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         num_allocated += global_rsv->size;
3508
3509         /*
3510          * in limited mode, we want to have some free space up to
3511          * about 1% of the FS size.
3512          */
3513         if (force == CHUNK_ALLOC_LIMITED) {
3514                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3515                 thresh = max_t(u64, 64 * 1024 * 1024,
3516                                div_factor_fine(thresh, 1));
3517
3518                 if (num_bytes - num_allocated < thresh)
3519                         return 1;
3520         }
3521         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3522
3523         /* 256MB or 2% of the FS */
3524         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 2));
3525         /* system chunks need a much small threshold */
3526         if (sinfo->flags & BTRFS_BLOCK_GROUP_SYSTEM)
3527                 thresh = 32 * 1024 * 1024;
3528
3529         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 8))
3530                 return 0;
3531         return 1;
3532 }
3533
3534 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3535 {
3536         u64 num_dev;
3537
3538         if (type & BTRFS_BLOCK_GROUP_RAID10 ||
3539             type & BTRFS_BLOCK_GROUP_RAID0)
3540                 num_dev = root->fs_info->fs_devices->rw_devices;
3541         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3542                 num_dev = 2;
3543         else
3544                 num_dev = 1;    /* DUP or single */
3545
3546         /* metadata for updaing devices and chunk tree */
3547         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3548 }
3549
3550 static void check_system_chunk(struct btrfs_trans_handle *trans,
3551                                struct btrfs_root *root, u64 type)
3552 {
3553         struct btrfs_space_info *info;
3554         u64 left;
3555         u64 thresh;
3556
3557         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3558         spin_lock(&info->lock);
3559         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3560                 info->bytes_reserved - info->bytes_readonly;
3561         spin_unlock(&info->lock);
3562
3563         thresh = get_system_chunk_thresh(root, type);
3564         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3565                 printk(KERN_INFO "left=%llu, need=%llu, flags=%llu\n",
3566                        left, thresh, type);
3567                 dump_space_info(info, 0, 0);
3568         }
3569
3570         if (left < thresh) {
3571                 u64 flags;
3572
3573                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3574                 btrfs_alloc_chunk(trans, root, flags);
3575         }
3576 }
3577
3578 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3579                           struct btrfs_root *extent_root, u64 alloc_bytes,
3580                           u64 flags, int force)
3581 {
3582         struct btrfs_space_info *space_info;
3583         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3584         int wait_for_alloc = 0;
3585         int ret = 0;
3586
3587         space_info = __find_space_info(extent_root->fs_info, flags);
3588         if (!space_info) {
3589                 ret = update_space_info(extent_root->fs_info, flags,
3590                                         0, 0, &space_info);
3591                 BUG_ON(ret); /* -ENOMEM */
3592         }
3593         BUG_ON(!space_info); /* Logic error */
3594
3595 again:
3596         spin_lock(&space_info->lock);
3597         if (force < space_info->force_alloc)
3598                 force = space_info->force_alloc;
3599         if (space_info->full) {
3600                 spin_unlock(&space_info->lock);
3601                 return 0;
3602         }
3603
3604         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3605                 spin_unlock(&space_info->lock);
3606                 return 0;
3607         } else if (space_info->chunk_alloc) {
3608                 wait_for_alloc = 1;
3609         } else {
3610                 space_info->chunk_alloc = 1;
3611         }
3612
3613         spin_unlock(&space_info->lock);
3614
3615         mutex_lock(&fs_info->chunk_mutex);
3616
3617         /*
3618          * The chunk_mutex is held throughout the entirety of a chunk
3619          * allocation, so once we've acquired the chunk_mutex we know that the
3620          * other guy is done and we need to recheck and see if we should
3621          * allocate.
3622          */
3623         if (wait_for_alloc) {
3624                 mutex_unlock(&fs_info->chunk_mutex);
3625                 wait_for_alloc = 0;
3626                 goto again;
3627         }
3628
3629         /*
3630          * If we have mixed data/metadata chunks we want to make sure we keep
3631          * allocating mixed chunks instead of individual chunks.
3632          */
3633         if (btrfs_mixed_space_info(space_info))
3634                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3635
3636         /*
3637          * if we're doing a data chunk, go ahead and make sure that
3638          * we keep a reasonable number of metadata chunks allocated in the
3639          * FS as well.
3640          */
3641         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3642                 fs_info->data_chunk_allocations++;
3643                 if (!(fs_info->data_chunk_allocations %
3644                       fs_info->metadata_ratio))
3645                         force_metadata_allocation(fs_info);
3646         }
3647
3648         /*
3649          * Check if we have enough space in SYSTEM chunk because we may need
3650          * to update devices.
3651          */
3652         check_system_chunk(trans, extent_root, flags);
3653
3654         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3655         if (ret < 0 && ret != -ENOSPC)
3656                 goto out;
3657
3658         spin_lock(&space_info->lock);
3659         if (ret)
3660                 space_info->full = 1;
3661         else
3662                 ret = 1;
3663
3664         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3665         space_info->chunk_alloc = 0;
3666         spin_unlock(&space_info->lock);
3667 out:
3668         mutex_unlock(&fs_info->chunk_mutex);
3669         return ret;
3670 }
3671
3672 /*
3673  * shrink metadata reservation for delalloc
3674  */
3675 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3676                             bool wait_ordered)
3677 {
3678         struct btrfs_block_rsv *block_rsv;
3679         struct btrfs_space_info *space_info;
3680         struct btrfs_trans_handle *trans;
3681         u64 delalloc_bytes;
3682         u64 max_reclaim;
3683         long time_left;
3684         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3685         int loops = 0;
3686
3687         trans = (struct btrfs_trans_handle *)current->journal_info;
3688         block_rsv = &root->fs_info->delalloc_block_rsv;
3689         space_info = block_rsv->space_info;
3690
3691         smp_mb();
3692         delalloc_bytes = root->fs_info->delalloc_bytes;
3693         if (delalloc_bytes == 0) {
3694                 if (trans)
3695                         return;
3696                 btrfs_wait_ordered_extents(root, 0, 0);
3697                 return;
3698         }
3699
3700         while (delalloc_bytes && loops < 3) {
3701                 max_reclaim = min(delalloc_bytes, to_reclaim);
3702                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3703                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3704                                                WB_REASON_FS_FREE_SPACE);
3705
3706                 spin_lock(&space_info->lock);
3707                 if (space_info->bytes_used + space_info->bytes_reserved +
3708                     space_info->bytes_pinned + space_info->bytes_readonly +
3709                     space_info->bytes_may_use + orig <=
3710                     space_info->total_bytes) {
3711                         spin_unlock(&space_info->lock);
3712                         break;
3713                 }
3714                 spin_unlock(&space_info->lock);
3715
3716                 loops++;
3717                 if (wait_ordered && !trans) {
3718                         btrfs_wait_ordered_extents(root, 0, 0);
3719                 } else {
3720                         time_left = schedule_timeout_killable(1);
3721                         if (time_left)
3722                                 break;
3723                 }
3724                 smp_mb();
3725                 delalloc_bytes = root->fs_info->delalloc_bytes;
3726         }
3727 }
3728
3729 /**
3730  * maybe_commit_transaction - possibly commit the transaction if its ok to
3731  * @root - the root we're allocating for
3732  * @bytes - the number of bytes we want to reserve
3733  * @force - force the commit
3734  *
3735  * This will check to make sure that committing the transaction will actually
3736  * get us somewhere and then commit the transaction if it does.  Otherwise it
3737  * will return -ENOSPC.
3738  */
3739 static int may_commit_transaction(struct btrfs_root *root,
3740                                   struct btrfs_space_info *space_info,
3741                                   u64 bytes, int force)
3742 {
3743         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3744         struct btrfs_trans_handle *trans;
3745
3746         trans = (struct btrfs_trans_handle *)current->journal_info;
3747         if (trans)
3748                 return -EAGAIN;
3749
3750         if (force)
3751                 goto commit;
3752
3753         /* See if there is enough pinned space to make this reservation */
3754         spin_lock(&space_info->lock);
3755         if (space_info->bytes_pinned >= bytes) {
3756                 spin_unlock(&space_info->lock);
3757                 goto commit;
3758         }
3759         spin_unlock(&space_info->lock);
3760
3761         /*
3762          * See if there is some space in the delayed insertion reservation for
3763          * this reservation.
3764          */
3765         if (space_info != delayed_rsv->space_info)
3766                 return -ENOSPC;
3767
3768         spin_lock(&space_info->lock);
3769         spin_lock(&delayed_rsv->lock);
3770         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3771                 spin_unlock(&delayed_rsv->lock);
3772                 spin_unlock(&space_info->lock);
3773                 return -ENOSPC;
3774         }
3775         spin_unlock(&delayed_rsv->lock);
3776         spin_unlock(&space_info->lock);
3777
3778 commit:
3779         trans = btrfs_join_transaction(root);
3780         if (IS_ERR(trans))
3781                 return -ENOSPC;
3782
3783         return btrfs_commit_transaction(trans, root);
3784 }
3785
3786 enum flush_state {
3787         FLUSH_DELALLOC          =       1,
3788         FLUSH_DELALLOC_WAIT     =       2,
3789         FLUSH_DELAYED_ITEMS_NR  =       3,
3790         FLUSH_DELAYED_ITEMS     =       4,
3791         COMMIT_TRANS            =       5,
3792 };
3793
3794 static int flush_space(struct btrfs_root *root,
3795                        struct btrfs_space_info *space_info, u64 num_bytes,
3796                        u64 orig_bytes, int state)
3797 {
3798         struct btrfs_trans_handle *trans;
3799         int nr;
3800         int ret = 0;
3801
3802         switch (state) {
3803         case FLUSH_DELALLOC:
3804         case FLUSH_DELALLOC_WAIT:
3805                 shrink_delalloc(root, num_bytes, orig_bytes,
3806                                 state == FLUSH_DELALLOC_WAIT);
3807                 break;
3808         case FLUSH_DELAYED_ITEMS_NR:
3809         case FLUSH_DELAYED_ITEMS:
3810                 if (state == FLUSH_DELAYED_ITEMS_NR) {
3811                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
3812
3813                         nr = (int)div64_u64(num_bytes, bytes);
3814                         if (!nr)
3815                                 nr = 1;
3816                         nr *= 2;
3817                 } else {
3818                         nr = -1;
3819                 }
3820                 trans = btrfs_join_transaction(root);
3821                 if (IS_ERR(trans)) {
3822                         ret = PTR_ERR(trans);
3823                         break;
3824                 }
3825                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
3826                 btrfs_end_transaction(trans, root);
3827                 break;
3828         case COMMIT_TRANS:
3829                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3830                 break;
3831         default:
3832                 ret = -ENOSPC;
3833                 break;
3834         }
3835
3836         return ret;
3837 }
3838 /**
3839  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3840  * @root - the root we're allocating for
3841  * @block_rsv - the block_rsv we're allocating for
3842  * @orig_bytes - the number of bytes we want
3843  * @flush - wether or not we can flush to make our reservation
3844  *
3845  * This will reserve orgi_bytes number of bytes from the space info associated
3846  * with the block_rsv.  If there is not enough space it will make an attempt to
3847  * flush out space to make room.  It will do this by flushing delalloc if
3848  * possible or committing the transaction.  If flush is 0 then no attempts to
3849  * regain reservations will be made and this will fail if there is not enough
3850  * space already.
3851  */
3852 static int reserve_metadata_bytes(struct btrfs_root *root,
3853                                   struct btrfs_block_rsv *block_rsv,
3854                                   u64 orig_bytes, int flush)
3855 {
3856         struct btrfs_space_info *space_info = block_rsv->space_info;
3857         u64 used;
3858         u64 num_bytes = orig_bytes;
3859         int flush_state = FLUSH_DELALLOC;
3860         int ret = 0;
3861         bool flushing = false;
3862         bool committed = false;
3863
3864 again:
3865         ret = 0;
3866         spin_lock(&space_info->lock);
3867         /*
3868          * We only want to wait if somebody other than us is flushing and we are
3869          * actually alloed to flush.
3870          */
3871         while (flush && !flushing && space_info->flush) {
3872                 spin_unlock(&space_info->lock);
3873                 /*
3874                  * If we have a trans handle we can't wait because the flusher
3875                  * may have to commit the transaction, which would mean we would
3876                  * deadlock since we are waiting for the flusher to finish, but
3877                  * hold the current transaction open.
3878                  */
3879                 if (current->journal_info)
3880                         return -EAGAIN;
3881                 ret = wait_event_killable(space_info->wait, !space_info->flush);
3882                 /* Must have been killed, return */
3883                 if (ret)
3884                         return -EINTR;
3885
3886                 spin_lock(&space_info->lock);
3887         }
3888
3889         ret = -ENOSPC;
3890         used = space_info->bytes_used + space_info->bytes_reserved +
3891                 space_info->bytes_pinned + space_info->bytes_readonly +
3892                 space_info->bytes_may_use;
3893
3894         /*
3895          * The idea here is that we've not already over-reserved the block group
3896          * then we can go ahead and save our reservation first and then start
3897          * flushing if we need to.  Otherwise if we've already overcommitted
3898          * lets start flushing stuff first and then come back and try to make
3899          * our reservation.
3900          */
3901         if (used <= space_info->total_bytes) {
3902                 if (used + orig_bytes <= space_info->total_bytes) {
3903                         space_info->bytes_may_use += orig_bytes;
3904                         trace_btrfs_space_reservation(root->fs_info,
3905                                 "space_info", space_info->flags, orig_bytes, 1);
3906                         ret = 0;
3907                 } else {
3908                         /*
3909                          * Ok set num_bytes to orig_bytes since we aren't
3910                          * overocmmitted, this way we only try and reclaim what
3911                          * we need.
3912                          */
3913                         num_bytes = orig_bytes;
3914                 }
3915         } else {
3916                 /*
3917                  * Ok we're over committed, set num_bytes to the overcommitted
3918                  * amount plus the amount of bytes that we need for this
3919                  * reservation.
3920                  */
3921                 num_bytes = used - space_info->total_bytes +
3922                         (orig_bytes * 2);
3923         }
3924
3925         if (ret) {
3926                 u64 profile = btrfs_get_alloc_profile(root, 0);
3927                 u64 avail;
3928
3929                 /*
3930                  * If we have a lot of space that's pinned, don't bother doing
3931                  * the overcommit dance yet and just commit the transaction.
3932                  */
3933                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3934                 do_div(avail, 10);
3935                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3936                         space_info->flush = 1;
3937                         flushing = true;
3938                         spin_unlock(&space_info->lock);
3939                         ret = may_commit_transaction(root, space_info,
3940                                                      orig_bytes, 1);
3941                         if (ret)
3942                                 goto out;
3943                         committed = true;
3944                         goto again;
3945                 }
3946
3947                 spin_lock(&root->fs_info->free_chunk_lock);
3948                 avail = root->fs_info->free_chunk_space;
3949
3950                 /*
3951                  * If we have dup, raid1 or raid10 then only half of the free
3952                  * space is actually useable.
3953                  */
3954                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3955                                BTRFS_BLOCK_GROUP_RAID1 |
3956                                BTRFS_BLOCK_GROUP_RAID10))
3957                         avail >>= 1;
3958
3959                 /*
3960                  * If we aren't flushing don't let us overcommit too much, say
3961                  * 1/8th of the space.  If we can flush, let it overcommit up to
3962                  * 1/2 of the space.
3963                  */
3964                 if (flush)
3965                         avail >>= 3;
3966                 else
3967                         avail >>= 1;
3968                  spin_unlock(&root->fs_info->free_chunk_lock);
3969
3970                 if (used + num_bytes < space_info->total_bytes + avail) {
3971                         space_info->bytes_may_use += orig_bytes;
3972                         trace_btrfs_space_reservation(root->fs_info,
3973                                 "space_info", space_info->flags, orig_bytes, 1);
3974                         ret = 0;
3975                 }
3976         }
3977
3978         /*
3979          * Couldn't make our reservation, save our place so while we're trying
3980          * to reclaim space we can actually use it instead of somebody else
3981          * stealing it from us.
