fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include "compat.h"
  28 #include "hash.h"
  29 #include "ctree.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "transaction.h"
  33 #include "volumes.h"
  34 #include "locking.h"
  35 #include "free-space-cache.h"
  36
  37 #undef SCRAMBLE_DELAYED_REFS
  38
  39 /*
  40  * control flags for do_chunk_alloc's force field
  41  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  42  * if we really need one.
  43  *
  44  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  45  * if we have very few chunks already allocated.  This is
  46  * used as part of the clustering code to help make sure
  47  * we have a good pool of storage to cluster in, without
  48  * filling the FS with empty chunks
  49  *
  50  * CHUNK_ALLOC_FORCE means it must try to allocate one
  51  *
  52  */
  53 enum {
  54         CHUNK_ALLOC_NO_FORCE = 0,
  55         CHUNK_ALLOC_LIMITED = 1,
  56         CHUNK_ALLOC_FORCE = 2,
  57 };
  58
  59 /*
  60  * Control how reservations are dealt with.
  61  *
  62  * RESERVE_FREE - freeing a reservation.
  63  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  64  *   ENOSPC accounting
  65  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  66  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  67  */
  68 enum {
  69         RESERVE_FREE = 0,
  70         RESERVE_ALLOC = 1,
  71         RESERVE_ALLOC_NO_ACCOUNT = 2,
  72 };
  73
  74 static int update_block_group(struct btrfs_trans_handle *trans,
  75                               struct btrfs_root *root,
  76                               u64 bytenr, u64 num_bytes, int alloc);
  77 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  78                                 struct btrfs_root *root,
  79                                 u64 bytenr, u64 num_bytes, u64 parent,
  80                                 u64 root_objectid, u64 owner_objectid,
  81                                 u64 owner_offset, int refs_to_drop,
  82                                 struct btrfs_delayed_extent_op *extra_op);
  83 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  84                                     struct extent_buffer *leaf,
  85                                     struct btrfs_extent_item *ei);
  86 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  87                                       struct btrfs_root *root,
  88                                       u64 parent, u64 root_objectid,
  89                                       u64 flags, u64 owner, u64 offset,
  90                                       struct btrfs_key *ins, int ref_mod);
  91 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  92                                      struct btrfs_root *root,
  93                                      u64 parent, u64 root_objectid,
  94                                      u64 flags, struct btrfs_disk_key *key,
  95                                      int level, struct btrfs_key *ins);
  96 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  97                           struct btrfs_root *extent_root, u64 flags,
  98                           int force);
  99 static int find_next_key(struct btrfs_path *path, int level,
 100                          struct btrfs_key *key);
 101 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 102                             int dump_block_groups);
 103 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 104                                        u64 num_bytes, int reserve);
 105
 106 static noinline int
 107 block_group_cache_done(struct btrfs_block_group_cache *cache)
 108 {
 109         smp_mb();
 110         return cache->cached == BTRFS_CACHE_FINISHED;
 111 }
 112
 113 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 114 {
 115         return (cache->flags & bits) == bits;
 116 }
 117
 118 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 119 {
 120         atomic_inc(&cache->count);
 121 }
 122
 123 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 124 {
 125         if (atomic_dec_and_test(&cache->count)) {
 126                 WARN_ON(cache->pinned > 0);
 127                 WARN_ON(cache->reserved > 0);
 128                 kfree(cache->free_space_ctl);
 129                 kfree(cache);
 130         }
 131 }
 132
 133 /*
 134  * this adds the block group to the fs_info rb tree for the block group
 135  * cache
 136  */
 137 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 138                                 struct btrfs_block_group_cache *block_group)
 139 {
 140         struct rb_node **p;
 141         struct rb_node *parent = NULL;
 142         struct btrfs_block_group_cache *cache;
 143
 144         spin_lock(&info->block_group_cache_lock);
 145         p = &info->block_group_cache_tree.rb_node;
 146
 147         while (*p) {
 148                 parent = *p;
 149                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 150                                  cache_node);
 151                 if (block_group->key.objectid < cache->key.objectid) {
 152                         p = &(*p)->rb_left;
 153                 } else if (block_group->key.objectid > cache->key.objectid) {
 154                         p = &(*p)->rb_right;
 155                 } else {
 156                         spin_unlock(&info->block_group_cache_lock);
 157                         return -EEXIST;
 158                 }
 159         }
 160
 161         rb_link_node(&block_group->cache_node, parent, p);
 162         rb_insert_color(&block_group->cache_node,
 163                         &info->block_group_cache_tree);
 164         spin_unlock(&info->block_group_cache_lock);
 165
 166         return 0;
 167 }
 168
 169 /*
 170  * This will return the block group at or after bytenr if contains is 0, else
 171  * it will return the block group that contains the bytenr
 172  */
 173 static struct btrfs_block_group_cache *
 174 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 175                               int contains)
 176 {
 177         struct btrfs_block_group_cache *cache, *ret = NULL;
 178         struct rb_node *n;
 179         u64 end, start;
 180
 181         spin_lock(&info->block_group_cache_lock);
 182         n = info->block_group_cache_tree.rb_node;
 183
 184         while (n) {
 185                 cache = rb_entry(n, struct btrfs_block_group_cache,
 186                                  cache_node);
 187                 end = cache->key.objectid + cache->key.offset - 1;
 188                 start = cache->key.objectid;
 189
 190                 if (bytenr < start) {
 191                         if (!contains && (!ret || start < ret->key.objectid))
 192                                 ret = cache;
 193                         n = n->rb_left;
 194                 } else if (bytenr > start) {
 195                         if (contains && bytenr <= end) {
 196                                 ret = cache;
 197                                 break;
 198                         }
 199                         n = n->rb_right;
 200                 } else {
 201                         ret = cache;
 202                         break;
 203                 }
 204         }
 205         if (ret)
 206                 btrfs_get_block_group(ret);
 207         spin_unlock(&info->block_group_cache_lock);
 208
 209         return ret;
 210 }
 211
 212 static int add_excluded_extent(struct btrfs_root *root,
 213                                u64 start, u64 num_bytes)
 214 {
 215         u64 end = start + num_bytes - 1;
 216         set_extent_bits(&root->fs_info->freed_extents[0],
 217                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 218         set_extent_bits(&root->fs_info->freed_extents[1],
 219                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 220         return 0;
 221 }
 222
 223 static void free_excluded_extents(struct btrfs_root *root,
 224                                   struct btrfs_block_group_cache *cache)
 225 {
 226         u64 start, end;
 227
 228         start = cache->key.objectid;
 229         end = start + cache->key.offset - 1;
 230
 231         clear_extent_bits(&root->fs_info->freed_extents[0],
 232                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 233         clear_extent_bits(&root->fs_info->freed_extents[1],
 234                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 235 }
 236
 237 static int exclude_super_stripes(struct btrfs_root *root,
 238                                  struct btrfs_block_group_cache *cache)
 239 {
 240         u64 bytenr;
 241         u64 *logical;
 242         int stripe_len;
 243         int i, nr, ret;
 244
 245         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 246                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 247                 cache->bytes_super += stripe_len;
 248                 ret = add_excluded_extent(root, cache->key.objectid,
 249                                           stripe_len);
 250                 BUG_ON(ret); /* -ENOMEM */
 251         }
 252
 253         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 254                 bytenr = btrfs_sb_offset(i);
 255                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 256                                        cache->key.objectid, bytenr,
 257                                        0, &logical, &nr, &stripe_len);
 258                 BUG_ON(ret); /* -ENOMEM */
 259
 260                 while (nr--) {
 261                         cache->bytes_super += stripe_len;
 262                         ret = add_excluded_extent(root, logical[nr],
 263                                                   stripe_len);
 264                         BUG_ON(ret); /* -ENOMEM */
 265                 }
 266
 267                 kfree(logical);
 268         }
 269         return 0;
 270 }
 271
 272 static struct btrfs_caching_control *
 273 get_caching_control(struct btrfs_block_group_cache *cache)
 274 {
 275         struct btrfs_caching_control *ctl;
 276
 277         spin_lock(&cache->lock);
 278         if (cache->cached != BTRFS_CACHE_STARTED) {
 279                 spin_unlock(&cache->lock);
 280                 return NULL;
 281         }
 282
 283         /* We're loading it the fast way, so we don't have a caching_ctl. */
 284         if (!cache->caching_ctl) {
 285                 spin_unlock(&cache->lock);
 286                 return NULL;
 287         }
 288
 289         ctl = cache->caching_ctl;
 290         atomic_inc(&ctl->count);
 291         spin_unlock(&cache->lock);
 292         return ctl;
 293 }
 294
 295 static void put_caching_control(struct btrfs_caching_control *ctl)
 296 {
 297         if (atomic_dec_and_test(&ctl->count))
 298                 kfree(ctl);
 299 }
 300
 301 /*
 302  * this is only called by cache_block_group, since we could have freed extents
 303  * we need to check the pinned_extents for any extents that can't be used yet
 304  * since their free space will be released as soon as the transaction commits.
 305  */
 306 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 307                               struct btrfs_fs_info *info, u64 start, u64 end)
 308 {
 309         u64 extent_start, extent_end, size, total_added = 0;
 310         int ret;
 311
 312         while (start < end) {
 313                 ret = find_first_extent_bit(info->pinned_extents, start,
 314                                             &extent_start, &extent_end,
 315                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 316                                             NULL);
 317                 if (ret)
 318                         break;
 319
 320                 if (extent_start <= start) {
 321                         start = extent_end + 1;
 322                 } else if (extent_start > start && extent_start < end) {
 323                         size = extent_start - start;
 324                         total_added += size;
 325                         ret = btrfs_add_free_space(block_group, start,
 326                                                    size);
 327                         BUG_ON(ret); /* -ENOMEM or logic error */
 328                         start = extent_end + 1;
 329                 } else {
 330                         break;
 331                 }
 332         }
 333
 334         if (start < end) {
 335                 size = end - start;
 336                 total_added += size;
 337                 ret = btrfs_add_free_space(block_group, start, size);
 338                 BUG_ON(ret); /* -ENOMEM or logic error */
 339         }
 340
 341         return total_added;
 342 }
 343
 344 static noinline void caching_thread(struct btrfs_work *work)
 345 {
 346         struct btrfs_block_group_cache *block_group;
 347         struct btrfs_fs_info *fs_info;
 348         struct btrfs_caching_control *caching_ctl;
 349         struct btrfs_root *extent_root;
 350         struct btrfs_path *path;
 351         struct extent_buffer *leaf;
 352         struct btrfs_key key;
 353         u64 total_found = 0;
 354         u64 last = 0;
 355         u32 nritems;
 356         int ret = 0;
 357
 358         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 359         block_group = caching_ctl->block_group;
 360         fs_info = block_group->fs_info;
 361         extent_root = fs_info->extent_root;
 362
 363         path = btrfs_alloc_path();
 364         if (!path)
 365                 goto out;
 366
 367         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 368
 369         /*
 370          * We don't want to deadlock with somebody trying to allocate a new
 371          * extent for the extent root while also trying to search the extent
 372          * root to add free space.  So we skip locking and search the commit
 373          * root, since its read-only
 374          */
 375         path->skip_locking = 1;
 376         path->search_commit_root = 1;
 377         path->reada = 1;
 378
 379         key.objectid = last;
 380         key.offset = 0;
 381         key.type = BTRFS_EXTENT_ITEM_KEY;
 382 again:
 383         mutex_lock(&caching_ctl->mutex);
 384         /* need to make sure the commit_root doesn't disappear */
 385         down_read(&fs_info->extent_commit_sem);
 386
 387         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 388         if (ret < 0)
 389                 goto err;
 390
 391         leaf = path->nodes[0];
 392         nritems = btrfs_header_nritems(leaf);
 393
 394         while (1) {
 395                 if (btrfs_fs_closing(fs_info) > 1) {
 396                         last = (u64)-1;
 397                         break;
 398                 }
 399
 400                 if (path->slots[0] < nritems) {
 401                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 402                 } else {
 403                         ret = find_next_key(path, 0, &key);
 404                         if (ret)
 405                                 break;
 406
 407                         if (need_resched() ||
 408                             btrfs_next_leaf(extent_root, path)) {
 409                                 caching_ctl->progress = last;
 410                                 btrfs_release_path(path);
 411                                 up_read(&fs_info->extent_commit_sem);
 412                                 mutex_unlock(&caching_ctl->mutex);
 413                                 cond_resched();
 414                                 goto again;
 415                         }
 416                         leaf = path->nodes[0];
 417                         nritems = btrfs_header_nritems(leaf);
 418                         continue;
 419                 }
 420
 421                 if (key.objectid < block_group->key.objectid) {
 422                         path->slots[0]++;
 423                         continue;
 424                 }
 425
 426                 if (key.objectid >= block_group->key.objectid +
 427                     block_group->key.offset)
 428                         break;
 429
 430                 if (key.type == BTRFS_EXTENT_ITEM_KEY) {
 431                         total_found += add_new_free_space(block_group,
 432                                                           fs_info, last,
 433                                                           key.objectid);
 434                         last = key.objectid + key.offset;
 435
 436                         if (total_found > (1024 * 1024 * 2)) {
 437                                 total_found = 0;
 438                                 wake_up(&caching_ctl->wait);
 439                         }
 440                 }
 441                 path->slots[0]++;
 442         }
 443         ret = 0;
 444
 445         total_found += add_new_free_space(block_group, fs_info, last,
 446                                           block_group->key.objectid +
 447                                           block_group->key.offset);
 448         caching_ctl->progress = (u64)-1;
 449
 450         spin_lock(&block_group->lock);
 451         block_group->caching_ctl = NULL;
 452         block_group->cached = BTRFS_CACHE_FINISHED;
 453         spin_unlock(&block_group->lock);
 454
 455 err:
 456         btrfs_free_path(path);
 457         up_read(&fs_info->extent_commit_sem);
 458
 459         free_excluded_extents(extent_root, block_group);
 460
 461         mutex_unlock(&caching_ctl->mutex);
 462 out:
 463         wake_up(&caching_ctl->wait);
 464
 465         put_caching_control(caching_ctl);
 466         btrfs_put_block_group(block_group);
 467 }
 468
 469 static int cache_block_group(struct btrfs_block_group_cache *cache,
 470                              struct btrfs_trans_handle *trans,
 471                              struct btrfs_root *root,
 472                              int load_cache_only)
 473 {
 474         DEFINE_WAIT(wait);
 475         struct btrfs_fs_info *fs_info = cache->fs_info;
 476         struct btrfs_caching_control *caching_ctl;
 477         int ret = 0;
 478
 479         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 480         if (!caching_ctl)
 481                 return -ENOMEM;
 482
 483         INIT_LIST_HEAD(&caching_ctl->list);
 484         mutex_init(&caching_ctl->mutex);
 485         init_waitqueue_head(&caching_ctl->wait);
 486         caching_ctl->block_group = cache;
 487         caching_ctl->progress = cache->key.objectid;
 488         atomic_set(&caching_ctl->count, 1);
 489         caching_ctl->work.func = caching_thread;
 490
 491         spin_lock(&cache->lock);
 492         /*
 493          * This should be a rare occasion, but this could happen I think in the
 494          * case where one thread starts to load the space cache info, and then
 495          * some other thread starts a transaction commit which tries to do an
 496          * allocation while the other thread is still loading the space cache
 497          * info.  The previous loop should have kept us from choosing this block
 498          * group, but if we've moved to the state where we will wait on caching
 499          * block groups we need to first check if we're doing a fast load here,
 500          * so we can wait for it to finish, otherwise we could end up allocating
 501          * from a block group who's cache gets evicted for one reason or
 502          * another.
 503          */
 504         while (cache->cached == BTRFS_CACHE_FAST) {
 505                 struct btrfs_caching_control *ctl;
 506
 507                 ctl = cache->caching_ctl;
 508                 atomic_inc(&ctl->count);
 509                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 510                 spin_unlock(&cache->lock);
 511
 512                 schedule();
 513
 514                 finish_wait(&ctl->wait, &wait);
 515                 put_caching_control(ctl);
 516                 spin_lock(&cache->lock);
 517         }
 518
 519         if (cache->cached != BTRFS_CACHE_NO) {
 520                 spin_unlock(&cache->lock);
 521                 kfree(caching_ctl);
 522                 return 0;
 523         }
 524         WARN_ON(cache->caching_ctl);
 525         cache->caching_ctl = caching_ctl;
 526         cache->cached = BTRFS_CACHE_FAST;
 527         spin_unlock(&cache->lock);
 528
 529         /*
 530          * We can't do the read from on-disk cache during a commit since we need
 531          * to have the normal tree locking.  Also if we are currently trying to
 532          * allocate blocks for the tree root we can't do the fast caching since
 533          * we likely hold important locks.
 534          */
 535         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 536                 ret = load_free_space_cache(fs_info, cache);
 537
 538                 spin_lock(&cache->lock);
 539                 if (ret == 1) {
 540                         cache->caching_ctl = NULL;
 541                         cache->cached = BTRFS_CACHE_FINISHED;
 542                         cache->last_byte_to_unpin = (u64)-1;
 543                 } else {
 544                         if (load_cache_only) {
 545                                 cache->caching_ctl = NULL;
 546                                 cache->cached = BTRFS_CACHE_NO;
 547                         } else {
 548                                 cache->cached = BTRFS_CACHE_STARTED;
 549                         }
 550                 }
 551                 spin_unlock(&cache->lock);
 552                 wake_up(&caching_ctl->wait);
 553                 if (ret == 1) {
 554                         put_caching_control(caching_ctl);
 555                         free_excluded_extents(fs_info->extent_root, cache);
 556                         return 0;
 557                 }
 558         } else {
 559                 /*
 560                  * We are not going to do the fast caching, set cached to the
 561                  * appropriate value and wakeup any waiters.
 562                  */
 563                 spin_lock(&cache->lock);
 564                 if (load_cache_only) {
 565                         cache->caching_ctl = NULL;
 566                         cache->cached = BTRFS_CACHE_NO;
 567                 } else {
 568                         cache->cached = BTRFS_CACHE_STARTED;
 569                 }
 570                 spin_unlock(&cache->lock);
 571                 wake_up(&caching_ctl->wait);
 572         }
 573
 574         if (load_cache_only) {
 575                 put_caching_control(caching_ctl);
 576                 return 0;
 577         }
 578
 579         down_write(&fs_info->extent_commit_sem);
 580         atomic_inc(&caching_ctl->count);
 581         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 582         up_write(&fs_info->extent_commit_sem);
 583
 584         btrfs_get_block_group(cache);
 585
 586         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 587
 588         return ret;
 589 }
 590
 591 /*
 592  * return the block group that starts at or after bytenr
 593  */
 594 static struct btrfs_block_group_cache *
 595 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 596 {
 597         struct btrfs_block_group_cache *cache;
 598
 599         cache = block_group_cache_tree_search(info, bytenr, 0);
 600
 601         return cache;
 602 }
 603
 604 /*
 605  * return the block group that contains the given bytenr
 606  */
 607 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 608                                                  struct btrfs_fs_info *info,
 609                                                  u64 bytenr)
 610 {
 611         struct btrfs_block_group_cache *cache;
 612
 613         cache = block_group_cache_tree_search(info, bytenr, 1);
 614
 615         return cache;
 616 }
 617
 618 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 619                                                   u64 flags)
 620 {
 621         struct list_head *head = &info->space_info;
 622         struct btrfs_space_info *found;
 623
 624         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 625
 626         rcu_read_lock();
 627         list_for_each_entry_rcu(found, head, list) {
 628                 if (found->flags & flags) {
 629                         rcu_read_unlock();
 630                         return found;
 631                 }
 632         }
 633         rcu_read_unlock();
 634         return NULL;
 635 }
 636
 637 /*
 638  * after adding space to the filesystem, we need to clear the full flags
 639  * on all the space infos.
 640  */
 641 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 642 {
 643         struct list_head *head = &info->space_info;
 644         struct btrfs_space_info *found;
 645
 646         rcu_read_lock();
 647         list_for_each_entry_rcu(found, head, list)
 648                 found->full = 0;
 649         rcu_read_unlock();
 650 }
 651
 652 static u64 div_factor(u64 num, int factor)
 653 {
 654         if (factor == 10)
 655                 return num;
 656         num *= factor;
 657         do_div(num, 10);
 658         return num;
 659 }
 660
 661 static u64 div_factor_fine(u64 num, int factor)
 662 {
 663         if (factor == 100)
 664                 return num;
 665         num *= factor;
 666         do_div(num, 100);
 667         return num;
 668 }
 669
 670 u64 btrfs_find_block_group(struct btrfs_root *root,
 671                            u64 search_start, u64 search_hint, int owner)
 672 {
 673         struct btrfs_block_group_cache *cache;
 674         u64 used;
 675         u64 last = max(search_hint, search_start);
 676         u64 group_start = 0;
 677         int full_search = 0;
 678         int factor = 9;
 679         int wrapped = 0;
 680 again:
 681         while (1) {
 682                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
 683                 if (!cache)
 684                         break;
 685
 686                 spin_lock(&cache->lock);
 687                 last = cache->key.objectid + cache->key.offset;
 688                 used = btrfs_block_group_used(&cache->item);
 689
 690                 if ((full_search || !cache->ro) &&
 691                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
 692                         if (used + cache->pinned + cache->reserved <
 693                             div_factor(cache->key.offset, factor)) {
 694                                 group_start = cache->key.objectid;
 695                                 spin_unlock(&cache->lock);
 696                                 btrfs_put_block_group(cache);
 697                                 goto found;
 698                         }
 699                 }
 700                 spin_unlock(&cache->lock);
 701                 btrfs_put_block_group(cache);
 702                 cond_resched();
 703         }
 704         if (!wrapped) {
 705                 last = search_start;
 706                 wrapped = 1;
 707                 goto again;
 708         }
 709         if (!full_search && factor < 10) {
 710                 last = search_start;
 711                 full_search = 1;
 712                 factor = 10;
 713                 goto again;
 714         }
 715 found:
 716         return group_start;
 717 }
 718
 719 /* simple helper to search for an existing extent at a given offset */
 720 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 721 {
 722         int ret;
 723         struct btrfs_key key;
 724         struct btrfs_path *path;
 725
 726         path = btrfs_alloc_path();
 727         if (!path)
 728                 return -ENOMEM;
 729
 730         key.objectid = start;
 731         key.offset = len;
 732         btrfs_set_key_type(&key, BTRFS_EXTENT_ITEM_KEY);
 733         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 734                                 0, 0);
 735         btrfs_free_path(path);
 736         return ret;
 737 }
 738
 739 /*
 740  * helper function to lookup reference count and flags of extent.
 741  *
 742  * the head node for delayed ref is used to store the sum of all the
 743  * reference count modifications queued up in the rbtree. the head
 744  * node may also store the extent flags to set. This way you can check
 745  * to see what the reference count and extent flags would be if all of
 746  * the delayed refs are not processed.
