fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include "compat.h"
  28 #include "hash.h"
  29 #include "ctree.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "transaction.h"
  33 #include "volumes.h"
  34 #include "locking.h"
  35 #include "free-space-cache.h"
  36
  37 /*
  38  * control flags for do_chunk_alloc's force field
  39  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  40  * if we really need one.
  41  *
  42  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  43  * if we have very few chunks already allocated.  This is
  44  * used as part of the clustering code to help make sure
  45  * we have a good pool of storage to cluster in, without
  46  * filling the FS with empty chunks
  47  *
  48  * CHUNK_ALLOC_FORCE means it must try to allocate one
  49  *
  50  */
  51 enum {
  52         CHUNK_ALLOC_NO_FORCE = 0,
  53         CHUNK_ALLOC_LIMITED = 1,
  54         CHUNK_ALLOC_FORCE = 2,
  55 };
  56
  57 /*
  58  * Control how reservations are dealt with.
  59  *
  60  * RESERVE_FREE - freeing a reservation.
  61  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  62  *   ENOSPC accounting
  63  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  64  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  65  */
  66 enum {
  67         RESERVE_FREE = 0,
  68         RESERVE_ALLOC = 1,
  69         RESERVE_ALLOC_NO_ACCOUNT = 2,
  70 };
  71
  72 static int update_block_group(struct btrfs_trans_handle *trans,
  73                               struct btrfs_root *root,
  74                               u64 bytenr, u64 num_bytes, int alloc);
  75 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  76                                 struct btrfs_root *root,
  77                                 u64 bytenr, u64 num_bytes, u64 parent,
  78                                 u64 root_objectid, u64 owner_objectid,
  79                                 u64 owner_offset, int refs_to_drop,
  80                                 struct btrfs_delayed_extent_op *extra_op);
  81 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  82                                     struct extent_buffer *leaf,
  83                                     struct btrfs_extent_item *ei);
  84 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  85                                       struct btrfs_root *root,
  86                                       u64 parent, u64 root_objectid,
  87                                       u64 flags, u64 owner, u64 offset,
  88                                       struct btrfs_key *ins, int ref_mod);
  89 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  90                                      struct btrfs_root *root,
  91                                      u64 parent, u64 root_objectid,
  92                                      u64 flags, struct btrfs_disk_key *key,
  93                                      int level, struct btrfs_key *ins);
  94 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  95                           struct btrfs_root *extent_root, u64 alloc_bytes,
  96                           u64 flags, int force);
  97 static int find_next_key(struct btrfs_path *path, int level,
  98                          struct btrfs_key *key);
  99 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 100                             int dump_block_groups);
 101 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 102                                        u64 num_bytes, int reserve);
 103
 104 static noinline int
 105 block_group_cache_done(struct btrfs_block_group_cache *cache)
 106 {
 107         smp_mb();
 108         return cache->cached == BTRFS_CACHE_FINISHED;
 109 }
 110
 111 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 112 {
 113         return (cache->flags & bits) == bits;
 114 }
 115
 116 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 117 {
 118         atomic_inc(&cache->count);
 119 }
 120
 121 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 122 {
 123         if (atomic_dec_and_test(&cache->count)) {
 124                 WARN_ON(cache->pinned > 0);
 125                 WARN_ON(cache->reserved > 0);
 126                 kfree(cache->free_space_ctl);
 127                 kfree(cache);
 128         }
 129 }
 130
 131 /*
 132  * this adds the block group to the fs_info rb tree for the block group
 133  * cache
 134  */
 135 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 136                                 struct btrfs_block_group_cache *block_group)
 137 {
 138         struct rb_node **p;
 139         struct rb_node *parent = NULL;
 140         struct btrfs_block_group_cache *cache;
 141
 142         spin_lock(&info->block_group_cache_lock);
 143         p = &info->block_group_cache_tree.rb_node;
 144
 145         while (*p) {
 146                 parent = *p;
 147                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 148                                  cache_node);
 149                 if (block_group->key.objectid < cache->key.objectid) {
 150                         p = &(*p)->rb_left;
 151                 } else if (block_group->key.objectid > cache->key.objectid) {
 152                         p = &(*p)->rb_right;
 153                 } else {
 154                         spin_unlock(&info->block_group_cache_lock);
 155                         return -EEXIST;
 156                 }
 157         }
 158
 159         rb_link_node(&block_group->cache_node, parent, p);
 160         rb_insert_color(&block_group->cache_node,
 161                         &info->block_group_cache_tree);
 162         spin_unlock(&info->block_group_cache_lock);
 163
 164         return 0;
 165 }
 166
 167 /*
 168  * This will return the block group at or after bytenr if contains is 0, else
 169  * it will return the block group that contains the bytenr
 170  */
 171 static struct btrfs_block_group_cache *
 172 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 173                               int contains)
 174 {
 175         struct btrfs_block_group_cache *cache, *ret = NULL;
 176         struct rb_node *n;
 177         u64 end, start;
 178
 179         spin_lock(&info->block_group_cache_lock);
 180         n = info->block_group_cache_tree.rb_node;
 181
 182         while (n) {
 183                 cache = rb_entry(n, struct btrfs_block_group_cache,
 184                                  cache_node);
 185                 end = cache->key.objectid + cache->key.offset - 1;
 186                 start = cache->key.objectid;
 187
 188                 if (bytenr < start) {
 189                         if (!contains && (!ret || start < ret->key.objectid))
 190                                 ret = cache;
 191                         n = n->rb_left;
 192                 } else if (bytenr > start) {
 193                         if (contains && bytenr <= end) {
 194                                 ret = cache;
 195                                 break;
 196                         }
 197                         n = n->rb_right;
 198                 } else {
 199                         ret = cache;
 200                         break;
 201                 }
 202         }
 203         if (ret)
 204                 btrfs_get_block_group(ret);
 205         spin_unlock(&info->block_group_cache_lock);
 206
 207         return ret;
 208 }
 209
 210 static int add_excluded_extent(struct btrfs_root *root,
 211                                u64 start, u64 num_bytes)
 212 {
 213         u64 end = start + num_bytes - 1;
 214         set_extent_bits(&root->fs_info->freed_extents[0],
 215                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 216         set_extent_bits(&root->fs_info->freed_extents[1],
 217                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 218         return 0;
 219 }
 220
 221 static void free_excluded_extents(struct btrfs_root *root,
 222                                   struct btrfs_block_group_cache *cache)
 223 {
 224         u64 start, end;
 225
 226         start = cache->key.objectid;
 227         end = start + cache->key.offset - 1;
 228
 229         clear_extent_bits(&root->fs_info->freed_extents[0],
 230                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 231         clear_extent_bits(&root->fs_info->freed_extents[1],
 232                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 233 }
 234
 235 static int exclude_super_stripes(struct btrfs_root *root,
 236                                  struct btrfs_block_group_cache *cache)
 237 {
 238         u64 bytenr;
 239         u64 *logical;
 240         int stripe_len;
 241         int i, nr, ret;
 242
 243         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 244                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 245                 cache->bytes_super += stripe_len;
 246                 ret = add_excluded_extent(root, cache->key.objectid,
 247                                           stripe_len);
 248                 BUG_ON(ret); /* -ENOMEM */
 249         }
 250
 251         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 252                 bytenr = btrfs_sb_offset(i);
 253                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 254                                        cache->key.objectid, bytenr,
 255                                        0, &logical, &nr, &stripe_len);
 256                 BUG_ON(ret); /* -ENOMEM */
 257
 258                 while (nr--) {
 259                         cache->bytes_super += stripe_len;
 260                         ret = add_excluded_extent(root, logical[nr],
 261                                                   stripe_len);
 262                         BUG_ON(ret); /* -ENOMEM */
 263                 }
 264
 265                 kfree(logical);
 266         }
 267         return 0;
 268 }
 269
 270 static struct btrfs_caching_control *
 271 get_caching_control(struct btrfs_block_group_cache *cache)
 272 {
 273         struct btrfs_caching_control *ctl;
 274
 275         spin_lock(&cache->lock);
 276         if (cache->cached != BTRFS_CACHE_STARTED) {
 277                 spin_unlock(&cache->lock);
 278                 return NULL;
 279         }
 280
 281         /* We're loading it the fast way, so we don't have a caching_ctl. */
 282         if (!cache->caching_ctl) {
 283                 spin_unlock(&cache->lock);
 284                 return NULL;
 285         }
 286
 287         ctl = cache->caching_ctl;
 288         atomic_inc(&ctl->count);
 289         spin_unlock(&cache->lock);
 290         return ctl;
 291 }
 292
 293 static void put_caching_control(struct btrfs_caching_control *ctl)
 294 {
 295         if (atomic_dec_and_test(&ctl->count))
 296                 kfree(ctl);
 297 }
 298
 299 /*
 300  * this is only called by cache_block_group, since we could have freed extents
 301  * we need to check the pinned_extents for any extents that can't be used yet
 302  * since their free space will be released as soon as the transaction commits.
 303  */
 304 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 305                               struct btrfs_fs_info *info, u64 start, u64 end)
 306 {
 307         u64 extent_start, extent_end, size, total_added = 0;
 308         int ret;
 309
 310         while (start < end) {
 311                 ret = find_first_extent_bit(info->pinned_extents, start,
 312                                             &extent_start, &extent_end,
 313                                             EXTENT_DIRTY | EXTENT_UPTODATE);
 314                 if (ret)
 315                         break;
 316
 317                 if (extent_start <= start) {
 318                         start = extent_end + 1;
 319                 } else if (extent_start > start && extent_start < end) {
 320                         size = extent_start - start;
 321                         total_added += size;
 322                         ret = btrfs_add_free_space(block_group, start,
 323                                                    size);
 324                         BUG_ON(ret); /* -ENOMEM or logic error */
 325                         start = extent_end + 1;
 326                 } else {
 327                         break;
 328                 }
 329         }
 330
 331         if (start < end) {
 332                 size = end - start;
 333                 total_added += size;
 334                 ret = btrfs_add_free_space(block_group, start, size);
 335                 BUG_ON(ret); /* -ENOMEM or logic error */
 336         }
 337
 338         return total_added;
 339 }
 340
 341 static noinline void caching_thread(struct btrfs_work *work)
 342 {
 343         struct btrfs_block_group_cache *block_group;
 344         struct btrfs_fs_info *fs_info;
 345         struct btrfs_caching_control *caching_ctl;
 346         struct btrfs_root *extent_root;
 347         struct btrfs_path *path;
 348         struct extent_buffer *leaf;
 349         struct btrfs_key key;
 350         u64 total_found = 0;
 351         u64 last = 0;
 352         u32 nritems;
 353         int ret = 0;
 354
 355         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 356         block_group = caching_ctl->block_group;
 357         fs_info = block_group->fs_info;
 358         extent_root = fs_info->extent_root;
 359
 360         path = btrfs_alloc_path();
 361         if (!path)
 362                 goto out;
 363
 364         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 365
 366         /*
 367          * We don't want to deadlock with somebody trying to allocate a new
 368          * extent for the extent root while also trying to search the extent
 369          * root to add free space.  So we skip locking and search the commit
 370          * root, since its read-only
 371          */
 372         path->skip_locking = 1;
 373         path->search_commit_root = 1;
 374         path->reada = 1;
 375
 376         key.objectid = last;
 377         key.offset = 0;
 378         key.type = BTRFS_EXTENT_ITEM_KEY;
 379 again:
 380         mutex_lock(&caching_ctl->mutex);
 381         /* need to make sure the commit_root doesn't disappear */
 382         down_read(&fs_info->extent_commit_sem);
 383
 384         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 385         if (ret < 0)
 386                 goto err;
 387
 388         leaf = path->nodes[0];
 389         nritems = btrfs_header_nritems(leaf);
 390
 391         while (1) {
 392                 if (btrfs_fs_closing(fs_info) > 1) {
 393                         last = (u64)-1;
 394                         break;
 395                 }
 396
 397                 if (path->slots[0] < nritems) {
 398                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 399                 } else {
 400                         ret = find_next_key(path, 0, &key);
 401                         if (ret)
 402                                 break;
 403
 404                         if (need_resched() ||
 405                             btrfs_next_leaf(extent_root, path)) {
 406                                 caching_ctl->progress = last;
 407                                 btrfs_release_path(path);
 408                                 up_read(&fs_info->extent_commit_sem);
 409                                 mutex_unlock(&caching_ctl->mutex);
 410                                 cond_resched();
 411                                 goto again;
 412                         }
 413                         leaf = path->nodes[0];
 414                         nritems = btrfs_header_nritems(leaf);
 415                         continue;
 416                 }
 417
 418                 if (key.objectid < block_group->key.objectid) {
 419                         path->slots[0]++;
 420                         continue;
 421                 }
 422
 423                 if (key.objectid >= block_group->key.objectid +
 424                     block_group->key.offset)
 425                         break;
 426
 427                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 428                         total_found += add_new_free_space(block_group,
 429                                                           fs_info, last,
 430                                                           key.objectid);
 431                         last = key.objectid + key.offset;
 432
 433                         if (total_found > (1024 * 1024 * 2)) {
 434                                 total_found = 0;
 435                                 wake_up(&caching_ctl->wait);
 436                         }
 437                 }
 438                 path->slots[0]++;
 439         }
 440         ret = 0;
 441
 442         total_found += add_new_free_space(block_group, fs_info, last,
 443                                           block_group->key.objectid +
 444                                           block_group->key.offset);
 445         caching_ctl->progress = (u64)-1;
 446
 447         spin_lock(&block_group->lock);
 448         block_group->caching_ctl = NULL;
 449         block_group->cached = BTRFS_CACHE_FINISHED;
 450         spin_unlock(&block_group->lock);
 451
 452 err:
 453         btrfs_free_path(path);
 454         up_read(&fs_info->extent_commit_sem);
 455
 456         free_excluded_extents(extent_root, block_group);
 457
 458         mutex_unlock(&caching_ctl->mutex);
 459 out:
 460         wake_up(&caching_ctl->wait);
 461
 462         put_caching_control(caching_ctl);
 463         btrfs_put_block_group(block_group);
 464 }
 465
 466 static int cache_block_group(struct btrfs_block_group_cache *cache,
 467                              struct btrfs_trans_handle *trans,
 468                              struct btrfs_root *root,
 469                              int load_cache_only)
 470 {
 471         DEFINE_WAIT(wait);
 472         struct btrfs_fs_info *fs_info = cache->fs_info;
 473         struct btrfs_caching_control *caching_ctl;
 474         int ret = 0;
 475
 476         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 477         if (!caching_ctl)
 478                 return -ENOMEM;
 479
 480         INIT_LIST_HEAD(&caching_ctl->list);
 481         mutex_init(&caching_ctl->mutex);
 482         init_waitqueue_head(&caching_ctl->wait);
 483         caching_ctl->block_group = cache;
 484         caching_ctl->progress = cache->key.objectid;
 485         atomic_set(&caching_ctl->count, 1);
 486         caching_ctl->work.func = caching_thread;
 487
 488         spin_lock(&cache->lock);
 489         /*
 490          * This should be a rare occasion, but this could happen I think in the
 491          * case where one thread starts to load the space cache info, and then
 492          * some other thread starts a transaction commit which tries to do an
 493          * allocation while the other thread is still loading the space cache
 494          * info.  The previous loop should have kept us from choosing this block
 495          * group, but if we've moved to the state where we will wait on caching
 496          * block groups we need to first check if we're doing a fast load here,
 497          * so we can wait for it to finish, otherwise we could end up allocating
 498          * from a block group who's cache gets evicted for one reason or
 499          * another.
 500          */
 501         while (cache->cached == BTRFS_CACHE_FAST) {
 502                 struct btrfs_caching_control *ctl;
 503
 504                 ctl = cache->caching_ctl;
 505                 atomic_inc(&ctl->count);
 506                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 507                 spin_unlock(&cache->lock);
 508
 509                 schedule();
 510
 511                 finish_wait(&ctl->wait, &wait);
 512                 put_caching_control(ctl);
 513                 spin_lock(&cache->lock);
 514         }
 515
 516         if (cache->cached != BTRFS_CACHE_NO) {
 517                 spin_unlock(&cache->lock);
 518                 kfree(caching_ctl);
 519                 return 0;
 520         }
 521         WARN_ON(cache->caching_ctl);
 522         cache->caching_ctl = caching_ctl;
 523         cache->cached = BTRFS_CACHE_FAST;
 524         spin_unlock(&cache->lock);
 525
 526         /*
 527          * We can't do the read from on-disk cache during a commit since we need
 528          * to have the normal tree locking.  Also if we are currently trying to
 529          * allocate blocks for the tree root we can't do the fast caching since
 530          * we likely hold important locks.
 531          */
 532         if (trans && (!trans->transaction->in_commit) &&
 533             (root && root != root->fs_info->tree_root) &&
 534             btrfs_test_opt(root, 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         int ret;
2223         int count = 0;
2224         int must_insert_reserved = 0;
2225
2226         delayed_refs = &trans->transaction->delayed_refs;
2227         while (1) {
2228                 if (!locked_ref) {
2229                         /* pick a new head ref from the cluster list */
2230                         if (list_empty(cluster))
2231                                 break;
2232
2233                         locked_ref = list_entry(cluster->next,
2234                                      struct btrfs_delayed_ref_head, cluster);
2235
2236                         /* grab the lock that says we are going to process
2237                          * all the refs for this head */
2238                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2239
2240                         /*
2241                          * we may have dropped the spin lock to get the head
2242                          * mutex lock, and that might have given someone else
2243                          * time to free the head.  If that's true, it has been
2244                          * removed from our list and we can move on.
2245                          */
2246                         if (ret == -EAGAIN) {
2247                                 locked_ref = NULL;
2248                                 count++;
2249                                 continue;
2250                         }
2251                 }
2252
2253                 /*
2254                  * locked_ref is the head node, so we have to go one
2255                  * node back for any delayed ref updates
2256                  */
2257                 ref = select_delayed_ref(locked_ref);
2258
2259                 if (ref && ref->seq &&
2260                     btrfs_check_delayed_seq(delayed_refs, ref->seq)) {
2261                         /*
2262                          * there are still refs with lower seq numbers in the
2263                          * process of being added. Don't run this ref yet.
2264                          */
2265                         list_del_init(&locked_ref->cluster);
2266                         mutex_unlock(&locked_ref->mutex);
2267                         locked_ref = NULL;
2268                         delayed_refs->num_heads_ready++;
2269                         spin_unlock(&delayed_refs->lock);
2270                         cond_resched();
2271                         spin_lock(&delayed_refs->lock);
2272                         continue;
2273                 }
2274
2275                 /*
2276                  * record the must insert reserved flag before we
2277                  * drop the spin lock.
2278                  */
2279                 must_insert_reserved = locked_ref->must_insert_reserved;
2280                 locked_ref->must_insert_reserved = 0;
2281
2282                 extent_op = locked_ref->extent_op;
2283                 locked_ref->extent_op = NULL;
2284
2285                 if (!ref) {
2286                         /* All delayed refs have been processed, Go ahead
2287                          * and send the head node to run_one_delayed_ref,
2288                          * so that any accounting fixes can happen
2289                          */
2290                         ref = &locked_ref->node;
2291
2292                         if (extent_op && must_insert_reserved) {
2293                                 kfree(extent_op);
2294                                 extent_op = NULL;
2295                         }
2296
2297                         if (extent_op) {
2298                                 spin_unlock(&delayed_refs->lock);
2299
2300                                 ret = run_delayed_extent_op(trans, root,
2301                                                             ref, extent_op);
2302                                 kfree(extent_op);
2303
2304                                 if (ret) {
2305                                         printk(KERN_DEBUG "btrfs: run_delayed_extent_op returned %d\n", ret);
2306                                         return ret;
2307                                 }
2308
2309                                 goto next;
2310                         }
2311
2312                         list_del_init(&locked_ref->cluster);
2313                         locked_ref = NULL;
2314                 }
2315
2316                 ref->in_tree = 0;
2317                 rb_erase(&ref->rb_node, &delayed_refs->root);
2318                 delayed_refs->num_entries--;
2319                 /*
2320                  * we modified num_entries, but as we're currently running
2321                  * delayed refs, skip
2322                  *     wake_up(&delayed_refs->seq_wait);
2323                  * here.
2324                  */
2325                 spin_unlock(&delayed_refs->lock);
2326
2327                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2328                                           must_insert_reserved);
2329
2330                 btrfs_put_delayed_ref(ref);
2331                 kfree(extent_op);
2332                 count++;
2333
2334                 if (ret) {
2335                         printk(KERN_DEBUG "btrfs: run_one_delayed_ref returned %d\n", ret);
2336                         return ret;
2337                 }
2338
2339 next:
2340                 do_chunk_alloc(trans, root->fs_info->extent_root,
2341                                2 * 1024 * 1024,
2342                                btrfs_get_alloc_profile(root, 0),
2343                                CHUNK_ALLOC_NO_FORCE);
2344                 cond_resched();
2345                 spin_lock(&delayed_refs->lock);
2346         }
2347         return count;
2348 }
2349
2350
2351 static void wait_for_more_refs(struct btrfs_delayed_ref_root *delayed_refs,
2352                         unsigned long num_refs)
2353 {
2354         struct list_head *first_seq = delayed_refs->seq_head.next;
2355
2356         spin_unlock(&delayed_refs->lock);
2357         pr_debug("waiting for more refs (num %ld, first %p)\n",
2358                  num_refs, first_seq);
2359         wait_event(delayed_refs->seq_wait,
2360                    num_refs != delayed_refs->num_entries ||
2361                    delayed_refs->seq_head.next != first_seq);
2362         pr_debug("done waiting for more refs (num %ld, first %p)\n",
2363                  delayed_refs->num_entries, delayed_refs->seq_head.next);
2364         spin_lock(&delayed_refs->lock);
2365 }
2366
2367 /*
2368  * this starts processing the delayed reference count updates and
2369  * extent insertions we have queued up so far.  count can be
2370  * 0, which means to process everything in the tree at the start
2371  * of the run (but not newly added entries), or it can be some target
2372  * number you'd like to process.
2373  *
2374  * Returns 0 on success or if called with an aborted transaction
2375  * Returns <0 on error and aborts the transaction
2376  */
2377 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2378                            struct btrfs_root *root, unsigned long count)
2379 {
2380         struct rb_node *node;
2381         struct btrfs_delayed_ref_root *delayed_refs;
2382         struct btrfs_delayed_ref_node *ref;
2383         struct list_head cluster;
2384         int ret;
2385         u64 delayed_start;
2386         int run_all = count == (unsigned long)-1;
2387         int run_most = 0;
2388         unsigned long num_refs = 0;
2389         int consider_waiting;
2390
2391         /* We'll clean this up in btrfs_cleanup_transaction */
2392         if (trans->aborted)
2393                 return 0;
2394
2395         if (root == root->fs_info->extent_root)
2396                 root = root->fs_info->tree_root;
2397
2398         do_chunk_alloc(trans, root->fs_info->extent_root,
2399                        2 * 1024 * 1024, btrfs_get_alloc_profile(root, 0),
2400                        CHUNK_ALLOC_NO_FORCE);
2401
2402         delayed_refs = &trans->transaction->delayed_refs;
2403         INIT_LIST_HEAD(&cluster);
2404 again:
2405         consider_waiting = 0;
2406         spin_lock(&delayed_refs->lock);
2407         if (count == 0) {
2408                 count = delayed_refs->num_entries * 2;
2409                 run_most = 1;
2410         }
2411         while (1) {
2412                 if (!(run_all || run_most) &&
2413                     delayed_refs->num_heads_ready < 64)
2414                         break;
2415
2416                 /*
2417                  * go find something we can process in the rbtree.  We start at
2418                  * the beginning of the tree, and then build a cluster
2419                  * of refs to process starting at the first one we are able to
2420                  * lock
2421                  */
2422                 delayed_start = delayed_refs->run_delayed_start;
2423                 ret = btrfs_find_ref_cluster(trans, &cluster,
2424                                              delayed_refs->run_delayed_start);
2425                 if (ret)
2426                         break;
2427
2428                 if (delayed_start >= delayed_refs->run_delayed_start) {
2429                         if (consider_waiting == 0) {
2430                                 /*
2431                                  * btrfs_find_ref_cluster looped. let's do one
2432                                  * more cycle. if we don't run any delayed ref
2433                                  * during that cycle (because we can't because
2434                                  * all of them are blocked) and if the number of
2435                                  * refs doesn't change, we avoid busy waiting.
2436                                  */
2437                                 consider_waiting = 1;
2438                                 num_refs = delayed_refs->num_entries;
2439                         } else {
2440                                 wait_for_more_refs(delayed_refs, num_refs);
2441                                 /*
2442                                  * after waiting, things have changed. we
2443                                  * dropped the lock and someone else might have
2444                                  * run some refs, built new clusters and so on.
2445                                  * therefore, we restart staleness detection.