3982          */
3983         if (ret && flush) {
3984                 flushing = true;
3985                 space_info->flush = 1;
3986         }
3987
3988         spin_unlock(&space_info->lock);
3989
3990         if (!ret || !flush)
3991                 goto out;
3992
3993         ret = flush_space(root, space_info, num_bytes, orig_bytes,
3994                           flush_state);
3995         flush_state++;
3996         if (!ret)
3997                 goto again;
3998         else if (flush_state <= COMMIT_TRANS)
3999                 goto again;
4000
4001 out:
4002         if (flushing) {
4003                 spin_lock(&space_info->lock);
4004                 space_info->flush = 0;
4005                 wake_up_all(&space_info->wait);
4006                 spin_unlock(&space_info->lock);
4007         }
4008         return ret;
4009 }
4010
4011 static struct btrfs_block_rsv *get_block_rsv(
4012                                         const struct btrfs_trans_handle *trans,
4013                                         const struct btrfs_root *root)
4014 {
4015         struct btrfs_block_rsv *block_rsv = NULL;
4016
4017         if (root->ref_cows)
4018                 block_rsv = trans->block_rsv;
4019
4020         if (root == root->fs_info->csum_root && trans->adding_csums)
4021                 block_rsv = trans->block_rsv;
4022
4023         if (!block_rsv)
4024                 block_rsv = root->block_rsv;
4025
4026         if (!block_rsv)
4027                 block_rsv = &root->fs_info->empty_block_rsv;
4028
4029         return block_rsv;
4030 }
4031
4032 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4033                                u64 num_bytes)
4034 {
4035         int ret = -ENOSPC;
4036         spin_lock(&block_rsv->lock);
4037         if (block_rsv->reserved >= num_bytes) {
4038                 block_rsv->reserved -= num_bytes;
4039                 if (block_rsv->reserved < block_rsv->size)
4040                         block_rsv->full = 0;
4041                 ret = 0;
4042         }
4043         spin_unlock(&block_rsv->lock);
4044         return ret;
4045 }
4046
4047 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4048                                 u64 num_bytes, int update_size)
4049 {
4050         spin_lock(&block_rsv->lock);
4051         block_rsv->reserved += num_bytes;
4052         if (update_size)
4053                 block_rsv->size += num_bytes;
4054         else if (block_rsv->reserved >= block_rsv->size)
4055                 block_rsv->full = 1;
4056         spin_unlock(&block_rsv->lock);
4057 }
4058
4059 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4060                                     struct btrfs_block_rsv *block_rsv,
4061                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4062 {
4063         struct btrfs_space_info *space_info = block_rsv->space_info;
4064
4065         spin_lock(&block_rsv->lock);
4066         if (num_bytes == (u64)-1)
4067                 num_bytes = block_rsv->size;
4068         block_rsv->size -= num_bytes;
4069         if (block_rsv->reserved >= block_rsv->size) {
4070                 num_bytes = block_rsv->reserved - block_rsv->size;
4071                 block_rsv->reserved = block_rsv->size;
4072                 block_rsv->full = 1;
4073         } else {
4074                 num_bytes = 0;
4075         }
4076         spin_unlock(&block_rsv->lock);
4077
4078         if (num_bytes > 0) {
4079                 if (dest) {
4080                         spin_lock(&dest->lock);
4081                         if (!dest->full) {
4082                                 u64 bytes_to_add;
4083
4084                                 bytes_to_add = dest->size - dest->reserved;
4085                                 bytes_to_add = min(num_bytes, bytes_to_add);
4086                                 dest->reserved += bytes_to_add;
4087                                 if (dest->reserved >= dest->size)
4088                                         dest->full = 1;
4089                                 num_bytes -= bytes_to_add;
4090                         }
4091                         spin_unlock(&dest->lock);
4092                 }
4093                 if (num_bytes) {
4094                         spin_lock(&space_info->lock);
4095                         space_info->bytes_may_use -= num_bytes;
4096                         trace_btrfs_space_reservation(fs_info, "space_info",
4097                                         space_info->flags, num_bytes, 0);
4098                         space_info->reservation_progress++;
4099                         spin_unlock(&space_info->lock);
4100                 }
4101         }
4102 }
4103
4104 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4105                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4106 {
4107         int ret;
4108
4109         ret = block_rsv_use_bytes(src, num_bytes);
4110         if (ret)
4111                 return ret;
4112
4113         block_rsv_add_bytes(dst, num_bytes, 1);
4114         return 0;
4115 }
4116
4117 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
4118 {
4119         memset(rsv, 0, sizeof(*rsv));
4120         spin_lock_init(&rsv->lock);
4121 }
4122
4123 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
4124 {
4125         struct btrfs_block_rsv *block_rsv;
4126         struct btrfs_fs_info *fs_info = root->fs_info;
4127
4128         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4129         if (!block_rsv)
4130                 return NULL;
4131
4132         btrfs_init_block_rsv(block_rsv);
4133         block_rsv->space_info = __find_space_info(fs_info,
4134                                                   BTRFS_BLOCK_GROUP_METADATA);
4135         return block_rsv;
4136 }
4137
4138 void btrfs_free_block_rsv(struct btrfs_root *root,
4139                           struct btrfs_block_rsv *rsv)
4140 {
4141         btrfs_block_rsv_release(root, rsv, (u64)-1);
4142         kfree(rsv);
4143 }
4144
4145 static inline int __block_rsv_add(struct btrfs_root *root,
4146                                   struct btrfs_block_rsv *block_rsv,
4147                                   u64 num_bytes, int flush)
4148 {
4149         int ret;
4150
4151         if (num_bytes == 0)
4152                 return 0;
4153
4154         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4155         if (!ret) {
4156                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4157                 return 0;
4158         }
4159
4160         return ret;
4161 }
4162
4163 int btrfs_block_rsv_add(struct btrfs_root *root,
4164                         struct btrfs_block_rsv *block_rsv,
4165                         u64 num_bytes)
4166 {
4167         return __block_rsv_add(root, block_rsv, num_bytes, 1);
4168 }
4169
4170 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
4171                                 struct btrfs_block_rsv *block_rsv,
4172                                 u64 num_bytes)
4173 {
4174         return __block_rsv_add(root, block_rsv, num_bytes, 0);
4175 }
4176
4177 int btrfs_block_rsv_check(struct btrfs_root *root,
4178                           struct btrfs_block_rsv *block_rsv, int min_factor)
4179 {
4180         u64 num_bytes = 0;
4181         int ret = -ENOSPC;
4182
4183         if (!block_rsv)
4184                 return 0;
4185
4186         spin_lock(&block_rsv->lock);
4187         num_bytes = div_factor(block_rsv->size, min_factor);
4188         if (block_rsv->reserved >= num_bytes)
4189                 ret = 0;
4190         spin_unlock(&block_rsv->lock);
4191
4192         return ret;
4193 }
4194
4195 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
4196                                            struct btrfs_block_rsv *block_rsv,
4197                                            u64 min_reserved, int flush)
4198 {
4199         u64 num_bytes = 0;
4200         int ret = -ENOSPC;
4201
4202         if (!block_rsv)
4203                 return 0;
4204
4205         spin_lock(&block_rsv->lock);
4206         num_bytes = min_reserved;
4207         if (block_rsv->reserved >= num_bytes)
4208                 ret = 0;
4209         else
4210                 num_bytes -= block_rsv->reserved;
4211         spin_unlock(&block_rsv->lock);
4212
4213         if (!ret)
4214                 return 0;
4215
4216         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4217         if (!ret) {
4218                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4219                 return 0;
4220         }
4221
4222         return ret;
4223 }
4224
4225 int btrfs_block_rsv_refill(struct btrfs_root *root,
4226                            struct btrfs_block_rsv *block_rsv,
4227                            u64 min_reserved)
4228 {
4229         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
4230 }
4231
4232 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
4233                                    struct btrfs_block_rsv *block_rsv,
4234                                    u64 min_reserved)
4235 {
4236         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
4237 }
4238
4239 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4240                             struct btrfs_block_rsv *dst_rsv,
4241                             u64 num_bytes)
4242 {
4243         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4244 }
4245
4246 void btrfs_block_rsv_release(struct btrfs_root *root,
4247                              struct btrfs_block_rsv *block_rsv,
4248                              u64 num_bytes)
4249 {
4250         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4251         if (global_rsv->full || global_rsv == block_rsv ||
4252             block_rsv->space_info != global_rsv->space_info)
4253                 global_rsv = NULL;
4254         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4255                                 num_bytes);
4256 }
4257
4258 /*
4259  * helper to calculate size of global block reservation.
4260  * the desired value is sum of space used by extent tree,
4261  * checksum tree and root tree
4262  */
4263 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4264 {
4265         struct btrfs_space_info *sinfo;
4266         u64 num_bytes;
4267         u64 meta_used;
4268         u64 data_used;
4269         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4270
4271         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4272         spin_lock(&sinfo->lock);
4273         data_used = sinfo->bytes_used;
4274         spin_unlock(&sinfo->lock);
4275
4276         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4277         spin_lock(&sinfo->lock);
4278         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4279                 data_used = 0;
4280         meta_used = sinfo->bytes_used;
4281         spin_unlock(&sinfo->lock);
4282
4283         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4284                     csum_size * 2;
4285         num_bytes += div64_u64(data_used + meta_used, 50);
4286
4287         if (num_bytes * 3 > meta_used)
4288                 num_bytes = div64_u64(meta_used, 3);
4289
4290         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4291 }
4292
4293 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4294 {
4295         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4296         struct btrfs_space_info *sinfo = block_rsv->space_info;
4297         u64 num_bytes;
4298
4299         num_bytes = calc_global_metadata_size(fs_info);
4300
4301         spin_lock(&sinfo->lock);
4302         spin_lock(&block_rsv->lock);
4303
4304         block_rsv->size = num_bytes;
4305
4306         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4307                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4308                     sinfo->bytes_may_use;
4309
4310         if (sinfo->total_bytes > num_bytes) {
4311                 num_bytes = sinfo->total_bytes - num_bytes;
4312                 block_rsv->reserved += num_bytes;
4313                 sinfo->bytes_may_use += num_bytes;
4314                 trace_btrfs_space_reservation(fs_info, "space_info",
4315                                       sinfo->flags, num_bytes, 1);
4316         }
4317
4318         if (block_rsv->reserved >= block_rsv->size) {
4319                 num_bytes = block_rsv->reserved - block_rsv->size;
4320                 sinfo->bytes_may_use -= num_bytes;
4321                 trace_btrfs_space_reservation(fs_info, "space_info",
4322                                       sinfo->flags, num_bytes, 0);
4323                 sinfo->reservation_progress++;
4324                 block_rsv->reserved = block_rsv->size;
4325                 block_rsv->full = 1;
4326         }
4327
4328         spin_unlock(&block_rsv->lock);
4329         spin_unlock(&sinfo->lock);
4330 }
4331
4332 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4333 {
4334         struct btrfs_space_info *space_info;
4335
4336         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4337         fs_info->chunk_block_rsv.space_info = space_info;
4338
4339         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4340         fs_info->global_block_rsv.space_info = space_info;
4341         fs_info->delalloc_block_rsv.space_info = space_info;
4342         fs_info->trans_block_rsv.space_info = space_info;
4343         fs_info->empty_block_rsv.space_info = space_info;
4344         fs_info->delayed_block_rsv.space_info = space_info;
4345
4346         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4347         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4348         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4349         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4350         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4351
4352         update_global_block_rsv(fs_info);
4353 }
4354
4355 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4356 {
4357         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4358                                 (u64)-1);
4359         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4360         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4361         WARN_ON(fs_info->trans_block_rsv.size > 0);
4362         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4363         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4364         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4365         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4366         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4367 }
4368
4369 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4370                                   struct btrfs_root *root)
4371 {
4372         if (!trans->block_rsv)
4373                 return;
4374
4375         if (!trans->bytes_reserved)
4376                 return;
4377
4378         trace_btrfs_space_reservation(root->fs_info, "transaction",
4379                                       trans->transid, trans->bytes_reserved, 0);
4380         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4381         trans->bytes_reserved = 0;
4382 }
4383
4384 /* Can only return 0 or -ENOSPC */
4385 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4386                                   struct inode *inode)
4387 {
4388         struct btrfs_root *root = BTRFS_I(inode)->root;
4389         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4390         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4391
4392         /*
4393          * We need to hold space in order to delete our orphan item once we've
4394          * added it, so this takes the reservation so we can release it later
4395          * when we are truly done with the orphan item.
4396          */
4397         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4398         trace_btrfs_space_reservation(root->fs_info, "orphan",
4399                                       btrfs_ino(inode), num_bytes, 1);
4400         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4401 }
4402
4403 void btrfs_orphan_release_metadata(struct inode *inode)
4404 {
4405         struct btrfs_root *root = BTRFS_I(inode)->root;
4406         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4407         trace_btrfs_space_reservation(root->fs_info, "orphan",
4408                                       btrfs_ino(inode), num_bytes, 0);
4409         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4410 }
4411
4412 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4413                                 struct btrfs_pending_snapshot *pending)
4414 {
4415         struct btrfs_root *root = pending->root;
4416         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4417         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4418         /*
4419          * two for root back/forward refs, two for directory entries
4420          * and one for root of the snapshot.
4421          */
4422         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4423         dst_rsv->space_info = src_rsv->space_info;
4424         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4425 }
4426
4427 /**
4428  * drop_outstanding_extent - drop an outstanding extent
4429  * @inode: the inode we're dropping the extent for
4430  *
4431  * This is called when we are freeing up an outstanding extent, either called
4432  * after an error or after an extent is written.  This will return the number of
4433  * reserved extents that need to be freed.  This must be called with
4434  * BTRFS_I(inode)->lock held.
4435  */
4436 static unsigned drop_outstanding_extent(struct inode *inode)
4437 {
4438         unsigned drop_inode_space = 0;
4439         unsigned dropped_extents = 0;
4440
4441         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4442         BTRFS_I(inode)->outstanding_extents--;
4443
4444         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4445             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4446                                &BTRFS_I(inode)->runtime_flags))
4447                 drop_inode_space = 1;
4448
4449         /*
4450          * If we have more or the same amount of outsanding extents than we have
4451          * reserved then we need to leave the reserved extents count alone.
4452          */
4453         if (BTRFS_I(inode)->outstanding_extents >=
4454             BTRFS_I(inode)->reserved_extents)
4455                 return drop_inode_space;
4456
4457         dropped_extents = BTRFS_I(inode)->reserved_extents -
4458                 BTRFS_I(inode)->outstanding_extents;
4459         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4460         return dropped_extents + drop_inode_space;
4461 }
4462
4463 /**
4464  * calc_csum_metadata_size - return the amount of metada space that must be
4465  *      reserved/free'd for the given bytes.
4466  * @inode: the inode we're manipulating
4467  * @num_bytes: the number of bytes in question
4468  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4469  *
4470  * This adjusts the number of csum_bytes in the inode and then returns the
4471  * correct amount of metadata that must either be reserved or freed.  We
4472  * calculate how many checksums we can fit into one leaf and then divide the
4473  * number of bytes that will need to be checksumed by this value to figure out
4474  * how many checksums will be required.  If we are adding bytes then the number
4475  * may go up and we will return the number of additional bytes that must be
4476  * reserved.  If it is going down we will return the number of bytes that must
4477  * be freed.
4478  *
4479  * This must be called with BTRFS_I(inode)->lock held.
4480  */
4481 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4482                                    int reserve)
4483 {
4484         struct btrfs_root *root = BTRFS_I(inode)->root;
4485         u64 csum_size;
4486         int num_csums_per_leaf;
4487         int num_csums;
4488         int old_csums;
4489
4490         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4491             BTRFS_I(inode)->csum_bytes == 0)
4492                 return 0;
4493
4494         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4495         if (reserve)
4496                 BTRFS_I(inode)->csum_bytes += num_bytes;
4497         else
4498                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4499         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4500         num_csums_per_leaf = (int)div64_u64(csum_size,
4501                                             sizeof(struct btrfs_csum_item) +
4502                                             sizeof(struct btrfs_disk_key));
4503         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4504         num_csums = num_csums + num_csums_per_leaf - 1;
4505         num_csums = num_csums / num_csums_per_leaf;
4506
4507         old_csums = old_csums + num_csums_per_leaf - 1;
4508         old_csums = old_csums / num_csums_per_leaf;
4509
4510         /* No change, no need to reserve more */
4511         if (old_csums == num_csums)
4512                 return 0;
4513
4514         if (reserve)
4515                 return btrfs_calc_trans_metadata_size(root,
4516                                                       num_csums - old_csums);
4517
4518         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4519 }
4520
4521 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4522 {
4523         struct btrfs_root *root = BTRFS_I(inode)->root;
4524         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4525         u64 to_reserve = 0;
4526         u64 csum_bytes;
4527         unsigned nr_extents = 0;
4528         int extra_reserve = 0;
4529         int flush = 1;
4530         int ret;
4531
4532         /* Need to be holding the i_mutex here if we aren't free space cache */
4533         if (btrfs_is_free_space_inode(inode))
4534                 flush = 0;
4535
4536         if (flush && btrfs_transaction_in_commit(root->fs_info))
4537                 schedule_timeout(1);
4538
4539         mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4540         num_bytes = ALIGN(num_bytes, root->sectorsize);
4541
4542         spin_lock(&BTRFS_I(inode)->lock);
4543         BTRFS_I(inode)->outstanding_extents++;
4544
4545         if (BTRFS_I(inode)->outstanding_extents >
4546             BTRFS_I(inode)->reserved_extents)
4547                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4548                         BTRFS_I(inode)->reserved_extents;
4549
4550         /*
4551          * Add an item to reserve for updating the inode when we complete the
4552          * delalloc io.
4553          */
4554         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4555                       &BTRFS_I(inode)->runtime_flags)) {
4556                 nr_extents++;
4557                 extra_reserve = 1;
4558         }
4559
4560         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4561         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4562         csum_bytes = BTRFS_I(inode)->csum_bytes;
4563         spin_unlock(&BTRFS_I(inode)->lock);
4564
4565         if (root->fs_info->quota_enabled) {
4566                 ret = btrfs_qgroup_reserve(root, num_bytes +
4567                                            nr_extents * root->leafsize);
4568                 if (ret) {
4569                         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4570                         return ret;
4571                 }
4572         }
4573
4574         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4575         if (ret) {
4576                 u64 to_free = 0;
4577                 unsigned dropped;
4578
4579                 spin_lock(&BTRFS_I(inode)->lock);
4580                 dropped = drop_outstanding_extent(inode);
4581                 /*
4582                  * If the inodes csum_bytes is the same as the original
4583                  * csum_bytes then we know we haven't raced with any free()ers
4584                  * so we can just reduce our inodes csum bytes and carry on.
4585                  * Otherwise we have to do the normal free thing to account for
4586                  * the case that the free side didn't free up its reserve
4587                  * because of this outstanding reservation.
4588                  */
4589                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4590                         calc_csum_metadata_size(inode, num_bytes, 0);
4591                 else
4592                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4593                 spin_unlock(&BTRFS_I(inode)->lock);
4594                 if (dropped)
4595                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4596
4597                 if (to_free) {
4598                         btrfs_block_rsv_release(root, block_rsv, to_free);
4599                         trace_btrfs_space_reservation(root->fs_info,
4600                                                       "delalloc",
4601                                                       btrfs_ino(inode),
4602                                                       to_free, 0);
4603                 }
4604                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4605                 return ret;
4606         }
4607
4608         spin_lock(&BTRFS_I(inode)->lock);
4609         if (extra_reserve) {
4610                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4611                         &BTRFS_I(inode)->runtime_flags);
4612                 nr_extents--;
4613         }
4614         BTRFS_I(inode)->reserved_extents += nr_extents;
4615         spin_unlock(&BTRFS_I(inode)->lock);
4616         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4617
4618         if (to_reserve)
4619                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4620                                               btrfs_ino(inode), to_reserve, 1);
4621         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4622
4623         return 0;
4624 }
4625
4626 /**
4627  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4628  * @inode: the inode to release the reservation for
4629  * @num_bytes: the number of bytes we're releasing
4630  *
4631  * This will release the metadata reservation for an inode.  This can be called
4632  * once we complete IO for a given set of bytes to release their metadata
4633  * reservations.
