
6829                 else
6830                         BUG_ON(root->root_key.objectid !=
6831                                btrfs_header_owner(eb));
6832         } else {
6833                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6834                         parent = path->nodes[level + 1]->start;
6835                 else
6836                         BUG_ON(root->root_key.objectid !=
6837                                btrfs_header_owner(path->nodes[level + 1]));
6838         }
6839
6840         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6841 out:
6842         wc->refs[level] = 0;
6843         wc->flags[level] = 0;
6844         return 0;
6845 }
6846
6847 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6848                                    struct btrfs_root *root,
6849                                    struct btrfs_path *path,
6850                                    struct walk_control *wc)
6851 {
6852         int level = wc->level;
6853         int lookup_info = 1;
6854         int ret;
6855
6856         while (level >= 0) {
6857                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6858                 if (ret > 0)
6859                         break;
6860
6861                 if (level == 0)
6862                         break;
6863
6864                 if (path->slots[level] >=
6865                     btrfs_header_nritems(path->nodes[level]))
6866                         break;
6867
6868                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6869                 if (ret > 0) {
6870                         path->slots[level]++;
6871                         continue;
6872                 } else if (ret < 0)
6873                         return ret;
6874                 level = wc->level;
6875         }
6876         return 0;
6877 }
6878
6879 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6880                                  struct btrfs_root *root,
6881                                  struct btrfs_path *path,
6882                                  struct walk_control *wc, int max_level)
6883 {
6884         int level = wc->level;
6885         int ret;
6886
6887         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6888         while (level < max_level && path->nodes[level]) {
6889                 wc->level = level;
6890                 if (path->slots[level] + 1 <
6891                     btrfs_header_nritems(path->nodes[level])) {
6892                         path->slots[level]++;
6893                         return 0;
6894                 } else {
6895                         ret = walk_up_proc(trans, root, path, wc);
6896                         if (ret > 0)
6897                                 return 0;
6898
6899                         if (path->locks[level]) {
6900                                 btrfs_tree_unlock_rw(path->nodes[level],
6901                                                      path->locks[level]);
6902                                 path->locks[level] = 0;
6903                         }
6904                         free_extent_buffer(path->nodes[level]);
6905                         path->nodes[level] = NULL;
6906                         level++;
6907                 }
6908         }
6909         return 1;
6910 }
6911
6912 /*
6913  * drop a subvolume tree.
6914  *
6915  * this function traverses the tree freeing any blocks that only
6916  * referenced by the tree.
6917  *
6918  * when a shared tree block is found. this function decreases its
6919  * reference count by one. if update_ref is true, this function
6920  * also make sure backrefs for the shared block and all lower level
6921  * blocks are properly updated.
6922  */
6923 int btrfs_drop_snapshot(struct btrfs_root *root,
6924                          struct btrfs_block_rsv *block_rsv, int update_ref,
6925                          int for_reloc)
6926 {
6927         struct btrfs_path *path;
6928         struct btrfs_trans_handle *trans;
6929         struct btrfs_root *tree_root = root->fs_info->tree_root;
6930         struct btrfs_root_item *root_item = &root->root_item;
6931         struct walk_control *wc;
6932         struct btrfs_key key;
6933         int err = 0;
6934         int ret;
6935         int level;
6936
6937         path = btrfs_alloc_path();
6938         if (!path) {
6939                 err = -ENOMEM;
6940                 goto out;
6941         }
6942
6943         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6944         if (!wc) {
6945                 btrfs_free_path(path);
6946                 err = -ENOMEM;
6947                 goto out;
6948         }
6949
6950         trans = btrfs_start_transaction(tree_root, 0);
6951         if (IS_ERR(trans)) {
6952                 err = PTR_ERR(trans);
6953                 goto out_free;
6954         }
6955
6956         if (block_rsv)
6957                 trans->block_rsv = block_rsv;
6958
6959         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6960                 level = btrfs_header_level(root->node);
6961                 path->nodes[level] = btrfs_lock_root_node(root);
6962                 btrfs_set_lock_blocking(path->nodes[level]);
6963                 path->slots[level] = 0;
6964                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6965                 memset(&wc->update_progress, 0,
6966                        sizeof(wc->update_progress));
6967         } else {
6968                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6969                 memcpy(&wc->update_progress, &key,
6970                        sizeof(wc->update_progress));
6971
6972                 level = root_item->drop_level;
6973                 BUG_ON(level == 0);
6974                 path->lowest_level = level;
6975                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6976                 path->lowest_level = 0;
6977                 if (ret < 0) {
6978                         err = ret;
6979                         goto out_end_trans;
6980                 }
6981                 WARN_ON(ret > 0);
6982
6983                 /*
6984                  * unlock our path, this is safe because only this
6985                  * function is allowed to delete this snapshot
6986                  */
6987                 btrfs_unlock_up_safe(path, 0);
6988
6989                 level = btrfs_header_level(root->node);
6990                 while (1) {
6991                         btrfs_tree_lock(path->nodes[level]);
6992                         btrfs_set_lock_blocking(path->nodes[level]);
6993
6994                         ret = btrfs_lookup_extent_info(trans, root,
6995                                                 path->nodes[level]->start,
6996                                                 path->nodes[level]->len,
6997                                                 &wc->refs[level],
6998                                                 &wc->flags[level]);
6999                         if (ret < 0) {
7000                                 err = ret;
7001                                 goto out_end_trans;
7002                         }
7003                         BUG_ON(wc->refs[level] == 0);
7004
7005                         if (level == root_item->drop_level)
7006                                 break;
7007
7008                         btrfs_tree_unlock(path->nodes[level]);
7009                         WARN_ON(wc->refs[level] != 1);