 747  */
 748 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 749                              struct btrfs_root *root, u64 bytenr,
 750                              u64 num_bytes, u64 *refs, u64 *flags)
 751 {
 752         struct btrfs_delayed_ref_head *head;
 753         struct btrfs_delayed_ref_root *delayed_refs;
 754         struct btrfs_path *path;
 755         struct btrfs_extent_item *ei;
 756         struct extent_buffer *leaf;
 757         struct btrfs_key key;
 758         u32 item_size;
 759         u64 num_refs;
 760         u64 extent_flags;
 761         int ret;
 762
 763         path = btrfs_alloc_path();
 764         if (!path)
 765                 return -ENOMEM;
 766
 767         key.objectid = bytenr;
 768         key.type = BTRFS_EXTENT_ITEM_KEY;
 769         key.offset = num_bytes;
 770         if (!trans) {
 771                 path->skip_locking = 1;
 772                 path->search_commit_root = 1;
 773         }
 774 again:
 775         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 776                                 &key, path, 0, 0);
 777         if (ret < 0)
 778                 goto out_free;
 779
 780         if (ret == 0) {
 781                 leaf = path->nodes[0];
 782                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 783                 if (item_size >= sizeof(*ei)) {
 784                         ei = btrfs_item_ptr(leaf, path->slots[0],
 785                                             struct btrfs_extent_item);
 786                         num_refs = btrfs_extent_refs(leaf, ei);
 787                         extent_flags = btrfs_extent_flags(leaf, ei);
 788                 } else {
 789 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 790                         struct btrfs_extent_item_v0 *ei0;
 791                         BUG_ON(item_size != sizeof(*ei0));
 792                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 793                                              struct btrfs_extent_item_v0);
 794                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 795                         /* FIXME: this isn't correct for data */
 796                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 797 #else
 798                         BUG();
 799 #endif
 800                 }
 801                 BUG_ON(num_refs == 0);
 802         } else {
 803                 num_refs = 0;
 804                 extent_flags = 0;
 805                 ret = 0;
 806         }
 807
 808         if (!trans)
 809                 goto out;
 810
 811         delayed_refs = &trans->transaction->delayed_refs;
 812         spin_lock(&delayed_refs->lock);
 813         head = btrfs_find_delayed_ref_head(trans, bytenr);
 814         if (head) {
 815                 if (!mutex_trylock(&head->mutex)) {
 816                         atomic_inc(&head->node.refs);
 817                         spin_unlock(&delayed_refs->lock);
 818
 819                         btrfs_release_path(path);
 820
 821                         /*
 822                          * Mutex was contended, block until it's released and try
 823                          * again
 824                          */
 825                         mutex_lock(&head->mutex);
 826                         mutex_unlock(&head->mutex);
 827                         btrfs_put_delayed_ref(&head->node);
 828                         goto again;
 829                 }
 830                 if (head->extent_op && head->extent_op->update_flags)
 831                         extent_flags |= head->extent_op->flags_to_set;
 832                 else
 833                         BUG_ON(num_refs == 0);
 834
 835                 num_refs += head->node.ref_mod;
 836                 mutex_unlock(&head->mutex);
 837         }
 838         spin_unlock(&delayed_refs->lock);
 839 out:
 840         WARN_ON(num_refs == 0);
 841         if (refs)
 842                 *refs = num_refs;
 843         if (flags)
 844                 *flags = extent_flags;
 845 out_free:
 846         btrfs_free_path(path);
 847         return ret;
 848 }
 849
 850 /*
 851  * Back reference rules.  Back refs have three main goals:
 852  *
 853  * 1) differentiate between all holders of references to an extent so that
 854  *    when a reference is dropped we can make sure it was a valid reference
 855  *    before freeing the extent.
 856  *
 857  * 2) Provide enough information to quickly find the holders of an extent
 858  *    if we notice a given block is corrupted or bad.
 859  *
 860  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 861  *    maintenance.  This is actually the same as #2, but with a slightly
 862  *    different use case.
 863  *
 864  * There are two kinds of back refs. The implicit back refs is optimized
 865  * for pointers in non-shared tree blocks. For a given pointer in a block,
 866  * back refs of this kind provide information about the block's owner tree
 867  * and the pointer's key. These information allow us to find the block by
 868  * b-tree searching. The full back refs is for pointers in tree blocks not
 869  * referenced by their owner trees. The location of tree block is recorded
 870  * in the back refs. Actually the full back refs is generic, and can be
 871  * used in all cases the implicit back refs is used. The major shortcoming
 872  * of the full back refs is its overhead. Every time a tree block gets
 873  * COWed, we have to update back refs entry for all pointers in it.
 874  *
 875  * For a newly allocated tree block, we use implicit back refs for
 876  * pointers in it. This means most tree related operations only involve
 877  * implicit back refs. For a tree block created in old transaction, the
 878  * only way to drop a reference to it is COW it. So we can detect the
 879  * event that tree block loses its owner tree's reference and do the
 880  * back refs conversion.
 881  *
 882  * When a tree block is COW'd through a tree, there are four cases:
 883  *
 884  * The reference count of the block is one and the tree is the block's
 885  * owner tree. Nothing to do in this case.
 886  *
 887  * The reference count of the block is one and the tree is not the
 888  * block's owner tree. In this case, full back refs is used for pointers
 889  * in the block. Remove these full back refs, add implicit back refs for
 890  * every pointers in the new block.
 891  *
 892  * The reference count of the block is greater than one and the tree is
 893  * the block's owner tree. In this case, implicit back refs is used for
 894  * pointers in the block. Add full back refs for every pointers in the
 895  * block, increase lower level extents' reference counts. The original
 896  * implicit back refs are entailed to the new block.
 897  *
 898  * The reference count of the block is greater than one and the tree is
 899  * not the block's owner tree. Add implicit back refs for every pointer in
 900  * the new block, increase lower level extents' reference count.
 901  *
 902  * Back Reference Key composing:
 903  *
 904  * The key objectid corresponds to the first byte in the extent,
 905  * The key type is used to differentiate between types of back refs.
 906  * There are different meanings of the key offset for different types
 907  * of back refs.
 908  *
 909  * File extents can be referenced by:
 910  *
 911  * - multiple snapshots, subvolumes, or different generations in one subvol
 912  * - different files inside a single subvolume
 913  * - different offsets inside a file (bookend extents in file.c)
 914  *
 915  * The extent ref structure for the implicit back refs has fields for:
 916  *
 917  * - Objectid of the subvolume root
 918  * - objectid of the file holding the reference
 919  * - original offset in the file
 920  * - how many bookend extents
 921  *
 922  * The key offset for the implicit back refs is hash of the first
 923  * three fields.
 924  *
 925  * The extent ref structure for the full back refs has field for:
 926  *
 927  * - number of pointers in the tree leaf
 928  *
 929  * The key offset for the implicit back refs is the first byte of
 930  * the tree leaf
 931  *
 932  * When a file extent is allocated, The implicit back refs is used.
 933  * the fields are filled in:
 934  *
 935  *     (root_key.objectid, inode objectid, offset in file, 1)
 936  *
 937  * When a file extent is removed file truncation, we find the
 938  * corresponding implicit back refs and check the following fields:
 939  *
 940  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 941  *
 942  * Btree extents can be referenced by:
 943  *
 944  * - Different subvolumes
 945  *
 946  * Both the implicit back refs and the full back refs for tree blocks
 947  * only consist of key. The key offset for the implicit back refs is
 948  * objectid of block's owner tree. The key offset for the full back refs
 949  * is the first byte of parent block.
 950  *
 951  * When implicit back refs is used, information about the lowest key and
 952  * level of the tree block are required. These information are stored in
 953  * tree block info structure.
 954  */
 955
 956 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 957 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 958                                   struct btrfs_root *root,
 959                                   struct btrfs_path *path,
 960                                   u64 owner, u32 extra_size)
 961 {
 962         struct btrfs_extent_item *item;
 963         struct btrfs_extent_item_v0 *ei0;
 964         struct btrfs_extent_ref_v0 *ref0;
 965         struct btrfs_tree_block_info *bi;
 966         struct extent_buffer *leaf;
 967         struct btrfs_key key;
 968         struct btrfs_key found_key;
 969         u32 new_size = sizeof(*item);
 970         u64 refs;
 971         int ret;
 972
 973         leaf = path->nodes[0];
 974         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 975
 976         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 977         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 978                              struct btrfs_extent_item_v0);
 979         refs = btrfs_extent_refs_v0(leaf, ei0);
 980
 981         if (owner == (u64)-1) {
 982                 while (1) {
 983                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
 984                                 ret = btrfs_next_leaf(root, path);
 985                                 if (ret < 0)
 986                                         return ret;
 987                                 BUG_ON(ret > 0); /* Corruption */
 988                                 leaf = path->nodes[0];
 989                         }
 990                         btrfs_item_key_to_cpu(leaf, &found_key,
 991                                               path->slots[0]);
 992                         BUG_ON(key.objectid != found_key.objectid);
 993                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
 994                                 path->slots[0]++;
 995                                 continue;
 996                         }
 997                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
 998                                               struct btrfs_extent_ref_v0);
 999                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1000                         break;
1001                 }
1002         }
1003         btrfs_release_path(path);
1004
1005         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1006                 new_size += sizeof(*bi);
1007
1008         new_size -= sizeof(*ei0);
1009         ret = btrfs_search_slot(trans, root, &key, path,
1010                                 new_size + extra_size, 1);
1011         if (ret < 0)
1012                 return ret;
1013         BUG_ON(ret); /* Corruption */
1014
1015         btrfs_extend_item(trans, root, path, new_size);
1016
1017         leaf = path->nodes[0];
1018         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1019         btrfs_set_extent_refs(leaf, item, refs);
1020         /* FIXME: get real generation */
1021         btrfs_set_extent_generation(leaf, item, 0);
1022         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1023                 btrfs_set_extent_flags(leaf, item,
1024                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1025                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1026                 bi = (struct btrfs_tree_block_info *)(item + 1);
1027                 /* FIXME: get first key of the block */
1028                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1029                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1030         } else {
1031                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1032         }
1033         btrfs_mark_buffer_dirty(leaf);
1034         return 0;
1035 }
1036 #endif
1037
1038 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1039 {
1040         u32 high_crc = ~(u32)0;
1041         u32 low_crc = ~(u32)0;
1042         __le64 lenum;
1043
1044         lenum = cpu_to_le64(root_objectid);
1045         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1046         lenum = cpu_to_le64(owner);
1047         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1048         lenum = cpu_to_le64(offset);
1049         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1050
1051         return ((u64)high_crc << 31) ^ (u64)low_crc;
1052 }
1053
1054 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1055                                      struct btrfs_extent_data_ref *ref)
1056 {
1057         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1058                                     btrfs_extent_data_ref_objectid(leaf, ref),
1059                                     btrfs_extent_data_ref_offset(leaf, ref));
1060 }
1061
1062 static int match_extent_data_ref(struct extent_buffer *leaf,
1063                                  struct btrfs_extent_data_ref *ref,
1064                                  u64 root_objectid, u64 owner, u64 offset)
1065 {
1066         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1067             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1068             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1069                 return 0;
1070         return 1;
1071 }
1072
1073 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1074                                            struct btrfs_root *root,
1075                                            struct btrfs_path *path,
1076                                            u64 bytenr, u64 parent,
1077                                            u64 root_objectid,
1078                                            u64 owner, u64 offset)
1079 {
1080         struct btrfs_key key;
1081         struct btrfs_extent_data_ref *ref;
1082         struct extent_buffer *leaf;
1083         u32 nritems;
1084         int ret;
1085         int recow;
1086         int err = -ENOENT;
1087
1088         key.objectid = bytenr;
1089         if (parent) {
1090                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1091                 key.offset = parent;
1092         } else {
1093                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1094                 key.offset = hash_extent_data_ref(root_objectid,
1095                                                   owner, offset);
1096         }
1097 again:
1098         recow = 0;
1099         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1100         if (ret < 0) {
1101                 err = ret;
1102                 goto fail;
1103         }
1104
1105         if (parent) {
1106                 if (!ret)
1107                         return 0;
1108 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1109                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1110                 btrfs_release_path(path);
1111                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1112                 if (ret < 0) {
1113                         err = ret;
1114                         goto fail;
1115                 }
1116                 if (!ret)
1117                         return 0;
1118 #endif
1119                 goto fail;
1120         }
1121
1122         leaf = path->nodes[0];
1123         nritems = btrfs_header_nritems(leaf);
1124         while (1) {
1125                 if (path->slots[0] >= nritems) {
1126                         ret = btrfs_next_leaf(root, path);
1127                         if (ret < 0)
1128                                 err = ret;
1129                         if (ret)
1130                                 goto fail;
1131
1132                         leaf = path->nodes[0];
1133                         nritems = btrfs_header_nritems(leaf);
1134                         recow = 1;
1135                 }
1136
1137                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1138                 if (key.objectid != bytenr ||
1139                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1140                         goto fail;
1141
1142                 ref = btrfs_item_ptr(leaf, path->slots[0],
1143                                      struct btrfs_extent_data_ref);
1144
1145                 if (match_extent_data_ref(leaf, ref, root_objectid,
1146                                           owner, offset)) {
1147                         if (recow) {
1148                                 btrfs_release_path(path);
1149                                 goto again;
1150                         }
1151                         err = 0;
1152                         break;
1153                 }
1154                 path->slots[0]++;
1155         }
1156 fail:
1157         return err;
1158 }
1159
1160 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1161                                            struct btrfs_root *root,
1162                                            struct btrfs_path *path,
1163                                            u64 bytenr, u64 parent,
1164                                            u64 root_objectid, u64 owner,
1165                                            u64 offset, int refs_to_add)
1166 {
1167         struct btrfs_key key;
1168         struct extent_buffer *leaf;
1169         u32 size;
1170         u32 num_refs;
1171         int ret;
1172
1173         key.objectid = bytenr;
1174         if (parent) {
1175                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1176                 key.offset = parent;
1177                 size = sizeof(struct btrfs_shared_data_ref);
1178         } else {
1179                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1180                 key.offset = hash_extent_data_ref(root_objectid,
1181                                                   owner, offset);
1182                 size = sizeof(struct btrfs_extent_data_ref);
1183         }
1184
1185         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1186         if (ret && ret != -EEXIST)
1187                 goto fail;
1188
1189         leaf = path->nodes[0];
1190         if (parent) {
1191                 struct btrfs_shared_data_ref *ref;
1192                 ref = btrfs_item_ptr(leaf, path->slots[0],
1193                                      struct btrfs_shared_data_ref);
1194                 if (ret == 0) {
1195                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1196                 } else {
1197                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1198                         num_refs += refs_to_add;
1199                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1200                 }
1201         } else {
1202                 struct btrfs_extent_data_ref *ref;
1203                 while (ret == -EEXIST) {
1204                         ref = btrfs_item_ptr(leaf, path->slots[0],
1205                                              struct btrfs_extent_data_ref);
1206                         if (match_extent_data_ref(leaf, ref, root_objectid,
1207                                                   owner, offset))
1208                                 break;
1209                         btrfs_release_path(path);
1210                         key.offset++;
1211                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1212                                                       size);
1213                         if (ret && ret != -EEXIST)
1214                                 goto fail;
1215
1216                         leaf = path->nodes[0];
1217                 }
1218                 ref = btrfs_item_ptr(leaf, path->slots[0],
1219                                      struct btrfs_extent_data_ref);
1220                 if (ret == 0) {
1221                         btrfs_set_extent_data_ref_root(leaf, ref,
1222                                                        root_objectid);
1223                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1224                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1225                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1226                 } else {
1227                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1228                         num_refs += refs_to_add;
1229                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1230                 }
1231         }
1232         btrfs_mark_buffer_dirty(leaf);
1233         ret = 0;
1234 fail:
1235         btrfs_release_path(path);
1236         return ret;
1237 }
1238
1239 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1240                                            struct btrfs_root *root,
1241                                            struct btrfs_path *path,
1242                                            int refs_to_drop)
1243 {
1244         struct btrfs_key key;
1245         struct btrfs_extent_data_ref *ref1 = NULL;
1246         struct btrfs_shared_data_ref *ref2 = NULL;
1247         struct extent_buffer *leaf;
1248         u32 num_refs = 0;
1249         int ret = 0;
1250
1251         leaf = path->nodes[0];
1252         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1253
1254         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1255                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1256                                       struct btrfs_extent_data_ref);
1257                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1258         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1259                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1260                                       struct btrfs_shared_data_ref);
1261                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1262 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1263         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1264                 struct btrfs_extent_ref_v0 *ref0;
1265                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1266                                       struct btrfs_extent_ref_v0);
1267                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1268 #endif
1269         } else {
1270                 BUG();
1271         }
1272
1273         BUG_ON(num_refs < refs_to_drop);
1274         num_refs -= refs_to_drop;
1275
1276         if (num_refs == 0) {
1277                 ret = btrfs_del_item(trans, root, path);
1278         } else {
1279                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1280                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1281                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1282                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1283 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1284                 else {
1285                         struct btrfs_extent_ref_v0 *ref0;
1286                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1287                                         struct btrfs_extent_ref_v0);
1288                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1289                 }
1290 #endif
1291                 btrfs_mark_buffer_dirty(leaf);
1292         }
1293         return ret;
1294 }
1295
1296 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1297                                           struct btrfs_path *path,
1298                                           struct btrfs_extent_inline_ref *iref)
1299 {
1300         struct btrfs_key key;
1301         struct extent_buffer *leaf;
1302         struct btrfs_extent_data_ref *ref1;
1303         struct btrfs_shared_data_ref *ref2;
1304         u32 num_refs = 0;
1305
1306         leaf = path->nodes[0];
1307         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1308         if (iref) {
1309                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1310                     BTRFS_EXTENT_DATA_REF_KEY) {
1311                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1312                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1313                 } else {
1314                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1315                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1316                 }
1317         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1318                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1319                                       struct btrfs_extent_data_ref);
1320                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1321         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1322                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1323                                       struct btrfs_shared_data_ref);
1324                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1325 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1326         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1327                 struct btrfs_extent_ref_v0 *ref0;
1328                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1329                                       struct btrfs_extent_ref_v0);
1330                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1331 #endif
1332         } else {
1333                 WARN_ON(1);
1334         }
1335         return num_refs;
1336 }
1337
1338 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1339                                           struct btrfs_root *root,
1340                                           struct btrfs_path *path,
1341                                           u64 bytenr, u64 parent,
1342                                           u64 root_objectid)
1343 {
1344         struct btrfs_key key;
1345         int ret;
1346
1347         key.objectid = bytenr;
1348         if (parent) {
1349                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1350                 key.offset = parent;
1351         } else {
1352                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1353                 key.offset = root_objectid;
1354         }
1355
1356         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1357         if (ret > 0)
1358                 ret = -ENOENT;
1359 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1360         if (ret == -ENOENT && parent) {
1361                 btrfs_release_path(path);
1362                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1363                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1364                 if (ret > 0)
1365                         ret = -ENOENT;
1366         }
1367 #endif
1368         return ret;
1369 }
1370
1371 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1372                                           struct btrfs_root *root,
1373                                           struct btrfs_path *path,
1374                                           u64 bytenr, u64 parent,
1375                                           u64 root_objectid)
1376 {
1377         struct btrfs_key key;
1378         int ret;
1379
1380         key.objectid = bytenr;
1381         if (parent) {
1382                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1383                 key.offset = parent;
1384         } else {
1385                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1386                 key.offset = root_objectid;
1387         }
1388
1389         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1390         btrfs_release_path(path);
1391         return ret;
1392 }
1393
1394 static inline int extent_ref_type(u64 parent, u64 owner)
1395 {
1396         int type;
1397         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1398                 if (parent > 0)
1399                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1400                 else
1401                         type = BTRFS_TREE_BLOCK_REF_KEY;
1402         } else {
1403                 if (parent > 0)
1404                         type = BTRFS_SHARED_DATA_REF_KEY;
1405                 else
1406                         type = BTRFS_EXTENT_DATA_REF_KEY;
1407         }
1408         return type;
1409 }
1410
1411 static int find_next_key(struct btrfs_path *path, int level,
1412                          struct btrfs_key *key)
1413
1414 {
1415         for (; level < BTRFS_MAX_LEVEL; level++) {
1416                 if (!path->nodes[level])
1417                         break;
1418                 if (path->slots[level] + 1 >=
1419                     btrfs_header_nritems(path->nodes[level]))
1420                         continue;
1421                 if (level == 0)
1422                         btrfs_item_key_to_cpu(path->nodes[level], key,
1423                                               path->slots[level] + 1);
1424                 else
1425                         btrfs_node_key_to_cpu(path->nodes[level], key,
1426                                               path->slots[level] + 1);
1427                 return 0;
1428         }
1429         return 1;
1430 }
1431
1432 /*
1433  * look for inline back ref. if back ref is found, *ref_ret is set
1434  * to the address of inline back ref, and 0 is returned.
1435  *
1436  * if back ref isn't found, *ref_ret is set to the address where it
1437  * should be inserted, and -ENOENT is returned.
1438  *
1439  * if insert is true and there are too many inline back refs, the path
1440  * points to the extent item, and -EAGAIN is returned.
1441  *
1442  * NOTE: inline back refs are ordered in the same way that back ref
1443  *       items in the tree are ordered.
