fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include "compat.h"
  28 #include "hash.h"
  29 #include "ctree.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "transaction.h"
  33 #include "volumes.h"
  34 #include "locking.h"
  35 #include "free-space-cache.h"
  36
  37 /*
  38  * control flags for do_chunk_alloc's force field
  39  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  40  * if we really need one.
  41  *
  42  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  43  * if we have very few chunks already allocated.  This is
  44  * used as part of the clustering code to help make sure
  45  * we have a good pool of storage to cluster in, without
  46  * filling the FS with empty chunks
  47  *
  48  * CHUNK_ALLOC_FORCE means it must try to allocate one
  49  *
  50  */
  51 enum {
  52         CHUNK_ALLOC_NO_FORCE = 0,
  53         CHUNK_ALLOC_LIMITED = 1,
  54         CHUNK_ALLOC_FORCE = 2,
  55 };
  56
  57 /*
  58  * Control how reservations are dealt with.
  59  *
  60  * RESERVE_FREE - freeing a reservation.
  61  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  62  *   ENOSPC accounting
  63  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  64  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  65  */
  66 enum {
  67         RESERVE_FREE = 0,
  68         RESERVE_ALLOC = 1,
  69         RESERVE_ALLOC_NO_ACCOUNT = 2,
  70 };
  71
  72 static int update_block_group(struct btrfs_trans_handle *trans,
  73                               struct btrfs_root *root,
  74                               u64 bytenr, u64 num_bytes, int alloc);
  75 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  76                                 struct btrfs_root *root,
  77                                 u64 bytenr, u64 num_bytes, u64 parent,
  78                                 u64 root_objectid, u64 owner_objectid,
  79                                 u64 owner_offset, int refs_to_drop,
  80                                 struct btrfs_delayed_extent_op *extra_op);
  81 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  82                                     struct extent_buffer *leaf,
  83                                     struct btrfs_extent_item *ei);
  84 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  85                                       struct btrfs_root *root,
  86                                       u64 parent, u64 root_objectid,
  87                                       u64 flags, u64 owner, u64 offset,
  88                                       struct btrfs_key *ins, int ref_mod);
  89 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  90                                      struct btrfs_root *root,
  91                                      u64 parent, u64 root_objectid,
  92                                      u64 flags, struct btrfs_disk_key *key,
  93                                      int level, struct btrfs_key *ins);
  94 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  95                           struct btrfs_root *extent_root, u64 alloc_bytes,
  96                           u64 flags, int force);
  97 static int find_next_key(struct btrfs_path *path, int level,
  98                          struct btrfs_key *key);
  99 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 100                             int dump_block_groups);
 101 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 102                                        u64 num_bytes, int reserve);
 103
 104 static noinline int
 105 block_group_cache_done(struct btrfs_block_group_cache *cache)
 106 {
 107         smp_mb();
 108         return cache->cached == BTRFS_CACHE_FINISHED;
 109 }
 110
 111 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 112 {
 113         return (cache->flags & bits) == bits;
 114 }
 115
 116 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 117 {
 118         atomic_inc(&cache->count);
 119 }
 120
 121 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 122 {
 123         if (atomic_dec_and_test(&cache->count)) {
 124                 WARN_ON(cache->pinned > 0);
 125                 WARN_ON(cache->reserved > 0);
 126                 kfree(cache->free_space_ctl);
 127                 kfree(cache);
 128         }
 129 }
 130
 131 /*
 132  * this adds the block group to the fs_info rb tree for the block group
 133  * cache
 134  */
 135 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 136                                 struct btrfs_block_group_cache *block_group)
 137 {
 138         struct rb_node **p;
 139         struct rb_node *parent = NULL;
 140         struct btrfs_block_group_cache *cache;
 141
 142         spin_lock(&info->block_group_cache_lock);
 143         p = &info->block_group_cache_tree.rb_node;
 144
 145         while (*p) {
 146                 parent = *p;
 147                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 148                                  cache_node);
 149                 if (block_group->key.objectid < cache->key.objectid) {
 150                         p = &(*p)->rb_left;
 151                 } else if (block_group->key.objectid > cache->key.objectid) {
 152                         p = &(*p)->rb_right;
 153                 } else {
 154                         spin_unlock(&info->block_group_cache_lock);
 155                         return -EEXIST;
 156                 }
 157         }
 158
 159         rb_link_node(&block_group->cache_node, parent, p);
 160         rb_insert_color(&block_group->cache_node,
 161                         &info->block_group_cache_tree);
 162         spin_unlock(&info->block_group_cache_lock);
 163
 164         return 0;
 165 }
 166
 167 /*
 168  * This will return the block group at or after bytenr if contains is 0, else
 169  * it will return the block group that contains the bytenr
 170  */
 171 static struct btrfs_block_group_cache *
 172 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 173                               int contains)
 174 {
 175         struct btrfs_block_group_cache *cache, *ret = NULL;
 176         struct rb_node *n;
 177         u64 end, start;
 178
 179         spin_lock(&info->block_group_cache_lock);
 180         n = info->block_group_cache_tree.rb_node;
 181
 182         while (n) {
 183                 cache = rb_entry(n, struct btrfs_block_group_cache,
 184                                  cache_node);
 185                 end = cache->key.objectid + cache->key.offset - 1;
 186                 start = cache->key.objectid;
 187
 188                 if (bytenr < start) {
 189                         if (!contains && (!ret || start < ret->key.objectid))
 190                                 ret = cache;
 191                         n = n->rb_left;
 192                 } else if (bytenr > start) {
 193                         if (contains && bytenr <= end) {
 194                                 ret = cache;
 195                                 break;
 196                         }
 197                         n = n->rb_right;
 198                 } else {
 199                         ret = cache;
 200                         break;
 201                 }
 202         }
 203         if (ret)
 204                 btrfs_get_block_group(ret);
 205         spin_unlock(&info->block_group_cache_lock);
 206
 207         return ret;
 208 }
 209
 210 static int add_excluded_extent(struct btrfs_root *root,
 211                                u64 start, u64 num_bytes)
 212 {
 213         u64 end = start + num_bytes - 1;
 214         set_extent_bits(&root->fs_info->freed_extents[0],
 215                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 216         set_extent_bits(&root->fs_info->freed_extents[1],
 217                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 218         return 0;
 219 }
 220
 221 static void free_excluded_extents(struct btrfs_root *root,
 222                                   struct btrfs_block_group_cache *cache)
 223 {
 224         u64 start, end;
 225
 226         start = cache->key.objectid;
 227         end = start + cache->key.offset - 1;
 228
 229         clear_extent_bits(&root->fs_info->freed_extents[0],
 230                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 231         clear_extent_bits(&root->fs_info->freed_extents[1],
 232                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 233 }
 234
 235 static int exclude_super_stripes(struct btrfs_root *root,
 236                                  struct btrfs_block_group_cache *cache)
 237 {
 238         u64 bytenr;
 239         u64 *logical;
 240         int stripe_len;
 241         int i, nr, ret;
 242
 243         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 244                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 245                 cache->bytes_super += stripe_len;
 246                 ret = add_excluded_extent(root, cache->key.objectid,
 247                                           stripe_len);
 248                 BUG_ON(ret); /* -ENOMEM */
 249         }
 250
 251         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 252                 bytenr = btrfs_sb_offset(i);
 253                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 254                                        cache->key.objectid, bytenr,
 255                                        0, &logical, &nr, &stripe_len);
 256                 BUG_ON(ret); /* -ENOMEM */
 257
 258                 while (nr--) {
 259                         cache->bytes_super += stripe_len;
 260                         ret = add_excluded_extent(root, logical[nr],
 261                                                   stripe_len);
 262                         BUG_ON(ret); /* -ENOMEM */
 263                 }
 264
 265                 kfree(logical);
 266         }
 267         return 0;
 268 }
 269
 270 static struct btrfs_caching_control *
 271 get_caching_control(struct btrfs_block_group_cache *cache)
 272 {
 273         struct btrfs_caching_control *ctl;
 274
 275         spin_lock(&cache->lock);
 276         if (cache->cached != BTRFS_CACHE_STARTED) {
 277                 spin_unlock(&cache->lock);
 278                 return NULL;
 279         }
 280
 281         /* We're loading it the fast way, so we don't have a caching_ctl. */
 282         if (!cache->caching_ctl) {
 283                 spin_unlock(&cache->lock);
 284                 return NULL;
 285         }
 286
 287         ctl = cache->caching_ctl;
 288         atomic_inc(&ctl->count);
 289         spin_unlock(&cache->lock);
 290         return ctl;
 291 }
 292
 293 static void put_caching_control(struct btrfs_caching_control *ctl)
 294 {
 295         if (atomic_dec_and_test(&ctl->count))
 296                 kfree(ctl);
 297 }
 298
 299 /*
 300  * this is only called by cache_block_group, since we could have freed extents
 301  * we need to check the pinned_extents for any extents that can't be used yet
 302  * since their free space will be released as soon as the transaction commits.
 303  */
 304 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 305                               struct btrfs_fs_info *info, u64 start, u64 end)
 306 {
 307         u64 extent_start, extent_end, size, total_added = 0;
 308         int ret;
 309
 310         while (start < end) {
 311                 ret = find_first_extent_bit(info->pinned_extents, start,
 312                                             &extent_start, &extent_end,
 313                                             EXTENT_DIRTY | EXTENT_UPTODATE);
 314                 if (ret)
 315                         break;
 316
 317                 if (extent_start <= start) {
 318                         start = extent_end + 1;
 319                 } else if (extent_start > start && extent_start < end) {
 320                         size = extent_start - start;
 321                         total_added += size;
 322                         ret = btrfs_add_free_space(block_group, start,
 323                                                    size);
 324                         BUG_ON(ret); /* -ENOMEM or logic error */
 325                         start = extent_end + 1;
 326                 } else {
 327                         break;
 328                 }
 329         }
 330
 331         if (start < end) {
 332                 size = end - start;
 333                 total_added += size;
 334                 ret = btrfs_add_free_space(block_group, start, size);
 335                 BUG_ON(ret); /* -ENOMEM or logic error */
 336         }
 337
 338         return total_added;
 339 }
 340
 341 static noinline void caching_thread(struct btrfs_work *work)
 342 {
 343         struct btrfs_block_group_cache *block_group;
 344         struct btrfs_fs_info *fs_info;
 345         struct btrfs_caching_control *caching_ctl;
 346         struct btrfs_root *extent_root;
 347         struct btrfs_path *path;
 348         struct extent_buffer *leaf;
 349         struct btrfs_key key;
 350         u64 total_found = 0;
 351         u64 last = 0;
 352         u32 nritems;
 353         int ret = 0;
 354
 355         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 356         block_group = caching_ctl->block_group;
 357         fs_info = block_group->fs_info;
 358         extent_root = fs_info->extent_root;
 359
 360         path = btrfs_alloc_path();
 361         if (!path)
 362                 goto out;
 363
 364         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 365
 366         /*
 367          * We don't want to deadlock with somebody trying to allocate a new
 368          * extent for the extent root while also trying to search the extent
 369          * root to add free space.  So we skip locking and search the commit
 370          * root, since its read-only
 371          */
 372         path->skip_locking = 1;
 373         path->search_commit_root = 1;
 374         path->reada = 1;
 375
 376         key.objectid = last;
 377         key.offset = 0;
 378         key.type = BTRFS_EXTENT_ITEM_KEY;
 379 again:
 380         mutex_lock(&caching_ctl->mutex);
 381         /* need to make sure the commit_root doesn't disappear */
 382         down_read(&fs_info->extent_commit_sem);
 383
 384         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 385         if (ret < 0)
 386                 goto err;
 387
 388         leaf = path->nodes[0];
 389         nritems = btrfs_header_nritems(leaf);
 390
 391         while (1) {
 392                 if (btrfs_fs_closing(fs_info) > 1) {
 393                         last = (u64)-1;
 394                         break;
 395                 }
 396
 397                 if (path->slots[0] < nritems) {
 398                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 399                 } else {
 400                         ret = find_next_key(path, 0, &key);
 401                         if (ret)
 402                                 break;
 403
 404                         if (need_resched() ||
 405                             btrfs_next_leaf(extent_root, path)) {
 406                                 caching_ctl->progress = last;
 407                                 btrfs_release_path(path);
 408                                 up_read(&fs_info->extent_commit_sem);
 409                                 mutex_unlock(&caching_ctl->mutex);
 410                                 cond_resched();
 411                                 goto again;
 412                         }
 413                         leaf = path->nodes[0];
 414                         nritems = btrfs_header_nritems(leaf);
 415                         continue;
 416                 }
 417
 418                 if (key.objectid < block_group->key.objectid) {
 419                         path->slots[0]++;
 420                         continue;
 421                 }
 422
 423                 if (key.objectid >= block_group->key.objectid +
 424                     block_group->key.offset)
 425                         break;
 426
 427                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 428                         total_found += add_new_free_space(block_group,
 429                                                           fs_info, last,
 430                                                           key.objectid);
 431                         last = key.objectid + key.offset;
 432
 433                         if (total_found > (1024 * 1024 * 2)) {
 434                                 total_found = 0;
 435                                 wake_up(&caching_ctl->wait);
 436                         }
 437                 }
 438                 path->slots[0]++;
 439         }
 440         ret = 0;
 441
 442         total_found += add_new_free_space(block_group, fs_info, last,
 443                                           block_group->key.objectid +
 444                                           block_group->key.offset);
 445         caching_ctl->progress = (u64)-1;
 446
 447         spin_lock(&block_group->lock);
 448         block_group->caching_ctl = NULL;
 449         block_group->cached = BTRFS_CACHE_FINISHED;
 450         spin_unlock(&block_group->lock);
 451
 452 err:
 453         btrfs_free_path(path);
 454         up_read(&fs_info->extent_commit_sem);
 455
 456         free_excluded_extents(extent_root, block_group);
 457
 458         mutex_unlock(&caching_ctl->mutex);
 459 out:
 460         wake_up(&caching_ctl->wait);
 461
 462         put_caching_control(caching_ctl);
 463         btrfs_put_block_group(block_group);
 464 }
 465
 466 static int cache_block_group(struct btrfs_block_group_cache *cache,
 467                              struct btrfs_trans_handle *trans,
 468                              struct btrfs_root *root,
 469                              int load_cache_only)
 470 {
 471         DEFINE_WAIT(wait);
 472         struct btrfs_fs_info *fs_info = cache->fs_info;
 473         struct btrfs_caching_control *caching_ctl;
 474         int ret = 0;
 475
 476         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 477         if (!caching_ctl)
 478                 return -ENOMEM;
 479
 480         INIT_LIST_HEAD(&caching_ctl->list);
 481         mutex_init(&caching_ctl->mutex);
 482         init_waitqueue_head(&caching_ctl->wait);
 483         caching_ctl->block_group = cache;
 484         caching_ctl->progress = cache->key.objectid;
 485         atomic_set(&caching_ctl->count, 1);
 486         caching_ctl->work.func = caching_thread;
 487
 488         spin_lock(&cache->lock);
 489         /*
 490          * This should be a rare occasion, but this could happen I think in the
 491          * case where one thread starts to load the space cache info, and then
 492          * some other thread starts a transaction commit which tries to do an
 493          * allocation while the other thread is still loading the space cache
 494          * info.  The previous loop should have kept us from choosing this block
 495          * group, but if we've moved to the state where we will wait on caching
 496          * block groups we need to first check if we're doing a fast load here,
 497          * so we can wait for it to finish, otherwise we could end up allocating
 498          * from a block group who's cache gets evicted for one reason or
 499          * another.
 500          */
 501         while (cache->cached == BTRFS_CACHE_FAST) {
 502                 struct btrfs_caching_control *ctl;
 503
 504                 ctl = cache->caching_ctl;
 505                 atomic_inc(&ctl->count);
 506                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 507                 spin_unlock(&cache->lock);
 508
 509                 schedule();
 510
 511                 finish_wait(&ctl->wait, &wait);
 512                 put_caching_control(ctl);
 513                 spin_lock(&cache->lock);
 514         }
 515
 516         if (cache->cached != BTRFS_CACHE_NO) {
 517                 spin_unlock(&cache->lock);
 518                 kfree(caching_ctl);
 519                 return 0;
 520         }
 521         WARN_ON(cache->caching_ctl);
 522         cache->caching_ctl = caching_ctl;
 523         cache->cached = BTRFS_CACHE_FAST;
 524         spin_unlock(&cache->lock);
 525
 526         /*
 527          * We can't do the read from on-disk cache during a commit since we need
 528          * to have the normal tree locking.  Also if we are currently trying to
 529          * allocate blocks for the tree root we can't do the fast caching since
 530          * we likely hold important locks.
 531          */
 532         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 533                 ret = load_free_space_cache(fs_info, cache);
 534
 535                 spin_lock(&cache->lock);
 536                 if (ret == 1) {
 537                         cache->caching_ctl = NULL;
 538                         cache->cached = BTRFS_CACHE_FINISHED;
 539                         cache->last_byte_to_unpin = (u64)-1;
 540                 } else {
 541                         if (load_cache_only) {
 542                                 cache->caching_ctl = NULL;
 543                                 cache->cached = BTRFS_CACHE_NO;
 544                         } else {
 545                                 cache->cached = BTRFS_CACHE_STARTED;
 546                         }
 547                 }
 548                 spin_unlock(&cache->lock);
 549                 wake_up(&caching_ctl->wait);
 550                 if (ret == 1) {
 551                         put_caching_control(caching_ctl);
 552                         free_excluded_extents(fs_info->extent_root, cache);
 553                         return 0;
 554                 }
 555         } else {
 556                 /*
 557                  * We are not going to do the fast caching, set cached to the
 558                  * appropriate value and wakeup any waiters.
 559                  */
 560                 spin_lock(&cache->lock);
 561                 if (load_cache_only) {
 562                         cache->caching_ctl = NULL;
 563                         cache->cached = BTRFS_CACHE_NO;
 564                 } else {
 565                         cache->cached = BTRFS_CACHE_STARTED;
 566                 }
 567                 spin_unlock(&cache->lock);
 568                 wake_up(&caching_ctl->wait);
 569         }
 570
 571         if (load_cache_only) {
 572                 put_caching_control(caching_ctl);
 573                 return 0;
 574         }
 575
 576         down_write(&fs_info->extent_commit_sem);
 577         atomic_inc(&caching_ctl->count);
 578         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 579         up_write(&fs_info->extent_commit_sem);
 580
 581         btrfs_get_block_group(cache);
 582
 583         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 584
 585         return ret;
 586 }
 587
 588 /*
 589  * return the block group that starts at or after bytenr
 590  */
 591 static struct btrfs_block_group_cache *
 592 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 593 {
 594         struct btrfs_block_group_cache *cache;
 595
 596         cache = block_group_cache_tree_search(info, bytenr, 0);
 597
 598         return cache;
 599 }
 600
 601 /*
 602  * return the block group that contains the given bytenr
 603  */
 604 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 605                                                  struct btrfs_fs_info *info,
 606                                                  u64 bytenr)
 607 {
 608         struct btrfs_block_group_cache *cache;
 609
 610         cache = block_group_cache_tree_search(info, bytenr, 1);
 611
 612         return cache;
 613 }
 614
 615 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 616                                                   u64 flags)
 617 {
 618         struct list_head *head = &info->space_info;
 619         struct btrfs_space_info *found;
 620
 621         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 622
 623         rcu_read_lock();
 624         list_for_each_entry_rcu(found, head, list) {
 625                 if (found->flags & flags) {
 626                         rcu_read_unlock();
 627                         return found;
 628                 }
 629         }
 630         rcu_read_unlock();
 631         return NULL;
 632 }
 633
 634 /*
 635  * after adding space to the filesystem, we need to clear the full flags
 636  * on all the space infos.
 637  */
 638 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 639 {
 640         struct list_head *head = &info->space_info;
 641         struct btrfs_space_info *found;
 642
 643         rcu_read_lock();
 644         list_for_each_entry_rcu(found, head, list)
 645                 found->full = 0;
 646         rcu_read_unlock();
 647 }
 648
 649 static u64 div_factor(u64 num, int factor)
 650 {
 651         if (factor == 10)
 652                 return num;
 653         num *= factor;
 654         do_div(num, 10);
 655         return num;
 656 }
 657
 658 static u64 div_factor_fine(u64 num, int factor)
 659 {
 660         if (factor == 100)
 661                 return num;
 662         num *= factor;
 663         do_div(num, 100);
 664         return num;
 665 }
 666
 667 u64 btrfs_find_block_group(struct btrfs_root *root,
 668                            u64 search_start, u64 search_hint, int owner)
 669 {
 670         struct btrfs_block_group_cache *cache;
 671         u64 used;
 672         u64 last = max(search_hint, search_start);
 673         u64 group_start = 0;
 674         int full_search = 0;
 675         int factor = 9;
 676         int wrapped = 0;
 677 again:
 678         while (1) {
 679                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
 680                 if (!cache)
 681                         break;
 682
 683                 spin_lock(&cache->lock);
 684                 last = cache->key.objectid + cache->key.offset;
 685                 used = btrfs_block_group_used(&cache->item);
 686
 687                 if ((full_search || !cache->ro) &&
 688                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
 689                         if (used + cache->pinned + cache->reserved <
 690                             div_factor(cache->key.offset, factor)) {
 691                                 group_start = cache->key.objectid;
 692                                 spin_unlock(&cache->lock);
 693                                 btrfs_put_block_group(cache);
 694                                 goto found;
 695                         }
 696                 }
 697                 spin_unlock(&cache->lock);
 698                 btrfs_put_block_group(cache);
 699                 cond_resched();
 700         }
 701         if (!wrapped) {
 702                 last = search_start;
 703                 wrapped = 1;
 704                 goto again;
 705         }
 706         if (!full_search && factor < 10) {
 707                 last = search_start;
 708                 full_search = 1;
 709                 factor = 10;
 710                 goto again;
 711         }
 712 found:
 713         return group_start;
 714 }
 715
 716 /* simple helper to search for an existing extent at a given offset */
 717 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 718 {
 719         int ret;
 720         struct btrfs_key key;
 721         struct btrfs_path *path;
 722
 723         path = btrfs_alloc_path();
 724         if (!path)
 725                 return -ENOMEM;
 726
 727         key.objectid = start;
 728         key.offset = len;
 729         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 730         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 731                                 0, 0);
 732         btrfs_free_path(path);
 733         return ret;
 734 }
 735
 736 /*
 737  * helper function to lookup reference count and flags of extent.
 738  *
 739  * the head node for delayed ref is used to store the sum of all the
 740  * reference count modifications queued up in the rbtree. the head
 741  * node may also store the extent flags to set. This way you can check
 742  * to see what the reference count and extent flags would be if all of
 743  * the delayed refs are not processed.
 744  */
 745 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 746                              struct btrfs_root *root, u64 bytenr,
 747                              u64 num_bytes, u64 *refs, u64 *flags)
 748 {
 749         struct btrfs_delayed_ref_head *head;
 750         struct btrfs_delayed_ref_root *delayed_refs;
 751         struct btrfs_path *path;
 752         struct btrfs_extent_item *ei;
 753         struct extent_buffer *leaf;
 754         struct btrfs_key key;
 755         u32 item_size;
 756         u64 num_refs;
 757         u64 extent_flags;
 758         int ret;
 759
 760         path = btrfs_alloc_path();
 761         if (!path)
 762                 return -ENOMEM;
 763
 764         key.objectid = bytenr;
 765         key.type = BTRFS_EXTENT_ITEM_KEY;
 766         key.offset = num_bytes;
 767         if (!trans) {
 768                 path->skip_locking = 1;
 769                 path->search_commit_root = 1;
 770         }
 771 again:
 772         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 773                                 &key, path, 0, 0);
 774         if (ret < 0)
 775                 goto out_free;
 776
 777         if (ret == 0) {
 778                 leaf = path->nodes[0];
 779                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 780                 if (item_size >= sizeof(*ei)) {
 781                         ei = btrfs_item_ptr(leaf, path->slots[0],
 782                                             struct btrfs_extent_item);
 783                         num_refs = btrfs_extent_refs(leaf, ei);
 784                         extent_flags = btrfs_extent_flags(leaf, ei);
 785                 } else {
 786 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 787                         struct btrfs_extent_item_v0 *ei0;
 788                         BUG_ON(item_size != sizeof(*ei0));
 789                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 790                                              struct btrfs_extent_item_v0);
 791                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 792                         /* FIXME: this isn't correct for data */
 793                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 794 #else
 795                         BUG();
 796 #endif
 797                 }
 798                 BUG_ON(num_refs == 0);
 799         } else {
 800                 num_refs = 0;
 801                 extent_flags = 0;
 802                 ret = 0;
 803         }
 804
 805         if (!trans)
 806                 goto out;
 807
 808         delayed_refs = &trans->transaction->delayed_refs;
 809         spin_lock(&delayed_refs->lock);
 810         head = btrfs_find_delayed_ref_head(trans, bytenr);
 811         if (head) {
 812                 if (!mutex_trylock(&head->mutex)) {
 813                         atomic_inc(&head->node.refs);
 814                         spin_unlock(&delayed_refs->lock);
 815
 816                         btrfs_release_path(path);
 817
 818                         /*
 819                          * Mutex was contended, block until it's released and try
 820                          * again
 821                          */
 822                         mutex_lock(&head->mutex);
 823                         mutex_unlock(&head->mutex);
 824                         btrfs_put_delayed_ref(&head->node);
 825                         goto again;
 826                 }
 827                 if (head->extent_op && head->extent_op->update_flags)
 828                         extent_flags |= head->extent_op->flags_to_set;
 829                 else
 830                         BUG_ON(num_refs == 0);
 831
 832                 num_refs += head->node.ref_mod;
 833                 mutex_unlock(&head->mutex);
 834         }
 835         spin_unlock(&delayed_refs->lock);
 836 out:
 837         WARN_ON(num_refs == 0);
 838         if (refs)
 839                 *refs = num_refs;
 840         if (flags)
 841                 *flags = extent_flags;
 842 out_free:
 843         btrfs_free_path(path);
 844         return ret;
 845 }
 846
 847 /*
 848  * Back reference rules.  Back refs have three main goals:
 849  *
 850  * 1) differentiate between all holders of references to an extent so that
 851  *    when a reference is dropped we can make sure it was a valid reference
 852  *    before freeing the extent.
 853  *
 854  * 2) Provide enough information to quickly find the holders of an extent
 855  *    if we notice a given block is corrupted or bad.
 856  *
 857  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 858  *    maintenance.  This is actually the same as #2, but with a slightly
 859  *    different use case.
 860  *
 861  * There are two kinds of back refs. The implicit back refs is optimized
 862  * for pointers in non-shared tree blocks. For a given pointer in a block,
 863  * back refs of this kind provide information about the block's owner tree
 864  * and the pointer's key. These information allow us to find the block by
 865  * b-tree searching. The full back refs is for pointers in tree blocks not
 866  * referenced by their owner trees. The location of tree block is recorded
 867  * in the back refs. Actually the full back refs is generic, and can be
 868  * used in all cases the implicit back refs is used. The major shortcoming
 869  * of the full back refs is its overhead. Every time a tree block gets
 870  * COWed, we have to update back refs entry for all pointers in it.
 871  *
 872  * For a newly allocated tree block, we use implicit back refs for
 873  * pointers in it. This means most tree related operations only involve
 874  * implicit back refs. For a tree block created in old transaction, the
 875  * only way to drop a reference to it is COW it. So we can detect the
 876  * event that tree block loses its owner tree's reference and do the
 877  * back refs conversion.
 878  *
 879  * When a tree block is COW'd through a tree, there are four cases:
 880  *
 881  * The reference count of the block is one and the tree is the block's
 882  * owner tree. Nothing to do in this case.
 883  *
 884  * The reference count of the block is one and the tree is not the
 885  * block's owner tree. In this case, full back refs is used for pointers
 886  * in the block. Remove these full back refs, add implicit back refs for
 887  * every pointers in the new block.
 888  *
 889  * The reference count of the block is greater than one and the tree is
 890  * the block's owner tree. In this case, implicit back refs is used for
 891  * pointers in the block. Add full back refs for every pointers in the
 892  * block, increase lower level extents' reference counts. The original
 893  * implicit back refs are entailed to the new block.
 894  *
 895  * The reference count of the block is greater than one and the tree is
 896  * not the block's owner tree. Add implicit back refs for every pointer in
 897  * the new block, increase lower level extents' reference count.
 898  *
 899  * Back Reference Key composing:
 900  *
 901  * The key objectid corresponds to the first byte in the extent,
 902  * The key type is used to differentiate between types of back refs.
 903  * There are different meanings of the key offset for different types
 904  * of back refs.
 905  *
 906  * File extents can be referenced by:
 907  *
 908  * - multiple snapshots, subvolumes, or different generations in one subvol
 909  * - different files inside a single subvolume
 910  * - different offsets inside a file (bookend extents in file.c)
 911  *
 912  * The extent ref structure for the implicit back refs has fields for:
 913  *
 914  * - Objectid of the subvolume root
 915  * - objectid of the file holding the reference
 916  * - original offset in the file
 917  * - how many bookend extents
 918  *
 919  * The key offset for the implicit back refs is hash of the first
 920  * three fields.
 921  *
 922  * The extent ref structure for the full back refs has field for:
 923  *
 924  * - number of pointers in the tree leaf
 925  *
 926  * The key offset for the implicit back refs is the first byte of
 927  * the tree leaf
 928  *
 929  * When a file extent is allocated, The implicit back refs is used.
 930  * the fields are filled in:
 931  *
 932  *     (root_key.objectid, inode objectid, offset in file, 1)
 933  *
 934  * When a file extent is removed file truncation, we find the
 935  * corresponding implicit back refs and check the following fields:
 936  *
 937  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 938  *
 939  * Btree extents can be referenced by:
 940  *
 941  * - Different subvolumes
 942  *
 943  * Both the implicit back refs and the full back refs for tree blocks
 944  * only consist of key. The key offset for the implicit back refs is
 945  * objectid of block's owner tree. The key offset for the full back refs
 946  * is the first byte of parent block.
 947  *
 948  * When implicit back refs is used, information about the lowest key and
 949  * level of the tree block are required. These information are stored in
 950  * tree block info structure.