7010                         level--;
7011                 }
7012         }
7013
7014         wc->level = level;
7015         wc->shared_level = -1;
7016         wc->stage = DROP_REFERENCE;
7017         wc->update_ref = update_ref;
7018         wc->keep_locks = 0;
7019         wc->for_reloc = for_reloc;
7020         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7021
7022         while (1) {
7023                 ret = walk_down_tree(trans, root, path, wc);
7024                 if (ret < 0) {
7025                         err = ret;
7026                         break;
7027                 }
7028
7029                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7030                 if (ret < 0) {
7031                         err = ret;
7032                         break;
7033                 }
7034
7035                 if (ret > 0) {
7036                         BUG_ON(wc->stage != DROP_REFERENCE);
7037                         break;
7038                 }
7039
7040                 if (wc->stage == DROP_REFERENCE) {
7041                         level = wc->level;
7042                         btrfs_node_key(path->nodes[level],
7043                                        &root_item->drop_progress,
7044                                        path->slots[level]);
7045                         root_item->drop_level = level;
7046                 }
7047
7048                 BUG_ON(wc->level == 0);
7049                 if (btrfs_should_end_transaction(trans, tree_root)) {
7050                         ret = btrfs_update_root(trans, tree_root,
7051                                                 &root->root_key,
7052                                                 root_item);
7053                         if (ret) {
7054                                 btrfs_abort_transaction(trans, tree_root, ret);
7055                                 err = ret;
7056                                 goto out_end_trans;
7057                         }
7058
7059                         btrfs_end_transaction_throttle(trans, tree_root);
7060                         trans = btrfs_start_transaction(tree_root, 0);
7061                         if (IS_ERR(trans)) {
7062                                 err = PTR_ERR(trans);
7063                                 goto out_free;
7064                         }
7065                         if (block_rsv)
7066                                 trans->block_rsv = block_rsv;
7067                 }
7068         }
7069         btrfs_release_path(path);
7070         if (err)
7071                 goto out_end_trans;
7072
7073         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7074         if (ret) {
7075                 btrfs_abort_transaction(trans, tree_root, ret);
7076                 goto out_end_trans;
7077         }
7078
7079         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7080                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
7081                                            NULL, NULL);
7082                 if (ret < 0) {
7083                         btrfs_abort_transaction(trans, tree_root, ret);
7084                         err = ret;
7085                         goto out_end_trans;
7086                 } else if (ret > 0) {
7087                         /* if we fail to delete the orphan item this time
7088                          * around, it'll get picked up the next time.
7089                          *
7090                          * The most common failure here is just -ENOENT.
7091                          */
7092                         btrfs_del_orphan_item(trans, tree_root,
7093                                               root->root_key.objectid);
7094                 }
7095         }
7096
7097         if (root->in_radix) {
7098                 btrfs_free_fs_root(tree_root->fs_info, root);
7099         } else {
7100                 free_extent_buffer(root->node);
7101                 free_extent_buffer(root->commit_root);
7102                 kfree(root);
7103         }
7104 out_end_trans:
7105         btrfs_end_transaction_throttle(trans, tree_root);
7106 out_free:
7107         kfree(wc);
7108         btrfs_free_path(path);
7109 out:
7110         if (err)
7111                 btrfs_std_error(root->fs_info, err);
7112         return err;
7113 }
7114
7115 /*
7116  * drop subtree rooted at tree block 'node'.
7117  *
7118  * NOTE: this function will unlock and release tree block 'node'
7119  * only used by relocation code
7120  */
7121 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7122                         struct btrfs_root *root,
7123                         struct extent_buffer *node,
7124                         struct extent_buffer *parent)
7125 {
7126         struct btrfs_path *path;
7127         struct walk_control *wc;
7128         int level;
7129         int parent_level;
7130         int ret = 0;
7131         int wret;
7132
7133         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7134
7135         path = btrfs_alloc_path();
7136         if (!path)
7137                 return -ENOMEM;
7138
7139         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7140         if (!wc) {
7141                 btrfs_free_path(path);
7142                 return -ENOMEM;
7143         }
7144
7145         btrfs_assert_tree_locked(parent);
7146         parent_level = btrfs_header_level(parent);
7147         extent_buffer_get(parent);
7148         path->nodes[parent_level] = parent;
7149         path->slots[parent_level] = btrfs_header_nritems(parent);
7150
7151         btrfs_assert_tree_locked(node);
7152         level = btrfs_header_level(node);
7153         path->nodes[level] = node;
7154         path->slots[level] = 0;
7155         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7156
7157         wc->refs[parent_level] = 1;
7158         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7159         wc->level = level;
7160         wc->shared_level = -1;
7161         wc->stage = DROP_REFERENCE;
7162         wc->update_ref = 0;
7163         wc->keep_locks = 1;
7164         wc->for_reloc = 1;
7165         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7166
7167         while (1) {
7168                 wret = walk_down_tree(trans, root, path, wc);
7169                 if (wret < 0) {
7170                         ret = wret;
7171                         break;
7172                 }
7173
7174                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7175                 if (wret < 0)
7176                         ret = wret;
7177                 if (wret != 0)
7178                         break;
7179         }
7180
7181         kfree(wc);
7182         btrfs_free_path(path);
7183         return ret;
7184 }
7185
7186 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7187 {
7188         u64 num_devices;
7189         u64 stripped;
7190
7191         /*
7192          * if restripe for this chunk_type is on pick target profile and
7193          * return, otherwise do the usual balance
7194          */
7195         stripped = get_restripe_target(root->fs_info, flags);
7196         if (stripped)
7197                 return extended_to_chunk(stripped);
7198
7199         /*
7200          * we add in the count of missing devices because we want
7201          * to make sure that any RAID levels on a degraded FS
7202          * continue to be honored.