2446                                  */
2447                                 consider_waiting = 0;
2448                         }
2449                 }
2450
2451                 ret = run_clustered_refs(trans, root, &cluster);
2452                 if (ret < 0) {
2453                         spin_unlock(&delayed_refs->lock);
2454                         btrfs_abort_transaction(trans, root, ret);
2455                         return ret;
2456                 }
2457
2458                 count -= min_t(unsigned long, ret, count);
2459
2460                 if (count == 0)
2461                         break;
2462
2463                 if (ret || delayed_refs->run_delayed_start == 0) {
2464                         /* refs were run, let's reset staleness detection */
2465                         consider_waiting = 0;
2466                 }
2467         }
2468
2469         if (run_all) {
2470                 node = rb_first(&delayed_refs->root);
2471                 if (!node)
2472                         goto out;
2473                 count = (unsigned long)-1;
2474
2475                 while (node) {
2476                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2477                                        rb_node);
2478                         if (btrfs_delayed_ref_is_head(ref)) {
2479                                 struct btrfs_delayed_ref_head *head;
2480
2481                                 head = btrfs_delayed_node_to_head(ref);
2482                                 atomic_inc(&ref->refs);
2483
2484                                 spin_unlock(&delayed_refs->lock);
2485                                 /*
2486                                  * Mutex was contended, block until it's
2487                                  * released and try again
2488                                  */
2489                                 mutex_lock(&head->mutex);
2490                                 mutex_unlock(&head->mutex);
2491
2492                                 btrfs_put_delayed_ref(ref);
2493                                 cond_resched();
2494                                 goto again;
2495                         }
2496                         node = rb_next(node);
2497                 }
2498                 spin_unlock(&delayed_refs->lock);
2499                 schedule_timeout(1);
2500                 goto again;
2501         }
2502 out:
2503         spin_unlock(&delayed_refs->lock);
2504         return 0;
2505 }
2506
2507 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2508                                 struct btrfs_root *root,
2509                                 u64 bytenr, u64 num_bytes, u64 flags,
2510                                 int is_data)
2511 {
2512         struct btrfs_delayed_extent_op *extent_op;
2513         int ret;
2514
2515         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2516         if (!extent_op)
2517                 return -ENOMEM;
2518
2519         extent_op->flags_to_set = flags;
2520         extent_op->update_flags = 1;
2521         extent_op->update_key = 0;
2522         extent_op->is_data = is_data ? 1 : 0;
2523
2524         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2525                                           num_bytes, extent_op);
2526         if (ret)
2527                 kfree(extent_op);
2528         return ret;
2529 }
2530
2531 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2532                                       struct btrfs_root *root,
2533                                       struct btrfs_path *path,
2534                                       u64 objectid, u64 offset, u64 bytenr)
2535 {
2536         struct btrfs_delayed_ref_head *head;
2537         struct btrfs_delayed_ref_node *ref;
2538         struct btrfs_delayed_data_ref *data_ref;
2539         struct btrfs_delayed_ref_root *delayed_refs;
2540         struct rb_node *node;
2541         int ret = 0;
2542
2543         ret = -ENOENT;
2544         delayed_refs = &trans->transaction->delayed_refs;
2545         spin_lock(&delayed_refs->lock);
2546         head = btrfs_find_delayed_ref_head(trans, bytenr);
2547         if (!head)
2548                 goto out;
2549
2550         if (!mutex_trylock(&head->mutex)) {
2551                 atomic_inc(&head->node.refs);
2552                 spin_unlock(&delayed_refs->lock);
2553
2554                 btrfs_release_path(path);
2555
2556                 /*
2557                  * Mutex was contended, block until it's released and let
2558                  * caller try again
2559                  */
2560                 mutex_lock(&head->mutex);
2561                 mutex_unlock(&head->mutex);
2562                 btrfs_put_delayed_ref(&head->node);
2563                 return -EAGAIN;
2564         }
2565
2566         node = rb_prev(&head->node.rb_node);
2567         if (!node)
2568                 goto out_unlock;
2569
2570         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2571
2572         if (ref->bytenr != bytenr)
2573                 goto out_unlock;
2574
2575         ret = 1;
2576         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2577                 goto out_unlock;
2578
2579         data_ref = btrfs_delayed_node_to_data_ref(ref);
2580
2581         node = rb_prev(node);
2582         if (node) {
2583                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2584                 if (ref->bytenr == bytenr)
2585                         goto out_unlock;
2586         }
2587
2588         if (data_ref->root != root->root_key.objectid ||
2589             data_ref->objectid != objectid || data_ref->offset != offset)
2590                 goto out_unlock;
2591
2592         ret = 0;
2593 out_unlock:
2594         mutex_unlock(&head->mutex);
2595 out:
2596         spin_unlock(&delayed_refs->lock);
2597         return ret;
2598 }
2599
2600 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2601                                         struct btrfs_root *root,
2602                                         struct btrfs_path *path,
2603                                         u64 objectid, u64 offset, u64 bytenr)
2604 {
2605         struct btrfs_root *extent_root = root->fs_info->extent_root;
2606         struct extent_buffer *leaf;
2607         struct btrfs_extent_data_ref *ref;
2608         struct btrfs_extent_inline_ref *iref;
2609         struct btrfs_extent_item *ei;
2610         struct btrfs_key key;
2611         u32 item_size;
2612         int ret;
2613
2614         key.objectid = bytenr;
2615         key.offset = (u64)-1;
2616         key.type = BTRFS_EXTENT_ITEM_KEY;
2617
2618         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2619         if (ret < 0)
2620                 goto out;
2621         BUG_ON(ret == 0); /* Corruption */
2622
2623         ret = -ENOENT;
2624         if (path->slots[0] == 0)
2625                 goto out;
2626
2627         path->slots[0]--;
2628         leaf = path->nodes[0];
2629         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2630
2631         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2632                 goto out;
2633
2634         ret = 1;
2635         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2636 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2637         if (item_size < sizeof(*ei)) {
2638                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2639                 goto out;
2640         }
2641 #endif
2642         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2643
2644         if (item_size != sizeof(*ei) +
2645             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2646                 goto out;
2647
2648         if (btrfs_extent_generation(leaf, ei) <=
2649             btrfs_root_last_snapshot(&root->root_item))
2650                 goto out;
2651
2652         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2653         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2654             BTRFS_EXTENT_DATA_REF_KEY)
2655                 goto out;
2656
2657         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2658         if (btrfs_extent_refs(leaf, ei) !=
2659             btrfs_extent_data_ref_count(leaf, ref) ||
2660             btrfs_extent_data_ref_root(leaf, ref) !=
2661             root->root_key.objectid ||
2662             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2663             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2664                 goto out;
2665
2666         ret = 0;
2667 out:
2668         return ret;
2669 }
2670
2671 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2672                           struct btrfs_root *root,
2673                           u64 objectid, u64 offset, u64 bytenr)
2674 {
2675         struct btrfs_path *path;
2676         int ret;
2677         int ret2;
2678
2679         path = btrfs_alloc_path();
2680         if (!path)
2681                 return -ENOENT;
2682
2683         do {
2684                 ret = check_committed_ref(trans, root, path, objectid,
2685                                           offset, bytenr);
2686                 if (ret && ret != -ENOENT)
2687                         goto out;
2688
2689                 ret2 = check_delayed_ref(trans, root, path, objectid,
2690                                          offset, bytenr);
2691         } while (ret2 == -EAGAIN);
2692
2693         if (ret2 && ret2 != -ENOENT) {
2694                 ret = ret2;
2695                 goto out;
2696         }
2697
2698         if (ret != -ENOENT || ret2 != -ENOENT)
2699                 ret = 0;
2700 out:
2701         btrfs_free_path(path);
2702         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2703                 WARN_ON(ret > 0);
2704         return ret;
2705 }
2706
2707 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2708                            struct btrfs_root *root,
2709                            struct extent_buffer *buf,
2710                            int full_backref, int inc, int for_cow)
2711 {
2712         u64 bytenr;
2713         u64 num_bytes;
2714         u64 parent;
2715         u64 ref_root;
2716         u32 nritems;
2717         struct btrfs_key key;
2718         struct btrfs_file_extent_item *fi;
2719         int i;
2720         int level;
2721         int ret = 0;
2722         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2723                             u64, u64, u64, u64, u64, u64, int);
2724
2725         ref_root = btrfs_header_owner(buf);
2726         nritems = btrfs_header_nritems(buf);
2727         level = btrfs_header_level(buf);
2728
2729         if (!root->ref_cows && level == 0)
2730                 return 0;
2731
2732         if (inc)
2733                 process_func = btrfs_inc_extent_ref;
2734         else
2735                 process_func = btrfs_free_extent;
2736
2737         if (full_backref)
2738                 parent = buf->start;
2739         else
2740                 parent = 0;
2741
2742         for (i = 0; i < nritems; i++) {
2743                 if (level == 0) {
2744                         btrfs_item_key_to_cpu(buf, &key, i);
2745                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2746                                 continue;
2747                         fi = btrfs_item_ptr(buf, i,
2748                                             struct btrfs_file_extent_item);
2749                         if (btrfs_file_extent_type(buf, fi) ==
2750                             BTRFS_FILE_EXTENT_INLINE)
2751                                 continue;
2752                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2753                         if (bytenr == 0)
2754                                 continue;
2755
2756                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2757                         key.offset -= btrfs_file_extent_offset(buf, fi);
2758                         ret = process_func(trans, root, bytenr, num_bytes,
2759                                            parent, ref_root, key.objectid,
2760                                            key.offset, for_cow);
2761                         if (ret)
2762                                 goto fail;
2763                 } else {
2764                         bytenr = btrfs_node_blockptr(buf, i);
2765                         num_bytes = btrfs_level_size(root, level - 1);
2766                         ret = process_func(trans, root, bytenr, num_bytes,
2767                                            parent, ref_root, level - 1, 0,
2768                                            for_cow);
2769                         if (ret)
2770                                 goto fail;
2771                 }
2772         }
2773         return 0;
2774 fail:
2775         return ret;
2776 }
2777
2778 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2779                   struct extent_buffer *buf, int full_backref, int for_cow)
2780 {
2781         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2782 }
2783
2784 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2785                   struct extent_buffer *buf, int full_backref, int for_cow)
2786 {
2787         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2788 }
2789
2790 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2791                                  struct btrfs_root *root,
2792                                  struct btrfs_path *path,
2793                                  struct btrfs_block_group_cache *cache)
2794 {
2795         int ret;
2796         struct btrfs_root *extent_root = root->fs_info->extent_root;
2797         unsigned long bi;
2798         struct extent_buffer *leaf;
2799
2800         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2801         if (ret < 0)
2802                 goto fail;
2803         BUG_ON(ret); /* Corruption */
2804
2805         leaf = path->nodes[0];
2806         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2807         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2808         btrfs_mark_buffer_dirty(leaf);
2809         btrfs_release_path(path);
2810 fail:
2811         if (ret) {
2812                 btrfs_abort_transaction(trans, root, ret);
2813                 return ret;
2814         }
2815         return 0;
2816
2817 }
2818
2819 static struct btrfs_block_group_cache *
2820 next_block_group(struct btrfs_root *root,
2821                  struct btrfs_block_group_cache *cache)
2822 {
2823         struct rb_node *node;
2824         spin_lock(&root->fs_info->block_group_cache_lock);
2825         node = rb_next(&cache->cache_node);
2826         btrfs_put_block_group(cache);
2827         if (node) {
2828                 cache = rb_entry(node, struct btrfs_block_group_cache,
2829                                  cache_node);
2830                 btrfs_get_block_group(cache);
2831         } else
2832                 cache = NULL;
2833         spin_unlock(&root->fs_info->block_group_cache_lock);
2834         return cache;
2835 }
2836
2837 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2838                             struct btrfs_trans_handle *trans,
2839                             struct btrfs_path *path)
2840 {
2841         struct btrfs_root *root = block_group->fs_info->tree_root;
2842         struct inode *inode = NULL;
2843         u64 alloc_hint = 0;
2844         int dcs = BTRFS_DC_ERROR;
2845         int num_pages = 0;
2846         int retries = 0;
2847         int ret = 0;
2848
2849         /*
2850          * If this block group is smaller than 100 megs don't bother caching the
2851          * block group.
2852          */
2853         if (block_group->key.offset < (100 * 1024 * 1024)) {
2854                 spin_lock(&block_group->lock);
2855                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2856                 spin_unlock(&block_group->lock);
2857                 return 0;
2858         }
2859
2860 again:
2861         inode = lookup_free_space_inode(root, block_group, path);
2862         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2863                 ret = PTR_ERR(inode);
2864                 btrfs_release_path(path);
2865                 goto out;
2866         }
2867
2868         if (IS_ERR(inode)) {
2869                 BUG_ON(retries);
2870                 retries++;
2871
2872                 if (block_group->ro)
2873                         goto out_free;
2874
2875                 ret = create_free_space_inode(root, trans, block_group, path);
2876                 if (ret)
2877                         goto out_free;
2878                 goto again;
2879         }
2880
2881         /* We've already setup this transaction, go ahead and exit */
2882         if (block_group->cache_generation == trans->transid &&
2883             i_size_read(inode)) {
2884                 dcs = BTRFS_DC_SETUP;
2885                 goto out_put;
2886         }
2887
2888         /*
2889          * We want to set the generation to 0, that way if anything goes wrong
2890          * from here on out we know not to trust this cache when we load up next
2891          * time.
2892          */
2893         BTRFS_I(inode)->generation = 0;
2894         ret = btrfs_update_inode(trans, root, inode);
2895         WARN_ON(ret);
2896
2897         if (i_size_read(inode) > 0) {
2898                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2899                                                       inode);
2900                 if (ret)
2901                         goto out_put;
2902         }
2903
2904         spin_lock(&block_group->lock);
2905         if (block_group->cached != BTRFS_CACHE_FINISHED) {
2906                 /* We're not cached, don't bother trying to write stuff out */
2907                 dcs = BTRFS_DC_WRITTEN;
2908                 spin_unlock(&block_group->lock);
2909                 goto out_put;
2910         }
2911         spin_unlock(&block_group->lock);
2912
2913         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2914         if (!num_pages)
2915                 num_pages = 1;
2916
2917         /*
2918          * Just to make absolutely sure we have enough space, we're going to
2919          * preallocate 12 pages worth of space for each block group.  In
2920          * practice we ought to use at most 8, but we need extra space so we can
2921          * add our header and have a terminator between the extents and the
2922          * bitmaps.
2923          */
2924         num_pages *= 16;
2925         num_pages *= PAGE_CACHE_SIZE;
2926
2927         ret = btrfs_check_data_free_space(inode, num_pages);
2928         if (ret)
2929                 goto out_put;
2930
2931         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2932                                               num_pages, num_pages,
2933                                               &alloc_hint);
2934         if (!ret)
2935                 dcs = BTRFS_DC_SETUP;
2936         btrfs_free_reserved_data_space(inode, num_pages);
2937
2938 out_put:
2939         iput(inode);
2940 out_free:
2941         btrfs_release_path(path);
2942 out:
2943         spin_lock(&block_group->lock);
2944         if (!ret && dcs == BTRFS_DC_SETUP)
2945                 block_group->cache_generation = trans->transid;
2946         block_group->disk_cache_state = dcs;
2947         spin_unlock(&block_group->lock);
2948
2949         return ret;
2950 }
2951
2952 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2953                                    struct btrfs_root *root)
2954 {
2955         struct btrfs_block_group_cache *cache;
2956         int err = 0;
2957         struct btrfs_path *path;
2958         u64 last = 0;
2959
2960         path = btrfs_alloc_path();
2961         if (!path)
2962                 return -ENOMEM;
2963
2964 again:
2965         while (1) {
2966                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2967                 while (cache) {
2968                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2969                                 break;
2970                         cache = next_block_group(root, cache);
2971                 }
2972                 if (!cache) {
2973                         if (last == 0)
2974                                 break;
2975                         last = 0;
2976                         continue;
2977                 }
2978                 err = cache_save_setup(cache, trans, path);
2979                 last = cache->key.objectid + cache->key.offset;
2980                 btrfs_put_block_group(cache);
2981         }
2982
2983         while (1) {
2984                 if (last == 0) {
2985                         err = btrfs_run_delayed_refs(trans, root,
2986                                                      (unsigned long)-1);
2987                         if (err) /* File system offline */
2988                                 goto out;
2989                 }
2990
2991                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2992                 while (cache) {
2993                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
2994                                 btrfs_put_block_group(cache);
2995                                 goto again;
2996                         }
2997
2998                         if (cache->dirty)
2999                                 break;
3000                         cache = next_block_group(root, cache);
3001                 }
3002                 if (!cache) {
3003                         if (last == 0)
3004                                 break;
3005                         last = 0;
3006                         continue;
3007                 }
3008
3009                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3010                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3011                 cache->dirty = 0;
3012                 last = cache->key.objectid + cache->key.offset;
3013
3014                 err = write_one_cache_group(trans, root, path, cache);
3015                 if (err) /* File system offline */
3016                         goto out;
3017
3018                 btrfs_put_block_group(cache);
3019         }
3020
3021         while (1) {
3022                 /*
3023                  * I don't think this is needed since we're just marking our
3024                  * preallocated extent as written, but just in case it can't
3025                  * hurt.
3026                  */
3027                 if (last == 0) {
3028                         err = btrfs_run_delayed_refs(trans, root,
3029                                                      (unsigned long)-1);
3030                         if (err) /* File system offline */
3031                                 goto out;
3032                 }
3033
3034                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3035                 while (cache) {
3036                         /*
3037                          * Really this shouldn't happen, but it could if we
3038                          * couldn't write the entire preallocated extent and
3039                          * splitting the extent resulted in a new block.
3040                          */
3041                         if (cache->dirty) {
3042                                 btrfs_put_block_group(cache);
3043                                 goto again;
3044                         }
3045                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3046                                 break;
3047                         cache = next_block_group(root, cache);
3048                 }
3049                 if (!cache) {
3050                         if (last == 0)
3051                                 break;
3052                         last = 0;
3053                         continue;
3054                 }
3055
3056                 err = btrfs_write_out_cache(root, trans, cache, path);
3057
3058                 /*
3059                  * If we didn't have an error then the cache state is still
3060                  * NEED_WRITE, so we can set it to WRITTEN.
3061                  */
3062                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3063                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3064                 last = cache->key.objectid + cache->key.offset;
3065                 btrfs_put_block_group(cache);
3066         }
3067 out:
3068
3069         btrfs_free_path(path);
3070         return err;
3071 }
3072
3073 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3074 {
3075         struct btrfs_block_group_cache *block_group;
3076         int readonly = 0;
3077
3078         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3079         if (!block_group || block_group->ro)
3080                 readonly = 1;
3081         if (block_group)
3082                 btrfs_put_block_group(block_group);
3083         return readonly;
3084 }
3085
3086 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3087                              u64 total_bytes, u64 bytes_used,
3088                              struct btrfs_space_info **space_info)
3089 {
3090         struct btrfs_space_info *found;
3091         int i;
3092         int factor;
3093
3094         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3095                      BTRFS_BLOCK_GROUP_RAID10))
3096                 factor = 2;
3097         else
3098                 factor = 1;
3099
3100         found = __find_space_info(info, flags);
3101         if (found) {
3102                 spin_lock(&found->lock);
3103                 found->total_bytes += total_bytes;
3104                 found->disk_total += total_bytes * factor;
3105                 found->bytes_used += bytes_used;
3106                 found->disk_used += bytes_used * factor;
3107                 found->full = 0;
3108                 spin_unlock(&found->lock);
3109                 *space_info = found;
3110                 return 0;
3111         }
3112         found = kzalloc(sizeof(*found), GFP_NOFS);
3113         if (!found)
3114                 return -ENOMEM;
3115
3116         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3117                 INIT_LIST_HEAD(&found->block_groups[i]);
3118         init_rwsem(&found->groups_sem);
3119         spin_lock_init(&found->lock);
3120         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3121         found->total_bytes = total_bytes;
3122         found->disk_total = total_bytes * factor;
3123         found->bytes_used = bytes_used;
3124         found->disk_used = bytes_used * factor;
3125         found->bytes_pinned = 0;
3126         found->bytes_reserved = 0;
3127         found->bytes_readonly = 0;
3128         found->bytes_may_use = 0;
3129         found->full = 0;
3130         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3131         found->chunk_alloc = 0;
3132         found->flush = 0;
3133         init_waitqueue_head(&found->wait);
3134         *space_info = found;
3135         list_add_rcu(&found->list, &info->space_info);
3136         return 0;
3137 }
3138
3139 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3140 {
3141         u64 extra_flags = flags & BTRFS_BLOCK_GROUP_PROFILE_MASK;
3142
3143         /* chunk -> extended profile */
3144         if (extra_flags == 0)
3145                 extra_flags = BTRFS_AVAIL_ALLOC_BIT_SINGLE;
3146
3147         if (flags & BTRFS_BLOCK_GROUP_DATA)
3148                 fs_info->avail_data_alloc_bits |= extra_flags;
3149         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3150                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3151         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3152                 fs_info->avail_system_alloc_bits |= extra_flags;
3153 }
3154
3155 /*
3156  * @flags: available profiles in extended format (see ctree.h)
3157  *
3158  * Returns reduced profile in chunk format.  If profile changing is in
3159  * progress (either running or paused) picks the target profile (if it's
3160  * already available), otherwise falls back to plain reducing.
3161  */
3162 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3163 {
3164         /*
3165          * we add in the count of missing devices because we want
3166          * to make sure that any RAID levels on a degraded FS
3167          * continue to be honored.
3168          */
3169         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3170                 root->fs_info->fs_devices->missing_devices;
3171
3172         /* pick restriper's target profile if it's available */
3173         spin_lock(&root->fs_info->balance_lock);
3174         if (root->fs_info->balance_ctl) {
3175                 struct btrfs_balance_control *bctl = root->fs_info->balance_ctl;
3176                 u64 tgt = 0;
3177
3178                 if ((flags & BTRFS_BLOCK_GROUP_DATA) &&
3179                     (bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) &&
3180                     (flags & bctl->data.target)) {
3181                         tgt = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3182                 } else if ((flags & BTRFS_BLOCK_GROUP_SYSTEM) &&
3183                            (bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) &&
3184                            (flags & bctl->sys.target)) {
3185                         tgt = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3186                 } else if ((flags & BTRFS_BLOCK_GROUP_METADATA) &&
3187                            (bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) &&
3188                            (flags & bctl->meta.target)) {
3189                         tgt = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3190                 }
3191
3192                 if (tgt) {
3193                         spin_unlock(&root->fs_info->balance_lock);
3194                         flags = tgt;
3195                         goto out;
3196                 }
3197         }
3198         spin_unlock(&root->fs_info->balance_lock);
3199
3200         if (num_devices == 1)
3201                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3202         if (num_devices < 4)
3203                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3204
3205         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3206             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3207                       BTRFS_BLOCK_GROUP_RAID10))) {
3208                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3209         }
3210
3211         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3212             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3213                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3214         }
3215
3216         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3217             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3218              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3219              (flags & BTRFS_BLOCK_GROUP_DUP))) {
3220                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3221         }
3222
3223 out:
3224         /* extended -> chunk profile */
3225         flags &= ~BTRFS_AVAIL_ALLOC_BIT_SINGLE;
3226         return flags;
3227 }
3228
3229 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3230 {
3231         if (flags & BTRFS_BLOCK_GROUP_DATA)
3232                 flags |= root->fs_info->avail_data_alloc_bits;
3233         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3234                 flags |= root->fs_info->avail_system_alloc_bits;
3235         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3236                 flags |= root->fs_info->avail_metadata_alloc_bits;
3237
3238         return btrfs_reduce_alloc_profile(root, flags);
3239 }
3240
3241 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3242 {
3243         u64 flags;
3244
3245         if (data)
3246                 flags = BTRFS_BLOCK_GROUP_DATA;
3247         else if (root == root->fs_info->chunk_root)
3248                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3249         else
3250                 flags = BTRFS_BLOCK_GROUP_METADATA;
3251
3252         return get_alloc_profile(root, flags);
3253 }
3254
3255 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3256 {
3257         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3258                                                        BTRFS_BLOCK_GROUP_DATA);
3259 }
3260
3261 /*
3262  * This will check the space that the inode allocates from to make sure we have
3263  * enough space for bytes.
3264  */
3265 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3266 {
3267         struct btrfs_space_info *data_sinfo;
3268         struct btrfs_root *root = BTRFS_I(inode)->root;
3269         u64 used;
3270         int ret = 0, committed = 0, alloc_chunk = 1;
3271
3272         /* make sure bytes are sectorsize aligned */
3273         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3274
3275         if (root == root->fs_info->tree_root ||
3276             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3277                 alloc_chunk = 0;
3278                 committed = 1;
3279         }
3280
3281         data_sinfo = BTRFS_I(inode)->space_info;
3282         if (!data_sinfo)
3283                 goto alloc;
3284
3285 again:
3286         /* make sure we have enough space to handle the data first */
3287         spin_lock(&data_sinfo->lock);
3288         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3289                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3290                 data_sinfo->bytes_may_use;
3291
3292         if (used + bytes > data_sinfo->total_bytes) {
3293                 struct btrfs_trans_handle *trans;
3294
3295                 /*
3296                  * if we don't have enough free bytes in this space then we need
3297                  * to alloc a new chunk.
3298                  */
3299                 if (!data_sinfo->full && alloc_chunk) {
3300                         u64 alloc_target;
3301
3302                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3303                         spin_unlock(&data_sinfo->lock);
3304 alloc:
3305                         alloc_target = btrfs_get_alloc_profile(root, 1);
3306                         trans = btrfs_join_transaction(root);
3307                         if (IS_ERR(trans))
3308                                 return PTR_ERR(trans);
3309
3310                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3311                                              bytes + 2 * 1024 * 1024,
3312                                              alloc_target,
3313                                              CHUNK_ALLOC_NO_FORCE);
3314                         btrfs_end_transaction(trans, root);
3315                         if (ret < 0) {
3316                                 if (ret != -ENOSPC)
3317                                         return ret;
3318                                 else
3319                                         goto commit_trans;
3320                         }
3321
3322                         if (!data_sinfo) {
3323                                 btrfs_set_inode_space_info(root, inode);
3324                                 data_sinfo = BTRFS_I(inode)->space_info;
3325                         }
3326                         goto again;
3327                 }
3328
3329                 /*
3330                  * If we have less pinned bytes than we want to allocate then
3331                  * don't bother committing the transaction, it won't help us.