 951  */
 952
 953 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 954 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 955                                   struct btrfs_root *root,
 956                                   struct btrfs_path *path,
 957                                   u64 owner, u32 extra_size)
 958 {
 959         struct btrfs_extent_item *item;
 960         struct btrfs_extent_item_v0 *ei0;
 961         struct btrfs_extent_ref_v0 *ref0;
 962         struct btrfs_tree_block_info *bi;
 963         struct extent_buffer *leaf;
 964         struct btrfs_key key;
 965         struct btrfs_key found_key;
 966         u32 new_size = sizeof(*item);
 967         u64 refs;
 968         int ret;
 969
 970         leaf = path->nodes[0];
 971         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 972
 973         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 974         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 975                              struct btrfs_extent_item_v0);
 976         refs = btrfs_extent_refs_v0(leaf, ei0);
 977
 978         if (owner == (u64)-1) {
 979                 while (1) {
 980                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 981                                 ret = btrfs_next_leaf(root, path);
 982                                 if (ret < 0)
 983                                         return ret;
 984                                 BUG_ON(ret > 0); /* Corruption */
 985                                 leaf = path->nodes[0];
 986                         }
 987                         btrfs_item_key_to_cpu(leaf, &found_key,
 988                                               path->slots[0]);
 989                         BUG_ON(key.objectid != found_key.objectid);
 990                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 991                                 path->slots[0]++;
 992                                 continue;
 993                         }
 994                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 995                                               struct btrfs_extent_ref_v0);
 996                         owner = btrfs_ref_objectid_v0(leaf, ref0);
 997                         break;
 998                 }
 999         }
1000         btrfs_release_path(path);
1001
1002         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1003                 new_size += sizeof(*bi);
1004
1005         new_size -= sizeof(*ei0);
1006         ret = btrfs_search_slot(trans, root, &key, path,
1007                                 new_size + extra_size, 1);
1008         if (ret < 0)
1009                 return ret;
1010         BUG_ON(ret); /* Corruption */
1011
1012         btrfs_extend_item(trans, root, path, new_size);
1013
1014         leaf = path->nodes[0];
1015         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1016         btrfs_set_extent_refs(leaf, item, refs);
1017         /* FIXME: get real generation */
1018         btrfs_set_extent_generation(leaf, item, 0);
1019         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1020                 btrfs_set_extent_flags(leaf, item,
1021                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1022                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1023                 bi = (struct btrfs_tree_block_info *)(item + 1);
1024                 /* FIXME: get first key of the block */
1025                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1026                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1027         } else {
1028                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1029         }
1030         btrfs_mark_buffer_dirty(leaf);
1031         return 0;
1032 }
1033 #endif
1034
1035 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1036 {
1037         u32 high_crc = ~(u32)0;
1038         u32 low_crc = ~(u32)0;
1039         __le64 lenum;
1040
1041         lenum = cpu_to_le64(root_objectid);
1042         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1043         lenum = cpu_to_le64(owner);
1044         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1045         lenum = cpu_to_le64(offset);
1046         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1047
1048         return ((u64)high_crc << 31) ^ (u64)low_crc;
1049 }
1050
1051 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1052                                      struct btrfs_extent_data_ref *ref)
1053 {
1054         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1055                                     btrfs_extent_data_ref_objectid(leaf, ref),
1056                                     btrfs_extent_data_ref_offset(leaf, ref));
1057 }
1058
1059 static int match_extent_data_ref(struct extent_buffer *leaf,
1060                                  struct btrfs_extent_data_ref *ref,
1061                                  u64 root_objectid, u64 owner, u64 offset)
1062 {
1063         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1064             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1065             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1066                 return 0;
1067         return 1;
1068 }
1069
1070 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1071                                            struct btrfs_root *root,
1072                                            struct btrfs_path *path,
1073                                            u64 bytenr, u64 parent,
1074                                            u64 root_objectid,
1075                                            u64 owner, u64 offset)
1076 {
1077         struct btrfs_key key;
1078         struct btrfs_extent_data_ref *ref;
1079         struct extent_buffer *leaf;
1080         u32 nritems;
1081         int ret;
1082         int recow;
1083         int err = -ENOENT;
1084
1085         key.objectid = bytenr;
1086         if (parent) {
1087                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1088                 key.offset = parent;
1089         } else {
1090                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1091                 key.offset = hash_extent_data_ref(root_objectid,
1092                                                   owner, offset);
1093         }
1094 again:
1095         recow = 0;
1096         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1097         if (ret < 0) {
1098                 err = ret;
1099                 goto fail;
1100         }
1101
1102         if (parent) {
1103                 if (!ret)
1104                         return 0;
1105 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1106                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1107                 btrfs_release_path(path);
1108                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1109                 if (ret < 0) {
1110                         err = ret;
1111                         goto fail;
1112                 }
1113                 if (!ret)
1114                         return 0;
1115 #endif
1116                 goto fail;
1117         }
1118
1119         leaf = path->nodes[0];
1120         nritems = btrfs_header_nritems(leaf);
1121         while (1) {
1122                 if (path->slots[0] >= nritems) {
1123                         ret = btrfs_next_leaf(root, path);
1124                         if (ret < 0)
1125                                 err = ret;
1126                         if (ret)
1127                                 goto fail;
1128
1129                         leaf = path->nodes[0];
1130                         nritems = btrfs_header_nritems(leaf);
1131                         recow = 1;
1132                 }
1133
1134                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1135                 if (key.objectid != bytenr ||
1136                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1137                         goto fail;
1138
1139                 ref = btrfs_item_ptr(leaf, path->slots[0],
1140                                      struct btrfs_extent_data_ref);
1141
1142                 if (match_extent_data_ref(leaf, ref, root_objectid,
1143                                           owner, offset)) {
1144                         if (recow) {
1145                                 btrfs_release_path(path);
1146                                 goto again;
1147                         }
1148                         err = 0;
1149                         break;
1150                 }
1151                 path->slots[0]++;
1152         }
1153 fail:
1154         return err;
1155 }
1156
1157 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1158                                            struct btrfs_root *root,
1159                                            struct btrfs_path *path,
1160                                            u64 bytenr, u64 parent,
1161                                            u64 root_objectid, u64 owner,
1162                                            u64 offset, int refs_to_add)
1163 {
1164         struct btrfs_key key;
1165         struct extent_buffer *leaf;
1166         u32 size;
1167         u32 num_refs;
1168         int ret;
1169
1170         key.objectid = bytenr;
1171         if (parent) {
1172                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1173                 key.offset = parent;
1174                 size = sizeof(struct btrfs_shared_data_ref);
1175         } else {
1176                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1177                 key.offset = hash_extent_data_ref(root_objectid,
1178                                                   owner, offset);
1179                 size = sizeof(struct btrfs_extent_data_ref);
1180         }
1181
1182         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1183         if (ret && ret != -EEXIST)
1184                 goto fail;
1185
1186         leaf = path->nodes[0];
1187         if (parent) {
1188                 struct btrfs_shared_data_ref *ref;
1189                 ref = btrfs_item_ptr(leaf, path->slots[0],
1190                                      struct btrfs_shared_data_ref);
1191                 if (ret == 0) {
1192                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1193                 } else {
1194                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1195                         num_refs += refs_to_add;
1196                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1197                 }
1198         } else {
1199                 struct btrfs_extent_data_ref *ref;
1200                 while (ret == -EEXIST) {
1201                         ref = btrfs_item_ptr(leaf, path->slots[0],
1202                                              struct btrfs_extent_data_ref);
1203                         if (match_extent_data_ref(leaf, ref, root_objectid,
1204                                                   owner, offset))
1205                                 break;
1206                         btrfs_release_path(path);
1207                         key.offset++;
1208                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1209                                                       size);
1210                         if (ret && ret != -EEXIST)
1211                                 goto fail;
1212
1213                         leaf = path->nodes[0];
1214                 }
1215                 ref = btrfs_item_ptr(leaf, path->slots[0],
1216                                      struct btrfs_extent_data_ref);
1217                 if (ret == 0) {
1218                         btrfs_set_extent_data_ref_root(leaf, ref,
1219                                                        root_objectid);
1220                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1221                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1222                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1223                 } else {
1224                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1225                         num_refs += refs_to_add;
1226                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1227                 }
1228         }
1229         btrfs_mark_buffer_dirty(leaf);
1230         ret = 0;
1231 fail:
1232         btrfs_release_path(path);
1233         return ret;
1234 }
1235
1236 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1237                                            struct btrfs_root *root,
1238                                            struct btrfs_path *path,
1239                                            int refs_to_drop)
1240 {
1241         struct btrfs_key key;
1242         struct btrfs_extent_data_ref *ref1 = NULL;
1243         struct btrfs_shared_data_ref *ref2 = NULL;
1244         struct extent_buffer *leaf;
1245         u32 num_refs = 0;
1246         int ret = 0;
1247
1248         leaf = path->nodes[0];
1249         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1250
1251         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1252                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1253                                       struct btrfs_extent_data_ref);
1254                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1255         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1256                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1257                                       struct btrfs_shared_data_ref);
1258                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1259 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1260         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1261                 struct btrfs_extent_ref_v0 *ref0;
1262                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1263                                       struct btrfs_extent_ref_v0);
1264                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1265 #endif
1266         } else {
1267                 BUG();
1268         }
1269
1270         BUG_ON(num_refs < refs_to_drop);
1271         num_refs -= refs_to_drop;
1272
1273         if (num_refs == 0) {
1274                 ret = btrfs_del_item(trans, root, path);
1275         } else {
1276                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1277                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1278                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1279                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1280 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1281                 else {
1282                         struct btrfs_extent_ref_v0 *ref0;
1283                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1284                                         struct btrfs_extent_ref_v0);
1285                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1286                 }
1287 #endif
1288                 btrfs_mark_buffer_dirty(leaf);
1289         }
1290         return ret;
1291 }
1292
1293 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1294                                           struct btrfs_path *path,
1295                                           struct btrfs_extent_inline_ref *iref)
1296 {
1297         struct btrfs_key key;
1298         struct extent_buffer *leaf;
1299         struct btrfs_extent_data_ref *ref1;
1300         struct btrfs_shared_data_ref *ref2;
1301         u32 num_refs = 0;
1302
1303         leaf = path->nodes[0];
1304         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1305         if (iref) {
1306                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1307                     BTRFS_EXTENT_DATA_REF_KEY) {
1308                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1309                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1310                 } else {
1311                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1312                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1313                 }
1314         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1315                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1316                                       struct btrfs_extent_data_ref);
1317                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1318         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1319                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1320                                       struct btrfs_shared_data_ref);
1321                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1322 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1323         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1324                 struct btrfs_extent_ref_v0 *ref0;
1325                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1326                                       struct btrfs_extent_ref_v0);
1327                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1328 #endif
1329         } else {
1330                 WARN_ON(1);
1331         }
1332         return num_refs;
1333 }
1334
1335 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1336                                           struct btrfs_root *root,
1337                                           struct btrfs_path *path,
1338                                           u64 bytenr, u64 parent,
1339                                           u64 root_objectid)
1340 {
1341         struct btrfs_key key;
1342         int ret;
1343
1344         key.objectid = bytenr;
1345         if (parent) {
1346                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1347                 key.offset = parent;
1348         } else {
1349                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1350                 key.offset = root_objectid;
1351         }
1352
1353         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1354         if (ret > 0)
1355                 ret = -ENOENT;
1356 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1357         if (ret == -ENOENT && parent) {
1358                 btrfs_release_path(path);
1359                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1360                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1361                 if (ret > 0)
1362                         ret = -ENOENT;
1363         }
1364 #endif
1365         return ret;
1366 }
1367
1368 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1369                                           struct btrfs_root *root,
1370                                           struct btrfs_path *path,
1371                                           u64 bytenr, u64 parent,
1372                                           u64 root_objectid)
1373 {
1374         struct btrfs_key key;
1375         int ret;
1376
1377         key.objectid = bytenr;
1378         if (parent) {
1379                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1380                 key.offset = parent;
1381         } else {
1382                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1383                 key.offset = root_objectid;
1384         }
1385
1386         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1387         btrfs_release_path(path);
1388         return ret;
1389 }
1390
1391 static inline int extent_ref_type(u64 parent, u64 owner)
1392 {
1393         int type;
1394         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1395                 if (parent > 0)
1396                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1397                 else
1398                         type = BTRFS_TREE_BLOCK_REF_KEY;
1399         } else {
1400                 if (parent > 0)
1401                         type = BTRFS_SHARED_DATA_REF_KEY;
1402                 else
1403                         type = BTRFS_EXTENT_DATA_REF_KEY;
1404         }
1405         return type;
1406 }
1407
1408 static int find_next_key(struct btrfs_path *path, int level,
1409                          struct btrfs_key *key)
1410
1411 {
1412         for (; level < BTRFS_MAX_LEVEL; level++) {
1413                 if (!path->nodes[level])
1414                         break;
1415                 if (path->slots[level] + 1 >=
1416                     btrfs_header_nritems(path->nodes[level]))
1417                         continue;
1418                 if (level == 0)
1419                         btrfs_item_key_to_cpu(path->nodes[level], key,
1420                                               path->slots[level] + 1);
1421                 else
1422                         btrfs_node_key_to_cpu(path->nodes[level], key,
1423                                               path->slots[level] + 1);
1424                 return 0;
1425         }
1426         return 1;
1427 }
1428
1429 /*
1430  * look for inline back ref. if back ref is found, *ref_ret is set
1431  * to the address of inline back ref, and 0 is returned.
1432  *
1433  * if back ref isn't found, *ref_ret is set to the address where it
1434  * should be inserted, and -ENOENT is returned.
1435  *
1436  * if insert is true and there are too many inline back refs, the path
1437  * points to the extent item, and -EAGAIN is returned.
1438  *
1439  * NOTE: inline back refs are ordered in the same way that back ref
1440  *       items in the tree are ordered.
1441  */
1442 static noinline_for_stack
1443 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1444                                  struct btrfs_root *root,
1445                                  struct btrfs_path *path,
1446                                  struct btrfs_extent_inline_ref **ref_ret,
1447                                  u64 bytenr, u64 num_bytes,
1448                                  u64 parent, u64 root_objectid,
1449                                  u64 owner, u64 offset, int insert)
1450 {
1451         struct btrfs_key key;
1452         struct extent_buffer *leaf;
1453         struct btrfs_extent_item *ei;
1454         struct btrfs_extent_inline_ref *iref;
1455         u64 flags;
1456         u64 item_size;
1457         unsigned long ptr;
1458         unsigned long end;
1459         int extra_size;
1460         int type;
1461         int want;
1462         int ret;
1463         int err = 0;
1464
1465         key.objectid = bytenr;
1466         key.type = BTRFS_EXTENT_ITEM_KEY;
1467         key.offset = num_bytes;
1468
1469         want = extent_ref_type(parent, owner);
1470         if (insert) {
1471                 extra_size = btrfs_extent_inline_ref_size(want);
1472                 path->keep_locks = 1;
1473         } else
1474                 extra_size = -1;
1475         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1476         if (ret < 0) {
1477                 err = ret;
1478                 goto out;
1479         }
1480         if (ret && !insert) {
1481                 err = -ENOENT;
1482                 goto out;
1483         }
1484         BUG_ON(ret); /* Corruption */
1485
1486         leaf = path->nodes[0];
1487         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1488 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1489         if (item_size < sizeof(*ei)) {
1490                 if (!insert) {
1491                         err = -ENOENT;
1492                         goto out;
1493                 }
1494                 ret = convert_extent_item_v0(trans, root, path, owner,
1495                                              extra_size);
1496                 if (ret < 0) {
1497                         err = ret;
1498                         goto out;
1499                 }
1500                 leaf = path->nodes[0];
1501                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1502         }
1503 #endif
1504         BUG_ON(item_size < sizeof(*ei));
1505
1506         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1507         flags = btrfs_extent_flags(leaf, ei);
1508
1509         ptr = (unsigned long)(ei + 1);
1510         end = (unsigned long)ei + item_size;
1511
1512         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1513                 ptr += sizeof(struct btrfs_tree_block_info);
1514                 BUG_ON(ptr > end);
1515         } else {
1516                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1517         }
1518
1519         err = -ENOENT;
1520         while (1) {
1521                 if (ptr >= end) {
1522                         WARN_ON(ptr > end);
1523                         break;
1524                 }
1525                 iref = (struct btrfs_extent_inline_ref *)ptr;
1526                 type = btrfs_extent_inline_ref_type(leaf, iref);
1527                 if (want < type)
1528                         break;
1529                 if (want > type) {
1530                         ptr += btrfs_extent_inline_ref_size(type);
1531                         continue;
1532                 }
1533
1534                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1535                         struct btrfs_extent_data_ref *dref;
1536                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1537                         if (match_extent_data_ref(leaf, dref, root_objectid,
1538                                                   owner, offset)) {
1539                                 err = 0;
1540                                 break;
1541                         }
1542                         if (hash_extent_data_ref_item(leaf, dref) <
1543                             hash_extent_data_ref(root_objectid, owner, offset))
1544                                 break;
1545                 } else {
1546                         u64 ref_offset;
1547                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1548                         if (parent > 0) {
1549                                 if (parent == ref_offset) {
1550                                         err = 0;
1551                                         break;
1552                                 }
1553                                 if (ref_offset < parent)
1554                                         break;
1555                         } else {
1556                                 if (root_objectid == ref_offset) {
1557                                         err = 0;
1558                                         break;
1559                                 }
1560                                 if (ref_offset < root_objectid)
1561                                         break;
1562                         }
1563                 }
1564                 ptr += btrfs_extent_inline_ref_size(type);
1565         }
1566         if (err == -ENOENT && insert) {
1567                 if (item_size + extra_size >=
1568                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1569                         err = -EAGAIN;
1570                         goto out;
1571                 }
1572                 /*
1573                  * To add new inline back ref, we have to make sure
1574                  * there is no corresponding back ref item.
1575                  * For simplicity, we just do not add new inline back
1576                  * ref if there is any kind of item for this block
1577                  */
1578                 if (find_next_key(path, 0, &key) == 0 &&
1579                     key.objectid == bytenr &&
1580                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1581                         err = -EAGAIN;
1582                         goto out;
1583                 }
1584         }
1585         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1586 out:
1587         if (insert) {
1588                 path->keep_locks = 0;
1589                 btrfs_unlock_up_safe(path, 1);
1590         }
1591         return err;
1592 }
1593
1594 /*
1595  * helper to add new inline back ref
1596  */
1597 static noinline_for_stack
1598 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1599                                  struct btrfs_root *root,
1600                                  struct btrfs_path *path,
1601                                  struct btrfs_extent_inline_ref *iref,
1602                                  u64 parent, u64 root_objectid,
1603                                  u64 owner, u64 offset, int refs_to_add,
1604                                  struct btrfs_delayed_extent_op *extent_op)
1605 {
1606         struct extent_buffer *leaf;
1607         struct btrfs_extent_item *ei;
1608         unsigned long ptr;
1609         unsigned long end;
1610         unsigned long item_offset;
1611         u64 refs;
1612         int size;
1613         int type;
1614
1615         leaf = path->nodes[0];
1616         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1617         item_offset = (unsigned long)iref - (unsigned long)ei;
1618
1619         type = extent_ref_type(parent, owner);
1620         size = btrfs_extent_inline_ref_size(type);
1621
1622         btrfs_extend_item(trans, root, path, size);
1623
1624         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1625         refs = btrfs_extent_refs(leaf, ei);
1626         refs += refs_to_add;
1627         btrfs_set_extent_refs(leaf, ei, refs);
1628         if (extent_op)
1629                 __run_delayed_extent_op(extent_op, leaf, ei);
1630
1631         ptr = (unsigned long)ei + item_offset;
1632         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1633         if (ptr < end - size)
1634                 memmove_extent_buffer(leaf, ptr + size, ptr,
1635                                       end - size - ptr);
1636
1637         iref = (struct btrfs_extent_inline_ref *)ptr;
1638         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1639         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1640                 struct btrfs_extent_data_ref *dref;
1641                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1642                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1643                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1644                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1645                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1646         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1647                 struct btrfs_shared_data_ref *sref;
1648                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1649                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1650                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1651         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1652                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1653         } else {
1654                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1655         }
1656         btrfs_mark_buffer_dirty(leaf);
1657 }
1658
1659 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1660                                  struct btrfs_root *root,
1661                                  struct btrfs_path *path,
1662                                  struct btrfs_extent_inline_ref **ref_ret,
1663                                  u64 bytenr, u64 num_bytes, u64 parent,
1664                                  u64 root_objectid, u64 owner, u64 offset)
1665 {
1666         int ret;
1667
1668         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1669                                            bytenr, num_bytes, parent,
1670                                            root_objectid, owner, offset, 0);
1671         if (ret != -ENOENT)
1672                 return ret;
1673
1674         btrfs_release_path(path);
1675         *ref_ret = NULL;
1676
1677         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1678                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1679                                             root_objectid);
1680         } else {
1681                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1682                                              root_objectid, owner, offset);
1683         }
1684         return ret;
1685 }
1686
1687 /*
1688  * helper to update/remove inline back ref
1689  */
1690 static noinline_for_stack
1691 void update_inline_extent_backref(struct btrfs_trans_handle *trans,
1692                                   struct btrfs_root *root,
1693                                   struct btrfs_path *path,
1694                                   struct btrfs_extent_inline_ref *iref,
1695                                   int refs_to_mod,
1696                                   struct btrfs_delayed_extent_op *extent_op)
1697 {
1698         struct extent_buffer *leaf;
1699         struct btrfs_extent_item *ei;
1700         struct btrfs_extent_data_ref *dref = NULL;
1701         struct btrfs_shared_data_ref *sref = NULL;
1702         unsigned long ptr;
1703         unsigned long end;
1704         u32 item_size;
1705         int size;
1706         int type;
1707         u64 refs;
1708
1709         leaf = path->nodes[0];
1710         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1711         refs = btrfs_extent_refs(leaf, ei);
1712         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1713         refs += refs_to_mod;
1714         btrfs_set_extent_refs(leaf, ei, refs);
1715         if (extent_op)
1716                 __run_delayed_extent_op(extent_op, leaf, ei);
1717
1718         type = btrfs_extent_inline_ref_type(leaf, iref);
1719
1720         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1721                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1722                 refs = btrfs_extent_data_ref_count(leaf, dref);
1723         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1724                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1725                 refs = btrfs_shared_data_ref_count(leaf, sref);
1726         } else {
1727                 refs = 1;
1728                 BUG_ON(refs_to_mod != -1);
1729         }
1730
1731         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1732         refs += refs_to_mod;
1733
1734         if (refs > 0) {
1735                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1736                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1737                 else
1738                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1739         } else {
1740                 size =  btrfs_extent_inline_ref_size(type);
1741                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1742                 ptr = (unsigned long)iref;
1743                 end = (unsigned long)ei + item_size;
1744                 if (ptr + size < end)
1745                         memmove_extent_buffer(leaf, ptr, ptr + size,
1746                                               end - ptr - size);
1747                 item_size -= size;
1748                 btrfs_truncate_item(trans, root, path, item_size, 1);
1749         }
1750         btrfs_mark_buffer_dirty(leaf);
1751 }
1752
1753 static noinline_for_stack
1754 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1755                                  struct btrfs_root *root,
1756                                  struct btrfs_path *path,
1757                                  u64 bytenr, u64 num_bytes, u64 parent,
1758                                  u64 root_objectid, u64 owner,
1759                                  u64 offset, int refs_to_add,
1760                                  struct btrfs_delayed_extent_op *extent_op)
1761 {
1762         struct btrfs_extent_inline_ref *iref;
1763         int ret;
1764
1765         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1766                                            bytenr, num_bytes, parent,
1767                                            root_objectid, owner, offset, 1);
1768         if (ret == 0) {
1769                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1770                 update_inline_extent_backref(trans, root, path, iref,
1771                                              refs_to_add, extent_op);
1772         } else if (ret == -ENOENT) {
1773                 setup_inline_extent_backref(trans, root, path, iref, parent,
1774                                             root_objectid, owner, offset,
1775                                             refs_to_add, extent_op);
1776                 ret = 0;
1777         }
1778         return ret;
1779 }
1780
1781 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1782                                  struct btrfs_root *root,
1783                                  struct btrfs_path *path,
1784                                  u64 bytenr, u64 parent, u64 root_objectid,
1785                                  u64 owner, u64 offset, int refs_to_add)
1786 {
1787         int ret;
1788         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1789                 BUG_ON(refs_to_add != 1);
1790                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1791                                             parent, root_objectid);
1792         } else {
1793                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1794                                              parent, root_objectid,
1795                                              owner, offset, refs_to_add);
1796         }
1797         return ret;
1798 }
1799
1800 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1801                                  struct btrfs_root *root,
1802                                  struct btrfs_path *path,
1803                                  struct btrfs_extent_inline_ref *iref,
1804                                  int refs_to_drop, int is_data)
1805 {
1806         int ret = 0;
1807
1808         BUG_ON(!is_data && refs_to_drop != 1);
1809         if (iref) {
1810                 update_inline_extent_backref(trans, root, path, iref,
1811                                              -refs_to_drop, NULL);
1812         } else if (is_data) {
1813                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1814         } else {
1815                 ret = btrfs_del_item(trans, root, path);
1816         }
1817         return ret;
1818 }
1819
1820 static int btrfs_issue_discard(struct block_device *bdev,
1821                                 u64 start, u64 len)
1822 {
1823         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1824 }
1825
1826 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1827                                 u64 num_bytes, u64 *actual_bytes)
1828 {
1829         int ret;
1830         u64 discarded_bytes = 0;
1831         struct btrfs_bio *bbio = NULL;
1832
1833
1834         /* Tell the block device(s) that the sectors can be discarded */
1835         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1836                               bytenr, &num_bytes, &bbio, 0);
1837         /* Error condition is -ENOMEM */
1838         if (!ret) {
1839                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1840                 int i;
1841
1842
1843                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1844                         if (!stripe->dev->can_discard)
1845                                 continue;
1846
1847                         ret = btrfs_issue_discard(stripe->dev->bdev,
1848                                                   stripe->physical,
1849                                                   stripe->length);
1850                         if (!ret)
1851                                 discarded_bytes += stripe->length;
1852                         else if (ret != -EOPNOTSUPP)
1853                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1854
1855                         /*
1856                          * Just in case we get back EOPNOTSUPP for some reason,
1857                          * just ignore the return value so we don't screw up
1858                          * people calling discard_extent.
1859                          */
1860                         ret = 0;
1861                 }
1862                 kfree(bbio);
1863         }
1864
1865         if (actual_bytes)
1866                 *actual_bytes = discarded_bytes;
1867
1868
1869         return ret;
1870 }
1871
1872 /* Can return -ENOMEM */
1873 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1874                          struct btrfs_root *root,
1875                          u64 bytenr, u64 num_bytes, u64 parent,
1876                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1877 {
1878         int ret;
1879         struct btrfs_fs_info *fs_info = root->fs_info;
1880
1881         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1882                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1883
1884         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1885                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1886                                         num_bytes,
1887                                         parent, root_objectid, (int)owner,
1888                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1889         } else {
1890                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1891                                         num_bytes,
1892                                         parent, root_objectid, owner, offset,
1893                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1894         }
1895         return ret;
1896 }
1897
1898 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1899                                   struct btrfs_root *root,
1900                                   u64 bytenr, u64 num_bytes,
1901                                   u64 parent, u64 root_objectid,
1902                                   u64 owner, u64 offset, int refs_to_add,
1903                                   struct btrfs_delayed_extent_op *extent_op)
1904 {
1905         struct btrfs_path *path;
1906         struct extent_buffer *leaf;
1907         struct btrfs_extent_item *item;
1908         u64 refs;
1909         int ret;
1910         int err = 0;
1911
1912         path = btrfs_alloc_path();
1913         if (!path)
1914                 return -ENOMEM;
1915
1916         path->reada = 1;
1917         path->leave_spinning = 1;
1918         /* this will setup the path even if it fails to insert the back ref */
1919         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1920                                            path, bytenr, num_bytes, parent,
1921                                            root_objectid, owner, offset,
1922                                            refs_to_add, extent_op);
1923         if (ret == 0)
1924                 goto out;
1925
1926         if (ret != -EAGAIN) {
1927                 err = ret;
1928                 goto out;
1929         }
1930
1931         leaf = path->nodes[0];
1932         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1933         refs = btrfs_extent_refs(leaf, item);
1934         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1935         if (extent_op)
1936                 __run_delayed_extent_op(extent_op, leaf, item);
1937
1938         btrfs_mark_buffer_dirty(leaf);
1939         btrfs_release_path(path);
1940
1941         path->reada = 1;
1942         path->leave_spinning = 1;
1943
1944         /* now insert the actual backref */
1945         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1946                                     path, bytenr, parent, root_objectid,
1947                                     owner, offset, refs_to_add);
1948         if (ret)
1949                 btrfs_abort_transaction(trans, root, ret);
1950 out:
1951         btrfs_free_path(path);
1952         return err;
1953 }
1954
1955 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1956                                 struct btrfs_root *root,
1957                                 struct btrfs_delayed_ref_node *node,
1958                                 struct btrfs_delayed_extent_op *extent_op,
1959                                 int insert_reserved)
1960 {
1961         int ret = 0;
1962         struct btrfs_delayed_data_ref *ref;
1963         struct btrfs_key ins;
1964         u64 parent = 0;
1965         u64 ref_root = 0;
1966         u64 flags = 0;
1967
1968         ins.objectid = node->bytenr;
1969         ins.offset = node->num_bytes;
1970         ins.type = BTRFS_EXTENT_ITEM_KEY;
1971
1972         ref = btrfs_delayed_node_to_data_ref(node);
1973         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1974                 parent = ref->parent;
1975         else
1976                 ref_root = ref->root;
1977
1978         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1979                 if (extent_op) {
1980                         BUG_ON(extent_op->update_key);
1981                         flags |= extent_op->flags_to_set;
1982                 }
1983                 ret = alloc_reserved_file_extent(trans, root,
1984                                                  parent, ref_root, flags,
1985                                                  ref->objectid, ref->offset,
1986                                                  &ins, node->ref_mod);
1987         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1988                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1989                                              node->num_bytes, parent,
1990                                              ref_root, ref->objectid,
1991                                              ref->offset, node->ref_mod,
1992                                              extent_op);
1993         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1994                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1995                                           node->num_bytes, parent,
1996                                           ref_root, ref->objectid,
1997                                           ref->offset, node->ref_mod,
1998                                           extent_op);
1999         } else {
2000                 BUG();
2001         }
2002         return ret;
2003 }
2004
2005 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2006                                     struct extent_buffer *leaf,
2007                                     struct btrfs_extent_item *ei)
2008 {
2009         u64 flags = btrfs_extent_flags(leaf, ei);
2010         if (extent_op->update_flags) {
2011                 flags |= extent_op->flags_to_set;
2012                 btrfs_set_extent_flags(leaf, ei, flags);
2013         }
2014
2015         if (extent_op->update_key) {
2016                 struct btrfs_tree_block_info *bi;
2017                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2018                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2019                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2020         }
2021 }
2022
2023 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2024                                  struct btrfs_root *root,
2025                                  struct btrfs_delayed_ref_node *node,
2026                                  struct btrfs_delayed_extent_op *extent_op)
2027 {
2028         struct btrfs_key key;
2029         struct btrfs_path *path;
2030         struct btrfs_extent_item *ei;
2031         struct extent_buffer *leaf;
2032         u32 item_size;
2033         int ret;
2034         int err = 0;
2035
2036         if (trans->aborted)
2037                 return 0;
2038
2039         path = btrfs_alloc_path();
2040         if (!path)
2041                 return -ENOMEM;
2042
2043         key.objectid = node->bytenr;
2044         key.type = BTRFS_EXTENT_ITEM_KEY;
2045         key.offset = node->num_bytes;
2046
2047         path->reada = 1;
2048         path->leave_spinning = 1;
2049         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2050                                 path, 0, 1);
2051         if (ret < 0) {
2052                 err = ret;
2053                 goto out;
2054         }
2055         if (ret > 0) {
2056                 err = -EIO;
2057                 goto out;
2058         }
2059
2060         leaf = path->nodes[0];
2061         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2062 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2063         if (item_size < sizeof(*ei)) {
2064                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2065                                              path, (u64)-1, 0);
2066                 if (ret < 0) {
2067                         err = ret;
2068                         goto out;
2069                 }
2070                 leaf = path->nodes[0];
2071                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2072         }
2073 #endif
2074         BUG_ON(item_size < sizeof(*ei));
2075         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2076         __run_delayed_extent_op(extent_op, leaf, ei);
2077
2078         btrfs_mark_buffer_dirty(leaf);
2079 out:
2080         btrfs_free_path(path);
2081         return err;
2082 }
2083
2084 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2085                                 struct btrfs_root *root,
2086                                 struct btrfs_delayed_ref_node *node,
2087                                 struct btrfs_delayed_extent_op *extent_op,
2088                                 int insert_reserved)
2089 {
2090         int ret = 0;
2091         struct btrfs_delayed_tree_ref *ref;
2092         struct btrfs_key ins;
2093         u64 parent = 0;
2094         u64 ref_root = 0;
2095
2096         ins.objectid = node->bytenr;
2097         ins.offset = node->num_bytes;
2098         ins.type = BTRFS_EXTENT_ITEM_KEY;
2099
2100         ref = btrfs_delayed_node_to_tree_ref(node);
2101         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2102                 parent = ref->parent;
2103         else
2104                 ref_root = ref->root;
2105
2106         BUG_ON(node->ref_mod != 1);
2107         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2108                 BUG_ON(!extent_op || !extent_op->update_flags ||
2109                        !extent_op->update_key);
2110                 ret = alloc_reserved_tree_block(trans, root,
2111                                                 parent, ref_root,
2112                                                 extent_op->flags_to_set,
2113                                                 &extent_op->key,
2114                                                 ref->level, &ins);
2115         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2116                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2117                                              node->num_bytes, parent, ref_root,
2118                                              ref->level, 0, 1, extent_op);
2119         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2120                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2121                                           node->num_bytes, parent, ref_root,
2122                                           ref->level, 0, 1, extent_op);
2123         } else {
2124                 BUG();
2125         }
2126         return ret;
2127 }
2128
2129 /* helper function to actually process a single delayed ref entry */
2130 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2131                                struct btrfs_root *root,
2132                                struct btrfs_delayed_ref_node *node,
2133                                struct btrfs_delayed_extent_op *extent_op,
2134                                int insert_reserved)
2135 {
2136         int ret = 0;
2137
2138         if (trans->aborted)
2139                 return 0;
2140
2141         if (btrfs_delayed_ref_is_head(node)) {
2142                 struct btrfs_delayed_ref_head *head;
2143                 /*
2144                  * we've hit the end of the chain and we were supposed
2145                  * to insert this extent into the tree.  But, it got
2146                  * deleted before we ever needed to insert it, so all
2147                  * we have to do is clean up the accounting
2148                  */
2149                 BUG_ON(extent_op);
2150                 head = btrfs_delayed_node_to_head(node);
2151                 if (insert_reserved) {
2152                         btrfs_pin_extent(root, node->bytenr,
2153                                          node->num_bytes, 1);
2154                         if (head->is_data) {
2155                                 ret = btrfs_del_csums(trans, root,
2156                                                       node->bytenr,
2157                                                       node->num_bytes);
2158                         }
2159                 }
2160                 mutex_unlock(&head->mutex);
2161                 return ret;
2162         }
2163
2164         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2165             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2166                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2167                                            insert_reserved);
2168         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2169                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2170                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2171                                            insert_reserved);
2172         else
2173                 BUG();
2174         return ret;
2175 }
2176
2177 static noinline struct btrfs_delayed_ref_node *
2178 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2179 {
2180         struct rb_node *node;
2181         struct btrfs_delayed_ref_node *ref;
2182         int action = BTRFS_ADD_DELAYED_REF;
2183 again:
2184         /*
2185          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2186          * this prevents ref count from going down to zero when
2187          * there still are pending delayed ref.