7203          */
7204         num_devices = root->fs_info->fs_devices->rw_devices +
7205                 root->fs_info->fs_devices->missing_devices;
7206
7207         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7208                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7209
7210         if (num_devices == 1) {
7211                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7212                 stripped = flags & ~stripped;
7213
7214                 /* turn raid0 into single device chunks */
7215                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7216                         return stripped;
7217
7218                 /* turn mirroring into duplication */
7219                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7220                              BTRFS_BLOCK_GROUP_RAID10))
7221                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7222         } else {
7223                 /* they already had raid on here, just return */
7224                 if (flags & stripped)
7225                         return flags;
7226
7227                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7228                 stripped = flags & ~stripped;
7229
7230                 /* switch duplicated blocks with raid1 */
7231                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7232                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7233
7234                 /* this is drive concat, leave it alone */
7235         }
7236
7237         return flags;
7238 }
7239
7240 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7241 {
7242         struct btrfs_space_info *sinfo = cache->space_info;
7243         u64 num_bytes;
7244         u64 min_allocable_bytes;
7245         int ret = -ENOSPC;
7246
7247
7248         /*
7249          * We need some metadata space and system metadata space for
7250          * allocating chunks in some corner cases until we force to set
7251          * it to be readonly.
7252          */
7253         if ((sinfo->flags &
7254              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7255             !force)
7256                 min_allocable_bytes = 1 * 1024 * 1024;
7257         else
7258                 min_allocable_bytes = 0;
7259
7260         spin_lock(&sinfo->lock);
7261         spin_lock(&cache->lock);
7262
7263         if (cache->ro) {
7264                 ret = 0;
7265                 goto out;
7266         }
7267
7268         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7269                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7270
7271         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7272             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7273             min_allocable_bytes <= sinfo->total_bytes) {
7274                 sinfo->bytes_readonly += num_bytes;
7275                 cache->ro = 1;
7276                 ret = 0;
7277         }
7278 out:
7279         spin_unlock(&cache->lock);
7280         spin_unlock(&sinfo->lock);
7281         return ret;
7282 }
7283
7284 int btrfs_set_block_group_ro(struct btrfs_root *root,
7285                              struct btrfs_block_group_cache *cache)
7286
7287 {
7288         struct btrfs_trans_handle *trans;
7289         u64 alloc_flags;
7290         int ret;
7291
7292         BUG_ON(cache->ro);
7293
7294         trans = btrfs_join_transaction(root);
7295         if (IS_ERR(trans))
7296                 return PTR_ERR(trans);
7297
7298         alloc_flags = update_block_group_flags(root, cache->flags);
7299         if (alloc_flags != cache->flags) {
7300                 ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7301                                      CHUNK_ALLOC_FORCE);
7302                 if (ret < 0)
7303                         goto out;
7304         }
7305
7306         ret = set_block_group_ro(cache, 0);
7307         if (!ret)
7308                 goto out;
7309         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7310         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7311                              CHUNK_ALLOC_FORCE);
7312         if (ret < 0)
7313                 goto out;
7314         ret = set_block_group_ro(cache, 0);
7315 out:
7316         btrfs_end_transaction(trans, root);
7317         return ret;
7318 }
7319
7320 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7321                             struct btrfs_root *root, u64 type)
7322 {
7323         u64 alloc_flags = get_alloc_profile(root, type);
7324         return do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7325                               CHUNK_ALLOC_FORCE);
7326 }
7327
7328 /*
7329  * helper to account the unused space of all the readonly block group in the
7330  * list. takes mirrors into account.
7331  */
7332 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7333 {
7334         struct btrfs_block_group_cache *block_group;
7335         u64 free_bytes = 0;
7336         int factor;
7337
7338         list_for_each_entry(block_group, groups_list, list) {
7339                 spin_lock(&block_group->lock);
7340
7341                 if (!block_group->ro) {
7342                         spin_unlock(&block_group->lock);
7343                         continue;
7344                 }
7345
7346                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7347                                           BTRFS_BLOCK_GROUP_RAID10 |
7348                                           BTRFS_BLOCK_GROUP_DUP))
7349                         factor = 2;
7350                 else
7351                         factor = 1;
7352
7353                 free_bytes += (block_group->key.offset -
7354                                btrfs_block_group_used(&block_group->item)) *
7355                                factor;
7356
7357                 spin_unlock(&block_group->lock);
7358         }
7359
7360         return free_bytes;
7361 }
7362
7363 /*
7364  * helper to account the unused space of all the readonly block group in the
7365  * space_info. takes mirrors into account.