1444  */
1445 static noinline_for_stack
1446 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1447                                  struct btrfs_root *root,
1448                                  struct btrfs_path *path,
1449                                  struct btrfs_extent_inline_ref **ref_ret,
1450                                  u64 bytenr, u64 num_bytes,
1451                                  u64 parent, u64 root_objectid,
1452                                  u64 owner, u64 offset, int insert)
1453 {
1454         struct btrfs_key key;
1455         struct extent_buffer *leaf;
1456         struct btrfs_extent_item *ei;
1457         struct btrfs_extent_inline_ref *iref;
1458         u64 flags;
1459         u64 item_size;
1460         unsigned long ptr;
1461         unsigned long end;
1462         int extra_size;
1463         int type;
1464         int want;
1465         int ret;
1466         int err = 0;
1467
1468         key.objectid = bytenr;
1469         key.type = BTRFS_EXTENT_ITEM_KEY;
1470         key.offset = num_bytes;
1471
1472         want = extent_ref_type(parent, owner);
1473         if (insert) {
1474                 extra_size = btrfs_extent_inline_ref_size(want);
1475                 path->keep_locks = 1;
1476         } else
1477                 extra_size = -1;
1478         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1479         if (ret < 0) {
1480                 err = ret;
1481                 goto out;
1482         }
1483         if (ret && !insert) {
1484                 err = -ENOENT;
1485                 goto out;
1486         }
1487         BUG_ON(ret); /* Corruption */
1488
1489         leaf = path->nodes[0];
1490         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1491 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1492         if (item_size < sizeof(*ei)) {
1493                 if (!insert) {
1494                         err = -ENOENT;
1495                         goto out;
1496                 }
1497                 ret = convert_extent_item_v0(trans, root, path, owner,
1498                                              extra_size);
1499                 if (ret < 0) {
1500                         err = ret;
1501                         goto out;
1502                 }
1503                 leaf = path->nodes[0];
1504                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1505         }
1506 #endif
1507         BUG_ON(item_size < sizeof(*ei));
1508
1509         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1510         flags = btrfs_extent_flags(leaf, ei);
1511
1512         ptr = (unsigned long)(ei + 1);
1513         end = (unsigned long)ei + item_size;
1514
1515         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK) {
1516                 ptr += sizeof(struct btrfs_tree_block_info);
1517                 BUG_ON(ptr > end);
1518         } else {
1519                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_DATA));
1520         }
1521
1522         err = -ENOENT;
1523         while (1) {
1524                 if (ptr >= end) {
1525                         WARN_ON(ptr > end);
1526                         break;
1527                 }
1528                 iref = (struct btrfs_extent_inline_ref *)ptr;
1529                 type = btrfs_extent_inline_ref_type(leaf, iref);
1530                 if (want < type)
1531                         break;
1532                 if (want > type) {
1533                         ptr += btrfs_extent_inline_ref_size(type);
1534                         continue;
1535                 }
1536
1537                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1538                         struct btrfs_extent_data_ref *dref;
1539                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1540                         if (match_extent_data_ref(leaf, dref, root_objectid,
1541                                                   owner, offset)) {
1542                                 err = 0;
1543                                 break;
1544                         }
1545                         if (hash_extent_data_ref_item(leaf, dref) <
1546                             hash_extent_data_ref(root_objectid, owner, offset))
1547                                 break;
1548                 } else {
1549                         u64 ref_offset;
1550                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1551                         if (parent > 0) {
1552                                 if (parent == ref_offset) {
1553                                         err = 0;
1554                                         break;
1555                                 }
1556                                 if (ref_offset < parent)
1557                                         break;
1558                         } else {
1559                                 if (root_objectid == ref_offset) {
1560                                         err = 0;
1561                                         break;
1562                                 }
1563                                 if (ref_offset < root_objectid)
1564                                         break;
1565                         }
1566                 }
1567                 ptr += btrfs_extent_inline_ref_size(type);
1568         }
1569         if (err == -ENOENT && insert) {
1570                 if (item_size + extra_size >=
1571                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1572                         err = -EAGAIN;
1573                         goto out;
1574                 }
1575                 /*
1576                  * To add new inline back ref, we have to make sure
1577                  * there is no corresponding back ref item.
1578                  * For simplicity, we just do not add new inline back
1579                  * ref if there is any kind of item for this block
1580                  */
1581                 if (find_next_key(path, 0, &key) == 0 &&
1582                     key.objectid == bytenr &&
1583                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1584                         err = -EAGAIN;
1585                         goto out;
1586                 }
1587         }
1588         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1589 out:
1590         if (insert) {
1591                 path->keep_locks = 0;
1592                 btrfs_unlock_up_safe(path, 1);
1593         }
1594         return err;
1595 }
1596
1597 /*
1598  * helper to add new inline back ref
1599  */
1600 static noinline_for_stack
1601 void setup_inline_extent_backref(struct btrfs_trans_handle *trans,
1602                                  struct btrfs_root *root,
1603                                  struct btrfs_path *path,
1604                                  struct btrfs_extent_inline_ref *iref,
1605                                  u64 parent, u64 root_objectid,
1606                                  u64 owner, u64 offset, int refs_to_add,
1607                                  struct btrfs_delayed_extent_op *extent_op)
1608 {
1609         struct extent_buffer *leaf;
1610         struct btrfs_extent_item *ei;
1611         unsigned long ptr;
1612         unsigned long end;
1613         unsigned long item_offset;
1614         u64 refs;
1615         int size;
1616         int type;
1617
1618         leaf = path->nodes[0];
1619         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1620         item_offset = (unsigned long)iref - (unsigned long)ei;
1621
1622         type = extent_ref_type(parent, owner);
1623         size = btrfs_extent_inline_ref_size(type);
1624
1625         btrfs_extend_item(trans, root, path, size);
1626
1627         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1628         refs = btrfs_extent_refs(leaf, ei);
1629         refs += refs_to_add;
1630         btrfs_set_extent_refs(leaf, ei, refs);
1631         if (extent_op)
1632                 __run_delayed_extent_op(extent_op, leaf, ei);
1633
1634         ptr = (unsigned long)ei + item_offset;
1635         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1636         if (ptr < end - size)
1637                 memmove_extent_buffer(leaf, ptr + size, ptr,
1638                                       end - size - ptr);
1639
1640         iref = (struct btrfs_extent_inline_ref *)ptr;
1641         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1642         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1643                 struct btrfs_extent_data_ref *dref;
1644                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1645                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1646                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1647                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1648                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1649         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1650                 struct btrfs_shared_data_ref *sref;
1651                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1652                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1653                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1654         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1655                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1656         } else {
1657                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1658         }
1659         btrfs_mark_buffer_dirty(leaf);
1660 }
1661
1662 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1663                                  struct btrfs_root *root,
1664                                  struct btrfs_path *path,
1665                                  struct btrfs_extent_inline_ref **ref_ret,
1666                                  u64 bytenr, u64 num_bytes, u64 parent,
1667                                  u64 root_objectid, u64 owner, u64 offset)
1668 {
1669         int ret;
1670
1671         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1672                                            bytenr, num_bytes, parent,
1673                                            root_objectid, owner, offset, 0);
1674         if (ret != -ENOENT)
1675                 return ret;
1676
1677         btrfs_release_path(path);
1678         *ref_ret = NULL;
1679
1680         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1681                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1682                                             root_objectid);
1683         } else {
1684                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1685                                              root_objectid, owner, offset);
1686         }
1687         return ret;
1688 }
1689
1690 /*
1691  * helper to update/remove inline back ref
1692  */
1693 static noinline_for_stack
1694 void update_inline_extent_backref(struct btrfs_trans_handle *trans,
1695                                   struct btrfs_root *root,
1696                                   struct btrfs_path *path,
1697                                   struct btrfs_extent_inline_ref *iref,
1698                                   int refs_to_mod,
1699                                   struct btrfs_delayed_extent_op *extent_op)
1700 {
1701         struct extent_buffer *leaf;
1702         struct btrfs_extent_item *ei;
1703         struct btrfs_extent_data_ref *dref = NULL;
1704         struct btrfs_shared_data_ref *sref = NULL;
1705         unsigned long ptr;
1706         unsigned long end;
1707         u32 item_size;
1708         int size;
1709         int type;
1710         u64 refs;
1711
1712         leaf = path->nodes[0];
1713         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1714         refs = btrfs_extent_refs(leaf, ei);
1715         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1716         refs += refs_to_mod;
1717         btrfs_set_extent_refs(leaf, ei, refs);
1718         if (extent_op)
1719                 __run_delayed_extent_op(extent_op, leaf, ei);
1720
1721         type = btrfs_extent_inline_ref_type(leaf, iref);
1722
1723         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1724                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1725                 refs = btrfs_extent_data_ref_count(leaf, dref);
1726         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1727                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1728                 refs = btrfs_shared_data_ref_count(leaf, sref);
1729         } else {
1730                 refs = 1;
1731                 BUG_ON(refs_to_mod != -1);
1732         }
1733
1734         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1735         refs += refs_to_mod;
1736
1737         if (refs > 0) {
1738                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1739                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1740                 else
1741                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1742         } else {
1743                 size =  btrfs_extent_inline_ref_size(type);
1744                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1745                 ptr = (unsigned long)iref;
1746                 end = (unsigned long)ei + item_size;
1747                 if (ptr + size < end)
1748                         memmove_extent_buffer(leaf, ptr, ptr + size,
1749                                               end - ptr - size);
1750                 item_size -= size;
1751                 btrfs_truncate_item(trans, root, path, item_size, 1);
1752         }
1753         btrfs_mark_buffer_dirty(leaf);
1754 }
1755
1756 static noinline_for_stack
1757 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1758                                  struct btrfs_root *root,
1759                                  struct btrfs_path *path,
1760                                  u64 bytenr, u64 num_bytes, u64 parent,
1761                                  u64 root_objectid, u64 owner,
1762                                  u64 offset, int refs_to_add,
1763                                  struct btrfs_delayed_extent_op *extent_op)
1764 {
1765         struct btrfs_extent_inline_ref *iref;
1766         int ret;
1767
1768         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1769                                            bytenr, num_bytes, parent,
1770                                            root_objectid, owner, offset, 1);
1771         if (ret == 0) {
1772                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1773                 update_inline_extent_backref(trans, root, path, iref,
1774                                              refs_to_add, extent_op);
1775         } else if (ret == -ENOENT) {
1776                 setup_inline_extent_backref(trans, root, path, iref, parent,
1777                                             root_objectid, owner, offset,
1778                                             refs_to_add, extent_op);
1779                 ret = 0;
1780         }
1781         return ret;
1782 }
1783
1784 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1785                                  struct btrfs_root *root,
1786                                  struct btrfs_path *path,
1787                                  u64 bytenr, u64 parent, u64 root_objectid,
1788                                  u64 owner, u64 offset, int refs_to_add)
1789 {
1790         int ret;
1791         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1792                 BUG_ON(refs_to_add != 1);
1793                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1794                                             parent, root_objectid);
1795         } else {
1796                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1797                                              parent, root_objectid,
1798                                              owner, offset, refs_to_add);
1799         }
1800         return ret;
1801 }
1802
1803 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1804                                  struct btrfs_root *root,
1805                                  struct btrfs_path *path,
1806                                  struct btrfs_extent_inline_ref *iref,
1807                                  int refs_to_drop, int is_data)
1808 {
1809         int ret = 0;
1810
1811         BUG_ON(!is_data && refs_to_drop != 1);
1812         if (iref) {
1813                 update_inline_extent_backref(trans, root, path, iref,
1814                                              -refs_to_drop, NULL);
1815         } else if (is_data) {
1816                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1817         } else {
1818                 ret = btrfs_del_item(trans, root, path);
1819         }
1820         return ret;
1821 }
1822
1823 static int btrfs_issue_discard(struct block_device *bdev,
1824                                 u64 start, u64 len)
1825 {
1826         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1827 }
1828
1829 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1830                                 u64 num_bytes, u64 *actual_bytes)
1831 {
1832         int ret;
1833         u64 discarded_bytes = 0;
1834         struct btrfs_bio *bbio = NULL;
1835
1836
1837         /* Tell the block device(s) that the sectors can be discarded */
1838         ret = btrfs_map_block(&root->fs_info->mapping_tree, REQ_DISCARD,
1839                               bytenr, &num_bytes, &bbio, 0);
1840         /* Error condition is -ENOMEM */
1841         if (!ret) {
1842                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1843                 int i;
1844
1845
1846                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1847                         if (!stripe->dev->can_discard)
1848                                 continue;
1849
1850                         ret = btrfs_issue_discard(stripe->dev->bdev,
1851                                                   stripe->physical,
1852                                                   stripe->length);
1853                         if (!ret)
1854                                 discarded_bytes += stripe->length;
1855                         else if (ret != -EOPNOTSUPP)
1856                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1857
1858                         /*
1859                          * Just in case we get back EOPNOTSUPP for some reason,
1860                          * just ignore the return value so we don't screw up
1861                          * people calling discard_extent.
1862                          */
1863                         ret = 0;
1864                 }
1865                 kfree(bbio);
1866         }
1867
1868         if (actual_bytes)
1869                 *actual_bytes = discarded_bytes;
1870
1871
1872         return ret;
1873 }
1874
1875 /* Can return -ENOMEM */
1876 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1877                          struct btrfs_root *root,
1878                          u64 bytenr, u64 num_bytes, u64 parent,
1879                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1880 {
1881         int ret;
1882         struct btrfs_fs_info *fs_info = root->fs_info;
1883
1884         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1885                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1886
1887         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1888                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1889                                         num_bytes,
1890                                         parent, root_objectid, (int)owner,
1891                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1892         } else {
1893                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1894                                         num_bytes,
1895                                         parent, root_objectid, owner, offset,
1896                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1897         }
1898         return ret;
1899 }
1900
1901 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1902                                   struct btrfs_root *root,
1903                                   u64 bytenr, u64 num_bytes,
1904                                   u64 parent, u64 root_objectid,
1905                                   u64 owner, u64 offset, int refs_to_add,
1906                                   struct btrfs_delayed_extent_op *extent_op)
1907 {
1908         struct btrfs_path *path;
1909         struct extent_buffer *leaf;
1910         struct btrfs_extent_item *item;
1911         u64 refs;
1912         int ret;
1913         int err = 0;
1914
1915         path = btrfs_alloc_path();
1916         if (!path)
1917                 return -ENOMEM;
1918
1919         path->reada = 1;
1920         path->leave_spinning = 1;
1921         /* this will setup the path even if it fails to insert the back ref */
1922         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
1923                                            path, bytenr, num_bytes, parent,
1924                                            root_objectid, owner, offset,
1925                                            refs_to_add, extent_op);
1926         if (ret == 0)
1927                 goto out;
1928
1929         if (ret != -EAGAIN) {
1930                 err = ret;
1931                 goto out;
1932         }
1933
1934         leaf = path->nodes[0];
1935         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1936         refs = btrfs_extent_refs(leaf, item);
1937         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
1938         if (extent_op)
1939                 __run_delayed_extent_op(extent_op, leaf, item);
1940
1941         btrfs_mark_buffer_dirty(leaf);
1942         btrfs_release_path(path);
1943
1944         path->reada = 1;
1945         path->leave_spinning = 1;
1946
1947         /* now insert the actual backref */
1948         ret = insert_extent_backref(trans, root->fs_info->extent_root,
1949                                     path, bytenr, parent, root_objectid,
1950                                     owner, offset, refs_to_add);
1951         if (ret)
1952                 btrfs_abort_transaction(trans, root, ret);
1953 out:
1954         btrfs_free_path(path);
1955         return err;
1956 }
1957
1958 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
1959                                 struct btrfs_root *root,
1960                                 struct btrfs_delayed_ref_node *node,
1961                                 struct btrfs_delayed_extent_op *extent_op,
1962                                 int insert_reserved)
1963 {
1964         int ret = 0;
1965         struct btrfs_delayed_data_ref *ref;
1966         struct btrfs_key ins;
1967         u64 parent = 0;
1968         u64 ref_root = 0;
1969         u64 flags = 0;
1970
1971         ins.objectid = node->bytenr;
1972         ins.offset = node->num_bytes;
1973         ins.type = BTRFS_EXTENT_ITEM_KEY;
1974
1975         ref = btrfs_delayed_node_to_data_ref(node);
1976         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
1977                 parent = ref->parent;
1978         else
1979                 ref_root = ref->root;
1980
1981         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
1982                 if (extent_op) {
1983                         BUG_ON(extent_op->update_key);
1984                         flags |= extent_op->flags_to_set;
1985                 }
1986                 ret = alloc_reserved_file_extent(trans, root,
1987                                                  parent, ref_root, flags,
1988                                                  ref->objectid, ref->offset,
1989                                                  &ins, node->ref_mod);
1990         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
1991                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
1992                                              node->num_bytes, parent,
1993                                              ref_root, ref->objectid,
1994                                              ref->offset, node->ref_mod,
1995                                              extent_op);
1996         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
1997                 ret = __btrfs_free_extent(trans, root, node->bytenr,
1998                                           node->num_bytes, parent,
1999                                           ref_root, ref->objectid,
2000                                           ref->offset, node->ref_mod,
2001                                           extent_op);
2002         } else {
2003                 BUG();
2004         }
2005         return ret;
2006 }
2007
2008 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2009                                     struct extent_buffer *leaf,
2010                                     struct btrfs_extent_item *ei)
2011 {
2012         u64 flags = btrfs_extent_flags(leaf, ei);
2013         if (extent_op->update_flags) {
2014                 flags |= extent_op->flags_to_set;
2015                 btrfs_set_extent_flags(leaf, ei, flags);
2016         }
2017
2018         if (extent_op->update_key) {
2019                 struct btrfs_tree_block_info *bi;
2020                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2021                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2022                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2023         }
2024 }
2025
2026 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2027                                  struct btrfs_root *root,
2028                                  struct btrfs_delayed_ref_node *node,
2029                                  struct btrfs_delayed_extent_op *extent_op)
2030 {
2031         struct btrfs_key key;
2032         struct btrfs_path *path;
2033         struct btrfs_extent_item *ei;
2034         struct extent_buffer *leaf;
2035         u32 item_size;
2036         int ret;
2037         int err = 0;
2038
2039         if (trans->aborted)
2040                 return 0;
2041
2042         path = btrfs_alloc_path();
2043         if (!path)
2044                 return -ENOMEM;
2045
2046         key.objectid = node->bytenr;
2047         key.type = BTRFS_EXTENT_ITEM_KEY;
2048         key.offset = node->num_bytes;
2049
2050         path->reada = 1;
2051         path->leave_spinning = 1;
2052         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2053                                 path, 0, 1);
2054         if (ret < 0) {
2055                 err = ret;
2056                 goto out;
2057         }
2058         if (ret > 0) {
2059                 err = -EIO;
2060                 goto out;
2061         }
2062
2063         leaf = path->nodes[0];
2064         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2065 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2066         if (item_size < sizeof(*ei)) {
2067                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2068                                              path, (u64)-1, 0);
2069                 if (ret < 0) {
2070                         err = ret;
2071                         goto out;
2072                 }
2073                 leaf = path->nodes[0];
2074                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2075         }
2076 #endif
2077         BUG_ON(item_size < sizeof(*ei));
2078         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2079         __run_delayed_extent_op(extent_op, leaf, ei);
2080
2081         btrfs_mark_buffer_dirty(leaf);
2082 out:
2083         btrfs_free_path(path);
2084         return err;
2085 }
2086
2087 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2088                                 struct btrfs_root *root,
2089                                 struct btrfs_delayed_ref_node *node,
2090                                 struct btrfs_delayed_extent_op *extent_op,
2091                                 int insert_reserved)
2092 {
2093         int ret = 0;
2094         struct btrfs_delayed_tree_ref *ref;
2095         struct btrfs_key ins;
2096         u64 parent = 0;
2097         u64 ref_root = 0;
2098
2099         ins.objectid = node->bytenr;
2100         ins.offset = node->num_bytes;
2101         ins.type = BTRFS_EXTENT_ITEM_KEY;
2102
2103         ref = btrfs_delayed_node_to_tree_ref(node);
2104         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2105                 parent = ref->parent;
2106         else
2107                 ref_root = ref->root;
2108
2109         BUG_ON(node->ref_mod != 1);
2110         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2111                 BUG_ON(!extent_op || !extent_op->update_flags ||
2112                        !extent_op->update_key);
2113                 ret = alloc_reserved_tree_block(trans, root,
2114                                                 parent, ref_root,
2115                                                 extent_op->flags_to_set,
2116                                                 &extent_op->key,
2117                                                 ref->level, &ins);
2118         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2119                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2120                                              node->num_bytes, parent, ref_root,
2121                                              ref->level, 0, 1, extent_op);
2122         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2123                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2124                                           node->num_bytes, parent, ref_root,
2125                                           ref->level, 0, 1, extent_op);
2126         } else {
2127                 BUG();
2128         }
2129         return ret;
2130 }
2131
2132 /* helper function to actually process a single delayed ref entry */
2133 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2134                                struct btrfs_root *root,
2135                                struct btrfs_delayed_ref_node *node,
2136                                struct btrfs_delayed_extent_op *extent_op,
2137                                int insert_reserved)
2138 {
2139         int ret = 0;
2140
2141         if (trans->aborted)
2142                 return 0;
2143
2144         if (btrfs_delayed_ref_is_head(node)) {
2145                 struct btrfs_delayed_ref_head *head;
2146                 /*
2147                  * we've hit the end of the chain and we were supposed
2148                  * to insert this extent into the tree.  But, it got
2149                  * deleted before we ever needed to insert it, so all
2150                  * we have to do is clean up the accounting
2151                  */
2152                 BUG_ON(extent_op);
2153                 head = btrfs_delayed_node_to_head(node);
2154                 if (insert_reserved) {
2155                         btrfs_pin_extent(root, node->bytenr,
2156                                          node->num_bytes, 1);
2157                         if (head->is_data) {
2158                                 ret = btrfs_del_csums(trans, root,
2159                                                       node->bytenr,
2160                                                       node->num_bytes);
2161                         }
2162                 }
2163                 mutex_unlock(&head->mutex);
2164                 return ret;
2165         }
2166
2167         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2168             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2169                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2170                                            insert_reserved);
2171         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2172                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2173                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2174                                            insert_reserved);
2175         else
2176                 BUG();
2177         return ret;
2178 }
2179
2180 static noinline struct btrfs_delayed_ref_node *
2181 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2182 {
2183         struct rb_node *node;
2184         struct btrfs_delayed_ref_node *ref;
2185         int action = BTRFS_ADD_DELAYED_REF;
2186 again:
2187         /*
2188          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2189          * this prevents ref count from going down to zero when
2190          * there still are pending delayed ref.
2191          */
2192         node = rb_prev(&head->node.rb_node);
2193         while (1) {
2194                 if (!node)
2195                         break;
2196                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2197                                 rb_node);
2198                 if (ref->bytenr != head->node.bytenr)
2199                         break;
2200                 if (ref->action == action)
2201                         return ref;
2202                 node = rb_prev(node);
2203         }
2204         if (action == BTRFS_ADD_DELAYED_REF) {
2205                 action = BTRFS_DROP_DELAYED_REF;
2206                 goto again;
2207         }
2208         return NULL;
2209 }
2210
2211 /*
2212  * Returns 0 on success or if called with an already aborted transaction.
2213  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2214  */
2215 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2216                                        struct btrfs_root *root,
2217                                        struct list_head *cluster)
2218 {
2219         struct btrfs_delayed_ref_root *delayed_refs;
2220         struct btrfs_delayed_ref_node *ref;
2221         struct btrfs_delayed_ref_head *locked_ref = NULL;
2222         struct btrfs_delayed_extent_op *extent_op;
2223         struct btrfs_fs_info *fs_info = root->fs_info;
2224         int ret;
2225         int count = 0;
2226         int must_insert_reserved = 0;
2227
2228         delayed_refs = &trans->transaction->delayed_refs;
2229         while (1) {
2230                 if (!locked_ref) {
2231                         /* pick a new head ref from the cluster list */
2232                         if (list_empty(cluster))
2233                                 break;
2234
2235                         locked_ref = list_entry(cluster->next,
2236                                      struct btrfs_delayed_ref_head, cluster);
2237
2238                         /* grab the lock that says we are going to process
2239                          * all the refs for this head */
2240                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2241
2242                         /*
2243                          * we may have dropped the spin lock to get the head
2244                          * mutex lock, and that might have given someone else
2245                          * time to free the head.  If that's true, it has been
2246                          * removed from our list and we can move on.