2188          */
2189         node = rb_prev(&head->node.rb_node);
2190         while (1) {
2191                 if (!node)
2192                         break;
2193                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2194                                 rb_node);
2195                 if (ref->bytenr != head->node.bytenr)
2196                         break;
2197                 if (ref->action == action)
2198                         return ref;
2199                 node = rb_prev(node);
2200         }
2201         if (action == BTRFS_ADD_DELAYED_REF) {
2202                 action = BTRFS_DROP_DELAYED_REF;
2203                 goto again;
2204         }
2205         return NULL;
2206 }
2207
2208 /*
2209  * Returns 0 on success or if called with an already aborted transaction.
2210  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2211  */
2212 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2213                                        struct btrfs_root *root,
2214                                        struct list_head *cluster)
2215 {
2216         struct btrfs_delayed_ref_root *delayed_refs;
2217         struct btrfs_delayed_ref_node *ref;
2218         struct btrfs_delayed_ref_head *locked_ref = NULL;
2219         struct btrfs_delayed_extent_op *extent_op;
2220         int ret;
2221         int count = 0;
2222         int must_insert_reserved = 0;
2223
2224         delayed_refs = &trans->transaction->delayed_refs;
2225         while (1) {
2226                 if (!locked_ref) {
2227                         /* pick a new head ref from the cluster list */
2228                         if (list_empty(cluster))
2229                                 break;
2230
2231                         locked_ref = list_entry(cluster->next,
2232                                      struct btrfs_delayed_ref_head, cluster);
2233
2234                         /* grab the lock that says we are going to process
2235                          * all the refs for this head */
2236                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2237
2238                         /*
2239                          * we may have dropped the spin lock to get the head
2240                          * mutex lock, and that might have given someone else
2241                          * time to free the head.  If that's true, it has been
2242                          * removed from our list and we can move on.
2243                          */
2244                         if (ret == -EAGAIN) {
2245                                 locked_ref = NULL;
2246                                 count++;
2247                                 continue;
2248                         }
2249                 }
2250
2251                 /*
2252                  * locked_ref is the head node, so we have to go one
2253                  * node back for any delayed ref updates
2254                  */
2255                 ref = select_delayed_ref(locked_ref);
2256
2257                 if (ref && ref->seq &&
2258                     btrfs_check_delayed_seq(delayed_refs, ref->seq)) {
2259                         /*
2260                          * there are still refs with lower seq numbers in the
2261                          * process of being added. Don't run this ref yet.
2262                          */
2263                         list_del_init(&locked_ref->cluster);
2264                         mutex_unlock(&locked_ref->mutex);
2265                         locked_ref = NULL;
2266                         delayed_refs->num_heads_ready++;
2267                         spin_unlock(&delayed_refs->lock);
2268                         cond_resched();
2269                         spin_lock(&delayed_refs->lock);
2270                         continue;
2271                 }
2272
2273                 /*
2274                  * record the must insert reserved flag before we
2275                  * drop the spin lock.
2276                  */
2277                 must_insert_reserved = locked_ref->must_insert_reserved;
2278                 locked_ref->must_insert_reserved = 0;
2279
2280                 extent_op = locked_ref->extent_op;
2281                 locked_ref->extent_op = NULL;
2282
2283                 if (!ref) {
2284                         /* All delayed refs have been processed, Go ahead
2285                          * and send the head node to run_one_delayed_ref,
2286                          * so that any accounting fixes can happen
2287                          */
2288                         ref = &locked_ref->node;
2289
2290                         if (extent_op && must_insert_reserved) {
2291                                 kfree(extent_op);
2292                                 extent_op = NULL;
2293                         }
2294
2295                         if (extent_op) {
2296                                 spin_unlock(&delayed_refs->lock);
2297
2298                                 ret = run_delayed_extent_op(trans, root,
2299                                                             ref, extent_op);
2300                                 kfree(extent_op);
2301
2302                                 if (ret) {
2303                                         printk(KERN_DEBUG "btrfs: run_delayed_extent_op returned %d\n", ret);
2304                                         spin_lock(&delayed_refs->lock);
2305                                         return ret;
2306                                 }
2307
2308                                 goto next;
2309                         }
2310
2311                         list_del_init(&locked_ref->cluster);
2312                         locked_ref = NULL;
2313                 }
2314
2315                 ref->in_tree = 0;
2316                 rb_erase(&ref->rb_node, &delayed_refs->root);
2317                 delayed_refs->num_entries--;
2318                 /*
2319                  * we modified num_entries, but as we're currently running
2320                  * delayed refs, skip
2321                  *     wake_up(&delayed_refs->seq_wait);
2322                  * here.
2323                  */
2324                 spin_unlock(&delayed_refs->lock);
2325
2326                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2327                                           must_insert_reserved);
2328
2329                 btrfs_put_delayed_ref(ref);
2330                 kfree(extent_op);
2331                 count++;
2332
2333                 if (ret) {
2334                         printk(KERN_DEBUG "btrfs: run_one_delayed_ref returned %d\n", ret);
2335                         spin_lock(&delayed_refs->lock);
2336                         return ret;
2337                 }
2338
2339 next:
2340                 do_chunk_alloc(trans, root->fs_info->extent_root,
2341                                2 * 1024 * 1024,
2342                                btrfs_get_alloc_profile(root, 0),
2343                                CHUNK_ALLOC_NO_FORCE);
2344                 cond_resched();
2345                 spin_lock(&delayed_refs->lock);
2346         }
2347         return count;
2348 }
2349
2350 static void wait_for_more_refs(struct btrfs_delayed_ref_root *delayed_refs,
2351                                unsigned long num_refs,
2352                                struct list_head *first_seq)
2353 {
2354         spin_unlock(&delayed_refs->lock);
2355         pr_debug("waiting for more refs (num %ld, first %p)\n",
2356                  num_refs, first_seq);
2357         wait_event(delayed_refs->seq_wait,
2358                    num_refs != delayed_refs->num_entries ||
2359                    delayed_refs->seq_head.next != first_seq);
2360         pr_debug("done waiting for more refs (num %ld, first %p)\n",
2361                  delayed_refs->num_entries, delayed_refs->seq_head.next);
2362         spin_lock(&delayed_refs->lock);
2363 }
2364
2365 /*
2366  * this starts processing the delayed reference count updates and
2367  * extent insertions we have queued up so far.  count can be
2368  * 0, which means to process everything in the tree at the start
2369  * of the run (but not newly added entries), or it can be some target
2370  * number you'd like to process.
2371  *
2372  * Returns 0 on success or if called with an aborted transaction
2373  * Returns <0 on error and aborts the transaction
2374  */
2375 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2376                            struct btrfs_root *root, unsigned long count)
2377 {
2378         struct rb_node *node;
2379         struct btrfs_delayed_ref_root *delayed_refs;
2380         struct btrfs_delayed_ref_node *ref;
2381         struct list_head cluster;
2382         struct list_head *first_seq = NULL;
2383         int ret;
2384         u64 delayed_start;
2385         int run_all = count == (unsigned long)-1;
2386         int run_most = 0;
2387         unsigned long num_refs = 0;
2388         int consider_waiting;
2389
2390         /* We'll clean this up in btrfs_cleanup_transaction */
2391         if (trans->aborted)
2392                 return 0;
2393
2394         if (root == root->fs_info->extent_root)
2395                 root = root->fs_info->tree_root;
2396
2397         do_chunk_alloc(trans, root->fs_info->extent_root,
2398                        2 * 1024 * 1024, btrfs_get_alloc_profile(root, 0),
2399                        CHUNK_ALLOC_NO_FORCE);
2400
2401         delayed_refs = &trans->transaction->delayed_refs;
2402         INIT_LIST_HEAD(&cluster);
2403 again:
2404         consider_waiting = 0;
2405         spin_lock(&delayed_refs->lock);
2406         if (count == 0) {
2407                 count = delayed_refs->num_entries * 2;
2408                 run_most = 1;
2409         }
2410         while (1) {
2411                 if (!(run_all || run_most) &&
2412                     delayed_refs->num_heads_ready < 64)
2413                         break;
2414
2415                 /*
2416                  * go find something we can process in the rbtree.  We start at
2417                  * the beginning of the tree, and then build a cluster
2418                  * of refs to process starting at the first one we are able to
2419                  * lock
2420                  */
2421                 delayed_start = delayed_refs->run_delayed_start;
2422                 ret = btrfs_find_ref_cluster(trans, &cluster,
2423                                              delayed_refs->run_delayed_start);
2424                 if (ret)
2425                         break;
2426
2427                 if (delayed_start >= delayed_refs->run_delayed_start) {
2428                         if (consider_waiting == 0) {
2429                                 /*
2430                                  * btrfs_find_ref_cluster looped. let's do one
2431                                  * more cycle. if we don't run any delayed ref
2432                                  * during that cycle (because we can't because
2433                                  * all of them are blocked) and if the number of
2434                                  * refs doesn't change, we avoid busy waiting.
2435                                  */
2436                                 consider_waiting = 1;
2437                                 num_refs = delayed_refs->num_entries;
2438                                 first_seq = root->fs_info->tree_mod_seq_list.next;
2439                         } else {
2440                                 wait_for_more_refs(delayed_refs,
2441                                                    num_refs, first_seq);
2442                                 /*
2443                                  * after waiting, things have changed. we
2444                                  * dropped the lock and someone else might have
2445                                  * run some refs, built new clusters and so on.
2446                                  * therefore, we restart staleness detection.
2447                                  */
2448                                 consider_waiting = 0;
2449                         }
2450                 }
2451
2452                 ret = run_clustered_refs(trans, root, &cluster);
2453                 if (ret < 0) {
2454                         spin_unlock(&delayed_refs->lock);
2455                         btrfs_abort_transaction(trans, root, ret);
2456                         return ret;
2457                 }
2458
2459                 count -= min_t(unsigned long, ret, count);
2460
2461                 if (count == 0)
2462                         break;
2463
2464                 if (ret || delayed_refs->run_delayed_start == 0) {
2465                         /* refs were run, let's reset staleness detection */
2466                         consider_waiting = 0;
2467                 }
2468         }
2469
2470         if (run_all) {
2471                 node = rb_first(&delayed_refs->root);
2472                 if (!node)
2473                         goto out;
2474                 count = (unsigned long)-1;
2475
2476                 while (node) {
2477                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2478                                        rb_node);
2479                         if (btrfs_delayed_ref_is_head(ref)) {
2480                                 struct btrfs_delayed_ref_head *head;
2481
2482                                 head = btrfs_delayed_node_to_head(ref);
2483                                 atomic_inc(&ref->refs);
2484
2485                                 spin_unlock(&delayed_refs->lock);
2486                                 /*
2487                                  * Mutex was contended, block until it's
2488                                  * released and try again
2489                                  */
2490                                 mutex_lock(&head->mutex);
2491                                 mutex_unlock(&head->mutex);
2492
2493                                 btrfs_put_delayed_ref(ref);
2494                                 cond_resched();
2495                                 goto again;
2496                         }
2497                         node = rb_next(node);
2498                 }
2499                 spin_unlock(&delayed_refs->lock);
2500                 schedule_timeout(1);
2501                 goto again;
2502         }
2503 out:
2504         spin_unlock(&delayed_refs->lock);
2505         return 0;
2506 }
2507
2508 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2509                                 struct btrfs_root *root,
2510                                 u64 bytenr, u64 num_bytes, u64 flags,
2511                                 int is_data)
2512 {
2513         struct btrfs_delayed_extent_op *extent_op;
2514         int ret;
2515
2516         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2517         if (!extent_op)
2518                 return -ENOMEM;
2519
2520         extent_op->flags_to_set = flags;
2521         extent_op->update_flags = 1;
2522         extent_op->update_key = 0;
2523         extent_op->is_data = is_data ? 1 : 0;
2524
2525         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2526                                           num_bytes, extent_op);
2527         if (ret)
2528                 kfree(extent_op);
2529         return ret;
2530 }
2531
2532 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2533                                       struct btrfs_root *root,
2534                                       struct btrfs_path *path,
2535                                       u64 objectid, u64 offset, u64 bytenr)
2536 {
2537         struct btrfs_delayed_ref_head *head;
2538         struct btrfs_delayed_ref_node *ref;
2539         struct btrfs_delayed_data_ref *data_ref;
2540         struct btrfs_delayed_ref_root *delayed_refs;
2541         struct rb_node *node;
2542         int ret = 0;
2543
2544         ret = -ENOENT;
2545         delayed_refs = &trans->transaction->delayed_refs;
2546         spin_lock(&delayed_refs->lock);
2547         head = btrfs_find_delayed_ref_head(trans, bytenr);
2548         if (!head)
2549                 goto out;
2550
2551         if (!mutex_trylock(&head->mutex)) {
2552                 atomic_inc(&head->node.refs);
2553                 spin_unlock(&delayed_refs->lock);
2554
2555                 btrfs_release_path(path);
2556
2557                 /*
2558                  * Mutex was contended, block until it's released and let
2559                  * caller try again
2560                  */
2561                 mutex_lock(&head->mutex);
2562                 mutex_unlock(&head->mutex);
2563                 btrfs_put_delayed_ref(&head->node);
2564                 return -EAGAIN;
2565         }
2566
2567         node = rb_prev(&head->node.rb_node);
2568         if (!node)
2569                 goto out_unlock;
2570
2571         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2572
2573         if (ref->bytenr != bytenr)
2574                 goto out_unlock;
2575
2576         ret = 1;
2577         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2578                 goto out_unlock;
2579
2580         data_ref = btrfs_delayed_node_to_data_ref(ref);
2581
2582         node = rb_prev(node);
2583         if (node) {
2584                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2585                 if (ref->bytenr == bytenr)
2586                         goto out_unlock;
2587         }
2588
2589         if (data_ref->root != root->root_key.objectid ||
2590             data_ref->objectid != objectid || data_ref->offset != offset)
2591                 goto out_unlock;
2592
2593         ret = 0;
2594 out_unlock:
2595         mutex_unlock(&head->mutex);
2596 out:
2597         spin_unlock(&delayed_refs->lock);
2598         return ret;
2599 }
2600
2601 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2602                                         struct btrfs_root *root,
2603                                         struct btrfs_path *path,
2604                                         u64 objectid, u64 offset, u64 bytenr)
2605 {
2606         struct btrfs_root *extent_root = root->fs_info->extent_root;
2607         struct extent_buffer *leaf;
2608         struct btrfs_extent_data_ref *ref;
2609         struct btrfs_extent_inline_ref *iref;
2610         struct btrfs_extent_item *ei;
2611         struct btrfs_key key;
2612         u32 item_size;
2613         int ret;
2614
2615         key.objectid = bytenr;
2616         key.offset = (u64)-1;
2617         key.type = BTRFS_EXTENT_ITEM_KEY;
2618
2619         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2620         if (ret < 0)
2621                 goto out;
2622         BUG_ON(ret == 0); /* Corruption */
2623
2624         ret = -ENOENT;
2625         if (path->slots[0] == 0)
2626                 goto out;
2627
2628         path->slots[0]--;
2629         leaf = path->nodes[0];
2630         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2631
2632         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2633                 goto out;
2634
2635         ret = 1;
2636         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2637 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2638         if (item_size < sizeof(*ei)) {
2639                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2640                 goto out;
2641         }
2642 #endif
2643         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2644
2645         if (item_size != sizeof(*ei) +
2646             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2647                 goto out;
2648
2649         if (btrfs_extent_generation(leaf, ei) <=
2650             btrfs_root_last_snapshot(&root->root_item))
2651                 goto out;
2652
2653         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2654         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2655             BTRFS_EXTENT_DATA_REF_KEY)
2656                 goto out;
2657
2658         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2659         if (btrfs_extent_refs(leaf, ei) !=
2660             btrfs_extent_data_ref_count(leaf, ref) ||
2661             btrfs_extent_data_ref_root(leaf, ref) !=
2662             root->root_key.objectid ||
2663             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2664             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2665                 goto out;
2666
2667         ret = 0;
2668 out:
2669         return ret;
2670 }
2671
2672 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2673                           struct btrfs_root *root,
2674                           u64 objectid, u64 offset, u64 bytenr)
2675 {
2676         struct btrfs_path *path;
2677         int ret;
2678         int ret2;
2679
2680         path = btrfs_alloc_path();
2681         if (!path)
2682                 return -ENOENT;
2683
2684         do {
2685                 ret = check_committed_ref(trans, root, path, objectid,
2686                                           offset, bytenr);
2687                 if (ret && ret != -ENOENT)
2688                         goto out;
2689
2690                 ret2 = check_delayed_ref(trans, root, path, objectid,
2691                                          offset, bytenr);
2692         } while (ret2 == -EAGAIN);
2693
2694         if (ret2 && ret2 != -ENOENT) {
2695                 ret = ret2;
2696                 goto out;
2697         }
2698
2699         if (ret != -ENOENT || ret2 != -ENOENT)
2700                 ret = 0;
2701 out:
2702         btrfs_free_path(path);
2703         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2704                 WARN_ON(ret > 0);
2705         return ret;
2706 }
2707
2708 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2709                            struct btrfs_root *root,
2710                            struct extent_buffer *buf,
2711                            int full_backref, int inc, int for_cow)
2712 {
2713         u64 bytenr;
2714         u64 num_bytes;
2715         u64 parent;
2716         u64 ref_root;
2717         u32 nritems;
2718         struct btrfs_key key;
2719         struct btrfs_file_extent_item *fi;
2720         int i;
2721         int level;
2722         int ret = 0;
2723         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2724                             u64, u64, u64, u64, u64, u64, int);
2725
2726         ref_root = btrfs_header_owner(buf);
2727         nritems = btrfs_header_nritems(buf);
2728         level = btrfs_header_level(buf);
2729
2730         if (!root->ref_cows && level == 0)
2731                 return 0;
2732
2733         if (inc)
2734                 process_func = btrfs_inc_extent_ref;
2735         else
2736                 process_func = btrfs_free_extent;
2737
2738         if (full_backref)
2739                 parent = buf->start;
2740         else
2741                 parent = 0;
2742
2743         for (i = 0; i < nritems; i++) {
2744                 if (level == 0) {
2745                         btrfs_item_key_to_cpu(buf, &key, i);
2746                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2747                                 continue;
2748                         fi = btrfs_item_ptr(buf, i,
2749                                             struct btrfs_file_extent_item);
2750                         if (btrfs_file_extent_type(buf, fi) ==
2751                             BTRFS_FILE_EXTENT_INLINE)
2752                                 continue;
2753                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2754                         if (bytenr == 0)
2755                                 continue;
2756
2757                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2758                         key.offset -= btrfs_file_extent_offset(buf, fi);
2759                         ret = process_func(trans, root, bytenr, num_bytes,
2760                                            parent, ref_root, key.objectid,
2761                                            key.offset, for_cow);
2762                         if (ret)
2763                                 goto fail;
2764                 } else {
2765                         bytenr = btrfs_node_blockptr(buf, i);
2766                         num_bytes = btrfs_level_size(root, level - 1);
2767                         ret = process_func(trans, root, bytenr, num_bytes,
2768                                            parent, ref_root, level - 1, 0,
2769                                            for_cow);
2770                         if (ret)
2771                                 goto fail;
2772                 }
2773         }
2774         return 0;
2775 fail:
2776         return ret;
2777 }
2778
2779 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2780                   struct extent_buffer *buf, int full_backref, int for_cow)
2781 {
2782         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2783 }
2784
2785 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2786                   struct extent_buffer *buf, int full_backref, int for_cow)
2787 {
2788         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2789 }
2790
2791 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2792                                  struct btrfs_root *root,
2793                                  struct btrfs_path *path,
2794                                  struct btrfs_block_group_cache *cache)
2795 {
2796         int ret;
2797         struct btrfs_root *extent_root = root->fs_info->extent_root;
2798         unsigned long bi;
2799         struct extent_buffer *leaf;
2800
2801         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2802         if (ret < 0)
2803                 goto fail;
2804         BUG_ON(ret); /* Corruption */
2805
2806         leaf = path->nodes[0];
2807         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2808         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2809         btrfs_mark_buffer_dirty(leaf);
2810         btrfs_release_path(path);
2811 fail:
2812         if (ret) {
2813                 btrfs_abort_transaction(trans, root, ret);
2814                 return ret;
2815         }
2816         return 0;
2817
2818 }
2819
2820 static struct btrfs_block_group_cache *
2821 next_block_group(struct btrfs_root *root,
2822                  struct btrfs_block_group_cache *cache)
2823 {
2824         struct rb_node *node;
2825         spin_lock(&root->fs_info->block_group_cache_lock);
2826         node = rb_next(&cache->cache_node);
2827         btrfs_put_block_group(cache);
2828         if (node) {
2829                 cache = rb_entry(node, struct btrfs_block_group_cache,
2830                                  cache_node);
2831                 btrfs_get_block_group(cache);
2832         } else
2833                 cache = NULL;
2834         spin_unlock(&root->fs_info->block_group_cache_lock);
2835         return cache;
2836 }
2837
2838 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2839                             struct btrfs_trans_handle *trans,
2840                             struct btrfs_path *path)
2841 {
2842         struct btrfs_root *root = block_group->fs_info->tree_root;
2843         struct inode *inode = NULL;
2844         u64 alloc_hint = 0;
2845         int dcs = BTRFS_DC_ERROR;
2846         int num_pages = 0;
2847         int retries = 0;
2848         int ret = 0;
2849
2850         /*
2851          * If this block group is smaller than 100 megs don't bother caching the
2852          * block group.
2853          */
2854         if (block_group->key.offset < (100 * 1024 * 1024)) {
2855                 spin_lock(&block_group->lock);
2856                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2857                 spin_unlock(&block_group->lock);
2858                 return 0;
2859         }
2860
2861 again:
2862         inode = lookup_free_space_inode(root, block_group, path);
2863         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2864                 ret = PTR_ERR(inode);
2865                 btrfs_release_path(path);
2866                 goto out;
2867         }
2868
2869         if (IS_ERR(inode)) {
2870                 BUG_ON(retries);
2871                 retries++;
2872
2873                 if (block_group->ro)
2874                         goto out_free;
2875
2876                 ret = create_free_space_inode(root, trans, block_group, path);
2877                 if (ret)
2878                         goto out_free;
2879                 goto again;
2880         }
2881
2882         /* We've already setup this transaction, go ahead and exit */
2883         if (block_group->cache_generation == trans->transid &&
2884             i_size_read(inode)) {
2885                 dcs = BTRFS_DC_SETUP;
2886                 goto out_put;
2887         }
2888
2889         /*
2890          * We want to set the generation to 0, that way if anything goes wrong
2891          * from here on out we know not to trust this cache when we load up next
2892          * time.
2893          */
2894         BTRFS_I(inode)->generation = 0;
2895         ret = btrfs_update_inode(trans, root, inode);
2896         WARN_ON(ret);
2897
2898         if (i_size_read(inode) > 0) {
2899                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2900                                                       inode);
2901                 if (ret)
2902                         goto out_put;
2903         }
2904
2905         spin_lock(&block_group->lock);
2906         if (block_group->cached != BTRFS_CACHE_FINISHED ||
2907             !btrfs_test_opt(root, SPACE_CACHE)) {
2908                 /*
2909                  * don't bother trying to write stuff out _if_
2910                  * a) we're not cached,
2911                  * b) we're with nospace_cache mount option.
2912                  */
2913                 dcs = BTRFS_DC_WRITTEN;
2914                 spin_unlock(&block_group->lock);
2915                 goto out_put;
2916         }
2917         spin_unlock(&block_group->lock);
2918
2919         num_pages = (int)div64_u64(block_group->key.offset, 1024 * 1024 * 1024);
2920         if (!num_pages)
2921                 num_pages = 1;
2922
2923         /*
2924          * Just to make absolutely sure we have enough space, we're going to
2925          * preallocate 12 pages worth of space for each block group.  In
2926          * practice we ought to use at most 8, but we need extra space so we can
2927          * add our header and have a terminator between the extents and the
2928          * bitmaps.
2929          */
2930         num_pages *= 16;
2931         num_pages *= PAGE_CACHE_SIZE;
2932
2933         ret = btrfs_check_data_free_space(inode, num_pages);
2934         if (ret)
2935                 goto out_put;
2936
2937         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
2938                                               num_pages, num_pages,
2939                                               &alloc_hint);
2940         if (!ret)
2941                 dcs = BTRFS_DC_SETUP;
2942         btrfs_free_reserved_data_space(inode, num_pages);
2943
2944 out_put:
2945         iput(inode);
2946 out_free:
2947         btrfs_release_path(path);
2948 out:
2949         spin_lock(&block_group->lock);
2950         if (!ret && dcs == BTRFS_DC_SETUP)
2951                 block_group->cache_generation = trans->transid;
2952         block_group->disk_cache_state = dcs;
2953         spin_unlock(&block_group->lock);
2954
2955         return ret;
2956 }
2957
2958 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
2959                                    struct btrfs_root *root)
2960 {
2961         struct btrfs_block_group_cache *cache;
2962         int err = 0;
2963         struct btrfs_path *path;
2964         u64 last = 0;
2965
2966         path = btrfs_alloc_path();
2967         if (!path)
2968                 return -ENOMEM;
2969
2970 again:
2971         while (1) {
2972                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2973                 while (cache) {
2974                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
2975                                 break;
2976                         cache = next_block_group(root, cache);
2977                 }
2978                 if (!cache) {
2979                         if (last == 0)
2980                                 break;
2981                         last = 0;
2982                         continue;
2983                 }
2984                 err = cache_save_setup(cache, trans, path);
2985                 last = cache->key.objectid + cache->key.offset;
2986                 btrfs_put_block_group(cache);
2987         }
2988
2989         while (1) {
2990                 if (last == 0) {
2991                         err = btrfs_run_delayed_refs(trans, root,
2992                                                      (unsigned long)-1);
2993                         if (err) /* File system offline */
2994                                 goto out;
2995                 }
2996
2997                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
2998                 while (cache) {
2999                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3000                                 btrfs_put_block_group(cache);
3001                                 goto again;
3002                         }
3003
3004                         if (cache->dirty)
3005                                 break;
3006                         cache = next_block_group(root, cache);
3007                 }
3008                 if (!cache) {
3009                         if (last == 0)
3010                                 break;
3011                         last = 0;
3012                         continue;
3013                 }
3014
3015                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3016                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3017                 cache->dirty = 0;
3018                 last = cache->key.objectid + cache->key.offset;
3019
3020                 err = write_one_cache_group(trans, root, path, cache);
3021                 if (err) /* File system offline */
3022                         goto out;
3023
3024                 btrfs_put_block_group(cache);
3025         }
3026
3027         while (1) {
3028                 /*
3029                  * I don't think this is needed since we're just marking our
3030                  * preallocated extent as written, but just in case it can't
3031                  * hurt.
3032                  */
3033                 if (last == 0) {
3034                         err = btrfs_run_delayed_refs(trans, root,
3035                                                      (unsigned long)-1);
3036                         if (err) /* File system offline */
3037                                 goto out;
3038                 }
3039
3040                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3041                 while (cache) {
3042                         /*
3043                          * Really this shouldn't happen, but it could if we
3044                          * couldn't write the entire preallocated extent and
3045                          * splitting the extent resulted in a new block.
3046                          */
3047                         if (cache->dirty) {
3048                                 btrfs_put_block_group(cache);
3049                                 goto again;
3050                         }
3051                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3052                                 break;
3053                         cache = next_block_group(root, cache);
3054                 }
3055                 if (!cache) {
3056                         if (last == 0)
3057                                 break;
3058                         last = 0;
3059                         continue;
3060                 }
3061
3062                 err = btrfs_write_out_cache(root, trans, cache, path);
3063
3064                 /*
3065                  * If we didn't have an error then the cache state is still
3066                  * NEED_WRITE, so we can set it to WRITTEN.