7366  */
7367 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7368 {
7369         int i;
7370         u64 free_bytes = 0;
7371
7372         spin_lock(&sinfo->lock);
7373
7374         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7375                 if (!list_empty(&sinfo->block_groups[i]))
7376                         free_bytes += __btrfs_get_ro_block_group_free_space(
7377                                                 &sinfo->block_groups[i]);
7378
7379         spin_unlock(&sinfo->lock);
7380
7381         return free_bytes;
7382 }
7383
7384 void btrfs_set_block_group_rw(struct btrfs_root *root,
7385                               struct btrfs_block_group_cache *cache)
7386 {
7387         struct btrfs_space_info *sinfo = cache->space_info;
7388         u64 num_bytes;
7389
7390         BUG_ON(!cache->ro);
7391
7392         spin_lock(&sinfo->lock);
7393         spin_lock(&cache->lock);
7394         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7395                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7396         sinfo->bytes_readonly -= num_bytes;
7397         cache->ro = 0;
7398         spin_unlock(&cache->lock);
7399         spin_unlock(&sinfo->lock);
7400 }
7401
7402 /*
7403  * checks to see if its even possible to relocate this block group.
7404  *
7405  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7406  * ok to go ahead and try.
7407  */
7408 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7409 {
7410         struct btrfs_block_group_cache *block_group;
7411         struct btrfs_space_info *space_info;
7412         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7413         struct btrfs_device *device;
7414         u64 min_free;
7415         u64 dev_min = 1;
7416         u64 dev_nr = 0;
7417         u64 target;
7418         int index;
7419         int full = 0;
7420         int ret = 0;
7421
7422         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7423
7424         /* odd, couldn't find the block group, leave it alone */
7425         if (!block_group)
7426                 return -1;
7427
7428         min_free = btrfs_block_group_used(&block_group->item);
7429
7430         /* no bytes used, we're good */
7431         if (!min_free)
7432                 goto out;
7433
7434         space_info = block_group->space_info;
7435         spin_lock(&space_info->lock);
7436
7437         full = space_info->full;
7438
7439         /*
7440          * if this is the last block group we have in this space, we can't
7441          * relocate it unless we're able to allocate a new chunk below.
7442          *
7443          * Otherwise, we need to make sure we have room in the space to handle
7444          * all of the extents from this block group.  If we can, we're good
7445          */
7446         if ((space_info->total_bytes != block_group->key.offset) &&
7447             (space_info->bytes_used + space_info->bytes_reserved +
7448              space_info->bytes_pinned + space_info->bytes_readonly +
7449              min_free < space_info->total_bytes)) {
7450                 spin_unlock(&space_info->lock);
7451                 goto out;
7452         }
7453         spin_unlock(&space_info->lock);
7454
7455         /*
7456          * ok we don't have enough space, but maybe we have free space on our
7457          * devices to allocate new chunks for relocation, so loop through our
7458          * alloc devices and guess if we have enough space.  if this block
7459          * group is going to be restriped, run checks against the target
7460          * profile instead of the current one.
7461          */
7462         ret = -1;
7463
7464         /*
7465          * index:
7466          *      0: raid10
7467          *      1: raid1
7468          *      2: dup
7469          *      3: raid0
7470          *      4: single
7471          */
7472         target = get_restripe_target(root->fs_info, block_group->flags);
7473         if (target) {
7474                 index = __get_block_group_index(extended_to_chunk(target));
7475         } else {
7476                 /*
7477                  * this is just a balance, so if we were marked as full
7478                  * we know there is no space for a new chunk
7479                  */
7480                 if (full)
7481                         goto out;
7482
7483                 index = get_block_group_index(block_group);
7484         }
7485
7486         if (index == 0) {
7487                 dev_min = 4;
7488                 /* Divide by 2 */
7489                 min_free >>= 1;
7490         } else if (index == 1) {
7491                 dev_min = 2;
7492         } else if (index == 2) {
7493                 /* Multiply by 2 */
7494                 min_free <<= 1;
7495         } else if (index == 3) {
7496                 dev_min = fs_devices->rw_devices;
7497                 do_div(min_free, dev_min);
7498         }
7499
7500         mutex_lock(&root->fs_info->chunk_mutex);
7501         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7502                 u64 dev_offset;
7503
7504                 /*
7505                  * check to make sure we can actually find a chunk with enough
7506                  * space to fit our block group in.