4634  */
4635 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4636 {
4637         struct btrfs_root *root = BTRFS_I(inode)->root;
4638         u64 to_free = 0;
4639         unsigned dropped;
4640
4641         num_bytes = ALIGN(num_bytes, root->sectorsize);
4642         spin_lock(&BTRFS_I(inode)->lock);
4643         dropped = drop_outstanding_extent(inode);
4644
4645         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4646         spin_unlock(&BTRFS_I(inode)->lock);
4647         if (dropped > 0)
4648                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4649
4650         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4651                                       btrfs_ino(inode), to_free, 0);
4652         if (root->fs_info->quota_enabled) {
4653                 btrfs_qgroup_free(root, num_bytes +
4654                                         dropped * root->leafsize);
4655         }
4656
4657         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4658                                 to_free);
4659 }
4660
4661 /**
4662  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4663  * @inode: inode we're writing to
4664  * @num_bytes: the number of bytes we want to allocate
4665  *
4666  * This will do the following things
4667  *
4668  * o reserve space in the data space info for num_bytes
4669  * o reserve space in the metadata space info based on number of outstanding
4670  *   extents and how much csums will be needed
4671  * o add to the inodes ->delalloc_bytes
4672  * o add it to the fs_info's delalloc inodes list.
4673  *
4674  * This will return 0 for success and -ENOSPC if there is no space left.
4675  */
4676 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4677 {
4678         int ret;
4679
4680         ret = btrfs_check_data_free_space(inode, num_bytes);
4681         if (ret)
4682                 return ret;
4683
4684         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4685         if (ret) {
4686                 btrfs_free_reserved_data_space(inode, num_bytes);
4687                 return ret;
4688         }
4689
4690         return 0;
4691 }
4692
4693 /**
4694  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4695  * @inode: inode we're releasing space for
4696  * @num_bytes: the number of bytes we want to free up
4697  *
4698  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4699  * called in the case that we don't need the metadata AND data reservations
4700  * anymore.  So if there is an error or we insert an inline extent.
4701  *
4702  * This function will release the metadata space that was not used and will
4703  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4704  * list if there are no delalloc bytes left.
4705  */
4706 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4707 {
4708         btrfs_delalloc_release_metadata(inode, num_bytes);
4709         btrfs_free_reserved_data_space(inode, num_bytes);
4710 }
4711
4712 static int update_block_group(struct btrfs_trans_handle *trans,
4713                               struct btrfs_root *root,
4714                               u64 bytenr, u64 num_bytes, int alloc)
4715 {
4716         struct btrfs_block_group_cache *cache = NULL;
4717         struct btrfs_fs_info *info = root->fs_info;
4718         u64 total = num_bytes;
4719         u64 old_val;
4720         u64 byte_in_group;
4721         int factor;
4722
4723         /* block accounting for super block */
4724         spin_lock(&info->delalloc_lock);
4725         old_val = btrfs_super_bytes_used(info->super_copy);
4726         if (alloc)
4727                 old_val += num_bytes;
4728         else
4729                 old_val -= num_bytes;
4730         btrfs_set_super_bytes_used(info->super_copy, old_val);
4731         spin_unlock(&info->delalloc_lock);
4732
4733         while (total) {
4734                 cache = btrfs_lookup_block_group(info, bytenr);
4735                 if (!cache)
4736                         return -ENOENT;
4737                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4738                                     BTRFS_BLOCK_GROUP_RAID1 |
4739                                     BTRFS_BLOCK_GROUP_RAID10))
4740                         factor = 2;
4741                 else
4742                         factor = 1;
4743                 /*
4744                  * If this block group has free space cache written out, we
4745                  * need to make sure to load it if we are removing space.  This
4746                  * is because we need the unpinning stage to actually add the
4747                  * space back to the block group, otherwise we will leak space.
4748                  */
4749                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4750                         cache_block_group(cache, trans, NULL, 1);
4751
4752                 byte_in_group = bytenr - cache->key.objectid;
4753                 WARN_ON(byte_in_group > cache->key.offset);
4754
4755                 spin_lock(&cache->space_info->lock);
4756                 spin_lock(&cache->lock);
4757
4758                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4759                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4760                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4761
4762                 cache->dirty = 1;
4763                 old_val = btrfs_block_group_used(&cache->item);
4764                 num_bytes = min(total, cache->key.offset - byte_in_group);
4765                 if (alloc) {
4766                         old_val += num_bytes;
4767                         btrfs_set_block_group_used(&cache->item, old_val);
4768                         cache->reserved -= num_bytes;
4769                         cache->space_info->bytes_reserved -= num_bytes;
4770                         cache->space_info->bytes_used += num_bytes;
4771                         cache->space_info->disk_used += num_bytes * factor;
4772                         spin_unlock(&cache->lock);
4773                         spin_unlock(&cache->space_info->lock);
4774                 } else {
4775                         old_val -= num_bytes;
4776                         btrfs_set_block_group_used(&cache->item, old_val);
4777                         cache->pinned += num_bytes;
4778                         cache->space_info->bytes_pinned += num_bytes;
4779                         cache->space_info->bytes_used -= num_bytes;
4780                         cache->space_info->disk_used -= num_bytes * factor;
4781                         spin_unlock(&cache->lock);
4782                         spin_unlock(&cache->space_info->lock);
4783
4784                         set_extent_dirty(info->pinned_extents,
4785                                          bytenr, bytenr + num_bytes - 1,
4786                                          GFP_NOFS | __GFP_NOFAIL);
4787                 }
4788                 btrfs_put_block_group(cache);
4789                 total -= num_bytes;
4790                 bytenr += num_bytes;
4791         }
4792         return 0;
4793 }
4794
4795 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4796 {
4797         struct btrfs_block_group_cache *cache;
4798         u64 bytenr;
4799
4800         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4801         if (!cache)
4802                 return 0;
4803
4804         bytenr = cache->key.objectid;
4805         btrfs_put_block_group(cache);
4806
4807         return bytenr;
4808 }
4809
4810 static int pin_down_extent(struct btrfs_root *root,
4811                            struct btrfs_block_group_cache *cache,
4812                            u64 bytenr, u64 num_bytes, int reserved)
4813 {
4814         spin_lock(&cache->space_info->lock);
4815         spin_lock(&cache->lock);
4816         cache->pinned += num_bytes;
4817         cache->space_info->bytes_pinned += num_bytes;
4818         if (reserved) {
4819                 cache->reserved -= num_bytes;
4820                 cache->space_info->bytes_reserved -= num_bytes;
4821         }
4822         spin_unlock(&cache->lock);
4823         spin_unlock(&cache->space_info->lock);
4824
4825         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4826                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4827         return 0;
4828 }
4829
4830 /*
4831  * this function must be called within transaction
4832  */
4833 int btrfs_pin_extent(struct btrfs_root *root,
4834                      u64 bytenr, u64 num_bytes, int reserved)
4835 {
4836         struct btrfs_block_group_cache *cache;
4837
4838         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4839         BUG_ON(!cache); /* Logic error */
4840
4841         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4842
4843         btrfs_put_block_group(cache);
4844         return 0;
4845 }
4846
4847 /*
4848  * this function must be called within transaction
4849  */
4850 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4851                                     struct btrfs_root *root,
4852                                     u64 bytenr, u64 num_bytes)
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         /*
4860          * pull in the free space cache (if any) so that our pin
4861          * removes the free space from the cache.  We have load_only set
4862          * to one because the slow code to read in the free extents does check
4863          * the pinned extents.
4864          */
4865         cache_block_group(cache, trans, root, 1);
4866
4867         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4868
4869         /* remove us from the free space cache (if we're there at all) */
4870         btrfs_remove_free_space(cache, bytenr, num_bytes);
4871         btrfs_put_block_group(cache);
4872         return 0;
4873 }
4874
4875 /**
4876  * btrfs_update_reserved_bytes - update the block_group and space info counters
4877  * @cache:      The cache we are manipulating
4878  * @num_bytes:  The number of bytes in question
4879  * @reserve:    One of the reservation enums
4880  *
4881  * This is called by the allocator when it reserves space, or by somebody who is
4882  * freeing space that was never actually used on disk.  For example if you
4883  * reserve some space for a new leaf in transaction A and before transaction A
4884  * commits you free that leaf, you call this with reserve set to 0 in order to
4885  * clear the reservation.
4886  *
4887  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4888  * ENOSPC accounting.  For data we handle the reservation through clearing the
4889  * delalloc bits in the io_tree.  We have to do this since we could end up
4890  * allocating less disk space for the amount of data we have reserved in the
4891  * case of compression.
4892  *
4893  * If this is a reservation and the block group has become read only we cannot
4894  * make the reservation and return -EAGAIN, otherwise this function always
4895  * succeeds.
4896  */
4897 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4898                                        u64 num_bytes, int reserve)
4899 {
4900         struct btrfs_space_info *space_info = cache->space_info;
4901         int ret = 0;
4902
4903         spin_lock(&space_info->lock);
4904         spin_lock(&cache->lock);
4905         if (reserve != RESERVE_FREE) {
4906                 if (cache->ro) {
4907                         ret = -EAGAIN;
4908                 } else {
4909                         cache->reserved += num_bytes;
4910                         space_info->bytes_reserved += num_bytes;
4911                         if (reserve == RESERVE_ALLOC) {
4912                                 trace_btrfs_space_reservation(cache->fs_info,
4913                                                 "space_info", space_info->flags,
4914                                                 num_bytes, 0);
4915                                 space_info->bytes_may_use -= num_bytes;
4916                         }
4917                 }
4918         } else {
4919                 if (cache->ro)
4920                         space_info->bytes_readonly += num_bytes;
4921                 cache->reserved -= num_bytes;
4922                 space_info->bytes_reserved -= num_bytes;
4923                 space_info->reservation_progress++;
4924         }
4925         spin_unlock(&cache->lock);
4926         spin_unlock(&space_info->lock);
4927         return ret;
4928 }
4929
4930 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4931                                 struct btrfs_root *root)
4932 {
4933         struct btrfs_fs_info *fs_info = root->fs_info;
4934         struct btrfs_caching_control *next;
4935         struct btrfs_caching_control *caching_ctl;
4936         struct btrfs_block_group_cache *cache;
4937
4938         down_write(&fs_info->extent_commit_sem);
4939
4940         list_for_each_entry_safe(caching_ctl, next,
4941                                  &fs_info->caching_block_groups, list) {
4942                 cache = caching_ctl->block_group;
4943                 if (block_group_cache_done(cache)) {
4944                         cache->last_byte_to_unpin = (u64)-1;
4945                         list_del_init(&caching_ctl->list);
4946                         put_caching_control(caching_ctl);
4947                 } else {
4948                         cache->last_byte_to_unpin = caching_ctl->progress;
4949                 }
4950         }
4951
4952         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4953                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4954         else
4955                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4956
4957         up_write(&fs_info->extent_commit_sem);
4958
4959         update_global_block_rsv(fs_info);
4960 }
4961
4962 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4963 {
4964         struct btrfs_fs_info *fs_info = root->fs_info;
4965         struct btrfs_block_group_cache *cache = NULL;
4966         u64 len;
4967
4968         while (start <= end) {
4969                 if (!cache ||
4970                     start >= cache->key.objectid + cache->key.offset) {
4971                         if (cache)
4972                                 btrfs_put_block_group(cache);
4973                         cache = btrfs_lookup_block_group(fs_info, start);
4974                         BUG_ON(!cache); /* Logic error */
4975                 }
4976
4977                 len = cache->key.objectid + cache->key.offset - start;
4978                 len = min(len, end + 1 - start);
4979
4980                 if (start < cache->last_byte_to_unpin) {
4981                         len = min(len, cache->last_byte_to_unpin - start);
4982                         btrfs_add_free_space(cache, start, len);
4983                 }
4984
4985                 start += len;
4986
4987                 spin_lock(&cache->space_info->lock);
4988                 spin_lock(&cache->lock);
4989                 cache->pinned -= len;
4990                 cache->space_info->bytes_pinned -= len;
4991                 if (cache->ro)
4992                         cache->space_info->bytes_readonly += len;
4993                 spin_unlock(&cache->lock);
4994                 spin_unlock(&cache->space_info->lock);
4995         }
4996
4997         if (cache)
4998                 btrfs_put_block_group(cache);
4999         return 0;
5000 }
5001
5002 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5003                                struct btrfs_root *root)
5004 {
5005         struct btrfs_fs_info *fs_info = root->fs_info;
5006         struct extent_io_tree *unpin;
5007         u64 start;
5008         u64 end;
5009         int ret;
5010
5011         if (trans->aborted)
5012                 return 0;
5013
5014         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5015                 unpin = &fs_info->freed_extents[1];
5016         else
5017                 unpin = &fs_info->freed_extents[0];
5018
5019         while (1) {
5020                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5021                                             EXTENT_DIRTY);
5022                 if (ret)
5023                         break;
5024
5025                 if (btrfs_test_opt(root, DISCARD))
5026                         ret = btrfs_discard_extent(root, start,
5027                                                    end + 1 - start, NULL);
5028
5029                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5030                 unpin_extent_range(root, start, end);
5031                 cond_resched();
5032         }
5033
5034         return 0;
5035 }
5036
5037 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5038                                 struct btrfs_root *root,
5039                                 u64 bytenr, u64 num_bytes, u64 parent,
5040                                 u64 root_objectid, u64 owner_objectid,
5041                                 u64 owner_offset, int refs_to_drop,
5042                                 struct btrfs_delayed_extent_op *extent_op)
5043 {
5044         struct btrfs_key key;
5045         struct btrfs_path *path;
5046         struct btrfs_fs_info *info = root->fs_info;
5047         struct btrfs_root *extent_root = info->extent_root;
5048         struct extent_buffer *leaf;
5049         struct btrfs_extent_item *ei;
5050         struct btrfs_extent_inline_ref *iref;
5051         int ret;
5052         int is_data;
5053         int extent_slot = 0;
5054         int found_extent = 0;
5055         int num_to_del = 1;
5056         u32 item_size;
5057         u64 refs;
5058
5059         path = btrfs_alloc_path();
5060         if (!path)
5061                 return -ENOMEM;
5062
5063         path->reada = 1;
5064         path->leave_spinning = 1;
5065
5066         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5067         BUG_ON(!is_data && refs_to_drop != 1);
5068
5069         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5070                                     bytenr, num_bytes, parent,
5071                                     root_objectid, owner_objectid,
5072                                     owner_offset);
5073         if (ret == 0) {
5074                 extent_slot = path->slots[0];
5075                 while (extent_slot >= 0) {
5076                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5077                                               extent_slot);
5078                         if (key.objectid != bytenr)
5079                                 break;
5080                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5081                             key.offset == num_bytes) {
5082                                 found_extent = 1;
5083                                 break;
5084                         }
5085                         if (path->slots[0] - extent_slot > 5)
5086                                 break;
5087                         extent_slot--;
5088                 }
5089 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5090                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5091                 if (found_extent && item_size < sizeof(*ei))
5092                         found_extent = 0;
5093 #endif
5094                 if (!found_extent) {
5095                         BUG_ON(iref);
5096                         ret = remove_extent_backref(trans, extent_root, path,
5097                                                     NULL, refs_to_drop,
5098                                                     is_data);
5099                         if (ret)
5100                                 goto abort;
5101                         btrfs_release_path(path);
5102                         path->leave_spinning = 1;
5103
5104                         key.objectid = bytenr;
5105                         key.type = BTRFS_EXTENT_ITEM_KEY;
5106                         key.offset = num_bytes;
5107
5108                         ret = btrfs_search_slot(trans, extent_root,
5109                                                 &key, path, -1, 1);
5110                         if (ret) {
5111                                 printk(KERN_ERR "umm, got %d back from search"
5112                                        ", was looking for %llu\n", ret,
5113                                        (unsigned long long)bytenr);
5114                                 if (ret > 0)
5115                                         btrfs_print_leaf(extent_root,
5116                                                          path->nodes[0]);
5117                         }
5118                         if (ret < 0)
5119                                 goto abort;
5120                         extent_slot = path->slots[0];
5121                 }
5122         } else if (ret == -ENOENT) {
5123                 btrfs_print_leaf(extent_root, path->nodes[0]);
5124                 WARN_ON(1);
5125                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
5126                        "parent %llu root %llu  owner %llu offset %llu\n",
5127                        (unsigned long long)bytenr,
5128                        (unsigned long long)parent,
5129                        (unsigned long long)root_objectid,
5130                        (unsigned long long)owner_objectid,
5131                        (unsigned long long)owner_offset);
5132         } else {
5133                 goto abort;
5134         }
5135
5136         leaf = path->nodes[0];
5137         item_size = btrfs_item_size_nr(leaf, extent_slot);
5138 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5139         if (item_size < sizeof(*ei)) {
5140                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5141                 ret = convert_extent_item_v0(trans, extent_root, path,
5142                                              owner_objectid, 0);
5143                 if (ret < 0)
5144                         goto abort;
5145
5146                 btrfs_release_path(path);
5147                 path->leave_spinning = 1;
5148
5149                 key.objectid = bytenr;
5150                 key.type = BTRFS_EXTENT_ITEM_KEY;
5151                 key.offset = num_bytes;
5152
5153                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5154                                         -1, 1);
5155                 if (ret) {
5156                         printk(KERN_ERR "umm, got %d back from search"
5157                                ", was looking for %llu\n", ret,
5158                                (unsigned long long)bytenr);
5159                         btrfs_print_leaf(extent_root, path->nodes[0]);
5160                 }
5161                 if (ret < 0)
5162                         goto abort;
5163                 extent_slot = path->slots[0];
5164                 leaf = path->nodes[0];
5165                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5166         }
5167 #endif
5168         BUG_ON(item_size < sizeof(*ei));
5169         ei = btrfs_item_ptr(leaf, extent_slot,
5170                             struct btrfs_extent_item);
5171         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5172                 struct btrfs_tree_block_info *bi;
5173                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5174                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5175                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5176         }
5177
5178         refs = btrfs_extent_refs(leaf, ei);
5179         BUG_ON(refs < refs_to_drop);
5180         refs -= refs_to_drop;
5181
5182         if (refs > 0) {
5183                 if (extent_op)
5184                         __run_delayed_extent_op(extent_op, leaf, ei);
5185                 /*
5186                  * In the case of inline back ref, reference count will
5187                  * be updated by remove_extent_backref
5188                  */
5189                 if (iref) {
5190                         BUG_ON(!found_extent);
5191                 } else {
5192                         btrfs_set_extent_refs(leaf, ei, refs);
5193                         btrfs_mark_buffer_dirty(leaf);
5194                 }
5195                 if (found_extent) {
5196                         ret = remove_extent_backref(trans, extent_root, path,
5197                                                     iref, refs_to_drop,
5198                                                     is_data);
5199                         if (ret)
5200                                 goto abort;
5201                 }
5202         } else {
5203                 if (found_extent) {
5204                         BUG_ON(is_data && refs_to_drop !=
5205                                extent_data_ref_count(root, path, iref));
5206                         if (iref) {
5207                                 BUG_ON(path->slots[0] != extent_slot);
5208                         } else {
5209                                 BUG_ON(path->slots[0] != extent_slot + 1);
5210                                 path->slots[0] = extent_slot;
5211                                 num_to_del = 2;
5212                         }
5213                 }
5214
5215                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5216                                       num_to_del);
5217                 if (ret)
5218                         goto abort;
5219                 btrfs_release_path(path);
5220
5221                 if (is_data) {
5222                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5223                         if (ret)
5224                                 goto abort;
5225                 }
5226
5227                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
5228                 if (ret)
5229                         goto abort;
5230         }
5231 out:
5232         btrfs_free_path(path);
5233         return ret;
5234
5235 abort:
5236         btrfs_abort_transaction(trans, extent_root, ret);
5237         goto out;
5238 }
5239
5240 /*
5241  * when we free an block, it is possible (and likely) that we free the last
5242  * delayed ref for that extent as well.  This searches the delayed ref tree for
5243  * a given extent, and if there are no other delayed refs to be processed, it
5244  * removes it from the tree.