3067                  */
3068                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3069                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3070                 last = cache->key.objectid + cache->key.offset;
3071                 btrfs_put_block_group(cache);
3072         }
3073 out:
3074
3075         btrfs_free_path(path);
3076         return err;
3077 }
3078
3079 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3080 {
3081         struct btrfs_block_group_cache *block_group;
3082         int readonly = 0;
3083
3084         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3085         if (!block_group || block_group->ro)
3086                 readonly = 1;
3087         if (block_group)
3088                 btrfs_put_block_group(block_group);
3089         return readonly;
3090 }
3091
3092 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3093                              u64 total_bytes, u64 bytes_used,
3094                              struct btrfs_space_info **space_info)
3095 {
3096         struct btrfs_space_info *found;
3097         int i;
3098         int factor;
3099
3100         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3101                      BTRFS_BLOCK_GROUP_RAID10))
3102                 factor = 2;
3103         else
3104                 factor = 1;
3105
3106         found = __find_space_info(info, flags);
3107         if (found) {
3108                 spin_lock(&found->lock);
3109                 found->total_bytes += total_bytes;
3110                 found->disk_total += total_bytes * factor;
3111                 found->bytes_used += bytes_used;
3112                 found->disk_used += bytes_used * factor;
3113                 found->full = 0;
3114                 spin_unlock(&found->lock);
3115                 *space_info = found;
3116                 return 0;
3117         }
3118         found = kzalloc(sizeof(*found), GFP_NOFS);
3119         if (!found)
3120                 return -ENOMEM;
3121
3122         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3123                 INIT_LIST_HEAD(&found->block_groups[i]);
3124         init_rwsem(&found->groups_sem);
3125         spin_lock_init(&found->lock);
3126         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3127         found->total_bytes = total_bytes;
3128         found->disk_total = total_bytes * factor;
3129         found->bytes_used = bytes_used;
3130         found->disk_used = bytes_used * factor;
3131         found->bytes_pinned = 0;
3132         found->bytes_reserved = 0;
3133         found->bytes_readonly = 0;
3134         found->bytes_may_use = 0;
3135         found->full = 0;
3136         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3137         found->chunk_alloc = 0;
3138         found->flush = 0;
3139         init_waitqueue_head(&found->wait);
3140         *space_info = found;
3141         list_add_rcu(&found->list, &info->space_info);
3142         return 0;
3143 }
3144
3145 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3146 {
3147         u64 extra_flags = chunk_to_extended(flags) &
3148                                 BTRFS_EXTENDED_PROFILE_MASK;
3149
3150         if (flags & BTRFS_BLOCK_GROUP_DATA)
3151                 fs_info->avail_data_alloc_bits |= extra_flags;
3152         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3153                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3154         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3155                 fs_info->avail_system_alloc_bits |= extra_flags;
3156 }
3157
3158 /*
3159  * returns target flags in extended format or 0 if restripe for this
3160  * chunk_type is not in progress
3161  *
3162  * should be called with either volume_mutex or balance_lock held
3163  */
3164 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3165 {
3166         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3167         u64 target = 0;
3168
3169         if (!bctl)
3170                 return 0;
3171
3172         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3173             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3174                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3175         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3176                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3177                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3178         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3179                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3180                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3181         }
3182
3183         return target;
3184 }
3185
3186 /*
3187  * @flags: available profiles in extended format (see ctree.h)
3188  *
3189  * Returns reduced profile in chunk format.  If profile changing is in
3190  * progress (either running or paused) picks the target profile (if it's
3191  * already available), otherwise falls back to plain reducing.
3192  */
3193 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3194 {
3195         /*
3196          * we add in the count of missing devices because we want
3197          * to make sure that any RAID levels on a degraded FS
3198          * continue to be honored.
3199          */
3200         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3201                 root->fs_info->fs_devices->missing_devices;
3202         u64 target;
3203
3204         /*
3205          * see if restripe for this chunk_type is in progress, if so
3206          * try to reduce to the target profile
3207          */
3208         spin_lock(&root->fs_info->balance_lock);
3209         target = get_restripe_target(root->fs_info, flags);
3210         if (target) {
3211                 /* pick target profile only if it's already available */
3212                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3213                         spin_unlock(&root->fs_info->balance_lock);
3214                         return extended_to_chunk(target);
3215                 }
3216         }
3217         spin_unlock(&root->fs_info->balance_lock);
3218
3219         if (num_devices == 1)
3220                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3221         if (num_devices < 4)
3222                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3223
3224         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3225             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3226                       BTRFS_BLOCK_GROUP_RAID10))) {
3227                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3228         }
3229
3230         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3231             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3232                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3233         }
3234
3235         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3236             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3237              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3238              (flags & BTRFS_BLOCK_GROUP_DUP))) {
3239                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3240         }
3241
3242         return extended_to_chunk(flags);
3243 }
3244
3245 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3246 {
3247         if (flags & BTRFS_BLOCK_GROUP_DATA)
3248                 flags |= root->fs_info->avail_data_alloc_bits;
3249         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3250                 flags |= root->fs_info->avail_system_alloc_bits;
3251         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3252                 flags |= root->fs_info->avail_metadata_alloc_bits;
3253
3254         return btrfs_reduce_alloc_profile(root, flags);
3255 }
3256
3257 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3258 {
3259         u64 flags;
3260
3261         if (data)
3262                 flags = BTRFS_BLOCK_GROUP_DATA;
3263         else if (root == root->fs_info->chunk_root)
3264                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3265         else
3266                 flags = BTRFS_BLOCK_GROUP_METADATA;
3267
3268         return get_alloc_profile(root, flags);
3269 }
3270
3271 void btrfs_set_inode_space_info(struct btrfs_root *root, struct inode *inode)
3272 {
3273         BTRFS_I(inode)->space_info = __find_space_info(root->fs_info,
3274                                                        BTRFS_BLOCK_GROUP_DATA);
3275 }
3276
3277 /*
3278  * This will check the space that the inode allocates from to make sure we have
3279  * enough space for bytes.
3280  */
3281 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3282 {
3283         struct btrfs_space_info *data_sinfo;
3284         struct btrfs_root *root = BTRFS_I(inode)->root;
3285         u64 used;
3286         int ret = 0, committed = 0, alloc_chunk = 1;
3287
3288         /* make sure bytes are sectorsize aligned */
3289         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3290
3291         if (root == root->fs_info->tree_root ||
3292             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3293                 alloc_chunk = 0;
3294                 committed = 1;
3295         }
3296
3297         data_sinfo = BTRFS_I(inode)->space_info;
3298         if (!data_sinfo)
3299                 goto alloc;
3300
3301 again:
3302         /* make sure we have enough space to handle the data first */
3303         spin_lock(&data_sinfo->lock);
3304         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3305                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3306                 data_sinfo->bytes_may_use;
3307
3308         if (used + bytes > data_sinfo->total_bytes) {
3309                 struct btrfs_trans_handle *trans;
3310
3311                 /*
3312                  * if we don't have enough free bytes in this space then we need
3313                  * to alloc a new chunk.
3314                  */
3315                 if (!data_sinfo->full && alloc_chunk) {
3316                         u64 alloc_target;
3317
3318                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3319                         spin_unlock(&data_sinfo->lock);
3320 alloc:
3321                         alloc_target = btrfs_get_alloc_profile(root, 1);
3322                         trans = btrfs_join_transaction(root);
3323                         if (IS_ERR(trans))
3324                                 return PTR_ERR(trans);
3325
3326                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3327                                              bytes + 2 * 1024 * 1024,
3328                                              alloc_target,
3329                                              CHUNK_ALLOC_NO_FORCE);
3330                         btrfs_end_transaction(trans, root);
3331                         if (ret < 0) {
3332                                 if (ret != -ENOSPC)
3333                                         return ret;
3334                                 else
3335                                         goto commit_trans;
3336                         }
3337
3338                         if (!data_sinfo) {
3339                                 btrfs_set_inode_space_info(root, inode);
3340                                 data_sinfo = BTRFS_I(inode)->space_info;
3341                         }
3342                         goto again;
3343                 }
3344
3345                 /*
3346                  * If we have less pinned bytes than we want to allocate then
3347                  * don't bother committing the transaction, it won't help us.
3348                  */
3349                 if (data_sinfo->bytes_pinned < bytes)
3350                         committed = 1;
3351                 spin_unlock(&data_sinfo->lock);
3352
3353                 /* commit the current transaction and try again */
3354 commit_trans:
3355                 if (!committed &&
3356                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3357                         committed = 1;
3358                         trans = btrfs_join_transaction(root);
3359                         if (IS_ERR(trans))
3360                                 return PTR_ERR(trans);
3361                         ret = btrfs_commit_transaction(trans, root);
3362                         if (ret)
3363                                 return ret;
3364                         goto again;
3365                 }
3366
3367                 return -ENOSPC;
3368         }
3369         data_sinfo->bytes_may_use += bytes;
3370         trace_btrfs_space_reservation(root->fs_info, "space_info",
3371                                       data_sinfo->flags, bytes, 1);
3372         spin_unlock(&data_sinfo->lock);
3373
3374         return 0;
3375 }
3376
3377 /*
3378  * Called if we need to clear a data reservation for this inode.
3379  */
3380 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3381 {
3382         struct btrfs_root *root = BTRFS_I(inode)->root;
3383         struct btrfs_space_info *data_sinfo;
3384
3385         /* make sure bytes are sectorsize aligned */
3386         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3387
3388         data_sinfo = BTRFS_I(inode)->space_info;
3389         spin_lock(&data_sinfo->lock);
3390         data_sinfo->bytes_may_use -= bytes;
3391         trace_btrfs_space_reservation(root->fs_info, "space_info",
3392                                       data_sinfo->flags, bytes, 0);
3393         spin_unlock(&data_sinfo->lock);
3394 }
3395
3396 static void force_metadata_allocation(struct btrfs_fs_info *info)
3397 {
3398         struct list_head *head = &info->space_info;
3399         struct btrfs_space_info *found;
3400
3401         rcu_read_lock();
3402         list_for_each_entry_rcu(found, head, list) {
3403                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3404                         found->force_alloc = CHUNK_ALLOC_FORCE;
3405         }
3406         rcu_read_unlock();
3407 }
3408
3409 static int should_alloc_chunk(struct btrfs_root *root,
3410                               struct btrfs_space_info *sinfo, u64 alloc_bytes,
3411                               int force)
3412 {
3413         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3414         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3415         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3416         u64 thresh;
3417
3418         if (force == CHUNK_ALLOC_FORCE)
3419                 return 1;
3420
3421         /*
3422          * We need to take into account the global rsv because for all intents
3423          * and purposes it's used space.  Don't worry about locking the
3424          * global_rsv, it doesn't change except when the transaction commits.
3425          */
3426         num_allocated += global_rsv->size;
3427
3428         /*
3429          * in limited mode, we want to have some free space up to
3430          * about 1% of the FS size.
3431          */
3432         if (force == CHUNK_ALLOC_LIMITED) {
3433                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3434                 thresh = max_t(u64, 64 * 1024 * 1024,
3435                                div_factor_fine(thresh, 1));
3436
3437                 if (num_bytes - num_allocated < thresh)
3438                         return 1;
3439         }
3440         thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3441
3442         /* 256MB or 2% of the FS */
3443         thresh = max_t(u64, 256 * 1024 * 1024, div_factor_fine(thresh, 2));
3444         /* system chunks need a much small threshold */
3445         if (sinfo->flags & BTRFS_BLOCK_GROUP_SYSTEM)
3446                 thresh = 32 * 1024 * 1024;
3447
3448         if (num_bytes > thresh && sinfo->bytes_used < div_factor(num_bytes, 8))
3449                 return 0;
3450         return 1;
3451 }
3452
3453 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3454 {
3455         u64 num_dev;
3456
3457         if (type & BTRFS_BLOCK_GROUP_RAID10 ||
3458             type & BTRFS_BLOCK_GROUP_RAID0)
3459                 num_dev = root->fs_info->fs_devices->rw_devices;
3460         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3461                 num_dev = 2;
3462         else
3463                 num_dev = 1;    /* DUP or single */
3464
3465         /* metadata for updaing devices and chunk tree */
3466         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3467 }
3468
3469 static void check_system_chunk(struct btrfs_trans_handle *trans,
3470                                struct btrfs_root *root, u64 type)
3471 {
3472         struct btrfs_space_info *info;
3473         u64 left;
3474         u64 thresh;
3475
3476         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3477         spin_lock(&info->lock);
3478         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3479                 info->bytes_reserved - info->bytes_readonly;
3480         spin_unlock(&info->lock);
3481
3482         thresh = get_system_chunk_thresh(root, type);
3483         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3484                 printk(KERN_INFO "left=%llu, need=%llu, flags=%llu\n",
3485                        left, thresh, type);
3486                 dump_space_info(info, 0, 0);
3487         }
3488
3489         if (left < thresh) {
3490                 u64 flags;
3491
3492                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3493                 btrfs_alloc_chunk(trans, root, flags);
3494         }
3495 }
3496
3497 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3498                           struct btrfs_root *extent_root, u64 alloc_bytes,
3499                           u64 flags, int force)
3500 {
3501         struct btrfs_space_info *space_info;
3502         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3503         int wait_for_alloc = 0;
3504         int ret = 0;
3505
3506         space_info = __find_space_info(extent_root->fs_info, flags);
3507         if (!space_info) {
3508                 ret = update_space_info(extent_root->fs_info, flags,
3509                                         0, 0, &space_info);
3510                 BUG_ON(ret); /* -ENOMEM */
3511         }
3512         BUG_ON(!space_info); /* Logic error */
3513
3514 again:
3515         spin_lock(&space_info->lock);
3516         if (force < space_info->force_alloc)
3517                 force = space_info->force_alloc;
3518         if (space_info->full) {
3519                 spin_unlock(&space_info->lock);
3520                 return 0;
3521         }
3522
3523         if (!should_alloc_chunk(extent_root, space_info, alloc_bytes, force)) {
3524                 spin_unlock(&space_info->lock);
3525                 return 0;
3526         } else if (space_info->chunk_alloc) {
3527                 wait_for_alloc = 1;
3528         } else {
3529                 space_info->chunk_alloc = 1;
3530         }
3531
3532         spin_unlock(&space_info->lock);
3533
3534         mutex_lock(&fs_info->chunk_mutex);
3535
3536         /*
3537          * The chunk_mutex is held throughout the entirety of a chunk
3538          * allocation, so once we've acquired the chunk_mutex we know that the
3539          * other guy is done and we need to recheck and see if we should
3540          * allocate.
3541          */
3542         if (wait_for_alloc) {
3543                 mutex_unlock(&fs_info->chunk_mutex);
3544                 wait_for_alloc = 0;
3545                 goto again;
3546         }
3547
3548         /*
3549          * If we have mixed data/metadata chunks we want to make sure we keep
3550          * allocating mixed chunks instead of individual chunks.
3551          */
3552         if (btrfs_mixed_space_info(space_info))
3553                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3554
3555         /*
3556          * if we're doing a data chunk, go ahead and make sure that
3557          * we keep a reasonable number of metadata chunks allocated in the
3558          * FS as well.
3559          */
3560         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3561                 fs_info->data_chunk_allocations++;
3562                 if (!(fs_info->data_chunk_allocations %
3563                       fs_info->metadata_ratio))
3564                         force_metadata_allocation(fs_info);
3565         }
3566
3567         /*
3568          * Check if we have enough space in SYSTEM chunk because we may need
3569          * to update devices.
3570          */
3571         check_system_chunk(trans, extent_root, flags);
3572
3573         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3574         if (ret < 0 && ret != -ENOSPC)
3575                 goto out;
3576
3577         spin_lock(&space_info->lock);
3578         if (ret)
3579                 space_info->full = 1;
3580         else
3581                 ret = 1;
3582
3583         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3584         space_info->chunk_alloc = 0;
3585         spin_unlock(&space_info->lock);
3586 out:
3587         mutex_unlock(&fs_info->chunk_mutex);
3588         return ret;
3589 }
3590
3591 /*
3592  * shrink metadata reservation for delalloc
3593  */
3594 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3595                             bool wait_ordered)
3596 {
3597         struct btrfs_block_rsv *block_rsv;
3598         struct btrfs_space_info *space_info;
3599         struct btrfs_trans_handle *trans;
3600         u64 delalloc_bytes;
3601         u64 max_reclaim;
3602         long time_left;
3603         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3604         int loops = 0;
3605
3606         trans = (struct btrfs_trans_handle *)current->journal_info;
3607         block_rsv = &root->fs_info->delalloc_block_rsv;
3608         space_info = block_rsv->space_info;
3609
3610         smp_mb();
3611         delalloc_bytes = root->fs_info->delalloc_bytes;
3612         if (delalloc_bytes == 0) {
3613                 if (trans)
3614                         return;
3615                 btrfs_wait_ordered_extents(root, 0, 0);
3616                 return;
3617         }
3618
3619         while (delalloc_bytes && loops < 3) {
3620                 max_reclaim = min(delalloc_bytes, to_reclaim);
3621                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3622                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3623                                                WB_REASON_FS_FREE_SPACE);
3624
3625                 spin_lock(&space_info->lock);
3626                 if (space_info->bytes_used + space_info->bytes_reserved +
3627                     space_info->bytes_pinned + space_info->bytes_readonly +
3628                     space_info->bytes_may_use + orig <=
3629                     space_info->total_bytes) {
3630                         spin_unlock(&space_info->lock);
3631                         break;
3632                 }
3633                 spin_unlock(&space_info->lock);
3634
3635                 loops++;
3636                 if (wait_ordered && !trans) {
3637                         btrfs_wait_ordered_extents(root, 0, 0);
3638                 } else {
3639                         time_left = schedule_timeout_killable(1);
3640                         if (time_left)
3641                                 break;
3642                 }
3643                 smp_mb();
3644                 delalloc_bytes = root->fs_info->delalloc_bytes;
3645         }
3646 }
3647
3648 /**
3649  * maybe_commit_transaction - possibly commit the transaction if its ok to
3650  * @root - the root we're allocating for
3651  * @bytes - the number of bytes we want to reserve
3652  * @force - force the commit
3653  *
3654  * This will check to make sure that committing the transaction will actually
3655  * get us somewhere and then commit the transaction if it does.  Otherwise it
3656  * will return -ENOSPC.
3657  */
3658 static int may_commit_transaction(struct btrfs_root *root,
3659                                   struct btrfs_space_info *space_info,
3660                                   u64 bytes, int force)
3661 {
3662         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3663         struct btrfs_trans_handle *trans;
3664
3665         trans = (struct btrfs_trans_handle *)current->journal_info;
3666         if (trans)
3667                 return -EAGAIN;
3668
3669         if (force)
3670                 goto commit;
3671
3672         /* See if there is enough pinned space to make this reservation */
3673         spin_lock(&space_info->lock);
3674         if (space_info->bytes_pinned >= bytes) {
3675                 spin_unlock(&space_info->lock);
3676                 goto commit;
3677         }
3678         spin_unlock(&space_info->lock);
3679
3680         /*
3681          * See if there is some space in the delayed insertion reservation for
3682          * this reservation.
3683          */
3684         if (space_info != delayed_rsv->space_info)
3685                 return -ENOSPC;
3686
3687         spin_lock(&space_info->lock);
3688         spin_lock(&delayed_rsv->lock);
3689         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3690                 spin_unlock(&delayed_rsv->lock);
3691                 spin_unlock(&space_info->lock);
3692                 return -ENOSPC;
3693         }
3694         spin_unlock(&delayed_rsv->lock);
3695         spin_unlock(&space_info->lock);
3696
3697 commit:
3698         trans = btrfs_join_transaction(root);
3699         if (IS_ERR(trans))
3700                 return -ENOSPC;
3701
3702         return btrfs_commit_transaction(trans, root);
3703 }
3704
3705 enum flush_state {
3706         FLUSH_DELALLOC          =       1,
3707         FLUSH_DELALLOC_WAIT     =       2,
3708         FLUSH_DELAYED_ITEMS_NR  =       3,
3709         FLUSH_DELAYED_ITEMS     =       4,
3710         COMMIT_TRANS            =       5,
3711 };
3712
3713 static int flush_space(struct btrfs_root *root,
3714                        struct btrfs_space_info *space_info, u64 num_bytes,
3715                        u64 orig_bytes, int state)
3716 {
3717         struct btrfs_trans_handle *trans;
3718         int nr;
3719         int ret = 0;
3720
3721         switch (state) {
3722         case FLUSH_DELALLOC:
3723         case FLUSH_DELALLOC_WAIT:
3724                 shrink_delalloc(root, num_bytes, orig_bytes,
3725                                 state == FLUSH_DELALLOC_WAIT);
3726                 break;
3727         case FLUSH_DELAYED_ITEMS_NR:
3728         case FLUSH_DELAYED_ITEMS:
3729                 if (state == FLUSH_DELAYED_ITEMS_NR) {
3730                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
3731
3732                         nr = (int)div64_u64(num_bytes, bytes);
3733                         if (!nr)
3734                                 nr = 1;
3735                         nr *= 2;
3736                 } else {
3737                         nr = -1;
3738                 }
3739                 trans = btrfs_join_transaction(root);
3740                 if (IS_ERR(trans)) {
3741                         ret = PTR_ERR(trans);
3742                         break;
3743                 }
3744                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
3745                 btrfs_end_transaction(trans, root);
3746                 break;
3747         case COMMIT_TRANS:
3748                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3749                 break;
3750         default:
3751                 ret = -ENOSPC;
3752                 break;
3753         }
3754
3755         return ret;
3756 }
3757 /**
3758  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3759  * @root - the root we're allocating for
3760  * @block_rsv - the block_rsv we're allocating for
3761  * @orig_bytes - the number of bytes we want
3762  * @flush - wether or not we can flush to make our reservation
3763  *
3764  * This will reserve orgi_bytes number of bytes from the space info associated
3765  * with the block_rsv.  If there is not enough space it will make an attempt to
3766  * flush out space to make room.  It will do this by flushing delalloc if
3767  * possible or committing the transaction.  If flush is 0 then no attempts to
3768  * regain reservations will be made and this will fail if there is not enough
3769  * space already.
3770  */
3771 static int reserve_metadata_bytes(struct btrfs_root *root,
3772                                   struct btrfs_block_rsv *block_rsv,
3773                                   u64 orig_bytes, int flush)
3774 {
3775         struct btrfs_space_info *space_info = block_rsv->space_info;
3776         u64 used;
3777         u64 num_bytes = orig_bytes;
3778         int flush_state = FLUSH_DELALLOC;
3779         int ret = 0;
3780         bool flushing = false;
3781         bool committed = false;
3782
3783 again:
3784         ret = 0;
3785         spin_lock(&space_info->lock);
3786         /*
3787          * We only want to wait if somebody other than us is flushing and we are
3788          * actually alloed to flush.
3789          */
3790         while (flush && !flushing && space_info->flush) {
3791                 spin_unlock(&space_info->lock);
3792                 /*
3793                  * If we have a trans handle we can't wait because the flusher
3794                  * may have to commit the transaction, which would mean we would
3795                  * deadlock since we are waiting for the flusher to finish, but
3796                  * hold the current transaction open.
3797                  */
3798                 if (current->journal_info)
3799                         return -EAGAIN;
3800                 ret = wait_event_killable(space_info->wait, !space_info->flush);
3801                 /* Must have been killed, return */
3802                 if (ret)
3803                         return -EINTR;
3804
3805                 spin_lock(&space_info->lock);
3806         }
3807
3808         ret = -ENOSPC;
3809         used = space_info->bytes_used + space_info->bytes_reserved +
3810                 space_info->bytes_pinned + space_info->bytes_readonly +
3811                 space_info->bytes_may_use;
3812
3813         /*
3814          * The idea here is that we've not already over-reserved the block group
3815          * then we can go ahead and save our reservation first and then start
3816          * flushing if we need to.  Otherwise if we've already overcommitted
3817          * lets start flushing stuff first and then come back and try to make
3818          * our reservation.
3819          */
3820         if (used <= space_info->total_bytes) {
3821                 if (used + orig_bytes <= space_info->total_bytes) {
3822                         space_info->bytes_may_use += orig_bytes;
3823                         trace_btrfs_space_reservation(root->fs_info,
3824                                 "space_info", space_info->flags, orig_bytes, 1);
3825                         ret = 0;
3826                 } else {
3827                         /*
3828                          * Ok set num_bytes to orig_bytes since we aren't
3829                          * overocmmitted, this way we only try and reclaim what
3830                          * we need.
3831                          */
3832                         num_bytes = orig_bytes;
3833                 }
3834         } else {
3835                 /*
3836                  * Ok we're over committed, set num_bytes to the overcommitted
3837                  * amount plus the amount of bytes that we need for this
3838                  * reservation.
3839                  */
3840                 num_bytes = used - space_info->total_bytes +
3841                         (orig_bytes * 2);
3842         }
3843
3844         if (ret) {
3845                 u64 profile = btrfs_get_alloc_profile(root, 0);
3846                 u64 avail;
3847
3848                 /*
3849                  * If we have a lot of space that's pinned, don't bother doing
3850                  * the overcommit dance yet and just commit the transaction.
3851                  */
3852                 avail = (space_info->total_bytes - space_info->bytes_used) * 8;
3853                 do_div(avail, 10);
3854                 if (space_info->bytes_pinned >= avail && flush && !committed) {
3855                         space_info->flush = 1;
3856                         flushing = true;
3857                         spin_unlock(&space_info->lock);
3858                         ret = may_commit_transaction(root, space_info,
3859                                                      orig_bytes, 1);
3860                         if (ret)
3861                                 goto out;
3862                         committed = true;
3863                         goto again;
3864                 }
3865
3866                 spin_lock(&root->fs_info->free_chunk_lock);
3867                 avail = root->fs_info->free_chunk_space;
3868
3869                 /*
3870                  * If we have dup, raid1 or raid10 then only half of the free
3871                  * space is actually useable.
3872                  */
3873                 if (profile & (BTRFS_BLOCK_GROUP_DUP |
3874                                BTRFS_BLOCK_GROUP_RAID1 |
3875                                BTRFS_BLOCK_GROUP_RAID10))
3876                         avail >>= 1;
3877
3878                 /*
3879                  * If we aren't flushing don't let us overcommit too much, say
3880                  * 1/8th of the space.  If we can flush, let it overcommit up to
3881                  * 1/2 of the space.
3882                  */
3883                 if (flush)
3884                         avail >>= 3;
3885                 else
3886                         avail >>= 1;
3887                  spin_unlock(&root->fs_info->free_chunk_lock);
3888
3889                 if (used + num_bytes < space_info->total_bytes + avail) {
3890                         space_info->bytes_may_use += orig_bytes;
3891                         trace_btrfs_space_reservation(root->fs_info,
3892                                 "space_info", space_info->flags, orig_bytes, 1);
3893                         ret = 0;
3894                 }
3895         }
3896
3897         /*
3898          * Couldn't make our reservation, save our place so while we're trying
3899          * to reclaim space we can actually use it instead of somebody else
3900          * stealing it from us.
3901          */
3902         if (ret && flush) {
3903                 flushing = true;
3904                 space_info->flush = 1;
3905         }
3906
3907         spin_unlock(&space_info->lock);
3908
3909         if (!ret || !flush)
3910                 goto out;
3911
3912         ret = flush_space(root, space_info, num_bytes, orig_bytes,
3913                           flush_state);
3914         flush_state++;
3915         if (!ret)
3916                 goto again;
3917         else if (flush_state <= COMMIT_TRANS)
3918                 goto again;
3919
3920 out:
3921         if (flushing) {
3922                 spin_lock(&space_info->lock);
3923                 space_info->flush = 0;
3924                 wake_up_all(&space_info->wait);
3925                 spin_unlock(&space_info->lock);
3926         }
3927         return ret;
3928 }
3929
3930 static struct btrfs_block_rsv *get_block_rsv(
3931                                         const struct btrfs_trans_handle *trans,
3932                                         const struct btrfs_root *root)
3933 {
3934         struct btrfs_block_rsv *block_rsv = NULL;
3935
3936         if (root->ref_cows)
3937                 block_rsv = trans->block_rsv;
3938
3939         if (root == root->fs_info->csum_root && trans->adding_csums)
3940                 block_rsv = trans->block_rsv;
3941
3942         if (!block_rsv)
3943                 block_rsv = root->block_rsv;
3944
3945         if (!block_rsv)
3946                 block_rsv = &root->fs_info->empty_block_rsv;
3947
3948         return block_rsv;
3949 }
3950
3951 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
3952                                u64 num_bytes)
3953 {
3954         int ret = -ENOSPC;
3955         spin_lock(&block_rsv->lock);
3956         if (block_rsv->reserved >= num_bytes) {
3957                 block_rsv->reserved -= num_bytes;
3958                 if (block_rsv->reserved < block_rsv->size)
3959                         block_rsv->full = 0;
3960                 ret = 0;
3961         }
3962         spin_unlock(&block_rsv->lock);
3963         return ret;
3964 }
3965
3966 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
3967                                 u64 num_bytes, int update_size)
3968 {
3969         spin_lock(&block_rsv->lock);
3970         block_rsv->reserved += num_bytes;
3971         if (update_size)
3972                 block_rsv->size += num_bytes;
3973         else if (block_rsv->reserved >= block_rsv->size)
3974                 block_rsv->full = 1;
3975         spin_unlock(&block_rsv->lock);
3976 }
3977
3978 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
3979                                     struct btrfs_block_rsv *block_rsv,
3980                                     struct btrfs_block_rsv *dest, u64 num_bytes)
3981 {
3982         struct btrfs_space_info *space_info = block_rsv->space_info;
3983
3984         spin_lock(&block_rsv->lock);
3985         if (num_bytes == (u64)-1)
3986                 num_bytes = block_rsv->size;
3987         block_rsv->size -= num_bytes;
3988         if (block_rsv->reserved >= block_rsv->size) {
3989                 num_bytes = block_rsv->reserved - block_rsv->size;
3990                 block_rsv->reserved = block_rsv->size;
3991                 block_rsv->full = 1;
3992         } else {
3993                 num_bytes = 0;
3994         }
3995         spin_unlock(&block_rsv->lock);
3996
3997         if (num_bytes > 0) {
3998                 if (dest) {
3999                         spin_lock(&dest->lock);
4000                         if (!dest->full) {
4001                                 u64 bytes_to_add;
4002
4003                                 bytes_to_add = dest->size - dest->reserved;
4004                                 bytes_to_add = min(num_bytes, bytes_to_add);
4005                                 dest->reserved += bytes_to_add;
4006                                 if (dest->reserved >= dest->size)
4007                                         dest->full = 1;
4008                                 num_bytes -= bytes_to_add;
4009                         }
4010                         spin_unlock(&dest->lock);
4011                 }
4012                 if (num_bytes) {
4013                         spin_lock(&space_info->lock);
4014                         space_info->bytes_may_use -= num_bytes;
4015                         trace_btrfs_space_reservation(fs_info, "space_info",
4016                                         space_info->flags, num_bytes, 0);
4017                         space_info->reservation_progress++;
4018                         spin_unlock(&space_info->lock);
4019                 }
4020         }
4021 }
4022
4023 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4024                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4025 {
4026         int ret;
4027
4028         ret = block_rsv_use_bytes(src, num_bytes);
4029         if (ret)
4030                 return ret;
4031
4032         block_rsv_add_bytes(dst, num_bytes, 1);
4033         return 0;
4034 }
4035
4036 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv)
4037 {
4038         memset(rsv, 0, sizeof(*rsv));
4039         spin_lock_init(&rsv->lock);
4040 }
4041
4042 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root)
4043 {
4044         struct btrfs_block_rsv *block_rsv;
4045         struct btrfs_fs_info *fs_info = root->fs_info;
4046
4047         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4048         if (!block_rsv)
4049                 return NULL;
4050
4051         btrfs_init_block_rsv(block_rsv);
4052         block_rsv->space_info = __find_space_info(fs_info,
4053                                                   BTRFS_BLOCK_GROUP_METADATA);
4054         return block_rsv;
4055 }
4056
4057 void btrfs_free_block_rsv(struct btrfs_root *root,
4058                           struct btrfs_block_rsv *rsv)
4059 {
4060         btrfs_block_rsv_release(root, rsv, (u64)-1);
4061         kfree(rsv);
4062 }
4063
4064 static inline int __block_rsv_add(struct btrfs_root *root,
4065                                   struct btrfs_block_rsv *block_rsv,
4066                                   u64 num_bytes, int flush)
4067 {
4068         int ret;
4069
4070         if (num_bytes == 0)
4071                 return 0;
4072
4073         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4074         if (!ret) {
4075                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4076                 return 0;
4077         }
4078
4079         return ret;
4080 }
4081
4082 int btrfs_block_rsv_add(struct btrfs_root *root,
4083                         struct btrfs_block_rsv *block_rsv,
4084                         u64 num_bytes)
4085 {
4086         return __block_rsv_add(root, block_rsv, num_bytes, 1);
4087 }
4088
4089 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
4090                                 struct btrfs_block_rsv *block_rsv,
4091                                 u64 num_bytes)
4092 {
4093         return __block_rsv_add(root, block_rsv, num_bytes, 0);
4094 }
4095
4096 int btrfs_block_rsv_check(struct btrfs_root *root,
4097                           struct btrfs_block_rsv *block_rsv, int min_factor)
4098 {
4099         u64 num_bytes = 0;
4100         int ret = -ENOSPC;
4101
4102         if (!block_rsv)
4103                 return 0;
4104
4105         spin_lock(&block_rsv->lock);
4106         num_bytes = div_factor(block_rsv->size, min_factor);
4107         if (block_rsv->reserved >= num_bytes)
4108                 ret = 0;
4109         spin_unlock(&block_rsv->lock);
4110
4111         return ret;
4112 }
4113
4114 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
4115                                            struct btrfs_block_rsv *block_rsv,
4116                                            u64 min_reserved, int flush)
4117 {
4118         u64 num_bytes = 0;
4119         int ret = -ENOSPC;
4120
4121         if (!block_rsv)
4122                 return 0;
4123
4124         spin_lock(&block_rsv->lock);
4125         num_bytes = min_reserved;
4126         if (block_rsv->reserved >= num_bytes)
4127                 ret = 0;
4128         else
4129                 num_bytes -= block_rsv->reserved;
4130         spin_unlock(&block_rsv->lock);
4131
4132         if (!ret)
4133                 return 0;
4134
4135         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4136         if (!ret) {
4137                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4138                 return 0;
4139         }
4140
4141         return ret;
4142 }
4143
4144 int btrfs_block_rsv_refill(struct btrfs_root *root,
4145                            struct btrfs_block_rsv *block_rsv,
4146                            u64 min_reserved)
4147 {
4148         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
4149 }
4150
4151 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
4152                                    struct btrfs_block_rsv *block_rsv,
4153                                    u64 min_reserved)
4154 {
4155         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
4156 }
4157
4158 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4159                             struct btrfs_block_rsv *dst_rsv,
4160                             u64 num_bytes)
4161 {
4162         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4163 }
4164
4165 void btrfs_block_rsv_release(struct btrfs_root *root,
4166                              struct btrfs_block_rsv *block_rsv,
4167                              u64 num_bytes)
4168 {
4169         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4170         if (global_rsv->full || global_rsv == block_rsv ||
4171             block_rsv->space_info != global_rsv->space_info)
4172                 global_rsv = NULL;
4173         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4174                                 num_bytes);
4175 }
4176
4177 /*
4178  * helper to calculate size of global block reservation.