7507                  */
7508                 if (device->total_bytes > device->bytes_used + min_free) {
7509                         ret = find_free_dev_extent(device, min_free,
7510                                                    &dev_offset, NULL);
7511                         if (!ret)
7512                                 dev_nr++;
7513
7514                         if (dev_nr >= dev_min)
7515                                 break;
7516
7517                         ret = -1;
7518                 }
7519         }
7520         mutex_unlock(&root->fs_info->chunk_mutex);
7521 out:
7522         btrfs_put_block_group(block_group);
7523         return ret;
7524 }
7525
7526 static int find_first_block_group(struct btrfs_root *root,
7527                 struct btrfs_path *path, struct btrfs_key *key)
7528 {
7529         int ret = 0;
7530         struct btrfs_key found_key;
7531         struct extent_buffer *leaf;
7532         int slot;
7533
7534         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7535         if (ret < 0)
7536                 goto out;
7537
7538         while (1) {
7539                 slot = path->slots[0];
7540                 leaf = path->nodes[0];
7541                 if (slot >= btrfs_header_nritems(leaf)) {
7542                         ret = btrfs_next_leaf(root, path);
7543                         if (ret == 0)
7544                                 continue;
7545                         if (ret < 0)
7546                                 goto out;
7547                         break;
7548                 }
7549                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7550
7551                 if (found_key.objectid >= key->objectid &&
7552                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7553                         ret = 0;
7554                         goto out;
7555                 }
7556                 path->slots[0]++;
7557         }
7558 out:
7559         return ret;
7560 }
7561
7562 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7563 {
7564         struct btrfs_block_group_cache *block_group;
7565         u64 last = 0;
7566
7567         while (1) {
7568                 struct inode *inode;
7569
7570                 block_group = btrfs_lookup_first_block_group(info, last);
7571                 while (block_group) {
7572                         spin_lock(&block_group->lock);
7573                         if (block_group->iref)
7574                                 break;
7575                         spin_unlock(&block_group->lock);
7576                         block_group = next_block_group(info->tree_root,
7577                                                        block_group);
7578                 }
7579                 if (!block_group) {
7580                         if (last == 0)
7581                                 break;
7582                         last = 0;
7583                         continue;
7584                 }
7585
7586                 inode = block_group->inode;
7587                 block_group->iref = 0;
7588                 block_group->inode = NULL;
7589                 spin_unlock(&block_group->lock);
7590                 iput(inode);
7591                 last = block_group->key.objectid + block_group->key.offset;
7592                 btrfs_put_block_group(block_group);
7593         }
7594 }
7595
7596 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7597 {
7598         struct btrfs_block_group_cache *block_group;
7599         struct btrfs_space_info *space_info;
7600         struct btrfs_caching_control *caching_ctl;
7601         struct rb_node *n;
7602
7603         down_write(&info->extent_commit_sem);
7604         while (!list_empty(&info->caching_block_groups)) {
7605                 caching_ctl = list_entry(info->caching_block_groups.next,
7606                                          struct btrfs_caching_control, list);
7607                 list_del(&caching_ctl->list);
7608                 put_caching_control(caching_ctl);
7609         }
7610         up_write(&info->extent_commit_sem);
7611
7612         spin_lock(&info->block_group_cache_lock);
7613         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7614                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7615                                        cache_node);
7616                 rb_erase(&block_group->cache_node,
7617                          &info->block_group_cache_tree);
7618                 spin_unlock(&info->block_group_cache_lock);
7619
7620                 down_write(&block_group->space_info->groups_sem);
7621                 list_del(&block_group->list);
7622                 up_write(&block_group->space_info->groups_sem);
7623
7624                 if (block_group->cached == BTRFS_CACHE_STARTED)
7625                         wait_block_group_cache_done(block_group);
7626
7627                 /*
7628                  * We haven't cached this block group, which means we could
7629                  * possibly have excluded extents on this block group.
7630                  */
7631                 if (block_group->cached == BTRFS_CACHE_NO)
7632                         free_excluded_extents(info->extent_root, block_group);
7633
7634                 btrfs_remove_free_space_cache(block_group);
7635                 btrfs_put_block_group(block_group);
7636
7637                 spin_lock(&info->block_group_cache_lock);
7638         }
7639         spin_unlock(&info->block_group_cache_lock);
7640
7641         /* now that all the block groups are freed, go through and
7642          * free all the space_info structs.  This is only called during
7643          * the final stages of unmount, and so we know nobody is
7644          * using them.  We call synchronize_rcu() once before we start,
7645          * just to be on the safe side.