2247                          */
2248                         if (ret == -EAGAIN) {
2249                                 locked_ref = NULL;
2250                                 count++;
2251                                 continue;
2252                         }
2253                 }
2254
2255                 /*
2256                  * We need to try and merge add/drops of the same ref since we
2257                  * can run into issues with relocate dropping the implicit ref
2258                  * and then it being added back again before the drop can
2259                  * finish.  If we merged anything we need to re-loop so we can
2260                  * get a good ref.
2261                  */
2262                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2263                                          locked_ref);
2264
2265                 /*
2266                  * locked_ref is the head node, so we have to go one
2267                  * node back for any delayed ref updates
2268                  */
2269                 ref = select_delayed_ref(locked_ref);
2270
2271                 if (ref && ref->seq &&
2272                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2273                         /*
2274                          * there are still refs with lower seq numbers in the
2275                          * process of being added. Don't run this ref yet.
2276                          */
2277                         list_del_init(&locked_ref->cluster);
2278                         mutex_unlock(&locked_ref->mutex);
2279                         locked_ref = NULL;
2280                         delayed_refs->num_heads_ready++;
2281                         spin_unlock(&delayed_refs->lock);
2282                         cond_resched();
2283                         spin_lock(&delayed_refs->lock);
2284                         continue;
2285                 }
2286
2287                 /*
2288                  * record the must insert reserved flag before we
2289                  * drop the spin lock.
2290                  */
2291                 must_insert_reserved = locked_ref->must_insert_reserved;
2292                 locked_ref->must_insert_reserved = 0;
2293
2294                 extent_op = locked_ref->extent_op;
2295                 locked_ref->extent_op = NULL;
2296
2297                 if (!ref) {
2298                         /* All delayed refs have been processed, Go ahead
2299                          * and send the head node to run_one_delayed_ref,
2300                          * so that any accounting fixes can happen
2301                          */
2302                         ref = &locked_ref->node;
2303
2304                         if (extent_op && must_insert_reserved) {
2305                                 kfree(extent_op);
2306                                 extent_op = NULL;
2307                         }
2308
2309                         if (extent_op) {
2310                                 spin_unlock(&delayed_refs->lock);
2311
2312                                 ret = run_delayed_extent_op(trans, root,
2313                                                             ref, extent_op);
2314                                 kfree(extent_op);
2315
2316                                 if (ret) {
2317                                         printk(KERN_DEBUG "btrfs: run_delayed_extent_op returned %d\n", ret);
2318                                         spin_lock(&delayed_refs->lock);
2319                                         return ret;
2320                                 }
2321
2322                                 goto next;
2323                         }
2324
2325                         list_del_init(&locked_ref->cluster);
2326                         locked_ref = NULL;
2327                 }
2328
2329                 ref->in_tree = 0;
2330                 rb_erase(&ref->rb_node, &delayed_refs->root);
2331                 delayed_refs->num_entries--;
2332                 if (locked_ref) {
2333                         /*
2334                          * when we play the delayed ref, also correct the
2335                          * ref_mod on head
2336                          */
2337                         switch (ref->action) {
2338                         case BTRFS_ADD_DELAYED_REF:
2339                         case BTRFS_ADD_DELAYED_EXTENT:
2340                                 locked_ref->node.ref_mod -= ref->ref_mod;
2341                                 break;
2342                         case BTRFS_DROP_DELAYED_REF:
2343                                 locked_ref->node.ref_mod += ref->ref_mod;
2344                                 break;
2345                         default:
2346                                 WARN_ON(1);
2347                         }
2348                 }
2349                 spin_unlock(&delayed_refs->lock);
2350
2351                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2352                                           must_insert_reserved);
2353
2354                 btrfs_put_delayed_ref(ref);
2355                 kfree(extent_op);
2356                 count++;
2357
2358                 if (ret) {
2359                         printk(KERN_DEBUG "btrfs: run_one_delayed_ref returned %d\n", ret);
2360                         spin_lock(&delayed_refs->lock);
2361                         return ret;
2362                 }
2363
2364 next:
2365                 cond_resched();
2366                 spin_lock(&delayed_refs->lock);
2367         }
2368         return count;
2369 }
2370
2371 #ifdef SCRAMBLE_DELAYED_REFS
2372 /*
2373  * Normally delayed refs get processed in ascending bytenr order. This
2374  * correlates in most cases to the order added. To expose dependencies on this
2375  * order, we start to process the tree in the middle instead of the beginning
2376  */
2377 static u64 find_middle(struct rb_root *root)
2378 {
2379         struct rb_node *n = root->rb_node;
2380         struct btrfs_delayed_ref_node *entry;
2381         int alt = 1;
2382         u64 middle;
2383         u64 first = 0, last = 0;
2384
2385         n = rb_first(root);
2386         if (n) {
2387                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2388                 first = entry->bytenr;
2389         }
2390         n = rb_last(root);
2391         if (n) {
2392                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2393                 last = entry->bytenr;
2394         }
2395         n = root->rb_node;
2396
2397         while (n) {
2398                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2399                 WARN_ON(!entry->in_tree);
2400
2401                 middle = entry->bytenr;
2402
2403                 if (alt)
2404                         n = n->rb_left;
2405                 else
2406                         n = n->rb_right;
2407
2408                 alt = 1 - alt;
2409         }
2410         return middle;
2411 }
2412 #endif
2413
2414 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2415                                          struct btrfs_fs_info *fs_info)
2416 {
2417         struct qgroup_update *qgroup_update;
2418         int ret = 0;
2419
2420         if (list_empty(&trans->qgroup_ref_list) !=
2421             !trans->delayed_ref_elem.seq) {
2422                 /* list without seq or seq without list */
2423                 printk(KERN_ERR "btrfs: qgroup accounting update error, list is%s empty, seq is %llu\n",
2424                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2425                         trans->delayed_ref_elem.seq);
2426                 BUG();
2427         }
2428
2429         if (!trans->delayed_ref_elem.seq)
2430                 return 0;
2431
2432         while (!list_empty(&trans->qgroup_ref_list)) {
2433                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2434                                                  struct qgroup_update, list);
2435                 list_del(&qgroup_update->list);
2436                 if (!ret)
2437                         ret = btrfs_qgroup_account_ref(
2438                                         trans, fs_info, qgroup_update->node,
2439                                         qgroup_update->extent_op);
2440                 kfree(qgroup_update);
2441         }
2442
2443         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2444
2445         return ret;
2446 }
2447
2448 /*
2449  * this starts processing the delayed reference count updates and
2450  * extent insertions we have queued up so far.  count can be
2451  * 0, which means to process everything in the tree at the start
2452  * of the run (but not newly added entries), or it can be some target
2453  * number you'd like to process.
2454  *
2455  * Returns 0 on success or if called with an aborted transaction
2456  * Returns <0 on error and aborts the transaction
2457  */
2458 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2459                            struct btrfs_root *root, unsigned long count)
2460 {
2461         struct rb_node *node;
2462         struct btrfs_delayed_ref_root *delayed_refs;
2463         struct btrfs_delayed_ref_node *ref;
2464         struct list_head cluster;
2465         int ret;
2466         u64 delayed_start;
2467         int run_all = count == (unsigned long)-1;
2468         int run_most = 0;
2469         int loops;
2470
2471         /* We'll clean this up in btrfs_cleanup_transaction */
2472         if (trans->aborted)
2473                 return 0;
2474
2475         if (root == root->fs_info->extent_root)
2476                 root = root->fs_info->tree_root;
2477
2478         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2479
2480         delayed_refs = &trans->transaction->delayed_refs;
2481         INIT_LIST_HEAD(&cluster);
2482 again:
2483         loops = 0;
2484         spin_lock(&delayed_refs->lock);
2485
2486 #ifdef SCRAMBLE_DELAYED_REFS
2487         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2488 #endif
2489
2490         if (count == 0) {
2491                 count = delayed_refs->num_entries * 2;
2492                 run_most = 1;
2493         }
2494         while (1) {
2495                 if (!(run_all || run_most) &&
2496                     delayed_refs->num_heads_ready < 64)
2497                         break;
2498
2499                 /*
2500                  * go find something we can process in the rbtree.  We start at
2501                  * the beginning of the tree, and then build a cluster
2502                  * of refs to process starting at the first one we are able to
2503                  * lock
2504                  */
2505                 delayed_start = delayed_refs->run_delayed_start;
2506                 ret = btrfs_find_ref_cluster(trans, &cluster,
2507                                              delayed_refs->run_delayed_start);
2508                 if (ret)
2509                         break;
2510
2511                 ret = run_clustered_refs(trans, root, &cluster);
2512                 if (ret < 0) {
2513                         spin_unlock(&delayed_refs->lock);
2514                         btrfs_abort_transaction(trans, root, ret);
2515                         return ret;
2516                 }
2517
2518                 count -= min_t(unsigned long, ret, count);
2519
2520                 if (count == 0)
2521                         break;
2522
2523                 if (delayed_start >= delayed_refs->run_delayed_start) {
2524                         if (loops == 0) {
2525                                 /*
2526                                  * btrfs_find_ref_cluster looped. let's do one
2527                                  * more cycle. if we don't run any delayed ref
2528                                  * during that cycle (because we can't because
2529                                  * all of them are blocked), bail out.
2530                                  */
2531                                 loops = 1;
2532                         } else {
2533                                 /*
2534                                  * no runnable refs left, stop trying
2535                                  */
2536                                 BUG_ON(run_all);
2537                                 break;
2538                         }
2539                 }
2540                 if (ret) {
2541                         /* refs were run, let's reset staleness detection */
2542                         loops = 0;
2543                 }
2544         }
2545
2546         if (run_all) {
2547                 if (!list_empty(&trans->new_bgs)) {
2548                         spin_unlock(&delayed_refs->lock);
2549                         btrfs_create_pending_block_groups(trans, root);
2550                         spin_lock(&delayed_refs->lock);
2551                 }
2552
2553                 node = rb_first(&delayed_refs->root);
2554                 if (!node)
2555                         goto out;
2556                 count = (unsigned long)-1;
2557
2558                 while (node) {
2559                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2560                                        rb_node);
2561                         if (btrfs_delayed_ref_is_head(ref)) {
2562                                 struct btrfs_delayed_ref_head *head;
2563
2564                                 head = btrfs_delayed_node_to_head(ref);
2565                                 atomic_inc(&ref->refs);
2566
2567                                 spin_unlock(&delayed_refs->lock);
2568                                 /*
2569                                  * Mutex was contended, block until it's
2570                                  * released and try again
2571                                  */
2572                                 mutex_lock(&head->mutex);
2573                                 mutex_unlock(&head->mutex);
2574
2575                                 btrfs_put_delayed_ref(ref);
2576                                 cond_resched();
2577                                 goto again;
2578                         }
2579                         node = rb_next(node);
2580                 }
2581                 spin_unlock(&delayed_refs->lock);
2582                 schedule_timeout(1);
2583                 goto again;
2584         }
2585 out:
2586         spin_unlock(&delayed_refs->lock);
2587         assert_qgroups_uptodate(trans);
2588         return 0;
2589 }
2590
2591 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2592                                 struct btrfs_root *root,
2593                                 u64 bytenr, u64 num_bytes, u64 flags,
2594                                 int is_data)
2595 {
2596         struct btrfs_delayed_extent_op *extent_op;
2597         int ret;
2598
2599         extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
2600         if (!extent_op)
2601                 return -ENOMEM;
2602
2603         extent_op->flags_to_set = flags;
2604         extent_op->update_flags = 1;
2605         extent_op->update_key = 0;
2606         extent_op->is_data = is_data ? 1 : 0;
2607
2608         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2609                                           num_bytes, extent_op);
2610         if (ret)
2611                 kfree(extent_op);
2612         return ret;
2613 }
2614
2615 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2616                                       struct btrfs_root *root,
2617                                       struct btrfs_path *path,
2618                                       u64 objectid, u64 offset, u64 bytenr)
2619 {
2620         struct btrfs_delayed_ref_head *head;
2621         struct btrfs_delayed_ref_node *ref;
2622         struct btrfs_delayed_data_ref *data_ref;
2623         struct btrfs_delayed_ref_root *delayed_refs;
2624         struct rb_node *node;
2625         int ret = 0;
2626
2627         ret = -ENOENT;
2628         delayed_refs = &trans->transaction->delayed_refs;
2629         spin_lock(&delayed_refs->lock);
2630         head = btrfs_find_delayed_ref_head(trans, bytenr);
2631         if (!head)
2632                 goto out;
2633
2634         if (!mutex_trylock(&head->mutex)) {
2635                 atomic_inc(&head->node.refs);
2636                 spin_unlock(&delayed_refs->lock);
2637
2638                 btrfs_release_path(path);
2639
2640                 /*
2641                  * Mutex was contended, block until it's released and let
2642                  * caller try again
2643                  */
2644                 mutex_lock(&head->mutex);
2645                 mutex_unlock(&head->mutex);
2646                 btrfs_put_delayed_ref(&head->node);
2647                 return -EAGAIN;
2648         }
2649
2650         node = rb_prev(&head->node.rb_node);
2651         if (!node)
2652                 goto out_unlock;
2653
2654         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2655
2656         if (ref->bytenr != bytenr)
2657                 goto out_unlock;
2658
2659         ret = 1;
2660         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2661                 goto out_unlock;
2662
2663         data_ref = btrfs_delayed_node_to_data_ref(ref);
2664
2665         node = rb_prev(node);
2666         if (node) {
2667                 int seq = ref->seq;
2668
2669                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2670                 if (ref->bytenr == bytenr && ref->seq == seq)
2671                         goto out_unlock;
2672         }
2673
2674         if (data_ref->root != root->root_key.objectid ||
2675             data_ref->objectid != objectid || data_ref->offset != offset)
2676                 goto out_unlock;
2677
2678         ret = 0;
2679 out_unlock:
2680         mutex_unlock(&head->mutex);
2681 out:
2682         spin_unlock(&delayed_refs->lock);
2683         return ret;
2684 }
2685
2686 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2687                                         struct btrfs_root *root,
2688                                         struct btrfs_path *path,
2689                                         u64 objectid, u64 offset, u64 bytenr)
2690 {
2691         struct btrfs_root *extent_root = root->fs_info->extent_root;
2692         struct extent_buffer *leaf;
2693         struct btrfs_extent_data_ref *ref;
2694         struct btrfs_extent_inline_ref *iref;
2695         struct btrfs_extent_item *ei;
2696         struct btrfs_key key;
2697         u32 item_size;
2698         int ret;
2699
2700         key.objectid = bytenr;
2701         key.offset = (u64)-1;
2702         key.type = BTRFS_EXTENT_ITEM_KEY;
2703
2704         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2705         if (ret < 0)
2706                 goto out;
2707         BUG_ON(ret == 0); /* Corruption */
2708
2709         ret = -ENOENT;
2710         if (path->slots[0] == 0)
2711                 goto out;
2712
2713         path->slots[0]--;
2714         leaf = path->nodes[0];
2715         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2716
2717         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2718                 goto out;
2719
2720         ret = 1;
2721         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2722 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2723         if (item_size < sizeof(*ei)) {
2724                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2725                 goto out;
2726         }
2727 #endif
2728         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2729
2730         if (item_size != sizeof(*ei) +
2731             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2732                 goto out;
2733
2734         if (btrfs_extent_generation(leaf, ei) <=
2735             btrfs_root_last_snapshot(&root->root_item))
2736                 goto out;
2737
2738         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2739         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2740             BTRFS_EXTENT_DATA_REF_KEY)
2741                 goto out;
2742
2743         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2744         if (btrfs_extent_refs(leaf, ei) !=
2745             btrfs_extent_data_ref_count(leaf, ref) ||
2746             btrfs_extent_data_ref_root(leaf, ref) !=
2747             root->root_key.objectid ||
2748             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2749             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2750                 goto out;
2751
2752         ret = 0;
2753 out:
2754         return ret;
2755 }
2756
2757 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2758                           struct btrfs_root *root,
2759                           u64 objectid, u64 offset, u64 bytenr)
2760 {
2761         struct btrfs_path *path;
2762         int ret;
2763         int ret2;
2764
2765         path = btrfs_alloc_path();
2766         if (!path)
2767                 return -ENOENT;
2768
2769         do {
2770                 ret = check_committed_ref(trans, root, path, objectid,
2771                                           offset, bytenr);
2772                 if (ret && ret != -ENOENT)
2773                         goto out;
2774
2775                 ret2 = check_delayed_ref(trans, root, path, objectid,
2776                                          offset, bytenr);
2777         } while (ret2 == -EAGAIN);
2778
2779         if (ret2 && ret2 != -ENOENT) {
2780                 ret = ret2;
2781                 goto out;
2782         }
2783
2784         if (ret != -ENOENT || ret2 != -ENOENT)
2785                 ret = 0;
2786 out:
2787         btrfs_free_path(path);
2788         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2789                 WARN_ON(ret > 0);
2790         return ret;
2791 }
2792
2793 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2794                            struct btrfs_root *root,
2795                            struct extent_buffer *buf,
2796                            int full_backref, int inc, int for_cow)
2797 {
2798         u64 bytenr;
2799         u64 num_bytes;
2800         u64 parent;
2801         u64 ref_root;
2802         u32 nritems;
2803         struct btrfs_key key;
2804         struct btrfs_file_extent_item *fi;
2805         int i;
2806         int level;
2807         int ret = 0;
2808         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2809                             u64, u64, u64, u64, u64, u64, int);
2810
2811         ref_root = btrfs_header_owner(buf);
2812         nritems = btrfs_header_nritems(buf);
2813         level = btrfs_header_level(buf);
2814
2815         if (!root->ref_cows && level == 0)
2816                 return 0;
2817
2818         if (inc)
2819                 process_func = btrfs_inc_extent_ref;
2820         else
2821                 process_func = btrfs_free_extent;
2822
2823         if (full_backref)
2824                 parent = buf->start;
2825         else
2826                 parent = 0;
2827
2828         for (i = 0; i < nritems; i++) {
2829                 if (level == 0) {
2830                         btrfs_item_key_to_cpu(buf, &key, i);
2831                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
2832                                 continue;
2833                         fi = btrfs_item_ptr(buf, i,
2834                                             struct btrfs_file_extent_item);
2835                         if (btrfs_file_extent_type(buf, fi) ==
2836                             BTRFS_FILE_EXTENT_INLINE)
2837                                 continue;
2838                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
2839                         if (bytenr == 0)
2840                                 continue;
2841
2842                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
2843                         key.offset -= btrfs_file_extent_offset(buf, fi);
2844                         ret = process_func(trans, root, bytenr, num_bytes,
2845                                            parent, ref_root, key.objectid,
2846                                            key.offset, for_cow);
2847                         if (ret)
2848                                 goto fail;
2849                 } else {
2850                         bytenr = btrfs_node_blockptr(buf, i);
2851                         num_bytes = btrfs_level_size(root, level - 1);
2852                         ret = process_func(trans, root, bytenr, num_bytes,
2853                                            parent, ref_root, level - 1, 0,
2854                                            for_cow);
2855                         if (ret)
2856                                 goto fail;
2857                 }
2858         }
2859         return 0;
2860 fail:
2861         return ret;
2862 }
2863
2864 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2865                   struct extent_buffer *buf, int full_backref, int for_cow)
2866 {
2867         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
2868 }
2869
2870 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
2871                   struct extent_buffer *buf, int full_backref, int for_cow)
2872 {
2873         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
2874 }
2875
2876 static int write_one_cache_group(struct btrfs_trans_handle *trans,
2877                                  struct btrfs_root *root,
2878                                  struct btrfs_path *path,
2879                                  struct btrfs_block_group_cache *cache)
2880 {
2881         int ret;
2882         struct btrfs_root *extent_root = root->fs_info->extent_root;
2883         unsigned long bi;
2884         struct extent_buffer *leaf;
2885
2886         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
2887         if (ret < 0)
2888                 goto fail;
2889         BUG_ON(ret); /* Corruption */
2890
2891         leaf = path->nodes[0];
2892         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
2893         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
2894         btrfs_mark_buffer_dirty(leaf);
2895         btrfs_release_path(path);
2896 fail:
2897         if (ret) {
2898                 btrfs_abort_transaction(trans, root, ret);
2899                 return ret;
2900         }
2901         return 0;
2902
2903 }
2904
2905 static struct btrfs_block_group_cache *
2906 next_block_group(struct btrfs_root *root,
2907                  struct btrfs_block_group_cache *cache)
2908 {
2909         struct rb_node *node;
2910         spin_lock(&root->fs_info->block_group_cache_lock);
2911         node = rb_next(&cache->cache_node);
2912         btrfs_put_block_group(cache);
2913         if (node) {
2914                 cache = rb_entry(node, struct btrfs_block_group_cache,
2915                                  cache_node);
2916                 btrfs_get_block_group(cache);
2917         } else
2918                 cache = NULL;
2919         spin_unlock(&root->fs_info->block_group_cache_lock);
2920         return cache;
2921 }
2922
2923 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
2924                             struct btrfs_trans_handle *trans,
2925                             struct btrfs_path *path)
2926 {
2927         struct btrfs_root *root = block_group->fs_info->tree_root;
2928         struct inode *inode = NULL;
2929         u64 alloc_hint = 0;
2930         int dcs = BTRFS_DC_ERROR;
2931         int num_pages = 0;
2932         int retries = 0;
2933         int ret = 0;
2934
2935         /*
2936          * If this block group is smaller than 100 megs don't bother caching the
2937          * block group.
2938          */
2939         if (block_group->key.offset < (100 * 1024 * 1024)) {
2940                 spin_lock(&block_group->lock);
2941                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
2942                 spin_unlock(&block_group->lock);
2943                 return 0;
2944         }
2945
2946 again:
2947         inode = lookup_free_space_inode(root, block_group, path);
2948         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
2949                 ret = PTR_ERR(inode);
2950                 btrfs_release_path(path);
2951                 goto out;
2952         }
2953
2954         if (IS_ERR(inode)) {
2955                 BUG_ON(retries);
2956                 retries++;
2957
2958                 if (block_group->ro)
2959                         goto out_free;
2960
2961                 ret = create_free_space_inode(root, trans, block_group, path);
2962                 if (ret)
2963                         goto out_free;
2964                 goto again;
2965         }
2966
2967         /* We've already setup this transaction, go ahead and exit */
2968         if (block_group->cache_generation == trans->transid &&
2969             i_size_read(inode)) {
2970                 dcs = BTRFS_DC_SETUP;
2971                 goto out_put;
2972         }
2973
2974         /*
2975          * We want to set the generation to 0, that way if anything goes wrong
2976          * from here on out we know not to trust this cache when we load up next
2977          * time.
2978          */
2979         BTRFS_I(inode)->generation = 0;
2980         ret = btrfs_update_inode(trans, root, inode);
2981         WARN_ON(ret);
2982
2983         if (i_size_read(inode) > 0) {
2984                 ret = btrfs_truncate_free_space_cache(root, trans, path,
2985                                                       inode);
2986                 if (ret)
2987                         goto out_put;
2988         }
2989
2990         spin_lock(&block_group->lock);
2991         if (block_group->cached != BTRFS_CACHE_FINISHED ||
2992             !btrfs_test_opt(root, SPACE_CACHE)) {
2993                 /*
2994                  * don't bother trying to write stuff out _if_
2995                  * a) we're not cached,
2996                  * b) we're with nospace_cache mount option.