4179  * the desired value is sum of space used by extent tree,
4180  * checksum tree and root tree
4181  */
4182 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4183 {
4184         struct btrfs_space_info *sinfo;
4185         u64 num_bytes;
4186         u64 meta_used;
4187         u64 data_used;
4188         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4189
4190         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4191         spin_lock(&sinfo->lock);
4192         data_used = sinfo->bytes_used;
4193         spin_unlock(&sinfo->lock);
4194
4195         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4196         spin_lock(&sinfo->lock);
4197         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4198                 data_used = 0;
4199         meta_used = sinfo->bytes_used;
4200         spin_unlock(&sinfo->lock);
4201
4202         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4203                     csum_size * 2;
4204         num_bytes += div64_u64(data_used + meta_used, 50);
4205
4206         if (num_bytes * 3 > meta_used)
4207                 num_bytes = div64_u64(meta_used, 3);
4208
4209         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4210 }
4211
4212 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4213 {
4214         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4215         struct btrfs_space_info *sinfo = block_rsv->space_info;
4216         u64 num_bytes;
4217
4218         num_bytes = calc_global_metadata_size(fs_info);
4219
4220         spin_lock(&sinfo->lock);
4221         spin_lock(&block_rsv->lock);
4222
4223         block_rsv->size = num_bytes;
4224
4225         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4226                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4227                     sinfo->bytes_may_use;
4228
4229         if (sinfo->total_bytes > num_bytes) {
4230                 num_bytes = sinfo->total_bytes - num_bytes;
4231                 block_rsv->reserved += num_bytes;
4232                 sinfo->bytes_may_use += num_bytes;
4233                 trace_btrfs_space_reservation(fs_info, "space_info",
4234                                       sinfo->flags, num_bytes, 1);
4235         }
4236
4237         if (block_rsv->reserved >= block_rsv->size) {
4238                 num_bytes = block_rsv->reserved - block_rsv->size;
4239                 sinfo->bytes_may_use -= num_bytes;
4240                 trace_btrfs_space_reservation(fs_info, "space_info",
4241                                       sinfo->flags, num_bytes, 0);
4242                 sinfo->reservation_progress++;
4243                 block_rsv->reserved = block_rsv->size;
4244                 block_rsv->full = 1;
4245         }
4246
4247         spin_unlock(&block_rsv->lock);
4248         spin_unlock(&sinfo->lock);
4249 }
4250
4251 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4252 {
4253         struct btrfs_space_info *space_info;
4254
4255         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4256         fs_info->chunk_block_rsv.space_info = space_info;
4257
4258         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4259         fs_info->global_block_rsv.space_info = space_info;
4260         fs_info->delalloc_block_rsv.space_info = space_info;
4261         fs_info->trans_block_rsv.space_info = space_info;
4262         fs_info->empty_block_rsv.space_info = space_info;
4263         fs_info->delayed_block_rsv.space_info = space_info;
4264
4265         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4266         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4267         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4268         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4269         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4270
4271         update_global_block_rsv(fs_info);
4272 }
4273
4274 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4275 {
4276         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4277                                 (u64)-1);
4278         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4279         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4280         WARN_ON(fs_info->trans_block_rsv.size > 0);
4281         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4282         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4283         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4284         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4285         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4286 }
4287
4288 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4289                                   struct btrfs_root *root)
4290 {
4291         if (!trans->block_rsv)
4292                 return;
4293
4294         if (!trans->bytes_reserved)
4295                 return;
4296
4297         trace_btrfs_space_reservation(root->fs_info, "transaction",
4298                                       trans->transid, trans->bytes_reserved, 0);
4299         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4300         trans->bytes_reserved = 0;
4301 }
4302
4303 /* Can only return 0 or -ENOSPC */
4304 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4305                                   struct inode *inode)
4306 {
4307         struct btrfs_root *root = BTRFS_I(inode)->root;
4308         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4309         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4310
4311         /*
4312          * We need to hold space in order to delete our orphan item once we've
4313          * added it, so this takes the reservation so we can release it later
4314          * when we are truly done with the orphan item.
4315          */
4316         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4317         trace_btrfs_space_reservation(root->fs_info, "orphan",
4318                                       btrfs_ino(inode), num_bytes, 1);
4319         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4320 }
4321
4322 void btrfs_orphan_release_metadata(struct inode *inode)
4323 {
4324         struct btrfs_root *root = BTRFS_I(inode)->root;
4325         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4326         trace_btrfs_space_reservation(root->fs_info, "orphan",
4327                                       btrfs_ino(inode), num_bytes, 0);
4328         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4329 }
4330
4331 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4332                                 struct btrfs_pending_snapshot *pending)
4333 {
4334         struct btrfs_root *root = pending->root;
4335         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4336         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4337         /*
4338          * two for root back/forward refs, two for directory entries
4339          * and one for root of the snapshot.
4340          */
4341         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 5);
4342         dst_rsv->space_info = src_rsv->space_info;
4343         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4344 }
4345
4346 /**
4347  * drop_outstanding_extent - drop an outstanding extent
4348  * @inode: the inode we're dropping the extent for
4349  *
4350  * This is called when we are freeing up an outstanding extent, either called
4351  * after an error or after an extent is written.  This will return the number of
4352  * reserved extents that need to be freed.  This must be called with
4353  * BTRFS_I(inode)->lock held.
4354  */
4355 static unsigned drop_outstanding_extent(struct inode *inode)
4356 {
4357         unsigned drop_inode_space = 0;
4358         unsigned dropped_extents = 0;
4359
4360         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4361         BTRFS_I(inode)->outstanding_extents--;
4362
4363         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4364             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4365                                &BTRFS_I(inode)->runtime_flags))
4366                 drop_inode_space = 1;
4367
4368         /*
4369          * If we have more or the same amount of outsanding extents than we have
4370          * reserved then we need to leave the reserved extents count alone.
4371          */
4372         if (BTRFS_I(inode)->outstanding_extents >=
4373             BTRFS_I(inode)->reserved_extents)
4374                 return drop_inode_space;
4375
4376         dropped_extents = BTRFS_I(inode)->reserved_extents -
4377                 BTRFS_I(inode)->outstanding_extents;
4378         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4379         return dropped_extents + drop_inode_space;
4380 }
4381
4382 /**
4383  * calc_csum_metadata_size - return the amount of metada space that must be
4384  *      reserved/free'd for the given bytes.
4385  * @inode: the inode we're manipulating
4386  * @num_bytes: the number of bytes in question
4387  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4388  *
4389  * This adjusts the number of csum_bytes in the inode and then returns the
4390  * correct amount of metadata that must either be reserved or freed.  We
4391  * calculate how many checksums we can fit into one leaf and then divide the
4392  * number of bytes that will need to be checksumed by this value to figure out
4393  * how many checksums will be required.  If we are adding bytes then the number
4394  * may go up and we will return the number of additional bytes that must be
4395  * reserved.  If it is going down we will return the number of bytes that must
4396  * be freed.
4397  *
4398  * This must be called with BTRFS_I(inode)->lock held.
4399  */
4400 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4401                                    int reserve)
4402 {
4403         struct btrfs_root *root = BTRFS_I(inode)->root;
4404         u64 csum_size;
4405         int num_csums_per_leaf;
4406         int num_csums;
4407         int old_csums;
4408
4409         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4410             BTRFS_I(inode)->csum_bytes == 0)
4411                 return 0;
4412
4413         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4414         if (reserve)
4415                 BTRFS_I(inode)->csum_bytes += num_bytes;
4416         else
4417                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4418         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4419         num_csums_per_leaf = (int)div64_u64(csum_size,
4420                                             sizeof(struct btrfs_csum_item) +
4421                                             sizeof(struct btrfs_disk_key));
4422         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4423         num_csums = num_csums + num_csums_per_leaf - 1;
4424         num_csums = num_csums / num_csums_per_leaf;
4425
4426         old_csums = old_csums + num_csums_per_leaf - 1;
4427         old_csums = old_csums / num_csums_per_leaf;
4428
4429         /* No change, no need to reserve more */
4430         if (old_csums == num_csums)
4431                 return 0;
4432
4433         if (reserve)
4434                 return btrfs_calc_trans_metadata_size(root,
4435                                                       num_csums - old_csums);
4436
4437         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4438 }
4439
4440 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4441 {
4442         struct btrfs_root *root = BTRFS_I(inode)->root;
4443         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4444         u64 to_reserve = 0;
4445         u64 csum_bytes;
4446         unsigned nr_extents = 0;
4447         int extra_reserve = 0;
4448         int flush = 1;
4449         int ret;
4450
4451         /* Need to be holding the i_mutex here if we aren't free space cache */
4452         if (btrfs_is_free_space_inode(inode))
4453                 flush = 0;
4454
4455         if (flush && btrfs_transaction_in_commit(root->fs_info))
4456                 schedule_timeout(1);
4457
4458         mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4459         num_bytes = ALIGN(num_bytes, root->sectorsize);
4460
4461         spin_lock(&BTRFS_I(inode)->lock);
4462         BTRFS_I(inode)->outstanding_extents++;
4463
4464         if (BTRFS_I(inode)->outstanding_extents >
4465             BTRFS_I(inode)->reserved_extents)
4466                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4467                         BTRFS_I(inode)->reserved_extents;
4468
4469         /*
4470          * Add an item to reserve for updating the inode when we complete the
4471          * delalloc io.
4472          */
4473         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4474                       &BTRFS_I(inode)->runtime_flags)) {
4475                 nr_extents++;
4476                 extra_reserve = 1;
4477         }
4478
4479         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4480         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4481         csum_bytes = BTRFS_I(inode)->csum_bytes;
4482         spin_unlock(&BTRFS_I(inode)->lock);
4483
4484         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4485         if (ret) {
4486                 u64 to_free = 0;
4487                 unsigned dropped;
4488
4489                 spin_lock(&BTRFS_I(inode)->lock);
4490                 dropped = drop_outstanding_extent(inode);
4491                 /*
4492                  * If the inodes csum_bytes is the same as the original
4493                  * csum_bytes then we know we haven't raced with any free()ers
4494                  * so we can just reduce our inodes csum bytes and carry on.
4495                  * Otherwise we have to do the normal free thing to account for
4496                  * the case that the free side didn't free up its reserve
4497                  * because of this outstanding reservation.
4498                  */
4499                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4500                         calc_csum_metadata_size(inode, num_bytes, 0);
4501                 else
4502                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4503                 spin_unlock(&BTRFS_I(inode)->lock);
4504                 if (dropped)
4505                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4506
4507                 if (to_free) {
4508                         btrfs_block_rsv_release(root, block_rsv, to_free);
4509                         trace_btrfs_space_reservation(root->fs_info,
4510                                                       "delalloc",
4511                                                       btrfs_ino(inode),
4512                                                       to_free, 0);
4513                 }
4514                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4515                 return ret;
4516         }
4517
4518         spin_lock(&BTRFS_I(inode)->lock);
4519         if (extra_reserve) {
4520                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4521                         &BTRFS_I(inode)->runtime_flags);
4522                 nr_extents--;
4523         }
4524         BTRFS_I(inode)->reserved_extents += nr_extents;
4525         spin_unlock(&BTRFS_I(inode)->lock);
4526         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4527
4528         if (to_reserve)
4529                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4530                                               btrfs_ino(inode), to_reserve, 1);
4531         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4532
4533         return 0;
4534 }
4535
4536 /**
4537  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4538  * @inode: the inode to release the reservation for
4539  * @num_bytes: the number of bytes we're releasing
4540  *
4541  * This will release the metadata reservation for an inode.  This can be called
4542  * once we complete IO for a given set of bytes to release their metadata
4543  * reservations.
4544  */
4545 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4546 {
4547         struct btrfs_root *root = BTRFS_I(inode)->root;
4548         u64 to_free = 0;
4549         unsigned dropped;
4550
4551         num_bytes = ALIGN(num_bytes, root->sectorsize);
4552         spin_lock(&BTRFS_I(inode)->lock);
4553         dropped = drop_outstanding_extent(inode);
4554
4555         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4556         spin_unlock(&BTRFS_I(inode)->lock);
4557         if (dropped > 0)
4558                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4559
4560         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4561                                       btrfs_ino(inode), to_free, 0);
4562         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4563                                 to_free);
4564 }
4565
4566 /**
4567  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4568  * @inode: inode we're writing to
4569  * @num_bytes: the number of bytes we want to allocate
4570  *
4571  * This will do the following things
4572  *
4573  * o reserve space in the data space info for num_bytes
4574  * o reserve space in the metadata space info based on number of outstanding
4575  *   extents and how much csums will be needed
4576  * o add to the inodes ->delalloc_bytes
4577  * o add it to the fs_info's delalloc inodes list.
4578  *
4579  * This will return 0 for success and -ENOSPC if there is no space left.
4580  */
4581 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4582 {
4583         int ret;
4584
4585         ret = btrfs_check_data_free_space(inode, num_bytes);
4586         if (ret)
4587                 return ret;
4588
4589         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4590         if (ret) {
4591                 btrfs_free_reserved_data_space(inode, num_bytes);
4592                 return ret;
4593         }
4594
4595         return 0;
4596 }
4597
4598 /**
4599  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4600  * @inode: inode we're releasing space for
4601  * @num_bytes: the number of bytes we want to free up
4602  *
4603  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4604  * called in the case that we don't need the metadata AND data reservations
4605  * anymore.  So if there is an error or we insert an inline extent.
4606  *
4607  * This function will release the metadata space that was not used and will
4608  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4609  * list if there are no delalloc bytes left.
4610  */
4611 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4612 {
4613         btrfs_delalloc_release_metadata(inode, num_bytes);
4614         btrfs_free_reserved_data_space(inode, num_bytes);
4615 }
4616
4617 static int update_block_group(struct btrfs_trans_handle *trans,
4618                               struct btrfs_root *root,
4619                               u64 bytenr, u64 num_bytes, int alloc)
4620 {
4621         struct btrfs_block_group_cache *cache = NULL;
4622         struct btrfs_fs_info *info = root->fs_info;
4623         u64 total = num_bytes;
4624         u64 old_val;
4625         u64 byte_in_group;
4626         int factor;
4627
4628         /* block accounting for super block */
4629         spin_lock(&info->delalloc_lock);
4630         old_val = btrfs_super_bytes_used(info->super_copy);
4631         if (alloc)
4632                 old_val += num_bytes;
4633         else
4634                 old_val -= num_bytes;
4635         btrfs_set_super_bytes_used(info->super_copy, old_val);
4636         spin_unlock(&info->delalloc_lock);
4637
4638         while (total) {
4639                 cache = btrfs_lookup_block_group(info, bytenr);
4640                 if (!cache)
4641                         return -ENOENT;
4642                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4643                                     BTRFS_BLOCK_GROUP_RAID1 |
4644                                     BTRFS_BLOCK_GROUP_RAID10))
4645                         factor = 2;
4646                 else
4647                         factor = 1;
4648                 /*
4649                  * If this block group has free space cache written out, we
4650                  * need to make sure to load it if we are removing space.  This
4651                  * is because we need the unpinning stage to actually add the
4652                  * space back to the block group, otherwise we will leak space.
4653                  */
4654                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4655                         cache_block_group(cache, trans, NULL, 1);
4656
4657                 byte_in_group = bytenr - cache->key.objectid;
4658                 WARN_ON(byte_in_group > cache->key.offset);
4659
4660                 spin_lock(&cache->space_info->lock);
4661                 spin_lock(&cache->lock);
4662
4663                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4664                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4665                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4666
4667                 cache->dirty = 1;
4668                 old_val = btrfs_block_group_used(&cache->item);
4669                 num_bytes = min(total, cache->key.offset - byte_in_group);
4670                 if (alloc) {
4671                         old_val += num_bytes;
4672                         btrfs_set_block_group_used(&cache->item, old_val);
4673                         cache->reserved -= num_bytes;
4674                         cache->space_info->bytes_reserved -= num_bytes;
4675                         cache->space_info->bytes_used += num_bytes;
4676                         cache->space_info->disk_used += num_bytes * factor;
4677                         spin_unlock(&cache->lock);
4678                         spin_unlock(&cache->space_info->lock);
4679                 } else {
4680                         old_val -= num_bytes;
4681                         btrfs_set_block_group_used(&cache->item, old_val);
4682                         cache->pinned += num_bytes;
4683                         cache->space_info->bytes_pinned += num_bytes;
4684                         cache->space_info->bytes_used -= num_bytes;
4685                         cache->space_info->disk_used -= num_bytes * factor;
4686                         spin_unlock(&cache->lock);
4687                         spin_unlock(&cache->space_info->lock);
4688
4689                         set_extent_dirty(info->pinned_extents,
4690                                          bytenr, bytenr + num_bytes - 1,
4691                                          GFP_NOFS | __GFP_NOFAIL);
4692                 }
4693                 btrfs_put_block_group(cache);
4694                 total -= num_bytes;
4695                 bytenr += num_bytes;
4696         }
4697         return 0;
4698 }
4699
4700 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4701 {
4702         struct btrfs_block_group_cache *cache;
4703         u64 bytenr;
4704
4705         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4706         if (!cache)
4707                 return 0;
4708
4709         bytenr = cache->key.objectid;
4710         btrfs_put_block_group(cache);
4711
4712         return bytenr;
4713 }
4714
4715 static int pin_down_extent(struct btrfs_root *root,
4716                            struct btrfs_block_group_cache *cache,
4717                            u64 bytenr, u64 num_bytes, int reserved)
4718 {
4719         spin_lock(&cache->space_info->lock);
4720         spin_lock(&cache->lock);
4721         cache->pinned += num_bytes;
4722         cache->space_info->bytes_pinned += num_bytes;
4723         if (reserved) {
4724                 cache->reserved -= num_bytes;
4725                 cache->space_info->bytes_reserved -= num_bytes;
4726         }
4727         spin_unlock(&cache->lock);
4728         spin_unlock(&cache->space_info->lock);
4729
4730         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4731                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4732         return 0;
4733 }
4734
4735 /*
4736  * this function must be called within transaction
4737  */
4738 int btrfs_pin_extent(struct btrfs_root *root,
4739                      u64 bytenr, u64 num_bytes, int reserved)
4740 {
4741         struct btrfs_block_group_cache *cache;
4742
4743         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4744         BUG_ON(!cache); /* Logic error */
4745
4746         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4747
4748         btrfs_put_block_group(cache);
4749         return 0;
4750 }
4751
4752 /*
4753  * this function must be called within transaction
4754  */
4755 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4756                                     struct btrfs_root *root,
4757                                     u64 bytenr, u64 num_bytes)
4758 {
4759         struct btrfs_block_group_cache *cache;
4760
4761         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4762         BUG_ON(!cache); /* Logic error */
4763
4764         /*
4765          * pull in the free space cache (if any) so that our pin
4766          * removes the free space from the cache.  We have load_only set
4767          * to one because the slow code to read in the free extents does check
4768          * the pinned extents.
4769          */
4770         cache_block_group(cache, trans, root, 1);
4771
4772         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4773
4774         /* remove us from the free space cache (if we're there at all) */
4775         btrfs_remove_free_space(cache, bytenr, num_bytes);
4776         btrfs_put_block_group(cache);
4777         return 0;
4778 }
4779
4780 /**
4781  * btrfs_update_reserved_bytes - update the block_group and space info counters
4782  * @cache:      The cache we are manipulating
4783  * @num_bytes:  The number of bytes in question
4784  * @reserve:    One of the reservation enums
4785  *
4786  * This is called by the allocator when it reserves space, or by somebody who is
4787  * freeing space that was never actually used on disk.  For example if you
4788  * reserve some space for a new leaf in transaction A and before transaction A
4789  * commits you free that leaf, you call this with reserve set to 0 in order to
4790  * clear the reservation.
4791  *
4792  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4793  * ENOSPC accounting.  For data we handle the reservation through clearing the
4794  * delalloc bits in the io_tree.  We have to do this since we could end up
4795  * allocating less disk space for the amount of data we have reserved in the
4796  * case of compression.
4797  *
4798  * If this is a reservation and the block group has become read only we cannot
4799  * make the reservation and return -EAGAIN, otherwise this function always
4800  * succeeds.
4801  */
4802 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4803                                        u64 num_bytes, int reserve)
4804 {
4805         struct btrfs_space_info *space_info = cache->space_info;
4806         int ret = 0;
4807
4808         spin_lock(&space_info->lock);
4809         spin_lock(&cache->lock);
4810         if (reserve != RESERVE_FREE) {
4811                 if (cache->ro) {
4812                         ret = -EAGAIN;
4813                 } else {
4814                         cache->reserved += num_bytes;
4815                         space_info->bytes_reserved += num_bytes;
4816                         if (reserve == RESERVE_ALLOC) {
4817                                 trace_btrfs_space_reservation(cache->fs_info,
4818                                                 "space_info", space_info->flags,
4819                                                 num_bytes, 0);
4820                                 space_info->bytes_may_use -= num_bytes;
4821                         }
4822                 }
4823         } else {
4824                 if (cache->ro)
4825                         space_info->bytes_readonly += num_bytes;
4826                 cache->reserved -= num_bytes;
4827                 space_info->bytes_reserved -= num_bytes;
4828                 space_info->reservation_progress++;
4829         }
4830         spin_unlock(&cache->lock);
4831         spin_unlock(&space_info->lock);
4832         return ret;
4833 }
4834
4835 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4836                                 struct btrfs_root *root)
4837 {
4838         struct btrfs_fs_info *fs_info = root->fs_info;
4839         struct btrfs_caching_control *next;
4840         struct btrfs_caching_control *caching_ctl;
4841         struct btrfs_block_group_cache *cache;
4842
4843         down_write(&fs_info->extent_commit_sem);
4844
4845         list_for_each_entry_safe(caching_ctl, next,
4846                                  &fs_info->caching_block_groups, list) {
4847                 cache = caching_ctl->block_group;
4848                 if (block_group_cache_done(cache)) {
4849                         cache->last_byte_to_unpin = (u64)-1;
4850                         list_del_init(&caching_ctl->list);
4851                         put_caching_control(caching_ctl);
4852                 } else {
4853                         cache->last_byte_to_unpin = caching_ctl->progress;
4854                 }
4855         }
4856
4857         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4858                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4859         else
4860                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4861
4862         up_write(&fs_info->extent_commit_sem);
4863
4864         update_global_block_rsv(fs_info);
4865 }
4866
4867 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4868 {
4869         struct btrfs_fs_info *fs_info = root->fs_info;
4870         struct btrfs_block_group_cache *cache = NULL;
4871         u64 len;
4872
4873         while (start <= end) {
4874                 if (!cache ||
4875                     start >= cache->key.objectid + cache->key.offset) {
4876                         if (cache)
4877                                 btrfs_put_block_group(cache);
4878                         cache = btrfs_lookup_block_group(fs_info, start);
4879                         BUG_ON(!cache); /* Logic error */
4880                 }
4881
4882                 len = cache->key.objectid + cache->key.offset - start;
4883                 len = min(len, end + 1 - start);
4884
4885                 if (start < cache->last_byte_to_unpin) {
4886                         len = min(len, cache->last_byte_to_unpin - start);
4887                         btrfs_add_free_space(cache, start, len);
4888                 }
4889
4890                 start += len;
4891
4892                 spin_lock(&cache->space_info->lock);
4893                 spin_lock(&cache->lock);
4894                 cache->pinned -= len;
4895                 cache->space_info->bytes_pinned -= len;
4896                 if (cache->ro)
4897                         cache->space_info->bytes_readonly += len;
4898                 spin_unlock(&cache->lock);
4899                 spin_unlock(&cache->space_info->lock);
4900         }
4901
4902         if (cache)
4903                 btrfs_put_block_group(cache);
4904         return 0;
4905 }
4906
4907 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
4908                                struct btrfs_root *root)
4909 {
4910         struct btrfs_fs_info *fs_info = root->fs_info;
4911         struct extent_io_tree *unpin;
4912         u64 start;
4913         u64 end;
4914         int ret;
4915
4916         if (trans->aborted)
4917                 return 0;
4918
4919         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4920                 unpin = &fs_info->freed_extents[1];
4921         else
4922                 unpin = &fs_info->freed_extents[0];
4923
4924         while (1) {
4925                 ret = find_first_extent_bit(unpin, 0, &start, &end,
4926                                             EXTENT_DIRTY);
4927                 if (ret)
4928                         break;
4929
4930                 if (btrfs_test_opt(root, DISCARD))
4931                         ret = btrfs_discard_extent(root, start,
4932                                                    end + 1 - start, NULL);
4933
4934                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
4935                 unpin_extent_range(root, start, end);
4936                 cond_resched();
4937         }
4938
4939         return 0;
4940 }
4941
4942 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
4943                                 struct btrfs_root *root,
4944                                 u64 bytenr, u64 num_bytes, u64 parent,
4945                                 u64 root_objectid, u64 owner_objectid,
4946                                 u64 owner_offset, int refs_to_drop,
4947                                 struct btrfs_delayed_extent_op *extent_op)
4948 {
4949         struct btrfs_key key;
4950         struct btrfs_path *path;
4951         struct btrfs_fs_info *info = root->fs_info;
4952         struct btrfs_root *extent_root = info->extent_root;
4953         struct extent_buffer *leaf;
4954         struct btrfs_extent_item *ei;
4955         struct btrfs_extent_inline_ref *iref;
4956         int ret;
4957         int is_data;
4958         int extent_slot = 0;
4959         int found_extent = 0;
4960         int num_to_del = 1;
4961         u32 item_size;
4962         u64 refs;
4963
4964         path = btrfs_alloc_path();
4965         if (!path)
4966                 return -ENOMEM;
4967
4968         path->reada = 1;
4969         path->leave_spinning = 1;
4970
4971         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
4972         BUG_ON(!is_data && refs_to_drop != 1);
4973
4974         ret = lookup_extent_backref(trans, extent_root, path, &iref,
4975                                     bytenr, num_bytes, parent,
4976                                     root_objectid, owner_objectid,
4977                                     owner_offset);
4978         if (ret == 0) {
4979                 extent_slot = path->slots[0];
4980                 while (extent_slot >= 0) {
4981                         btrfs_item_key_to_cpu(path->nodes[0], &key,
4982                                               extent_slot);
4983                         if (key.objectid != bytenr)
4984                                 break;
4985                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
4986                             key.offset == num_bytes) {
4987                                 found_extent = 1;
4988                                 break;
4989                         }
4990                         if (path->slots[0] - extent_slot > 5)
4991                                 break;
4992                         extent_slot--;
4993                 }
4994 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
4995                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
4996                 if (found_extent && item_size < sizeof(*ei))
4997                         found_extent = 0;
4998 #endif
4999                 if (!found_extent) {
5000                         BUG_ON(iref);
5001                         ret = remove_extent_backref(trans, extent_root, path,
5002                                                     NULL, refs_to_drop,
5003                                                     is_data);
5004                         if (ret)
5005                                 goto abort;
5006                         btrfs_release_path(path);
5007                         path->leave_spinning = 1;
5008
5009                         key.objectid = bytenr;
5010                         key.type = BTRFS_EXTENT_ITEM_KEY;
5011                         key.offset = num_bytes;
5012
5013                         ret = btrfs_search_slot(trans, extent_root,
5014                                                 &key, path, -1, 1);
5015                         if (ret) {
5016                                 printk(KERN_ERR "umm, got %d back from search"
5017                                        ", was looking for %llu\n", ret,
5018                                        (unsigned long long)bytenr);
5019                                 if (ret > 0)
5020                                         btrfs_print_leaf(extent_root,
5021                                                          path->nodes[0]);
5022                         }
5023                         if (ret < 0)
5024                                 goto abort;
5025                         extent_slot = path->slots[0];
5026                 }
5027         } else if (ret == -ENOENT) {
5028                 btrfs_print_leaf(extent_root, path->nodes[0]);
5029                 WARN_ON(1);
5030                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
5031                        "parent %llu root %llu  owner %llu offset %llu\n",
5032                        (unsigned long long)bytenr,
5033                        (unsigned long long)parent,
5034                        (unsigned long long)root_objectid,
5035                        (unsigned long long)owner_objectid,
5036                        (unsigned long long)owner_offset);
5037         } else {
5038                 goto abort;
5039         }
5040
5041         leaf = path->nodes[0];
5042         item_size = btrfs_item_size_nr(leaf, extent_slot);
5043 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5044         if (item_size < sizeof(*ei)) {
5045                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5046                 ret = convert_extent_item_v0(trans, extent_root, path,
5047                                              owner_objectid, 0);
5048                 if (ret < 0)
5049                         goto abort;
5050
5051                 btrfs_release_path(path);
5052                 path->leave_spinning = 1;
5053
5054                 key.objectid = bytenr;
5055                 key.type = BTRFS_EXTENT_ITEM_KEY;
5056                 key.offset = num_bytes;
5057
5058                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5059                                         -1, 1);
5060                 if (ret) {
5061                         printk(KERN_ERR "umm, got %d back from search"
5062                                ", was looking for %llu\n", ret,
5063                                (unsigned long long)bytenr);
5064                         btrfs_print_leaf(extent_root, path->nodes[0]);
5065                 }
5066                 if (ret < 0)
5067                         goto abort;
5068                 extent_slot = path->slots[0];
5069                 leaf = path->nodes[0];
5070                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5071         }
5072 #endif
5073         BUG_ON(item_size < sizeof(*ei));
5074         ei = btrfs_item_ptr(leaf, extent_slot,
5075                             struct btrfs_extent_item);
5076         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5077                 struct btrfs_tree_block_info *bi;
5078                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5079                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5080                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5081         }
5082
5083         refs = btrfs_extent_refs(leaf, ei);
5084         BUG_ON(refs < refs_to_drop);
5085         refs -= refs_to_drop;
5086
5087         if (refs > 0) {
5088                 if (extent_op)
5089                         __run_delayed_extent_op(extent_op, leaf, ei);
5090                 /*
5091                  * In the case of inline back ref, reference count will
5092                  * be updated by remove_extent_backref
5093                  */
5094                 if (iref) {
5095                         BUG_ON(!found_extent);
5096                 } else {
5097                         btrfs_set_extent_refs(leaf, ei, refs);
5098                         btrfs_mark_buffer_dirty(leaf);
5099                 }
5100                 if (found_extent) {
5101                         ret = remove_extent_backref(trans, extent_root, path,
5102                                                     iref, refs_to_drop,
5103                                                     is_data);
5104                         if (ret)
5105                                 goto abort;
5106                 }
5107         } else {
5108                 if (found_extent) {
5109                         BUG_ON(is_data && refs_to_drop !=
5110                                extent_data_ref_count(root, path, iref));
5111                         if (iref) {
5112                                 BUG_ON(path->slots[0] != extent_slot);
5113                         } else {
5114                                 BUG_ON(path->slots[0] != extent_slot + 1);
5115                                 path->slots[0] = extent_slot;
5116                                 num_to_del = 2;
5117                         }
5118                 }
5119
5120                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5121                                       num_to_del);
5122                 if (ret)
5123                         goto abort;
5124                 btrfs_release_path(path);
5125
5126                 if (is_data) {
5127                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5128                         if (ret)
5129                                 goto abort;
5130                 }
5131
5132                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
5133                 if (ret)
5134                         goto abort;
5135         }
5136 out:
5137         btrfs_free_path(path);
5138         return ret;
5139
5140 abort:
5141         btrfs_abort_transaction(trans, extent_root, ret);
5142         goto out;
5143 }
5144
5145 /*
5146  * when we free an block, it is possible (and likely) that we free the last
5147  * delayed ref for that extent as well.  This searches the delayed ref tree for
5148  * a given extent, and if there are no other delayed refs to be processed, it
5149  * removes it from the tree.