7646          */
7647         synchronize_rcu();
7648
7649         release_global_block_rsv(info);
7650
7651         while(!list_empty(&info->space_info)) {
7652                 space_info = list_entry(info->space_info.next,
7653                                         struct btrfs_space_info,
7654                                         list);
7655                 if (space_info->bytes_pinned > 0 ||
7656                     space_info->bytes_reserved > 0 ||
7657                     space_info->bytes_may_use > 0) {
7658                         WARN_ON(1);
7659                         dump_space_info(space_info, 0, 0);
7660                 }
7661                 list_del(&space_info->list);
7662                 kfree(space_info);
7663         }
7664         return 0;
7665 }
7666
7667 static void __link_block_group(struct btrfs_space_info *space_info,
7668                                struct btrfs_block_group_cache *cache)
7669 {
7670         int index = get_block_group_index(cache);
7671
7672         down_write(&space_info->groups_sem);
7673         list_add_tail(&cache->list, &space_info->block_groups[index]);
7674         up_write(&space_info->groups_sem);
7675 }
7676
7677 int btrfs_read_block_groups(struct btrfs_root *root)
7678 {
7679         struct btrfs_path *path;
7680         int ret;
7681         struct btrfs_block_group_cache *cache;
7682         struct btrfs_fs_info *info = root->fs_info;
7683         struct btrfs_space_info *space_info;
7684         struct btrfs_key key;
7685         struct btrfs_key found_key;
7686         struct extent_buffer *leaf;
7687         int need_clear = 0;
7688         u64 cache_gen;
7689
7690         root = info->extent_root;
7691         key.objectid = 0;
7692         key.offset = 0;
7693         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7694         path = btrfs_alloc_path();
7695         if (!path)
7696                 return -ENOMEM;
7697         path->reada = 1;
7698
7699         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7700         if (btrfs_test_opt(root, SPACE_CACHE) &&
7701             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7702                 need_clear = 1;
7703         if (btrfs_test_opt(root, CLEAR_CACHE))
7704                 need_clear = 1;
7705
7706         while (1) {
7707                 ret = find_first_block_group(root, path, &key);
7708                 if (ret > 0)
7709                         break;
7710                 if (ret != 0)
7711                         goto error;
7712                 leaf = path->nodes[0];
7713                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7714                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7715                 if (!cache) {
7716                         ret = -ENOMEM;
7717                         goto error;
7718                 }
7719                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7720                                                 GFP_NOFS);
7721                 if (!cache->free_space_ctl) {
7722                         kfree(cache);
7723                         ret = -ENOMEM;
7724                         goto error;
7725                 }
7726
7727                 atomic_set(&cache->count, 1);
7728                 spin_lock_init(&cache->lock);
7729                 cache->fs_info = info;
7730                 INIT_LIST_HEAD(&cache->list);
7731                 INIT_LIST_HEAD(&cache->cluster_list);
7732
7733                 if (need_clear) {
7734                         /*
7735                          * When we mount with old space cache, we need to
7736                          * set BTRFS_DC_CLEAR and set dirty flag.
7737                          *
7738                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
7739                          *    truncate the old free space cache inode and
7740                          *    setup a new one.
7741                          * b) Setting 'dirty flag' makes sure that we flush
7742                          *    the new space cache info onto disk.
7743                          */
7744                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7745                         if (btrfs_test_opt(root, SPACE_CACHE))
7746                                 cache->dirty = 1;
7747                 }
7748
7749                 read_extent_buffer(leaf, &cache->item,
7750                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7751                                    sizeof(cache->item));
7752                 memcpy(&cache->key, &found_key, sizeof(found_key));
7753
7754                 key.objectid = found_key.objectid + found_key.offset;
7755                 btrfs_release_path(path);
7756                 cache->flags = btrfs_block_group_flags(&cache->item);
7757                 cache->sectorsize = root->sectorsize;
7758
7759                 btrfs_init_free_space_ctl(cache);
7760
7761                 /*
7762                  * We need to exclude the super stripes now so that the space
7763                  * info has super bytes accounted for, otherwise we'll think
7764                  * we have more space than we actually do.
7765                  */
7766                 exclude_super_stripes(root, cache);
7767
7768                 /*
7769                  * check for two cases, either we are full, and therefore
7770                  * don't need to bother with the caching work since we won't
7771                  * find any space, or we are empty, and we can just add all
7772                  * the space in and be done with it.  This saves us _alot_ of
7773                  * time, particularly in the full case.
7774                  */
7775                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7776                         cache->last_byte_to_unpin = (u64)-1;
7777                         cache->cached = BTRFS_CACHE_FINISHED;
7778                         free_excluded_extents(root, cache);
7779                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7780                         cache->last_byte_to_unpin = (u64)-1;
7781                         cache->cached = BTRFS_CACHE_FINISHED;
7782                         add_new_free_space(cache, root->fs_info,
7783                                            found_key.objectid,
7784                                            found_key.objectid +
7785                                            found_key.offset);
7786                         free_excluded_extents(root, cache);
7787                 }
7788
7789                 ret = update_space_info(info, cache->flags, found_key.offset,
7790                                         btrfs_block_group_used(&cache->item),
7791                                         &space_info);
7792                 BUG_ON(ret); /* -ENOMEM */
7793                 cache->space_info = space_info;
7794                 spin_lock(&cache->space_info->lock);
7795                 cache->space_info->bytes_readonly += cache->bytes_super;
7796                 spin_unlock(&cache->space_info->lock);
7797
7798                 __link_block_group(space_info, cache);
7799
7800                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7801                 BUG_ON(ret); /* Logic error */
7802
7803                 set_avail_alloc_bits(root->fs_info, cache->flags);
7804                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7805                         set_block_group_ro(cache, 1);
7806         }
7807
7808         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7809                 if (!(get_alloc_profile(root, space_info->flags) &
7810                       (BTRFS_BLOCK_GROUP_RAID10 |
7811                        BTRFS_BLOCK_GROUP_RAID1 |
7812                        BTRFS_BLOCK_GROUP_DUP)))
7813                         continue;
7814                 /*
7815                  * avoid allocating from un-mirrored block group if there are
7816                  * mirrored block groups.