5245  */
5246 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5247                                       struct btrfs_root *root, u64 bytenr)
5248 {
5249         struct btrfs_delayed_ref_head *head;
5250         struct btrfs_delayed_ref_root *delayed_refs;
5251         struct btrfs_delayed_ref_node *ref;
5252         struct rb_node *node;
5253         int ret = 0;
5254
5255         delayed_refs = &trans->transaction->delayed_refs;
5256         spin_lock(&delayed_refs->lock);
5257         head = btrfs_find_delayed_ref_head(trans, bytenr);
5258         if (!head)
5259                 goto out;
5260
5261         node = rb_prev(&head->node.rb_node);
5262         if (!node)
5263                 goto out;
5264
5265         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5266
5267         /* there are still entries for this ref, we can't drop it */
5268         if (ref->bytenr == bytenr)
5269                 goto out;
5270
5271         if (head->extent_op) {
5272                 if (!head->must_insert_reserved)
5273                         goto out;
5274                 kfree(head->extent_op);
5275                 head->extent_op = NULL;
5276         }
5277
5278         /*
5279          * waiting for the lock here would deadlock.  If someone else has it
5280          * locked they are already in the process of dropping it anyway
5281          */
5282         if (!mutex_trylock(&head->mutex))
5283                 goto out;
5284
5285         /*
5286          * at this point we have a head with no other entries.  Go
5287          * ahead and process it.
5288          */
5289         head->node.in_tree = 0;
5290         rb_erase(&head->node.rb_node, &delayed_refs->root);
5291
5292         delayed_refs->num_entries--;
5293
5294         /*
5295          * we don't take a ref on the node because we're removing it from the
5296          * tree, so we just steal the ref the tree was holding.
5297          */
5298         delayed_refs->num_heads--;
5299         if (list_empty(&head->cluster))
5300                 delayed_refs->num_heads_ready--;
5301
5302         list_del_init(&head->cluster);
5303         spin_unlock(&delayed_refs->lock);
5304
5305         BUG_ON(head->extent_op);
5306         if (head->must_insert_reserved)
5307                 ret = 1;
5308
5309         mutex_unlock(&head->mutex);
5310         btrfs_put_delayed_ref(&head->node);
5311         return ret;
5312 out:
5313         spin_unlock(&delayed_refs->lock);
5314         return 0;
5315 }
5316
5317 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5318                            struct btrfs_root *root,
5319                            struct extent_buffer *buf,
5320                            u64 parent, int last_ref)
5321 {
5322         struct btrfs_block_group_cache *cache = NULL;
5323         int ret;
5324
5325         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5326                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5327                                         buf->start, buf->len,
5328                                         parent, root->root_key.objectid,
5329                                         btrfs_header_level(buf),
5330                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5331                 BUG_ON(ret); /* -ENOMEM */
5332         }
5333
5334         if (!last_ref)
5335                 return;
5336
5337         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5338
5339         if (btrfs_header_generation(buf) == trans->transid) {
5340                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5341                         ret = check_ref_cleanup(trans, root, buf->start);
5342                         if (!ret)
5343                                 goto out;
5344                 }
5345
5346                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5347                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5348                         goto out;
5349                 }
5350
5351                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5352
5353                 btrfs_add_free_space(cache, buf->start, buf->len);
5354                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5355         }
5356 out:
5357         /*
5358          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5359          * anymore.
5360          */
5361         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5362         btrfs_put_block_group(cache);
5363 }
5364
5365 /* Can return -ENOMEM */
5366 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5367                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5368                       u64 owner, u64 offset, int for_cow)
5369 {
5370         int ret;
5371         struct btrfs_fs_info *fs_info = root->fs_info;
5372
5373         /*
5374          * tree log blocks never actually go into the extent allocation
5375          * tree, just update pinning info and exit early.
5376          */
5377         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5378                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5379                 /* unlocks the pinned mutex */
5380                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5381                 ret = 0;
5382         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5383                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5384                                         num_bytes,
5385                                         parent, root_objectid, (int)owner,
5386                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5387         } else {
5388                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5389                                                 num_bytes,
5390                                                 parent, root_objectid, owner,
5391                                                 offset, BTRFS_DROP_DELAYED_REF,
5392                                                 NULL, for_cow);
5393         }
5394         return ret;
5395 }
5396
5397 static u64 stripe_align(struct btrfs_root *root, u64 val)
5398 {
5399         u64 mask = ((u64)root->stripesize - 1);
5400         u64 ret = (val + mask) & ~mask;
5401         return ret;
5402 }
5403
5404 /*
5405  * when we wait for progress in the block group caching, its because
5406  * our allocation attempt failed at least once.  So, we must sleep
5407  * and let some progress happen before we try again.
5408  *
5409  * This function will sleep at least once waiting for new free space to
5410  * show up, and then it will check the block group free space numbers
5411  * for our min num_bytes.  Another option is to have it go ahead
5412  * and look in the rbtree for a free extent of a given size, but this
5413  * is a good start.
5414  */
5415 static noinline int
5416 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5417                                 u64 num_bytes)
5418 {
5419         struct btrfs_caching_control *caching_ctl;
5420         DEFINE_WAIT(wait);
5421
5422         caching_ctl = get_caching_control(cache);
5423         if (!caching_ctl)
5424                 return 0;
5425
5426         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5427                    (cache->free_space_ctl->free_space >= num_bytes));
5428
5429         put_caching_control(caching_ctl);
5430         return 0;
5431 }
5432
5433 static noinline int
5434 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5435 {
5436         struct btrfs_caching_control *caching_ctl;
5437         DEFINE_WAIT(wait);
5438
5439         caching_ctl = get_caching_control(cache);
5440         if (!caching_ctl)
5441                 return 0;
5442
5443         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5444
5445         put_caching_control(caching_ctl);
5446         return 0;
5447 }
5448
5449 static int __get_block_group_index(u64 flags)
5450 {
5451         int index;
5452
5453         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5454                 index = 0;
5455         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5456                 index = 1;
5457         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5458                 index = 2;
5459         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5460                 index = 3;
5461         else
5462                 index = 4;
5463
5464         return index;
5465 }
5466
5467 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5468 {
5469         return __get_block_group_index(cache->flags);
5470 }
5471
5472 enum btrfs_loop_type {
5473         LOOP_CACHING_NOWAIT = 0,
5474         LOOP_CACHING_WAIT = 1,
5475         LOOP_ALLOC_CHUNK = 2,
5476         LOOP_NO_EMPTY_SIZE = 3,
5477 };
5478
5479 /*
5480  * walks the btree of allocated extents and find a hole of a given size.
5481  * The key ins is changed to record the hole:
5482  * ins->objectid == block start
5483  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5484  * ins->offset == number of blocks
5485  * Any available blocks before search_start are skipped.
5486  */
5487 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5488                                      struct btrfs_root *orig_root,
5489                                      u64 num_bytes, u64 empty_size,
5490                                      u64 hint_byte, struct btrfs_key *ins,
5491                                      u64 data)
5492 {
5493         int ret = 0;
5494         struct btrfs_root *root = orig_root->fs_info->extent_root;
5495         struct btrfs_free_cluster *last_ptr = NULL;
5496         struct btrfs_block_group_cache *block_group = NULL;
5497         struct btrfs_block_group_cache *used_block_group;
5498         u64 search_start = 0;
5499         int empty_cluster = 2 * 1024 * 1024;
5500         int allowed_chunk_alloc = 0;
5501         int done_chunk_alloc = 0;
5502         struct btrfs_space_info *space_info;
5503         int loop = 0;
5504         int index = 0;
5505         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5506                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5507         bool found_uncached_bg = false;
5508         bool failed_cluster_refill = false;
5509         bool failed_alloc = false;
5510         bool use_cluster = true;
5511         bool have_caching_bg = false;
5512
5513         WARN_ON(num_bytes < root->sectorsize);
5514         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5515         ins->objectid = 0;
5516         ins->offset = 0;
5517
5518         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5519
5520         space_info = __find_space_info(root->fs_info, data);
5521         if (!space_info) {
5522                 printk(KERN_ERR "No space info for %llu\n", data);
5523                 return -ENOSPC;
5524         }
5525
5526         /*
5527          * If the space info is for both data and metadata it means we have a
5528          * small filesystem and we can't use the clustering stuff.
5529          */
5530         if (btrfs_mixed_space_info(space_info))
5531                 use_cluster = false;
5532
5533         if (orig_root->ref_cows || empty_size)
5534                 allowed_chunk_alloc = 1;
5535
5536         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5537                 last_ptr = &root->fs_info->meta_alloc_cluster;
5538                 if (!btrfs_test_opt(root, SSD))
5539                         empty_cluster = 64 * 1024;
5540         }
5541
5542         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5543             btrfs_test_opt(root, SSD)) {
5544                 last_ptr = &root->fs_info->data_alloc_cluster;
5545         }
5546
5547         if (last_ptr) {
5548                 spin_lock(&last_ptr->lock);
5549                 if (last_ptr->block_group)
5550                         hint_byte = last_ptr->window_start;
5551                 spin_unlock(&last_ptr->lock);
5552         }
5553
5554         search_start = max(search_start, first_logical_byte(root, 0));
5555         search_start = max(search_start, hint_byte);
5556
5557         if (!last_ptr)
5558                 empty_cluster = 0;
5559
5560         if (search_start == hint_byte) {
5561                 block_group = btrfs_lookup_block_group(root->fs_info,
5562                                                        search_start);
5563                 used_block_group = block_group;
5564                 /*
5565                  * we don't want to use the block group if it doesn't match our
5566                  * allocation bits, or if its not cached.
5567                  *
5568                  * However if we are re-searching with an ideal block group
5569                  * picked out then we don't care that the block group is cached.
5570                  */
5571                 if (block_group && block_group_bits(block_group, data) &&
5572                     block_group->cached != BTRFS_CACHE_NO) {
5573                         down_read(&space_info->groups_sem);
5574                         if (list_empty(&block_group->list) ||
5575                             block_group->ro) {
5576                                 /*
5577                                  * someone is removing this block group,
5578                                  * we can't jump into the have_block_group
5579                                  * target because our list pointers are not
5580                                  * valid
5581                                  */
5582                                 btrfs_put_block_group(block_group);
5583                                 up_read(&space_info->groups_sem);
5584                         } else {
5585                                 index = get_block_group_index(block_group);
5586                                 goto have_block_group;
5587                         }
5588                 } else if (block_group) {
5589                         btrfs_put_block_group(block_group);
5590                 }
5591         }
5592 search:
5593         have_caching_bg = false;
5594         down_read(&space_info->groups_sem);
5595         list_for_each_entry(block_group, &space_info->block_groups[index],
5596                             list) {
5597                 u64 offset;
5598                 int cached;
5599
5600                 used_block_group = block_group;
5601                 btrfs_get_block_group(block_group);
5602                 search_start = block_group->key.objectid;
5603
5604                 /*
5605                  * this can happen if we end up cycling through all the
5606                  * raid types, but we want to make sure we only allocate
5607                  * for the proper type.
5608                  */
5609                 if (!block_group_bits(block_group, data)) {
5610                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5611                                 BTRFS_BLOCK_GROUP_RAID1 |
5612                                 BTRFS_BLOCK_GROUP_RAID10;
5613
5614                         /*
5615                          * if they asked for extra copies and this block group
5616                          * doesn't provide them, bail.  This does allow us to
5617                          * fill raid0 from raid1.
5618                          */
5619                         if ((data & extra) && !(block_group->flags & extra))
5620                                 goto loop;
5621                 }
5622
5623 have_block_group:
5624                 cached = block_group_cache_done(block_group);
5625                 if (unlikely(!cached)) {
5626                         found_uncached_bg = true;
5627                         ret = cache_block_group(block_group, trans,
5628                                                 orig_root, 0);
5629                         BUG_ON(ret < 0);
5630                         ret = 0;
5631                 }
5632
5633                 if (unlikely(block_group->ro))
5634                         goto loop;
5635
5636                 /*
5637                  * Ok we want to try and use the cluster allocator, so
5638                  * lets look there
5639                  */
5640                 if (last_ptr) {
5641                         /*
5642                          * the refill lock keeps out other
5643                          * people trying to start a new cluster
5644                          */
5645                         spin_lock(&last_ptr->refill_lock);
5646                         used_block_group = last_ptr->block_group;
5647                         if (used_block_group != block_group &&
5648                             (!used_block_group ||
5649                              used_block_group->ro ||
5650                              !block_group_bits(used_block_group, data))) {
5651                                 used_block_group = block_group;
5652                                 goto refill_cluster;
5653                         }
5654
5655                         if (used_block_group != block_group)
5656                                 btrfs_get_block_group(used_block_group);
5657
5658                         offset = btrfs_alloc_from_cluster(used_block_group,
5659                           last_ptr, num_bytes, used_block_group->key.objectid);
5660                         if (offset) {
5661                                 /* we have a block, we're done */
5662                                 spin_unlock(&last_ptr->refill_lock);
5663                                 trace_btrfs_reserve_extent_cluster(root,
5664                                         block_group, search_start, num_bytes);
5665                                 goto checks;
5666                         }
5667
5668                         WARN_ON(last_ptr->block_group != used_block_group);
5669                         if (used_block_group != block_group) {
5670                                 btrfs_put_block_group(used_block_group);
5671                                 used_block_group = block_group;
5672                         }
5673 refill_cluster:
5674                         BUG_ON(used_block_group != block_group);
5675                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5676                          * set up a new clusters, so lets just skip it
5677                          * and let the allocator find whatever block
5678                          * it can find.  If we reach this point, we
5679                          * will have tried the cluster allocator
5680                          * plenty of times and not have found
5681                          * anything, so we are likely way too
5682                          * fragmented for the clustering stuff to find
5683                          * anything.
5684                          *
5685                          * However, if the cluster is taken from the
5686                          * current block group, release the cluster
5687                          * first, so that we stand a better chance of
5688                          * succeeding in the unclustered
5689                          * allocation.  */
5690                         if (loop >= LOOP_NO_EMPTY_SIZE &&
5691                             last_ptr->block_group != block_group) {
5692                                 spin_unlock(&last_ptr->refill_lock);
5693                                 goto unclustered_alloc;
5694                         }
5695
5696                         /*
5697                          * this cluster didn't work out, free it and
5698                          * start over
5699                          */
5700                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5701
5702                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5703                                 spin_unlock(&last_ptr->refill_lock);
5704                                 goto unclustered_alloc;
5705                         }
5706
5707                         /* allocate a cluster in this block group */
5708                         ret = btrfs_find_space_cluster(trans, root,
5709                                                block_group, last_ptr,
5710                                                search_start, num_bytes,
5711                                                empty_cluster + empty_size);
5712                         if (ret == 0) {
5713                                 /*
5714                                  * now pull our allocation out of this
5715                                  * cluster
5716                                  */
5717                                 offset = btrfs_alloc_from_cluster(block_group,
5718                                                   last_ptr, num_bytes,
5719                                                   search_start);
5720                                 if (offset) {
5721                                         /* we found one, proceed */
5722                                         spin_unlock(&last_ptr->refill_lock);
5723                                         trace_btrfs_reserve_extent_cluster(root,
5724                                                 block_group, search_start,
5725                                                 num_bytes);
5726                                         goto checks;
5727                                 }
5728                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5729                                    && !failed_cluster_refill) {
5730                                 spin_unlock(&last_ptr->refill_lock);
5731
5732                                 failed_cluster_refill = true;
5733                                 wait_block_group_cache_progress(block_group,
5734                                        num_bytes + empty_cluster + empty_size);
5735                                 goto have_block_group;
5736                         }
5737
5738                         /*
5739                          * at this point we either didn't find a cluster
5740                          * or we weren't able to allocate a block from our
5741                          * cluster.  Free the cluster we've been trying
5742                          * to use, and go to the next block group
5743                          */
5744                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5745                         spin_unlock(&last_ptr->refill_lock);
5746                         goto loop;
5747                 }
5748
5749 unclustered_alloc:
5750                 spin_lock(&block_group->free_space_ctl->tree_lock);
5751                 if (cached &&
5752                     block_group->free_space_ctl->free_space <
5753                     num_bytes + empty_cluster + empty_size) {
5754                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5755                         goto loop;
5756                 }
5757                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5758
5759                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5760                                                     num_bytes, empty_size);
5761                 /*
5762                  * If we didn't find a chunk, and we haven't failed on this
5763                  * block group before, and this block group is in the middle of
5764                  * caching and we are ok with waiting, then go ahead and wait
5765                  * for progress to be made, and set failed_alloc to true.
5766                  *
5767                  * If failed_alloc is true then we've already waited on this
5768                  * block group once and should move on to the next block group.