5150  */
5151 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5152                                       struct btrfs_root *root, u64 bytenr)
5153 {
5154         struct btrfs_delayed_ref_head *head;
5155         struct btrfs_delayed_ref_root *delayed_refs;
5156         struct btrfs_delayed_ref_node *ref;
5157         struct rb_node *node;
5158         int ret = 0;
5159
5160         delayed_refs = &trans->transaction->delayed_refs;
5161         spin_lock(&delayed_refs->lock);
5162         head = btrfs_find_delayed_ref_head(trans, bytenr);
5163         if (!head)
5164                 goto out;
5165
5166         node = rb_prev(&head->node.rb_node);
5167         if (!node)
5168                 goto out;
5169
5170         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5171
5172         /* there are still entries for this ref, we can't drop it */
5173         if (ref->bytenr == bytenr)
5174                 goto out;
5175
5176         if (head->extent_op) {
5177                 if (!head->must_insert_reserved)
5178                         goto out;
5179                 kfree(head->extent_op);
5180                 head->extent_op = NULL;
5181         }
5182
5183         /*
5184          * waiting for the lock here would deadlock.  If someone else has it
5185          * locked they are already in the process of dropping it anyway
5186          */
5187         if (!mutex_trylock(&head->mutex))
5188                 goto out;
5189
5190         /*
5191          * at this point we have a head with no other entries.  Go
5192          * ahead and process it.
5193          */
5194         head->node.in_tree = 0;
5195         rb_erase(&head->node.rb_node, &delayed_refs->root);
5196
5197         delayed_refs->num_entries--;
5198         if (waitqueue_active(&delayed_refs->seq_wait))
5199                 wake_up(&delayed_refs->seq_wait);
5200
5201         /*
5202          * we don't take a ref on the node because we're removing it from the
5203          * tree, so we just steal the ref the tree was holding.
5204          */
5205         delayed_refs->num_heads--;
5206         if (list_empty(&head->cluster))
5207                 delayed_refs->num_heads_ready--;
5208
5209         list_del_init(&head->cluster);
5210         spin_unlock(&delayed_refs->lock);
5211
5212         BUG_ON(head->extent_op);
5213         if (head->must_insert_reserved)
5214                 ret = 1;
5215
5216         mutex_unlock(&head->mutex);
5217         btrfs_put_delayed_ref(&head->node);
5218         return ret;
5219 out:
5220         spin_unlock(&delayed_refs->lock);
5221         return 0;
5222 }
5223
5224 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5225                            struct btrfs_root *root,
5226                            struct extent_buffer *buf,
5227                            u64 parent, int last_ref)
5228 {
5229         struct btrfs_block_group_cache *cache = NULL;
5230         int ret;
5231
5232         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5233                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5234                                         buf->start, buf->len,
5235                                         parent, root->root_key.objectid,
5236                                         btrfs_header_level(buf),
5237                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5238                 BUG_ON(ret); /* -ENOMEM */
5239         }
5240
5241         if (!last_ref)
5242                 return;
5243
5244         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5245
5246         if (btrfs_header_generation(buf) == trans->transid) {
5247                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5248                         ret = check_ref_cleanup(trans, root, buf->start);
5249                         if (!ret)
5250                                 goto out;
5251                 }
5252
5253                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5254                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5255                         goto out;
5256                 }
5257
5258                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5259
5260                 btrfs_add_free_space(cache, buf->start, buf->len);
5261                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5262         }
5263 out:
5264         /*
5265          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5266          * anymore.
5267          */
5268         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5269         btrfs_put_block_group(cache);
5270 }
5271
5272 /* Can return -ENOMEM */
5273 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5274                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5275                       u64 owner, u64 offset, int for_cow)
5276 {
5277         int ret;
5278         struct btrfs_fs_info *fs_info = root->fs_info;
5279
5280         /*
5281          * tree log blocks never actually go into the extent allocation
5282          * tree, just update pinning info and exit early.
5283          */
5284         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5285                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5286                 /* unlocks the pinned mutex */
5287                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5288                 ret = 0;
5289         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5290                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5291                                         num_bytes,
5292                                         parent, root_objectid, (int)owner,
5293                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5294         } else {
5295                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5296                                                 num_bytes,
5297                                                 parent, root_objectid, owner,
5298                                                 offset, BTRFS_DROP_DELAYED_REF,
5299                                                 NULL, for_cow);
5300         }
5301         return ret;
5302 }
5303
5304 static u64 stripe_align(struct btrfs_root *root, u64 val)
5305 {
5306         u64 mask = ((u64)root->stripesize - 1);
5307         u64 ret = (val + mask) & ~mask;
5308         return ret;
5309 }
5310
5311 /*
5312  * when we wait for progress in the block group caching, its because
5313  * our allocation attempt failed at least once.  So, we must sleep
5314  * and let some progress happen before we try again.
5315  *
5316  * This function will sleep at least once waiting for new free space to
5317  * show up, and then it will check the block group free space numbers
5318  * for our min num_bytes.  Another option is to have it go ahead
5319  * and look in the rbtree for a free extent of a given size, but this
5320  * is a good start.
5321  */
5322 static noinline int
5323 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5324                                 u64 num_bytes)
5325 {
5326         struct btrfs_caching_control *caching_ctl;
5327         DEFINE_WAIT(wait);
5328
5329         caching_ctl = get_caching_control(cache);
5330         if (!caching_ctl)
5331                 return 0;
5332
5333         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5334                    (cache->free_space_ctl->free_space >= num_bytes));
5335
5336         put_caching_control(caching_ctl);
5337         return 0;
5338 }
5339
5340 static noinline int
5341 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5342 {
5343         struct btrfs_caching_control *caching_ctl;
5344         DEFINE_WAIT(wait);
5345
5346         caching_ctl = get_caching_control(cache);
5347         if (!caching_ctl)
5348                 return 0;
5349
5350         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5351
5352         put_caching_control(caching_ctl);
5353         return 0;
5354 }
5355
5356 static int __get_block_group_index(u64 flags)
5357 {
5358         int index;
5359
5360         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5361                 index = 0;
5362         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5363                 index = 1;
5364         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5365                 index = 2;
5366         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5367                 index = 3;
5368         else
5369                 index = 4;
5370
5371         return index;
5372 }
5373
5374 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5375 {
5376         return __get_block_group_index(cache->flags);
5377 }
5378
5379 enum btrfs_loop_type {
5380         LOOP_CACHING_NOWAIT = 0,
5381         LOOP_CACHING_WAIT = 1,
5382         LOOP_ALLOC_CHUNK = 2,
5383         LOOP_NO_EMPTY_SIZE = 3,
5384 };
5385
5386 /*
5387  * walks the btree of allocated extents and find a hole of a given size.
5388  * The key ins is changed to record the hole:
5389  * ins->objectid == block start
5390  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5391  * ins->offset == number of blocks
5392  * Any available blocks before search_start are skipped.
5393  */
5394 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5395                                      struct btrfs_root *orig_root,
5396                                      u64 num_bytes, u64 empty_size,
5397                                      u64 hint_byte, struct btrfs_key *ins,
5398                                      u64 data)
5399 {
5400         int ret = 0;
5401         struct btrfs_root *root = orig_root->fs_info->extent_root;
5402         struct btrfs_free_cluster *last_ptr = NULL;
5403         struct btrfs_block_group_cache *block_group = NULL;
5404         struct btrfs_block_group_cache *used_block_group;
5405         u64 search_start = 0;
5406         int empty_cluster = 2 * 1024 * 1024;
5407         int allowed_chunk_alloc = 0;
5408         int done_chunk_alloc = 0;
5409         struct btrfs_space_info *space_info;
5410         int loop = 0;
5411         int index = 0;
5412         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5413                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5414         bool found_uncached_bg = false;
5415         bool failed_cluster_refill = false;
5416         bool failed_alloc = false;
5417         bool use_cluster = true;
5418         bool have_caching_bg = false;
5419
5420         WARN_ON(num_bytes < root->sectorsize);
5421         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5422         ins->objectid = 0;
5423         ins->offset = 0;
5424
5425         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5426
5427         space_info = __find_space_info(root->fs_info, data);
5428         if (!space_info) {
5429                 printk(KERN_ERR "No space info for %llu\n", data);
5430                 return -ENOSPC;
5431         }
5432
5433         /*
5434          * If the space info is for both data and metadata it means we have a
5435          * small filesystem and we can't use the clustering stuff.
5436          */
5437         if (btrfs_mixed_space_info(space_info))
5438                 use_cluster = false;
5439
5440         if (orig_root->ref_cows || empty_size)
5441                 allowed_chunk_alloc = 1;
5442
5443         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5444                 last_ptr = &root->fs_info->meta_alloc_cluster;
5445                 if (!btrfs_test_opt(root, SSD))
5446                         empty_cluster = 64 * 1024;
5447         }
5448
5449         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5450             btrfs_test_opt(root, SSD)) {
5451                 last_ptr = &root->fs_info->data_alloc_cluster;
5452         }
5453
5454         if (last_ptr) {
5455                 spin_lock(&last_ptr->lock);
5456                 if (last_ptr->block_group)
5457                         hint_byte = last_ptr->window_start;
5458                 spin_unlock(&last_ptr->lock);
5459         }
5460
5461         search_start = max(search_start, first_logical_byte(root, 0));
5462         search_start = max(search_start, hint_byte);
5463
5464         if (!last_ptr)
5465                 empty_cluster = 0;
5466
5467         if (search_start == hint_byte) {
5468                 block_group = btrfs_lookup_block_group(root->fs_info,
5469                                                        search_start);
5470                 used_block_group = block_group;
5471                 /*
5472                  * we don't want to use the block group if it doesn't match our
5473                  * allocation bits, or if its not cached.
5474                  *
5475                  * However if we are re-searching with an ideal block group
5476                  * picked out then we don't care that the block group is cached.
5477                  */
5478                 if (block_group && block_group_bits(block_group, data) &&
5479                     block_group->cached != BTRFS_CACHE_NO) {
5480                         down_read(&space_info->groups_sem);
5481                         if (list_empty(&block_group->list) ||
5482                             block_group->ro) {
5483                                 /*
5484                                  * someone is removing this block group,
5485                                  * we can't jump into the have_block_group
5486                                  * target because our list pointers are not
5487                                  * valid
5488                                  */
5489                                 btrfs_put_block_group(block_group);
5490                                 up_read(&space_info->groups_sem);
5491                         } else {
5492                                 index = get_block_group_index(block_group);
5493                                 goto have_block_group;
5494                         }
5495                 } else if (block_group) {
5496                         btrfs_put_block_group(block_group);
5497                 }
5498         }
5499 search:
5500         have_caching_bg = false;
5501         down_read(&space_info->groups_sem);
5502         list_for_each_entry(block_group, &space_info->block_groups[index],
5503                             list) {
5504                 u64 offset;
5505                 int cached;
5506
5507                 used_block_group = block_group;
5508                 btrfs_get_block_group(block_group);
5509                 search_start = block_group->key.objectid;
5510
5511                 /*
5512                  * this can happen if we end up cycling through all the
5513                  * raid types, but we want to make sure we only allocate
5514                  * for the proper type.
5515                  */
5516                 if (!block_group_bits(block_group, data)) {
5517                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5518                                 BTRFS_BLOCK_GROUP_RAID1 |
5519                                 BTRFS_BLOCK_GROUP_RAID10;
5520
5521                         /*
5522                          * if they asked for extra copies and this block group
5523                          * doesn't provide them, bail.  This does allow us to
5524                          * fill raid0 from raid1.
5525                          */
5526                         if ((data & extra) && !(block_group->flags & extra))
5527                                 goto loop;
5528                 }
5529
5530 have_block_group:
5531                 cached = block_group_cache_done(block_group);
5532                 if (unlikely(!cached)) {
5533                         found_uncached_bg = true;
5534                         ret = cache_block_group(block_group, trans,
5535                                                 orig_root, 0);
5536                         BUG_ON(ret < 0);
5537                         ret = 0;
5538                 }
5539
5540                 if (unlikely(block_group->ro))
5541                         goto loop;
5542
5543                 /*
5544                  * Ok we want to try and use the cluster allocator, so
5545                  * lets look there
5546                  */
5547                 if (last_ptr) {
5548                         /*
5549                          * the refill lock keeps out other
5550                          * people trying to start a new cluster
5551                          */
5552                         spin_lock(&last_ptr->refill_lock);
5553                         used_block_group = last_ptr->block_group;
5554                         if (used_block_group != block_group &&
5555                             (!used_block_group ||
5556                              used_block_group->ro ||
5557                              !block_group_bits(used_block_group, data))) {
5558                                 used_block_group = block_group;
5559                                 goto refill_cluster;
5560                         }
5561
5562                         if (used_block_group != block_group)
5563                                 btrfs_get_block_group(used_block_group);
5564
5565                         offset = btrfs_alloc_from_cluster(used_block_group,
5566                           last_ptr, num_bytes, used_block_group->key.objectid);
5567                         if (offset) {
5568                                 /* we have a block, we're done */
5569                                 spin_unlock(&last_ptr->refill_lock);
5570                                 trace_btrfs_reserve_extent_cluster(root,
5571                                         block_group, search_start, num_bytes);
5572                                 goto checks;
5573                         }
5574
5575                         WARN_ON(last_ptr->block_group != used_block_group);
5576                         if (used_block_group != block_group) {
5577                                 btrfs_put_block_group(used_block_group);
5578                                 used_block_group = block_group;
5579                         }
5580 refill_cluster:
5581                         BUG_ON(used_block_group != block_group);
5582                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5583                          * set up a new clusters, so lets just skip it
5584                          * and let the allocator find whatever block
5585                          * it can find.  If we reach this point, we
5586                          * will have tried the cluster allocator
5587                          * plenty of times and not have found
5588                          * anything, so we are likely way too
5589                          * fragmented for the clustering stuff to find
5590                          * anything.
5591                          *
5592                          * However, if the cluster is taken from the
5593                          * current block group, release the cluster
5594                          * first, so that we stand a better chance of
5595                          * succeeding in the unclustered
5596                          * allocation.  */
5597                         if (loop >= LOOP_NO_EMPTY_SIZE &&
5598                             last_ptr->block_group != block_group) {
5599                                 spin_unlock(&last_ptr->refill_lock);
5600                                 goto unclustered_alloc;
5601                         }
5602
5603                         /*
5604                          * this cluster didn't work out, free it and
5605                          * start over
5606                          */
5607                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5608
5609                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5610                                 spin_unlock(&last_ptr->refill_lock);
5611                                 goto unclustered_alloc;
5612                         }
5613
5614                         /* allocate a cluster in this block group */
5615                         ret = btrfs_find_space_cluster(trans, root,
5616                                                block_group, last_ptr,
5617                                                search_start, num_bytes,
5618                                                empty_cluster + empty_size);
5619                         if (ret == 0) {
5620                                 /*
5621                                  * now pull our allocation out of this
5622                                  * cluster
5623                                  */
5624                                 offset = btrfs_alloc_from_cluster(block_group,
5625                                                   last_ptr, num_bytes,
5626                                                   search_start);
5627                                 if (offset) {
5628                                         /* we found one, proceed */
5629                                         spin_unlock(&last_ptr->refill_lock);
5630                                         trace_btrfs_reserve_extent_cluster(root,
5631                                                 block_group, search_start,
5632                                                 num_bytes);
5633                                         goto checks;
5634                                 }
5635                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5636                                    && !failed_cluster_refill) {
5637                                 spin_unlock(&last_ptr->refill_lock);
5638
5639                                 failed_cluster_refill = true;
5640                                 wait_block_group_cache_progress(block_group,
5641                                        num_bytes + empty_cluster + empty_size);
5642                                 goto have_block_group;
5643                         }
5644
5645                         /*
5646                          * at this point we either didn't find a cluster
5647                          * or we weren't able to allocate a block from our
5648                          * cluster.  Free the cluster we've been trying
5649                          * to use, and go to the next block group
5650                          */
5651                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5652                         spin_unlock(&last_ptr->refill_lock);
5653                         goto loop;
5654                 }
5655
5656 unclustered_alloc:
5657                 spin_lock(&block_group->free_space_ctl->tree_lock);
5658                 if (cached &&
5659                     block_group->free_space_ctl->free_space <
5660                     num_bytes + empty_cluster + empty_size) {
5661                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5662                         goto loop;
5663                 }
5664                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5665
5666                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5667                                                     num_bytes, empty_size);
5668                 /*
5669                  * If we didn't find a chunk, and we haven't failed on this
5670                  * block group before, and this block group is in the middle of
5671                  * caching and we are ok with waiting, then go ahead and wait
5672                  * for progress to be made, and set failed_alloc to true.
5673                  *
5674                  * If failed_alloc is true then we've already waited on this
5675                  * block group once and should move on to the next block group.
5676                  */
5677                 if (!offset && !failed_alloc && !cached &&
5678                     loop > LOOP_CACHING_NOWAIT) {
5679                         wait_block_group_cache_progress(block_group,
5680                                                 num_bytes + empty_size);
5681                         failed_alloc = true;
5682                         goto have_block_group;
5683                 } else if (!offset) {
5684                         if (!cached)
5685                                 have_caching_bg = true;
5686                         goto loop;
5687                 }
5688 checks:
5689                 search_start = stripe_align(root, offset);
5690
5691                 /* move on to the next group */
5692                 if (search_start + num_bytes >
5693                     used_block_group->key.objectid + used_block_group->key.offset) {
5694                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5695                         goto loop;
5696                 }
5697
5698                 if (offset < search_start)
5699                         btrfs_add_free_space(used_block_group, offset,
5700                                              search_start - offset);
5701                 BUG_ON(offset > search_start);
5702
5703                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
5704                                                   alloc_type);
5705                 if (ret == -EAGAIN) {
5706                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5707                         goto loop;
5708                 }
5709
5710                 /* we are all good, lets return */
5711                 ins->objectid = search_start;
5712                 ins->offset = num_bytes;
5713
5714                 trace_btrfs_reserve_extent(orig_root, block_group,
5715                                            search_start, num_bytes);
5716                 if (offset < search_start)
5717                         btrfs_add_free_space(used_block_group, offset,
5718                                              search_start - offset);
5719                 BUG_ON(offset > search_start);
5720                 if (used_block_group != block_group)
5721                         btrfs_put_block_group(used_block_group);
5722                 btrfs_put_block_group(block_group);
5723                 break;
5724 loop:
5725                 failed_cluster_refill = false;
5726                 failed_alloc = false;
5727                 BUG_ON(index != get_block_group_index(block_group));
5728                 if (used_block_group != block_group)
5729                         btrfs_put_block_group(used_block_group);
5730                 btrfs_put_block_group(block_group);
5731         }
5732         up_read(&space_info->groups_sem);
5733
5734         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
5735                 goto search;
5736
5737         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5738                 goto search;
5739
5740         /*
5741          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5742          *                      caching kthreads as we move along
5743          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5744          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5745          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5746          *                      again
5747          */
5748         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5749                 index = 0;
5750                 loop++;
5751                 if (loop == LOOP_ALLOC_CHUNK) {
5752                        if (allowed_chunk_alloc) {
5753                                 ret = do_chunk_alloc(trans, root, num_bytes +
5754                                                      2 * 1024 * 1024, data,
5755                                                      CHUNK_ALLOC_LIMITED);
5756                                 /*
5757                                  * Do not bail out on ENOSPC since we
5758                                  * can do more things.
5759                                  */
5760                                 if (ret < 0 && ret != -ENOSPC) {
5761                                         btrfs_abort_transaction(trans,
5762                                                                 root, ret);
5763                                         goto out;
5764                                 }
5765                                 allowed_chunk_alloc = 0;
5766                                 if (ret == 1)
5767                                         done_chunk_alloc = 1;
5768                         } else if (!done_chunk_alloc &&
5769                                    space_info->force_alloc ==
5770                                    CHUNK_ALLOC_NO_FORCE) {
5771                                 space_info->force_alloc = CHUNK_ALLOC_LIMITED;
5772                         }
5773
5774                        /*
5775                         * We didn't allocate a chunk, go ahead and drop the
5776                         * empty size and loop again.
5777                         */
5778                        if (!done_chunk_alloc)
5779                                loop = LOOP_NO_EMPTY_SIZE;
5780                 }
5781
5782                 if (loop == LOOP_NO_EMPTY_SIZE) {
5783                         empty_size = 0;
5784                         empty_cluster = 0;
5785                 }
5786
5787                 goto search;
5788         } else if (!ins->objectid) {
5789                 ret = -ENOSPC;
5790         } else if (ins->objectid) {
5791                 ret = 0;
5792         }
5793 out:
5794
5795         return ret;
5796 }
5797
5798 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5799                             int dump_block_groups)
5800 {
5801         struct btrfs_block_group_cache *cache;
5802         int index = 0;
5803
5804         spin_lock(&info->lock);
5805         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5806                (unsigned long long)info->flags,
5807                (unsigned long long)(info->total_bytes - info->bytes_used -
5808                                     info->bytes_pinned - info->bytes_reserved -
5809                                     info->bytes_readonly),
5810                (info->full) ? "" : "not ");
5811         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5812                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5813                (unsigned long long)info->total_bytes,
5814                (unsigned long long)info->bytes_used,
5815                (unsigned long long)info->bytes_pinned,
5816                (unsigned long long)info->bytes_reserved,
5817                (unsigned long long)info->bytes_may_use,
5818                (unsigned long long)info->bytes_readonly);
5819         spin_unlock(&info->lock);
5820
5821         if (!dump_block_groups)
5822                 return;
5823
5824         down_read(&info->groups_sem);
5825 again:
5826         list_for_each_entry(cache, &info->block_groups[index], list) {
5827                 spin_lock(&cache->lock);
5828                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used "
5829                        "%llu pinned %llu reserved\n",
5830                        (unsigned long long)cache->key.objectid,
5831                        (unsigned long long)cache->key.offset,
5832                        (unsigned long long)btrfs_block_group_used(&cache->item),
5833                        (unsigned long long)cache->pinned,
5834                        (unsigned long long)cache->reserved);
5835                 btrfs_dump_free_space(cache, bytes);
5836                 spin_unlock(&cache->lock);
5837         }
5838         if (++index < BTRFS_NR_RAID_TYPES)
5839                 goto again;
5840         up_read(&info->groups_sem);
5841 }
5842
5843 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5844                          struct btrfs_root *root,
5845                          u64 num_bytes, u64 min_alloc_size,
5846                          u64 empty_size, u64 hint_byte,
5847                          struct btrfs_key *ins, u64 data)
5848 {
5849         bool final_tried = false;
5850         int ret;
5851
5852         data = btrfs_get_alloc_profile(root, data);
5853 again:
5854         /*
5855          * the only place that sets empty_size is btrfs_realloc_node, which
5856          * is not called recursively on allocations
5857          */
5858         if (empty_size || root->ref_cows) {
5859                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5860                                      num_bytes + 2 * 1024 * 1024, data,
5861                                      CHUNK_ALLOC_NO_FORCE);
5862                 if (ret < 0 && ret != -ENOSPC) {
5863                         btrfs_abort_transaction(trans, root, ret);
5864                         return ret;
5865                 }
5866         }
5867
5868         WARN_ON(num_bytes < root->sectorsize);
5869         ret = find_free_extent(trans, root, num_bytes, empty_size,
5870                                hint_byte, ins, data);
5871
5872         if (ret == -ENOSPC) {
5873                 if (!final_tried) {
5874                         num_bytes = num_bytes >> 1;
5875                         num_bytes = num_bytes & ~(root->sectorsize - 1);
5876                         num_bytes = max(num_bytes, min_alloc_size);
5877                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
5878                                        num_bytes, data, CHUNK_ALLOC_FORCE);
5879                         if (ret < 0 && ret != -ENOSPC) {
5880                                 btrfs_abort_transaction(trans, root, ret);
5881                                 return ret;
5882                         }
5883                         if (num_bytes == min_alloc_size)
5884                                 final_tried = true;
5885                         goto again;
5886                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
5887                         struct btrfs_space_info *sinfo;
5888
5889                         sinfo = __find_space_info(root->fs_info, data);
5890                         printk(KERN_ERR "btrfs allocation failed flags %llu, "
5891                                "wanted %llu\n", (unsigned long long)data,
5892                                (unsigned long long)num_bytes);
5893                         if (sinfo)
5894                                 dump_space_info(sinfo, num_bytes, 1);
5895                 }
5896         }
5897
5898         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
5899
5900         return ret;
5901 }
5902
5903 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
5904                                         u64 start, u64 len, int pin)
5905 {
5906         struct btrfs_block_group_cache *cache;
5907         int ret = 0;
5908
5909         cache = btrfs_lookup_block_group(root->fs_info, start);
5910         if (!cache) {
5911                 printk(KERN_ERR "Unable to find block group for %llu\n",
5912                        (unsigned long long)start);
5913                 return -ENOSPC;
5914         }
5915
5916         if (btrfs_test_opt(root, DISCARD))
5917                 ret = btrfs_discard_extent(root, start, len, NULL);
5918
5919         if (pin)
5920                 pin_down_extent(root, cache, start, len, 1);
5921         else {
5922                 btrfs_add_free_space(cache, start, len);
5923                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
5924         }
5925         btrfs_put_block_group(cache);
5926
5927         trace_btrfs_reserved_extent_free(root, start, len);
5928
5929         return ret;
5930 }
5931
5932 int btrfs_free_reserved_extent(struct btrfs_root *root,
5933                                         u64 start, u64 len)
5934 {
5935         return __btrfs_free_reserved_extent(root, start, len, 0);
5936 }
5937
5938 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
5939                                        u64 start, u64 len)
5940 {
5941         return __btrfs_free_reserved_extent(root, start, len, 1);
5942 }
5943
5944 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
5945                                       struct btrfs_root *root,
5946                                       u64 parent, u64 root_objectid,
5947                                       u64 flags, u64 owner, u64 offset,
5948                                       struct btrfs_key *ins, int ref_mod)
5949 {
5950         int ret;
5951         struct btrfs_fs_info *fs_info = root->fs_info;
5952         struct btrfs_extent_item *extent_item;
5953         struct btrfs_extent_inline_ref *iref;
5954         struct btrfs_path *path;
5955         struct extent_buffer *leaf;
5956         int type;
5957         u32 size;
5958
5959         if (parent > 0)
5960                 type = BTRFS_SHARED_DATA_REF_KEY;
5961         else
5962                 type = BTRFS_EXTENT_DATA_REF_KEY;
5963
5964         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
5965
5966         path = btrfs_alloc_path();
5967         if (!path)
5968                 return -ENOMEM;
5969
5970         path->leave_spinning = 1;
5971         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
5972                                       ins, size);
5973         if (ret) {
5974                 btrfs_free_path(path);
5975                 return ret;
5976         }
5977
5978         leaf = path->nodes[0];
5979         extent_item = btrfs_item_ptr(leaf, path->slots[0],
5980                                      struct btrfs_extent_item);
5981         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
5982         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
5983         btrfs_set_extent_flags(leaf, extent_item,
5984                                flags | BTRFS_EXTENT_FLAG_DATA);
5985
5986         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
5987         btrfs_set_extent_inline_ref_type(leaf, iref, type);
5988         if (parent > 0) {
5989                 struct btrfs_shared_data_ref *ref;
5990                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
5991                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
5992                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
5993         } else {
5994                 struct btrfs_extent_data_ref *ref;
5995                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
5996                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
5997                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
5998                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
5999                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6000         }
6001
6002         btrfs_mark_buffer_dirty(path->nodes[0]);
6003         btrfs_free_path(path);
6004
6005         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
6006         if (ret) { /* -ENOENT, logic error */
6007                 printk(KERN_ERR "btrfs update block group failed for %llu "
6008                        "%llu\n", (unsigned long long)ins->objectid,
6009                        (unsigned long long)ins->offset);
6010                 BUG();
6011         }
6012         return ret;
6013 }
6014
6015 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6016                                      struct btrfs_root *root,
6017                                      u64 parent, u64 root_objectid,
6018                                      u64 flags, struct btrfs_disk_key *key,
6019                                      int level, struct btrfs_key *ins)
6020 {
6021         int ret;
6022         struct btrfs_fs_info *fs_info = root->fs_info;
6023         struct btrfs_extent_item *extent_item;
6024         struct btrfs_tree_block_info *block_info;
6025         struct btrfs_extent_inline_ref *iref;
6026         struct btrfs_path *path;
6027         struct extent_buffer *leaf;
6028         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
6029
6030         path = btrfs_alloc_path();
6031         if (!path)
6032                 return -ENOMEM;
6033
6034         path->leave_spinning = 1;
6035         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6036                                       ins, size);
6037         if (ret) {
6038                 btrfs_free_path(path);
6039                 return ret;
6040         }
6041
6042         leaf = path->nodes[0];
6043         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6044                                      struct btrfs_extent_item);
6045         btrfs_set_extent_refs(leaf, extent_item, 1);
6046         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6047         btrfs_set_extent_flags(leaf, extent_item,
6048                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6049         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6050
6051         btrfs_set_tree_block_key(leaf, block_info, key);
6052         btrfs_set_tree_block_level(leaf, block_info, level);
6053
6054         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6055         if (parent > 0) {
6056                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6057                 btrfs_set_extent_inline_ref_type(leaf, iref,
6058                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6059                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6060         } else {
6061                 btrfs_set_extent_inline_ref_type(leaf, iref,
6062                                                  BTRFS_TREE_BLOCK_REF_KEY);
6063                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6064         }
6065
6066         btrfs_mark_buffer_dirty(leaf);
6067         btrfs_free_path(path);
6068
6069         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
6070         if (ret) { /* -ENOENT, logic error */
6071                 printk(KERN_ERR "btrfs update block group failed for %llu "
6072                        "%llu\n", (unsigned long long)ins->objectid,
6073                        (unsigned long long)ins->offset);
6074                 BUG();
6075         }
6076         return ret;
6077 }
6078
6079 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6080                                      struct btrfs_root *root,
6081                                      u64 root_objectid, u64 owner,
6082                                      u64 offset, struct btrfs_key *ins)
6083 {
6084         int ret;
6085
6086         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6087
6088         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6089                                          ins->offset, 0,
6090                                          root_objectid, owner, offset,
6091                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6092         return ret;
6093 }
6094
6095 /*
6096  * this is used by the tree logging recovery code.  It records that
6097  * an extent has been allocated and makes sure to clear the free
6098  * space cache bits as well
6099  */
6100 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6101                                    struct btrfs_root *root,
6102                                    u64 root_objectid, u64 owner, u64 offset,
6103                                    struct btrfs_key *ins)
6104 {
6105         int ret;
6106         struct btrfs_block_group_cache *block_group;
6107         struct btrfs_caching_control *caching_ctl;
6108         u64 start = ins->objectid;
6109         u64 num_bytes = ins->offset;
6110
6111         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6112         cache_block_group(block_group, trans, NULL, 0);
6113         caching_ctl = get_caching_control(block_group);
6114
6115         if (!caching_ctl) {
6116                 BUG_ON(!block_group_cache_done(block_group));
6117                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6118                 BUG_ON(ret); /* -ENOMEM */
6119         } else {
6120                 mutex_lock(&caching_ctl->mutex);
6121
6122                 if (start >= caching_ctl->progress) {
6123                         ret = add_excluded_extent(root, start, num_bytes);
6124                         BUG_ON(ret); /* -ENOMEM */
6125                 } else if (start + num_bytes <= caching_ctl->progress) {
6126                         ret = btrfs_remove_free_space(block_group,
6127                                                       start, num_bytes);
6128                         BUG_ON(ret); /* -ENOMEM */
6129                 } else {
6130                         num_bytes = caching_ctl->progress - start;
6131                         ret = btrfs_remove_free_space(block_group,
6132                                                       start, num_bytes);
6133                         BUG_ON(ret); /* -ENOMEM */
6134
6135                         start = caching_ctl->progress;
6136                         num_bytes = ins->objectid + ins->offset -
6137                                     caching_ctl->progress;
6138                         ret = add_excluded_extent(root, start, num_bytes);
6139                         BUG_ON(ret); /* -ENOMEM */
6140                 }
6141
6142                 mutex_unlock(&caching_ctl->mutex);
6143                 put_caching_control(caching_ctl);
6144         }
6145
6146         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6147                                           RESERVE_ALLOC_NO_ACCOUNT);
6148         BUG_ON(ret); /* logic error */
6149         btrfs_put_block_group(block_group);
6150         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6151                                          0, owner, offset, ins, 1);
6152         return ret;
6153 }
6154
6155 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
6156                                             struct btrfs_root *root,
6157                                             u64 bytenr, u32 blocksize,
6158                                             int level)
6159 {
6160         struct extent_buffer *buf;
6161
6162         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6163         if (!buf)
6164                 return ERR_PTR(-ENOMEM);
6165         btrfs_set_header_generation(buf, trans->transid);
6166         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6167         btrfs_tree_lock(buf);
6168         clean_tree_block(trans, root, buf);
6169         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6170
6171         btrfs_set_lock_blocking(buf);
6172         btrfs_set_buffer_uptodate(buf);
6173
6174         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6175                 /*
6176                  * we allow two log transactions at a time, use different
6177                  * EXENT bit to differentiate dirty pages.