7817                  */
7818                 list_for_each_entry(cache, &space_info->block_groups[3], list)
7819                         set_block_group_ro(cache, 1);
7820                 list_for_each_entry(cache, &space_info->block_groups[4], list)
7821                         set_block_group_ro(cache, 1);
7822         }
7823
7824         init_global_block_rsv(info);
7825         ret = 0;
7826 error:
7827         btrfs_free_path(path);
7828         return ret;
7829 }
7830
7831 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7832                            struct btrfs_root *root, u64 bytes_used,
7833                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7834                            u64 size)
7835 {
7836         int ret;
7837         struct btrfs_root *extent_root;
7838         struct btrfs_block_group_cache *cache;
7839
7840         extent_root = root->fs_info->extent_root;
7841
7842         root->fs_info->last_trans_log_full_commit = trans->transid;
7843
7844         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7845         if (!cache)
7846                 return -ENOMEM;
7847         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7848                                         GFP_NOFS);
7849         if (!cache->free_space_ctl) {
7850                 kfree(cache);
7851                 return -ENOMEM;
7852         }
7853
7854         cache->key.objectid = chunk_offset;
7855         cache->key.offset = size;
7856         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7857         cache->sectorsize = root->sectorsize;
7858         cache->fs_info = root->fs_info;
7859
7860         atomic_set(&cache->count, 1);
7861         spin_lock_init(&cache->lock);
7862         INIT_LIST_HEAD(&cache->list);
7863         INIT_LIST_HEAD(&cache->cluster_list);
7864
7865         btrfs_init_free_space_ctl(cache);
7866
7867         btrfs_set_block_group_used(&cache->item, bytes_used);
7868         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7869         cache->flags = type;
7870         btrfs_set_block_group_flags(&cache->item, type);
7871
7872         cache->last_byte_to_unpin = (u64)-1;
7873         cache->cached = BTRFS_CACHE_FINISHED;
7874         exclude_super_stripes(root, cache);
7875
7876         add_new_free_space(cache, root->fs_info, chunk_offset,
7877                            chunk_offset + size);
7878
7879         free_excluded_extents(root, cache);
7880
7881         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7882                                 &cache->space_info);
7883         BUG_ON(ret); /* -ENOMEM */
7884         update_global_block_rsv(root->fs_info);
7885
7886         spin_lock(&cache->space_info->lock);
7887         cache->space_info->bytes_readonly += cache->bytes_super;
7888         spin_unlock(&cache->space_info->lock);
7889
7890         __link_block_group(cache->space_info, cache);
7891
7892         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7893         BUG_ON(ret); /* Logic error */
7894
7895         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
7896                                 sizeof(cache->item));
7897         if (ret) {
7898                 btrfs_abort_transaction(trans, extent_root, ret);
7899                 return ret;
7900         }
7901
7902         set_avail_alloc_bits(extent_root->fs_info, type);
7903
7904         return 0;
7905 }
7906
7907 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
7908 {
7909         u64 extra_flags = chunk_to_extended(flags) &
7910                                 BTRFS_EXTENDED_PROFILE_MASK;
7911
7912         if (flags & BTRFS_BLOCK_GROUP_DATA)
7913                 fs_info->avail_data_alloc_bits &= ~extra_flags;
7914         if (flags & BTRFS_BLOCK_GROUP_METADATA)
7915                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
7916         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
7917                 fs_info->avail_system_alloc_bits &= ~extra_flags;
7918 }
7919
7920 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7921                              struct btrfs_root *root, u64 group_start)
7922 {
7923         struct btrfs_path *path;
7924         struct btrfs_block_group_cache *block_group;
7925         struct btrfs_free_cluster *cluster;
7926         struct btrfs_root *tree_root = root->fs_info->tree_root;
7927         struct btrfs_key key;
7928         struct inode *inode;
7929         int ret;
7930         int index;
7931         int factor;
7932
7933         root = root->fs_info->extent_root;
7934
7935         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7936         BUG_ON(!block_group);
7937         BUG_ON(!block_group->ro);
7938
7939         /*
7940          * Free the reserved super bytes from this block group before
7941          * remove it.
7942          */
7943         free_excluded_extents(root, block_group);
7944
7945         memcpy(&key, &block_group->key, sizeof(key));
7946         index = get_block_group_index(block_group);
7947         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7948                                   BTRFS_BLOCK_GROUP_RAID1 |
7949                                   BTRFS_BLOCK_GROUP_RAID10))
7950                 factor = 2;
7951         else
7952                 factor = 1;
7953
7954         /* make sure this block group isn't part of an allocation cluster */
7955         cluster = &root->fs_info->data_alloc_cluster;
7956         spin_lock(&cluster->refill_lock);
7957         btrfs_return_cluster_to_free_space(block_group, cluster);
7958         spin_unlock(&cluster->refill_lock);
7959
7960         /*
7961          * make sure this block group isn't part of a metadata
7962          * allocation cluster
7963          */
7964         cluster = &root->fs_info->meta_alloc_cluster;
7965         spin_lock(&cluster->refill_lock);
7966         btrfs_return_cluster_to_free_space(block_group, cluster);
7967         spin_unlock(&cluster->refill_lock);
7968
7969         path = btrfs_alloc_path();
7970         if (!path) {
7971                 ret = -ENOMEM;
7972                 goto out;
7973         }
7974
7975         inode = lookup_free_space_inode(tree_root, block_group, path);
7976         if (!IS_ERR(inode)) {
7977                 ret = btrfs_orphan_add(trans, inode);
7978                 if (ret) {
7979                         btrfs_add_delayed_iput(inode);
7980                         goto out;
7981                 }
7982                 clear_nlink(inode);
7983                 /* One for the block groups ref */
7984                 spin_lock(&block_group->lock);
7985                 if (block_group->iref) {
7986                         block_group->iref = 0;
7987                         block_group->inode = NULL;