3332                  */
3333                 if (data_sinfo->bytes_pinned < bytes)
3334                         committed = 1;
3335                 spin_unlock(&data_sinfo->lock);
3336
3337                 /* commit the current transaction and try again */
3338 commit_trans:
3339                 if (!committed &&
3340                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3341                         committed = 1;
3342                         trans = btrfs_join_transaction(root);
3343                         if (IS_ERR(trans))
3344                                 return PTR_ERR(trans);
3345                         ret = btrfs_commit_transaction(trans, root);
3346                         if (ret)
3347                                 return ret;
3348                         goto again;
3349                 }
3350
3351                 return -ENOSPC;
3352         }
3353         data_sinfo->bytes_may_use += bytes;
3354         trace_btrfs_space_reservation(root->fs_info, "space_info",
3355                                       (u64)(unsigned long)data_sinfo,
3356                                       bytes, 1);
3357         spin_unlock(&data_sinfo->lock);
3358
3359         return 0;
3360 }
3361
3362 /*
3363  * Called if we need to clear a data reservation for this inode.
3364  */
3365 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3366 {
3367         struct btrfs_root *root = BTRFS_I(inode)->root;
3368         struct btrfs_space_info *data_sinfo;
3369
3370         /* make sure bytes are sectorsize aligned */
3371         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3372
3373         data_sinfo = BTRFS_I(inode)->space_info;
3374         spin_lock(&data_sinfo->lock);
3375         data_sinfo->bytes_may_use -= bytes;
3376         trace_btrfs_space_reservation(root->fs_info, "space_info",
3377                                       (u64)(unsigned long)data_sinfo,
3378                                       bytes, 0);
3379         spin_unlock(&data_sinfo->lock);
3380 }
3381
3382 static void force_metadata_allocation(struct btrfs_fs_info *info)
3383 {
3384         struct list_head *head = &info->space_info;
3385         struct btrfs_space_info *found;
3386
3387         rcu_read_lock();
3388         list_for_each_entry_rcu(found, head, list) {
3389                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3390                         found->force_alloc = CHUNK_ALLOC_FORCE;
3391         }
3392         rcu_read_unlock();
3393 }
3394
3395 static int should_alloc_chunk(struct btrfs_root *root,
3396                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3397                               int force)
3398 {
3399         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3400         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3401         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3402         u64 thresh;
3403
3404         if (force == CHUNK_ALLOC_FORCE)
3405                 return 1;
3406
3407         /*
3408          * We need to take into account the global rsv because for all intents
3409          * and purposes it's used space.  Don't worry about locking the
3410          * global_rsv, it doesn't change except when the transaction commits.
3411          */
3412         num_allocated += global_rsv->size;
3413
3414         /*
3415          * in limited mode, we want to have some free space up to
3416          * about 1% of the FS size.
3417          */
3418         if (force == CHUNK_ALLOC_LIMITED) {
3419                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3420                 thresh = max_t(u64, 64 * 1024 * 1024,
3421                                div_factor_fine(thresh, 1));
3422
3423                 if (num_bytes - num_allocated < thresh)
3424                         return 1;
3425         }
3426         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3427
3428         /* 256MB or 2% of the FS */
3429         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 2));
3430         /* system chunks need a much small threshold */
3431         if (sinfo->flags & BTRFS_BLOCK_GROUP_SYSTEM)
3432                 thresh = 32 * 1024 * 1024;
3433
3434         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 8))
3435                 return 0;
3436         return 1;
3437 }
3438
3439 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3440                           struct btrfs_root *extent_root, u64 alloc_bytes,
3441                           u64 flags, int force)
3442 {
3443         struct btrfs_space_info *space_info;
3444         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3445         int wait_for_alloc = 0;
3446         int ret = 0;
3447
3448         BUG_ON(!profile_is_valid(flags, 0));
3449
3450         space_info = __find_space_info(extent_root->fs_info, flags);
3451         if (!space_info) {
3452                 ret = update_space_info(extent_root->fs_info, flags,
3453                                         0, 0, &space_info);
3454                 BUG_ON(ret); /* -ENOMEM */
3455         }
3456         BUG_ON(!space_info); /* Logic error */
3457
3458 again:
3459         spin_lock(&space_info->lock);
3460         if (force < space_info->force_alloc)
3461                 force = space_info->force_alloc;
3462         if (space_info->full) {
3463                 spin_unlock(&space_info->lock);
3464                 return 0;
3465         }
3466
3467         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3468                 spin_unlock(&space_info->lock);
3469                 return 0;
3470         } else if (space_info->chunk_alloc) {
3471                 wait_for_alloc = 1;
3472         } else {
3473                 space_info->chunk_alloc = 1;
3474         }
3475
3476         spin_unlock(&space_info->lock);
3477
3478         mutex_lock(&fs_info->chunk_mutex);
3479
3480         /*
3481          * The chunk_mutex is held throughout the entirety of a chunk
3482          * allocation, so once we've acquired the chunk_mutex we know that the
3483          * other guy is done and we need to recheck and see if we should
3484          * allocate.
3485          */
3486         if (wait_for_alloc) {
3487                 mutex_unlock(&fs_info->chunk_mutex);
3488                 wait_for_alloc = 0;
3489                 goto again;
3490         }
3491
3492         /*
3493          * If we have mixed data/metadata chunks we want to make sure we keep
3494          * allocating mixed chunks instead of individual chunks.
3495          */
3496         if (btrfs_mixed_space_info(space_info))
3497                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3498
3499         /*
3500          * if we're doing a data chunk, go ahead and make sure that
3501          * we keep a reasonable number of metadata chunks allocated in the
3502          * FS as well.
3503          */
3504         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3505                 fs_info->data_chunk_allocations++;
3506                 if (!(fs_info->data_chunk_allocations %
3507                       fs_info->metadata_ratio))
3508                         force_metadata_allocation(fs_info);
3509         }
3510
3511         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3512         if (ret < 0 && ret != -ENOSPC)
3513                 goto out;
3514
3515         spin_lock(&space_info->lock);
3516         if (ret)
3517                 space_info->full = 1;
3518         else
3519                 ret = 1;
3520
3521         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3522         space_info->chunk_alloc = 0;
3523         spin_unlock(&space_info->lock);
3524 out:
3525         mutex_unlock(&extent_root->fs_info->chunk_mutex);
3526         return ret;
3527 }
3528
3529 /*
3530  * shrink metadata reservation for delalloc
3531  */
3532 static int shrink_delalloc(struct btrfs_root *root, u64 to_reclaim,
3533                            bool wait_ordered)
3534 {
3535         struct btrfs_block_rsv *block_rsv;
3536         struct btrfs_space_info *space_info;
3537         struct btrfs_trans_handle *trans;
3538         u64 reserved;
3539         u64 max_reclaim;
3540         u64 reclaimed = 0;
3541         long time_left;
3542         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3543         int loops = 0;
3544         unsigned long progress;
3545
3546         trans = (struct btrfs_trans_handle *)current->journal_info;
3547         block_rsv = &root->fs_info->delalloc_block_rsv;
3548         space_info = block_rsv->space_info;
3549
3550         smp_mb();
3551         reserved = space_info->bytes_may_use;
3552         progress = space_info->reservation_progress;
3553
3554         if (reserved == 0)
3555                 return 0;
3556
3557         smp_mb();
3558         if (root->fs_info->delalloc_bytes == 0) {
3559                 if (trans)
3560                         return 0;
3561                 btrfs_wait_ordered_extents(root, 0, 0);
3562                 return 0;
3563         }
3564
3565         max_reclaim = min(reserved, to_reclaim);
3566         nr_pages = max_t(unsigned long, nr_pages,
3567                          max_reclaim >> PAGE_CACHE_SHIFT);
3568         while (loops < 1024) {
3569                 /* have the flusher threads jump in and do some IO */
3570                 smp_mb();
3571                 nr_pages = min_t(unsigned long, nr_pages,
3572                        root->fs_info->delalloc_bytes >> PAGE_CACHE_SHIFT);
3573                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3574                                                 WB_REASON_FS_FREE_SPACE);
3575
3576                 spin_lock(&space_info->lock);
3577                 if (reserved > space_info->bytes_may_use)
3578                         reclaimed += reserved - space_info->bytes_may_use;
3579                 reserved = space_info->bytes_may_use;
3580                 spin_unlock(&space_info->lock);
3581
3582                 loops++;
3583
3584                 if (reserved == 0 || reclaimed >= max_reclaim)
3585                         break;
3586
3587                 if (trans && trans->transaction->blocked)
3588                         return -EAGAIN;
3589
3590                 if (wait_ordered && !trans) {
3591                         btrfs_wait_ordered_extents(root, 0, 0);
3592                 } else {
3593                         time_left = schedule_timeout_interruptible(1);
3594
3595                         /* We were interrupted, exit */
3596                         if (time_left)
3597                                 break;
3598                 }
3599
3600                 /* we've kicked the IO a few times, if anything has been freed,
3601                  * exit.  There is no sense in looping here for a long time
3602                  * when we really need to commit the transaction, or there are
3603                  * just too many writers without enough free space
3604                  */
3605
3606                 if (loops > 3) {
3607                         smp_mb();
3608                         if (progress != space_info->reservation_progress)
3609                                 break;
3610                 }
3611
3612         }
3613
3614         return reclaimed >= to_reclaim;
3615 }
3616
3617 /**
3618  * maybe_commit_transaction - possibly commit the transaction if its ok to
3619  * @root - the root we're allocating for
3620  * @bytes - the number of bytes we want to reserve
3621  * @force - force the commit
3622  *
3623  * This will check to make sure that committing the transaction will actually
3624  * get us somewhere and then commit the transaction if it does.  Otherwise it
3625  * will return -ENOSPC.
3626  */
3627 static int may_commit_transaction(struct btrfs_root *root,
3628                                   struct btrfs_space_info *space_info,
3629                                   u64 bytes, int force)
3630 {
3631         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3632         struct btrfs_trans_handle *trans;
3633
3634         trans = (struct btrfs_trans_handle *)current->journal_info;
3635         if (trans)
3636                 return -EAGAIN;
3637
3638         if (force)
3639                 goto commit;
3640
3641         /* See if there is enough pinned space to make this reservation */
3642         spin_lock(&space_info->lock);
3643         if (space_info->bytes_pinned >= bytes) {
3644                 spin_unlock(&space_info->lock);
3645                 goto commit;
3646         }
3647         spin_unlock(&space_info->lock);
3648
3649         /*
3650          * See if there is some space in the delayed insertion reservation for
3651          * this reservation.
3652          */
3653         if (space_info != delayed_rsv->space_info)
3654                 return -ENOSPC;
3655
3656         spin_lock(&space_info->lock);
3657         spin_lock(&delayed_rsv->lock);
3658         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3659                 spin_unlock(&delayed_rsv->lock);
3660                 spin_unlock(&space_info->lock);
3661                 return -ENOSPC;
3662         }
3663         spin_unlock(&delayed_rsv->lock);
3664         spin_unlock(&space_info->lock);
3665
3666 commit:
3667         trans = btrfs_join_transaction(root);
3668         if (IS_ERR(trans))
3669                 return -ENOSPC;
3670
3671         return btrfs_commit_transaction(trans, root);
3672 }
3673
3674 /**
3675  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3676  * @root - the root we're allocating for
3677  * @block_rsv - the block_rsv we're allocating for
3678  * @orig_bytes - the number of bytes we want
3679  * @flush - wether or not we can flush to make our reservation
3680  *
3681  * This will reserve orgi_bytes number of bytes from the space info associated
3682  * with the block_rsv.  If there is not enough space it will make an attempt to
3683  * flush out space to make room.  It will do this by flushing delalloc if
3684  * possible or committing the transaction.  If flush is 0 then no attempts to
3685  * regain reservations will be made and this will fail if there is not enough
3686  * space already.
3687  */
3688 static int reserve_metadata_bytes(struct btrfs_root *root,
3689                                   struct btrfs_block_rsv *block_rsv,
3690                                   u64 orig_bytes, int flush)
3691 {
3692         struct btrfs_space_info *space_info = block_rsv->space_info;
3693         u64 used;
3694         u64 num_bytes = orig_bytes;
3695         int retries = 0;
3696         int ret = 0;
3697         bool committed = false;
3698         bool flushing = false;
3699         bool wait_ordered = false;
3700
3701 again:
3702         ret = 0;
3703         spin_lock(&space_info->lock);
3704         /*
3705          * We only want to wait if somebody other than us is flushing and we are
3706          * actually alloed to flush.
3707          */
3708         while (flush && !flushing && space_info->flush) {
3709                 spin_unlock(&space_info->lock);
3710                 /*
3711                  * If we have a trans handle we can't wait because the flusher
3712                  * may have to commit the transaction, which would mean we would
3713                  * deadlock since we are waiting for the flusher to finish, but
3714                  * hold the current transaction open.
3715                  */
3716                 if (current->journal_info)
3717                         return -EAGAIN;
3718                 ret = wait_event_interruptible(space_info->wait,
3719                                                !space_info->flush);
3720                 /* Must have been interrupted, return */
3721                 if (ret) {
3722                         printk(KERN_DEBUG "btrfs: %s returning -EINTR\n", __func__);
3723                         return -EINTR;
3724                 }
3725
3726                 spin_lock(&space_info->lock);
3727         }
3728
3729         ret = -ENOSPC;
3730         used = space_info->bytes_used + space_info->bytes_reserved +
3731                 space_info->bytes_pinned + space_info->bytes_readonly +
3732                 space_info->bytes_may_use;
3733
3734         /*
3735          * The idea here is that we've not already over-reserved the block group
3736          * then we can go ahead and save our reservation first and then start
3737          * flushing if we need to.  Otherwise if we've already overcommitted
3738          * lets start flushing stuff first and then come back and try to make
3739          * our reservation.
3740          */
3741         if (used <= space_info->total_bytes) {
3742                 if (used + orig_bytes <= space_info->total_bytes) {
3743                         space_info->bytes_may_use += orig_bytes;
3744                         trace_btrfs_space_reservation(root->fs_info,
3745                                               "space_info",
3746                                               (u64)(unsigned long)space_info,
3747                                               orig_bytes, 1);
3748                         ret = 0;
3749                 } else {
3750                         /*
3751                          * Ok set num_bytes to orig_bytes since we aren't
3752                          * overocmmitted, this way we only try and reclaim what
3753                          * we need.
3754                          */
3755                         num_bytes = orig_bytes;
3756                 }
3757         } else {
3758                 /*
3759                  * Ok we're over committed, set num_bytes to the overcommitted
3760                  * amount plus the amount of bytes that we need for this
3761                  * reservation.
3762                  */
3763                 wait_ordered = true;
3764                 num_bytes = used - space_info->total_bytes +
3765                         (orig_bytes * (retries + 1));
3766         }
3767
3768         if (ret) {
3769                 u64 profile = btrfs_get_alloc_profile(root, 0);
3770                 u64 avail;
3771
3772                 /*
3773                  * If we have a lot of space that's pinned, don't bother doing
3774                  * the overcommit dance yet and just commit the transaction.
3775                  */
3776                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3777                 do_div(avail, 10);
3778                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3779                         space_info->flush = 1;
3780                         flushing = true;
3781                         spin_unlock(&space_info->lock);
3782                         ret = may_commit_transaction(root, space_info,
3783                                                      orig_bytes, 1);
3784                         if (ret)
3785                                 goto out;
3786                         committed = true;
3787                         goto again;
3788                 }
3789
3790                 spin_lock(&root->fs_info->free_chunk_lock);
3791                 avail = root->fs_info->free_chunk_space;
3792
3793                 /*
3794                  * If we have dup, raid1 or raid10 then only half of the free
3795                  * space is actually useable.
3796                  */
3797                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3798                                BTRFS_BLOCK_GROUP_RAID1 |
3799                                BTRFS_BLOCK_GROUP_RAID10))
3800                         avail >>= 1;
3801
3802                 /*
3803                  * If we aren't flushing don't let us overcommit too much, say
3804                  * 1/8th of the space.  If we can flush, let it overcommit up to
3805                  * 1/2 of the space.
3806                  */
3807                 if (flush)
3808                         avail >>= 3;
3809                 else
3810                         avail >>= 1;
3811                  spin_unlock(&root->fs_info->free_chunk_lock);
3812
3813                 if (used + num_bytes < space_info->total_bytes + avail) {
3814                         space_info->bytes_may_use += orig_bytes;
3815                         trace_btrfs_space_reservation(root->fs_info,
3816                                               "space_info",
3817                                               (u64)(unsigned long)space_info,
3818                                               orig_bytes, 1);
3819                         ret = 0;
3820                 } else {
3821                         wait_ordered = true;
3822                 }
3823         }
3824
3825         /*
3826          * Couldn't make our reservation, save our place so while we're trying
3827          * to reclaim space we can actually use it instead of somebody else
3828          * stealing it from us.
3829          */
3830         if (ret && flush) {
3831                 flushing = true;
3832                 space_info->flush = 1;
3833         }
3834
3835         spin_unlock(&space_info->lock);
3836
3837         if (!ret || !flush)
3838                 goto out;
3839
3840         /*
3841          * We do synchronous shrinking since we don't actually unreserve
3842          * metadata until after the IO is completed.
3843          */
3844         ret = shrink_delalloc(root, num_bytes, wait_ordered);
3845         if (ret < 0)
3846                 goto out;
3847
3848         ret = 0;
3849
3850         /*
3851          * So if we were overcommitted it's possible that somebody else flushed
3852          * out enough space and we simply didn't have enough space to reclaim,
3853          * so go back around and try again.
3854          */
3855         if (retries < 2) {
3856                 wait_ordered = true;
3857                 retries++;
3858                 goto again;
3859         }
3860
3861         ret = -ENOSPC;
3862         if (committed)
3863                 goto out;
3864
3865         ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3866         if (!ret) {
3867                 committed = true;
3868                 goto again;
3869         }
3870
3871 out:
3872         if (flushing) {
3873                 spin_lock(&space_info->lock);
3874                 space_info->flush = 0;
3875                 wake_up_all(&space_info->wait);
3876                 spin_unlock(&space_info->lock);
3877         }
3878         return ret;
3879 }
3880
3881 static struct btrfs_block_rsv *get_block_rsv(
3882                                         const struct btrfs_trans_handle *trans,
3883                                         const struct btrfs_root *root)
3884 {
3885         struct btrfs_block_rsv *block_rsv = NULL;
3886
3887         if (root->ref_cows || root == root->fs_info->csum_root)
3888                 block_rsv = trans->block_rsv;
3889
3890         if (!block_rsv)
3891                 block_rsv = root->block_rsv;
3892
3893         if (!block_rsv)
3894                 block_rsv = &root->fs_info->empty_block_rsv;
3895
3896         return block_rsv;
3897 }
3898
3899 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3900                                u64 num_bytes)
3901 {
3902         int ret = -ENOSPC;
3903         spin_lock(&block_rsv->lock);
3904         if (block_rsv->reserved >= num_bytes) {
3905                 block_rsv->reserved -= num_bytes;
3906                 if (block_rsv->reserved < block_rsv->size)
3907                         block_rsv->full = 0;
3908                 ret = 0;
3909         }
3910         spin_unlock(&block_rsv->lock);
3911         return ret;
3912 }
3913
3914 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3915                                 u64 num_bytes, int update_size)
3916 {
3917         spin_lock(&block_rsv->lock);
3918         block_rsv->reserved += num_bytes;
3919         if (update_size)
3920                 block_rsv->size += num_bytes;
3921         else if (block_rsv->reserved >= block_rsv->size)
3922                 block_rsv->full = 1;
3923         spin_unlock(&block_rsv->lock);
3924 }
3925
3926 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
3927                                     struct btrfs_block_rsv *block_rsv,
3928                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3929 {
3930         struct btrfs_space_info *space_info = block_rsv->space_info;
3931
3932         spin_lock(&block_rsv->lock);
3933         if (num_bytes == (u64)-1)
3934                 num_bytes = block_rsv->size;
3935         block_rsv->size -= num_bytes;
3936         if (block_rsv->reserved >= block_rsv->size) {
3937                 num_bytes = block_rsv->reserved - block_rsv->size;
3938                 block_rsv->reserved = block_rsv->size;
3939                 block_rsv->full = 1;
3940         } else {
3941                 num_bytes = 0;
3942         }
3943         spin_unlock(&block_rsv->lock);
3944
3945         if (num_bytes > 0) {
3946                 if (dest) {
3947                         spin_lock(&dest->lock);
3948                         if (!dest->full) {
3949                                 u64 bytes_to_add;
3950
3951                                 bytes_to_add = dest->size - dest->reserved;
3952                                 bytes_to_add = min(num_bytes, bytes_to_add);
3953                                 dest->reserved += bytes_to_add;
3954                                 if (dest->reserved >= dest->size)
3955                                         dest->full = 1;
3956                                 num_bytes -= bytes_to_add;
3957                         }
3958                         spin_unlock(&dest->lock);
3959                 }
3960                 if (num_bytes) {
3961                         spin_lock(&space_info->lock);
3962                         space_info->bytes_may_use -= num_bytes;
3963                         trace_btrfs_space_reservation(fs_info, "space_info",
3964                                               (u64)(unsigned long)space_info,
3965                                               num_bytes, 0);
3966                         space_info->reservation_progress++;
3967                         spin_unlock(&space_info->lock);
3968                 }
3969         }
3970 }
3971
3972 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
3973                                    struct btrfs_block_rsv *dst, u64 num_bytes)
3974 {
3975         int ret;
3976
3977         ret = block_rsv_use_bytes(src, num_bytes);
3978         if (ret)
3979                 return ret;
3980
3981         block_rsv_add_bytes(dst, num_bytes, 1);
3982         return 0;
3983 }
3984
3985 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
3986 {
3987         memset(rsv, 0, sizeof(*rsv));
3988         spin_lock_init(&rsv->lock);
3989 }
3990
3991 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
3992 {
3993         struct btrfs_block_rsv *block_rsv;
3994         struct btrfs_fs_info *fs_info = root->fs_info;
3995
3996         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
3997         if (!block_rsv)
3998                 return NULL;
3999
4000         btrfs_init_block_rsv(block_rsv);
4001         block_rsv->space_info = __find_space_info(fs_info,
4002                                                   BTRFS_BLOCK_GROUP_METADATA);
4003         return block_rsv;
4004 }
4005
4006 void btrfs_free_block_rsv(struct btrfs_root *root,
4007                           struct btrfs_block_rsv *rsv)
4008 {
4009         btrfs_block_rsv_release(root, rsv, (u64)-1);
4010         kfree(rsv);
4011 }
4012
4013 static inline int __block_rsv_add(struct btrfs_root *root,
4014                                   struct btrfs_block_rsv *block_rsv,
4015                                   u64 num_bytes, int flush)
4016 {
4017         int ret;
4018
4019         if (num_bytes == 0)
4020                 return 0;
4021
4022         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4023         if (!ret) {
4024                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4025                 return 0;
4026         }
4027
4028         return ret;
4029 }
4030
4031 int btrfs_block_rsv_add(struct btrfs_root *root,
4032                         struct btrfs_block_rsv *block_rsv,
4033                         u64 num_bytes)
4034 {
4035         return __block_rsv_add(root, block_rsv, num_bytes, 1);
4036 }
4037
4038 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
4039                                 struct btrfs_block_rsv *block_rsv,
4040                                 u64 num_bytes)
4041 {
4042         return __block_rsv_add(root, block_rsv, num_bytes, 0);
4043 }
4044
4045 int btrfs_block_rsv_check(struct btrfs_root *root,
4046                           struct btrfs_block_rsv *block_rsv, int min_factor)
4047 {
4048         u64 num_bytes = 0;
4049         int ret = -ENOSPC;
4050
4051         if (!block_rsv)
4052                 return 0;
4053
4054         spin_lock(&block_rsv->lock);
4055         num_bytes = div_factor(block_rsv->size, min_factor);
4056         if (block_rsv->reserved >= num_bytes)
4057                 ret = 0;
4058         spin_unlock(&block_rsv->lock);
4059
4060         return ret;
4061 }
4062
4063 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
4064                                            struct btrfs_block_rsv *block_rsv,
4065                                            u64 min_reserved, int flush)
4066 {
4067         u64 num_bytes = 0;
4068         int ret = -ENOSPC;
4069
4070         if (!block_rsv)
4071                 return 0;
4072
4073         spin_lock(&block_rsv->lock);
4074         num_bytes = min_reserved;
4075         if (block_rsv->reserved >= num_bytes)
4076                 ret = 0;
4077         else
4078                 num_bytes -= block_rsv->reserved;
4079         spin_unlock(&block_rsv->lock);
4080
4081         if (!ret)
4082                 return 0;
4083
4084         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4085         if (!ret) {
4086                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4087                 return 0;
4088         }
4089
4090         return ret;
4091 }
4092
4093 int btrfs_block_rsv_refill(struct btrfs_root *root,
4094                            struct btrfs_block_rsv *block_rsv,
4095                            u64 min_reserved)
4096 {
4097         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
4098 }
4099
4100 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
4101                                    struct btrfs_block_rsv *block_rsv,
4102                                    u64 min_reserved)
4103 {
4104         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
4105 }
4106
4107 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4108                             struct btrfs_block_rsv *dst_rsv,
4109                             u64 num_bytes)
4110 {
4111         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4112 }
4113
4114 void btrfs_block_rsv_release(struct btrfs_root *root,
4115                              struct btrfs_block_rsv *block_rsv,
4116                              u64 num_bytes)
4117 {
4118         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4119         if (global_rsv->full || global_rsv == block_rsv ||
4120             block_rsv->space_info != global_rsv->space_info)
4121                 global_rsv = NULL;
4122         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4123                                 num_bytes);
4124 }
4125
4126 /*
4127  * helper to calculate size of global block reservation.