6178                  */
6179                 if (root->log_transid % 2 == 0)
6180                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6181                                         buf->start + buf->len - 1, GFP_NOFS);
6182                 else
6183                         set_extent_new(&root->dirty_log_pages, buf->start,
6184                                         buf->start + buf->len - 1, GFP_NOFS);
6185         } else {
6186                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6187                          buf->start + buf->len - 1, GFP_NOFS);
6188         }
6189         trans->blocks_used++;
6190         /* this returns a buffer locked for blocking */
6191         return buf;
6192 }
6193
6194 static struct btrfs_block_rsv *
6195 use_block_rsv(struct btrfs_trans_handle *trans,
6196               struct btrfs_root *root, u32 blocksize)
6197 {
6198         struct btrfs_block_rsv *block_rsv;
6199         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6200         int ret;
6201
6202         block_rsv = get_block_rsv(trans, root);
6203
6204         if (block_rsv->size == 0) {
6205                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
6206                 /*
6207                  * If we couldn't reserve metadata bytes try and use some from
6208                  * the global reserve.
6209                  */
6210                 if (ret && block_rsv != global_rsv) {
6211                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6212                         if (!ret)
6213                                 return global_rsv;
6214                         return ERR_PTR(ret);
6215                 } else if (ret) {
6216                         return ERR_PTR(ret);
6217                 }
6218                 return block_rsv;
6219         }
6220
6221         ret = block_rsv_use_bytes(block_rsv, blocksize);
6222         if (!ret)
6223                 return block_rsv;
6224         if (ret) {
6225                 static DEFINE_RATELIMIT_STATE(_rs,
6226                                 DEFAULT_RATELIMIT_INTERVAL,
6227                                 /*DEFAULT_RATELIMIT_BURST*/ 2);
6228                 if (__ratelimit(&_rs)) {
6229                         printk(KERN_DEBUG "btrfs: block rsv returned %d\n", ret);
6230                         WARN_ON(1);
6231                 }
6232                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
6233                 if (!ret) {
6234                         return block_rsv;
6235                 } else if (ret && block_rsv != global_rsv) {
6236                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6237                         if (!ret)
6238                                 return global_rsv;
6239                 }
6240         }
6241
6242         return ERR_PTR(-ENOSPC);
6243 }
6244
6245 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6246                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6247 {
6248         block_rsv_add_bytes(block_rsv, blocksize, 0);
6249         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6250 }
6251
6252 /*
6253  * finds a free extent and does all the dirty work required for allocation
6254  * returns the key for the extent through ins, and a tree buffer for
6255  * the first block of the extent through buf.
6256  *
6257  * returns the tree buffer or NULL.
6258  */
6259 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6260                                         struct btrfs_root *root, u32 blocksize,
6261                                         u64 parent, u64 root_objectid,
6262                                         struct btrfs_disk_key *key, int level,
6263                                         u64 hint, u64 empty_size)
6264 {
6265         struct btrfs_key ins;
6266         struct btrfs_block_rsv *block_rsv;
6267         struct extent_buffer *buf;
6268         u64 flags = 0;
6269         int ret;
6270
6271
6272         block_rsv = use_block_rsv(trans, root, blocksize);
6273         if (IS_ERR(block_rsv))
6274                 return ERR_CAST(block_rsv);
6275
6276         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6277                                    empty_size, hint, &ins, 0);
6278         if (ret) {
6279                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6280                 return ERR_PTR(ret);
6281         }
6282
6283         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6284                                     blocksize, level);
6285         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6286
6287         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6288                 if (parent == 0)
6289                         parent = ins.objectid;
6290                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6291         } else
6292                 BUG_ON(parent > 0);
6293
6294         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6295                 struct btrfs_delayed_extent_op *extent_op;
6296                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
6297                 BUG_ON(!extent_op); /* -ENOMEM */
6298                 if (key)
6299                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6300                 else
6301                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6302                 extent_op->flags_to_set = flags;
6303                 extent_op->update_key = 1;
6304                 extent_op->update_flags = 1;
6305                 extent_op->is_data = 0;
6306
6307                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6308                                         ins.objectid,
6309                                         ins.offset, parent, root_objectid,
6310                                         level, BTRFS_ADD_DELAYED_EXTENT,
6311                                         extent_op, 0);
6312                 BUG_ON(ret); /* -ENOMEM */
6313         }
6314         return buf;
6315 }
6316
6317 struct walk_control {
6318         u64 refs[BTRFS_MAX_LEVEL];
6319         u64 flags[BTRFS_MAX_LEVEL];
6320         struct btrfs_key update_progress;
6321         int stage;
6322         int level;
6323         int shared_level;
6324         int update_ref;
6325         int keep_locks;
6326         int reada_slot;
6327         int reada_count;
6328         int for_reloc;
6329 };
6330
6331 #define DROP_REFERENCE  1
6332 #define UPDATE_BACKREF  2
6333
6334 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6335                                      struct btrfs_root *root,
6336                                      struct walk_control *wc,
6337                                      struct btrfs_path *path)
6338 {
6339         u64 bytenr;
6340         u64 generation;
6341         u64 refs;
6342         u64 flags;
6343         u32 nritems;
6344         u32 blocksize;
6345         struct btrfs_key key;
6346         struct extent_buffer *eb;
6347         int ret;
6348         int slot;
6349         int nread = 0;
6350
6351         if (path->slots[wc->level] < wc->reada_slot) {
6352                 wc->reada_count = wc->reada_count * 2 / 3;
6353                 wc->reada_count = max(wc->reada_count, 2);
6354         } else {
6355                 wc->reada_count = wc->reada_count * 3 / 2;
6356                 wc->reada_count = min_t(int, wc->reada_count,
6357                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6358         }
6359
6360         eb = path->nodes[wc->level];
6361         nritems = btrfs_header_nritems(eb);
6362         blocksize = btrfs_level_size(root, wc->level - 1);
6363
6364         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6365                 if (nread >= wc->reada_count)
6366                         break;
6367
6368                 cond_resched();
6369                 bytenr = btrfs_node_blockptr(eb, slot);
6370                 generation = btrfs_node_ptr_generation(eb, slot);
6371
6372                 if (slot == path->slots[wc->level])
6373                         goto reada;
6374
6375                 if (wc->stage == UPDATE_BACKREF &&
6376                     generation <= root->root_key.offset)
6377                         continue;
6378
6379                 /* We don't lock the tree block, it's OK to be racy here */
6380                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6381                                                &refs, &flags);
6382                 /* We don't care about errors in readahead. */
6383                 if (ret < 0)
6384                         continue;
6385                 BUG_ON(refs == 0);
6386
6387                 if (wc->stage == DROP_REFERENCE) {
6388                         if (refs == 1)
6389                                 goto reada;
6390
6391                         if (wc->level == 1 &&
6392                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6393                                 continue;
6394                         if (!wc->update_ref ||
6395                             generation <= root->root_key.offset)
6396                                 continue;
6397                         btrfs_node_key_to_cpu(eb, &key, slot);
6398                         ret = btrfs_comp_cpu_keys(&key,
6399                                                   &wc->update_progress);
6400                         if (ret < 0)
6401                                 continue;
6402                 } else {
6403                         if (wc->level == 1 &&
6404                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6405                                 continue;
6406                 }
6407 reada:
6408                 ret = readahead_tree_block(root, bytenr, blocksize,
6409                                            generation);
6410                 if (ret)
6411                         break;
6412                 nread++;
6413         }
6414         wc->reada_slot = slot;
6415 }
6416
6417 /*
6418  * hepler to process tree block while walking down the tree.
6419  *
6420  * when wc->stage == UPDATE_BACKREF, this function updates
6421  * back refs for pointers in the block.
6422  *
6423  * NOTE: return value 1 means we should stop walking down.
6424  */
6425 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6426                                    struct btrfs_root *root,
6427                                    struct btrfs_path *path,
6428                                    struct walk_control *wc, int lookup_info)
6429 {
6430         int level = wc->level;
6431         struct extent_buffer *eb = path->nodes[level];
6432         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6433         int ret;
6434
6435         if (wc->stage == UPDATE_BACKREF &&
6436             btrfs_header_owner(eb) != root->root_key.objectid)
6437                 return 1;
6438
6439         /*
6440          * when reference count of tree block is 1, it won't increase
6441          * again. once full backref flag is set, we never clear it.
6442          */
6443         if (lookup_info &&
6444             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6445              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6446                 BUG_ON(!path->locks[level]);
6447                 ret = btrfs_lookup_extent_info(trans, root,
6448                                                eb->start, eb->len,
6449                                                &wc->refs[level],
6450                                                &wc->flags[level]);
6451                 BUG_ON(ret == -ENOMEM);
6452                 if (ret)
6453                         return ret;
6454                 BUG_ON(wc->refs[level] == 0);
6455         }
6456
6457         if (wc->stage == DROP_REFERENCE) {
6458                 if (wc->refs[level] > 1)
6459                         return 1;
6460
6461                 if (path->locks[level] && !wc->keep_locks) {
6462                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6463                         path->locks[level] = 0;
6464                 }
6465                 return 0;
6466         }
6467
6468         /* wc->stage == UPDATE_BACKREF */
6469         if (!(wc->flags[level] & flag)) {
6470                 BUG_ON(!path->locks[level]);
6471                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
6472                 BUG_ON(ret); /* -ENOMEM */
6473                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
6474                 BUG_ON(ret); /* -ENOMEM */
6475                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6476                                                   eb->len, flag, 0);
6477                 BUG_ON(ret); /* -ENOMEM */
6478                 wc->flags[level] |= flag;
6479         }
6480
6481         /*
6482          * the block is shared by multiple trees, so it's not good to
6483          * keep the tree lock
6484          */
6485         if (path->locks[level] && level > 0) {
6486                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6487                 path->locks[level] = 0;
6488         }
6489         return 0;
6490 }
6491
6492 /*
6493  * hepler to process tree block pointer.
6494  *
6495  * when wc->stage == DROP_REFERENCE, this function checks
6496  * reference count of the block pointed to. if the block
6497  * is shared and we need update back refs for the subtree
6498  * rooted at the block, this function changes wc->stage to
6499  * UPDATE_BACKREF. if the block is shared and there is no
6500  * need to update back, this function drops the reference
6501  * to the block.
6502  *
6503  * NOTE: return value 1 means we should stop walking down.
6504  */
6505 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6506                                  struct btrfs_root *root,
6507                                  struct btrfs_path *path,
6508                                  struct walk_control *wc, int *lookup_info)
6509 {
6510         u64 bytenr;
6511         u64 generation;
6512         u64 parent;
6513         u32 blocksize;
6514         struct btrfs_key key;
6515         struct extent_buffer *next;
6516         int level = wc->level;
6517         int reada = 0;
6518         int ret = 0;
6519
6520         generation = btrfs_node_ptr_generation(path->nodes[level],
6521                                                path->slots[level]);
6522         /*
6523          * if the lower level block was created before the snapshot
6524          * was created, we know there is no need to update back refs
6525          * for the subtree
6526          */
6527         if (wc->stage == UPDATE_BACKREF &&
6528             generation <= root->root_key.offset) {
6529                 *lookup_info = 1;
6530                 return 1;
6531         }
6532
6533         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6534         blocksize = btrfs_level_size(root, level - 1);
6535
6536         next = btrfs_find_tree_block(root, bytenr, blocksize);
6537         if (!next) {
6538                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6539                 if (!next)
6540                         return -ENOMEM;
6541                 reada = 1;
6542         }
6543         btrfs_tree_lock(next);
6544         btrfs_set_lock_blocking(next);
6545
6546         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6547                                        &wc->refs[level - 1],
6548                                        &wc->flags[level - 1]);
6549         if (ret < 0) {
6550                 btrfs_tree_unlock(next);
6551                 return ret;
6552         }
6553
6554         BUG_ON(wc->refs[level - 1] == 0);
6555         *lookup_info = 0;
6556
6557         if (wc->stage == DROP_REFERENCE) {
6558                 if (wc->refs[level - 1] > 1) {
6559                         if (level == 1 &&
6560                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6561                                 goto skip;
6562
6563                         if (!wc->update_ref ||
6564                             generation <= root->root_key.offset)
6565                                 goto skip;
6566
6567                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6568                                               path->slots[level]);
6569                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6570                         if (ret < 0)
6571                                 goto skip;
6572
6573                         wc->stage = UPDATE_BACKREF;
6574                         wc->shared_level = level - 1;
6575                 }
6576         } else {
6577                 if (level == 1 &&
6578                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6579                         goto skip;
6580         }
6581
6582         if (!btrfs_buffer_uptodate(next, generation, 0)) {
6583                 btrfs_tree_unlock(next);
6584                 free_extent_buffer(next);
6585                 next = NULL;
6586                 *lookup_info = 1;
6587         }
6588
6589         if (!next) {
6590                 if (reada && level == 1)
6591                         reada_walk_down(trans, root, wc, path);
6592                 next = read_tree_block(root, bytenr, blocksize, generation);
6593                 if (!next)
6594                         return -EIO;
6595                 btrfs_tree_lock(next);
6596                 btrfs_set_lock_blocking(next);
6597         }
6598
6599         level--;
6600         BUG_ON(level != btrfs_header_level(next));
6601         path->nodes[level] = next;
6602         path->slots[level] = 0;
6603         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6604         wc->level = level;
6605         if (wc->level == 1)
6606                 wc->reada_slot = 0;
6607         return 0;
6608 skip:
6609         wc->refs[level - 1] = 0;
6610         wc->flags[level - 1] = 0;
6611         if (wc->stage == DROP_REFERENCE) {
6612                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6613                         parent = path->nodes[level]->start;
6614                 } else {
6615                         BUG_ON(root->root_key.objectid !=
6616                                btrfs_header_owner(path->nodes[level]));
6617                         parent = 0;
6618                 }
6619
6620                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6621                                 root->root_key.objectid, level - 1, 0, 0);
6622                 BUG_ON(ret); /* -ENOMEM */
6623         }
6624         btrfs_tree_unlock(next);
6625         free_extent_buffer(next);
6626         *lookup_info = 1;
6627         return 1;
6628 }
6629
6630 /*
6631  * hepler to process tree block while walking up the tree.
6632  *
6633  * when wc->stage == DROP_REFERENCE, this function drops
6634  * reference count on the block.
6635  *
6636  * when wc->stage == UPDATE_BACKREF, this function changes
6637  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6638  * to UPDATE_BACKREF previously while processing the block.
6639  *
6640  * NOTE: return value 1 means we should stop walking up.
6641  */
6642 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6643                                  struct btrfs_root *root,
6644                                  struct btrfs_path *path,
6645                                  struct walk_control *wc)
6646 {
6647         int ret;
6648         int level = wc->level;
6649         struct extent_buffer *eb = path->nodes[level];
6650         u64 parent = 0;
6651
6652         if (wc->stage == UPDATE_BACKREF) {
6653                 BUG_ON(wc->shared_level < level);
6654                 if (level < wc->shared_level)
6655                         goto out;
6656
6657                 ret = find_next_key(path, level + 1, &wc->update_progress);
6658                 if (ret > 0)
6659                         wc->update_ref = 0;
6660
6661                 wc->stage = DROP_REFERENCE;
6662                 wc->shared_level = -1;
6663                 path->slots[level] = 0;
6664
6665                 /*
6666                  * check reference count again if the block isn't locked.
6667                  * we should start walking down the tree again if reference
6668                  * count is one.
6669                  */
6670                 if (!path->locks[level]) {
6671                         BUG_ON(level == 0);
6672                         btrfs_tree_lock(eb);
6673                         btrfs_set_lock_blocking(eb);
6674                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6675
6676                         ret = btrfs_lookup_extent_info(trans, root,
6677                                                        eb->start, eb->len,
6678                                                        &wc->refs[level],
6679                                                        &wc->flags[level]);
6680                         if (ret < 0) {
6681                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6682                                 return ret;
6683                         }
6684                         BUG_ON(wc->refs[level] == 0);
6685                         if (wc->refs[level] == 1) {
6686                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6687                                 return 1;
6688                         }
6689                 }
6690         }
6691
6692         /* wc->stage == DROP_REFERENCE */
6693         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6694
6695         if (wc->refs[level] == 1) {
6696                 if (level == 0) {
6697                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6698                                 ret = btrfs_dec_ref(trans, root, eb, 1,
6699                                                     wc->for_reloc);
6700                         else
6701                                 ret = btrfs_dec_ref(trans, root, eb, 0,
6702                                                     wc->for_reloc);
6703                         BUG_ON(ret); /* -ENOMEM */
6704                 }
6705                 /* make block locked assertion in clean_tree_block happy */
6706                 if (!path->locks[level] &&
6707                     btrfs_header_generation(eb) == trans->transid) {
6708                         btrfs_tree_lock(eb);
6709                         btrfs_set_lock_blocking(eb);
6710                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6711                 }
6712                 clean_tree_block(trans, root, eb);
6713         }
6714
6715         if (eb == root->node) {
6716                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6717                         parent = eb->start;
6718                 else
6719                         BUG_ON(root->root_key.objectid !=
6720                                btrfs_header_owner(eb));
6721         } else {
6722                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6723                         parent = path->nodes[level + 1]->start;
6724                 else
6725                         BUG_ON(root->root_key.objectid !=
6726                                btrfs_header_owner(path->nodes[level + 1]));
6727         }
6728
6729         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6730 out:
6731         wc->refs[level] = 0;
6732         wc->flags[level] = 0;
6733         return 0;
6734 }
6735
6736 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6737                                    struct btrfs_root *root,
6738                                    struct btrfs_path *path,
6739                                    struct walk_control *wc)
6740 {
6741         int level = wc->level;
6742         int lookup_info = 1;
6743         int ret;
6744
6745         while (level >= 0) {
6746                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6747                 if (ret > 0)
6748                         break;
6749
6750                 if (level == 0)
6751                         break;
6752
6753                 if (path->slots[level] >=
6754                     btrfs_header_nritems(path->nodes[level]))
6755                         break;
6756
6757                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6758                 if (ret > 0) {
6759                         path->slots[level]++;
6760                         continue;
6761                 } else if (ret < 0)
6762                         return ret;
6763                 level = wc->level;
6764         }
6765         return 0;
6766 }
6767
6768 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6769                                  struct btrfs_root *root,
6770                                  struct btrfs_path *path,
6771                                  struct walk_control *wc, int max_level)
6772 {
6773         int level = wc->level;
6774         int ret;
6775
6776         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6777         while (level < max_level && path->nodes[level]) {
6778                 wc->level = level;
6779                 if (path->slots[level] + 1 <
6780                     btrfs_header_nritems(path->nodes[level])) {
6781                         path->slots[level]++;
6782                         return 0;
6783                 } else {
6784                         ret = walk_up_proc(trans, root, path, wc);
6785                         if (ret > 0)
6786                                 return 0;
6787
6788                         if (path->locks[level]) {
6789                                 btrfs_tree_unlock_rw(path->nodes[level],
6790                                                      path->locks[level]);
6791                                 path->locks[level] = 0;
6792                         }
6793                         free_extent_buffer(path->nodes[level]);
6794                         path->nodes[level] = NULL;
6795                         level++;
6796                 }
6797         }
6798         return 1;
6799 }
6800
6801 /*
6802  * drop a subvolume tree.
6803  *
6804  * this function traverses the tree freeing any blocks that only
6805  * referenced by the tree.
6806  *
6807  * when a shared tree block is found. this function decreases its
6808  * reference count by one. if update_ref is true, this function
6809  * also make sure backrefs for the shared block and all lower level
6810  * blocks are properly updated.
6811  */
6812 int btrfs_drop_snapshot(struct btrfs_root *root,
6813                          struct btrfs_block_rsv *block_rsv, int update_ref,
6814                          int for_reloc)
6815 {
6816         struct btrfs_path *path;
6817         struct btrfs_trans_handle *trans;
6818         struct btrfs_root *tree_root = root->fs_info->tree_root;
6819         struct btrfs_root_item *root_item = &root->root_item;
6820         struct walk_control *wc;
6821         struct btrfs_key key;
6822         int err = 0;
6823         int ret;
6824         int level;
6825
6826         path = btrfs_alloc_path();
6827         if (!path) {
6828                 err = -ENOMEM;
6829                 goto out;
6830         }
6831
6832         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6833         if (!wc) {
6834                 btrfs_free_path(path);
6835                 err = -ENOMEM;
6836                 goto out;
6837         }
6838
6839         trans = btrfs_start_transaction(tree_root, 0);
6840         if (IS_ERR(trans)) {
6841                 err = PTR_ERR(trans);
6842                 goto out_free;
6843         }
6844
6845         if (block_rsv)
6846                 trans->block_rsv = block_rsv;
6847
6848         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6849                 level = btrfs_header_level(root->node);
6850                 path->nodes[level] = btrfs_lock_root_node(root);
6851                 btrfs_set_lock_blocking(path->nodes[level]);
6852                 path->slots[level] = 0;
6853                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6854                 memset(&wc->update_progress, 0,
6855                        sizeof(wc->update_progress));
6856         } else {
6857                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6858                 memcpy(&wc->update_progress, &key,
6859                        sizeof(wc->update_progress));
6860
6861                 level = root_item->drop_level;
6862                 BUG_ON(level == 0);
6863                 path->lowest_level = level;
6864                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6865                 path->lowest_level = 0;
6866                 if (ret < 0) {
6867                         err = ret;
6868                         goto out_end_trans;
6869                 }
6870                 WARN_ON(ret > 0);
6871
6872                 /*
6873                  * unlock our path, this is safe because only this
6874                  * function is allowed to delete this snapshot
6875                  */
6876                 btrfs_unlock_up_safe(path, 0);
6877
6878                 level = btrfs_header_level(root->node);
6879                 while (1) {
6880                         btrfs_tree_lock(path->nodes[level]);
6881                         btrfs_set_lock_blocking(path->nodes[level]);
6882
6883                         ret = btrfs_lookup_extent_info(trans, root,
6884                                                 path->nodes[level]->start,
6885                                                 path->nodes[level]->len,
6886                                                 &wc->refs[level],
6887                                                 &wc->flags[level]);
6888                         if (ret < 0) {
6889                                 err = ret;
6890                                 goto out_end_trans;
6891                         }
6892                         BUG_ON(wc->refs[level] == 0);
6893
6894                         if (level == root_item->drop_level)
6895                                 break;
6896
6897                         btrfs_tree_unlock(path->nodes[level]);
6898                         WARN_ON(wc->refs[level] != 1);
6899                         level--;
6900                 }
6901         }
6902
6903         wc->level = level;
6904         wc->shared_level = -1;
6905         wc->stage = DROP_REFERENCE;
6906         wc->update_ref = update_ref;
6907         wc->keep_locks = 0;
6908         wc->for_reloc = for_reloc;
6909         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6910
6911         while (1) {
6912                 ret = walk_down_tree(trans, root, path, wc);
6913                 if (ret < 0) {
6914                         err = ret;
6915                         break;
6916                 }
6917
6918                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
6919                 if (ret < 0) {
6920                         err = ret;
6921                         break;
6922                 }
6923
6924                 if (ret > 0) {
6925                         BUG_ON(wc->stage != DROP_REFERENCE);
6926                         break;
6927                 }
6928
6929                 if (wc->stage == DROP_REFERENCE) {
6930                         level = wc->level;
6931                         btrfs_node_key(path->nodes[level],
6932                                        &root_item->drop_progress,
6933                                        path->slots[level]);
6934                         root_item->drop_level = level;
6935                 }
6936
6937                 BUG_ON(wc->level == 0);
6938                 if (btrfs_should_end_transaction(trans, tree_root)) {
6939                         ret = btrfs_update_root(trans, tree_root,
6940                                                 &root->root_key,
6941                                                 root_item);
6942                         if (ret) {
6943                                 btrfs_abort_transaction(trans, tree_root, ret);
6944                                 err = ret;
6945                                 goto out_end_trans;
6946                         }
6947
6948                         btrfs_end_transaction_throttle(trans, tree_root);
6949                         trans = btrfs_start_transaction(tree_root, 0);
6950                         if (IS_ERR(trans)) {
6951                                 err = PTR_ERR(trans);
6952                                 goto out_free;
6953                         }
6954                         if (block_rsv)
6955                                 trans->block_rsv = block_rsv;
6956                 }
6957         }
6958         btrfs_release_path(path);
6959         if (err)
6960                 goto out_end_trans;
6961
6962         ret = btrfs_del_root(trans, tree_root, &root->root_key);
6963         if (ret) {
6964                 btrfs_abort_transaction(trans, tree_root, ret);
6965                 goto out_end_trans;
6966         }
6967
6968         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
6969                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
6970                                            NULL, NULL);
6971                 if (ret < 0) {
6972                         btrfs_abort_transaction(trans, tree_root, ret);
6973                         err = ret;
6974                         goto out_end_trans;
6975                 } else if (ret > 0) {
6976                         /* if we fail to delete the orphan item this time
6977                          * around, it'll get picked up the next time.
6978                          *
6979                          * The most common failure here is just -ENOENT.
6980                          */
6981                         btrfs_del_orphan_item(trans, tree_root,
6982                                               root->root_key.objectid);
6983                 }
6984         }
6985
6986         if (root->in_radix) {
6987                 btrfs_free_fs_root(tree_root->fs_info, root);
6988         } else {
6989                 free_extent_buffer(root->node);
6990                 free_extent_buffer(root->commit_root);
6991                 kfree(root);
6992         }
6993 out_end_trans:
6994         btrfs_end_transaction_throttle(trans, tree_root);
6995 out_free:
6996         kfree(wc);
6997         btrfs_free_path(path);
6998 out:
6999         if (err)
7000                 btrfs_std_error(root->fs_info, err);
7001         return err;
7002 }
7003
7004 /*
7005  * drop subtree rooted at tree block 'node'.