7988                         spin_unlock(&block_group->lock);
7989                         iput(inode);
7990                 } else {
7991                         spin_unlock(&block_group->lock);
7992                 }
7993                 /* One for our lookup ref */
7994                 btrfs_add_delayed_iput(inode);
7995         }
7996
7997         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
7998         key.offset = block_group->key.objectid;
7999         key.type = 0;
8000
8001         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
8002         if (ret < 0)
8003                 goto out;
8004         if (ret > 0)
8005                 btrfs_release_path(path);
8006         if (ret == 0) {
8007                 ret = btrfs_del_item(trans, tree_root, path);
8008                 if (ret)
8009                         goto out;
8010                 btrfs_release_path(path);
8011         }
8012
8013         spin_lock(&root->fs_info->block_group_cache_lock);
8014         rb_erase(&block_group->cache_node,
8015                  &root->fs_info->block_group_cache_tree);
8016         spin_unlock(&root->fs_info->block_group_cache_lock);
8017
8018         down_write(&block_group->space_info->groups_sem);
8019         /*
8020          * we must use list_del_init so people can check to see if they
8021          * are still on the list after taking the semaphore
8022          */
8023         list_del_init(&block_group->list);
8024         if (list_empty(&block_group->space_info->block_groups[index]))
8025                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
8026         up_write(&block_group->space_info->groups_sem);
8027
8028         if (block_group->cached == BTRFS_CACHE_STARTED)
8029                 wait_block_group_cache_done(block_group);
8030
8031         btrfs_remove_free_space_cache(block_group);
8032
8033         spin_lock(&block_group->space_info->lock);
8034         block_group->space_info->total_bytes -= block_group->key.offset;
8035         block_group->space_info->bytes_readonly -= block_group->key.offset;
8036         block_group->space_info->disk_total -= block_group->key.offset * factor;
8037         spin_unlock(&block_group->space_info->lock);
8038
8039         memcpy(&key, &block_group->key, sizeof(key));
8040
8041         btrfs_clear_space_info_full(root->fs_info);
8042
8043         btrfs_put_block_group(block_group);
8044         btrfs_put_block_group(block_group);
8045
8046         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8047         if (ret > 0)
8048                 ret = -EIO;
8049         if (ret < 0)
8050                 goto out;
8051
8052         ret = btrfs_del_item(trans, root, path);
8053 out:
8054         btrfs_free_path(path);
8055         return ret;
8056 }
8057
8058 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
8059 {
8060         struct btrfs_space_info *space_info;
8061         struct btrfs_super_block *disk_super;
8062         u64 features;
8063         u64 flags;
8064         int mixed = 0;
8065         int ret;
8066
8067         disk_super = fs_info->super_copy;
8068         if (!btrfs_super_root(disk_super))
8069                 return 1;
8070
8071         features = btrfs_super_incompat_flags(disk_super);
8072         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
8073                 mixed = 1;
8074
8075         flags = BTRFS_BLOCK_GROUP_SYSTEM;
8076         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8077         if (ret)
8078                 goto out;
8079
8080         if (mixed) {
8081                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
8082                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8083         } else {
8084                 flags = BTRFS_BLOCK_GROUP_METADATA;
8085                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8086                 if (ret)
8087                         goto out;
8088
8089                 flags = BTRFS_BLOCK_GROUP_DATA;
8090                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8091         }
8092 out:
8093         return ret;
8094 }
8095
8096 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8097 {
8098         return unpin_extent_range(root, start, end);
8099 }
8100
8101 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8102                                u64 num_bytes, u64 *actual_bytes)
8103 {
8104         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8105 }
8106
8107 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8108 {
8109         struct btrfs_fs_info *fs_info = root->fs_info;
8110         struct btrfs_block_group_cache *cache = NULL;
8111         u64 group_trimmed;
8112         u64 start;
8113         u64 end;
8114         u64 trimmed = 0;
8115         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8116         int ret = 0;
8117
8118         /*
8119          * try to trim all FS space, our block group may start from non-zero.
8120          */
8121         if (range->len == total_bytes)
8122                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8123         else
8124                 cache = btrfs_lookup_block_group(fs_info, range->start);
8125
8126         while (cache) {
8127                 if (cache->key.objectid >= (range->start + range->len)) {
8128                         btrfs_put_block_group(cache);
8129                         break;
8130                 }
8131
8132                 start = max(range->start, cache->key.objectid);
8133                 end = min(range->start + range->len,
8134                                 cache->key.objectid + cache->key.offset);
8135
8136                 if (end - start >= range->minlen) {
8137                         if (!block_group_cache_done(cache)) {
8138                                 ret = cache_block_group(cache, NULL, root, 0);
8139                                 if (!ret)
8140                                         wait_block_group_cache_done(cache);
8141                         }
8142                         ret = btrfs_trim_block_group(cache,
8143                                                      &group_trimmed,
8144                                                      start,
8145                                                      end,
8146                                                      range->minlen);
8147
8148                         trimmed += group_trimmed;
8149                         if (ret) {
8150                                 btrfs_put_block_group(cache);
8151                                 break;
8152                         }
8153                 }
8154
8155                 cache = next_block_group(fs_info->tree_root, cache);
8156         }
8157
8158         range->len = trimmed;
8159         return ret;
8160 }