4128  * the desired value is sum of space used by extent tree,
4129  * checksum tree and root tree
4130  */
4131 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4132 {
4133         struct btrfs_space_info *sinfo;
4134         u64 num_bytes;
4135         u64 meta_used;
4136         u64 data_used;
4137         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4138
4139         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4140         spin_lock(&sinfo->lock);
4141         data_used = sinfo->bytes_used;
4142         spin_unlock(&sinfo->lock);
4143
4144         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4145         spin_lock(&sinfo->lock);
4146         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4147                 data_used = 0;
4148         meta_used = sinfo->bytes_used;
4149         spin_unlock(&sinfo->lock);
4150
4151         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4152                     csum_size * 2;
4153         num_bytes += div64_u64(data_used + meta_used, 50);
4154
4155         if (num_bytes * 3 > meta_used)
4156                 num_bytes = div64_u64(meta_used, 3) * 2;
4157
4158         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4159 }
4160
4161 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4162 {
4163         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4164         struct btrfs_space_info *sinfo = block_rsv->space_info;
4165         u64 num_bytes;
4166
4167         num_bytes = calc_global_metadata_size(fs_info);
4168
4169         spin_lock(&block_rsv->lock);
4170         spin_lock(&sinfo->lock);
4171
4172         block_rsv->size = num_bytes;
4173
4174         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4175                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4176                     sinfo->bytes_may_use;
4177
4178         if (sinfo->total_bytes > num_bytes) {
4179                 num_bytes = sinfo->total_bytes - num_bytes;
4180                 block_rsv->reserved += num_bytes;
4181                 sinfo->bytes_may_use += num_bytes;
4182                 trace_btrfs_space_reservation(fs_info, "space_info",
4183                                       (u64)(unsigned long)sinfo, num_bytes, 1);
4184         }
4185
4186         if (block_rsv->reserved >= block_rsv->size) {
4187                 num_bytes = block_rsv->reserved - block_rsv->size;
4188                 sinfo->bytes_may_use -= num_bytes;
4189                 trace_btrfs_space_reservation(fs_info, "space_info",
4190                                       (u64)(unsigned long)sinfo, num_bytes, 0);
4191                 sinfo->reservation_progress++;
4192                 block_rsv->reserved = block_rsv->size;
4193                 block_rsv->full = 1;
4194         }
4195
4196         spin_unlock(&sinfo->lock);
4197         spin_unlock(&block_rsv->lock);
4198 }
4199
4200 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4201 {
4202         struct btrfs_space_info *space_info;
4203
4204         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4205         fs_info->chunk_block_rsv.space_info = space_info;
4206
4207         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4208         fs_info->global_block_rsv.space_info = space_info;
4209         fs_info->delalloc_block_rsv.space_info = space_info;
4210         fs_info->trans_block_rsv.space_info = space_info;
4211         fs_info->empty_block_rsv.space_info = space_info;
4212         fs_info->delayed_block_rsv.space_info = space_info;
4213
4214         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4215         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4216         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4217         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4218         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4219
4220         update_global_block_rsv(fs_info);
4221 }
4222
4223 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4224 {
4225         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4226                                 (u64)-1);
4227         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4228         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4229         WARN_ON(fs_info->trans_block_rsv.size > 0);
4230         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4231         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4232         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4233         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4234         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4235 }
4236
4237 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4238                                   struct btrfs_root *root)
4239 {
4240         if (!trans->bytes_reserved)
4241                 return;
4242
4243         trace_btrfs_space_reservation(root->fs_info, "transaction",
4244                                       (u64)(unsigned long)trans,
4245                                       trans->bytes_reserved, 0);
4246         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4247         trans->bytes_reserved = 0;
4248 }
4249
4250 /* Can only return 0 or -ENOSPC */
4251 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4252                                   struct inode *inode)
4253 {
4254         struct btrfs_root *root = BTRFS_I(inode)->root;
4255         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4256         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4257
4258         /*
4259          * We need to hold space in order to delete our orphan item once we've
4260          * added it, so this takes the reservation so we can release it later
4261          * when we are truly done with the orphan item.
4262          */
4263         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4264         trace_btrfs_space_reservation(root->fs_info, "orphan",
4265                                       btrfs_ino(inode), num_bytes, 1);
4266         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4267 }
4268
4269 void btrfs_orphan_release_metadata(struct inode *inode)
4270 {
4271         struct btrfs_root *root = BTRFS_I(inode)->root;
4272         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4273         trace_btrfs_space_reservation(root->fs_info, "orphan",
4274                                       btrfs_ino(inode), num_bytes, 0);
4275         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4276 }
4277
4278 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4279                                 struct btrfs_pending_snapshot *pending)
4280 {
4281         struct btrfs_root *root = pending->root;
4282         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4283         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4284         /*
4285          * two for root back/forward refs, two for directory entries
4286          * and one for root of the snapshot.
4287          */
4288         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4289         dst_rsv->space_info = src_rsv->space_info;
4290         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4291 }
4292
4293 /**
4294  * drop_outstanding_extent - drop an outstanding extent
4295  * @inode: the inode we're dropping the extent for
4296  *
4297  * This is called when we are freeing up an outstanding extent, either called
4298  * after an error or after an extent is written.  This will return the number of
4299  * reserved extents that need to be freed.  This must be called with
4300  * BTRFS_I(inode)->lock held.
4301  */
4302 static unsigned drop_outstanding_extent(struct inode *inode)
4303 {
4304         unsigned drop_inode_space = 0;
4305         unsigned dropped_extents = 0;
4306
4307         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4308         BTRFS_I(inode)->outstanding_extents--;
4309
4310         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4311             BTRFS_I(inode)->delalloc_meta_reserved) {
4312                 drop_inode_space = 1;
4313                 BTRFS_I(inode)->delalloc_meta_reserved = 0;
4314         }
4315
4316         /*
4317          * If we have more or the same amount of outsanding extents than we have
4318          * reserved then we need to leave the reserved extents count alone.
4319          */
4320         if (BTRFS_I(inode)->outstanding_extents >=
4321             BTRFS_I(inode)->reserved_extents)
4322                 return drop_inode_space;
4323
4324         dropped_extents = BTRFS_I(inode)->reserved_extents -
4325                 BTRFS_I(inode)->outstanding_extents;
4326         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4327         return dropped_extents + drop_inode_space;
4328 }
4329
4330 /**
4331  * calc_csum_metadata_size - return the amount of metada space that must be
4332  *      reserved/free'd for the given bytes.
4333  * @inode: the inode we're manipulating
4334  * @num_bytes: the number of bytes in question
4335  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4336  *
4337  * This adjusts the number of csum_bytes in the inode and then returns the
4338  * correct amount of metadata that must either be reserved or freed.  We
4339  * calculate how many checksums we can fit into one leaf and then divide the
4340  * number of bytes that will need to be checksumed by this value to figure out
4341  * how many checksums will be required.  If we are adding bytes then the number
4342  * may go up and we will return the number of additional bytes that must be
4343  * reserved.  If it is going down we will return the number of bytes that must
4344  * be freed.
4345  *
4346  * This must be called with BTRFS_I(inode)->lock held.
4347  */
4348 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4349                                    int reserve)
4350 {
4351         struct btrfs_root *root = BTRFS_I(inode)->root;
4352         u64 csum_size;
4353         int num_csums_per_leaf;
4354         int num_csums;
4355         int old_csums;
4356
4357         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4358             BTRFS_I(inode)->csum_bytes == 0)
4359                 return 0;
4360
4361         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4362         if (reserve)
4363                 BTRFS_I(inode)->csum_bytes += num_bytes;
4364         else
4365                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4366         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4367         num_csums_per_leaf = (int)div64_u64(csum_size,
4368                                             sizeof(struct btrfs_csum_item) +
4369                                             sizeof(struct btrfs_disk_key));
4370         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4371         num_csums = num_csums + num_csums_per_leaf - 1;
4372         num_csums = num_csums / num_csums_per_leaf;
4373
4374         old_csums = old_csums + num_csums_per_leaf - 1;
4375         old_csums = old_csums / num_csums_per_leaf;
4376
4377         /* No change, no need to reserve more */
4378         if (old_csums == num_csums)
4379                 return 0;
4380
4381         if (reserve)
4382                 return btrfs_calc_trans_metadata_size(root,
4383                                                       num_csums - old_csums);
4384
4385         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4386 }
4387
4388 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4389 {
4390         struct btrfs_root *root = BTRFS_I(inode)->root;
4391         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4392         u64 to_reserve = 0;
4393         u64 csum_bytes;
4394         unsigned nr_extents = 0;
4395         int extra_reserve = 0;
4396         int flush = 1;
4397         int ret;
4398
4399         /* Need to be holding the i_mutex here if we aren't free space cache */
4400         if (btrfs_is_free_space_inode(root, inode))
4401                 flush = 0;
4402
4403         if (flush && btrfs_transaction_in_commit(root->fs_info))
4404                 schedule_timeout(1);
4405
4406         mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4407         num_bytes = ALIGN(num_bytes, root->sectorsize);
4408
4409         spin_lock(&BTRFS_I(inode)->lock);
4410         BTRFS_I(inode)->outstanding_extents++;
4411
4412         if (BTRFS_I(inode)->outstanding_extents >
4413             BTRFS_I(inode)->reserved_extents)
4414                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4415                         BTRFS_I(inode)->reserved_extents;
4416
4417         /*
4418          * Add an item to reserve for updating the inode when we complete the
4419          * delalloc io.
4420          */
4421         if (!BTRFS_I(inode)->delalloc_meta_reserved) {
4422                 nr_extents++;
4423                 extra_reserve = 1;
4424         }
4425
4426         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4427         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4428         csum_bytes = BTRFS_I(inode)->csum_bytes;
4429         spin_unlock(&BTRFS_I(inode)->lock);
4430
4431         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4432         if (ret) {
4433                 u64 to_free = 0;
4434                 unsigned dropped;
4435
4436                 spin_lock(&BTRFS_I(inode)->lock);
4437                 dropped = drop_outstanding_extent(inode);
4438                 /*
4439                  * If the inodes csum_bytes is the same as the original
4440                  * csum_bytes then we know we haven't raced with any free()ers
4441                  * so we can just reduce our inodes csum bytes and carry on.
4442                  * Otherwise we have to do the normal free thing to account for
4443                  * the case that the free side didn't free up its reserve
4444                  * because of this outstanding reservation.
4445                  */
4446                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4447                         calc_csum_metadata_size(inode, num_bytes, 0);
4448                 else
4449                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4450                 spin_unlock(&BTRFS_I(inode)->lock);
4451                 if (dropped)
4452                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4453
4454                 if (to_free) {
4455                         btrfs_block_rsv_release(root, block_rsv, to_free);
4456                         trace_btrfs_space_reservation(root->fs_info,
4457                                                       "delalloc",
4458                                                       btrfs_ino(inode),
4459                                                       to_free, 0);
4460                 }
4461                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4462                 return ret;
4463         }
4464
4465         spin_lock(&BTRFS_I(inode)->lock);
4466         if (extra_reserve) {
4467                 BTRFS_I(inode)->delalloc_meta_reserved = 1;
4468                 nr_extents--;
4469         }
4470         BTRFS_I(inode)->reserved_extents += nr_extents;
4471         spin_unlock(&BTRFS_I(inode)->lock);
4472         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4473
4474         if (to_reserve)
4475                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4476                                               btrfs_ino(inode), to_reserve, 1);
4477         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4478
4479         return 0;
4480 }
4481
4482 /**
4483  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4484  * @inode: the inode to release the reservation for
4485  * @num_bytes: the number of bytes we're releasing
4486  *
4487  * This will release the metadata reservation for an inode.  This can be called
4488  * once we complete IO for a given set of bytes to release their metadata
4489  * reservations.
4490  */
4491 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4492 {
4493         struct btrfs_root *root = BTRFS_I(inode)->root;
4494         u64 to_free = 0;
4495         unsigned dropped;
4496
4497         num_bytes = ALIGN(num_bytes, root->sectorsize);
4498         spin_lock(&BTRFS_I(inode)->lock);
4499         dropped = drop_outstanding_extent(inode);
4500
4501         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4502         spin_unlock(&BTRFS_I(inode)->lock);
4503         if (dropped > 0)
4504                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4505
4506         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4507                                       btrfs_ino(inode), to_free, 0);
4508         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4509                                 to_free);
4510 }
4511
4512 /**
4513  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4514  * @inode: inode we're writing to
4515  * @num_bytes: the number of bytes we want to allocate
4516  *
4517  * This will do the following things
4518  *
4519  * o reserve space in the data space info for num_bytes
4520  * o reserve space in the metadata space info based on number of outstanding
4521  *   extents and how much csums will be needed
4522  * o add to the inodes ->delalloc_bytes
4523  * o add it to the fs_info's delalloc inodes list.
4524  *
4525  * This will return 0 for success and -ENOSPC if there is no space left.
4526  */
4527 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4528 {
4529         int ret;
4530
4531         ret = btrfs_check_data_free_space(inode, num_bytes);
4532         if (ret)
4533                 return ret;
4534
4535         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4536         if (ret) {
4537                 btrfs_free_reserved_data_space(inode, num_bytes);
4538                 return ret;
4539         }
4540
4541         return 0;
4542 }
4543
4544 /**
4545  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4546  * @inode: inode we're releasing space for
4547  * @num_bytes: the number of bytes we want to free up
4548  *
4549  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4550  * called in the case that we don't need the metadata AND data reservations
4551  * anymore.  So if there is an error or we insert an inline extent.
4552  *
4553  * This function will release the metadata space that was not used and will
4554  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4555  * list if there are no delalloc bytes left.
4556  */
4557 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4558 {
4559         btrfs_delalloc_release_metadata(inode, num_bytes);
4560         btrfs_free_reserved_data_space(inode, num_bytes);
4561 }
4562
4563 static int update_block_group(struct btrfs_trans_handle *trans,
4564                               struct btrfs_root *root,
4565                               u64 bytenr, u64 num_bytes, int alloc)
4566 {
4567         struct btrfs_block_group_cache *cache = NULL;
4568         struct btrfs_fs_info *info = root->fs_info;
4569         u64 total = num_bytes;
4570         u64 old_val;
4571         u64 byte_in_group;
4572         int factor;
4573
4574         /* block accounting for super block */
4575         spin_lock(&info->delalloc_lock);
4576         old_val = btrfs_super_bytes_used(info->super_copy);
4577         if (alloc)
4578                 old_val += num_bytes;
4579         else
4580                 old_val -= num_bytes;
4581         btrfs_set_super_bytes_used(info->super_copy, old_val);
4582         spin_unlock(&info->delalloc_lock);
4583
4584         while (total) {
4585                 cache = btrfs_lookup_block_group(info, bytenr);
4586                 if (!cache)
4587                         return -ENOENT;
4588                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4589                                     BTRFS_BLOCK_GROUP_RAID1 |
4590                                     BTRFS_BLOCK_GROUP_RAID10))
4591                         factor = 2;
4592                 else
4593                         factor = 1;
4594                 /*
4595                  * If this block group has free space cache written out, we
4596                  * need to make sure to load it if we are removing space.  This
4597                  * is because we need the unpinning stage to actually add the
4598                  * space back to the block group, otherwise we will leak space.
4599                  */
4600                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4601                         cache_block_group(cache, trans, NULL, 1);
4602
4603                 byte_in_group = bytenr - cache->key.objectid;
4604                 WARN_ON(byte_in_group > cache->key.offset);
4605
4606                 spin_lock(&cache->space_info->lock);
4607                 spin_lock(&cache->lock);
4608
4609                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4610                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4611                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4612
4613                 cache->dirty = 1;
4614                 old_val = btrfs_block_group_used(&cache->item);
4615                 num_bytes = min(total, cache->key.offset - byte_in_group);
4616                 if (alloc) {
4617                         old_val += num_bytes;
4618                         btrfs_set_block_group_used(&cache->item, old_val);
4619                         cache->reserved -= num_bytes;
4620                         cache->space_info->bytes_reserved -= num_bytes;
4621                         cache->space_info->bytes_used += num_bytes;
4622                         cache->space_info->disk_used += num_bytes * factor;
4623                         spin_unlock(&cache->lock);
4624                         spin_unlock(&cache->space_info->lock);
4625                 } else {
4626                         old_val -= num_bytes;
4627                         btrfs_set_block_group_used(&cache->item, old_val);
4628                         cache->pinned += num_bytes;
4629                         cache->space_info->bytes_pinned += num_bytes;
4630                         cache->space_info->bytes_used -= num_bytes;
4631                         cache->space_info->disk_used -= num_bytes * factor;
4632                         spin_unlock(&cache->lock);
4633                         spin_unlock(&cache->space_info->lock);
4634
4635                         set_extent_dirty(info->pinned_extents,
4636                                          bytenr, bytenr + num_bytes - 1,
4637                                          GFP_NOFS | __GFP_NOFAIL);
4638                 }
4639                 btrfs_put_block_group(cache);
4640                 total -= num_bytes;
4641                 bytenr += num_bytes;
4642         }
4643         return 0;
4644 }
4645
4646 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4647 {
4648         struct btrfs_block_group_cache *cache;
4649         u64 bytenr;
4650
4651         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4652         if (!cache)
4653                 return 0;
4654
4655         bytenr = cache->key.objectid;
4656         btrfs_put_block_group(cache);
4657
4658         return bytenr;
4659 }
4660
4661 static int pin_down_extent(struct btrfs_root *root,
4662                            struct btrfs_block_group_cache *cache,
4663                            u64 bytenr, u64 num_bytes, int reserved)
4664 {
4665         spin_lock(&cache->space_info->lock);
4666         spin_lock(&cache->lock);
4667         cache->pinned += num_bytes;
4668         cache->space_info->bytes_pinned += num_bytes;
4669         if (reserved) {
4670                 cache->reserved -= num_bytes;
4671                 cache->space_info->bytes_reserved -= num_bytes;
4672         }
4673         spin_unlock(&cache->lock);
4674         spin_unlock(&cache->space_info->lock);
4675
4676         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4677                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4678         return 0;
4679 }
4680
4681 /*
4682  * this function must be called within transaction
4683  */
4684 int btrfs_pin_extent(struct btrfs_root *root,
4685                      u64 bytenr, u64 num_bytes, int reserved)
4686 {
4687         struct btrfs_block_group_cache *cache;
4688
4689         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4690         BUG_ON(!cache); /* Logic error */
4691
4692         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4693
4694         btrfs_put_block_group(cache);
4695         return 0;
4696 }
4697
4698 /*
4699  * this function must be called within transaction
4700  */
4701 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4702                                     struct btrfs_root *root,
4703                                     u64 bytenr, u64 num_bytes)
4704 {
4705         struct btrfs_block_group_cache *cache;
4706
4707         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4708         BUG_ON(!cache); /* Logic error */
4709
4710         /*
4711          * pull in the free space cache (if any) so that our pin
4712          * removes the free space from the cache.  We have load_only set
4713          * to one because the slow code to read in the free extents does check
4714          * the pinned extents.
4715          */
4716         cache_block_group(cache, trans, root, 1);
4717
4718         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4719
4720         /* remove us from the free space cache (if we're there at all) */
4721         btrfs_remove_free_space(cache, bytenr, num_bytes);
4722         btrfs_put_block_group(cache);
4723         return 0;
4724 }
4725
4726 /**
4727  * btrfs_update_reserved_bytes - update the block_group and space info counters
4728  * @cache:      The cache we are manipulating
4729  * @num_bytes:  The number of bytes in question
4730  * @reserve:    One of the reservation enums
4731  *
4732  * This is called by the allocator when it reserves space, or by somebody who is
4733  * freeing space that was never actually used on disk.  For example if you
4734  * reserve some space for a new leaf in transaction A and before transaction A
4735  * commits you free that leaf, you call this with reserve set to 0 in order to
4736  * clear the reservation.
4737  *
4738  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4739  * ENOSPC accounting.  For data we handle the reservation through clearing the
4740  * delalloc bits in the io_tree.  We have to do this since we could end up
4741  * allocating less disk space for the amount of data we have reserved in the
4742  * case of compression.
4743  *
4744  * If this is a reservation and the block group has become read only we cannot
4745  * make the reservation and return -EAGAIN, otherwise this function always
4746  * succeeds.