7006  *
7007  * NOTE: this function will unlock and release tree block 'node'
7008  * only used by relocation code
7009  */
7010 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7011                         struct btrfs_root *root,
7012                         struct extent_buffer *node,
7013                         struct extent_buffer *parent)
7014 {
7015         struct btrfs_path *path;
7016         struct walk_control *wc;
7017         int level;
7018         int parent_level;
7019         int ret = 0;
7020         int wret;
7021
7022         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7023
7024         path = btrfs_alloc_path();
7025         if (!path)
7026                 return -ENOMEM;
7027
7028         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7029         if (!wc) {
7030                 btrfs_free_path(path);
7031                 return -ENOMEM;
7032         }
7033
7034         btrfs_assert_tree_locked(parent);
7035         parent_level = btrfs_header_level(parent);
7036         extent_buffer_get(parent);
7037         path->nodes[parent_level] = parent;
7038         path->slots[parent_level] = btrfs_header_nritems(parent);
7039
7040         btrfs_assert_tree_locked(node);
7041         level = btrfs_header_level(node);
7042         path->nodes[level] = node;
7043         path->slots[level] = 0;
7044         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7045
7046         wc->refs[parent_level] = 1;
7047         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7048         wc->level = level;
7049         wc->shared_level = -1;
7050         wc->stage = DROP_REFERENCE;
7051         wc->update_ref = 0;
7052         wc->keep_locks = 1;
7053         wc->for_reloc = 1;
7054         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7055
7056         while (1) {
7057                 wret = walk_down_tree(trans, root, path, wc);
7058                 if (wret < 0) {
7059                         ret = wret;
7060                         break;
7061                 }
7062
7063                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7064                 if (wret < 0)
7065                         ret = wret;
7066                 if (wret != 0)
7067                         break;
7068         }
7069
7070         kfree(wc);
7071         btrfs_free_path(path);
7072         return ret;
7073 }
7074
7075 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7076 {
7077         u64 num_devices;
7078         u64 stripped;
7079
7080         /*
7081          * if restripe for this chunk_type is on pick target profile and
7082          * return, otherwise do the usual balance
7083          */
7084         stripped = get_restripe_target(root->fs_info, flags);
7085         if (stripped)
7086                 return extended_to_chunk(stripped);
7087
7088         /*
7089          * we add in the count of missing devices because we want
7090          * to make sure that any RAID levels on a degraded FS
7091          * continue to be honored.
7092          */
7093         num_devices = root->fs_info->fs_devices->rw_devices +
7094                 root->fs_info->fs_devices->missing_devices;
7095
7096         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7097                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7098
7099         if (num_devices == 1) {
7100                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7101                 stripped = flags & ~stripped;
7102
7103                 /* turn raid0 into single device chunks */
7104                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7105                         return stripped;
7106
7107                 /* turn mirroring into duplication */
7108                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7109                              BTRFS_BLOCK_GROUP_RAID10))
7110                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7111         } else {
7112                 /* they already had raid on here, just return */
7113                 if (flags & stripped)
7114                         return flags;
7115
7116                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7117                 stripped = flags & ~stripped;
7118
7119                 /* switch duplicated blocks with raid1 */
7120                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7121                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7122
7123                 /* this is drive concat, leave it alone */
7124         }
7125
7126         return flags;
7127 }
7128
7129 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7130 {
7131         struct btrfs_space_info *sinfo = cache->space_info;
7132         u64 num_bytes;
7133         u64 min_allocable_bytes;
7134         int ret = -ENOSPC;
7135
7136
7137         /*
7138          * We need some metadata space and system metadata space for
7139          * allocating chunks in some corner cases until we force to set
7140          * it to be readonly.
7141          */
7142         if ((sinfo->flags &
7143              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7144             !force)
7145                 min_allocable_bytes = 1 * 1024 * 1024;
7146         else
7147                 min_allocable_bytes = 0;
7148
7149         spin_lock(&sinfo->lock);
7150         spin_lock(&cache->lock);
7151
7152         if (cache->ro) {
7153                 ret = 0;
7154                 goto out;
7155         }
7156
7157         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7158                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7159
7160         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7161             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7162             min_allocable_bytes <= sinfo->total_bytes) {
7163                 sinfo->bytes_readonly += num_bytes;
7164                 cache->ro = 1;
7165                 ret = 0;
7166         }
7167 out:
7168         spin_unlock(&cache->lock);
7169         spin_unlock(&sinfo->lock);
7170         return ret;
7171 }
7172
7173 int btrfs_set_block_group_ro(struct btrfs_root *root,
7174                              struct btrfs_block_group_cache *cache)
7175
7176 {
7177         struct btrfs_trans_handle *trans;
7178         u64 alloc_flags;
7179         int ret;
7180
7181         BUG_ON(cache->ro);
7182
7183         trans = btrfs_join_transaction(root);
7184         if (IS_ERR(trans))
7185                 return PTR_ERR(trans);
7186
7187         alloc_flags = update_block_group_flags(root, cache->flags);
7188         if (alloc_flags != cache->flags) {
7189                 ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7190                                      CHUNK_ALLOC_FORCE);
7191                 if (ret < 0)
7192                         goto out;
7193         }
7194
7195         ret = set_block_group_ro(cache, 0);
7196         if (!ret)
7197                 goto out;
7198         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7199         ret = do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7200                              CHUNK_ALLOC_FORCE);
7201         if (ret < 0)
7202                 goto out;
7203         ret = set_block_group_ro(cache, 0);
7204 out:
7205         btrfs_end_transaction(trans, root);
7206         return ret;
7207 }
7208
7209 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7210                             struct btrfs_root *root, u64 type)
7211 {
7212         u64 alloc_flags = get_alloc_profile(root, type);
7213         return do_chunk_alloc(trans, root, 2 * 1024 * 1024, alloc_flags,
7214                               CHUNK_ALLOC_FORCE);
7215 }
7216
7217 /*
7218  * helper to account the unused space of all the readonly block group in the
7219  * list. takes mirrors into account.
7220  */
7221 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7222 {
7223         struct btrfs_block_group_cache *block_group;
7224         u64 free_bytes = 0;
7225         int factor;
7226
7227         list_for_each_entry(block_group, groups_list, list) {
7228                 spin_lock(&block_group->lock);
7229
7230                 if (!block_group->ro) {
7231                         spin_unlock(&block_group->lock);
7232                         continue;
7233                 }
7234
7235                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7236                                           BTRFS_BLOCK_GROUP_RAID10 |
7237                                           BTRFS_BLOCK_GROUP_DUP))
7238                         factor = 2;
7239                 else
7240                         factor = 1;
7241
7242                 free_bytes += (block_group->key.offset -
7243                                btrfs_block_group_used(&block_group->item)) *
7244                                factor;
7245
7246                 spin_unlock(&block_group->lock);
7247         }
7248
7249         return free_bytes;
7250 }
7251
7252 /*
7253  * helper to account the unused space of all the readonly block group in the
7254  * space_info. takes mirrors into account.
7255  */
7256 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7257 {
7258         int i;
7259         u64 free_bytes = 0;
7260
7261         spin_lock(&sinfo->lock);
7262
7263         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7264                 if (!list_empty(&sinfo->block_groups[i]))
7265                         free_bytes += __btrfs_get_ro_block_group_free_space(
7266                                                 &sinfo->block_groups[i]);
7267
7268         spin_unlock(&sinfo->lock);
7269
7270         return free_bytes;
7271 }
7272
7273 void btrfs_set_block_group_rw(struct btrfs_root *root,
7274                               struct btrfs_block_group_cache *cache)
7275 {
7276         struct btrfs_space_info *sinfo = cache->space_info;
7277         u64 num_bytes;
7278
7279         BUG_ON(!cache->ro);
7280
7281         spin_lock(&sinfo->lock);
7282         spin_lock(&cache->lock);
7283         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7284                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7285         sinfo->bytes_readonly -= num_bytes;
7286         cache->ro = 0;
7287         spin_unlock(&cache->lock);
7288         spin_unlock(&sinfo->lock);
7289 }
7290
7291 /*
7292  * checks to see if its even possible to relocate this block group.
7293  *
7294  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7295  * ok to go ahead and try.
7296  */
7297 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7298 {
7299         struct btrfs_block_group_cache *block_group;
7300         struct btrfs_space_info *space_info;
7301         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7302         struct btrfs_device *device;
7303         u64 min_free;
7304         u64 dev_min = 1;
7305         u64 dev_nr = 0;
7306         u64 target;
7307         int index;
7308         int full = 0;
7309         int ret = 0;
7310
7311         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7312
7313         /* odd, couldn't find the block group, leave it alone */
7314         if (!block_group)
7315                 return -1;
7316
7317         min_free = btrfs_block_group_used(&block_group->item);
7318
7319         /* no bytes used, we're good */
7320         if (!min_free)
7321                 goto out;
7322
7323         space_info = block_group->space_info;
7324         spin_lock(&space_info->lock);
7325
7326         full = space_info->full;
7327
7328         /*
7329          * if this is the last block group we have in this space, we can't
7330          * relocate it unless we're able to allocate a new chunk below.
7331          *
7332          * Otherwise, we need to make sure we have room in the space to handle
7333          * all of the extents from this block group.  If we can, we're good
7334          */
7335         if ((space_info->total_bytes != block_group->key.offset) &&
7336             (space_info->bytes_used + space_info->bytes_reserved +
7337              space_info->bytes_pinned + space_info->bytes_readonly +
7338              min_free < space_info->total_bytes)) {
7339                 spin_unlock(&space_info->lock);
7340                 goto out;
7341         }
7342         spin_unlock(&space_info->lock);
7343
7344         /*
7345          * ok we don't have enough space, but maybe we have free space on our
7346          * devices to allocate new chunks for relocation, so loop through our
7347          * alloc devices and guess if we have enough space.  if this block
7348          * group is going to be restriped, run checks against the target
7349          * profile instead of the current one.
7350          */
7351         ret = -1;
7352
7353         /*
7354          * index:
7355          *      0: raid10
7356          *      1: raid1
7357          *      2: dup
7358          *      3: raid0
7359          *      4: single
7360          */
7361         target = get_restripe_target(root->fs_info, block_group->flags);
7362         if (target) {
7363                 index = __get_block_group_index(extended_to_chunk(target));
7364         } else {
7365                 /*
7366                  * this is just a balance, so if we were marked as full
7367                  * we know there is no space for a new chunk
7368                  */
7369                 if (full)
7370                         goto out;
7371
7372                 index = get_block_group_index(block_group);
7373         }
7374
7375         if (index == 0) {
7376                 dev_min = 4;
7377                 /* Divide by 2 */
7378                 min_free >>= 1;
7379         } else if (index == 1) {
7380                 dev_min = 2;
7381         } else if (index == 2) {
7382                 /* Multiply by 2 */
7383                 min_free <<= 1;
7384         } else if (index == 3) {
7385                 dev_min = fs_devices->rw_devices;
7386                 do_div(min_free, dev_min);
7387         }
7388
7389         mutex_lock(&root->fs_info->chunk_mutex);
7390         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7391                 u64 dev_offset;
7392
7393                 /*
7394                  * check to make sure we can actually find a chunk with enough
7395                  * space to fit our block group in.
7396                  */
7397                 if (device->total_bytes > device->bytes_used + min_free) {
7398                         ret = find_free_dev_extent(device, min_free,
7399                                                    &dev_offset, NULL);
7400                         if (!ret)
7401                                 dev_nr++;
7402
7403                         if (dev_nr >= dev_min)
7404                                 break;
7405
7406                         ret = -1;
7407                 }
7408         }
7409         mutex_unlock(&root->fs_info->chunk_mutex);
7410 out:
7411         btrfs_put_block_group(block_group);
7412         return ret;
7413 }
7414
7415 static int find_first_block_group(struct btrfs_root *root,
7416                 struct btrfs_path *path, struct btrfs_key *key)
7417 {
7418         int ret = 0;
7419         struct btrfs_key found_key;
7420         struct extent_buffer *leaf;
7421         int slot;
7422
7423         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7424         if (ret < 0)
7425                 goto out;
7426
7427         while (1) {
7428                 slot = path->slots[0];
7429                 leaf = path->nodes[0];
7430                 if (slot >= btrfs_header_nritems(leaf)) {
7431                         ret = btrfs_next_leaf(root, path);
7432                         if (ret == 0)
7433                                 continue;
7434                         if (ret < 0)
7435                                 goto out;
7436                         break;
7437                 }
7438                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7439
7440                 if (found_key.objectid >= key->objectid &&
7441                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7442                         ret = 0;
7443                         goto out;
7444                 }
7445                 path->slots[0]++;
7446         }
7447 out:
7448         return ret;
7449 }
7450
7451 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7452 {
7453         struct btrfs_block_group_cache *block_group;
7454         u64 last = 0;
7455
7456         while (1) {
7457                 struct inode *inode;
7458
7459                 block_group = btrfs_lookup_first_block_group(info, last);
7460                 while (block_group) {
7461                         spin_lock(&block_group->lock);
7462                         if (block_group->iref)
7463                                 break;
7464                         spin_unlock(&block_group->lock);
7465                         block_group = next_block_group(info->tree_root,
7466                                                        block_group);
7467                 }
7468                 if (!block_group) {
7469                         if (last == 0)
7470                                 break;
7471                         last = 0;
7472                         continue;
7473                 }
7474
7475                 inode = block_group->inode;
7476                 block_group->iref = 0;
7477                 block_group->inode = NULL;
7478                 spin_unlock(&block_group->lock);
7479                 iput(inode);
7480                 last = block_group->key.objectid + block_group->key.offset;
7481                 btrfs_put_block_group(block_group);
7482         }
7483 }
7484
7485 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7486 {
7487         struct btrfs_block_group_cache *block_group;
7488         struct btrfs_space_info *space_info;
7489         struct btrfs_caching_control *caching_ctl;
7490         struct rb_node *n;
7491
7492         down_write(&info->extent_commit_sem);
7493         while (!list_empty(&info->caching_block_groups)) {
7494                 caching_ctl = list_entry(info->caching_block_groups.next,
7495                                          struct btrfs_caching_control, list);
7496                 list_del(&caching_ctl->list);
7497                 put_caching_control(caching_ctl);
7498         }
7499         up_write(&info->extent_commit_sem);
7500
7501         spin_lock(&info->block_group_cache_lock);
7502         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7503                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7504                                        cache_node);
7505                 rb_erase(&block_group->cache_node,
7506                          &info->block_group_cache_tree);
7507                 spin_unlock(&info->block_group_cache_lock);
7508
7509                 down_write(&block_group->space_info->groups_sem);
7510                 list_del(&block_group->list);
7511                 up_write(&block_group->space_info->groups_sem);
7512
7513                 if (block_group->cached == BTRFS_CACHE_STARTED)
7514                         wait_block_group_cache_done(block_group);
7515
7516                 /*
7517                  * We haven't cached this block group, which means we could
7518                  * possibly have excluded extents on this block group.
7519                  */
7520                 if (block_group->cached == BTRFS_CACHE_NO)
7521                         free_excluded_extents(info->extent_root, block_group);
7522
7523                 btrfs_remove_free_space_cache(block_group);
7524                 btrfs_put_block_group(block_group);
7525
7526                 spin_lock(&info->block_group_cache_lock);
7527         }
7528         spin_unlock(&info->block_group_cache_lock);
7529
7530         /* now that all the block groups are freed, go through and
7531          * free all the space_info structs.  This is only called during
7532          * the final stages of unmount, and so we know nobody is
7533          * using them.  We call synchronize_rcu() once before we start,
7534          * just to be on the safe side.
7535          */
7536         synchronize_rcu();
7537
7538         release_global_block_rsv(info);
7539
7540         while(!list_empty(&info->space_info)) {
7541                 space_info = list_entry(info->space_info.next,
7542                                         struct btrfs_space_info,
7543                                         list);
7544                 if (space_info->bytes_pinned > 0 ||
7545                     space_info->bytes_reserved > 0 ||
7546                     space_info->bytes_may_use > 0) {
7547                         WARN_ON(1);
7548                         dump_space_info(space_info, 0, 0);
7549                 }
7550                 list_del(&space_info->list);
7551                 kfree(space_info);
7552         }
7553         return 0;
7554 }
7555
7556 static void __link_block_group(struct btrfs_space_info *space_info,
7557                                struct btrfs_block_group_cache *cache)
7558 {
7559         int index = get_block_group_index(cache);
7560
7561         down_write(&space_info->groups_sem);
7562         list_add_tail(&cache->list, &space_info->block_groups[index]);
7563         up_write(&space_info->groups_sem);
7564 }
7565
7566 int btrfs_read_block_groups(struct btrfs_root *root)
7567 {
7568         struct btrfs_path *path;
7569         int ret;
7570         struct btrfs_block_group_cache *cache;
7571         struct btrfs_fs_info *info = root->fs_info;
7572         struct btrfs_space_info *space_info;
7573         struct btrfs_key key;
7574         struct btrfs_key found_key;
7575         struct extent_buffer *leaf;
7576         int need_clear = 0;
7577         u64 cache_gen;
7578
7579         root = info->extent_root;
7580         key.objectid = 0;
7581         key.offset = 0;
7582         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7583         path = btrfs_alloc_path();
7584         if (!path)
7585                 return -ENOMEM;
7586         path->reada = 1;
7587
7588         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7589         if (btrfs_test_opt(root, SPACE_CACHE) &&
7590             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7591                 need_clear = 1;
7592         if (btrfs_test_opt(root, CLEAR_CACHE))
7593                 need_clear = 1;
7594
7595         while (1) {
7596                 ret = find_first_block_group(root, path, &key);
7597                 if (ret > 0)
7598                         break;
7599                 if (ret != 0)
7600                         goto error;
7601                 leaf = path->nodes[0];
7602                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7603                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7604                 if (!cache) {
7605                         ret = -ENOMEM;
7606                         goto error;
7607                 }
7608                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7609                                                 GFP_NOFS);
7610                 if (!cache->free_space_ctl) {
7611                         kfree(cache);
7612                         ret = -ENOMEM;
7613                         goto error;
7614                 }
7615
7616                 atomic_set(&cache->count, 1);
7617                 spin_lock_init(&cache->lock);
7618                 cache->fs_info = info;
7619                 INIT_LIST_HEAD(&cache->list);
7620                 INIT_LIST_HEAD(&cache->cluster_list);
7621
7622                 if (need_clear) {
7623                         /*
7624                          * When we mount with old space cache, we need to
7625                          * set BTRFS_DC_CLEAR and set dirty flag.
7626                          *
7627                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
7628                          *    truncate the old free space cache inode and
7629                          *    setup a new one.
7630                          * b) Setting 'dirty flag' makes sure that we flush
7631                          *    the new space cache info onto disk.
7632                          */
7633                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7634                         if (btrfs_test_opt(root, SPACE_CACHE))
7635                                 cache->dirty = 1;
7636                 }
7637
7638                 read_extent_buffer(leaf, &cache->item,
7639                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7640                                    sizeof(cache->item));
7641                 memcpy(&cache->key, &found_key, sizeof(found_key));
7642
7643                 key.objectid = found_key.objectid + found_key.offset;
7644                 btrfs_release_path(path);
7645                 cache->flags = btrfs_block_group_flags(&cache->item);
7646                 cache->sectorsize = root->sectorsize;
7647
7648                 btrfs_init_free_space_ctl(cache);
7649
7650                 /*
7651                  * We need to exclude the super stripes now so that the space
7652                  * info has super bytes accounted for, otherwise we'll think
7653                  * we have more space than we actually do.
7654                  */
7655                 exclude_super_stripes(root, cache);
7656
7657                 /*
7658                  * check for two cases, either we are full, and therefore
7659                  * don't need to bother with the caching work since we won't
7660                  * find any space, or we are empty, and we can just add all
7661                  * the space in and be done with it.  This saves us _alot_ of
7662                  * time, particularly in the full case.
7663                  */
7664                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7665                         cache->last_byte_to_unpin = (u64)-1;
7666                         cache->cached = BTRFS_CACHE_FINISHED;
7667                         free_excluded_extents(root, cache);
7668                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7669                         cache->last_byte_to_unpin = (u64)-1;
7670                         cache->cached = BTRFS_CACHE_FINISHED;
7671                         add_new_free_space(cache, root->fs_info,
7672                                            found_key.objectid,
7673                                            found_key.objectid +
7674                                            found_key.offset);
7675                         free_excluded_extents(root, cache);
7676                 }
7677
7678                 ret = update_space_info(info, cache->flags, found_key.offset,
7679                                         btrfs_block_group_used(&cache->item),
7680                                         &space_info);
7681                 BUG_ON(ret); /* -ENOMEM */
7682                 cache->space_info = space_info;
7683                 spin_lock(&cache->space_info->lock);
7684                 cache->space_info->bytes_readonly += cache->bytes_super;
7685                 spin_unlock(&cache->space_info->lock);
7686
7687                 __link_block_group(space_info, cache);
7688
7689                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7690                 BUG_ON(ret); /* Logic error */
7691
7692                 set_avail_alloc_bits(root->fs_info, cache->flags);
7693                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7694                         set_block_group_ro(cache, 1);
7695         }
7696
7697         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7698                 if (!(get_alloc_profile(root, space_info->flags) &
7699                       (BTRFS_BLOCK_GROUP_RAID10 |
7700                        BTRFS_BLOCK_GROUP_RAID1 |
7701                        BTRFS_BLOCK_GROUP_DUP)))
7702                         continue;
7703                 /*
7704                  * avoid allocating from un-mirrored block group if there are
7705                  * mirrored block groups.
7706                  */
7707                 list_for_each_entry(cache, &space_info->block_groups[3], list)
7708                         set_block_group_ro(cache, 1);
7709                 list_for_each_entry(cache, &space_info->block_groups[4], list)
7710                         set_block_group_ro(cache, 1);
7711         }
7712
7713         init_global_block_rsv(info);
7714         ret = 0;
7715 error:
7716         btrfs_free_path(path);
7717         return ret;
7718 }
7719
7720 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7721                            struct btrfs_root *root, u64 bytes_used,
7722                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7723                            u64 size)
7724 {
7725         int ret;
7726         struct btrfs_root *extent_root;
7727         struct btrfs_block_group_cache *cache;
7728
7729         extent_root = root->fs_info->extent_root;
7730
7731         root->fs_info->last_trans_log_full_commit = trans->transid;
7732
7733         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7734         if (!cache)
7735                 return -ENOMEM;
7736         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7737                                         GFP_NOFS);
7738         if (!cache->free_space_ctl) {
7739                 kfree(cache);
7740                 return -ENOMEM;
7741         }
7742
7743         cache->key.objectid = chunk_offset;
7744         cache->key.offset = size;
7745         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7746         cache->sectorsize = root->sectorsize;
7747         cache->fs_info = root->fs_info;
7748
7749         atomic_set(&cache->count, 1);
7750         spin_lock_init(&cache->lock);
7751         INIT_LIST_HEAD(&cache->list);
7752         INIT_LIST_HEAD(&cache->cluster_list);
7753
7754         btrfs_init_free_space_ctl(cache);
7755
7756         btrfs_set_block_group_used(&cache->item, bytes_used);
7757         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7758         cache->flags = type;
7759         btrfs_set_block_group_flags(&cache->item, type);
7760
7761         cache->last_byte_to_unpin = (u64)-1;
7762         cache->cached = BTRFS_CACHE_FINISHED;
7763         exclude_super_stripes(root, cache);
7764
7765         add_new_free_space(cache, root->fs_info, chunk_offset,
7766                            chunk_offset + size);
7767
7768         free_excluded_extents(root, cache);
7769
7770         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7771                                 &cache->space_info);
7772         BUG_ON(ret); /* -ENOMEM */
7773         update_global_block_rsv(root->fs_info);
7774
7775         spin_lock(&cache->space_info->lock);
7776         cache->space_info->bytes_readonly += cache->bytes_super;
7777         spin_unlock(&cache->space_info->lock);
7778
7779         __link_block_group(cache->space_info, cache);
7780
7781         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7782         BUG_ON(ret); /* Logic error */
7783
7784         ret = btrfs_insert_item(trans, extent_root, &cache->key, &cache->item,
7785                                 sizeof(cache->item));
7786         if (ret) {
7787                 btrfs_abort_transaction(trans, extent_root, ret);
7788                 return ret;
7789         }
7790
7791         set_avail_alloc_bits(extent_root->fs_info, type);
7792
7793         return 0;
7794 }
7795
7796 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
7797 {
7798         u64 extra_flags = chunk_to_extended(flags) &
7799                                 BTRFS_EXTENDED_PROFILE_MASK;
7800
7801         if (flags & BTRFS_BLOCK_GROUP_DATA)
7802                 fs_info->avail_data_alloc_bits &= ~extra_flags;
7803         if (flags & BTRFS_BLOCK_GROUP_METADATA)
7804                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
7805         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
7806                 fs_info->avail_system_alloc_bits &= ~extra_flags;
7807 }
7808
7809 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7810                              struct btrfs_root *root, u64 group_start)
7811 {
7812         struct btrfs_path *path;
7813         struct btrfs_block_group_cache *block_group;
7814         struct btrfs_free_cluster *cluster;
7815         struct btrfs_root *tree_root = root->fs_info->tree_root;
7816         struct btrfs_key key;
7817         struct inode *inode;
7818         int ret;
7819         int index;
7820         int factor;
7821
7822         root = root->fs_info->extent_root;
7823
7824         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7825         BUG_ON(!block_group);
7826         BUG_ON(!block_group->ro);
7827
7828         /*
7829          * Free the reserved super bytes from this block group before
7830          * remove it.
7831          */
7832         free_excluded_extents(root, block_group);
7833
7834         memcpy(&key, &block_group->key, sizeof(key));
7835         index = get_block_group_index(block_group);
7836         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7837                                   BTRFS_BLOCK_GROUP_RAID1 |
7838                                   BTRFS_BLOCK_GROUP_RAID10))
7839                 factor = 2;
7840         else
7841                 factor = 1;
7842
7843         /* make sure this block group isn't part of an allocation cluster */
7844         cluster = &root->fs_info->data_alloc_cluster;
7845         spin_lock(&cluster->refill_lock);
7846         btrfs_return_cluster_to_free_space(block_group, cluster);
7847         spin_unlock(&cluster->refill_lock);
7848
7849         /*
7850          * make sure this block group isn't part of a metadata
7851          * allocation cluster
7852          */
7853         cluster = &root->fs_info->meta_alloc_cluster;
7854         spin_lock(&cluster->refill_lock);
7855         btrfs_return_cluster_to_free_space(block_group, cluster);
7856         spin_unlock(&cluster->refill_lock);
7857
7858         path = btrfs_alloc_path();
7859         if (!path) {
7860                 ret = -ENOMEM;
7861                 goto out;
7862         }
7863
7864         inode = lookup_free_space_inode(tree_root, block_group, path);
7865         if (!IS_ERR(inode)) {
7866                 ret = btrfs_orphan_add(trans, inode);
7867                 if (ret) {
7868                         btrfs_add_delayed_iput(inode);
7869                         goto out;
7870                 }
7871                 clear_nlink(inode);
7872                 /* One for the block groups ref */
7873                 spin_lock(&block_group->lock);
7874                 if (block_group->iref) {
7875                         block_group->iref = 0;
7876                         block_group->inode = NULL;
7877                         spin_unlock(&block_group->lock);
7878                         iput(inode);
7879                 } else {
7880                         spin_unlock(&block_group->lock);
7881                 }
7882                 /* One for our lookup ref */
7883                 btrfs_add_delayed_iput(inode);
7884         }
7885
7886         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
7887         key.offset = block_group->key.objectid;
7888         key.type = 0;
7889
7890         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
7891         if (ret < 0)
7892                 goto out;
7893         if (ret > 0)
7894                 btrfs_release_path(path);
7895         if (ret == 0) {
7896                 ret = btrfs_del_item(trans, tree_root, path);
7897                 if (ret)
7898                         goto out;
7899                 btrfs_release_path(path);
7900         }
7901
7902         spin_lock(&root->fs_info->block_group_cache_lock);
7903         rb_erase(&block_group->cache_node,
7904                  &root->fs_info->block_group_cache_tree);
7905         spin_unlock(&root->fs_info->block_group_cache_lock);
7906
7907         down_write(&block_group->space_info->groups_sem);
7908         /*
7909          * we must use list_del_init so people can check to see if they
7910          * are still on the list after taking the semaphore
7911          */
7912         list_del_init(&block_group->list);
7913         if (list_empty(&block_group->space_info->block_groups[index]))
7914                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
7915         up_write(&block_group->space_info->groups_sem);
7916
7917         if (block_group->cached == BTRFS_CACHE_STARTED)
7918                 wait_block_group_cache_done(block_group);
7919
7920         btrfs_remove_free_space_cache(block_group);
7921
7922         spin_lock(&block_group->space_info->lock);
7923         block_group->space_info->total_bytes -= block_group->key.offset;
7924         block_group->space_info->bytes_readonly -= block_group->key.offset;
7925         block_group->space_info->disk_total -= block_group->key.offset * factor;
7926         spin_unlock(&block_group->space_info->lock);
7927
7928         memcpy(&key, &block_group->key, sizeof(key));
7929
7930         btrfs_clear_space_info_full(root->fs_info);
7931
7932         btrfs_put_block_group(block_group);
7933         btrfs_put_block_group(block_group);
7934
7935         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
7936         if (ret > 0)
7937                 ret = -EIO;
7938         if (ret < 0)
7939                 goto out;
7940
7941         ret = btrfs_del_item(trans, root, path);
7942 out:
7943         btrfs_free_path(path);
7944         return ret;
7945 }
7946
7947 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
7948 {
7949         struct btrfs_space_info *space_info;
7950         struct btrfs_super_block *disk_super;
7951         u64 features;
7952         u64 flags;
7953         int mixed = 0;
7954         int ret;
7955
7956         disk_super = fs_info->super_copy;
7957         if (!btrfs_super_root(disk_super))
7958                 return 1;
7959
7960         features = btrfs_super_incompat_flags(disk_super);
7961         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
7962                 mixed = 1;
7963
7964         flags = BTRFS_BLOCK_GROUP_SYSTEM;
7965         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7966         if (ret)
7967                 goto out;
7968
7969         if (mixed) {
7970                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
7971                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7972         } else {
7973                 flags = BTRFS_BLOCK_GROUP_METADATA;
7974                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7975                 if (ret)
7976                         goto out;
7977
7978                 flags = BTRFS_BLOCK_GROUP_DATA;
7979                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
7980         }
7981 out:
7982         return ret;
7983 }
7984
7985 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
7986 {
7987         return unpin_extent_range(root, start, end);
7988 }
7989
7990 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
7991                                u64 num_bytes, u64 *actual_bytes)
7992 {
7993         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
7994 }
7995
7996 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
7997 {
7998         struct btrfs_fs_info *fs_info = root->fs_info;
7999         struct btrfs_block_group_cache *cache = NULL;
8000         u64 group_trimmed;
8001         u64 start;
8002         u64 end;
8003         u64 trimmed = 0;
8004         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8005         int ret = 0;
8006
8007         /*
8008          * try to trim all FS space, our block group may start from non-zero.
8009          */
8010         if (range->len == total_bytes)
8011                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8012         else
8013                 cache = btrfs_lookup_block_group(fs_info, range->start);
8014
8015         while (cache) {
8016                 if (cache->key.objectid >= (range->start + range->len)) {
8017                         btrfs_put_block_group(cache);
8018                         break;
8019                 }
8020
8021                 start = max(range->start, cache->key.objectid);
8022                 end = min(range->start + range->len,
8023                                 cache->key.objectid + cache->key.offset);
8024
8025                 if (end - start >= range->minlen) {
8026                         if (!block_group_cache_done(cache)) {
8027                                 ret = cache_block_group(cache, NULL, root, 0);
8028                                 if (!ret)
8029                                         wait_block_group_cache_done(cache);
8030                         }
8031                         ret = btrfs_trim_block_group(cache,
8032                                                      &group_trimmed,
8033                                                      start,
8034                                                      end,
8035                                                      range->minlen);
8036
8037                         trimmed += group_trimmed;
8038                         if (ret) {
8039                                 btrfs_put_block_group(cache);
8040                                 break;
8041                         }
8042                 }
8043
8044                 cache = next_block_group(fs_info->tree_root, cache);
8045         }
8046
8047         range->len = trimmed;
8048         return ret;
8049 }