2997                  */
2998                 dcs = BTRFS_DC_WRITTEN;
2999                 spin_unlock(&block_group->lock);
3000                 goto out_put;
3001         }
3002         spin_unlock(&block_group->lock);
3003
3004         /*
3005          * Try to preallocate enough space based on how big the block group is.
3006          * Keep in mind this has to include any pinned space which could end up
3007          * taking up quite a bit since it's not folded into the other space
3008          * cache.
3009          */
3010         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3011         if (!num_pages)
3012                 num_pages = 1;
3013
3014         num_pages *= 16;
3015         num_pages *= PAGE_CACHE_SIZE;
3016
3017         ret = btrfs_check_data_free_space(inode, num_pages);
3018         if (ret)
3019                 goto out_put;
3020
3021         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3022                                               num_pages, num_pages,
3023                                               &alloc_hint);
3024         if (!ret)
3025                 dcs = BTRFS_DC_SETUP;
3026         btrfs_free_reserved_data_space(inode, num_pages);
3027
3028 out_put:
3029         iput(inode);
3030 out_free:
3031         btrfs_release_path(path);
3032 out:
3033         spin_lock(&block_group->lock);
3034         if (!ret && dcs == BTRFS_DC_SETUP)
3035                 block_group->cache_generation = trans->transid;
3036         block_group->disk_cache_state = dcs;
3037         spin_unlock(&block_group->lock);
3038
3039         return ret;
3040 }
3041
3042 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3043                                    struct btrfs_root *root)
3044 {
3045         struct btrfs_block_group_cache *cache;
3046         int err = 0;
3047         struct btrfs_path *path;
3048         u64 last = 0;
3049
3050         path = btrfs_alloc_path();
3051         if (!path)
3052                 return -ENOMEM;
3053
3054 again:
3055         while (1) {
3056                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3057                 while (cache) {
3058                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3059                                 break;
3060                         cache = next_block_group(root, cache);
3061                 }
3062                 if (!cache) {
3063                         if (last == 0)
3064                                 break;
3065                         last = 0;
3066                         continue;
3067                 }
3068                 err = cache_save_setup(cache, trans, path);
3069                 last = cache->key.objectid + cache->key.offset;
3070                 btrfs_put_block_group(cache);
3071         }
3072
3073         while (1) {
3074                 if (last == 0) {
3075                         err = btrfs_run_delayed_refs(trans, root,
3076                                                      (unsigned long)-1);
3077                         if (err) /* File system offline */
3078                                 goto out;
3079                 }
3080
3081                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3082                 while (cache) {
3083                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3084                                 btrfs_put_block_group(cache);
3085                                 goto again;
3086                         }
3087
3088                         if (cache->dirty)
3089                                 break;
3090                         cache = next_block_group(root, cache);
3091                 }
3092                 if (!cache) {
3093                         if (last == 0)
3094                                 break;
3095                         last = 0;
3096                         continue;
3097                 }
3098
3099                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3100                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3101                 cache->dirty = 0;
3102                 last = cache->key.objectid + cache->key.offset;
3103
3104                 err = write_one_cache_group(trans, root, path, cache);
3105                 if (err) /* File system offline */
3106                         goto out;
3107
3108                 btrfs_put_block_group(cache);
3109         }
3110
3111         while (1) {
3112                 /*
3113                  * I don't think this is needed since we're just marking our
3114                  * preallocated extent as written, but just in case it can't
3115                  * hurt.
3116                  */
3117                 if (last == 0) {
3118                         err = btrfs_run_delayed_refs(trans, root,
3119                                                      (unsigned long)-1);
3120                         if (err) /* File system offline */
3121                                 goto out;
3122                 }
3123
3124                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3125                 while (cache) {
3126                         /*
3127                          * Really this shouldn't happen, but it could if we
3128                          * couldn't write the entire preallocated extent and
3129                          * splitting the extent resulted in a new block.
3130                          */
3131                         if (cache->dirty) {
3132                                 btrfs_put_block_group(cache);
3133                                 goto again;
3134                         }
3135                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3136                                 break;
3137                         cache = next_block_group(root, cache);
3138                 }
3139                 if (!cache) {
3140                         if (last == 0)
3141                                 break;
3142                         last = 0;
3143                         continue;
3144                 }
3145
3146                 err = btrfs_write_out_cache(root, trans, cache, path);
3147
3148                 /*
3149                  * If we didn't have an error then the cache state is still
3150                  * NEED_WRITE, so we can set it to WRITTEN.
3151                  */
3152                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3153                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3154                 last = cache->key.objectid + cache->key.offset;
3155                 btrfs_put_block_group(cache);
3156         }
3157 out:
3158
3159         btrfs_free_path(path);
3160         return err;
3161 }
3162
3163 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3164 {
3165         struct btrfs_block_group_cache *block_group;
3166         int readonly = 0;
3167
3168         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3169         if (!block_group || block_group->ro)
3170                 readonly = 1;
3171         if (block_group)
3172                 btrfs_put_block_group(block_group);
3173         return readonly;
3174 }
3175
3176 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3177                              u64 total_bytes, u64 bytes_used,
3178                              struct btrfs_space_info **space_info)
3179 {
3180         struct btrfs_space_info *found;
3181         int i;
3182         int factor;
3183
3184         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3185                      BTRFS_BLOCK_GROUP_RAID10))
3186                 factor = 2;
3187         else
3188                 factor = 1;
3189
3190         found = __find_space_info(info, flags);
3191         if (found) {
3192                 spin_lock(&found->lock);
3193                 found->total_bytes += total_bytes;
3194                 found->disk_total += total_bytes * factor;
3195                 found->bytes_used += bytes_used;
3196                 found->disk_used += bytes_used * factor;
3197                 found->full = 0;
3198                 spin_unlock(&found->lock);
3199                 *space_info = found;
3200                 return 0;
3201         }
3202         found = kzalloc(sizeof(*found), GFP_NOFS);
3203         if (!found)
3204                 return -ENOMEM;
3205
3206         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3207                 INIT_LIST_HEAD(&found->block_groups[i]);
3208         init_rwsem(&found->groups_sem);
3209         spin_lock_init(&found->lock);
3210         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3211         found->total_bytes = total_bytes;
3212         found->disk_total = total_bytes * factor;
3213         found->bytes_used = bytes_used;
3214         found->disk_used = bytes_used * factor;
3215         found->bytes_pinned = 0;
3216         found->bytes_reserved = 0;
3217         found->bytes_readonly = 0;
3218         found->bytes_may_use = 0;
3219         found->full = 0;
3220         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3221         found->chunk_alloc = 0;
3222         found->flush = 0;
3223         init_waitqueue_head(&found->wait);
3224         *space_info = found;
3225         list_add_rcu(&found->list, &info->space_info);
3226         if (flags & BTRFS_BLOCK_GROUP_DATA)
3227                 info->data_sinfo = found;
3228         return 0;
3229 }
3230
3231 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3232 {
3233         u64 extra_flags = chunk_to_extended(flags) &
3234                                 BTRFS_EXTENDED_PROFILE_MASK;
3235
3236         if (flags & BTRFS_BLOCK_GROUP_DATA)
3237                 fs_info->avail_data_alloc_bits |= extra_flags;
3238         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3239                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3240         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3241                 fs_info->avail_system_alloc_bits |= extra_flags;
3242 }
3243
3244 /*
3245  * returns target flags in extended format or 0 if restripe for this
3246  * chunk_type is not in progress
3247  *
3248  * should be called with either volume_mutex or balance_lock held
3249  */
3250 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3251 {
3252         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3253         u64 target = 0;
3254
3255         if (!bctl)
3256                 return 0;
3257
3258         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3259             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3260                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3261         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3262                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3263                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3264         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3265                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3266                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3267         }
3268
3269         return target;
3270 }
3271
3272 /*
3273  * @flags: available profiles in extended format (see ctree.h)
3274  *
3275  * Returns reduced profile in chunk format.  If profile changing is in
3276  * progress (either running or paused) picks the target profile (if it's
3277  * already available), otherwise falls back to plain reducing.
3278  */
3279 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3280 {
3281         /*
3282          * we add in the count of missing devices because we want
3283          * to make sure that any RAID levels on a degraded FS
3284          * continue to be honored.
3285          */
3286         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3287                 root->fs_info->fs_devices->missing_devices;
3288         u64 target;
3289
3290         /*
3291          * see if restripe for this chunk_type is in progress, if so
3292          * try to reduce to the target profile
3293          */
3294         spin_lock(&root->fs_info->balance_lock);
3295         target = get_restripe_target(root->fs_info, flags);
3296         if (target) {
3297                 /* pick target profile only if it's already available */
3298                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3299                         spin_unlock(&root->fs_info->balance_lock);
3300                         return extended_to_chunk(target);
3301                 }
3302         }
3303         spin_unlock(&root->fs_info->balance_lock);
3304
3305         if (num_devices == 1)
3306                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0);
3307         if (num_devices < 4)
3308                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3309
3310         if ((flags & BTRFS_BLOCK_GROUP_DUP) &&
3311             (flags & (BTRFS_BLOCK_GROUP_RAID1 |
3312                       BTRFS_BLOCK_GROUP_RAID10))) {
3313                 flags &= ~BTRFS_BLOCK_GROUP_DUP;
3314         }
3315
3316         if ((flags & BTRFS_BLOCK_GROUP_RAID1) &&
3317             (flags & BTRFS_BLOCK_GROUP_RAID10)) {
3318                 flags &= ~BTRFS_BLOCK_GROUP_RAID1;
3319         }
3320
3321         if ((flags & BTRFS_BLOCK_GROUP_RAID0) &&
3322             ((flags & BTRFS_BLOCK_GROUP_RAID1) |
3323              (flags & BTRFS_BLOCK_GROUP_RAID10) |
3324              (flags & BTRFS_BLOCK_GROUP_DUP))) {
3325                 flags &= ~BTRFS_BLOCK_GROUP_RAID0;
3326         }
3327
3328         return extended_to_chunk(flags);
3329 }
3330
3331 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3332 {
3333         if (flags & BTRFS_BLOCK_GROUP_DATA)
3334                 flags |= root->fs_info->avail_data_alloc_bits;
3335         else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3336                 flags |= root->fs_info->avail_system_alloc_bits;
3337         else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3338                 flags |= root->fs_info->avail_metadata_alloc_bits;
3339
3340         return btrfs_reduce_alloc_profile(root, flags);
3341 }
3342
3343 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3344 {
3345         u64 flags;
3346
3347         if (data)
3348                 flags = BTRFS_BLOCK_GROUP_DATA;
3349         else if (root == root->fs_info->chunk_root)
3350                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3351         else
3352                 flags = BTRFS_BLOCK_GROUP_METADATA;
3353
3354         return get_alloc_profile(root, flags);
3355 }
3356
3357 /*
3358  * This will check the space that the inode allocates from to make sure we have
3359  * enough space for bytes.
3360  */
3361 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3362 {
3363         struct btrfs_space_info *data_sinfo;
3364         struct btrfs_root *root = BTRFS_I(inode)->root;
3365         struct btrfs_fs_info *fs_info = root->fs_info;
3366         u64 used;
3367         int ret = 0, committed = 0, alloc_chunk = 1;
3368
3369         /* make sure bytes are sectorsize aligned */
3370         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3371
3372         if (root == root->fs_info->tree_root ||
3373             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3374                 alloc_chunk = 0;
3375                 committed = 1;
3376         }
3377
3378         data_sinfo = fs_info->data_sinfo;
3379         if (!data_sinfo)
3380                 goto alloc;
3381
3382 again:
3383         /* make sure we have enough space to handle the data first */
3384         spin_lock(&data_sinfo->lock);
3385         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3386                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3387                 data_sinfo->bytes_may_use;
3388
3389         if (used + bytes > data_sinfo->total_bytes) {
3390                 struct btrfs_trans_handle *trans;
3391
3392                 /*
3393                  * if we don't have enough free bytes in this space then we need
3394                  * to alloc a new chunk.
3395                  */
3396                 if (!data_sinfo->full && alloc_chunk) {
3397                         u64 alloc_target;
3398
3399                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3400                         spin_unlock(&data_sinfo->lock);
3401 alloc:
3402                         alloc_target = btrfs_get_alloc_profile(root, 1);
3403                         trans = btrfs_join_transaction(root);
3404                         if (IS_ERR(trans))
3405                                 return PTR_ERR(trans);
3406
3407                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3408                                              alloc_target,
3409                                              CHUNK_ALLOC_NO_FORCE);
3410                         btrfs_end_transaction(trans, root);
3411                         if (ret < 0) {
3412                                 if (ret != -ENOSPC)
3413                                         return ret;
3414                                 else
3415                                         goto commit_trans;
3416                         }
3417
3418                         if (!data_sinfo)
3419                                 data_sinfo = fs_info->data_sinfo;
3420
3421                         goto again;
3422                 }
3423
3424                 /*
3425                  * If we have less pinned bytes than we want to allocate then
3426                  * don't bother committing the transaction, it won't help us.
3427                  */
3428                 if (data_sinfo->bytes_pinned < bytes)
3429                         committed = 1;
3430                 spin_unlock(&data_sinfo->lock);
3431
3432                 /* commit the current transaction and try again */
3433 commit_trans:
3434                 if (!committed &&
3435                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3436                         committed = 1;
3437                         trans = btrfs_join_transaction(root);
3438                         if (IS_ERR(trans))
3439                                 return PTR_ERR(trans);
3440                         ret = btrfs_commit_transaction(trans, root);
3441                         if (ret)
3442                                 return ret;
3443                         goto again;
3444                 }
3445
3446                 return -ENOSPC;
3447         }
3448         data_sinfo->bytes_may_use += bytes;
3449         trace_btrfs_space_reservation(root->fs_info, "space_info",
3450                                       data_sinfo->flags, bytes, 1);
3451         spin_unlock(&data_sinfo->lock);
3452
3453         return 0;
3454 }
3455
3456 /*
3457  * Called if we need to clear a data reservation for this inode.
3458  */
3459 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3460 {
3461         struct btrfs_root *root = BTRFS_I(inode)->root;
3462         struct btrfs_space_info *data_sinfo;
3463
3464         /* make sure bytes are sectorsize aligned */
3465         bytes = (bytes + root->sectorsize - 1) & ~((u64)root->sectorsize - 1);
3466
3467         data_sinfo = root->fs_info->data_sinfo;
3468         spin_lock(&data_sinfo->lock);
3469         data_sinfo->bytes_may_use -= bytes;
3470         trace_btrfs_space_reservation(root->fs_info, "space_info",
3471                                       data_sinfo->flags, bytes, 0);
3472         spin_unlock(&data_sinfo->lock);
3473 }
3474
3475 static void force_metadata_allocation(struct btrfs_fs_info *info)
3476 {
3477         struct list_head *head = &info->space_info;
3478         struct btrfs_space_info *found;
3479
3480         rcu_read_lock();
3481         list_for_each_entry_rcu(found, head, list) {
3482                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3483                         found->force_alloc = CHUNK_ALLOC_FORCE;
3484         }
3485         rcu_read_unlock();
3486 }
3487
3488 static int should_alloc_chunk(struct btrfs_root *root,
3489                               struct btrfs_space_info *sinfo, int force)
3490 {
3491         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3492         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3493         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3494         u64 thresh;
3495
3496         if (force == CHUNK_ALLOC_FORCE)
3497                 return 1;
3498
3499         /*
3500          * We need to take into account the global rsv because for all intents
3501          * and purposes it's used space.  Don't worry about locking the
3502          * global_rsv, it doesn't change except when the transaction commits.
3503          */
3504         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3505                 num_allocated += global_rsv->size;
3506
3507         /*
3508          * in limited mode, we want to have some free space up to
3509          * about 1% of the FS size.
3510          */
3511         if (force == CHUNK_ALLOC_LIMITED) {
3512                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3513                 thresh = max_t(u64, 64 * 1024 * 1024,
3514                                div_factor_fine(thresh, 1));
3515
3516                 if (num_bytes - num_allocated < thresh)
3517                         return 1;
3518         }
3519
3520         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3521                 return 0;
3522         return 1;
3523 }
3524
3525 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3526 {
3527         u64 num_dev;
3528
3529         if (type & BTRFS_BLOCK_GROUP_RAID10 ||
3530             type & BTRFS_BLOCK_GROUP_RAID0)
3531                 num_dev = root->fs_info->fs_devices->rw_devices;
3532         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3533                 num_dev = 2;
3534         else
3535                 num_dev = 1;    /* DUP or single */
3536
3537         /* metadata for updaing devices and chunk tree */
3538         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3539 }
3540
3541 static void check_system_chunk(struct btrfs_trans_handle *trans,
3542                                struct btrfs_root *root, u64 type)
3543 {
3544         struct btrfs_space_info *info;
3545         u64 left;
3546         u64 thresh;
3547
3548         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3549         spin_lock(&info->lock);
3550         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3551                 info->bytes_reserved - info->bytes_readonly;
3552         spin_unlock(&info->lock);
3553
3554         thresh = get_system_chunk_thresh(root, type);
3555         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3556                 printk(KERN_INFO "left=%llu, need=%llu, flags=%llu\n",
3557                        left, thresh, type);
3558                 dump_space_info(info, 0, 0);
3559         }
3560
3561         if (left < thresh) {
3562                 u64 flags;
3563
3564                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3565                 btrfs_alloc_chunk(trans, root, flags);
3566         }
3567 }
3568
3569 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3570                           struct btrfs_root *extent_root, u64 flags, int force)
3571 {
3572         struct btrfs_space_info *space_info;
3573         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3574         int wait_for_alloc = 0;
3575         int ret = 0;
3576
3577         space_info = __find_space_info(extent_root->fs_info, flags);
3578         if (!space_info) {
3579                 ret = update_space_info(extent_root->fs_info, flags,
3580                                         0, 0, &space_info);
3581                 BUG_ON(ret); /* -ENOMEM */
3582         }
3583         BUG_ON(!space_info); /* Logic error */
3584
3585 again:
3586         spin_lock(&space_info->lock);
3587         if (force < space_info->force_alloc)
3588                 force = space_info->force_alloc;
3589         if (space_info->full) {
3590                 spin_unlock(&space_info->lock);
3591                 return 0;
3592         }
3593
3594         if (!should_alloc_chunk(extent_root, space_info, force)) {
3595                 spin_unlock(&space_info->lock);
3596                 return 0;
3597         } else if (space_info->chunk_alloc) {
3598                 wait_for_alloc = 1;
3599         } else {
3600                 space_info->chunk_alloc = 1;
3601         }
3602
3603         spin_unlock(&space_info->lock);
3604
3605         mutex_lock(&fs_info->chunk_mutex);
3606
3607         /*
3608          * The chunk_mutex is held throughout the entirety of a chunk
3609          * allocation, so once we've acquired the chunk_mutex we know that the
3610          * other guy is done and we need to recheck and see if we should
3611          * allocate.
3612          */
3613         if (wait_for_alloc) {
3614                 mutex_unlock(&fs_info->chunk_mutex);
3615                 wait_for_alloc = 0;
3616                 goto again;
3617         }
3618
3619         /*
3620          * If we have mixed data/metadata chunks we want to make sure we keep
3621          * allocating mixed chunks instead of individual chunks.
3622          */
3623         if (btrfs_mixed_space_info(space_info))
3624                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3625
3626         /*
3627          * if we're doing a data chunk, go ahead and make sure that
3628          * we keep a reasonable number of metadata chunks allocated in the
3629          * FS as well.
3630          */
3631         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3632                 fs_info->data_chunk_allocations++;
3633                 if (!(fs_info->data_chunk_allocations %
3634                       fs_info->metadata_ratio))
3635                         force_metadata_allocation(fs_info);
3636         }
3637
3638         /*
3639          * Check if we have enough space in SYSTEM chunk because we may need
3640          * to update devices.
3641          */
3642         check_system_chunk(trans, extent_root, flags);
3643
3644         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3645         if (ret < 0 && ret != -ENOSPC)
3646                 goto out;
3647
3648         spin_lock(&space_info->lock);
3649         if (ret)
3650                 space_info->full = 1;
3651         else
3652                 ret = 1;
3653
3654         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3655         space_info->chunk_alloc = 0;
3656         spin_unlock(&space_info->lock);
3657 out:
3658         mutex_unlock(&fs_info->chunk_mutex);
3659         return ret;
3660 }
3661
3662 static int can_overcommit(struct btrfs_root *root,
3663                           struct btrfs_space_info *space_info, u64 bytes,
3664                           int flush)
3665 {
3666         u64 profile = btrfs_get_alloc_profile(root, 0);
3667         u64 avail;
3668         u64 used;
3669
3670         used = space_info->bytes_used + space_info->bytes_reserved +
3671                 space_info->bytes_pinned + space_info->bytes_readonly +
3672                 space_info->bytes_may_use;
3673
3674         spin_lock(&root->fs_info->free_chunk_lock);
3675         avail = root->fs_info->free_chunk_space;
3676         spin_unlock(&root->fs_info->free_chunk_lock);
3677
3678         /*
3679          * If we have dup, raid1 or raid10 then only half of the free
3680          * space is actually useable.
3681          */
3682         if (profile & (BTRFS_BLOCK_GROUP_DUP |
3683                        BTRFS_BLOCK_GROUP_RAID1 |
3684                        BTRFS_BLOCK_GROUP_RAID10))
3685                 avail >>= 1;
3686
3687         /*
3688          * If we aren't flushing don't let us overcommit too much, say
3689          * 1/8th of the space.  If we can flush, let it overcommit up to
3690          * 1/2 of the space.
3691          */
3692         if (flush)
3693                 avail >>= 3;
3694         else
3695                 avail >>= 1;
3696
3697         if (used + bytes < space_info->total_bytes + avail)
3698                 return 1;
3699         return 0;
3700 }
3701
3702 /*
3703  * shrink metadata reservation for delalloc
3704  */
3705 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3706                             bool wait_ordered)
3707 {
3708         struct btrfs_block_rsv *block_rsv;
3709         struct btrfs_space_info *space_info;
3710         struct btrfs_trans_handle *trans;
3711         u64 delalloc_bytes;
3712         u64 max_reclaim;
3713         long time_left;
3714         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3715         int loops = 0;
3716
3717         trans = (struct btrfs_trans_handle *)current->journal_info;
3718         block_rsv = &root->fs_info->delalloc_block_rsv;
3719         space_info = block_rsv->space_info;
3720
3721         smp_mb();
3722         delalloc_bytes = root->fs_info->delalloc_bytes;
3723         if (delalloc_bytes == 0) {
3724                 if (trans)
3725                         return;
3726                 btrfs_wait_ordered_extents(root, 0);
3727                 return;
3728         }
3729
3730         while (delalloc_bytes && loops < 3) {
3731                 max_reclaim = min(delalloc_bytes, to_reclaim);
3732                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3733                 writeback_inodes_sb_nr_if_idle(root->fs_info->sb, nr_pages,
3734                                                WB_REASON_FS_FREE_SPACE);
3735
3736                 /*
3737                  * We need to wait for the async pages to actually start before
3738                  * we do anything.