4747  */
4748 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4749                                        u64 num_bytes, int reserve)
4750 {
4751         struct btrfs_space_info *space_info = cache->space_info;
4752         int ret = 0;
4753
4754         spin_lock(&space_info->lock);
4755         spin_lock(&cache->lock);
4756         if (reserve != RESERVE_FREE) {
4757                 if (cache->ro) {
4758                         ret = -EAGAIN;
4759                 } else {
4760                         cache->reserved += num_bytes;
4761                         space_info->bytes_reserved += num_bytes;
4762                         if (reserve == RESERVE_ALLOC) {
4763                                 trace_btrfs_space_reservation(cache->fs_info,
4764                                               "space_info",
4765                                               (u64)(unsigned long)space_info,
4766                                               num_bytes, 0);
4767                                 space_info->bytes_may_use -= num_bytes;
4768                         }
4769                 }
4770         } else {
4771                 if (cache->ro)
4772                         space_info->bytes_readonly += num_bytes;
4773                 cache->reserved -= num_bytes;
4774                 space_info->bytes_reserved -= num_bytes;
4775                 space_info->reservation_progress++;
4776         }
4777         spin_unlock(&cache->lock);
4778         spin_unlock(&space_info->lock);
4779         return ret;
4780 }
4781
4782 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4783                                 struct btrfs_root *root)
4784 {
4785         struct btrfs_fs_info *fs_info = root->fs_info;
4786         struct btrfs_caching_control *next;
4787         struct btrfs_caching_control *caching_ctl;
4788         struct btrfs_block_group_cache *cache;
4789
4790         down_write(&fs_info->extent_commit_sem);
4791
4792         list_for_each_entry_safe(caching_ctl, next,
4793                                  &fs_info->caching_block_groups, list) {
4794                 cache = caching_ctl->block_group;
4795                 if (block_group_cache_done(cache)) {
4796                         cache->last_byte_to_unpin = (u64)-1;
4797                         list_del_init(&caching_ctl->list);
4798                         put_caching_control(caching_ctl);
4799                 } else {
4800                         cache->last_byte_to_unpin = caching_ctl->progress;
4801                 }
4802         }
4803
4804         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4805                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4806         else
4807                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4808
4809         up_write(&fs_info->extent_commit_sem);
4810
4811         update_global_block_rsv(fs_info);
4812 }
4813
4814 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4815 {
4816         struct btrfs_fs_info *fs_info = root->fs_info;
4817         struct btrfs_block_group_cache *cache = NULL;
4818         u64 len;
4819
4820         while (start <= end) {
4821                 if (!cache ||
4822                     start >= cache->key.objectid + cache->key.offset) {
4823                         if (cache)
4824                                 btrfs_put_block_group(cache);
4825                         cache = btrfs_lookup_block_group(fs_info, start);
4826                         BUG_ON(!cache); /* Logic error */
4827                 }
4828
4829                 len = cache->key.objectid + cache->key.offset - start;
4830                 len = min(len, end + 1 - start);
4831
4832                 if (start < cache->last_byte_to_unpin) {
4833                         len = min(len, cache->last_byte_to_unpin - start);
4834                         btrfs_add_free_space(cache, start, len);
4835                 }
4836
4837                 start += len;
4838
4839                 spin_lock(&cache->space_info->lock);
4840                 spin_lock(&cache->lock);
4841                 cache->pinned -= len;
4842                 cache->space_info->bytes_pinned -= len;
4843                 if (cache->ro)
4844                         cache->space_info->bytes_readonly += len;
4845                 spin_unlock(&cache->lock);
4846                 spin_unlock(&cache->space_info->lock);
4847         }
4848
4849         if (cache)
4850                 btrfs_put_block_group(cache);
4851         return 0;
4852 }
4853
4854 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4855                                struct btrfs_root *root)
4856 {
4857         struct btrfs_fs_info *fs_info = root->fs_info;
4858         struct extent_io_tree *unpin;
4859         u64 start;
4860         u64 end;
4861         int ret;
4862
4863         if (trans->aborted)
4864                 return 0;
4865
4866         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4867                 unpin = &fs_info->freed_extents[1];
4868         else
4869                 unpin = &fs_info->freed_extents[0];
4870
4871         while (1) {
4872                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4873                                             EXTENT_DIRTY);
4874                 if (ret)
4875                         break;
4876
4877                 if (btrfs_test_opt(root, DISCARD))
4878                         ret = btrfs_discard_extent(root, start,
4879                                                    end + 1 - start, NULL);
4880
4881                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4882                 unpin_extent_range(root, start, end);
4883                 cond_resched();
4884         }
4885
4886         return 0;
4887 }
4888
4889 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4890                                 struct btrfs_root *root,
4891                                 u64 bytenr, u64 num_bytes, u64 parent,
4892                                 u64 root_objectid, u64 owner_objectid,
4893                                 u64 owner_offset, int refs_to_drop,
4894                                 struct btrfs_delayed_extent_op *extent_op)
4895 {
4896         struct btrfs_key key;
4897         struct btrfs_path *path;
4898         struct btrfs_fs_info *info = root->fs_info;
4899         struct btrfs_root *extent_root = info->extent_root;
4900         struct extent_buffer *leaf;
4901         struct btrfs_extent_item *ei;
4902         struct btrfs_extent_inline_ref *iref;
4903         int ret;
4904         int is_data;
4905         int extent_slot = 0;
4906         int found_extent = 0;
4907         int num_to_del = 1;
4908         u32 item_size;
4909         u64 refs;
4910
4911         path = btrfs_alloc_path();
4912         if (!path)
4913                 return -ENOMEM;
4914
4915         path->reada = 1;
4916         path->leave_spinning = 1;
4917
4918         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4919         BUG_ON(!is_data && refs_to_drop != 1);
4920
4921         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4922                                     bytenr, num_bytes, parent,
4923                                     root_objectid, owner_objectid,
4924                                     owner_offset);
4925         if (ret == 0) {
4926                 extent_slot = path->slots[0];
4927                 while (extent_slot >= 0) {
4928                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4929                                               extent_slot);
4930                         if (key.objectid != bytenr)
4931                                 break;
4932                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4933                             key.offset == num_bytes) {
4934                                 found_extent = 1;
4935                                 break;
4936                         }
4937                         if (path->slots[0] - extent_slot > 5)
4938                                 break;
4939                         extent_slot--;
4940                 }
4941 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4942                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4943                 if (found_extent && item_size < sizeof(*ei))
4944                         found_extent = 0;
4945 #endif
4946                 if (!found_extent) {
4947                         BUG_ON(iref);
4948                         ret = remove_extent_backref(trans, extent_root, path,
4949                                                     NULL, refs_to_drop,
4950                                                     is_data);
4951                         if (ret)
4952                                 goto abort;
4953                         btrfs_release_path(path);
4954                         path->leave_spinning = 1;
4955
4956                         key.objectid = bytenr;
4957                         key.type = BTRFS_EXTENT_ITEM_KEY;
4958                         key.offset = num_bytes;
4959
4960                         ret = btrfs_search_slot(trans, extent_root,
4961                                                 &key, path, -1, 1);
4962                         if (ret) {
4963                                 printk(KERN_ERR "umm, got %d back from search"
4964                                        ", was looking for %llu\n", ret,
4965                                        (unsigned long long)bytenr);
4966                                 if (ret > 0)
4967                                         btrfs_print_leaf(extent_root,
4968                                                          path->nodes[0]);
4969                         }
4970                         if (ret < 0)
4971                                 goto abort;
4972                         extent_slot = path->slots[0];
4973                 }
4974         } else if (ret == -ENOENT) {
4975                 btrfs_print_leaf(extent_root, path->nodes[0]);
4976                 WARN_ON(1);
4977                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
4978                        "parent %llu root %llu  owner %llu offset %llu\n",
4979                        (unsigned long long)bytenr,
4980                        (unsigned long long)parent,
4981                        (unsigned long long)root_objectid,
4982                        (unsigned long long)owner_objectid,
4983                        (unsigned long long)owner_offset);
4984         } else {
4985                 goto abort;
4986         }
4987
4988         leaf = path->nodes[0];
4989         item_size = btrfs_item_size_nr(leaf, extent_slot);
4990 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4991         if (item_size < sizeof(*ei)) {
4992                 BUG_ON(found_extent || extent_slot != path->slots[0]);
4993                 ret = convert_extent_item_v0(trans, extent_root, path,
4994                                              owner_objectid, 0);
4995                 if (ret < 0)
4996                         goto abort;
4997
4998                 btrfs_release_path(path);
4999                 path->leave_spinning = 1;
5000
5001                 key.objectid = bytenr;
5002                 key.type = BTRFS_EXTENT_ITEM_KEY;
5003                 key.offset = num_bytes;
5004
5005                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5006                                         -1, 1);
5007                 if (ret) {
5008                         printk(KERN_ERR "umm, got %d back from search"
5009                                ", was looking for %llu\n", ret,
5010                                (unsigned long long)bytenr);
5011                         btrfs_print_leaf(extent_root, path->nodes[0]);
5012                 }
5013                 if (ret < 0)
5014                         goto abort;
5015                 extent_slot = path->slots[0];
5016                 leaf = path->nodes[0];
5017                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5018         }
5019 #endif
5020         BUG_ON(item_size < sizeof(*ei));
5021         ei = btrfs_item_ptr(leaf, extent_slot,
5022                             struct btrfs_extent_item);
5023         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5024                 struct btrfs_tree_block_info *bi;
5025                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5026                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5027                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5028         }
5029
5030         refs = btrfs_extent_refs(leaf, ei);
5031         BUG_ON(refs < refs_to_drop);
5032         refs -= refs_to_drop;
5033
5034         if (refs > 0) {
5035                 if (extent_op)
5036                         __run_delayed_extent_op(extent_op, leaf, ei);
5037                 /*
5038                  * In the case of inline back ref, reference count will
5039                  * be updated by remove_extent_backref
5040                  */
5041                 if (iref) {
5042                         BUG_ON(!found_extent);
5043                 } else {
5044                         btrfs_set_extent_refs(leaf, ei, refs);
5045                         btrfs_mark_buffer_dirty(leaf);
5046                 }
5047                 if (found_extent) {
5048                         ret = remove_extent_backref(trans, extent_root, path,
5049                                                     iref, refs_to_drop,
5050                                                     is_data);
5051                         if (ret)
5052                                 goto abort;
5053                 }
5054         } else {
5055                 if (found_extent) {
5056                         BUG_ON(is_data && refs_to_drop !=
5057                                extent_data_ref_count(root, path, iref));
5058                         if (iref) {
5059                                 BUG_ON(path->slots[0] != extent_slot);
5060                         } else {
5061                                 BUG_ON(path->slots[0] != extent_slot + 1);
5062                                 path->slots[0] = extent_slot;
5063                                 num_to_del = 2;
5064                         }
5065                 }
5066
5067                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5068                                       num_to_del);
5069                 if (ret)
5070                         goto abort;
5071                 btrfs_release_path(path);
5072
5073                 if (is_data) {
5074                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5075                         if (ret)
5076                                 goto abort;
5077                 } else {
5078                         invalidate_mapping_pages(info->btree_inode->i_mapping,
5079                              bytenr >> PAGE_CACHE_SHIFT,
5080                              (bytenr + num_bytes - 1) >> PAGE_CACHE_SHIFT);
5081                 }
5082
5083                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
5084                 if (ret)
5085                         goto abort;
5086         }
5087 out:
5088         btrfs_free_path(path);
5089         return ret;
5090
5091 abort:
5092         btrfs_abort_transaction(trans, extent_root, ret);
5093         goto out;
5094 }
5095
5096 /*
5097  * when we free an block, it is possible (and likely) that we free the last
5098  * delayed ref for that extent as well.  This searches the delayed ref tree for
5099  * a given extent, and if there are no other delayed refs to be processed, it
5100  * removes it from the tree.
5101  */
5102 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5103                                       struct btrfs_root *root, u64 bytenr)
5104 {
5105         struct btrfs_delayed_ref_head *head;
5106         struct btrfs_delayed_ref_root *delayed_refs;
5107         struct btrfs_delayed_ref_node *ref;
5108         struct rb_node *node;
5109         int ret = 0;
5110
5111         delayed_refs = &trans->transaction->delayed_refs;
5112         spin_lock(&delayed_refs->lock);
5113         head = btrfs_find_delayed_ref_head(trans, bytenr);
5114         if (!head)
5115                 goto out;
5116
5117         node = rb_prev(&head->node.rb_node);
5118         if (!node)
5119                 goto out;
5120
5121         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5122
5123         /* there are still entries for this ref, we can't drop it */
5124         if (ref->bytenr == bytenr)
5125                 goto out;
5126
5127         if (head->extent_op) {
5128                 if (!head->must_insert_reserved)
5129                         goto out;
5130                 kfree(head->extent_op);
5131                 head->extent_op = NULL;
5132         }
5133
5134         /*
5135          * waiting for the lock here would deadlock.  If someone else has it
5136          * locked they are already in the process of dropping it anyway
5137          */
5138         if (!mutex_trylock(&head->mutex))
5139                 goto out;
5140
5141         /*
5142          * at this point we have a head with no other entries.  Go
5143          * ahead and process it.
5144          */
5145         head->node.in_tree = 0;
5146         rb_erase(&head->node.rb_node, &delayed_refs->root);
5147
5148         delayed_refs->num_entries--;
5149         if (waitqueue_active(&delayed_refs->seq_wait))
5150                 wake_up(&delayed_refs->seq_wait);
5151
5152         /*
5153          * we don't take a ref on the node because we're removing it from the
5154          * tree, so we just steal the ref the tree was holding.
5155          */
5156         delayed_refs->num_heads--;
5157         if (list_empty(&head->cluster))
5158                 delayed_refs->num_heads_ready--;
5159
5160         list_del_init(&head->cluster);
5161         spin_unlock(&delayed_refs->lock);
5162
5163         BUG_ON(head->extent_op);
5164         if (head->must_insert_reserved)
5165                 ret = 1;
5166
5167         mutex_unlock(&head->mutex);
5168         btrfs_put_delayed_ref(&head->node);
5169         return ret;
5170 out:
5171         spin_unlock(&delayed_refs->lock);
5172         return 0;
5173 }
5174
5175 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5176                            struct btrfs_root *root,
5177                            struct extent_buffer *buf,
5178                            u64 parent, int last_ref, int for_cow)
5179 {
5180         struct btrfs_block_group_cache *cache = NULL;
5181         int ret;
5182
5183         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5184                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5185                                         buf->start, buf->len,
5186                                         parent, root->root_key.objectid,
5187                                         btrfs_header_level(buf),
5188                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5189                 BUG_ON(ret); /* -ENOMEM */
5190         }
5191
5192         if (!last_ref)
5193                 return;
5194
5195         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5196
5197         if (btrfs_header_generation(buf) == trans->transid) {
5198                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5199                         ret = check_ref_cleanup(trans, root, buf->start);
5200                         if (!ret)
5201                                 goto out;
5202                 }
5203
5204                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5205                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5206                         goto out;
5207                 }
5208
5209                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5210
5211                 btrfs_add_free_space(cache, buf->start, buf->len);
5212                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5213         }
5214 out:
5215         /*
5216          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5217          * anymore.
5218          */
5219         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5220         btrfs_put_block_group(cache);
5221 }
5222
5223 /* Can return -ENOMEM */
5224 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5225                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5226                       u64 owner, u64 offset, int for_cow)
5227 {
5228         int ret;
5229         struct btrfs_fs_info *fs_info = root->fs_info;
5230
5231         /*
5232          * tree log blocks never actually go into the extent allocation
5233          * tree, just update pinning info and exit early.
5234          */
5235         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5236                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5237                 /* unlocks the pinned mutex */
5238                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5239                 ret = 0;
5240         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5241                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5242                                         num_bytes,
5243                                         parent, root_objectid, (int)owner,
5244                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5245         } else {
5246                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5247                                                 num_bytes,
5248                                                 parent, root_objectid, owner,
5249                                                 offset, BTRFS_DROP_DELAYED_REF,
5250                                                 NULL, for_cow);
5251         }
5252         return ret;
5253 }
5254
5255 static u64 stripe_align(struct btrfs_root *root, u64 val)
5256 {
5257         u64 mask = ((u64)root->stripesize - 1);
5258         u64 ret = (val + mask) & ~mask;
5259         return ret;
5260 }
5261
5262 /*
5263  * when we wait for progress in the block group caching, its because
5264  * our allocation attempt failed at least once.  So, we must sleep
5265  * and let some progress happen before we try again.
5266  *
5267  * This function will sleep at least once waiting for new free space to
5268  * show up, and then it will check the block group free space numbers
5269  * for our min num_bytes.  Another option is to have it go ahead
5270  * and look in the rbtree for a free extent of a given size, but this
5271  * is a good start.
5272  */
5273 static noinline int
5274 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5275                                 u64 num_bytes)
5276 {
5277         struct btrfs_caching_control *caching_ctl;
5278         DEFINE_WAIT(wait);
5279
5280         caching_ctl = get_caching_control(cache);
5281         if (!caching_ctl)
5282                 return 0;
5283
5284         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5285                    (cache->free_space_ctl->free_space >= num_bytes));
5286
5287         put_caching_control(caching_ctl);
5288         return 0;
5289 }
5290
5291 static noinline int
5292 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5293 {
5294         struct btrfs_caching_control *caching_ctl;
5295         DEFINE_WAIT(wait);
5296
5297         caching_ctl = get_caching_control(cache);
5298         if (!caching_ctl)
5299                 return 0;
5300
5301         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5302
5303         put_caching_control(caching_ctl);
5304         return 0;
5305 }
5306
5307 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5308 {
5309         int index;
5310         if (cache->flags & BTRFS_BLOCK_GROUP_RAID10)
5311                 index = 0;
5312         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID1)
5313                 index = 1;
5314         else if (cache->flags & BTRFS_BLOCK_GROUP_DUP)
5315                 index = 2;
5316         else if (cache->flags & BTRFS_BLOCK_GROUP_RAID0)
5317                 index = 3;
5318         else
5319                 index = 4;
5320         return index;
5321 }
5322
5323 enum btrfs_loop_type {
5324         LOOP_FIND_IDEAL = 0,
5325         LOOP_CACHING_NOWAIT = 1,
5326         LOOP_CACHING_WAIT = 2,
5327         LOOP_ALLOC_CHUNK = 3,
5328         LOOP_NO_EMPTY_SIZE = 4,
5329 };
5330
5331 /*
5332  * walks the btree of allocated extents and find a hole of a given size.
5333  * The key ins is changed to record the hole:
5334  * ins->objectid == block start
5335  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5336  * ins->offset == number of blocks
5337  * Any available blocks before search_start are skipped.
5338  */
5339 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5340                                      struct btrfs_root *orig_root,
5341                                      u64 num_bytes, u64 empty_size,
5342                                      u64 search_start, u64 search_end,
5343                                      u64 hint_byte, struct btrfs_key *ins,
5344                                      u64 data)
5345 {
5346         int ret = 0;
5347         struct btrfs_root *root = orig_root->fs_info->extent_root;
5348         struct btrfs_free_cluster *last_ptr = NULL;
5349         struct btrfs_block_group_cache *block_group = NULL;
5350         struct btrfs_block_group_cache *used_block_group;
5351         int empty_cluster = 2 * 1024 * 1024;
5352         int allowed_chunk_alloc = 0;
5353         int done_chunk_alloc = 0;
5354         struct btrfs_space_info *space_info;
5355         int loop = 0;
5356         int index = 0;
5357         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5358                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5359         bool found_uncached_bg = false;
5360         bool failed_cluster_refill = false;
5361         bool failed_alloc = false;
5362         bool use_cluster = true;
5363         bool have_caching_bg = false;
5364         u64 ideal_cache_percent = 0;
5365         u64 ideal_cache_offset = 0;
5366
5367         WARN_ON(num_bytes < root->sectorsize);
5368         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5369         ins->objectid = 0;
5370         ins->offset = 0;
5371
5372         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5373
5374         space_info = __find_space_info(root->fs_info, data);
5375         if (!space_info) {
5376                 printk(KERN_ERR "No space info for %llu\n", data);
5377                 return -ENOSPC;
5378         }
5379
5380         /*
5381          * If the space info is for both data and metadata it means we have a
5382          * small filesystem and we can't use the clustering stuff.
5383          */
5384         if (btrfs_mixed_space_info(space_info))
5385                 use_cluster = false;
5386
5387         if (orig_root->ref_cows || empty_size)
5388                 allowed_chunk_alloc = 1;
5389
5390         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5391                 last_ptr = &root->fs_info->meta_alloc_cluster;
5392                 if (!btrfs_test_opt(root, SSD))
5393                         empty_cluster = 64 * 1024;
5394         }
5395
5396         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5397             btrfs_test_opt(root, SSD)) {
5398                 last_ptr = &root->fs_info->data_alloc_cluster;
5399         }
5400
5401         if (last_ptr) {
5402                 spin_lock(&last_ptr->lock);
5403                 if (last_ptr->block_group)
5404                         hint_byte = last_ptr->window_start;
5405                 spin_unlock(&last_ptr->lock);
5406         }
5407
5408         search_start = max(search_start, first_logical_byte(root, 0));
5409         search_start = max(search_start, hint_byte);
5410
5411         if (!last_ptr)
5412                 empty_cluster = 0;
5413
5414         if (search_start == hint_byte) {
5415 ideal_cache:
5416                 block_group = btrfs_lookup_block_group(root->fs_info,
5417                                                        search_start);
5418                 used_block_group = block_group;
5419                 /*
5420                  * we don't want to use the block group if it doesn't match our
5421                  * allocation bits, or if its not cached.
5422                  *
5423                  * However if we are re-searching with an ideal block group
5424                  * picked out then we don't care that the block group is cached.
5425                  */
5426                 if (block_group && block_group_bits(block_group, data) &&
5427                     (block_group->cached != BTRFS_CACHE_NO ||
5428                      search_start == ideal_cache_offset)) {
5429                         down_read(&space_info->groups_sem);
5430                         if (list_empty(&block_group->list) ||
5431                             block_group->ro) {
5432                                 /*
5433                                  * someone is removing this block group,
5434                                  * we can't jump into the have_block_group
5435                                  * target because our list pointers are not
5436                                  * valid
5437                                  */
5438                                 btrfs_put_block_group(block_group);
5439                                 up_read(&space_info->groups_sem);
5440                         } else {
5441                                 index = get_block_group_index(block_group);
5442                                 goto have_block_group;
5443                         }
5444                 } else if (block_group) {
5445                         btrfs_put_block_group(block_group);
5446                 }
5447         }
5448 search:
5449         have_caching_bg = false;
5450         down_read(&space_info->groups_sem);
5451         list_for_each_entry(block_group, &space_info->block_groups[index],
5452                             list) {
5453                 u64 offset;
5454                 int cached;
5455
5456                 used_block_group = block_group;
5457                 btrfs_get_block_group(block_group);
5458                 search_start = block_group->key.objectid;
5459
5460                 /*
5461                  * this can happen if we end up cycling through all the
5462                  * raid types, but we want to make sure we only allocate
5463                  * for the proper type.
5464                  */
5465                 if (!block_group_bits(block_group, data)) {
5466                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5467                                 BTRFS_BLOCK_GROUP_RAID1 |
5468                                 BTRFS_BLOCK_GROUP_RAID10;
5469
5470                         /*
5471                          * if they asked for extra copies and this block group
5472                          * doesn't provide them, bail.  This does allow us to
5473                          * fill raid0 from raid1.
5474                          */
5475                         if ((data & extra) && !(block_group->flags & extra))
5476                                 goto loop;
5477                 }
5478
5479 have_block_group:
5480                 cached = block_group_cache_done(block_group);
5481                 if (unlikely(!cached)) {
5482                         u64 free_percent;
5483
5484                         found_uncached_bg = true;
5485                         ret = cache_block_group(block_group, trans,
5486                                                 orig_root, 1);
5487                         BUG_ON(ret < 0); /* -ENOMEM */
5488                         if (block_group->cached == BTRFS_CACHE_FINISHED)
5489                                 goto alloc;
5490
5491                         free_percent = btrfs_block_group_used(&block_group->item);
5492                         free_percent *= 100;
5493                         free_percent = div64_u64(free_percent,
5494                                                  block_group->key.offset);
5495                         free_percent = 100 - free_percent;
5496                         if (free_percent > ideal_cache_percent &&
5497                             likely(!block_group->ro)) {
5498                                 ideal_cache_offset = block_group->key.objectid;
5499                                 ideal_cache_percent = free_percent;
5500                         }
5501
5502                         /*
5503                          * The caching workers are limited to 2 threads, so we
5504                          * can queue as much work as we care to.
5505                          */
5506                         if (loop > LOOP_FIND_IDEAL) {
5507                                 ret = cache_block_group(block_group, trans,
5508                                                         orig_root, 0);
5509                                 BUG_ON(ret); /* -ENOMEM */
5510                         }
5511
5512                         /*
5513                          * If loop is set for cached only, try the next block
5514                          * group.
5515                          */
5516                         if (loop == LOOP_FIND_IDEAL)
5517                                 goto loop;
5518                 }
5519
5520 alloc:
5521                 if (unlikely(block_group->ro))
5522                         goto loop;
5523
5524                 /*
5525                  * Ok we want to try and use the cluster allocator, so
5526                  * lets look there
5527                  */
5528                 if (last_ptr) {
5529                         /*
5530                          * the refill lock keeps out other
5531                          * people trying to start a new cluster
5532                          */
5533                         spin_lock(&last_ptr->refill_lock);
5534                         used_block_group = last_ptr->block_group;
5535                         if (used_block_group != block_group &&
5536                             (!used_block_group ||
5537                              used_block_group->ro ||
5538                              !block_group_bits(used_block_group, data))) {
5539                                 used_block_group = block_group;
5540                                 goto refill_cluster;
5541                         }
5542
5543                         if (used_block_group != block_group)
5544                                 btrfs_get_block_group(used_block_group);
5545
5546                         offset = btrfs_alloc_from_cluster(used_block_group,
5547                           last_ptr, num_bytes, used_block_group->key.objectid);
5548                         if (offset) {
5549                                 /* we have a block, we're done */
5550                                 spin_unlock(&last_ptr->refill_lock);
5551                                 trace_btrfs_reserve_extent_cluster(root,
5552                                         block_group, search_start, num_bytes);
5553                                 goto checks;
5554                         }
5555
5556                         WARN_ON(last_ptr->block_group != used_block_group);
5557                         if (used_block_group != block_group) {
5558                                 btrfs_put_block_group(used_block_group);
5559                                 used_block_group = block_group;
5560                         }
5561 refill_cluster:
5562                         BUG_ON(used_block_group != block_group);
5563                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5564                          * set up a new clusters, so lets just skip it
5565                          * and let the allocator find whatever block
5566                          * it can find.  If we reach this point, we
5567                          * will have tried the cluster allocator
5568                          * plenty of times and not have found
5569                          * anything, so we are likely way too
5570                          * fragmented for the clustering stuff to find
5571                          * anything.
5572                          *
5573                          * However, if the cluster is taken from the
5574                          * current block group, release the cluster
5575                          * first, so that we stand a better chance of
5576                          * succeeding in the unclustered
5577                          * allocation.  */
5578                         if (loop >= LOOP_NO_EMPTY_SIZE &&
5579                             last_ptr->block_group != block_group) {
5580                                 spin_unlock(&last_ptr->refill_lock);
5581                                 goto unclustered_alloc;
5582                         }
5583
5584                         /*
5585                          * this cluster didn't work out, free it and
5586                          * start over
5587                          */
5588                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5589
5590                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5591                                 spin_unlock(&last_ptr->refill_lock);
5592                                 goto unclustered_alloc;
5593                         }
5594
5595                         /* allocate a cluster in this block group */
5596                         ret = btrfs_find_space_cluster(trans, root,
5597                                                block_group, last_ptr,
5598                                                search_start, num_bytes,
5599                                                empty_cluster + empty_size);
5600                         if (ret == 0) {
5601                                 /*
5602                                  * now pull our allocation out of this
5603                                  * cluster
5604                                  */
5605                                 offset = btrfs_alloc_from_cluster(block_group,
5606                                                   last_ptr, num_bytes,
5607                                                   search_start);
5608                                 if (offset) {
5609                                         /* we found one, proceed */
5610                                         spin_unlock(&last_ptr->refill_lock);
5611                                         trace_btrfs_reserve_extent_cluster(root,
5612                                                 block_group, search_start,
5613                                                 num_bytes);
5614                                         goto checks;
5615                                 }
5616                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5617                                    && !failed_cluster_refill) {
5618                                 spin_unlock(&last_ptr->refill_lock);
5619
5620                                 failed_cluster_refill = true;
5621                                 wait_block_group_cache_progress(block_group,
5622                                        num_bytes + empty_cluster + empty_size);
5623                                 goto have_block_group;
5624                         }
5625
5626                         /*
5627                          * at this point we either didn't find a cluster
5628                          * or we weren't able to allocate a block from our
5629                          * cluster.  Free the cluster we've been trying
5630                          * to use, and go to the next block group
5631                          */
5632                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5633                         spin_unlock(&last_ptr->refill_lock);
5634                         goto loop;
5635                 }
5636
5637 unclustered_alloc:
5638                 spin_lock(&block_group->free_space_ctl->tree_lock);
5639                 if (cached &&
5640                     block_group->free_space_ctl->free_space <
5641                     num_bytes + empty_cluster + empty_size) {
5642                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5643                         goto loop;
5644                 }
5645                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5646
5647                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5648                                                     num_bytes, empty_size);
5649                 /*
5650                  * If we didn't find a chunk, and we haven't failed on this
5651                  * block group before, and this block group is in the middle of
5652                  * caching and we are ok with waiting, then go ahead and wait
5653                  * for progress to be made, and set failed_alloc to true.