5769                  */
5770                 if (!offset && !failed_alloc && !cached &&
5771                     loop > LOOP_CACHING_NOWAIT) {
5772                         wait_block_group_cache_progress(block_group,
5773                                                 num_bytes + empty_size);
5774                         failed_alloc = true;
5775                         goto have_block_group;
5776                 } else if (!offset) {
5777                         if (!cached)
5778                                 have_caching_bg = true;
5779                         goto loop;
5780                 }
5781 checks:
5782                 search_start = stripe_align(root, offset);
5783
5784                 /* move on to the next group */
5785                 if (search_start + num_bytes >
5786                     used_block_group->key.objectid + used_block_group->key.offset) {
5787                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5788                         goto loop;
5789                 }
5790
5791                 if (offset < search_start)
5792                         btrfs_add_free_space(used_block_group, offset,
5793                                              search_start - offset);
5794                 BUG_ON(offset > search_start);
5795
5796                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
5797                                                   alloc_type);
5798                 if (ret == -EAGAIN) {
5799                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5800                         goto loop;
5801                 }
5802
5803                 /* we are all good, lets return */
5804                 ins->objectid = search_start;
5805                 ins->offset = num_bytes;
5806
5807                 trace_btrfs_reserve_extent(orig_root, block_group,
5808                                            search_start, num_bytes);
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                 if (used_block_group != block_group)
5814                         btrfs_put_block_group(used_block_group);
5815                 btrfs_put_block_group(block_group);
5816                 break;
5817 loop:
5818                 failed_cluster_refill = false;
5819                 failed_alloc = false;
5820                 BUG_ON(index != get_block_group_index(block_group));
5821                 if (used_block_group != block_group)
5822                         btrfs_put_block_group(used_block_group);
5823                 btrfs_put_block_group(block_group);
5824         }
5825         up_read(&space_info->groups_sem);
5826
5827         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
5828                 goto search;
5829
5830         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5831                 goto search;
5832
5833         /*
5834          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5835          *                      caching kthreads as we move along
5836          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5837          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5838          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5839          *                      again
5840          */
5841         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5842                 index = 0;
5843                 loop++;
5844                 if (loop == LOOP_ALLOC_CHUNK) {
5845                        if (allowed_chunk_alloc) {
5846                                 ret = do_chunk_alloc(trans, root, num_bytes +
5847                                                      2 * 1024 * 1024, data,
5848                                                      CHUNK_ALLOC_LIMITED);
5849                                 /*
5850                                  * Do not bail out on ENOSPC since we
5851                                  * can do more things.
5852                                  */
5853                                 if (ret < 0 && ret != -ENOSPC) {
5854                                         btrfs_abort_transaction(trans,
5855                                                                 root, ret);
5856                                         goto out;
5857                                 }
5858                                 allowed_chunk_alloc = 0;
5859                                 if (ret == 1)
5860                                         done_chunk_alloc = 1;
5861                         } else if (!done_chunk_alloc &&
5862                                    space_info->force_alloc ==
5863                                    CHUNK_ALLOC_NO_FORCE) {
5864                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5865                         }
5866
5867                        /*
5868                         * We didn't allocate a chunk, go ahead and drop the
5869                         * empty size and loop again.
5870                         */
5871                        if (!done_chunk_alloc)
5872                                loop = LOOP_NO_EMPTY_SIZE;
5873                 }
5874
5875                 if (loop == LOOP_NO_EMPTY_SIZE) {
5876                         empty_size = 0;
5877                         empty_cluster = 0;
5878                 }
5879
5880                 goto search;
5881         } else if (!ins->objectid) {
5882                 ret = -ENOSPC;
5883         } else if (ins->objectid) {
5884                 ret = 0;
5885         }
5886 out:
5887
5888         return ret;
5889 }
5890
5891 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5892                             int dump_block_groups)
5893 {
5894         struct btrfs_block_group_cache *cache;
5895         int index = 0;
5896
5897         spin_lock(&info->lock);
5898         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5899                (unsigned long long)info->flags,
5900                (unsigned long long)(info->total_bytes - info->bytes_used -
5901                                     info->bytes_pinned - info->bytes_reserved -
5902                                     info->bytes_readonly),
5903                (info->full) ? "" : "not ");
5904         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5905                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5906                (unsigned long long)info->total_bytes,
5907                (unsigned long long)info->bytes_used,
5908                (unsigned long long)info->bytes_pinned,
5909                (unsigned long long)info->bytes_reserved,
5910                (unsigned long long)info->bytes_may_use,
5911                (unsigned long long)info->bytes_readonly);
5912         spin_unlock(&info->lock);
5913
5914         if (!dump_block_groups)
5915                 return;
5916
5917         down_read(&info->groups_sem);
5918 again:
5919         list_for_each_entry(cache, &info->block_groups[index], list) {
5920                 spin_lock(&cache->lock);
5921                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s\n",
5922                        (unsigned long long)cache->key.objectid,
5923                        (unsigned long long)cache->key.offset,
5924                        (unsigned long long)btrfs_block_group_used(&cache->item),
5925                        (unsigned long long)cache->pinned,
5926                        (unsigned long long)cache->reserved,
5927                        cache->ro ? "[readonly]" : "");
5928                 btrfs_dump_free_space(cache, bytes);
5929                 spin_unlock(&cache->lock);
5930         }
5931         if (++index < BTRFS_NR_RAID_TYPES)
5932                 goto again;
5933         up_read(&info->groups_sem);
5934 }
5935
5936 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5937                          struct btrfs_root *root,
5938                          u64 num_bytes, u64 min_alloc_size,
5939                          u64 empty_size, u64 hint_byte,
5940                          struct btrfs_key *ins, u64 data)
5941 {
5942         bool final_tried = false;
5943         int ret;
5944
5945         data = btrfs_get_alloc_profile(root, data);
5946 again:
5947         /*
5948          * the only place that sets empty_size is btrfs_realloc_node, which
5949          * is not called recursively on allocations
5950          */
5951         if (empty_size || root->ref_cows) {
5952                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5953                                      num_bytes + 2 * 1024 * 1024, data,
5954                                      CHUNK_ALLOC_NO_FORCE);
5955                 if (ret < 0 && ret != -ENOSPC) {
5956                         btrfs_abort_transaction(trans, root, ret);
5957                         return ret;
5958                 }
5959         }
5960
5961         WARN_ON(num_bytes < root->sectorsize);
5962         ret = find_free_extent(trans, root, num_bytes, empty_size,
5963                                hint_byte, ins, data);
5964
5965         if (ret == -ENOSPC) {
5966                 if (!final_tried) {
5967                         num_bytes = num_bytes >> 1;
5968                         num_bytes = num_bytes & ~(root->sectorsize - 1);
5969                         num_bytes = max(num_bytes, min_alloc_size);
5970                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5971                                        num_bytes, data, CHUNK_ALLOC_FORCE);
5972                         if (ret < 0 && ret != -ENOSPC) {
5973                                 btrfs_abort_transaction(trans, root, ret);
5974                                 return ret;
5975                         }
5976                         if (num_bytes == min_alloc_size)
5977                                 final_tried = true;
5978                         goto again;
5979                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
5980                         struct btrfs_space_info *sinfo;
5981
5982                         sinfo = __find_space_info(root->fs_info, data);
5983                         printk(KERN_ERR "btrfs allocation failed flags %llu, "
5984                                "wanted %llu\n", (unsigned long long)data,
5985                                (unsigned long long)num_bytes);
5986                         if (sinfo)
5987                                 dump_space_info(sinfo, num_bytes, 1);
5988                 }
5989         }
5990
5991         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
5992
5993         return ret;
5994 }
5995
5996 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
5997                                         u64 start, u64 len, int pin)
5998 {
5999         struct btrfs_block_group_cache *cache;
6000         int ret = 0;
6001
6002         cache = btrfs_lookup_block_group(root->fs_info, start);
6003         if (!cache) {
6004                 printk(KERN_ERR "Unable to find block group for %llu\n",
6005                        (unsigned long long)start);
6006                 return -ENOSPC;
6007         }
6008
6009         if (btrfs_test_opt(root, DISCARD))
6010                 ret = btrfs_discard_extent(root, start, len, NULL);
6011
6012         if (pin)
6013                 pin_down_extent(root, cache, start, len, 1);
6014         else {
6015                 btrfs_add_free_space(cache, start, len);
6016                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6017         }
6018         btrfs_put_block_group(cache);
6019
6020         trace_btrfs_reserved_extent_free(root, start, len);
6021
6022         return ret;
6023 }
6024
6025 int btrfs_free_reserved_extent(struct btrfs_root *root,
6026                                         u64 start, u64 len)
6027 {
6028         return __btrfs_free_reserved_extent(root, start, len, 0);
6029 }
6030
6031 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6032                                        u64 start, u64 len)
6033 {
6034         return __btrfs_free_reserved_extent(root, start, len, 1);
6035 }
6036
6037 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6038                                       struct btrfs_root *root,
6039                                       u64 parent, u64 root_objectid,
6040                                       u64 flags, u64 owner, u64 offset,
6041                                       struct btrfs_key *ins, int ref_mod)
6042 {
6043         int ret;
6044         struct btrfs_fs_info *fs_info = root->fs_info;
6045         struct btrfs_extent_item *extent_item;
6046         struct btrfs_extent_inline_ref *iref;
6047         struct btrfs_path *path;
6048         struct extent_buffer *leaf;
6049         int type;
6050         u32 size;
6051
6052         if (parent > 0)
6053                 type = BTRFS_SHARED_DATA_REF_KEY;
6054         else
6055                 type = BTRFS_EXTENT_DATA_REF_KEY;
6056
6057         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6058
6059         path = btrfs_alloc_path();
6060         if (!path)
6061                 return -ENOMEM;
6062
6063         path->leave_spinning = 1;
6064         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6065                                       ins, size);
6066         if (ret) {
6067                 btrfs_free_path(path);
6068                 return ret;
6069         }
6070
6071         leaf = path->nodes[0];
6072         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6073                                      struct btrfs_extent_item);
6074         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6075         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6076         btrfs_set_extent_flags(leaf, extent_item,
6077                                flags | BTRFS_EXTENT_FLAG_DATA);
6078
6079         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6080         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6081         if (parent > 0) {
6082                 struct btrfs_shared_data_ref *ref;
6083                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6084                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6085                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6086         } else {
6087                 struct btrfs_extent_data_ref *ref;
6088                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6089                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6090                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6091                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6092                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6093         }
6094
6095         btrfs_mark_buffer_dirty(path->nodes[0]);
6096         btrfs_free_path(path);
6097
6098         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
6099         if (ret) { /* -ENOENT, logic error */
6100                 printk(KERN_ERR "btrfs update block group failed for %llu "
6101                        "%llu\n", (unsigned long long)ins->objectid,
6102                        (unsigned long long)ins->offset);
6103                 BUG();
6104         }
6105         return ret;
6106 }
6107
6108 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6109                                      struct btrfs_root *root,
6110                                      u64 parent, u64 root_objectid,
6111                                      u64 flags, struct btrfs_disk_key *key,
6112                                      int level, struct btrfs_key *ins)
6113 {
6114         int ret;
6115         struct btrfs_fs_info *fs_info = root->fs_info;
6116         struct btrfs_extent_item *extent_item;
6117         struct btrfs_tree_block_info *block_info;
6118         struct btrfs_extent_inline_ref *iref;
6119         struct btrfs_path *path;
6120         struct extent_buffer *leaf;
6121         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
6122
6123         path = btrfs_alloc_path();
6124         if (!path)
6125                 return -ENOMEM;
6126
6127         path->leave_spinning = 1;
6128         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6129                                       ins, size);
6130         if (ret) {
6131                 btrfs_free_path(path);
6132                 return ret;
6133         }
6134
6135         leaf = path->nodes[0];
6136         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6137                                      struct btrfs_extent_item);
6138         btrfs_set_extent_refs(leaf, extent_item, 1);
6139         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6140         btrfs_set_extent_flags(leaf, extent_item,
6141                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6142         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6143
6144         btrfs_set_tree_block_key(leaf, block_info, key);
6145         btrfs_set_tree_block_level(leaf, block_info, level);
6146
6147         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6148         if (parent > 0) {
6149                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6150                 btrfs_set_extent_inline_ref_type(leaf, iref,
6151                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6152                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6153         } else {
6154                 btrfs_set_extent_inline_ref_type(leaf, iref,
6155                                                  BTRFS_TREE_BLOCK_REF_KEY);
6156                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6157         }
6158
6159         btrfs_mark_buffer_dirty(leaf);
6160         btrfs_free_path(path);
6161
6162         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
6163         if (ret) { /* -ENOENT, logic error */
6164                 printk(KERN_ERR "btrfs update block group failed for %llu "
6165                        "%llu\n", (unsigned long long)ins->objectid,
6166                        (unsigned long long)ins->offset);
6167                 BUG();
6168         }
6169         return ret;
6170 }
6171
6172 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6173                                      struct btrfs_root *root,
6174                                      u64 root_objectid, u64 owner,
6175                                      u64 offset, struct btrfs_key *ins)
6176 {
6177         int ret;
6178
6179         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6180
6181         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6182                                          ins->offset, 0,
6183                                          root_objectid, owner, offset,
6184                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6185         return ret;
6186 }
6187
6188 /*
6189  * this is used by the tree logging recovery code.  It records that
6190  * an extent has been allocated and makes sure to clear the free
6191  * space cache bits as well
6192  */
6193 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6194                                    struct btrfs_root *root,
6195                                    u64 root_objectid, u64 owner, u64 offset,
6196                                    struct btrfs_key *ins)
6197 {
6198         int ret;
6199         struct btrfs_block_group_cache *block_group;
6200         struct btrfs_caching_control *caching_ctl;
6201         u64 start = ins->objectid;
6202         u64 num_bytes = ins->offset;
6203
6204         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6205         cache_block_group(block_group, trans, NULL, 0);
6206         caching_ctl = get_caching_control(block_group);
6207
6208         if (!caching_ctl) {
6209                 BUG_ON(!block_group_cache_done(block_group));
6210                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6211                 BUG_ON(ret); /* -ENOMEM */
6212         } else {
6213                 mutex_lock(&caching_ctl->mutex);
6214
6215                 if (start >= caching_ctl->progress) {
6216                         ret = add_excluded_extent(root, start, num_bytes);
6217                         BUG_ON(ret); /* -ENOMEM */
6218                 } else if (start + num_bytes <= caching_ctl->progress) {
6219                         ret = btrfs_remove_free_space(block_group,
6220                                                       start, num_bytes);
6221                         BUG_ON(ret); /* -ENOMEM */
6222                 } else {
6223                         num_bytes = caching_ctl->progress - start;
6224                         ret = btrfs_remove_free_space(block_group,
6225                                                       start, num_bytes);
6226                         BUG_ON(ret); /* -ENOMEM */
6227
6228                         start = caching_ctl->progress;
6229                         num_bytes = ins->objectid + ins->offset -
6230                                     caching_ctl->progress;
6231                         ret = add_excluded_extent(root, start, num_bytes);
6232                         BUG_ON(ret); /* -ENOMEM */
6233                 }
6234
6235                 mutex_unlock(&caching_ctl->mutex);
6236                 put_caching_control(caching_ctl);
6237         }
6238
6239         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6240                                           RESERVE_ALLOC_NO_ACCOUNT);
6241         BUG_ON(ret); /* logic error */
6242         btrfs_put_block_group(block_group);
6243         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6244                                          0, owner, offset, ins, 1);
6245         return ret;
6246 }
6247
6248 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
6249                                             struct btrfs_root *root,
6250                                             u64 bytenr, u32 blocksize,
6251                                             int level)
6252 {
6253         struct extent_buffer *buf;
6254
6255         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6256         if (!buf)
6257                 return ERR_PTR(-ENOMEM);
6258         btrfs_set_header_generation(buf, trans->transid);
6259         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6260         btrfs_tree_lock(buf);
6261         clean_tree_block(trans, root, buf);
6262         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6263
6264         btrfs_set_lock_blocking(buf);
6265         btrfs_set_buffer_uptodate(buf);
6266
6267         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6268                 /*
6269                  * we allow two log transactions at a time, use different
6270                  * EXENT bit to differentiate dirty pages.
6271                  */
6272                 if (root->log_transid % 2 == 0)
6273                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6274                                         buf->start + buf->len - 1, GFP_NOFS);
6275                 else
6276                         set_extent_new(&root->dirty_log_pages, buf->start,
6277                                         buf->start + buf->len - 1, GFP_NOFS);
6278         } else {
6279                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6280                          buf->start + buf->len - 1, GFP_NOFS);
6281         }
6282         trans->blocks_used++;
6283         /* this returns a buffer locked for blocking */
6284         return buf;
6285 }
6286
6287 static struct btrfs_block_rsv *
6288 use_block_rsv(struct btrfs_trans_handle *trans,
6289               struct btrfs_root *root, u32 blocksize)
6290 {
6291         struct btrfs_block_rsv *block_rsv;
6292         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6293         int ret;
6294
6295         block_rsv = get_block_rsv(trans, root);
6296
6297         if (block_rsv->size == 0) {
6298                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
6299                 /*
6300                  * If we couldn't reserve metadata bytes try and use some from
6301                  * the global reserve.
6302                  */
6303                 if (ret && block_rsv != global_rsv) {
6304                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6305                         if (!ret)
6306                                 return global_rsv;
6307                         return ERR_PTR(ret);
6308                 } else if (ret) {
6309                         return ERR_PTR(ret);
6310                 }
6311                 return block_rsv;
6312         }
6313
6314         ret = block_rsv_use_bytes(block_rsv, blocksize);
6315         if (!ret)
6316                 return block_rsv;
6317         if (ret) {
6318                 static DEFINE_RATELIMIT_STATE(_rs,
6319                                 DEFAULT_RATELIMIT_INTERVAL,
6320                                 /*DEFAULT_RATELIMIT_BURST*/ 2);
6321                 if (__ratelimit(&_rs)) {
6322                         printk(KERN_DEBUG "btrfs: block rsv returned %d\n", ret);
6323                         WARN_ON(1);
6324                 }
6325                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
6326                 if (!ret) {
6327                         return block_rsv;
6328                 } else if (ret && block_rsv != global_rsv) {
6329                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6330                         if (!ret)
6331                                 return global_rsv;
6332                 }
6333         }
6334
6335         return ERR_PTR(-ENOSPC);
6336 }
6337
6338 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6339                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6340 {
6341         block_rsv_add_bytes(block_rsv, blocksize, 0);
6342         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6343 }
6344
6345 /*
6346  * finds a free extent and does all the dirty work required for allocation
6347  * returns the key for the extent through ins, and a tree buffer for
6348  * the first block of the extent through buf.
6349  *
6350  * returns the tree buffer or NULL.