3739                  */
3740                 wait_event(root->fs_info->async_submit_wait,
3741                            !atomic_read(&root->fs_info->async_delalloc_pages));
3742
3743                 spin_lock(&space_info->lock);
3744                 if (can_overcommit(root, space_info, orig, !trans)) {
3745                         spin_unlock(&space_info->lock);
3746                         break;
3747                 }
3748                 spin_unlock(&space_info->lock);
3749
3750                 loops++;
3751                 if (wait_ordered && !trans) {
3752                         btrfs_wait_ordered_extents(root, 0);
3753                 } else {
3754                         time_left = schedule_timeout_killable(1);
3755                         if (time_left)
3756                                 break;
3757                 }
3758                 smp_mb();
3759                 delalloc_bytes = root->fs_info->delalloc_bytes;
3760         }
3761 }
3762
3763 /**
3764  * maybe_commit_transaction - possibly commit the transaction if its ok to
3765  * @root - the root we're allocating for
3766  * @bytes - the number of bytes we want to reserve
3767  * @force - force the commit
3768  *
3769  * This will check to make sure that committing the transaction will actually
3770  * get us somewhere and then commit the transaction if it does.  Otherwise it
3771  * will return -ENOSPC.
3772  */
3773 static int may_commit_transaction(struct btrfs_root *root,
3774                                   struct btrfs_space_info *space_info,
3775                                   u64 bytes, int force)
3776 {
3777         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
3778         struct btrfs_trans_handle *trans;
3779
3780         trans = (struct btrfs_trans_handle *)current->journal_info;
3781         if (trans)
3782                 return -EAGAIN;
3783
3784         if (force)
3785                 goto commit;
3786
3787         /* See if there is enough pinned space to make this reservation */
3788         spin_lock(&space_info->lock);
3789         if (space_info->bytes_pinned >= bytes) {
3790                 spin_unlock(&space_info->lock);
3791                 goto commit;
3792         }
3793         spin_unlock(&space_info->lock);
3794
3795         /*
3796          * See if there is some space in the delayed insertion reservation for
3797          * this reservation.
3798          */
3799         if (space_info != delayed_rsv->space_info)
3800                 return -ENOSPC;
3801
3802         spin_lock(&space_info->lock);
3803         spin_lock(&delayed_rsv->lock);
3804         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
3805                 spin_unlock(&delayed_rsv->lock);
3806                 spin_unlock(&space_info->lock);
3807                 return -ENOSPC;
3808         }
3809         spin_unlock(&delayed_rsv->lock);
3810         spin_unlock(&space_info->lock);
3811
3812 commit:
3813         trans = btrfs_join_transaction(root);
3814         if (IS_ERR(trans))
3815                 return -ENOSPC;
3816
3817         return btrfs_commit_transaction(trans, root);
3818 }
3819
3820 enum flush_state {
3821         FLUSH_DELAYED_ITEMS_NR  =       1,
3822         FLUSH_DELAYED_ITEMS     =       2,
3823         FLUSH_DELALLOC          =       3,
3824         FLUSH_DELALLOC_WAIT     =       4,
3825         ALLOC_CHUNK             =       5,
3826         COMMIT_TRANS            =       6,
3827 };
3828
3829 static int flush_space(struct btrfs_root *root,
3830                        struct btrfs_space_info *space_info, u64 num_bytes,
3831                        u64 orig_bytes, int state)
3832 {
3833         struct btrfs_trans_handle *trans;
3834         int nr;
3835         int ret = 0;
3836
3837         switch (state) {
3838         case FLUSH_DELAYED_ITEMS_NR:
3839         case FLUSH_DELAYED_ITEMS:
3840                 if (state == FLUSH_DELAYED_ITEMS_NR) {
3841                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
3842
3843                         nr = (int)div64_u64(num_bytes, bytes);
3844                         if (!nr)
3845                                 nr = 1;
3846                         nr *= 2;
3847                 } else {
3848                         nr = -1;
3849                 }
3850                 trans = btrfs_join_transaction(root);
3851                 if (IS_ERR(trans)) {
3852                         ret = PTR_ERR(trans);
3853                         break;
3854                 }
3855                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
3856                 btrfs_end_transaction(trans, root);
3857                 break;
3858         case FLUSH_DELALLOC:
3859         case FLUSH_DELALLOC_WAIT:
3860                 shrink_delalloc(root, num_bytes, orig_bytes,
3861                                 state == FLUSH_DELALLOC_WAIT);
3862                 break;
3863         case ALLOC_CHUNK:
3864                 trans = btrfs_join_transaction(root);
3865                 if (IS_ERR(trans)) {
3866                         ret = PTR_ERR(trans);
3867                         break;
3868                 }
3869                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3870                                      btrfs_get_alloc_profile(root, 0),
3871                                      CHUNK_ALLOC_NO_FORCE);
3872                 btrfs_end_transaction(trans, root);
3873                 if (ret == -ENOSPC)
3874                         ret = 0;
3875                 break;
3876         case COMMIT_TRANS:
3877                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
3878                 break;
3879         default:
3880                 ret = -ENOSPC;
3881                 break;
3882         }
3883
3884         return ret;
3885 }
3886 /**
3887  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
3888  * @root - the root we're allocating for
3889  * @block_rsv - the block_rsv we're allocating for
3890  * @orig_bytes - the number of bytes we want
3891  * @flush - wether or not we can flush to make our reservation
3892  *
3893  * This will reserve orgi_bytes number of bytes from the space info associated
3894  * with the block_rsv.  If there is not enough space it will make an attempt to
3895  * flush out space to make room.  It will do this by flushing delalloc if
3896  * possible or committing the transaction.  If flush is 0 then no attempts to
3897  * regain reservations will be made and this will fail if there is not enough
3898  * space already.
3899  */
3900 static int reserve_metadata_bytes(struct btrfs_root *root,
3901                                   struct btrfs_block_rsv *block_rsv,
3902                                   u64 orig_bytes, int flush)
3903 {
3904         struct btrfs_space_info *space_info = block_rsv->space_info;
3905         u64 used;
3906         u64 num_bytes = orig_bytes;
3907         int flush_state = FLUSH_DELAYED_ITEMS_NR;
3908         int ret = 0;
3909         bool flushing = false;
3910
3911 again:
3912         ret = 0;
3913         spin_lock(&space_info->lock);
3914         /*
3915          * We only want to wait if somebody other than us is flushing and we are
3916          * actually alloed to flush.
3917          */
3918         while (flush && !flushing && space_info->flush) {
3919                 spin_unlock(&space_info->lock);
3920                 /*
3921                  * If we have a trans handle we can't wait because the flusher
3922                  * may have to commit the transaction, which would mean we would
3923                  * deadlock since we are waiting for the flusher to finish, but
3924                  * hold the current transaction open.
3925                  */
3926                 if (current->journal_info)
3927                         return -EAGAIN;
3928                 ret = wait_event_killable(space_info->wait, !space_info->flush);
3929                 /* Must have been killed, return */
3930                 if (ret)
3931                         return -EINTR;
3932
3933                 spin_lock(&space_info->lock);
3934         }
3935
3936         ret = -ENOSPC;
3937         used = space_info->bytes_used + space_info->bytes_reserved +
3938                 space_info->bytes_pinned + space_info->bytes_readonly +
3939                 space_info->bytes_may_use;
3940
3941         /*
3942          * The idea here is that we've not already over-reserved the block group
3943          * then we can go ahead and save our reservation first and then start
3944          * flushing if we need to.  Otherwise if we've already overcommitted
3945          * lets start flushing stuff first and then come back and try to make
3946          * our reservation.
3947          */
3948         if (used <= space_info->total_bytes) {
3949                 if (used + orig_bytes <= space_info->total_bytes) {
3950                         space_info->bytes_may_use += orig_bytes;
3951                         trace_btrfs_space_reservation(root->fs_info,
3952                                 "space_info", space_info->flags, orig_bytes, 1);
3953                         ret = 0;
3954                 } else {
3955                         /*
3956                          * Ok set num_bytes to orig_bytes since we aren't
3957                          * overocmmitted, this way we only try and reclaim what
3958                          * we need.
3959                          */
3960                         num_bytes = orig_bytes;
3961                 }
3962         } else {
3963                 /*
3964                  * Ok we're over committed, set num_bytes to the overcommitted
3965                  * amount plus the amount of bytes that we need for this
3966                  * reservation.
3967                  */
3968                 num_bytes = used - space_info->total_bytes +
3969                         (orig_bytes * 2);
3970         }
3971
3972         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
3973                 space_info->bytes_may_use += orig_bytes;
3974                 trace_btrfs_space_reservation(root->fs_info, "space_info",
3975                                               space_info->flags, orig_bytes,
3976                                               1);
3977                 ret = 0;
3978         }
3979
3980         /*
3981          * Couldn't make our reservation, save our place so while we're trying
3982          * to reclaim space we can actually use it instead of somebody else
3983          * stealing it from us.
3984          */
3985         if (ret && flush) {
3986                 flushing = true;
3987                 space_info->flush = 1;
3988         }
3989
3990         spin_unlock(&space_info->lock);
3991
3992         if (!ret || !flush)
3993                 goto out;
3994
3995         ret = flush_space(root, space_info, num_bytes, orig_bytes,
3996                           flush_state);
3997         flush_state++;
3998         if (!ret)
3999                 goto again;
4000         else if (flush_state <= COMMIT_TRANS)
4001                 goto again;
4002
4003 out:
4004         if (flushing) {
4005                 spin_lock(&space_info->lock);
4006                 space_info->flush = 0;
4007                 wake_up_all(&space_info->wait);
4008                 spin_unlock(&space_info->lock);
4009         }
4010         return ret;
4011 }
4012
4013 static struct btrfs_block_rsv *get_block_rsv(
4014                                         const struct btrfs_trans_handle *trans,
4015                                         const struct btrfs_root *root)
4016 {
4017         struct btrfs_block_rsv *block_rsv = NULL;
4018
4019         if (root->ref_cows)
4020                 block_rsv = trans->block_rsv;
4021
4022         if (root == root->fs_info->csum_root && trans->adding_csums)
4023                 block_rsv = trans->block_rsv;
4024
4025         if (!block_rsv)
4026                 block_rsv = root->block_rsv;
4027
4028         if (!block_rsv)
4029                 block_rsv = &root->fs_info->empty_block_rsv;
4030
4031         return block_rsv;
4032 }
4033
4034 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4035                                u64 num_bytes)
4036 {
4037         int ret = -ENOSPC;
4038         spin_lock(&block_rsv->lock);
4039         if (block_rsv->reserved >= num_bytes) {
4040                 block_rsv->reserved -= num_bytes;
4041                 if (block_rsv->reserved < block_rsv->size)
4042                         block_rsv->full = 0;
4043                 ret = 0;
4044         }
4045         spin_unlock(&block_rsv->lock);
4046         return ret;
4047 }
4048
4049 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4050                                 u64 num_bytes, int update_size)
4051 {
4052         spin_lock(&block_rsv->lock);
4053         block_rsv->reserved += num_bytes;
4054         if (update_size)
4055                 block_rsv->size += num_bytes;
4056         else if (block_rsv->reserved >= block_rsv->size)
4057                 block_rsv->full = 1;
4058         spin_unlock(&block_rsv->lock);
4059 }
4060
4061 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4062                                     struct btrfs_block_rsv *block_rsv,
4063                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4064 {
4065         struct btrfs_space_info *space_info = block_rsv->space_info;
4066
4067         spin_lock(&block_rsv->lock);
4068         if (num_bytes == (u64)-1)
4069                 num_bytes = block_rsv->size;
4070         block_rsv->size -= num_bytes;
4071         if (block_rsv->reserved >= block_rsv->size) {
4072                 num_bytes = block_rsv->reserved - block_rsv->size;
4073                 block_rsv->reserved = block_rsv->size;
4074                 block_rsv->full = 1;
4075         } else {
4076                 num_bytes = 0;
4077         }
4078         spin_unlock(&block_rsv->lock);
4079
4080         if (num_bytes > 0) {
4081                 if (dest) {
4082                         spin_lock(&dest->lock);
4083                         if (!dest->full) {
4084                                 u64 bytes_to_add;
4085
4086                                 bytes_to_add = dest->size - dest->reserved;
4087                                 bytes_to_add = min(num_bytes, bytes_to_add);
4088                                 dest->reserved += bytes_to_add;
4089                                 if (dest->reserved >= dest->size)
4090                                         dest->full = 1;
4091                                 num_bytes -= bytes_to_add;
4092                         }
4093                         spin_unlock(&dest->lock);
4094                 }
4095                 if (num_bytes) {
4096                         spin_lock(&space_info->lock);
4097                         space_info->bytes_may_use -= num_bytes;
4098                         trace_btrfs_space_reservation(fs_info, "space_info",
4099                                         space_info->flags, num_bytes, 0);
4100                         space_info->reservation_progress++;
4101                         spin_unlock(&space_info->lock);
4102                 }
4103         }
4104 }
4105
4106 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4107                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4108 {
4109         int ret;
4110
4111         ret = block_rsv_use_bytes(src, num_bytes);
4112         if (ret)
4113                 return ret;
4114
4115         block_rsv_add_bytes(dst, num_bytes, 1);
4116         return 0;
4117 }
4118
4119 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4120 {
4121         memset(rsv, 0, sizeof(*rsv));
4122         spin_lock_init(&rsv->lock);
4123         rsv->type = type;
4124 }
4125
4126 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4127                                               unsigned short type)
4128 {
4129         struct btrfs_block_rsv *block_rsv;
4130         struct btrfs_fs_info *fs_info = root->fs_info;
4131
4132         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4133         if (!block_rsv)
4134                 return NULL;
4135
4136         btrfs_init_block_rsv(block_rsv, type);
4137         block_rsv->space_info = __find_space_info(fs_info,
4138                                                   BTRFS_BLOCK_GROUP_METADATA);
4139         return block_rsv;
4140 }
4141
4142 void btrfs_free_block_rsv(struct btrfs_root *root,
4143                           struct btrfs_block_rsv *rsv)
4144 {
4145         if (!rsv)
4146                 return;
4147         btrfs_block_rsv_release(root, rsv, (u64)-1);
4148         kfree(rsv);
4149 }
4150
4151 static inline int __block_rsv_add(struct btrfs_root *root,
4152                                   struct btrfs_block_rsv *block_rsv,
4153                                   u64 num_bytes, int flush)
4154 {
4155         int ret;
4156
4157         if (num_bytes == 0)
4158                 return 0;
4159
4160         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4161         if (!ret) {
4162                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4163                 return 0;
4164         }
4165
4166         return ret;
4167 }
4168
4169 int btrfs_block_rsv_add(struct btrfs_root *root,
4170                         struct btrfs_block_rsv *block_rsv,
4171                         u64 num_bytes)
4172 {
4173         return __block_rsv_add(root, block_rsv, num_bytes, 1);
4174 }
4175
4176 int btrfs_block_rsv_add_noflush(struct btrfs_root *root,
4177                                 struct btrfs_block_rsv *block_rsv,
4178                                 u64 num_bytes)
4179 {
4180         return __block_rsv_add(root, block_rsv, num_bytes, 0);
4181 }
4182
4183 int btrfs_block_rsv_check(struct btrfs_root *root,
4184                           struct btrfs_block_rsv *block_rsv, int min_factor)
4185 {
4186         u64 num_bytes = 0;
4187         int ret = -ENOSPC;
4188
4189         if (!block_rsv)
4190                 return 0;
4191
4192         spin_lock(&block_rsv->lock);
4193         num_bytes = div_factor(block_rsv->size, min_factor);
4194         if (block_rsv->reserved >= num_bytes)
4195                 ret = 0;
4196         spin_unlock(&block_rsv->lock);
4197
4198         return ret;
4199 }
4200
4201 static inline int __btrfs_block_rsv_refill(struct btrfs_root *root,
4202                                            struct btrfs_block_rsv *block_rsv,
4203                                            u64 min_reserved, int flush)
4204 {
4205         u64 num_bytes = 0;
4206         int ret = -ENOSPC;
4207
4208         if (!block_rsv)
4209                 return 0;
4210
4211         spin_lock(&block_rsv->lock);
4212         num_bytes = min_reserved;
4213         if (block_rsv->reserved >= num_bytes)
4214                 ret = 0;
4215         else
4216                 num_bytes -= block_rsv->reserved;
4217         spin_unlock(&block_rsv->lock);
4218
4219         if (!ret)
4220                 return 0;
4221
4222         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4223         if (!ret) {
4224                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4225                 return 0;
4226         }
4227
4228         return ret;
4229 }
4230
4231 int btrfs_block_rsv_refill(struct btrfs_root *root,
4232                            struct btrfs_block_rsv *block_rsv,
4233                            u64 min_reserved)
4234 {
4235         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 1);
4236 }
4237
4238 int btrfs_block_rsv_refill_noflush(struct btrfs_root *root,
4239                                    struct btrfs_block_rsv *block_rsv,
4240                                    u64 min_reserved)
4241 {
4242         return __btrfs_block_rsv_refill(root, block_rsv, min_reserved, 0);
4243 }
4244
4245 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4246                             struct btrfs_block_rsv *dst_rsv,
4247                             u64 num_bytes)
4248 {
4249         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4250 }
4251
4252 void btrfs_block_rsv_release(struct btrfs_root *root,
4253                              struct btrfs_block_rsv *block_rsv,
4254                              u64 num_bytes)
4255 {
4256         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4257         if (global_rsv->full || global_rsv == block_rsv ||
4258             block_rsv->space_info != global_rsv->space_info)
4259                 global_rsv = NULL;
4260         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4261                                 num_bytes);
4262 }
4263
4264 /*
4265  * helper to calculate size of global block reservation.
4266  * the desired value is sum of space used by extent tree,
4267  * checksum tree and root tree
4268  */
4269 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4270 {
4271         struct btrfs_space_info *sinfo;
4272         u64 num_bytes;
4273         u64 meta_used;
4274         u64 data_used;
4275         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4276
4277         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4278         spin_lock(&sinfo->lock);
4279         data_used = sinfo->bytes_used;
4280         spin_unlock(&sinfo->lock);
4281
4282         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4283         spin_lock(&sinfo->lock);
4284         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4285                 data_used = 0;
4286         meta_used = sinfo->bytes_used;
4287         spin_unlock(&sinfo->lock);
4288
4289         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4290                     csum_size * 2;
4291         num_bytes += div64_u64(data_used + meta_used, 50);
4292
4293         if (num_bytes * 3 > meta_used)
4294                 num_bytes = div64_u64(meta_used, 3);
4295
4296         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4297 }
4298
4299 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4300 {
4301         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4302         struct btrfs_space_info *sinfo = block_rsv->space_info;
4303         u64 num_bytes;
4304
4305         num_bytes = calc_global_metadata_size(fs_info);
4306
4307         spin_lock(&sinfo->lock);
4308         spin_lock(&block_rsv->lock);
4309
4310         block_rsv->size = num_bytes;
4311
4312         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4313                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4314                     sinfo->bytes_may_use;
4315
4316         if (sinfo->total_bytes > num_bytes) {
4317                 num_bytes = sinfo->total_bytes - num_bytes;
4318                 block_rsv->reserved += num_bytes;
4319                 sinfo->bytes_may_use += num_bytes;
4320                 trace_btrfs_space_reservation(fs_info, "space_info",
4321                                       sinfo->flags, num_bytes, 1);
4322         }
4323
4324         if (block_rsv->reserved >= block_rsv->size) {
4325                 num_bytes = block_rsv->reserved - block_rsv->size;
4326                 sinfo->bytes_may_use -= num_bytes;
4327                 trace_btrfs_space_reservation(fs_info, "space_info",
4328                                       sinfo->flags, num_bytes, 0);
4329                 sinfo->reservation_progress++;
4330                 block_rsv->reserved = block_rsv->size;
4331                 block_rsv->full = 1;
4332         }
4333
4334         spin_unlock(&block_rsv->lock);
4335         spin_unlock(&sinfo->lock);
4336 }
4337
4338 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4339 {
4340         struct btrfs_space_info *space_info;
4341
4342         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4343         fs_info->chunk_block_rsv.space_info = space_info;
4344
4345         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4346         fs_info->global_block_rsv.space_info = space_info;
4347         fs_info->delalloc_block_rsv.space_info = space_info;
4348         fs_info->trans_block_rsv.space_info = space_info;
4349         fs_info->empty_block_rsv.space_info = space_info;
4350         fs_info->delayed_block_rsv.space_info = space_info;
4351
4352         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4353         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4354         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4355         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4356         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4357
4358         update_global_block_rsv(fs_info);
4359 }
4360
4361 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4362 {
4363         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4364                                 (u64)-1);
4365         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4366         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4367         WARN_ON(fs_info->trans_block_rsv.size > 0);
4368         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4369         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4370         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4371         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4372         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4373 }
4374
4375 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4376                                   struct btrfs_root *root)
4377 {
4378         if (!trans->block_rsv)
4379                 return;
4380
4381         if (!trans->bytes_reserved)
4382                 return;
4383
4384         trace_btrfs_space_reservation(root->fs_info, "transaction",
4385                                       trans->transid, trans->bytes_reserved, 0);
4386         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4387         trans->bytes_reserved = 0;
4388 }
4389
4390 /* Can only return 0 or -ENOSPC */
4391 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4392                                   struct inode *inode)
4393 {
4394         struct btrfs_root *root = BTRFS_I(inode)->root;
4395         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4396         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4397
4398         /*
4399          * We need to hold space in order to delete our orphan item once we've
4400          * added it, so this takes the reservation so we can release it later
4401          * when we are truly done with the orphan item.
4402          */
4403         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4404         trace_btrfs_space_reservation(root->fs_info, "orphan",
4405                                       btrfs_ino(inode), num_bytes, 1);
4406         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4407 }
4408
4409 void btrfs_orphan_release_metadata(struct inode *inode)
4410 {
4411         struct btrfs_root *root = BTRFS_I(inode)->root;
4412         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4413         trace_btrfs_space_reservation(root->fs_info, "orphan",
4414                                       btrfs_ino(inode), num_bytes, 0);
4415         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4416 }
4417
4418 int btrfs_snap_reserve_metadata(struct btrfs_trans_handle *trans,
4419                                 struct btrfs_pending_snapshot *pending)
4420 {
4421         struct btrfs_root *root = pending->root;
4422         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4423         struct btrfs_block_rsv *dst_rsv = &pending->block_rsv;
4424         /*
4425          * two for root back/forward refs, two for directory entries,
4426          * one for root of the snapshot and one for parent inode.