5654                  *
5655                  * If failed_alloc is true then we've already waited on this
5656                  * block group once and should move on to the next block group.
5657                  */
5658                 if (!offset && !failed_alloc && !cached &&
5659                     loop > LOOP_CACHING_NOWAIT) {
5660                         wait_block_group_cache_progress(block_group,
5661                                                 num_bytes + empty_size);
5662                         failed_alloc = true;
5663                         goto have_block_group;
5664                 } else if (!offset) {
5665                         if (!cached)
5666                                 have_caching_bg = true;
5667                         goto loop;
5668                 }
5669 checks:
5670                 search_start = stripe_align(root, offset);
5671                 /* move on to the next group */
5672                 if (search_start + num_bytes >= search_end) {
5673                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5674                         goto loop;
5675                 }
5676
5677                 /* move on to the next group */
5678                 if (search_start + num_bytes >
5679                     used_block_group->key.objectid + used_block_group->key.offset) {
5680                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5681                         goto loop;
5682                 }
5683
5684                 if (offset < search_start)
5685                         btrfs_add_free_space(used_block_group, offset,
5686                                              search_start - offset);
5687                 BUG_ON(offset > search_start);
5688
5689                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
5690                                                   alloc_type);
5691                 if (ret == -EAGAIN) {
5692                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5693                         goto loop;
5694                 }
5695
5696                 /* we are all good, lets return */
5697                 ins->objectid = search_start;
5698                 ins->offset = num_bytes;
5699
5700                 trace_btrfs_reserve_extent(orig_root, block_group,
5701                                            search_start, num_bytes);
5702                 if (offset < search_start)
5703                         btrfs_add_free_space(used_block_group, offset,
5704                                              search_start - offset);
5705                 BUG_ON(offset > search_start);
5706                 if (used_block_group != block_group)
5707                         btrfs_put_block_group(used_block_group);
5708                 btrfs_put_block_group(block_group);
5709                 break;
5710 loop:
5711                 failed_cluster_refill = false;
5712                 failed_alloc = false;
5713                 BUG_ON(index != get_block_group_index(block_group));
5714                 if (used_block_group != block_group)
5715                         btrfs_put_block_group(used_block_group);
5716                 btrfs_put_block_group(block_group);
5717         }
5718         up_read(&space_info->groups_sem);
5719
5720         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
5721                 goto search;
5722
5723         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5724                 goto search;
5725
5726         /* LOOP_FIND_IDEAL, only search caching/cached bg's, and don't wait for
5727          *                      for them to make caching progress.  Also
5728          *                      determine the best possible bg to cache
5729          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5730          *                      caching kthreads as we move along
5731          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5732          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5733          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5734          *                      again
5735          */
5736         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5737                 index = 0;
5738                 if (loop == LOOP_FIND_IDEAL && found_uncached_bg) {
5739                         found_uncached_bg = false;
5740                         loop++;
5741                         if (!ideal_cache_percent)
5742                                 goto search;
5743
5744                         /*
5745                          * 1 of the following 2 things have happened so far
5746                          *
5747                          * 1) We found an ideal block group for caching that
5748                          * is mostly full and will cache quickly, so we might
5749                          * as well wait for it.
5750                          *
5751                          * 2) We searched for cached only and we didn't find
5752                          * anything, and we didn't start any caching kthreads
5753                          * either, so chances are we will loop through and
5754                          * start a couple caching kthreads, and then come back
5755                          * around and just wait for them.  This will be slower
5756                          * because we will have 2 caching kthreads reading at
5757                          * the same time when we could have just started one
5758                          * and waited for it to get far enough to give us an
5759                          * allocation, so go ahead and go to the wait caching
5760                          * loop.
5761                          */
5762                         loop = LOOP_CACHING_WAIT;
5763                         search_start = ideal_cache_offset;
5764                         ideal_cache_percent = 0;
5765                         goto ideal_cache;
5766                 } else if (loop == LOOP_FIND_IDEAL) {
5767                         /*
5768                          * Didn't find a uncached bg, wait on anything we find
5769                          * next.
5770                          */
5771                         loop = LOOP_CACHING_WAIT;
5772                         goto search;
5773                 }
5774
5775                 loop++;
5776
5777                 if (loop == LOOP_ALLOC_CHUNK) {
5778                        if (allowed_chunk_alloc) {
5779                                 ret = do_chunk_alloc(trans, root, num_bytes +
5780                                                      2 * 1024 * 1024, data,
5781                                                      CHUNK_ALLOC_LIMITED);
5782                                 if (ret < 0) {
5783                                         btrfs_abort_transaction(trans,
5784                                                                 root, ret);
5785                                         goto out;
5786                                 }
5787                                 allowed_chunk_alloc = 0;
5788                                 if (ret == 1)
5789                                         done_chunk_alloc = 1;
5790                         } else if (!done_chunk_alloc &&
5791                                    space_info->force_alloc ==
5792                                    CHUNK_ALLOC_NO_FORCE) {
5793                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5794                         }
5795
5796                        /*
5797                         * We didn't allocate a chunk, go ahead and drop the
5798                         * empty size and loop again.
5799                         */
5800                        if (!done_chunk_alloc)
5801                                loop = LOOP_NO_EMPTY_SIZE;
5802                 }
5803
5804                 if (loop == LOOP_NO_EMPTY_SIZE) {
5805                         empty_size = 0;
5806                         empty_cluster = 0;
5807                 }
5808
5809                 goto search;
5810         } else if (!ins->objectid) {
5811                 ret = -ENOSPC;
5812         } else if (ins->objectid) {
5813                 ret = 0;
5814         }
5815 out:
5816
5817         return ret;
5818 }
5819
5820 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5821                             int dump_block_groups)
5822 {
5823         struct btrfs_block_group_cache *cache;
5824         int index = 0;
5825
5826         spin_lock(&info->lock);
5827         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5828                (unsigned long long)info->flags,
5829                (unsigned long long)(info->total_bytes - info->bytes_used -
5830                                     info->bytes_pinned - info->bytes_reserved -
5831                                     info->bytes_readonly),
5832                (info->full) ? "" : "not ");
5833         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5834                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5835                (unsigned long long)info->total_bytes,
5836                (unsigned long long)info->bytes_used,
5837                (unsigned long long)info->bytes_pinned,
5838                (unsigned long long)info->bytes_reserved,
5839                (unsigned long long)info->bytes_may_use,
5840                (unsigned long long)info->bytes_readonly);
5841         spin_unlock(&info->lock);
5842
5843         if (!dump_block_groups)
5844                 return;
5845
5846         down_read(&info->groups_sem);
5847 again:
5848         list_for_each_entry(cache, &info->block_groups[index], list) {
5849                 spin_lock(&cache->lock);
5850                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5851                        "%llu pinned %llu reserved\n",
5852                        (unsigned long long)cache->key.objectid,
5853                        (unsigned long long)cache->key.offset,
5854                        (unsigned long long)btrfs_block_group_used(&cache->item),
5855                        (unsigned long long)cache->pinned,
5856                        (unsigned long long)cache->reserved);
5857                 btrfs_dump_free_space(cache, bytes);
5858                 spin_unlock(&cache->lock);
5859         }
5860         if (++index < BTRFS_NR_RAID_TYPES)
5861                 goto again;
5862         up_read(&info->groups_sem);
5863 }
5864
5865 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5866                          struct btrfs_root *root,
5867                          u64 num_bytes, u64 min_alloc_size,
5868                          u64 empty_size, u64 hint_byte,
5869                          u64 search_end, struct btrfs_key *ins,
5870                          u64 data)
5871 {
5872         bool final_tried = false;
5873         int ret;
5874         u64 search_start = 0;
5875
5876         data = btrfs_get_alloc_profile(root, data);
5877 again:
5878         /*
5879          * the only place that sets empty_size is btrfs_realloc_node, which
5880          * is not called recursively on allocations
5881          */
5882         if (empty_size || root->ref_cows) {
5883                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5884                                      num_bytes + 2 * 1024 * 1024, data,
5885                                      CHUNK_ALLOC_NO_FORCE);
5886                 if (ret < 0 && ret != -ENOSPC) {
5887                         btrfs_abort_transaction(trans, root, ret);
5888                         return ret;
5889                 }
5890         }
5891
5892         WARN_ON(num_bytes < root->sectorsize);
5893         ret = find_free_extent(trans, root, num_bytes, empty_size,
5894                                search_start, search_end, hint_byte,
5895                                ins, data);
5896
5897         if (ret == -ENOSPC) {
5898                 if (!final_tried) {
5899                         num_bytes = num_bytes >> 1;
5900                         num_bytes = num_bytes & ~(root->sectorsize - 1);
5901                         num_bytes = max(num_bytes, min_alloc_size);
5902                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5903                                        num_bytes, data, CHUNK_ALLOC_FORCE);
5904                         if (ret < 0 && ret != -ENOSPC) {
5905                                 btrfs_abort_transaction(trans, root, ret);
5906                                 return ret;
5907                         }
5908                         if (num_bytes == min_alloc_size)
5909                                 final_tried = true;
5910                         goto again;
5911                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
5912                         struct btrfs_space_info *sinfo;
5913
5914                         sinfo = __find_space_info(root->fs_info, data);
5915                         printk(KERN_ERR "btrfs allocation failed flags %llu, "
5916                                "wanted %llu\n", (unsigned long long)data,
5917                                (unsigned long long)num_bytes);
5918                         if (sinfo)
5919                                 dump_space_info(sinfo, num_bytes, 1);
5920                 }
5921         }
5922
5923         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
5924
5925         return ret;
5926 }
5927
5928 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
5929                                         u64 start, u64 len, int pin)
5930 {
5931         struct btrfs_block_group_cache *cache;
5932         int ret = 0;
5933
5934         cache = btrfs_lookup_block_group(root->fs_info, start);
5935         if (!cache) {
5936                 printk(KERN_ERR "Unable to find block group for %llu\n",
5937                        (unsigned long long)start);
5938                 return -ENOSPC;
5939         }
5940
5941         if (btrfs_test_opt(root, DISCARD))
5942                 ret = btrfs_discard_extent(root, start, len, NULL);
5943
5944         if (pin)
5945                 pin_down_extent(root, cache, start, len, 1);
5946         else {
5947                 btrfs_add_free_space(cache, start, len);
5948                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
5949         }
5950         btrfs_put_block_group(cache);
5951
5952         trace_btrfs_reserved_extent_free(root, start, len);
5953
5954         return ret;
5955 }
5956
5957 int btrfs_free_reserved_extent(struct btrfs_root *root,
5958                                         u64 start, u64 len)
5959 {
5960         return __btrfs_free_reserved_extent(root, start, len, 0);
5961 }
5962
5963 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
5964                                        u64 start, u64 len)
5965 {
5966         return __btrfs_free_reserved_extent(root, start, len, 1);
5967 }
5968
5969 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5970                                       struct btrfs_root *root,
5971                                       u64 parent, u64 root_objectid,
5972                                       u64 flags, u64 owner, u64 offset,
5973                                       struct btrfs_key *ins, int ref_mod)
5974 {
5975         int ret;
5976         struct btrfs_fs_info *fs_info = root->fs_info;
5977         struct btrfs_extent_item *extent_item;
5978         struct btrfs_extent_inline_ref *iref;
5979         struct btrfs_path *path;
5980         struct extent_buffer *leaf;
5981         int type;
5982         u32 size;
5983
5984         if (parent > 0)
5985                 type = BTRFS_SHARED_DATA_REF_KEY;
5986         else
5987                 type = BTRFS_EXTENT_DATA_REF_KEY;
5988
5989         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5990
5991         path = btrfs_alloc_path();
5992         if (!path)
5993                 return -ENOMEM;
5994
5995         path->leave_spinning = 1;
5996         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5997                                       ins, size);
5998         if (ret) {
5999                 btrfs_free_path(path);
6000                 return ret;
6001         }
6002
6003         leaf = path->nodes[0];
6004         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6005                                      struct btrfs_extent_item);
6006         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6007         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6008         btrfs_set_extent_flags(leaf, extent_item,
6009                                flags | BTRFS_EXTENT_FLAG_DATA);
6010
6011         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6012         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6013         if (parent > 0) {
6014                 struct btrfs_shared_data_ref *ref;
6015                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6016                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6017                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6018         } else {
6019                 struct btrfs_extent_data_ref *ref;
6020                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6021                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6022                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6023                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6024                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6025         }
6026
6027         btrfs_mark_buffer_dirty(path->nodes[0]);
6028         btrfs_free_path(path);
6029
6030         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
6031         if (ret) { /* -ENOENT, logic error */
6032                 printk(KERN_ERR "btrfs update block group failed for %llu "
6033                        "%llu\n", (unsigned long long)ins->objectid,
6034                        (unsigned long long)ins->offset);
6035                 BUG();
6036         }
6037         return ret;
6038 }
6039
6040 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6041                                      struct btrfs_root *root,
6042                                      u64 parent, u64 root_objectid,
6043                                      u64 flags, struct btrfs_disk_key *key,
6044                                      int level, struct btrfs_key *ins)
6045 {
6046         int ret;
6047         struct btrfs_fs_info *fs_info = root->fs_info;
6048         struct btrfs_extent_item *extent_item;
6049         struct btrfs_tree_block_info *block_info;
6050         struct btrfs_extent_inline_ref *iref;
6051         struct btrfs_path *path;
6052         struct extent_buffer *leaf;
6053         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
6054
6055         path = btrfs_alloc_path();
6056         if (!path)
6057                 return -ENOMEM;
6058
6059         path->leave_spinning = 1;
6060         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6061                                       ins, size);
6062         if (ret) {
6063                 btrfs_free_path(path);
6064                 return ret;
6065         }
6066
6067         leaf = path->nodes[0];
6068         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6069                                      struct btrfs_extent_item);
6070         btrfs_set_extent_refs(leaf, extent_item, 1);
6071         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6072         btrfs_set_extent_flags(leaf, extent_item,
6073                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6074         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6075
6076         btrfs_set_tree_block_key(leaf, block_info, key);
6077         btrfs_set_tree_block_level(leaf, block_info, level);
6078
6079         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6080         if (parent > 0) {
6081                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6082                 btrfs_set_extent_inline_ref_type(leaf, iref,
6083                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6084                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6085         } else {
6086                 btrfs_set_extent_inline_ref_type(leaf, iref,
6087                                                  BTRFS_TREE_BLOCK_REF_KEY);
6088                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6089         }
6090
6091         btrfs_mark_buffer_dirty(leaf);
6092         btrfs_free_path(path);
6093
6094         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
6095         if (ret) { /* -ENOENT, logic error */
6096                 printk(KERN_ERR "btrfs update block group failed for %llu "
6097                        "%llu\n", (unsigned long long)ins->objectid,
6098                        (unsigned long long)ins->offset);
6099                 BUG();
6100         }
6101         return ret;
6102 }
6103
6104 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6105                                      struct btrfs_root *root,
6106                                      u64 root_objectid, u64 owner,
6107                                      u64 offset, struct btrfs_key *ins)
6108 {
6109         int ret;
6110
6111         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6112
6113         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6114                                          ins->offset, 0,
6115                                          root_objectid, owner, offset,
6116                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6117         return ret;
6118 }
6119
6120 /*
6121  * this is used by the tree logging recovery code.  It records that
6122  * an extent has been allocated and makes sure to clear the free
6123  * space cache bits as well
6124  */
6125 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6126                                    struct btrfs_root *root,
6127                                    u64 root_objectid, u64 owner, u64 offset,
6128                                    struct btrfs_key *ins)
6129 {
6130         int ret;
6131         struct btrfs_block_group_cache *block_group;
6132         struct btrfs_caching_control *caching_ctl;
6133         u64 start = ins->objectid;
6134         u64 num_bytes = ins->offset;
6135
6136         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6137         cache_block_group(block_group, trans, NULL, 0);
6138         caching_ctl = get_caching_control(block_group);
6139
6140         if (!caching_ctl) {
6141                 BUG_ON(!block_group_cache_done(block_group));
6142                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6143                 BUG_ON(ret); /* -ENOMEM */
6144         } else {
6145                 mutex_lock(&caching_ctl->mutex);
6146
6147                 if (start >= caching_ctl->progress) {
6148                         ret = add_excluded_extent(root, start, num_bytes);
6149                         BUG_ON(ret); /* -ENOMEM */
6150                 } else if (start + num_bytes <= caching_ctl->progress) {
6151                         ret = btrfs_remove_free_space(block_group,
6152                                                       start, num_bytes);
6153                         BUG_ON(ret); /* -ENOMEM */
6154                 } else {
6155                         num_bytes = caching_ctl->progress - start;
6156                         ret = btrfs_remove_free_space(block_group,
6157                                                       start, num_bytes);
6158                         BUG_ON(ret); /* -ENOMEM */
6159
6160                         start = caching_ctl->progress;
6161                         num_bytes = ins->objectid + ins->offset -
6162                                     caching_ctl->progress;
6163                         ret = add_excluded_extent(root, start, num_bytes);
6164                         BUG_ON(ret); /* -ENOMEM */
6165                 }
6166
6167                 mutex_unlock(&caching_ctl->mutex);
6168                 put_caching_control(caching_ctl);
6169         }
6170
6171         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6172                                           RESERVE_ALLOC_NO_ACCOUNT);
6173         BUG_ON(ret); /* logic error */
6174         btrfs_put_block_group(block_group);
6175         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6176                                          0, owner, offset, ins, 1);
6177         return ret;
6178 }
6179
6180 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
6181                                             struct btrfs_root *root,
6182                                             u64 bytenr, u32 blocksize,
6183                                             int level)
6184 {
6185         struct extent_buffer *buf;
6186
6187         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6188         if (!buf)
6189                 return ERR_PTR(-ENOMEM);
6190         btrfs_set_header_generation(buf, trans->transid);
6191         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6192         btrfs_tree_lock(buf);
6193         clean_tree_block(trans, root, buf);
6194
6195         btrfs_set_lock_blocking(buf);
6196         btrfs_set_buffer_uptodate(buf);
6197
6198         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6199                 /*
6200                  * we allow two log transactions at a time, use different
6201                  * EXENT bit to differentiate dirty pages.
6202                  */
6203                 if (root->log_transid % 2 == 0)
6204                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6205                                         buf->start + buf->len - 1, GFP_NOFS);
6206                 else
6207                         set_extent_new(&root->dirty_log_pages, buf->start,
6208                                         buf->start + buf->len - 1, GFP_NOFS);
6209         } else {
6210                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6211                          buf->start + buf->len - 1, GFP_NOFS);
6212         }
6213         trans->blocks_used++;
6214         /* this returns a buffer locked for blocking */
6215         return buf;
6216 }
6217
6218 static struct btrfs_block_rsv *
6219 use_block_rsv(struct btrfs_trans_handle *trans,
6220               struct btrfs_root *root, u32 blocksize)
6221 {
6222         struct btrfs_block_rsv *block_rsv;
6223         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6224         int ret;
6225
6226         block_rsv = get_block_rsv(trans, root);
6227
6228         if (block_rsv->size == 0) {
6229                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
6230                 /*
6231                  * If we couldn't reserve metadata bytes try and use some from
6232                  * the global reserve.
6233                  */
6234                 if (ret && block_rsv != global_rsv) {
6235                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6236                         if (!ret)
6237                                 return global_rsv;
6238                         return ERR_PTR(ret);
6239                 } else if (ret) {
6240                         return ERR_PTR(ret);
6241                 }
6242                 return block_rsv;
6243         }
6244
6245         ret = block_rsv_use_bytes(block_rsv, blocksize);
6246         if (!ret)
6247                 return block_rsv;
6248         if (ret) {
6249                 static DEFINE_RATELIMIT_STATE(_rs,
6250                                 DEFAULT_RATELIMIT_INTERVAL,
6251                                 /*DEFAULT_RATELIMIT_BURST*/ 2);
6252                 if (__ratelimit(&_rs)) {
6253                         printk(KERN_DEBUG "btrfs: block rsv returned %d\n", ret);
6254                         WARN_ON(1);
6255                 }
6256                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
6257                 if (!ret) {
6258                         return block_rsv;
6259                 } else if (ret && block_rsv != global_rsv) {
6260                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6261                         if (!ret)
6262                                 return global_rsv;
6263                 }
6264         }
6265
6266         return ERR_PTR(-ENOSPC);
6267 }
6268
6269 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6270                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6271 {
6272         block_rsv_add_bytes(block_rsv, blocksize, 0);
6273         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6274 }
6275
6276 /*
6277  * finds a free extent and does all the dirty work required for allocation
6278  * returns the key for the extent through ins, and a tree buffer for
6279  * the first block of the extent through buf.
6280  *
6281  * returns the tree buffer or NULL.
6282  */
6283 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6284                                         struct btrfs_root *root, u32 blocksize,
6285                                         u64 parent, u64 root_objectid,
6286                                         struct btrfs_disk_key *key, int level,
6287                                         u64 hint, u64 empty_size, int for_cow)
6288 {
6289         struct btrfs_key ins;
6290         struct btrfs_block_rsv *block_rsv;
6291         struct extent_buffer *buf;
6292         u64 flags = 0;
6293         int ret;
6294
6295
6296         block_rsv = use_block_rsv(trans, root, blocksize);
6297         if (IS_ERR(block_rsv))
6298                 return ERR_CAST(block_rsv);
6299
6300         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6301                                    empty_size, hint, (u64)-1, &ins, 0);
6302         if (ret) {
6303                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6304                 return ERR_PTR(ret);
6305         }
6306
6307         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6308                                     blocksize, level);
6309         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6310
6311         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6312                 if (parent == 0)
6313                         parent = ins.objectid;
6314                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6315         } else
6316                 BUG_ON(parent > 0);
6317
6318         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6319                 struct btrfs_delayed_extent_op *extent_op;
6320                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
6321                 BUG_ON(!extent_op); /* -ENOMEM */
6322                 if (key)
6323                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6324                 else
6325                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6326                 extent_op->flags_to_set = flags;
6327                 extent_op->update_key = 1;
6328                 extent_op->update_flags = 1;
6329                 extent_op->is_data = 0;
6330
6331                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6332                                         ins.objectid,
6333                                         ins.offset, parent, root_objectid,
6334                                         level, BTRFS_ADD_DELAYED_EXTENT,
6335                                         extent_op, for_cow);
6336                 BUG_ON(ret); /* -ENOMEM */
6337         }
6338         return buf;
6339 }
6340
6341 struct walk_control {
6342         u64 refs[BTRFS_MAX_LEVEL];
6343         u64 flags[BTRFS_MAX_LEVEL];
6344         struct btrfs_key update_progress;
6345         int stage;
6346         int level;
6347         int shared_level;
6348         int update_ref;
6349         int keep_locks;
6350         int reada_slot;
6351         int reada_count;
6352         int for_reloc;
6353 };
6354
6355 #define DROP_REFERENCE  1
6356 #define UPDATE_BACKREF  2
6357
6358 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6359                                      struct btrfs_root *root,
6360                                      struct walk_control *wc,
6361                                      struct btrfs_path *path)
6362 {
6363         u64 bytenr;
6364         u64 generation;
6365         u64 refs;
6366         u64 flags;
6367         u32 nritems;
6368         u32 blocksize;
6369         struct btrfs_key key;
6370         struct extent_buffer *eb;
6371         int ret;
6372         int slot;
6373         int nread = 0;
6374
6375         if (path->slots[wc->level] < wc->reada_slot) {
6376                 wc->reada_count = wc->reada_count * 2 / 3;
6377                 wc->reada_count = max(wc->reada_count, 2);
6378         } else {
6379                 wc->reada_count = wc->reada_count * 3 / 2;
6380                 wc->reada_count = min_t(int, wc->reada_count,
6381                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6382         }
6383
6384         eb = path->nodes[wc->level];
6385         nritems = btrfs_header_nritems(eb);
6386         blocksize = btrfs_level_size(root, wc->level - 1);
6387
6388         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6389                 if (nread >= wc->reada_count)
6390                         break;
6391
6392                 cond_resched();
6393                 bytenr = btrfs_node_blockptr(eb, slot);
6394                 generation = btrfs_node_ptr_generation(eb, slot);
6395
6396                 if (slot == path->slots[wc->level])
6397                         goto reada;
6398
6399                 if (wc->stage == UPDATE_BACKREF &&
6400                     generation <= root->root_key.offset)
6401                         continue;
6402
6403                 /* We don't lock the tree block, it's OK to be racy here */
6404                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6405                                                &refs, &flags);
6406                 /* We don't care about errors in readahead. */
6407                 if (ret < 0)
6408                         continue;
6409                 BUG_ON(refs == 0);
6410
6411                 if (wc->stage == DROP_REFERENCE) {
6412                         if (refs == 1)
6413                                 goto reada;
6414
6415                         if (wc->level == 1 &&
6416                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6417                                 continue;
6418                         if (!wc->update_ref ||
6419                             generation <= root->root_key.offset)
6420                                 continue;
6421                         btrfs_node_key_to_cpu(eb, &key, slot);
6422                         ret = btrfs_comp_cpu_keys(&key,
6423                                                   &wc->update_progress);
6424                         if (ret < 0)
6425                                 continue;
6426                 } else {
6427                         if (wc->level == 1 &&
6428                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6429                                 continue;
6430                 }
6431 reada:
6432                 ret = readahead_tree_block(root, bytenr, blocksize,
6433                                            generation);
6434                 if (ret)
6435                         break;
6436                 nread++;
6437         }
6438         wc->reada_slot = slot;
6439 }
6440
6441 /*
6442  * hepler to process tree block while walking down the tree.