6351  */
6352 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6353                                         struct btrfs_root *root, u32 blocksize,
6354                                         u64 parent, u64 root_objectid,
6355                                         struct btrfs_disk_key *key, int level,
6356                                         u64 hint, u64 empty_size)
6357 {
6358         struct btrfs_key ins;
6359         struct btrfs_block_rsv *block_rsv;
6360         struct extent_buffer *buf;
6361         u64 flags = 0;
6362         int ret;
6363
6364
6365         block_rsv = use_block_rsv(trans, root, blocksize);
6366         if (IS_ERR(block_rsv))
6367                 return ERR_CAST(block_rsv);
6368
6369         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6370                                    empty_size, hint, &ins, 0);
6371         if (ret) {
6372                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6373                 return ERR_PTR(ret);
6374         }
6375
6376         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6377                                     blocksize, level);
6378         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6379
6380         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6381                 if (parent == 0)
6382                         parent = ins.objectid;
6383                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6384         } else
6385                 BUG_ON(parent > 0);
6386
6387         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6388                 struct btrfs_delayed_extent_op *extent_op;
6389                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
6390                 BUG_ON(!extent_op); /* -ENOMEM */
6391                 if (key)
6392                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6393                 else
6394                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6395                 extent_op->flags_to_set = flags;
6396                 extent_op->update_key = 1;
6397                 extent_op->update_flags = 1;
6398                 extent_op->is_data = 0;
6399
6400                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6401                                         ins.objectid,
6402                                         ins.offset, parent, root_objectid,
6403                                         level, BTRFS_ADD_DELAYED_EXTENT,
6404                                         extent_op, 0);
6405                 BUG_ON(ret); /* -ENOMEM */
6406         }
6407         return buf;
6408 }
6409
6410 struct walk_control {
6411         u64 refs[BTRFS_MAX_LEVEL];
6412         u64 flags[BTRFS_MAX_LEVEL];
6413         struct btrfs_key update_progress;
6414         int stage;
6415         int level;
6416         int shared_level;
6417         int update_ref;
6418         int keep_locks;
6419         int reada_slot;
6420         int reada_count;
6421         int for_reloc;
6422 };
6423
6424 #define DROP_REFERENCE  1
6425 #define UPDATE_BACKREF  2
6426
6427 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6428                                      struct btrfs_root *root,
6429                                      struct walk_control *wc,
6430                                      struct btrfs_path *path)
6431 {
6432         u64 bytenr;
6433         u64 generation;
6434         u64 refs;
6435         u64 flags;
6436         u32 nritems;
6437         u32 blocksize;
6438         struct btrfs_key key;
6439         struct extent_buffer *eb;
6440         int ret;
6441         int slot;
6442         int nread = 0;
6443
6444         if (path->slots[wc->level] < wc->reada_slot) {
6445                 wc->reada_count = wc->reada_count * 2 / 3;
6446                 wc->reada_count = max(wc->reada_count, 2);
6447         } else {
6448                 wc->reada_count = wc->reada_count * 3 / 2;
6449                 wc->reada_count = min_t(int, wc->reada_count,
6450                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6451         }
6452
6453         eb = path->nodes[wc->level];
6454         nritems = btrfs_header_nritems(eb);
6455         blocksize = btrfs_level_size(root, wc->level - 1);
6456
6457         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6458                 if (nread >= wc->reada_count)
6459                         break;
6460
6461                 cond_resched();
6462                 bytenr = btrfs_node_blockptr(eb, slot);
6463                 generation = btrfs_node_ptr_generation(eb, slot);
6464
6465                 if (slot == path->slots[wc->level])
6466                         goto reada;
6467
6468                 if (wc->stage == UPDATE_BACKREF &&
6469                     generation <= root->root_key.offset)
6470                         continue;
6471
6472                 /* We don't lock the tree block, it's OK to be racy here */
6473                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6474                                                &refs, &flags);
6475                 /* We don't care about errors in readahead. */
6476                 if (ret < 0)
6477                         continue;
6478                 BUG_ON(refs == 0);
6479
6480                 if (wc->stage == DROP_REFERENCE) {
6481                         if (refs == 1)
6482                                 goto reada;
6483
6484                         if (wc->level == 1 &&
6485                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6486                                 continue;
6487                         if (!wc->update_ref ||
6488                             generation <= root->root_key.offset)
6489                                 continue;
6490                         btrfs_node_key_to_cpu(eb, &key, slot);
6491                         ret = btrfs_comp_cpu_keys(&key,
6492                                                   &wc->update_progress);
6493                         if (ret < 0)
6494                                 continue;
6495                 } else {
6496                         if (wc->level == 1 &&
6497                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6498                                 continue;
6499                 }
6500 reada:
6501                 ret = readahead_tree_block(root, bytenr, blocksize,
6502                                            generation);
6503                 if (ret)
6504                         break;
6505                 nread++;
6506         }
6507         wc->reada_slot = slot;
6508 }
6509
6510 /*
6511  * hepler to process tree block while walking down the tree.
6512  *
6513  * when wc->stage == UPDATE_BACKREF, this function updates
6514  * back refs for pointers in the block.
6515  *
6516  * NOTE: return value 1 means we should stop walking down.
6517  */
6518 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6519                                    struct btrfs_root *root,
6520                                    struct btrfs_path *path,
6521                                    struct walk_control *wc, int lookup_info)
6522 {
6523         int level = wc->level;
6524         struct extent_buffer *eb = path->nodes[level];
6525         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6526         int ret;
6527
6528         if (wc->stage == UPDATE_BACKREF &&
6529             btrfs_header_owner(eb) != root->root_key.objectid)
6530                 return 1;
6531
6532         /*
6533          * when reference count of tree block is 1, it won't increase
6534          * again. once full backref flag is set, we never clear it.
6535          */
6536         if (lookup_info &&
6537             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6538              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6539                 BUG_ON(!path->locks[level]);
6540                 ret = btrfs_lookup_extent_info(trans, root,
6541                                                eb->start, eb->len,
6542                                                &wc->refs[level],
6543                                                &wc->flags[level]);
6544                 BUG_ON(ret == -ENOMEM);
6545                 if (ret)
6546                         return ret;
6547                 BUG_ON(wc->refs[level] == 0);
6548         }
6549
6550         if (wc->stage == DROP_REFERENCE) {
6551                 if (wc->refs[level] > 1)
6552                         return 1;
6553
6554                 if (path->locks[level] && !wc->keep_locks) {
6555                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6556                         path->locks[level] = 0;
6557                 }
6558                 return 0;
6559         }
6560
6561         /* wc->stage == UPDATE_BACKREF */
6562         if (!(wc->flags[level] & flag)) {
6563                 BUG_ON(!path->locks[level]);
6564                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
6565                 BUG_ON(ret); /* -ENOMEM */
6566                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
6567                 BUG_ON(ret); /* -ENOMEM */
6568                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6569                                                   eb->len, flag, 0);
6570                 BUG_ON(ret); /* -ENOMEM */
6571                 wc->flags[level] |= flag;
6572         }
6573
6574         /*
6575          * the block is shared by multiple trees, so it's not good to
6576          * keep the tree lock
6577          */
6578         if (path->locks[level] && level > 0) {
6579                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6580                 path->locks[level] = 0;
6581         }
6582         return 0;
6583 }
6584
6585 /*
6586  * hepler to process tree block pointer.
6587  *
6588  * when wc->stage == DROP_REFERENCE, this function checks
6589  * reference count of the block pointed to. if the block
6590  * is shared and we need update back refs for the subtree
6591  * rooted at the block, this function changes wc->stage to
6592  * UPDATE_BACKREF. if the block is shared and there is no
6593  * need to update back, this function drops the reference
6594  * to the block.
6595  *
6596  * NOTE: return value 1 means we should stop walking down.
6597  */
6598 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6599                                  struct btrfs_root *root,
6600                                  struct btrfs_path *path,
6601                                  struct walk_control *wc, int *lookup_info)
6602 {
6603         u64 bytenr;
6604         u64 generation;
6605         u64 parent;
6606         u32 blocksize;
6607         struct btrfs_key key;
6608         struct extent_buffer *next;
6609         int level = wc->level;
6610         int reada = 0;
6611         int ret = 0;
6612
6613         generation = btrfs_node_ptr_generation(path->nodes[level],
6614                                                path->slots[level]);
6615         /*
6616          * if the lower level block was created before the snapshot
6617          * was created, we know there is no need to update back refs
6618          * for the subtree
6619          */
6620         if (wc->stage == UPDATE_BACKREF &&
6621             generation <= root->root_key.offset) {
6622                 *lookup_info = 1;
6623                 return 1;
6624         }
6625
6626         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6627         blocksize = btrfs_level_size(root, level - 1);
6628
6629         next = btrfs_find_tree_block(root, bytenr, blocksize);
6630         if (!next) {
6631                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6632                 if (!next)
6633                         return -ENOMEM;
6634                 reada = 1;
6635         }
6636         btrfs_tree_lock(next);
6637         btrfs_set_lock_blocking(next);
6638
6639         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6640                                        &wc->refs[level - 1],
6641                                        &wc->flags[level - 1]);
6642         if (ret < 0) {
6643                 btrfs_tree_unlock(next);
6644                 return ret;
6645         }
6646
6647         BUG_ON(wc->refs[level - 1] == 0);
6648         *lookup_info = 0;
6649
6650         if (wc->stage == DROP_REFERENCE) {
6651                 if (wc->refs[level - 1] > 1) {
6652                         if (level == 1 &&
6653                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6654                                 goto skip;
6655
6656                         if (!wc->update_ref ||
6657                             generation <= root->root_key.offset)
6658                                 goto skip;
6659
6660                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6661                                               path->slots[level]);
6662                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6663                         if (ret < 0)
6664                                 goto skip;
6665
6666                         wc->stage = UPDATE_BACKREF;
6667                         wc->shared_level = level - 1;
6668                 }
6669         } else {
6670                 if (level == 1 &&
6671                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6672                         goto skip;
6673         }
6674
6675         if (!btrfs_buffer_uptodate(next, generation, 0)) {
6676                 btrfs_tree_unlock(next);
6677                 free_extent_buffer(next);
6678                 next = NULL;
6679                 *lookup_info = 1;
6680         }
6681
6682         if (!next) {
6683                 if (reada && level == 1)
6684                         reada_walk_down(trans, root, wc, path);
6685                 next = read_tree_block(root, bytenr, blocksize, generation);
6686                 if (!next)
6687                         return -EIO;
6688                 btrfs_tree_lock(next);
6689                 btrfs_set_lock_blocking(next);
6690         }
6691
6692         level--;
6693         BUG_ON(level != btrfs_header_level(next));
6694         path->nodes[level] = next;
6695         path->slots[level] = 0;
6696         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6697         wc->level = level;
6698         if (wc->level == 1)
6699                 wc->reada_slot = 0;
6700         return 0;
6701 skip:
6702         wc->refs[level - 1] = 0;
6703         wc->flags[level - 1] = 0;
6704         if (wc->stage == DROP_REFERENCE) {
6705                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6706                         parent = path->nodes[level]->start;
6707                 } else {
6708                         BUG_ON(root->root_key.objectid !=
6709                                btrfs_header_owner(path->nodes[level]));
6710                         parent = 0;
6711                 }
6712
6713                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6714                                 root->root_key.objectid, level - 1, 0, 0);
6715                 BUG_ON(ret); /* -ENOMEM */
6716         }
6717         btrfs_tree_unlock(next);
6718         free_extent_buffer(next);
6719         *lookup_info = 1;
6720         return 1;
6721 }
6722
6723 /*
6724  * hepler to process tree block while walking up the tree.
6725  *
6726  * when wc->stage == DROP_REFERENCE, this function drops
6727  * reference count on the block.
6728  *
6729  * when wc->stage == UPDATE_BACKREF, this function changes
6730  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6731  * to UPDATE_BACKREF previously while processing the block.
6732  *
6733  * NOTE: return value 1 means we should stop walking up.
6734  */
6735 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6736                                  struct btrfs_root *root,
6737                                  struct btrfs_path *path,
6738                                  struct walk_control *wc)
6739 {
6740         int ret;
6741         int level = wc->level;
6742         struct extent_buffer *eb = path->nodes[level];
6743         u64 parent = 0;
6744
6745         if (wc->stage == UPDATE_BACKREF) {
6746                 BUG_ON(wc->shared_level < level);
6747                 if (level < wc->shared_level)
6748                         goto out;
6749
6750                 ret = find_next_key(path, level + 1, &wc->update_progress);
6751                 if (ret > 0)
6752                         wc->update_ref = 0;
6753
6754                 wc->stage = DROP_REFERENCE;
6755                 wc->shared_level = -1;
6756                 path->slots[level] = 0;
6757
6758                 /*
6759                  * check reference count again if the block isn't locked.
6760                  * we should start walking down the tree again if reference
6761                  * count is one.
6762                  */
6763                 if (!path->locks[level]) {
6764                         BUG_ON(level == 0);
6765                         btrfs_tree_lock(eb);
6766                         btrfs_set_lock_blocking(eb);
6767                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6768
6769                         ret = btrfs_lookup_extent_info(trans, root,
6770                                                        eb->start, eb->len,
6771                                                        &wc->refs[level],
6772                                                        &wc->flags[level]);
6773                         if (ret < 0) {
6774                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6775                                 return ret;
6776                         }
6777                         BUG_ON(wc->refs[level] == 0);
6778                         if (wc->refs[level] == 1) {
6779                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6780                                 return 1;
6781                         }
6782                 }
6783         }
6784
6785         /* wc->stage == DROP_REFERENCE */
6786         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6787
6788         if (wc->refs[level] == 1) {
6789                 if (level == 0) {
6790                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6791                                 ret = btrfs_dec_ref(trans, root, eb, 1,
6792                                                     wc->for_reloc);
6793                         else
6794                                 ret = btrfs_dec_ref(trans, root, eb, 0,
6795                                                     wc->for_reloc);
6796                         BUG_ON(ret); /* -ENOMEM */
6797                 }
6798                 /* make block locked assertion in clean_tree_block happy */
6799                 if (!path->locks[level] &&
6800                     btrfs_header_generation(eb) == trans->transid) {
6801                         btrfs_tree_lock(eb);
6802                         btrfs_set_lock_blocking(eb);
6803                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6804                 }
6805                 clean_tree_block(trans, root, eb);
6806         }
6807
6808         if (eb == root->node) {
6809                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6810                         parent = eb->start;
6811                 else
6812                         BUG_ON(root->root_key.objectid !=
6813                                btrfs_header_owner(eb));
6814         } else {
6815                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6816                         parent = path->nodes[level + 1]->start;
6817                 else
6818                         BUG_ON(root->root_key.objectid !=
6819                                btrfs_header_owner(path->nodes[level + 1]));
6820         }
6821
6822         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6823 out:
6824         wc->refs[level] = 0;
6825         wc->flags[level] = 0;
6826         return 0;
6827 }
6828
6829 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6830                                    struct btrfs_root *root,
6831                                    struct btrfs_path *path,
6832                                    struct walk_control *wc)
6833 {
6834         int level = wc->level;
6835         int lookup_info = 1;
6836         int ret;
6837
6838         while (level >= 0) {
6839                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6840                 if (ret > 0)
6841                         break;
6842
6843                 if (level == 0)
6844                         break;
6845
6846                 if (path->slots[level] >=
6847                     btrfs_header_nritems(path->nodes[level]))
6848                         break;
6849
6850                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6851                 if (ret > 0) {
6852                         path->slots[level]++;
6853                         continue;
6854                 } else if (ret < 0)
6855                         return ret;
6856                 level = wc->level;
6857         }
6858         return 0;
6859 }
6860
6861 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6862                                  struct btrfs_root *root,
6863                                  struct btrfs_path *path,
6864                                  struct walk_control *wc, int max_level)
6865 {
6866         int level = wc->level;
6867         int ret;
6868
6869         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6870         while (level < max_level && path->nodes[level]) {
6871                 wc->level = level;
6872                 if (path->slots[level] + 1 <
6873                     btrfs_header_nritems(path->nodes[level])) {
6874                         path->slots[level]++;
6875                         return 0;
6876                 } else {
6877                         ret = walk_up_proc(trans, root, path, wc);
6878                         if (ret > 0)
6879                                 return 0;
6880
6881                         if (path->locks[level]) {
6882                                 btrfs_tree_unlock_rw(path->nodes[level],
6883                                                      path->locks[level]);
6884                                 path->locks[level] = 0;
6885                         }
6886                         free_extent_buffer(path->nodes[level]);
6887                         path->nodes[level] = NULL;
6888                         level++;
6889                 }
6890         }
6891         return 1;
6892 }
6893
6894 /*
6895  * drop a subvolume tree.
6896  *
6897  * this function traverses the tree freeing any blocks that only
6898  * referenced by the tree.
6899  *
6900  * when a shared tree block is found. this function decreases its
6901  * reference count by one. if update_ref is true, this function
6902  * also make sure backrefs for the shared block and all lower level
6903  * blocks are properly updated.