4427          */
4428         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 6);
4429         dst_rsv->space_info = src_rsv->space_info;
4430         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4431 }
4432
4433 /**
4434  * drop_outstanding_extent - drop an outstanding extent
4435  * @inode: the inode we're dropping the extent for
4436  *
4437  * This is called when we are freeing up an outstanding extent, either called
4438  * after an error or after an extent is written.  This will return the number of
4439  * reserved extents that need to be freed.  This must be called with
4440  * BTRFS_I(inode)->lock held.
4441  */
4442 static unsigned drop_outstanding_extent(struct inode *inode)
4443 {
4444         unsigned drop_inode_space = 0;
4445         unsigned dropped_extents = 0;
4446
4447         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4448         BTRFS_I(inode)->outstanding_extents--;
4449
4450         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4451             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4452                                &BTRFS_I(inode)->runtime_flags))
4453                 drop_inode_space = 1;
4454
4455         /*
4456          * If we have more or the same amount of outsanding extents than we have
4457          * reserved then we need to leave the reserved extents count alone.
4458          */
4459         if (BTRFS_I(inode)->outstanding_extents >=
4460             BTRFS_I(inode)->reserved_extents)
4461                 return drop_inode_space;
4462
4463         dropped_extents = BTRFS_I(inode)->reserved_extents -
4464                 BTRFS_I(inode)->outstanding_extents;
4465         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4466         return dropped_extents + drop_inode_space;
4467 }
4468
4469 /**
4470  * calc_csum_metadata_size - return the amount of metada space that must be
4471  *      reserved/free'd for the given bytes.
4472  * @inode: the inode we're manipulating
4473  * @num_bytes: the number of bytes in question
4474  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4475  *
4476  * This adjusts the number of csum_bytes in the inode and then returns the
4477  * correct amount of metadata that must either be reserved or freed.  We
4478  * calculate how many checksums we can fit into one leaf and then divide the
4479  * number of bytes that will need to be checksumed by this value to figure out
4480  * how many checksums will be required.  If we are adding bytes then the number
4481  * may go up and we will return the number of additional bytes that must be
4482  * reserved.  If it is going down we will return the number of bytes that must
4483  * be freed.
4484  *
4485  * This must be called with BTRFS_I(inode)->lock held.
4486  */
4487 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4488                                    int reserve)
4489 {
4490         struct btrfs_root *root = BTRFS_I(inode)->root;
4491         u64 csum_size;
4492         int num_csums_per_leaf;
4493         int num_csums;
4494         int old_csums;
4495
4496         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4497             BTRFS_I(inode)->csum_bytes == 0)
4498                 return 0;
4499
4500         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4501         if (reserve)
4502                 BTRFS_I(inode)->csum_bytes += num_bytes;
4503         else
4504                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4505         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4506         num_csums_per_leaf = (int)div64_u64(csum_size,
4507                                             sizeof(struct btrfs_csum_item) +
4508                                             sizeof(struct btrfs_disk_key));
4509         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4510         num_csums = num_csums + num_csums_per_leaf - 1;
4511         num_csums = num_csums / num_csums_per_leaf;
4512
4513         old_csums = old_csums + num_csums_per_leaf - 1;
4514         old_csums = old_csums / num_csums_per_leaf;
4515
4516         /* No change, no need to reserve more */
4517         if (old_csums == num_csums)
4518                 return 0;
4519
4520         if (reserve)
4521                 return btrfs_calc_trans_metadata_size(root,
4522                                                       num_csums - old_csums);
4523
4524         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4525 }
4526
4527 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4528 {
4529         struct btrfs_root *root = BTRFS_I(inode)->root;
4530         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4531         u64 to_reserve = 0;
4532         u64 csum_bytes;
4533         unsigned nr_extents = 0;
4534         int extra_reserve = 0;
4535         int flush = 1;
4536         int ret;
4537
4538         /* Need to be holding the i_mutex here if we aren't free space cache */
4539         if (btrfs_is_free_space_inode(inode))
4540                 flush = 0;
4541
4542         if (flush && btrfs_transaction_in_commit(root->fs_info))
4543                 schedule_timeout(1);
4544
4545         mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4546         num_bytes = ALIGN(num_bytes, root->sectorsize);
4547
4548         spin_lock(&BTRFS_I(inode)->lock);
4549         BTRFS_I(inode)->outstanding_extents++;
4550
4551         if (BTRFS_I(inode)->outstanding_extents >
4552             BTRFS_I(inode)->reserved_extents)
4553                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4554                         BTRFS_I(inode)->reserved_extents;
4555
4556         /*
4557          * Add an item to reserve for updating the inode when we complete the
4558          * delalloc io.
4559          */
4560         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4561                       &BTRFS_I(inode)->runtime_flags)) {
4562                 nr_extents++;
4563                 extra_reserve = 1;
4564         }
4565
4566         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4567         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4568         csum_bytes = BTRFS_I(inode)->csum_bytes;
4569         spin_unlock(&BTRFS_I(inode)->lock);
4570
4571         if (root->fs_info->quota_enabled) {
4572                 ret = btrfs_qgroup_reserve(root, num_bytes +
4573                                            nr_extents * root->leafsize);
4574                 if (ret) {
4575                         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4576                         return ret;
4577                 }
4578         }
4579
4580         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4581         if (ret) {
4582                 u64 to_free = 0;
4583                 unsigned dropped;
4584
4585                 spin_lock(&BTRFS_I(inode)->lock);
4586                 dropped = drop_outstanding_extent(inode);
4587                 /*
4588                  * If the inodes csum_bytes is the same as the original
4589                  * csum_bytes then we know we haven't raced with any free()ers
4590                  * so we can just reduce our inodes csum bytes and carry on.
4591                  * Otherwise we have to do the normal free thing to account for
4592                  * the case that the free side didn't free up its reserve
4593                  * because of this outstanding reservation.
4594                  */
4595                 if (BTRFS_I(inode)->csum_bytes == csum_bytes)
4596                         calc_csum_metadata_size(inode, num_bytes, 0);
4597                 else
4598                         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4599                 spin_unlock(&BTRFS_I(inode)->lock);
4600                 if (dropped)
4601                         to_free += btrfs_calc_trans_metadata_size(root, dropped);
4602
4603                 if (to_free) {
4604                         btrfs_block_rsv_release(root, block_rsv, to_free);
4605                         trace_btrfs_space_reservation(root->fs_info,
4606                                                       "delalloc",
4607                                                       btrfs_ino(inode),
4608                                                       to_free, 0);
4609                 }
4610                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4611                 return ret;
4612         }
4613
4614         spin_lock(&BTRFS_I(inode)->lock);
4615         if (extra_reserve) {
4616                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4617                         &BTRFS_I(inode)->runtime_flags);
4618                 nr_extents--;
4619         }
4620         BTRFS_I(inode)->reserved_extents += nr_extents;
4621         spin_unlock(&BTRFS_I(inode)->lock);
4622         mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4623
4624         if (to_reserve)
4625                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4626                                               btrfs_ino(inode), to_reserve, 1);
4627         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4628
4629         return 0;
4630 }
4631
4632 /**
4633  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4634  * @inode: the inode to release the reservation for
4635  * @num_bytes: the number of bytes we're releasing
4636  *
4637  * This will release the metadata reservation for an inode.  This can be called
4638  * once we complete IO for a given set of bytes to release their metadata
4639  * reservations.
4640  */
4641 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4642 {
4643         struct btrfs_root *root = BTRFS_I(inode)->root;
4644         u64 to_free = 0;
4645         unsigned dropped;
4646
4647         num_bytes = ALIGN(num_bytes, root->sectorsize);
4648         spin_lock(&BTRFS_I(inode)->lock);
4649         dropped = drop_outstanding_extent(inode);
4650
4651         to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4652         spin_unlock(&BTRFS_I(inode)->lock);
4653         if (dropped > 0)
4654                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4655
4656         trace_btrfs_space_reservation(root->fs_info, "delalloc",
4657                                       btrfs_ino(inode), to_free, 0);
4658         if (root->fs_info->quota_enabled) {
4659                 btrfs_qgroup_free(root, num_bytes +
4660                                         dropped * root->leafsize);
4661         }
4662
4663         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
4664                                 to_free);
4665 }
4666
4667 /**
4668  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
4669  * @inode: inode we're writing to
4670  * @num_bytes: the number of bytes we want to allocate
4671  *
4672  * This will do the following things
4673  *
4674  * o reserve space in the data space info for num_bytes
4675  * o reserve space in the metadata space info based on number of outstanding
4676  *   extents and how much csums will be needed
4677  * o add to the inodes ->delalloc_bytes
4678  * o add it to the fs_info's delalloc inodes list.
4679  *
4680  * This will return 0 for success and -ENOSPC if there is no space left.
4681  */
4682 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
4683 {
4684         int ret;
4685
4686         ret = btrfs_check_data_free_space(inode, num_bytes);
4687         if (ret)
4688                 return ret;
4689
4690         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
4691         if (ret) {
4692                 btrfs_free_reserved_data_space(inode, num_bytes);
4693                 return ret;
4694         }
4695
4696         return 0;
4697 }
4698
4699 /**
4700  * btrfs_delalloc_release_space - release data and metadata space for delalloc
4701  * @inode: inode we're releasing space for
4702  * @num_bytes: the number of bytes we want to free up
4703  *
4704  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
4705  * called in the case that we don't need the metadata AND data reservations
4706  * anymore.  So if there is an error or we insert an inline extent.
4707  *
4708  * This function will release the metadata space that was not used and will
4709  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
4710  * list if there are no delalloc bytes left.
4711  */
4712 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
4713 {
4714         btrfs_delalloc_release_metadata(inode, num_bytes);
4715         btrfs_free_reserved_data_space(inode, num_bytes);
4716 }
4717
4718 static int update_block_group(struct btrfs_trans_handle *trans,
4719                               struct btrfs_root *root,
4720                               u64 bytenr, u64 num_bytes, int alloc)
4721 {
4722         struct btrfs_block_group_cache *cache = NULL;
4723         struct btrfs_fs_info *info = root->fs_info;
4724         u64 total = num_bytes;
4725         u64 old_val;
4726         u64 byte_in_group;
4727         int factor;
4728
4729         /* block accounting for super block */
4730         spin_lock(&info->delalloc_lock);
4731         old_val = btrfs_super_bytes_used(info->super_copy);
4732         if (alloc)
4733                 old_val += num_bytes;
4734         else
4735                 old_val -= num_bytes;
4736         btrfs_set_super_bytes_used(info->super_copy, old_val);
4737         spin_unlock(&info->delalloc_lock);
4738
4739         while (total) {
4740                 cache = btrfs_lookup_block_group(info, bytenr);
4741                 if (!cache)
4742                         return -ENOENT;
4743                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
4744                                     BTRFS_BLOCK_GROUP_RAID1 |
4745                                     BTRFS_BLOCK_GROUP_RAID10))
4746                         factor = 2;
4747                 else
4748                         factor = 1;
4749                 /*
4750                  * If this block group has free space cache written out, we
4751                  * need to make sure to load it if we are removing space.  This
4752                  * is because we need the unpinning stage to actually add the
4753                  * space back to the block group, otherwise we will leak space.
4754                  */
4755                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
4756                         cache_block_group(cache, trans, NULL, 1);
4757
4758                 byte_in_group = bytenr - cache->key.objectid;
4759                 WARN_ON(byte_in_group > cache->key.offset);
4760
4761                 spin_lock(&cache->space_info->lock);
4762                 spin_lock(&cache->lock);
4763
4764                 if (btrfs_test_opt(root, SPACE_CACHE) &&
4765                     cache->disk_cache_state < BTRFS_DC_CLEAR)
4766                         cache->disk_cache_state = BTRFS_DC_CLEAR;
4767
4768                 cache->dirty = 1;
4769                 old_val = btrfs_block_group_used(&cache->item);
4770                 num_bytes = min(total, cache->key.offset - byte_in_group);
4771                 if (alloc) {
4772                         old_val += num_bytes;
4773                         btrfs_set_block_group_used(&cache->item, old_val);
4774                         cache->reserved -= num_bytes;
4775                         cache->space_info->bytes_reserved -= num_bytes;
4776                         cache->space_info->bytes_used += num_bytes;
4777                         cache->space_info->disk_used += num_bytes * factor;
4778                         spin_unlock(&cache->lock);
4779                         spin_unlock(&cache->space_info->lock);
4780                 } else {
4781                         old_val -= num_bytes;
4782                         btrfs_set_block_group_used(&cache->item, old_val);
4783                         cache->pinned += num_bytes;
4784                         cache->space_info->bytes_pinned += num_bytes;
4785                         cache->space_info->bytes_used -= num_bytes;
4786                         cache->space_info->disk_used -= num_bytes * factor;
4787                         spin_unlock(&cache->lock);
4788                         spin_unlock(&cache->space_info->lock);
4789
4790                         set_extent_dirty(info->pinned_extents,
4791                                          bytenr, bytenr + num_bytes - 1,
4792                                          GFP_NOFS | __GFP_NOFAIL);
4793                 }
4794                 btrfs_put_block_group(cache);
4795                 total -= num_bytes;
4796                 bytenr += num_bytes;
4797         }
4798         return 0;
4799 }
4800
4801 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
4802 {
4803         struct btrfs_block_group_cache *cache;
4804         u64 bytenr;
4805
4806         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
4807         if (!cache)
4808                 return 0;
4809
4810         bytenr = cache->key.objectid;
4811         btrfs_put_block_group(cache);
4812
4813         return bytenr;
4814 }
4815
4816 static int pin_down_extent(struct btrfs_root *root,
4817                            struct btrfs_block_group_cache *cache,
4818                            u64 bytenr, u64 num_bytes, int reserved)
4819 {
4820         spin_lock(&cache->space_info->lock);
4821         spin_lock(&cache->lock);
4822         cache->pinned += num_bytes;
4823         cache->space_info->bytes_pinned += num_bytes;
4824         if (reserved) {
4825                 cache->reserved -= num_bytes;
4826                 cache->space_info->bytes_reserved -= num_bytes;
4827         }
4828         spin_unlock(&cache->lock);
4829         spin_unlock(&cache->space_info->lock);
4830
4831         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
4832                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
4833         return 0;
4834 }
4835
4836 /*
4837  * this function must be called within transaction
4838  */
4839 int btrfs_pin_extent(struct btrfs_root *root,
4840                      u64 bytenr, u64 num_bytes, int reserved)
4841 {
4842         struct btrfs_block_group_cache *cache;
4843
4844         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4845         BUG_ON(!cache); /* Logic error */
4846
4847         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
4848
4849         btrfs_put_block_group(cache);
4850         return 0;
4851 }
4852
4853 /*
4854  * this function must be called within transaction
4855  */
4856 int btrfs_pin_extent_for_log_replay(struct btrfs_trans_handle *trans,
4857                                     struct btrfs_root *root,
4858                                     u64 bytenr, u64 num_bytes)
4859 {
4860         struct btrfs_block_group_cache *cache;
4861
4862         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
4863         BUG_ON(!cache); /* Logic error */
4864
4865         /*
4866          * pull in the free space cache (if any) so that our pin
4867          * removes the free space from the cache.  We have load_only set
4868          * to one because the slow code to read in the free extents does check
4869          * the pinned extents.
4870          */
4871         cache_block_group(cache, trans, root, 1);
4872
4873         pin_down_extent(root, cache, bytenr, num_bytes, 0);
4874
4875         /* remove us from the free space cache (if we're there at all) */
4876         btrfs_remove_free_space(cache, bytenr, num_bytes);
4877         btrfs_put_block_group(cache);
4878         return 0;
4879 }
4880
4881 /**
4882  * btrfs_update_reserved_bytes - update the block_group and space info counters
4883  * @cache:      The cache we are manipulating
4884  * @num_bytes:  The number of bytes in question
4885  * @reserve:    One of the reservation enums
4886  *
4887  * This is called by the allocator when it reserves space, or by somebody who is
4888  * freeing space that was never actually used on disk.  For example if you
4889  * reserve some space for a new leaf in transaction A and before transaction A
4890  * commits you free that leaf, you call this with reserve set to 0 in order to
4891  * clear the reservation.
4892  *
4893  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
4894  * ENOSPC accounting.  For data we handle the reservation through clearing the
4895  * delalloc bits in the io_tree.  We have to do this since we could end up
4896  * allocating less disk space for the amount of data we have reserved in the
4897  * case of compression.
4898  *
4899  * If this is a reservation and the block group has become read only we cannot
4900  * make the reservation and return -EAGAIN, otherwise this function always
4901  * succeeds.
4902  */
4903 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
4904                                        u64 num_bytes, int reserve)
4905 {
4906         struct btrfs_space_info *space_info = cache->space_info;
4907         int ret = 0;
4908
4909         spin_lock(&space_info->lock);
4910         spin_lock(&cache->lock);
4911         if (reserve != RESERVE_FREE) {
4912                 if (cache->ro) {
4913                         ret = -EAGAIN;
4914                 } else {
4915                         cache->reserved += num_bytes;
4916                         space_info->bytes_reserved += num_bytes;
4917                         if (reserve == RESERVE_ALLOC) {
4918                                 trace_btrfs_space_reservation(cache->fs_info,
4919                                                 "space_info", space_info->flags,
4920                                                 num_bytes, 0);
4921                                 space_info->bytes_may_use -= num_bytes;
4922                         }
4923                 }
4924         } else {
4925                 if (cache->ro)
4926                         space_info->bytes_readonly += num_bytes;
4927                 cache->reserved -= num_bytes;
4928                 space_info->bytes_reserved -= num_bytes;
4929                 space_info->reservation_progress++;
4930         }
4931         spin_unlock(&cache->lock);
4932         spin_unlock(&space_info->lock);
4933         return ret;
4934 }
4935
4936 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
4937                                 struct btrfs_root *root)
4938 {
4939         struct btrfs_fs_info *fs_info = root->fs_info;
4940         struct btrfs_caching_control *next;
4941         struct btrfs_caching_control *caching_ctl;
4942         struct btrfs_block_group_cache *cache;
4943
4944         down_write(&fs_info->extent_commit_sem);
4945
4946         list_for_each_entry_safe(caching_ctl, next,
4947                                  &fs_info->caching_block_groups, list) {
4948                 cache = caching_ctl->block_group;
4949                 if (block_group_cache_done(cache)) {
4950                         cache->last_byte_to_unpin = (u64)-1;
4951                         list_del_init(&caching_ctl->list);
4952                         put_caching_control(caching_ctl);
4953                 } else {
4954                         cache->last_byte_to_unpin = caching_ctl->progress;
4955                 }
4956         }
4957
4958         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
4959                 fs_info->pinned_extents = &fs_info->freed_extents[1];
4960         else
4961                 fs_info->pinned_extents = &fs_info->freed_extents[0];
4962
4963         up_write(&fs_info->extent_commit_sem);
4964
4965         update_global_block_rsv(fs_info);
4966 }
4967
4968 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
4969 {
4970         struct btrfs_fs_info *fs_info = root->fs_info;
4971         struct btrfs_block_group_cache *cache = NULL;
4972         u64 len;
4973
4974         while (start <= end) {
4975                 if (!cache ||
4976                     start >= cache->key.objectid + cache->key.offset) {
4977                         if (cache)
4978                                 btrfs_put_block_group(cache);
4979                         cache = btrfs_lookup_block_group(fs_info, start);
4980                         BUG_ON(!cache); /* Logic error */
4981                 }
4982
4983                 len = cache->key.objectid + cache->key.offset - start;
4984                 len = min(len, end + 1 - start);
4985
4986                 if (start < cache->last_byte_to_unpin) {
4987                         len = min(len, cache->last_byte_to_unpin - start);
4988                         btrfs_add_free_space(cache, start, len);
4989                 }
4990
4991                 start += len;
4992
4993                 spin_lock(&cache->space_info->lock);
4994                 spin_lock(&cache->lock);
4995                 cache->pinned -= len;
4996                 cache->space_info->bytes_pinned -= len;
4997                 if (cache->ro)
4998                         cache->space_info->bytes_readonly += len;
4999                 spin_unlock(&cache->lock);
5000                 spin_unlock(&cache->space_info->lock);
5001         }
5002
5003         if (cache)
5004                 btrfs_put_block_group(cache);
5005         return 0;
5006 }
5007
5008 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5009                                struct btrfs_root *root)
5010 {
5011         struct btrfs_fs_info *fs_info = root->fs_info;
5012         struct extent_io_tree *unpin;
5013         u64 start;
5014         u64 end;
5015         int ret;
5016
5017         if (trans->aborted)
5018                 return 0;
5019
5020         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5021                 unpin = &fs_info->freed_extents[1];
5022         else
5023                 unpin = &fs_info->freed_extents[0];
5024
5025         while (1) {
5026                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5027                                             EXTENT_DIRTY, NULL);
5028                 if (ret)
5029                         break;
5030
5031                 if (btrfs_test_opt(root, DISCARD))
5032                         ret = btrfs_discard_extent(root, start,
5033                                                    end + 1 - start, NULL);
5034
5035                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5036                 unpin_extent_range(root, start, end);
5037                 cond_resched();
5038         }
5039
5040         return 0;
5041 }
5042
5043 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5044                                 struct btrfs_root *root,
5045                                 u64 bytenr, u64 num_bytes, u64 parent,
5046                                 u64 root_objectid, u64 owner_objectid,
5047                                 u64 owner_offset, int refs_to_drop,
5048                                 struct btrfs_delayed_extent_op *extent_op)
5049 {
5050         struct btrfs_key key;
5051         struct btrfs_path *path;
5052         struct btrfs_fs_info *info = root->fs_info;
5053         struct btrfs_root *extent_root = info->extent_root;
5054         struct extent_buffer *leaf;
5055         struct btrfs_extent_item *ei;
5056         struct btrfs_extent_inline_ref *iref;
5057         int ret;
5058         int is_data;
5059         int extent_slot = 0;
5060         int found_extent = 0;
5061         int num_to_del = 1;
5062         u32 item_size;
5063         u64 refs;
5064
5065         path = btrfs_alloc_path();
5066         if (!path)
5067                 return -ENOMEM;
5068
5069         path->reada = 1;
5070         path->leave_spinning = 1;
5071
5072         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5073         BUG_ON(!is_data && refs_to_drop != 1);
5074
5075         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5076                                     bytenr, num_bytes, parent,
5077                                     root_objectid, owner_objectid,
5078                                     owner_offset);
5079         if (ret == 0) {
5080                 extent_slot = path->slots[0];
5081                 while (extent_slot >= 0) {
5082                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5083                                               extent_slot);
5084                         if (key.objectid != bytenr)
5085                                 break;
5086                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5087                             key.offset == num_bytes) {
5088                                 found_extent = 1;
5089                                 break;
5090                         }
5091                         if (path->slots[0] - extent_slot > 5)
5092                                 break;
5093                         extent_slot--;
5094                 }
5095 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5096                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5097                 if (found_extent && item_size < sizeof(*ei))
5098                         found_extent = 0;
5099 #endif
5100                 if (!found_extent) {
5101                         BUG_ON(iref);
5102                         ret = remove_extent_backref(trans, extent_root, path,
5103                                                     NULL, refs_to_drop,
5104                                                     is_data);
5105                         if (ret) {
5106                                 btrfs_abort_transaction(trans, extent_root, ret);
5107                                 goto out;
5108                         }
5109                         btrfs_release_path(path);
5110                         path->leave_spinning = 1;
5111
5112                         key.objectid = bytenr;
5113                         key.type = BTRFS_EXTENT_ITEM_KEY;
5114                         key.offset = num_bytes;
5115
5116                         ret = btrfs_search_slot(trans, extent_root,
5117                                                 &key, path, -1, 1);
5118                         if (ret) {
5119                                 printk(KERN_ERR "umm, got %d back from search"
5120                                        ", was looking for %llu\n", ret,
5121                                        (unsigned long long)bytenr);
5122                                 if (ret > 0)
5123                                         btrfs_print_leaf(extent_root,
5124                                                          path->nodes[0]);
5125                         }
5126                         if (ret < 0) {
5127                                 btrfs_abort_transaction(trans, extent_root, ret);
5128                                 goto out;
5129                         }
5130                         extent_slot = path->slots[0];
5131                 }
5132         } else if (ret == -ENOENT) {
5133                 btrfs_print_leaf(extent_root, path->nodes[0]);
5134                 WARN_ON(1);
5135                 printk(KERN_ERR "btrfs unable to find ref byte nr %llu "
5136                        "parent %llu root %llu  owner %llu offset %llu\n",
5137                        (unsigned long long)bytenr,
5138                        (unsigned long long)parent,
5139                        (unsigned long long)root_objectid,
5140                        (unsigned long long)owner_objectid,
5141                        (unsigned long long)owner_offset);
5142         } else {
5143                 btrfs_abort_transaction(trans, extent_root, ret);
5144                 goto out;
5145         }
5146
5147         leaf = path->nodes[0];
5148         item_size = btrfs_item_size_nr(leaf, extent_slot);
5149 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5150         if (item_size < sizeof(*ei)) {
5151                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5152                 ret = convert_extent_item_v0(trans, extent_root, path,
5153                                              owner_objectid, 0);
5154                 if (ret < 0) {
5155                         btrfs_abort_transaction(trans, extent_root, ret);
5156                         goto out;
5157                 }
5158
5159                 btrfs_release_path(path);
5160                 path->leave_spinning = 1;
5161
5162                 key.objectid = bytenr;
5163                 key.type = BTRFS_EXTENT_ITEM_KEY;
5164                 key.offset = num_bytes;
5165
5166                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5167                                         -1, 1);
5168                 if (ret) {
5169                         printk(KERN_ERR "umm, got %d back from search"
5170                                ", was looking for %llu\n", ret,
5171                                (unsigned long long)bytenr);
5172                         btrfs_print_leaf(extent_root, path->nodes[0]);
5173                 }
5174                 if (ret < 0) {
5175                         btrfs_abort_transaction(trans, extent_root, ret);
5176                         goto out;
5177                 }
5178
5179                 extent_slot = path->slots[0];
5180                 leaf = path->nodes[0];
5181                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5182         }
5183 #endif
5184         BUG_ON(item_size < sizeof(*ei));
5185         ei = btrfs_item_ptr(leaf, extent_slot,
5186                             struct btrfs_extent_item);
5187         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID) {
5188                 struct btrfs_tree_block_info *bi;
5189                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5190                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5191                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5192         }
5193
5194         refs = btrfs_extent_refs(leaf, ei);
5195         BUG_ON(refs < refs_to_drop);
5196         refs -= refs_to_drop;
5197
5198         if (refs > 0) {
5199                 if (extent_op)
5200                         __run_delayed_extent_op(extent_op, leaf, ei);
5201                 /*
5202                  * In the case of inline back ref, reference count will
5203                  * be updated by remove_extent_backref
5204                  */
5205                 if (iref) {
5206                         BUG_ON(!found_extent);
5207                 } else {
5208                         btrfs_set_extent_refs(leaf, ei, refs);
5209                         btrfs_mark_buffer_dirty(leaf);
5210                 }
5211                 if (found_extent) {
5212                         ret = remove_extent_backref(trans, extent_root, path,
5213                                                     iref, refs_to_drop,
5214                                                     is_data);
5215                         if (ret) {
5216                                 btrfs_abort_transaction(trans, extent_root, ret);
5217                                 goto out;
5218                         }
5219                 }
5220         } else {
5221                 if (found_extent) {
5222                         BUG_ON(is_data && refs_to_drop !=
5223                                extent_data_ref_count(root, path, iref));
5224                         if (iref) {
5225                                 BUG_ON(path->slots[0] != extent_slot);
5226                         } else {
5227                                 BUG_ON(path->slots[0] != extent_slot + 1);
5228                                 path->slots[0] = extent_slot;
5229                                 num_to_del = 2;
5230                         }
5231                 }
5232
5233                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5234                                       num_to_del);
5235                 if (ret) {
5236                         btrfs_abort_transaction(trans, extent_root, ret);
5237                         goto out;
5238                 }
5239                 btrfs_release_path(path);
5240
5241                 if (is_data) {
5242                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5243                         if (ret) {
5244                                 btrfs_abort_transaction(trans, extent_root, ret);
5245                                 goto out;
5246                         }
5247                 }
5248
5249                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
5250                 if (ret) {
5251                         btrfs_abort_transaction(trans, extent_root, ret);
5252                         goto out;
5253                 }
5254         }
5255 out:
5256         btrfs_free_path(path);
5257         return ret;
5258 }
5259
5260 /*
5261  * when we free an block, it is possible (and likely) that we free the last
5262  * delayed ref for that extent as well.  This searches the delayed ref tree for
5263  * a given extent, and if there are no other delayed refs to be processed, it
5264  * removes it from the tree.