6443  *
6444  * when wc->stage == UPDATE_BACKREF, this function updates
6445  * back refs for pointers in the block.
6446  *
6447  * NOTE: return value 1 means we should stop walking down.
6448  */
6449 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6450                                    struct btrfs_root *root,
6451                                    struct btrfs_path *path,
6452                                    struct walk_control *wc, int lookup_info)
6453 {
6454         int level = wc->level;
6455         struct extent_buffer *eb = path->nodes[level];
6456         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6457         int ret;
6458
6459         if (wc->stage == UPDATE_BACKREF &&
6460             btrfs_header_owner(eb) != root->root_key.objectid)
6461                 return 1;
6462
6463         /*
6464          * when reference count of tree block is 1, it won't increase
6465          * again. once full backref flag is set, we never clear it.
6466          */
6467         if (lookup_info &&
6468             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6469              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6470                 BUG_ON(!path->locks[level]);
6471                 ret = btrfs_lookup_extent_info(trans, root,
6472                                                eb->start, eb->len,
6473                                                &wc->refs[level],
6474                                                &wc->flags[level]);
6475                 BUG_ON(ret == -ENOMEM);
6476                 if (ret)
6477                         return ret;
6478                 BUG_ON(wc->refs[level] == 0);
6479         }
6480
6481         if (wc->stage == DROP_REFERENCE) {
6482                 if (wc->refs[level] > 1)
6483                         return 1;
6484
6485                 if (path->locks[level] && !wc->keep_locks) {
6486                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6487                         path->locks[level] = 0;
6488                 }
6489                 return 0;
6490         }
6491
6492         /* wc->stage == UPDATE_BACKREF */
6493         if (!(wc->flags[level] & flag)) {
6494                 BUG_ON(!path->locks[level]);
6495                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
6496                 BUG_ON(ret); /* -ENOMEM */
6497                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
6498                 BUG_ON(ret); /* -ENOMEM */
6499                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6500                                                   eb->len, flag, 0);
6501                 BUG_ON(ret); /* -ENOMEM */
6502                 wc->flags[level] |= flag;
6503         }
6504
6505         /*
6506          * the block is shared by multiple trees, so it's not good to
6507          * keep the tree lock
6508          */
6509         if (path->locks[level] && level > 0) {
6510                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6511                 path->locks[level] = 0;
6512         }
6513         return 0;
6514 }
6515
6516 /*
6517  * hepler to process tree block pointer.
6518  *
6519  * when wc->stage == DROP_REFERENCE, this function checks
6520  * reference count of the block pointed to. if the block
6521  * is shared and we need update back refs for the subtree
6522  * rooted at the block, this function changes wc->stage to
6523  * UPDATE_BACKREF. if the block is shared and there is no
6524  * need to update back, this function drops the reference
6525  * to the block.
6526  *
6527  * NOTE: return value 1 means we should stop walking down.
6528  */
6529 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6530                                  struct btrfs_root *root,
6531                                  struct btrfs_path *path,
6532                                  struct walk_control *wc, int *lookup_info)
6533 {
6534         u64 bytenr;
6535         u64 generation;
6536         u64 parent;
6537         u32 blocksize;
6538         struct btrfs_key key;
6539         struct extent_buffer *next;
6540         int level = wc->level;
6541         int reada = 0;
6542         int ret = 0;
6543
6544         generation = btrfs_node_ptr_generation(path->nodes[level],
6545                                                path->slots[level]);
6546         /*
6547          * if the lower level block was created before the snapshot
6548          * was created, we know there is no need to update back refs
6549          * for the subtree
6550          */
6551         if (wc->stage == UPDATE_BACKREF &&
6552             generation <= root->root_key.offset) {
6553                 *lookup_info = 1;
6554                 return 1;
6555         }
6556
6557         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6558         blocksize = btrfs_level_size(root, level - 1);
6559
6560         next = btrfs_find_tree_block(root, bytenr, blocksize);
6561         if (!next) {
6562                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6563                 if (!next)
6564                         return -ENOMEM;
6565                 reada = 1;
6566         }
6567         btrfs_tree_lock(next);
6568         btrfs_set_lock_blocking(next);
6569
6570         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6571                                        &wc->refs[level - 1],
6572                                        &wc->flags[level - 1]);
6573         if (ret < 0) {
6574                 btrfs_tree_unlock(next);
6575                 return ret;
6576         }
6577
6578         BUG_ON(wc->refs[level - 1] == 0);
6579         *lookup_info = 0;
6580
6581         if (wc->stage == DROP_REFERENCE) {
6582                 if (wc->refs[level - 1] > 1) {
6583                         if (level == 1 &&
6584                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6585                                 goto skip;
6586
6587                         if (!wc->update_ref ||
6588                             generation <= root->root_key.offset)
6589                                 goto skip;
6590
6591                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6592                                               path->slots[level]);
6593                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6594                         if (ret < 0)
6595                                 goto skip;
6596
6597                         wc->stage = UPDATE_BACKREF;
6598                         wc->shared_level = level - 1;
6599                 }
6600         } else {
6601                 if (level == 1 &&
6602                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6603                         goto skip;
6604         }
6605
6606         if (!btrfs_buffer_uptodate(next, generation)) {
6607                 btrfs_tree_unlock(next);
6608                 free_extent_buffer(next);
6609                 next = NULL;
6610                 *lookup_info = 1;
6611         }
6612
6613         if (!next) {
6614                 if (reada && level == 1)
6615                         reada_walk_down(trans, root, wc, path);
6616                 next = read_tree_block(root, bytenr, blocksize, generation);
6617                 if (!next)
6618                         return -EIO;
6619                 btrfs_tree_lock(next);
6620                 btrfs_set_lock_blocking(next);
6621         }
6622
6623         level--;
6624         BUG_ON(level != btrfs_header_level(next));
6625         path->nodes[level] = next;
6626         path->slots[level] = 0;
6627         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6628         wc->level = level;
6629         if (wc->level == 1)
6630                 wc->reada_slot = 0;
6631         return 0;
6632 skip:
6633         wc->refs[level - 1] = 0;
6634         wc->flags[level - 1] = 0;
6635         if (wc->stage == DROP_REFERENCE) {
6636                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6637                         parent = path->nodes[level]->start;
6638                 } else {
6639                         BUG_ON(root->root_key.objectid !=
6640                                btrfs_header_owner(path->nodes[level]));
6641                         parent = 0;
6642                 }
6643
6644                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6645                                 root->root_key.objectid, level - 1, 0, 0);
6646                 BUG_ON(ret); /* -ENOMEM */
6647         }
6648         btrfs_tree_unlock(next);
6649         free_extent_buffer(next);
6650         *lookup_info = 1;
6651         return 1;
6652 }
6653
6654 /*
6655  * hepler to process tree block while walking up the tree.
6656  *
6657  * when wc->stage == DROP_REFERENCE, this function drops
6658  * reference count on the block.
6659  *
6660  * when wc->stage == UPDATE_BACKREF, this function changes
6661  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6662  * to UPDATE_BACKREF previously while processing the block.
6663  *
6664  * NOTE: return value 1 means we should stop walking up.
6665  */
6666 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6667                                  struct btrfs_root *root,
6668                                  struct btrfs_path *path,
6669                                  struct walk_control *wc)
6670 {
6671         int ret;
6672         int level = wc->level;
6673         struct extent_buffer *eb = path->nodes[level];
6674         u64 parent = 0;
6675
6676         if (wc->stage == UPDATE_BACKREF) {
6677                 BUG_ON(wc->shared_level < level);
6678                 if (level < wc->shared_level)
6679                         goto out;
6680
6681                 ret = find_next_key(path, level + 1, &wc->update_progress);
6682                 if (ret > 0)
6683                         wc->update_ref = 0;
6684
6685                 wc->stage = DROP_REFERENCE;
6686                 wc->shared_level = -1;
6687                 path->slots[level] = 0;
6688
6689                 /*
6690                  * check reference count again if the block isn't locked.
6691                  * we should start walking down the tree again if reference
6692                  * count is one.
6693                  */
6694                 if (!path->locks[level]) {
6695                         BUG_ON(level == 0);
6696                         btrfs_tree_lock(eb);
6697                         btrfs_set_lock_blocking(eb);
6698                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6699
6700                         ret = btrfs_lookup_extent_info(trans, root,
6701                                                        eb->start, eb->len,
6702                                                        &wc->refs[level],
6703                                                        &wc->flags[level]);
6704                         if (ret < 0) {
6705                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6706                                 return ret;
6707                         }
6708                         BUG_ON(wc->refs[level] == 0);
6709                         if (wc->refs[level] == 1) {
6710                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6711                                 return 1;
6712                         }
6713                 }
6714         }
6715
6716         /* wc->stage == DROP_REFERENCE */
6717         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6718
6719         if (wc->refs[level] == 1) {
6720                 if (level == 0) {
6721                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6722                                 ret = btrfs_dec_ref(trans, root, eb, 1,
6723                                                     wc->for_reloc);
6724                         else
6725                                 ret = btrfs_dec_ref(trans, root, eb, 0,
6726                                                     wc->for_reloc);
6727                         BUG_ON(ret); /* -ENOMEM */
6728                 }
6729                 /* make block locked assertion in clean_tree_block happy */
6730                 if (!path->locks[level] &&
6731                     btrfs_header_generation(eb) == trans->transid) {
6732                         btrfs_tree_lock(eb);
6733                         btrfs_set_lock_blocking(eb);
6734                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6735                 }
6736                 clean_tree_block(trans, root, eb);
6737         }
6738
6739         if (eb == root->node) {
6740                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6741                         parent = eb->start;
6742                 else
6743                         BUG_ON(root->root_key.objectid !=
6744                                btrfs_header_owner(eb));
6745         } else {
6746                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6747                         parent = path->nodes[level + 1]->start;
6748                 else
6749                         BUG_ON(root->root_key.objectid !=
6750                                btrfs_header_owner(path->nodes[level + 1]));
6751         }
6752
6753         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1, 0);
6754 out:
6755         wc->refs[level] = 0;
6756         wc->flags[level] = 0;
6757         return 0;
6758 }
6759
6760 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6761                                    struct btrfs_root *root,
6762                                    struct btrfs_path *path,
6763                                    struct walk_control *wc)
6764 {
6765         int level = wc->level;
6766         int lookup_info = 1;
6767         int ret;
6768
6769         while (level >= 0) {
6770                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6771                 if (ret > 0)
6772                         break;
6773
6774                 if (level == 0)
6775                         break;
6776
6777                 if (path->slots[level] >=
6778                     btrfs_header_nritems(path->nodes[level]))
6779                         break;
6780
6781                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6782                 if (ret > 0) {
6783                         path->slots[level]++;
6784                         continue;
6785                 } else if (ret < 0)
6786                         return ret;
6787                 level = wc->level;
6788         }
6789         return 0;
6790 }
6791
6792 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6793                                  struct btrfs_root *root,
6794                                  struct btrfs_path *path,
6795                                  struct walk_control *wc, int max_level)
6796 {
6797         int level = wc->level;
6798         int ret;
6799
6800         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6801         while (level < max_level && path->nodes[level]) {
6802                 wc->level = level;
6803                 if (path->slots[level] + 1 <
6804                     btrfs_header_nritems(path->nodes[level])) {
6805                         path->slots[level]++;
6806                         return 0;
6807                 } else {
6808                         ret = walk_up_proc(trans, root, path, wc);
6809                         if (ret > 0)
6810                                 return 0;
6811
6812                         if (path->locks[level]) {
6813                                 btrfs_tree_unlock_rw(path->nodes[level],
6814                                                      path->locks[level]);
6815                                 path->locks[level] = 0;
6816                         }
6817                         free_extent_buffer(path->nodes[level]);
6818                         path->nodes[level] = NULL;
6819                         level++;
6820                 }
6821         }
6822         return 1;
6823 }
6824
6825 /*
6826  * drop a subvolume tree.
6827  *
6828  * this function traverses the tree freeing any blocks that only
6829  * referenced by the tree.
6830  *
6831  * when a shared tree block is found. this function decreases its
6832  * reference count by one. if update_ref is true, this function
6833  * also make sure backrefs for the shared block and all lower level
6834  * blocks are properly updated.
6835  */
6836 int btrfs_drop_snapshot(struct btrfs_root *root,
6837                          struct btrfs_block_rsv *block_rsv, int update_ref,
6838                          int for_reloc)
6839 {
6840         struct btrfs_path *path;
6841         struct btrfs_trans_handle *trans;
6842         struct btrfs_root *tree_root = root->fs_info->tree_root;
6843         struct btrfs_root_item *root_item = &root->root_item;
6844         struct walk_control *wc;
6845         struct btrfs_key key;
6846         int err = 0;
6847         int ret;
6848         int level;
6849
6850         path = btrfs_alloc_path();
6851         if (!path) {
6852                 err = -ENOMEM;
6853                 goto out;
6854         }
6855
6856         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6857         if (!wc) {
6858                 btrfs_free_path(path);
6859                 err = -ENOMEM;
6860                 goto out;
6861         }
6862
6863         trans = btrfs_start_transaction(tree_root, 0);
6864         if (IS_ERR(trans)) {
6865                 err = PTR_ERR(trans);
6866                 goto out_free;
6867         }
6868
6869         if (block_rsv)
6870                 trans->block_rsv = block_rsv;
6871
6872         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6873                 level = btrfs_header_level(root->node);
6874                 path->nodes[level] = btrfs_lock_root_node(root);
6875                 btrfs_set_lock_blocking(path->nodes[level]);
6876                 path->slots[level] = 0;
6877                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6878                 memset(&wc->update_progress, 0,
6879                        sizeof(wc->update_progress));
6880         } else {
6881                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6882                 memcpy(&wc->update_progress, &key,
6883                        sizeof(wc->update_progress));
6884
6885                 level = root_item->drop_level;
6886                 BUG_ON(level == 0);
6887                 path->lowest_level = level;
6888                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6889                 path->lowest_level = 0;
6890                 if (ret < 0) {
6891                         err = ret;
6892                         goto out_end_trans;
6893                 }
6894                 WARN_ON(ret > 0);
6895
6896                 /*
6897                  * unlock our path, this is safe because only this
6898                  * function is allowed to delete this snapshot
6899                  */
6900                 btrfs_unlock_up_safe(path, 0);
6901
6902                 level = btrfs_header_level(root->node);
6903                 while (1) {
6904                         btrfs_tree_lock(path->nodes[level]);
6905                         btrfs_set_lock_blocking(path->nodes[level]);
6906
6907                         ret = btrfs_lookup_extent_info(trans, root,
6908                                                 path->nodes[level]->start,
6909                                                 path->nodes[level]->len,
6910                                                 &wc->refs[level],
6911                                                 &wc->flags[level]);
6912                         if (ret < 0) {
6913                                 err = ret;
6914                                 goto out_end_trans;
6915                         }
6916                         BUG_ON(wc->refs[level] == 0);
6917
6918                         if (level == root_item->drop_level)
6919                                 break;
6920
6921                         btrfs_tree_unlock(path->nodes[level]);
6922                         WARN_ON(wc->refs[level] != 1);
6923                         level--;
6924                 }
6925         }
6926
6927         wc->level = level;
6928         wc->shared_level = -1;
6929         wc->stage = DROP_REFERENCE;
6930         wc->update_ref = update_ref;
6931         wc->keep_locks = 0;
6932         wc->for_reloc = for_reloc;
6933         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6934
6935         while (1) {
6936                 ret = walk_down_tree(trans, root, path, wc);
6937                 if (ret < 0) {
6938                         err = ret;
6939                         break;
6940                 }
6941
6942                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
6943                 if (ret < 0) {
6944                         err = ret;
6945                         break;
6946                 }
6947
6948                 if (ret > 0) {
6949                         BUG_ON(wc->stage != DROP_REFERENCE);
6950                         break;
6951                 }
6952
6953                 if (wc->stage == DROP_REFERENCE) {
6954                         level = wc->level;
6955                         btrfs_node_key(path->nodes[level],
6956                                        &root_item->drop_progress,
6957                                        path->slots[level]);
6958                         root_item->drop_level = level;
6959                 }
6960
6961                 BUG_ON(wc->level == 0);
6962                 if (btrfs_should_end_transaction(trans, tree_root)) {
6963                         ret = btrfs_update_root(trans, tree_root,
6964                                                 &root->root_key,
6965                                                 root_item);
6966                         if (ret) {
6967                                 btrfs_abort_transaction(trans, tree_root, ret);
6968                                 err = ret;
6969                                 goto out_end_trans;
6970                         }
6971
6972                         btrfs_end_transaction_throttle(trans, tree_root);
6973                         trans = btrfs_start_transaction(tree_root, 0);
6974                         if (IS_ERR(trans)) {
6975                                 err = PTR_ERR(trans);
6976                                 goto out_free;
6977                         }
6978                         if (block_rsv)
6979                                 trans->block_rsv = block_rsv;
6980                 }
6981         }
6982         btrfs_release_path(path);
6983         if (err)
6984                 goto out_end_trans;
6985
6986         ret = btrfs_del_root(trans, tree_root, &root->root_key);
6987         if (ret) {
6988                 btrfs_abort_transaction(trans, tree_root, ret);
6989                 goto out_end_trans;
6990         }
6991
6992         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
6993                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
6994                                            NULL, NULL);
6995                 if (ret < 0) {
6996                         btrfs_abort_transaction(trans, tree_root, ret);
6997                         err = ret;
6998                         goto out_end_trans;
6999                 } else if (ret > 0) {
7000                         /* if we fail to delete the orphan item this time
7001                          * around, it'll get picked up the next time.
7002                          *
7003                          * The most common failure here is just -ENOENT.
7004                          */
7005                         btrfs_del_orphan_item(trans, tree_root,
7006                                               root->root_key.objectid);
7007                 }
7008         }
7009
7010         if (root->in_radix) {
7011                 btrfs_free_fs_root(tree_root->fs_info, root);
7012         } else {
7013                 free_extent_buffer(root->node);
7014                 free_extent_buffer(root->commit_root);
7015                 kfree(root);
7016         }
7017 out_end_trans:
7018         btrfs_end_transaction_throttle(trans, tree_root);
7019 out_free:
7020         kfree(wc);
7021         btrfs_free_path(path);
7022 out:
7023         if (err)
7024                 btrfs_std_error(root->fs_info, err);
7025         return err;
7026 }
7027
7028 /*
7029  * drop subtree rooted at tree block 'node'.
7030  *
7031  * NOTE: this function will unlock and release tree block 'node'
7032  * only used by relocation code
7033  */
7034 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7035                         struct btrfs_root *root,
7036                         struct extent_buffer *node,
7037                         struct extent_buffer *parent)
7038 {
7039         struct btrfs_path *path;
7040         struct walk_control *wc;
7041         int level;
7042         int parent_level;
7043         int ret = 0;
7044         int wret;
7045
7046         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7047
7048         path = btrfs_alloc_path();
7049         if (!path)
7050                 return -ENOMEM;
7051
7052         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7053         if (!wc) {
7054                 btrfs_free_path(path);
7055                 return -ENOMEM;
7056         }
7057
7058         btrfs_assert_tree_locked(parent);
7059         parent_level = btrfs_header_level(parent);
7060         extent_buffer_get(parent);
7061         path->nodes[parent_level] = parent;
7062         path->slots[parent_level] = btrfs_header_nritems(parent);
7063
7064         btrfs_assert_tree_locked(node);
7065         level = btrfs_header_level(node);
7066         path->nodes[level] = node;
7067         path->slots[level] = 0;
7068         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7069
7070         wc->refs[parent_level] = 1;
7071         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7072         wc->level = level;
7073         wc->shared_level = -1;
7074         wc->stage = DROP_REFERENCE;
7075         wc->update_ref = 0;
7076         wc->keep_locks = 1;
7077         wc->for_reloc = 1;
7078         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7079
7080         while (1) {
7081                 wret = walk_down_tree(trans, root, path, wc);
7082                 if (wret < 0) {
7083                         ret = wret;
7084                         break;
7085                 }
7086
7087                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7088                 if (wret < 0)
7089                         ret = wret;
7090                 if (wret != 0)
7091                         break;
7092         }
7093
7094         kfree(wc);
7095         btrfs_free_path(path);
7096         return ret;
7097 }
7098
7099 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7100 {
7101         u64 num_devices;
7102         u64 stripped = BTRFS_BLOCK_GROUP_RAID0 |
7103                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7104
7105         if (root->fs_info->balance_ctl) {
7106                 struct btrfs_balance_control *bctl = root->fs_info->balance_ctl;
7107                 u64 tgt = 0;
7108
7109                 /* pick restriper's target profile and return */
7110                 if (flags & BTRFS_BLOCK_GROUP_DATA &&
7111                     bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
7112                         tgt = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
7113                 } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
7114                            bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
7115                         tgt = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
7116                 } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
7117                            bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
7118                         tgt = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
7119                 }
7120
7121                 if (tgt) {
7122                         /* extended -> chunk profile */
7123                         tgt &= ~BTRFS_AVAIL_ALLOC_BIT_SINGLE;
7124                         return tgt;
7125                 }
7126         }
7127
7128         /*
7129          * we add in the count of missing devices because we want
7130          * to make sure that any RAID levels on a degraded FS
7131          * continue to be honored.
7132          */
7133         num_devices = root->fs_info->fs_devices->rw_devices +
7134                 root->fs_info->fs_devices->missing_devices;
7135
7136         if (num_devices == 1) {
7137                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7138                 stripped = flags & ~stripped;
7139
7140                 /* turn raid0 into single device chunks */
7141                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7142                         return stripped;
7143
7144                 /* turn mirroring into duplication */
7145                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7146                              BTRFS_BLOCK_GROUP_RAID10))
7147                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7148                 return flags;
7149         } else {
7150                 /* they already had raid on here, just return */
7151                 if (flags & stripped)
7152                         return flags;
7153
7154                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7155                 stripped = flags & ~stripped;
7156
7157                 /* switch duplicated blocks with raid1 */
7158                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7159                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7160
7161                 /* turn single device chunks into raid0 */
7162                 return stripped | BTRFS_BLOCK_GROUP_RAID0;
7163         }
7164         return flags;
7165 }
7166
7167 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7168 {
7169         struct btrfs_space_info *sinfo = cache->space_info;
7170         u64 num_bytes;
7171         u64 min_allocable_bytes;
7172         int ret = -ENOSPC;
7173
7174
7175         /*
7176          * We need some metadata space and system metadata space for
7177          * allocating chunks in some corner cases until we force to set
7178          * it to be readonly.
7179          */
7180         if ((sinfo->flags &
7181              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7182             !force)
7183                 min_allocable_bytes = 1 * 1024 * 1024;
7184         else
7185                 min_allocable_bytes = 0;
7186
7187         spin_lock(&sinfo->lock);
7188         spin_lock(&cache->lock);
7189
7190         if (cache->ro) {
7191                 ret = 0;
7192                 goto out;
7193         }
7194
7195         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7196                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7197
7198         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7199             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7200             min_allocable_bytes <= sinfo->total_bytes) {
7201                 sinfo->bytes_readonly += num_bytes;
7202                 cache->ro = 1;
7203                 ret = 0;
7204         }
7205 out:
7206         spin_unlock(&cache->lock);
7207         spin_unlock(&sinfo->lock);
7208         return ret;
7209 }
7210
7211 int btrfs_set_block_group_ro(struct btrfs_root *root,
7212                              struct btrfs_block_group_cache *cache)
7213
7214 {
7215         struct btrfs_trans_handle *trans;
7216         u64 alloc_flags;
7217         int ret;
7218
7219         BUG_ON(cache->ro);
7220
7221         trans = btrfs_join_transaction(root);
7222         if (IS_ERR(trans))
7223                 return PTR_ERR(trans);
7224
7225         alloc_flags = update_block_group_flags(root, cache->flags);
7226         if (alloc_flags != cache->flags) {
7227                 ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7228                                      CHUNK_ALLOC_FORCE);
7229                 if (ret < 0)
7230                         goto out;
7231         }
7232
7233         ret = set_block_group_ro(cache, 0);
7234         if (!ret)
7235                 goto out;
7236         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7237         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7238                              CHUNK_ALLOC_FORCE);
7239         if (ret < 0)
7240                 goto out;
7241         ret = set_block_group_ro(cache, 0);
7242 out:
7243         btrfs_end_transaction(trans, root);
7244         return ret;
7245 }
7246
7247 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7248                             struct btrfs_root *root, u64 type)
7249 {
7250         u64 alloc_flags = get_alloc_profile(root, type);
7251         return do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7252                               CHUNK_ALLOC_FORCE);
7253 }
7254
7255 /*
7256  * helper to account the unused space of all the readonly block group in the
7257  * list. takes mirrors into account.