6904  */
6905 int btrfs_drop_snapshot(struct btrfs_root *root,
6906                          struct btrfs_block_rsv *block_rsv, int update_ref,
6907                          int for_reloc)
6908 {
6909         struct btrfs_path *path;
6910         struct btrfs_trans_handle *trans;
6911         struct btrfs_root *tree_root = root->fs_info->tree_root;
6912         struct btrfs_root_item *root_item = &root->root_item;
6913         struct walk_control *wc;
6914         struct btrfs_key key;
6915         int err = 0;
6916         int ret;
6917         int level;
6918
6919         path = btrfs_alloc_path();
6920         if (!path) {
6921                 err = -ENOMEM;
6922                 goto out;
6923         }
6924
6925         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6926         if (!wc) {
6927                 btrfs_free_path(path);
6928                 err = -ENOMEM;
6929                 goto out;
6930         }
6931
6932         trans = btrfs_start_transaction(tree_root, 0);
6933         if (IS_ERR(trans)) {
6934                 err = PTR_ERR(trans);
6935                 goto out_free;
6936         }
6937
6938         if (block_rsv)
6939                 trans->block_rsv = block_rsv;
6940
6941         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6942                 level = btrfs_header_level(root->node);
6943                 path->nodes[level] = btrfs_lock_root_node(root);
6944                 btrfs_set_lock_blocking(path->nodes[level]);
6945                 path->slots[level] = 0;
6946                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6947                 memset(&wc->update_progress, 0,
6948                        sizeof(wc->update_progress));
6949         } else {
6950                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6951                 memcpy(&wc->update_progress, &key,
6952                        sizeof(wc->update_progress));
6953
6954                 level = root_item->drop_level;
6955                 BUG_ON(level == 0);
6956                 path->lowest_level = level;
6957                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6958                 path->lowest_level = 0;
6959                 if (ret < 0) {
6960                         err = ret;
6961                         goto out_end_trans;
6962                 }
6963                 WARN_ON(ret > 0);
6964
6965                 /*
6966                  * unlock our path, this is safe because only this
6967                  * function is allowed to delete this snapshot
6968                  */
6969                 btrfs_unlock_up_safe(path, 0);
6970
6971                 level = btrfs_header_level(root->node);
6972                 while (1) {
6973                         btrfs_tree_lock(path->nodes[level]);
6974                         btrfs_set_lock_blocking(path->nodes[level]);
6975
6976                         ret = btrfs_lookup_extent_info(trans, root,
6977                                                 path->nodes[level]->start,
6978                                                 path->nodes[level]->len,
6979                                                 &wc->refs[level],
6980                                                 &wc->flags[level]);
6981                         if (ret < 0) {
6982                                 err = ret;
6983                                 goto out_end_trans;
6984                         }
6985                         BUG_ON(wc->refs[level] == 0);
6986
6987                         if (level == root_item->drop_level)
6988                                 break;
6989
6990                         btrfs_tree_unlock(path->nodes[level]);
6991                         WARN_ON(wc->refs[level] != 1);
6992                         level--;
6993                 }
6994         }
6995
6996         wc->level = level;
6997         wc->shared_level = -1;
6998         wc->stage = DROP_REFERENCE;
6999         wc->update_ref = update_ref;
7000         wc->keep_locks = 0;
7001         wc->for_reloc = for_reloc;
7002         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7003
7004         while (1) {
7005                 ret = walk_down_tree(trans, root, path, wc);
7006                 if (ret < 0) {
7007                         err = ret;
7008                         break;
7009                 }
7010
7011                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7012                 if (ret < 0) {
7013                         err = ret;
7014                         break;
7015                 }
7016
7017                 if (ret > 0) {
7018                         BUG_ON(wc->stage != DROP_REFERENCE);
7019                         break;
7020                 }
7021
7022                 if (wc->stage == DROP_REFERENCE) {
7023                         level = wc->level;
7024                         btrfs_node_key(path->nodes[level],
7025                                        &root_item->drop_progress,
7026                                        path->slots[level]);
7027                         root_item->drop_level = level;
7028                 }
7029
7030                 BUG_ON(wc->level == 0);
7031                 if (btrfs_should_end_transaction(trans, tree_root)) {
7032                         ret = btrfs_update_root(trans, tree_root,
7033                                                 &root->root_key,
7034                                                 root_item);
7035                         if (ret) {
7036                                 btrfs_abort_transaction(trans, tree_root, ret);
7037                                 err = ret;
7038                                 goto out_end_trans;
7039                         }
7040
7041                         btrfs_end_transaction_throttle(trans, tree_root);
7042                         trans = btrfs_start_transaction(tree_root, 0);
7043                         if (IS_ERR(trans)) {
7044                                 err = PTR_ERR(trans);
7045                                 goto out_free;
7046                         }
7047                         if (block_rsv)
7048                                 trans->block_rsv = block_rsv;
7049                 }
7050         }
7051         btrfs_release_path(path);
7052         if (err)
7053                 goto out_end_trans;
7054
7055         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7056         if (ret) {
7057                 btrfs_abort_transaction(trans, tree_root, ret);
7058                 goto out_end_trans;
7059         }
7060
7061         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7062                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
7063                                            NULL, NULL);
7064                 if (ret < 0) {
7065                         btrfs_abort_transaction(trans, tree_root, ret);
7066                         err = ret;
7067                         goto out_end_trans;
7068                 } else if (ret > 0) {
7069                         /* if we fail to delete the orphan item this time
7070                          * around, it'll get picked up the next time.
7071                          *
7072                          * The most common failure here is just -ENOENT.
7073                          */
7074                         btrfs_del_orphan_item(trans, tree_root,
7075                                               root->root_key.objectid);
7076                 }
7077         }
7078
7079         if (root->in_radix) {
7080                 btrfs_free_fs_root(tree_root->fs_info, root);
7081         } else {
7082                 free_extent_buffer(root->node);
7083                 free_extent_buffer(root->commit_root);
7084                 kfree(root);
7085         }
7086 out_end_trans:
7087         btrfs_end_transaction_throttle(trans, tree_root);
7088 out_free:
7089         kfree(wc);
7090         btrfs_free_path(path);
7091 out:
7092         if (err)
7093                 btrfs_std_error(root->fs_info, err);
7094         return err;
7095 }
7096
7097 /*
7098  * drop subtree rooted at tree block 'node'.
7099  *
7100  * NOTE: this function will unlock and release tree block 'node'
7101  * only used by relocation code
7102  */
7103 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7104                         struct btrfs_root *root,
7105                         struct extent_buffer *node,
7106                         struct extent_buffer *parent)
7107 {
7108         struct btrfs_path *path;
7109         struct walk_control *wc;
7110         int level;
7111         int parent_level;
7112         int ret = 0;
7113         int wret;
7114
7115         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7116
7117         path = btrfs_alloc_path();
7118         if (!path)
7119                 return -ENOMEM;
7120
7121         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7122         if (!wc) {
7123                 btrfs_free_path(path);
7124                 return -ENOMEM;
7125         }
7126
7127         btrfs_assert_tree_locked(parent);
7128         parent_level = btrfs_header_level(parent);
7129         extent_buffer_get(parent);
7130         path->nodes[parent_level] = parent;
7131         path->slots[parent_level] = btrfs_header_nritems(parent);
7132
7133         btrfs_assert_tree_locked(node);
7134         level = btrfs_header_level(node);
7135         path->nodes[level] = node;
7136         path->slots[level] = 0;
7137         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7138
7139         wc->refs[parent_level] = 1;
7140         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7141         wc->level = level;
7142         wc->shared_level = -1;
7143         wc->stage = DROP_REFERENCE;
7144         wc->update_ref = 0;
7145         wc->keep_locks = 1;
7146         wc->for_reloc = 1;
7147         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7148
7149         while (1) {
7150                 wret = walk_down_tree(trans, root, path, wc);
7151                 if (wret < 0) {
7152                         ret = wret;
7153                         break;
7154                 }
7155
7156                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7157                 if (wret < 0)
7158                         ret = wret;
7159                 if (wret != 0)
7160                         break;
7161         }
7162
7163         kfree(wc);
7164         btrfs_free_path(path);
7165         return ret;
7166 }
7167
7168 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7169 {
7170         u64 num_devices;
7171         u64 stripped;
7172
7173         /*
7174          * if restripe for this chunk_type is on pick target profile and
7175          * return, otherwise do the usual balance
7176          */
7177         stripped = get_restripe_target(root->fs_info, flags);
7178         if (stripped)
7179                 return extended_to_chunk(stripped);
7180
7181         /*
7182          * we add in the count of missing devices because we want
7183          * to make sure that any RAID levels on a degraded FS
7184          * continue to be honored.
7185          */
7186         num_devices = root->fs_info->fs_devices->rw_devices +
7187                 root->fs_info->fs_devices->missing_devices;
7188
7189         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7190                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7191
7192         if (num_devices == 1) {
7193                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7194                 stripped = flags & ~stripped;
7195
7196                 /* turn raid0 into single device chunks */
7197                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7198                         return stripped;
7199
7200                 /* turn mirroring into duplication */
7201                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7202                              BTRFS_BLOCK_GROUP_RAID10))
7203                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7204         } else {
7205                 /* they already had raid on here, just return */
7206                 if (flags & stripped)
7207                         return flags;
7208
7209                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7210                 stripped = flags & ~stripped;
7211
7212                 /* switch duplicated blocks with raid1 */
7213                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7214                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7215
7216                 /* this is drive concat, leave it alone */
7217         }
7218
7219         return flags;
7220 }
7221
7222 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7223 {
7224         struct btrfs_space_info *sinfo = cache->space_info;
7225         u64 num_bytes;
7226         u64 min_allocable_bytes;
7227         int ret = -ENOSPC;
7228
7229
7230         /*
7231          * We need some metadata space and system metadata space for
7232          * allocating chunks in some corner cases until we force to set
7233          * it to be readonly.
7234          */
7235         if ((sinfo->flags &
7236              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7237             !force)
7238                 min_allocable_bytes = 1 * 1024 * 1024;
7239         else
7240                 min_allocable_bytes = 0;
7241
7242         spin_lock(&sinfo->lock);
7243         spin_lock(&cache->lock);
7244
7245         if (cache->ro) {
7246                 ret = 0;
7247                 goto out;
7248         }
7249
7250         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7251                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7252
7253         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7254             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7255             min_allocable_bytes <= sinfo->total_bytes) {
7256                 sinfo->bytes_readonly += num_bytes;
7257                 cache->ro = 1;
7258                 ret = 0;
7259         }
7260 out:
7261         spin_unlock(&cache->lock);
7262         spin_unlock(&sinfo->lock);
7263         return ret;
7264 }
7265
7266 int btrfs_set_block_group_ro(struct btrfs_root *root,
7267                              struct btrfs_block_group_cache *cache)
7268
7269 {
7270         struct btrfs_trans_handle *trans;
7271         u64 alloc_flags;
7272         int ret;
7273
7274         BUG_ON(cache->ro);
7275
7276         trans = btrfs_join_transaction(root);
7277         if (IS_ERR(trans))
7278                 return PTR_ERR(trans);
7279
7280         alloc_flags = update_block_group_flags(root, cache->flags);
7281         if (alloc_flags != cache->flags) {
7282                 ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7283                                      CHUNK_ALLOC_FORCE);
7284                 if (ret < 0)
7285                         goto out;
7286         }
7287
7288         ret = set_block_group_ro(cache, 0);
7289         if (!ret)
7290                 goto out;
7291         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7292         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7293                              CHUNK_ALLOC_FORCE);
7294         if (ret < 0)
7295                 goto out;
7296         ret = set_block_group_ro(cache, 0);
7297 out:
7298         btrfs_end_transaction(trans, root);
7299         return ret;
7300 }
7301
7302 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7303                             struct btrfs_root *root, u64 type)
7304 {
7305         u64 alloc_flags = get_alloc_profile(root, type);
7306         return do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7307                               CHUNK_ALLOC_FORCE);
7308 }
7309
7310 /*
7311  * helper to account the unused space of all the readonly block group in the
7312  * list. takes mirrors into account.
7313  */
7314 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7315 {
7316         struct btrfs_block_group_cache *block_group;
7317         u64 free_bytes = 0;
7318         int factor;
7319
7320         list_for_each_entry(block_group, groups_list, list) {
7321                 spin_lock(&block_group->lock);
7322
7323                 if (!block_group->ro) {
7324                         spin_unlock(&block_group->lock);
7325                         continue;
7326                 }
7327
7328                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7329                                           BTRFS_BLOCK_GROUP_RAID10 |
7330                                           BTRFS_BLOCK_GROUP_DUP))
7331                         factor = 2;
7332                 else
7333                         factor = 1;
7334
7335                 free_bytes += (block_group->key.offset -
7336                                btrfs_block_group_used(&block_group->item)) *
7337                                factor;
7338
7339                 spin_unlock(&block_group->lock);
7340         }
7341
7342         return free_bytes;
7343 }
7344
7345 /*
7346  * helper to account the unused space of all the readonly block group in the
7347  * space_info. takes mirrors into account.
7348  */
7349 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7350 {
7351         int i;
7352         u64 free_bytes = 0;
7353
7354         spin_lock(&sinfo->lock);
7355
7356         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7357                 if (!list_empty(&sinfo->block_groups[i]))
7358                         free_bytes += __btrfs_get_ro_block_group_free_space(
7359                                                 &sinfo->block_groups[i]);
7360
7361         spin_unlock(&sinfo->lock);
7362
7363         return free_bytes;
7364 }
7365
7366 void btrfs_set_block_group_rw(struct btrfs_root *root,
7367                               struct btrfs_block_group_cache *cache)
7368 {
7369         struct btrfs_space_info *sinfo = cache->space_info;
7370         u64 num_bytes;
7371
7372         BUG_ON(!cache->ro);
7373
7374         spin_lock(&sinfo->lock);
7375         spin_lock(&cache->lock);
7376         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7377                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7378         sinfo->bytes_readonly -= num_bytes;
7379         cache->ro = 0;
7380         spin_unlock(&cache->lock);
7381         spin_unlock(&sinfo->lock);
7382 }
7383
7384 /*
7385  * checks to see if its even possible to relocate this block group.
7386  *
7387  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7388  * ok to go ahead and try.
7389  */
7390 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7391 {
7392         struct btrfs_block_group_cache *block_group;
7393         struct btrfs_space_info *space_info;
7394         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7395         struct btrfs_device *device;
7396         u64 min_free;
7397         u64 dev_min = 1;
7398         u64 dev_nr = 0;
7399         u64 target;
7400         int index;
7401         int full = 0;
7402         int ret = 0;
7403
7404         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7405
7406         /* odd, couldn't find the block group, leave it alone */
7407         if (!block_group)
7408                 return -1;
7409
7410         min_free = btrfs_block_group_used(&block_group->item);
7411
7412         /* no bytes used, we're good */
7413         if (!min_free)
7414                 goto out;
7415
7416         space_info = block_group->space_info;
7417         spin_lock(&space_info->lock);
7418
7419         full = space_info->full;
7420
7421         /*
7422          * if this is the last block group we have in this space, we can't
7423          * relocate it unless we're able to allocate a new chunk below.
7424          *
7425          * Otherwise, we need to make sure we have room in the space to handle
7426          * all of the extents from this block group.  If we can, we're good
7427          */
7428         if ((space_info->total_bytes != block_group->key.offset) &&
7429             (space_info->bytes_used + space_info->bytes_reserved +
7430              space_info->bytes_pinned + space_info->bytes_readonly +
7431              min_free < space_info->total_bytes)) {
7432                 spin_unlock(&space_info->lock);
7433                 goto out;
7434         }
7435         spin_unlock(&space_info->lock);
7436
7437         /*
7438          * ok we don't have enough space, but maybe we have free space on our
7439          * devices to allocate new chunks for relocation, so loop through our
7440          * alloc devices and guess if we have enough space.  if this block
7441          * group is going to be restriped, run checks against the target
7442          * profile instead of the current one.
7443          */
7444         ret = -1;
7445
7446         /*
7447          * index:
7448          *      0: raid10
7449          *      1: raid1
7450          *      2: dup
7451          *      3: raid0
7452          *      4: single
7453          */
7454         target = get_restripe_target(root->fs_info, block_group->flags);
7455         if (target) {
7456                 index = __get_block_group_index(extended_to_chunk(target));
7457         } else {
7458                 /*
7459                  * this is just a balance, so if we were marked as full
7460                  * we know there is no space for a new chunk
7461                  */
7462                 if (full)
7463                         goto out;
7464
7465                 index = get_block_group_index(block_group);
7466         }
7467
7468         if (index == 0) {
7469                 dev_min = 4;
7470                 /* Divide by 2 */
7471                 min_free >>= 1;
7472         } else if (index == 1) {
7473                 dev_min = 2;
7474         } else if (index == 2) {
7475                 /* Multiply by 2 */
7476                 min_free <<= 1;
7477         } else if (index == 3) {
7478                 dev_min = fs_devices->rw_devices;
7479                 do_div(min_free, dev_min);
7480         }
7481
7482         mutex_lock(&root->fs_info->chunk_mutex);
7483         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7484                 u64 dev_offset;
7485
7486                 /*
7487                  * check to make sure we can actually find a chunk with enough
7488                  * space to fit our block group in.
7489                  */
7490                 if (device->total_bytes > device->bytes_used + min_free) {
7491                         ret = find_free_dev_extent(device, min_free,
7492                                                    &dev_offset, NULL);
7493                         if (!ret)
7494                                 dev_nr++;
7495
7496                         if (dev_nr >= dev_min)
7497                                 break;
7498
7499                         ret = -1;
7500                 }
7501         }
7502         mutex_unlock(&root->fs_info->chunk_mutex);
7503 out:
7504         btrfs_put_block_group(block_group);
7505         return ret;
7506 }
7507
7508 static int find_first_block_group(struct btrfs_root *root,
7509                 struct btrfs_path *path, struct btrfs_key *key)
7510 {
7511         int ret = 0;
7512         struct btrfs_key found_key;
7513         struct extent_buffer *leaf;
7514         int slot;
7515
7516         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7517         if (ret < 0)
7518                 goto out;
7519
7520         while (1) {
7521                 slot = path->slots[0];
7522                 leaf = path->nodes[0];
7523                 if (slot >= btrfs_header_nritems(leaf)) {
7524                         ret = btrfs_next_leaf(root, path);
7525                         if (ret == 0)
7526                                 continue;
7527                         if (ret < 0)
7528                                 goto out;
7529                         break;
7530                 }
7531                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7532
7533                 if (found_key.objectid >= key->objectid &&
7534                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7535                         ret = 0;
7536                         goto out;
7537                 }
7538                 path->slots[0]++;
7539         }
7540 out:
7541         return ret;
7542 }
7543
7544 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7545 {
7546         struct btrfs_block_group_cache *block_group;
7547         u64 last = 0;
7548
7549         while (1) {
7550                 struct inode *inode;
7551
7552                 block_group = btrfs_lookup_first_block_group(info, last);
7553                 while (block_group) {
7554                         spin_lock(&block_group->lock);
7555                         if (block_group->iref)
7556                                 break;
7557                         spin_unlock(&block_group->lock);
7558                         block_group = next_block_group(info->tree_root,
7559                                                        block_group);
7560                 }
7561                 if (!block_group) {
7562                         if (last == 0)
7563                                 break;
7564                         last = 0;
7565                         continue;
7566                 }
7567
7568                 inode = block_group->inode;
7569                 block_group->iref = 0;
7570                 block_group->inode = NULL;
7571                 spin_unlock(&block_group->lock);
7572                 iput(inode);
7573                 last = block_group->key.objectid + block_group->key.offset;
7574                 btrfs_put_block_group(block_group);
7575         }
7576 }
7577
7578 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7579 {
7580         struct btrfs_block_group_cache *block_group;
7581         struct btrfs_space_info *space_info;
7582         struct btrfs_caching_control *caching_ctl;
7583         struct rb_node *n;
7584
7585         down_write(&info->extent_commit_sem);
7586         while (!list_empty(&info->caching_block_groups)) {
7587                 caching_ctl = list_entry(info->caching_block_groups.next,
7588                                          struct btrfs_caching_control, list);
7589                 list_del(&caching_ctl->list);
7590                 put_caching_control(caching_ctl);
7591         }
7592         up_write(&info->extent_commit_sem);
7593
7594         spin_lock(&info->block_group_cache_lock);
7595         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7596                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7597                                        cache_node);
7598                 rb_erase(&block_group->cache_node,
7599                          &info->block_group_cache_tree);
7600                 spin_unlock(&info->block_group_cache_lock);
7601
7602                 down_write(&block_group->space_info->groups_sem);
7603                 list_del(&block_group->list);
7604                 up_write(&block_group->space_info->groups_sem);
7605
7606                 if (block_group->cached == BTRFS_CACHE_STARTED)
7607                         wait_block_group_cache_done(block_group);
7608
7609                 /*
7610                  * We haven't cached this block group, which means we could
7611                  * possibly have excluded extents on this block group.