7258  */
7259 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7260 {
7261         struct btrfs_block_group_cache *block_group;
7262         u64 free_bytes = 0;
7263         int factor;
7264
7265         list_for_each_entry(block_group, groups_list, list) {
7266                 spin_lock(&block_group->lock);
7267
7268                 if (!block_group->ro) {
7269                         spin_unlock(&block_group->lock);
7270                         continue;
7271                 }
7272
7273                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7274                                           BTRFS_BLOCK_GROUP_RAID10 |
7275                                           BTRFS_BLOCK_GROUP_DUP))
7276                         factor = 2;
7277                 else
7278                         factor = 1;
7279
7280                 free_bytes += (block_group->key.offset -
7281                                btrfs_block_group_used(&block_group->item)) *
7282                                factor;
7283
7284                 spin_unlock(&block_group->lock);
7285         }
7286
7287         return free_bytes;
7288 }
7289
7290 /*
7291  * helper to account the unused space of all the readonly block group in the
7292  * space_info. takes mirrors into account.
7293  */
7294 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7295 {
7296         int i;
7297         u64 free_bytes = 0;
7298
7299         spin_lock(&sinfo->lock);
7300
7301         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7302                 if (!list_empty(&sinfo->block_groups[i]))
7303                         free_bytes += __btrfs_get_ro_block_group_free_space(
7304                                                 &sinfo->block_groups[i]);
7305
7306         spin_unlock(&sinfo->lock);
7307
7308         return free_bytes;
7309 }
7310
7311 void btrfs_set_block_group_rw(struct btrfs_root *root,
7312                               struct btrfs_block_group_cache *cache)
7313 {
7314         struct btrfs_space_info *sinfo = cache->space_info;
7315         u64 num_bytes;
7316
7317         BUG_ON(!cache->ro);
7318
7319         spin_lock(&sinfo->lock);
7320         spin_lock(&cache->lock);
7321         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7322                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7323         sinfo->bytes_readonly -= num_bytes;
7324         cache->ro = 0;
7325         spin_unlock(&cache->lock);
7326         spin_unlock(&sinfo->lock);
7327 }
7328
7329 /*
7330  * checks to see if its even possible to relocate this block group.
7331  *
7332  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7333  * ok to go ahead and try.
7334  */
7335 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7336 {
7337         struct btrfs_block_group_cache *block_group;
7338         struct btrfs_space_info *space_info;
7339         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7340         struct btrfs_device *device;
7341         u64 min_free;
7342         u64 dev_min = 1;
7343         u64 dev_nr = 0;
7344         int index;
7345         int full = 0;
7346         int ret = 0;
7347
7348         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7349
7350         /* odd, couldn't find the block group, leave it alone */
7351         if (!block_group)
7352                 return -1;
7353
7354         min_free = btrfs_block_group_used(&block_group->item);
7355
7356         /* no bytes used, we're good */
7357         if (!min_free)
7358                 goto out;
7359
7360         space_info = block_group->space_info;
7361         spin_lock(&space_info->lock);
7362
7363         full = space_info->full;
7364
7365         /*
7366          * if this is the last block group we have in this space, we can't
7367          * relocate it unless we're able to allocate a new chunk below.
7368          *
7369          * Otherwise, we need to make sure we have room in the space to handle
7370          * all of the extents from this block group.  If we can, we're good
7371          */
7372         if ((space_info->total_bytes != block_group->key.offset) &&
7373             (space_info->bytes_used + space_info->bytes_reserved +
7374              space_info->bytes_pinned + space_info->bytes_readonly +
7375              min_free < space_info->total_bytes)) {
7376                 spin_unlock(&space_info->lock);
7377                 goto out;
7378         }
7379         spin_unlock(&space_info->lock);
7380
7381         /*
7382          * ok we don't have enough space, but maybe we have free space on our
7383          * devices to allocate new chunks for relocation, so loop through our
7384          * alloc devices and guess if we have enough space.  However, if we
7385          * were marked as full, then we know there aren't enough chunks, and we
7386          * can just return.
7387          */
7388         ret = -1;
7389         if (full)
7390                 goto out;
7391
7392         /*
7393          * index:
7394          *      0: raid10
7395          *      1: raid1
7396          *      2: dup
7397          *      3: raid0
7398          *      4: single
7399          */
7400         index = get_block_group_index(block_group);
7401         if (index == 0) {
7402                 dev_min = 4;
7403                 /* Divide by 2 */
7404                 min_free >>= 1;
7405         } else if (index == 1) {
7406                 dev_min = 2;
7407         } else if (index == 2) {
7408                 /* Multiply by 2 */
7409                 min_free <<= 1;
7410         } else if (index == 3) {
7411                 dev_min = fs_devices->rw_devices;
7412                 do_div(min_free, dev_min);
7413         }
7414
7415         mutex_lock(&root->fs_info->chunk_mutex);
7416         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7417                 u64 dev_offset;
7418
7419                 /*
7420                  * check to make sure we can actually find a chunk with enough
7421                  * space to fit our block group in.
7422                  */
7423                 if (device->total_bytes > device->bytes_used + min_free) {
7424                         ret = find_free_dev_extent(device, min_free,
7425                                                    &dev_offset, NULL);
7426                         if (!ret)
7427                                 dev_nr++;
7428
7429                         if (dev_nr >= dev_min)
7430                                 break;
7431
7432                         ret = -1;
7433                 }
7434         }
7435         mutex_unlock(&root->fs_info->chunk_mutex);
7436 out:
7437         btrfs_put_block_group(block_group);
7438         return ret;
7439 }
7440
7441 static int find_first_block_group(struct btrfs_root *root,
7442                 struct btrfs_path *path, struct btrfs_key *key)
7443 {
7444         int ret = 0;
7445         struct btrfs_key found_key;
7446         struct extent_buffer *leaf;
7447         int slot;
7448
7449         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7450         if (ret < 0)
7451                 goto out;
7452
7453         while (1) {
7454                 slot = path->slots[0];
7455                 leaf = path->nodes[0];
7456                 if (slot >= btrfs_header_nritems(leaf)) {
7457                         ret = btrfs_next_leaf(root, path);
7458                         if (ret == 0)
7459                                 continue;
7460                         if (ret < 0)
7461                                 goto out;
7462                         break;
7463                 }
7464                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7465
7466                 if (found_key.objectid >= key->objectid &&
7467                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7468                         ret = 0;
7469                         goto out;
7470                 }
7471                 path->slots[0]++;
7472         }
7473 out:
7474         return ret;
7475 }
7476
7477 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7478 {
7479         struct btrfs_block_group_cache *block_group;
7480         u64 last = 0;
7481
7482         while (1) {
7483                 struct inode *inode;
7484
7485                 block_group = btrfs_lookup_first_block_group(info, last);
7486                 while (block_group) {
7487                         spin_lock(&block_group->lock);
7488                         if (block_group->iref)
7489                                 break;
7490                         spin_unlock(&block_group->lock);
7491                         block_group = next_block_group(info->tree_root,
7492                                                        block_group);
7493                 }
7494                 if (!block_group) {
7495                         if (last == 0)
7496                                 break;
7497                         last = 0;
7498                         continue;
7499                 }
7500
7501                 inode = block_group->inode;
7502                 block_group->iref = 0;
7503                 block_group->inode = NULL;
7504                 spin_unlock(&block_group->lock);
7505                 iput(inode);
7506                 last = block_group->key.objectid + block_group->key.offset;
7507                 btrfs_put_block_group(block_group);
7508         }
7509 }
7510
7511 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7512 {
7513         struct btrfs_block_group_cache *block_group;
7514         struct btrfs_space_info *space_info;
7515         struct btrfs_caching_control *caching_ctl;
7516         struct rb_node *n;
7517
7518         down_write(&info->extent_commit_sem);
7519         while (!list_empty(&info->caching_block_groups)) {
7520                 caching_ctl = list_entry(info->caching_block_groups.next,
7521                                          struct btrfs_caching_control, list);
7522                 list_del(&caching_ctl->list);
7523                 put_caching_control(caching_ctl);
7524         }
7525         up_write(&info->extent_commit_sem);
7526
7527         spin_lock(&info->block_group_cache_lock);
7528         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7529                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7530                                        cache_node);
7531                 rb_erase(&block_group->cache_node,
7532                          &info->block_group_cache_tree);
7533                 spin_unlock(&info->block_group_cache_lock);
7534
7535                 down_write(&block_group->space_info->groups_sem);
7536                 list_del(&block_group->list);
7537                 up_write(&block_group->space_info->groups_sem);
7538
7539                 if (block_group->cached == BTRFS_CACHE_STARTED)
7540                         wait_block_group_cache_done(block_group);
7541
7542                 /*
7543                  * We haven't cached this block group, which means we could
7544                  * possibly have excluded extents on this block group.
7545                  */
7546                 if (block_group->cached == BTRFS_CACHE_NO)
7547                         free_excluded_extents(info->extent_root, block_group);
7548
7549                 btrfs_remove_free_space_cache(block_group);
7550                 btrfs_put_block_group(block_group);
7551
7552                 spin_lock(&info->block_group_cache_lock);
7553         }
7554         spin_unlock(&info->block_group_cache_lock);
7555
7556         /* now that all the block groups are freed, go through and
7557          * free all the space_info structs.  This is only called during
7558          * the final stages of unmount, and so we know nobody is
7559          * using them.  We call synchronize_rcu() once before we start,
7560          * just to be on the safe side.
7561          */
7562         synchronize_rcu();
7563
7564         release_global_block_rsv(info);
7565
7566         while(!list_empty(&info->space_info)) {
7567                 space_info = list_entry(info->space_info.next,
7568                                         struct btrfs_space_info,
7569                                         list);
7570                 if (space_info->bytes_pinned > 0 ||
7571                     space_info->bytes_reserved > 0 ||
7572                     space_info->bytes_may_use > 0) {
7573                         WARN_ON(1);
7574                         dump_space_info(space_info, 0, 0);
7575                 }
7576                 list_del(&space_info->list);
7577                 kfree(space_info);
7578         }
7579         return 0;
7580 }
7581
7582 static void __link_block_group(struct btrfs_space_info *space_info,
7583                                struct btrfs_block_group_cache *cache)
7584 {
7585         int index = get_block_group_index(cache);
7586
7587         down_write(&space_info->groups_sem);
7588         list_add_tail(&cache->list, &space_info->block_groups[index]);
7589         up_write(&space_info->groups_sem);
7590 }
7591
7592 int btrfs_read_block_groups(struct btrfs_root *root)
7593 {
7594         struct btrfs_path *path;
7595         int ret;
7596         struct btrfs_block_group_cache *cache;
7597         struct btrfs_fs_info *info = root->fs_info;
7598         struct btrfs_space_info *space_info;
7599         struct btrfs_key key;
7600         struct btrfs_key found_key;
7601         struct extent_buffer *leaf;
7602         int need_clear = 0;
7603         u64 cache_gen;
7604
7605         root = info->extent_root;
7606         key.objectid = 0;
7607         key.offset = 0;
7608         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7609         path = btrfs_alloc_path();
7610         if (!path)
7611                 return -ENOMEM;
7612         path->reada = 1;
7613
7614         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7615         if (btrfs_test_opt(root, SPACE_CACHE) &&
7616             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7617                 need_clear = 1;
7618         if (btrfs_test_opt(root, CLEAR_CACHE))
7619                 need_clear = 1;
7620
7621         while (1) {
7622                 ret = find_first_block_group(root, path, &key);
7623                 if (ret > 0)
7624                         break;
7625                 if (ret != 0)
7626                         goto error;
7627                 leaf = path->nodes[0];
7628                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7629                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7630                 if (!cache) {
7631                         ret = -ENOMEM;
7632                         goto error;
7633                 }
7634                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7635                                                 GFP_NOFS);
7636                 if (!cache->free_space_ctl) {
7637                         kfree(cache);
7638                         ret = -ENOMEM;
7639                         goto error;
7640                 }
7641
7642                 atomic_set(&cache->count, 1);
7643                 spin_lock_init(&cache->lock);
7644                 cache->fs_info = info;
7645                 INIT_LIST_HEAD(&cache->list);
7646                 INIT_LIST_HEAD(&cache->cluster_list);
7647
7648                 if (need_clear)
7649                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7650
7651                 read_extent_buffer(leaf, &cache->item,
7652                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7653                                    sizeof(cache->item));
7654                 memcpy(&cache->key, &found_key, sizeof(found_key));
7655
7656                 key.objectid = found_key.objectid + found_key.offset;
7657                 btrfs_release_path(path);
7658                 cache->flags = btrfs_block_group_flags(&cache->item);
7659                 cache->sectorsize = root->sectorsize;
7660
7661                 btrfs_init_free_space_ctl(cache);
7662
7663                 /*
7664                  * We need to exclude the super stripes now so that the space
7665                  * info has super bytes accounted for, otherwise we'll think
7666                  * we have more space than we actually do.
7667                  */
7668                 exclude_super_stripes(root, cache);
7669
7670                 /*
7671                  * check for two cases, either we are full, and therefore
7672                  * don't need to bother with the caching work since we won't
7673                  * find any space, or we are empty, and we can just add all
7674                  * the space in and be done with it.  This saves us _alot_ of
7675                  * time, particularly in the full case.
7676                  */
7677                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7678                         cache->last_byte_to_unpin = (u64)-1;
7679                         cache->cached = BTRFS_CACHE_FINISHED;
7680                         free_excluded_extents(root, cache);
7681                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7682                         cache->last_byte_to_unpin = (u64)-1;
7683                         cache->cached = BTRFS_CACHE_FINISHED;
7684                         add_new_free_space(cache, root->fs_info,
7685                                            found_key.objectid,
7686                                            found_key.objectid +
7687                                            found_key.offset);
7688                         free_excluded_extents(root, cache);
7689                 }
7690
7691                 ret = update_space_info(info, cache->flags, found_key.offset,
7692                                         btrfs_block_group_used(&cache->item),
7693                                         &space_info);
7694                 BUG_ON(ret); /* -ENOMEM */
7695                 cache->space_info = space_info;
7696                 spin_lock(&cache->space_info->lock);
7697                 cache->space_info->bytes_readonly += cache->bytes_super;
7698                 spin_unlock(&cache->space_info->lock);
7699
7700                 __link_block_group(space_info, cache);
7701
7702                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7703                 BUG_ON(ret); /* Logic error */
7704
7705                 set_avail_alloc_bits(root->fs_info, cache->flags);
7706                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7707                         set_block_group_ro(cache, 1);
7708         }
7709
7710         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7711                 if (!(get_alloc_profile(root, space_info->flags) &
7712                       (BTRFS_BLOCK_GROUP_RAID10 |
7713                        BTRFS_BLOCK_GROUP_RAID1 |
7714                        BTRFS_BLOCK_GROUP_DUP)))
7715                         continue;
7716                 /*
7717                  * avoid allocating from un-mirrored block group if there are
7718                  * mirrored block groups.
7719                  */
7720                 list_for_each_entry(cache, &space_info->block_groups[3], list)
7721                         set_block_group_ro(cache, 1);
7722                 list_for_each_entry(cache, &space_info->block_groups[4], list)
7723                         set_block_group_ro(cache, 1);
7724         }
7725
7726         init_global_block_rsv(info);
7727         ret = 0;
7728 error:
7729         btrfs_free_path(path);
7730         return ret;
7731 }
7732
7733 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7734                            struct btrfs_root *root, u64 bytes_used,
7735                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7736                            u64 size)
7737 {
7738         int ret;
7739         struct btrfs_root *extent_root;
7740         struct btrfs_block_group_cache *cache;
7741
7742         extent_root = root->fs_info->extent_root;
7743
7744         root->fs_info->last_trans_log_full_commit = trans->transid;
7745
7746         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7747         if (!cache)
7748                 return -ENOMEM;
7749         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7750                                         GFP_NOFS);
7751         if (!cache->free_space_ctl) {
7752                 kfree(cache);
7753                 return -ENOMEM;
7754         }
7755
7756         cache->key.objectid = chunk_offset;
7757         cache->key.offset = size;
7758         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7759         cache->sectorsize = root->sectorsize;
7760         cache->fs_info = root->fs_info;
7761
7762         atomic_set(&cache->count, 1);
7763         spin_lock_init(&cache->lock);
7764         INIT_LIST_HEAD(&cache->list);
7765         INIT_LIST_HEAD(&cache->cluster_list);
7766
7767         btrfs_init_free_space_ctl(cache);
7768
7769         btrfs_set_block_group_used(&cache->item, bytes_used);
7770         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7771         cache->flags = type;
7772         btrfs_set_block_group_flags(&cache->item, type);
7773
7774         cache->last_byte_to_unpin = (u64)-1;
7775         cache->cached = BTRFS_CACHE_FINISHED;
7776         exclude_super_stripes(root, cache);
7777
7778         add_new_free_space(cache, root->fs_info, chunk_offset,
7779                            chunk_offset + size);
7780
7781         free_excluded_extents(root, cache);
7782
7783         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7784                                 &cache->space_info);
7785         BUG_ON(ret); /* -ENOMEM */
7786         update_global_block_rsv(root->fs_info);
7787
7788         spin_lock(&cache->space_info->lock);
7789         cache->space_info->bytes_readonly += cache->bytes_super;
7790         spin_unlock(&cache->space_info->lock);
7791
7792         __link_block_group(cache->space_info, cache);
7793
7794         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7795         BUG_ON(ret); /* Logic error */
7796
7797         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
7798                                 sizeof(cache->item));
7799         if (ret) {
7800                 btrfs_abort_transaction(trans, extent_root, ret);
7801                 return ret;
7802         }
7803
7804         set_avail_alloc_bits(extent_root->fs_info, type);
7805
7806         return 0;
7807 }
7808
7809 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
7810 {
7811         u64 extra_flags = flags & BTRFS_BLOCK_GROUP_PROFILE_MASK;
7812
7813         /* chunk -> extended profile */
7814         if (extra_flags == 0)
7815                 extra_flags = BTRFS_AVAIL_ALLOC_BIT_SINGLE;
7816
7817         if (flags & BTRFS_BLOCK_GROUP_DATA)
7818                 fs_info->avail_data_alloc_bits &= ~extra_flags;
7819         if (flags & BTRFS_BLOCK_GROUP_METADATA)
7820                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
7821         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
7822                 fs_info->avail_system_alloc_bits &= ~extra_flags;
7823 }
7824
7825 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7826                              struct btrfs_root *root, u64 group_start)
7827 {
7828         struct btrfs_path *path;
7829         struct btrfs_block_group_cache *block_group;
7830         struct btrfs_free_cluster *cluster;
7831         struct btrfs_root *tree_root = root->fs_info->tree_root;
7832         struct btrfs_key key;
7833         struct inode *inode;
7834         int ret;
7835         int index;
7836         int factor;
7837
7838         root = root->fs_info->extent_root;
7839
7840         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7841         BUG_ON(!block_group);
7842         BUG_ON(!block_group->ro);
7843
7844         /*
7845          * Free the reserved super bytes from this block group before
7846          * remove it.
7847          */
7848         free_excluded_extents(root, block_group);
7849
7850         memcpy(&key, &block_group->key, sizeof(key));
7851         index = get_block_group_index(block_group);
7852         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7853                                   BTRFS_BLOCK_GROUP_RAID1 |
7854                                   BTRFS_BLOCK_GROUP_RAID10))
7855                 factor = 2;
7856         else
7857                 factor = 1;
7858
7859         /* make sure this block group isn't part of an allocation cluster */
7860         cluster = &root->fs_info->data_alloc_cluster;
7861         spin_lock(&cluster->refill_lock);
7862         btrfs_return_cluster_to_free_space(block_group, cluster);
7863         spin_unlock(&cluster->refill_lock);
7864
7865         /*
7866          * make sure this block group isn't part of a metadata
7867          * allocation cluster
7868          */
7869         cluster = &root->fs_info->meta_alloc_cluster;
7870         spin_lock(&cluster->refill_lock);
7871         btrfs_return_cluster_to_free_space(block_group, cluster);
7872         spin_unlock(&cluster->refill_lock);
7873
7874         path = btrfs_alloc_path();
7875         if (!path) {
7876                 ret = -ENOMEM;
7877                 goto out;
7878         }
7879
7880         inode = lookup_free_space_inode(tree_root, block_group, path);
7881         if (!IS_ERR(inode)) {
7882                 ret = btrfs_orphan_add(trans, inode);
7883                 if (ret) {
7884                         btrfs_add_delayed_iput(inode);
7885                         goto out;
7886                 }
7887                 clear_nlink(inode);
7888                 /* One for the block groups ref */
7889                 spin_lock(&block_group->lock);
7890                 if (block_group->iref) {
7891                         block_group->iref = 0;
7892                         block_group->inode = NULL;
7893                         spin_unlock(&block_group->lock);
7894                         iput(inode);
7895                 } else {
7896                         spin_unlock(&block_group->lock);
7897                 }
7898                 /* One for our lookup ref */
7899                 btrfs_add_delayed_iput(inode);
7900         }
7901
7902         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
7903         key.offset = block_group->key.objectid;
7904         key.type = 0;
7905
7906         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
7907         if (ret < 0)
7908                 goto out;
7909         if (ret > 0)
7910                 btrfs_release_path(path);
7911         if (ret == 0) {
7912                 ret = btrfs_del_item(trans, tree_root, path);
7913                 if (ret)
7914                         goto out;
7915                 btrfs_release_path(path);
7916         }
7917
7918         spin_lock(&root->fs_info->block_group_cache_lock);
7919         rb_erase(&block_group->cache_node,
7920                  &root->fs_info->block_group_cache_tree);
7921         spin_unlock(&root->fs_info->block_group_cache_lock);
7922
7923         down_write(&block_group->space_info->groups_sem);
7924         /*
7925          * we must use list_del_init so people can check to see if they
7926          * are still on the list after taking the semaphore
7927          */
7928         list_del_init(&block_group->list);
7929         if (list_empty(&block_group->space_info->block_groups[index]))
7930                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
7931         up_write(&block_group->space_info->groups_sem);
7932
7933         if (block_group->cached == BTRFS_CACHE_STARTED)
7934                 wait_block_group_cache_done(block_group);
7935
7936         btrfs_remove_free_space_cache(block_group);
7937
7938         spin_lock(&block_group->space_info->lock);
7939         block_group->space_info->total_bytes -= block_group->key.offset;
7940         block_group->space_info->bytes_readonly -= block_group->key.offset;
7941         block_group->space_info->disk_total -= block_group->key.offset * factor;
7942         spin_unlock(&block_group->space_info->lock);
7943
7944         memcpy(&key, &block_group->key, sizeof(key));
7945
7946         btrfs_clear_space_info_full(root->fs_info);
7947
7948         btrfs_put_block_group(block_group);
7949         btrfs_put_block_group(block_group);
7950
7951         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
7952         if (ret > 0)
7953                 ret = -EIO;
7954         if (ret < 0)
7955                 goto out;
7956
7957         ret = btrfs_del_item(trans, root, path);
7958 out:
7959         btrfs_free_path(path);
7960         return ret;
7961 }
7962
7963 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
7964 {
7965         struct btrfs_space_info *space_info;
7966         struct btrfs_super_block *disk_super;
7967         u64 features;
7968         u64 flags;
7969         int mixed = 0;
7970         int ret;
7971
7972         disk_super = fs_info->super_copy;
7973         if (!btrfs_super_root(disk_super))
7974                 return 1;
7975
7976         features = btrfs_super_incompat_flags(disk_super);
7977         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
7978                 mixed = 1;
7979
7980         flags = BTRFS_BLOCK_GROUP_SYSTEM;
7981         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7982         if (ret)
7983                 goto out;
7984
7985         if (mixed) {
7986                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
7987                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7988         } else {
7989                 flags = BTRFS_BLOCK_GROUP_METADATA;
7990                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7991                 if (ret)
7992                         goto out;
7993
7994                 flags = BTRFS_BLOCK_GROUP_DATA;
7995                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7996         }
7997 out:
7998         return ret;
7999 }
8000
8001 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8002 {
8003         return unpin_extent_range(root, start, end);
8004 }
8005
8006 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8007                                u64 num_bytes, u64 *actual_bytes)
8008 {
8009         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8010 }
8011
8012 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8013 {
8014         struct btrfs_fs_info *fs_info = root->fs_info;
8015         struct btrfs_block_group_cache *cache = NULL;
8016         u64 group_trimmed;
8017         u64 start;
8018         u64 end;
8019         u64 trimmed = 0;
8020         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8021         int ret = 0;
8022
8023         /*
8024          * try to trim all FS space, our block group may start from non-zero.
8025          */
8026         if (range->len == total_bytes)
8027                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8028         else
8029                 cache = btrfs_lookup_block_group(fs_info, range->start);
8030
8031         while (cache) {
8032                 if (cache->key.objectid >= (range->start + range->len)) {
8033                         btrfs_put_block_group(cache);
8034                         break;
8035                 }
8036
8037                 start = max(range->start, cache->key.objectid);
8038                 end = min(range->start + range->len,
8039                                 cache->key.objectid + cache->key.offset);
8040
8041                 if (end - start >= range->minlen) {
8042                         if (!block_group_cache_done(cache)) {
8043                                 ret = cache_block_group(cache, NULL, root, 0);
8044                                 if (!ret)
8045                                         wait_block_group_cache_done(cache);
8046                         }
8047                         ret = btrfs_trim_block_group(cache,
8048                                                      &group_trimmed,
8049                                                      start,
8050                                                      end,
8051                                                      range->minlen);
8052
8053                         trimmed += group_trimmed;
8054                         if (ret) {
8055                                 btrfs_put_block_group(cache);
8056                                 break;
8057                         }
8058                 }
8059
8060                 cache = next_block_group(fs_info->tree_root, cache);
8061         }
8062
8063         range->len = trimmed;
8064         return ret;
8065 }