7612                  */
7613                 if (block_group->cached == BTRFS_CACHE_NO)
7614                         free_excluded_extents(info->extent_root, block_group);
7615
7616                 btrfs_remove_free_space_cache(block_group);
7617                 btrfs_put_block_group(block_group);
7618
7619                 spin_lock(&info->block_group_cache_lock);
7620         }
7621         spin_unlock(&info->block_group_cache_lock);
7622
7623         /* now that all the block groups are freed, go through and
7624          * free all the space_info structs.  This is only called during
7625          * the final stages of unmount, and so we know nobody is
7626          * using them.  We call synchronize_rcu() once before we start,
7627          * just to be on the safe side.
7628          */
7629         synchronize_rcu();
7630
7631         release_global_block_rsv(info);
7632
7633         while(!list_empty(&info->space_info)) {
7634                 space_info = list_entry(info->space_info.next,
7635                                         struct btrfs_space_info,
7636                                         list);
7637                 if (space_info->bytes_pinned > 0 ||
7638                     space_info->bytes_reserved > 0 ||
7639                     space_info->bytes_may_use > 0) {
7640                         WARN_ON(1);
7641                         dump_space_info(space_info, 0, 0);
7642                 }
7643                 list_del(&space_info->list);
7644                 kfree(space_info);
7645         }
7646         return 0;
7647 }
7648
7649 static void __link_block_group(struct btrfs_space_info *space_info,
7650                                struct btrfs_block_group_cache *cache)
7651 {
7652         int index = get_block_group_index(cache);
7653
7654         down_write(&space_info->groups_sem);
7655         list_add_tail(&cache->list, &space_info->block_groups[index]);
7656         up_write(&space_info->groups_sem);
7657 }
7658
7659 int btrfs_read_block_groups(struct btrfs_root *root)
7660 {
7661         struct btrfs_path *path;
7662         int ret;
7663         struct btrfs_block_group_cache *cache;
7664         struct btrfs_fs_info *info = root->fs_info;
7665         struct btrfs_space_info *space_info;
7666         struct btrfs_key key;
7667         struct btrfs_key found_key;
7668         struct extent_buffer *leaf;
7669         int need_clear = 0;
7670         u64 cache_gen;
7671
7672         root = info->extent_root;
7673         key.objectid = 0;
7674         key.offset = 0;
7675         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7676         path = btrfs_alloc_path();
7677         if (!path)
7678                 return -ENOMEM;
7679         path->reada = 1;
7680
7681         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7682         if (btrfs_test_opt(root, SPACE_CACHE) &&
7683             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7684                 need_clear = 1;
7685         if (btrfs_test_opt(root, CLEAR_CACHE))
7686                 need_clear = 1;
7687
7688         while (1) {
7689                 ret = find_first_block_group(root, path, &key);
7690                 if (ret > 0)
7691                         break;
7692                 if (ret != 0)
7693                         goto error;
7694                 leaf = path->nodes[0];
7695                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7696                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7697                 if (!cache) {
7698                         ret = -ENOMEM;
7699                         goto error;
7700                 }
7701                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7702                                                 GFP_NOFS);
7703                 if (!cache->free_space_ctl) {
7704                         kfree(cache);
7705                         ret = -ENOMEM;
7706                         goto error;
7707                 }
7708
7709                 atomic_set(&cache->count, 1);
7710                 spin_lock_init(&cache->lock);
7711                 cache->fs_info = info;
7712                 INIT_LIST_HEAD(&cache->list);
7713                 INIT_LIST_HEAD(&cache->cluster_list);
7714
7715                 if (need_clear) {
7716                         /*
7717                          * When we mount with old space cache, we need to
7718                          * set BTRFS_DC_CLEAR and set dirty flag.
7719                          *
7720                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
7721                          *    truncate the old free space cache inode and
7722                          *    setup a new one.
7723                          * b) Setting 'dirty flag' makes sure that we flush
7724                          *    the new space cache info onto disk.
7725                          */
7726                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7727                         if (btrfs_test_opt(root, SPACE_CACHE))
7728                                 cache->dirty = 1;
7729                 }
7730
7731                 read_extent_buffer(leaf, &cache->item,
7732                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7733                                    sizeof(cache->item));
7734                 memcpy(&cache->key, &found_key, sizeof(found_key));
7735
7736                 key.objectid = found_key.objectid + found_key.offset;
7737                 btrfs_release_path(path);
7738                 cache->flags = btrfs_block_group_flags(&cache->item);
7739                 cache->sectorsize = root->sectorsize;
7740
7741                 btrfs_init_free_space_ctl(cache);
7742
7743                 /*
7744                  * We need to exclude the super stripes now so that the space
7745                  * info has super bytes accounted for, otherwise we'll think
7746                  * we have more space than we actually do.
7747                  */
7748                 exclude_super_stripes(root, cache);
7749
7750                 /*
7751                  * check for two cases, either we are full, and therefore
7752                  * don't need to bother with the caching work since we won't
7753                  * find any space, or we are empty, and we can just add all
7754                  * the space in and be done with it.  This saves us _alot_ of
7755                  * time, particularly in the full case.
7756                  */
7757                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7758                         cache->last_byte_to_unpin = (u64)-1;
7759                         cache->cached = BTRFS_CACHE_FINISHED;
7760                         free_excluded_extents(root, cache);
7761                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7762                         cache->last_byte_to_unpin = (u64)-1;
7763                         cache->cached = BTRFS_CACHE_FINISHED;
7764                         add_new_free_space(cache, root->fs_info,
7765                                            found_key.objectid,
7766                                            found_key.objectid +
7767                                            found_key.offset);
7768                         free_excluded_extents(root, cache);
7769                 }
7770
7771                 ret = update_space_info(info, cache->flags, found_key.offset,
7772                                         btrfs_block_group_used(&cache->item),
7773                                         &space_info);
7774                 BUG_ON(ret); /* -ENOMEM */
7775                 cache->space_info = space_info;
7776                 spin_lock(&cache->space_info->lock);
7777                 cache->space_info->bytes_readonly += cache->bytes_super;
7778                 spin_unlock(&cache->space_info->lock);
7779
7780                 __link_block_group(space_info, cache);
7781
7782                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7783                 BUG_ON(ret); /* Logic error */
7784
7785                 set_avail_alloc_bits(root->fs_info, cache->flags);
7786                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7787                         set_block_group_ro(cache, 1);
7788         }
7789
7790         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7791                 if (!(get_alloc_profile(root, space_info->flags) &
7792                       (BTRFS_BLOCK_GROUP_RAID10 |
7793                        BTRFS_BLOCK_GROUP_RAID1 |
7794                        BTRFS_BLOCK_GROUP_DUP)))
7795                         continue;
7796                 /*
7797                  * avoid allocating from un-mirrored block group if there are
7798                  * mirrored block groups.
7799                  */
7800                 list_for_each_entry(cache, &space_info->block_groups[3], list)
7801                         set_block_group_ro(cache, 1);
7802                 list_for_each_entry(cache, &space_info->block_groups[4], list)
7803                         set_block_group_ro(cache, 1);
7804         }
7805
7806         init_global_block_rsv(info);
7807         ret = 0;
7808 error:
7809         btrfs_free_path(path);
7810         return ret;
7811 }
7812
7813 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7814                            struct btrfs_root *root, u64 bytes_used,
7815                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7816                            u64 size)
7817 {
7818         int ret;
7819         struct btrfs_root *extent_root;
7820         struct btrfs_block_group_cache *cache;
7821
7822         extent_root = root->fs_info->extent_root;
7823
7824         root->fs_info->last_trans_log_full_commit = trans->transid;
7825
7826         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7827         if (!cache)
7828                 return -ENOMEM;
7829         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7830                                         GFP_NOFS);
7831         if (!cache->free_space_ctl) {
7832                 kfree(cache);
7833                 return -ENOMEM;
7834         }
7835
7836         cache->key.objectid = chunk_offset;
7837         cache->key.offset = size;
7838         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7839         cache->sectorsize = root->sectorsize;
7840         cache->fs_info = root->fs_info;
7841
7842         atomic_set(&cache->count, 1);
7843         spin_lock_init(&cache->lock);
7844         INIT_LIST_HEAD(&cache->list);
7845         INIT_LIST_HEAD(&cache->cluster_list);
7846
7847         btrfs_init_free_space_ctl(cache);
7848
7849         btrfs_set_block_group_used(&cache->item, bytes_used);
7850         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7851         cache->flags = type;
7852         btrfs_set_block_group_flags(&cache->item, type);
7853
7854         cache->last_byte_to_unpin = (u64)-1;
7855         cache->cached = BTRFS_CACHE_FINISHED;
7856         exclude_super_stripes(root, cache);
7857
7858         add_new_free_space(cache, root->fs_info, chunk_offset,
7859                            chunk_offset + size);
7860
7861         free_excluded_extents(root, cache);
7862
7863         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7864                                 &cache->space_info);
7865         BUG_ON(ret); /* -ENOMEM */
7866         update_global_block_rsv(root->fs_info);
7867
7868         spin_lock(&cache->space_info->lock);
7869         cache->space_info->bytes_readonly += cache->bytes_super;
7870         spin_unlock(&cache->space_info->lock);
7871
7872         __link_block_group(cache->space_info, cache);
7873
7874         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7875         BUG_ON(ret); /* Logic error */
7876
7877         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
7878                                 sizeof(cache->item));
7879         if (ret) {
7880                 btrfs_abort_transaction(trans, extent_root, ret);
7881                 return ret;
7882         }
7883
7884         set_avail_alloc_bits(extent_root->fs_info, type);
7885
7886         return 0;
7887 }
7888
7889 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
7890 {
7891         u64 extra_flags = chunk_to_extended(flags) &
7892                                 BTRFS_EXTENDED_PROFILE_MASK;
7893
7894         if (flags & BTRFS_BLOCK_GROUP_DATA)
7895                 fs_info->avail_data_alloc_bits &= ~extra_flags;
7896         if (flags & BTRFS_BLOCK_GROUP_METADATA)
7897                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
7898         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
7899                 fs_info->avail_system_alloc_bits &= ~extra_flags;
7900 }
7901
7902 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7903                              struct btrfs_root *root, u64 group_start)
7904 {
7905         struct btrfs_path *path;
7906         struct btrfs_block_group_cache *block_group;
7907         struct btrfs_free_cluster *cluster;
7908         struct btrfs_root *tree_root = root->fs_info->tree_root;
7909         struct btrfs_key key;
7910         struct inode *inode;
7911         int ret;
7912         int index;
7913         int factor;
7914
7915         root = root->fs_info->extent_root;
7916
7917         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7918         BUG_ON(!block_group);
7919         BUG_ON(!block_group->ro);
7920
7921         /*
7922          * Free the reserved super bytes from this block group before
7923          * remove it.
7924          */
7925         free_excluded_extents(root, block_group);
7926
7927         memcpy(&key, &block_group->key, sizeof(key));
7928         index = get_block_group_index(block_group);
7929         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7930                                   BTRFS_BLOCK_GROUP_RAID1 |
7931                                   BTRFS_BLOCK_GROUP_RAID10))
7932                 factor = 2;
7933         else
7934                 factor = 1;
7935
7936         /* make sure this block group isn't part of an allocation cluster */
7937         cluster = &root->fs_info->data_alloc_cluster;
7938         spin_lock(&cluster->refill_lock);
7939         btrfs_return_cluster_to_free_space(block_group, cluster);
7940         spin_unlock(&cluster->refill_lock);
7941
7942         /*
7943          * make sure this block group isn't part of a metadata
7944          * allocation cluster
7945          */
7946         cluster = &root->fs_info->meta_alloc_cluster;
7947         spin_lock(&cluster->refill_lock);
7948         btrfs_return_cluster_to_free_space(block_group, cluster);
7949         spin_unlock(&cluster->refill_lock);
7950
7951         path = btrfs_alloc_path();
7952         if (!path) {
7953                 ret = -ENOMEM;
7954                 goto out;
7955         }
7956
7957         inode = lookup_free_space_inode(tree_root, block_group, path);
7958         if (!IS_ERR(inode)) {
7959                 ret = btrfs_orphan_add(trans, inode);
7960                 if (ret) {
7961                         btrfs_add_delayed_iput(inode);
7962                         goto out;
7963                 }
7964                 clear_nlink(inode);
7965                 /* One for the block groups ref */
7966                 spin_lock(&block_group->lock);
7967                 if (block_group->iref) {
7968                         block_group->iref = 0;
7969                         block_group->inode = NULL;
7970                         spin_unlock(&block_group->lock);
7971                         iput(inode);
7972                 } else {
7973                         spin_unlock(&block_group->lock);
7974                 }
7975                 /* One for our lookup ref */
7976                 btrfs_add_delayed_iput(inode);
7977         }
7978
7979         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
7980         key.offset = block_group->key.objectid;
7981         key.type = 0;
7982
7983         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
7984         if (ret < 0)
7985                 goto out;
7986         if (ret > 0)
7987                 btrfs_release_path(path);
7988         if (ret == 0) {
7989                 ret = btrfs_del_item(trans, tree_root, path);
7990                 if (ret)
7991                         goto out;
7992                 btrfs_release_path(path);
7993         }
7994
7995         spin_lock(&root->fs_info->block_group_cache_lock);
7996         rb_erase(&block_group->cache_node,
7997                  &root->fs_info->block_group_cache_tree);
7998         spin_unlock(&root->fs_info->block_group_cache_lock);
7999
8000         down_write(&block_group->space_info->groups_sem);
8001         /*
8002          * we must use list_del_init so people can check to see if they
8003          * are still on the list after taking the semaphore
8004          */
8005         list_del_init(&block_group->list);
8006         if (list_empty(&block_group->space_info->block_groups[index]))
8007                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
8008         up_write(&block_group->space_info->groups_sem);
8009
8010         if (block_group->cached == BTRFS_CACHE_STARTED)
8011                 wait_block_group_cache_done(block_group);
8012
8013         btrfs_remove_free_space_cache(block_group);
8014
8015         spin_lock(&block_group->space_info->lock);
8016         block_group->space_info->total_bytes -= block_group->key.offset;
8017         block_group->space_info->bytes_readonly -= block_group->key.offset;
8018         block_group->space_info->disk_total -= block_group->key.offset * factor;
8019         spin_unlock(&block_group->space_info->lock);
8020
8021         memcpy(&key, &block_group->key, sizeof(key));
8022
8023         btrfs_clear_space_info_full(root->fs_info);
8024
8025         btrfs_put_block_group(block_group);
8026         btrfs_put_block_group(block_group);
8027
8028         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8029         if (ret > 0)
8030                 ret = -EIO;
8031         if (ret < 0)
8032                 goto out;
8033
8034         ret = btrfs_del_item(trans, root, path);
8035 out:
8036         btrfs_free_path(path);
8037         return ret;
8038 }
8039
8040 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
8041 {
8042         struct btrfs_space_info *space_info;
8043         struct btrfs_super_block *disk_super;
8044         u64 features;
8045         u64 flags;
8046         int mixed = 0;
8047         int ret;
8048
8049         disk_super = fs_info->super_copy;
8050         if (!btrfs_super_root(disk_super))
8051                 return 1;
8052
8053         features = btrfs_super_incompat_flags(disk_super);
8054         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
8055                 mixed = 1;
8056
8057         flags = BTRFS_BLOCK_GROUP_SYSTEM;
8058         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8059         if (ret)
8060                 goto out;
8061
8062         if (mixed) {
8063                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
8064                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8065         } else {
8066                 flags = BTRFS_BLOCK_GROUP_METADATA;
8067                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8068                 if (ret)
8069                         goto out;
8070
8071                 flags = BTRFS_BLOCK_GROUP_DATA;
8072                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8073         }
8074 out:
8075         return ret;
8076 }
8077
8078 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8079 {
8080         return unpin_extent_range(root, start, end);
8081 }
8082
8083 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8084                                u64 num_bytes, u64 *actual_bytes)
8085 {
8086         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8087 }
8088
8089 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8090 {
8091         struct btrfs_fs_info *fs_info = root->fs_info;
8092         struct btrfs_block_group_cache *cache = NULL;
8093         u64 group_trimmed;
8094         u64 start;
8095         u64 end;
8096         u64 trimmed = 0;
8097         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8098         int ret = 0;
8099
8100         /*
8101          * try to trim all FS space, our block group may start from non-zero.
8102          */
8103         if (range->len == total_bytes)
8104                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8105         else
8106                 cache = btrfs_lookup_block_group(fs_info, range->start);
8107
8108         while (cache) {
8109                 if (cache->key.objectid >= (range->start + range->len)) {
8110                         btrfs_put_block_group(cache);
8111                         break;
8112                 }
8113
8114                 start = max(range->start, cache->key.objectid);
8115                 end = min(range->start + range->len,
8116                                 cache->key.objectid + cache->key.offset);
8117
8118                 if (end - start >= range->minlen) {
8119                         if (!block_group_cache_done(cache)) {
8120                                 ret = cache_block_group(cache, NULL, root, 0);
8121                                 if (!ret)
8122                                         wait_block_group_cache_done(cache);
8123                         }
8124                         ret = btrfs_trim_block_group(cache,
8125                                                      &group_trimmed,
8126                                                      start,
8127                                                      end,
8128                                                      range->minlen);
8129
8130                         trimmed += group_trimmed;
8131                         if (ret) {
8132                                 btrfs_put_block_group(cache);
8133                                 break;
8134                         }
8135                 }
8136
8137                 cache = next_block_group(fs_info->tree_root, cache);
8138         }
8139
8140         range->len = trimmed;
8141         return ret;
8142 }