5265  */
5266 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5267                                       struct btrfs_root *root, u64 bytenr)
5268 {
5269         struct btrfs_delayed_ref_head *head;
5270         struct btrfs_delayed_ref_root *delayed_refs;
5271         struct btrfs_delayed_ref_node *ref;
5272         struct rb_node *node;
5273         int ret = 0;
5274
5275         delayed_refs = &trans->transaction->delayed_refs;
5276         spin_lock(&delayed_refs->lock);
5277         head = btrfs_find_delayed_ref_head(trans, bytenr);
5278         if (!head)
5279                 goto out;
5280
5281         node = rb_prev(&head->node.rb_node);
5282         if (!node)
5283                 goto out;
5284
5285         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5286
5287         /* there are still entries for this ref, we can't drop it */
5288         if (ref->bytenr == bytenr)
5289                 goto out;
5290
5291         if (head->extent_op) {
5292                 if (!head->must_insert_reserved)
5293                         goto out;
5294                 kfree(head->extent_op);
5295                 head->extent_op = NULL;
5296         }
5297
5298         /*
5299          * waiting for the lock here would deadlock.  If someone else has it
5300          * locked they are already in the process of dropping it anyway
5301          */
5302         if (!mutex_trylock(&head->mutex))
5303                 goto out;
5304
5305         /*
5306          * at this point we have a head with no other entries.  Go
5307          * ahead and process it.
5308          */
5309         head->node.in_tree = 0;
5310         rb_erase(&head->node.rb_node, &delayed_refs->root);
5311
5312         delayed_refs->num_entries--;
5313
5314         /*
5315          * we don't take a ref on the node because we're removing it from the
5316          * tree, so we just steal the ref the tree was holding.
5317          */
5318         delayed_refs->num_heads--;
5319         if (list_empty(&head->cluster))
5320                 delayed_refs->num_heads_ready--;
5321
5322         list_del_init(&head->cluster);
5323         spin_unlock(&delayed_refs->lock);
5324
5325         BUG_ON(head->extent_op);
5326         if (head->must_insert_reserved)
5327                 ret = 1;
5328
5329         mutex_unlock(&head->mutex);
5330         btrfs_put_delayed_ref(&head->node);
5331         return ret;
5332 out:
5333         spin_unlock(&delayed_refs->lock);
5334         return 0;
5335 }
5336
5337 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5338                            struct btrfs_root *root,
5339                            struct extent_buffer *buf,
5340                            u64 parent, int last_ref)
5341 {
5342         struct btrfs_block_group_cache *cache = NULL;
5343         int ret;
5344
5345         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5346                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5347                                         buf->start, buf->len,
5348                                         parent, root->root_key.objectid,
5349                                         btrfs_header_level(buf),
5350                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5351                 BUG_ON(ret); /* -ENOMEM */
5352         }
5353
5354         if (!last_ref)
5355                 return;
5356
5357         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5358
5359         if (btrfs_header_generation(buf) == trans->transid) {
5360                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5361                         ret = check_ref_cleanup(trans, root, buf->start);
5362                         if (!ret)
5363                                 goto out;
5364                 }
5365
5366                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5367                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5368                         goto out;
5369                 }
5370
5371                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5372
5373                 btrfs_add_free_space(cache, buf->start, buf->len);
5374                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5375         }
5376 out:
5377         /*
5378          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5379          * anymore.
5380          */
5381         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5382         btrfs_put_block_group(cache);
5383 }
5384
5385 /* Can return -ENOMEM */
5386 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5387                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5388                       u64 owner, u64 offset, int for_cow)
5389 {
5390         int ret;
5391         struct btrfs_fs_info *fs_info = root->fs_info;
5392
5393         /*
5394          * tree log blocks never actually go into the extent allocation
5395          * tree, just update pinning info and exit early.
5396          */
5397         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5398                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5399                 /* unlocks the pinned mutex */
5400                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5401                 ret = 0;
5402         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5403                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5404                                         num_bytes,
5405                                         parent, root_objectid, (int)owner,
5406                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5407         } else {
5408                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5409                                                 num_bytes,
5410                                                 parent, root_objectid, owner,
5411                                                 offset, BTRFS_DROP_DELAYED_REF,
5412                                                 NULL, for_cow);
5413         }
5414         return ret;
5415 }
5416
5417 static u64 stripe_align(struct btrfs_root *root, u64 val)
5418 {
5419         u64 mask = ((u64)root->stripesize - 1);
5420         u64 ret = (val + mask) & ~mask;
5421         return ret;
5422 }
5423
5424 /*
5425  * when we wait for progress in the block group caching, its because
5426  * our allocation attempt failed at least once.  So, we must sleep
5427  * and let some progress happen before we try again.
5428  *
5429  * This function will sleep at least once waiting for new free space to
5430  * show up, and then it will check the block group free space numbers
5431  * for our min num_bytes.  Another option is to have it go ahead
5432  * and look in the rbtree for a free extent of a given size, but this
5433  * is a good start.
5434  */
5435 static noinline int
5436 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5437                                 u64 num_bytes)
5438 {
5439         struct btrfs_caching_control *caching_ctl;
5440         DEFINE_WAIT(wait);
5441
5442         caching_ctl = get_caching_control(cache);
5443         if (!caching_ctl)
5444                 return 0;
5445
5446         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5447                    (cache->free_space_ctl->free_space >= num_bytes));
5448
5449         put_caching_control(caching_ctl);
5450         return 0;
5451 }
5452
5453 static noinline int
5454 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5455 {
5456         struct btrfs_caching_control *caching_ctl;
5457         DEFINE_WAIT(wait);
5458
5459         caching_ctl = get_caching_control(cache);
5460         if (!caching_ctl)
5461                 return 0;
5462
5463         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5464
5465         put_caching_control(caching_ctl);
5466         return 0;
5467 }
5468
5469 static int __get_block_group_index(u64 flags)
5470 {
5471         int index;
5472
5473         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5474                 index = 0;
5475         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5476                 index = 1;
5477         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5478                 index = 2;
5479         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5480                 index = 3;
5481         else
5482                 index = 4;
5483
5484         return index;
5485 }
5486
5487 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5488 {
5489         return __get_block_group_index(cache->flags);
5490 }
5491
5492 enum btrfs_loop_type {
5493         LOOP_CACHING_NOWAIT = 0,
5494         LOOP_CACHING_WAIT = 1,
5495         LOOP_ALLOC_CHUNK = 2,
5496         LOOP_NO_EMPTY_SIZE = 3,
5497 };
5498
5499 /*
5500  * walks the btree of allocated extents and find a hole of a given size.
5501  * The key ins is changed to record the hole:
5502  * ins->objectid == block start
5503  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5504  * ins->offset == number of blocks
5505  * Any available blocks before search_start are skipped.
5506  */
5507 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5508                                      struct btrfs_root *orig_root,
5509                                      u64 num_bytes, u64 empty_size,
5510                                      u64 hint_byte, struct btrfs_key *ins,
5511                                      u64 data)
5512 {
5513         int ret = 0;
5514         struct btrfs_root *root = orig_root->fs_info->extent_root;
5515         struct btrfs_free_cluster *last_ptr = NULL;
5516         struct btrfs_block_group_cache *block_group = NULL;
5517         struct btrfs_block_group_cache *used_block_group;
5518         u64 search_start = 0;
5519         int empty_cluster = 2 * 1024 * 1024;
5520         struct btrfs_space_info *space_info;
5521         int loop = 0;
5522         int index = 0;
5523         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5524                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5525         bool found_uncached_bg = false;
5526         bool failed_cluster_refill = false;
5527         bool failed_alloc = false;
5528         bool use_cluster = true;
5529         bool have_caching_bg = false;
5530
5531         WARN_ON(num_bytes < root->sectorsize);
5532         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5533         ins->objectid = 0;
5534         ins->offset = 0;
5535
5536         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5537
5538         space_info = __find_space_info(root->fs_info, data);
5539         if (!space_info) {
5540                 printk(KERN_ERR "No space info for %llu\n", data);
5541                 return -ENOSPC;
5542         }
5543
5544         /*
5545          * If the space info is for both data and metadata it means we have a
5546          * small filesystem and we can't use the clustering stuff.
5547          */
5548         if (btrfs_mixed_space_info(space_info))
5549                 use_cluster = false;
5550
5551         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5552                 last_ptr = &root->fs_info->meta_alloc_cluster;
5553                 if (!btrfs_test_opt(root, SSD))
5554                         empty_cluster = 64 * 1024;
5555         }
5556
5557         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5558             btrfs_test_opt(root, SSD)) {
5559                 last_ptr = &root->fs_info->data_alloc_cluster;
5560         }
5561
5562         if (last_ptr) {
5563                 spin_lock(&last_ptr->lock);
5564                 if (last_ptr->block_group)
5565                         hint_byte = last_ptr->window_start;
5566                 spin_unlock(&last_ptr->lock);
5567         }
5568
5569         search_start = max(search_start, first_logical_byte(root, 0));
5570         search_start = max(search_start, hint_byte);
5571
5572         if (!last_ptr)
5573                 empty_cluster = 0;
5574
5575         if (search_start == hint_byte) {
5576                 block_group = btrfs_lookup_block_group(root->fs_info,
5577                                                        search_start);
5578                 used_block_group = block_group;
5579                 /*
5580                  * we don't want to use the block group if it doesn't match our
5581                  * allocation bits, or if its not cached.
5582                  *
5583                  * However if we are re-searching with an ideal block group
5584                  * picked out then we don't care that the block group is cached.
5585                  */
5586                 if (block_group && block_group_bits(block_group, data) &&
5587                     block_group->cached != BTRFS_CACHE_NO) {
5588                         down_read(&space_info->groups_sem);
5589                         if (list_empty(&block_group->list) ||
5590                             block_group->ro) {
5591                                 /*
5592                                  * someone is removing this block group,
5593                                  * we can't jump into the have_block_group
5594                                  * target because our list pointers are not
5595                                  * valid
5596                                  */
5597                                 btrfs_put_block_group(block_group);
5598                                 up_read(&space_info->groups_sem);
5599                         } else {
5600                                 index = get_block_group_index(block_group);
5601                                 goto have_block_group;
5602                         }
5603                 } else if (block_group) {
5604                         btrfs_put_block_group(block_group);
5605                 }
5606         }
5607 search:
5608         have_caching_bg = false;
5609         down_read(&space_info->groups_sem);
5610         list_for_each_entry(block_group, &space_info->block_groups[index],
5611                             list) {
5612                 u64 offset;
5613                 int cached;
5614
5615                 used_block_group = block_group;
5616                 btrfs_get_block_group(block_group);
5617                 search_start = block_group->key.objectid;
5618
5619                 /*
5620                  * this can happen if we end up cycling through all the
5621                  * raid types, but we want to make sure we only allocate
5622                  * for the proper type.
5623                  */
5624                 if (!block_group_bits(block_group, data)) {
5625                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
5626                                 BTRFS_BLOCK_GROUP_RAID1 |
5627                                 BTRFS_BLOCK_GROUP_RAID10;
5628
5629                         /*
5630                          * if they asked for extra copies and this block group
5631                          * doesn't provide them, bail.  This does allow us to
5632                          * fill raid0 from raid1.
5633                          */
5634                         if ((data & extra) && !(block_group->flags & extra))
5635                                 goto loop;
5636                 }
5637
5638 have_block_group:
5639                 cached = block_group_cache_done(block_group);
5640                 if (unlikely(!cached)) {
5641                         found_uncached_bg = true;
5642                         ret = cache_block_group(block_group, trans,
5643                                                 orig_root, 0);
5644                         BUG_ON(ret < 0);
5645                         ret = 0;
5646                 }
5647
5648                 if (unlikely(block_group->ro))
5649                         goto loop;
5650
5651                 /*
5652                  * Ok we want to try and use the cluster allocator, so
5653                  * lets look there
5654                  */
5655                 if (last_ptr) {
5656                         /*
5657                          * the refill lock keeps out other
5658                          * people trying to start a new cluster
5659                          */
5660                         spin_lock(&last_ptr->refill_lock);
5661                         used_block_group = last_ptr->block_group;
5662                         if (used_block_group != block_group &&
5663                             (!used_block_group ||
5664                              used_block_group->ro ||
5665                              !block_group_bits(used_block_group, data))) {
5666                                 used_block_group = block_group;
5667                                 goto refill_cluster;
5668                         }
5669
5670                         if (used_block_group != block_group)
5671                                 btrfs_get_block_group(used_block_group);
5672
5673                         offset = btrfs_alloc_from_cluster(used_block_group,
5674                           last_ptr, num_bytes, used_block_group->key.objectid);
5675                         if (offset) {
5676                                 /* we have a block, we're done */
5677                                 spin_unlock(&last_ptr->refill_lock);
5678                                 trace_btrfs_reserve_extent_cluster(root,
5679                                         block_group, search_start, num_bytes);
5680                                 goto checks;
5681                         }
5682
5683                         WARN_ON(last_ptr->block_group != used_block_group);
5684                         if (used_block_group != block_group) {
5685                                 btrfs_put_block_group(used_block_group);
5686                                 used_block_group = block_group;
5687                         }
5688 refill_cluster:
5689                         BUG_ON(used_block_group != block_group);
5690                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
5691                          * set up a new clusters, so lets just skip it
5692                          * and let the allocator find whatever block
5693                          * it can find.  If we reach this point, we
5694                          * will have tried the cluster allocator
5695                          * plenty of times and not have found
5696                          * anything, so we are likely way too
5697                          * fragmented for the clustering stuff to find
5698                          * anything.
5699                          *
5700                          * However, if the cluster is taken from the
5701                          * current block group, release the cluster
5702                          * first, so that we stand a better chance of
5703                          * succeeding in the unclustered
5704                          * allocation.  */
5705                         if (loop >= LOOP_NO_EMPTY_SIZE &&
5706                             last_ptr->block_group != block_group) {
5707                                 spin_unlock(&last_ptr->refill_lock);
5708                                 goto unclustered_alloc;
5709                         }
5710
5711                         /*
5712                          * this cluster didn't work out, free it and
5713                          * start over
5714                          */
5715                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5716
5717                         if (loop >= LOOP_NO_EMPTY_SIZE) {
5718                                 spin_unlock(&last_ptr->refill_lock);
5719                                 goto unclustered_alloc;
5720                         }
5721
5722                         /* allocate a cluster in this block group */
5723                         ret = btrfs_find_space_cluster(trans, root,
5724                                                block_group, last_ptr,
5725                                                search_start, num_bytes,
5726                                                empty_cluster + empty_size);
5727                         if (ret == 0) {
5728                                 /*
5729                                  * now pull our allocation out of this
5730                                  * cluster
5731                                  */
5732                                 offset = btrfs_alloc_from_cluster(block_group,
5733                                                   last_ptr, num_bytes,
5734                                                   search_start);
5735                                 if (offset) {
5736                                         /* we found one, proceed */
5737                                         spin_unlock(&last_ptr->refill_lock);
5738                                         trace_btrfs_reserve_extent_cluster(root,
5739                                                 block_group, search_start,
5740                                                 num_bytes);
5741                                         goto checks;
5742                                 }
5743                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
5744                                    && !failed_cluster_refill) {
5745                                 spin_unlock(&last_ptr->refill_lock);
5746
5747                                 failed_cluster_refill = true;
5748                                 wait_block_group_cache_progress(block_group,
5749                                        num_bytes + empty_cluster + empty_size);
5750                                 goto have_block_group;
5751                         }
5752
5753                         /*
5754                          * at this point we either didn't find a cluster
5755                          * or we weren't able to allocate a block from our
5756                          * cluster.  Free the cluster we've been trying
5757                          * to use, and go to the next block group
5758                          */
5759                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
5760                         spin_unlock(&last_ptr->refill_lock);
5761                         goto loop;
5762                 }
5763
5764 unclustered_alloc:
5765                 spin_lock(&block_group->free_space_ctl->tree_lock);
5766                 if (cached &&
5767                     block_group->free_space_ctl->free_space <
5768                     num_bytes + empty_cluster + empty_size) {
5769                         spin_unlock(&block_group->free_space_ctl->tree_lock);
5770                         goto loop;
5771                 }
5772                 spin_unlock(&block_group->free_space_ctl->tree_lock);
5773
5774                 offset = btrfs_find_space_for_alloc(block_group, search_start,
5775                                                     num_bytes, empty_size);
5776                 /*
5777                  * If we didn't find a chunk, and we haven't failed on this
5778                  * block group before, and this block group is in the middle of
5779                  * caching and we are ok with waiting, then go ahead and wait
5780                  * for progress to be made, and set failed_alloc to true.
5781                  *
5782                  * If failed_alloc is true then we've already waited on this
5783                  * block group once and should move on to the next block group.
5784                  */
5785                 if (!offset && !failed_alloc && !cached &&
5786                     loop > LOOP_CACHING_NOWAIT) {
5787                         wait_block_group_cache_progress(block_group,
5788                                                 num_bytes + empty_size);
5789                         failed_alloc = true;
5790                         goto have_block_group;
5791                 } else if (!offset) {
5792                         if (!cached)
5793                                 have_caching_bg = true;
5794                         goto loop;
5795                 }
5796 checks:
5797                 search_start = stripe_align(root, offset);
5798
5799                 /* move on to the next group */
5800                 if (search_start + num_bytes >
5801                     used_block_group->key.objectid + used_block_group->key.offset) {
5802                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5803                         goto loop;
5804                 }
5805
5806                 if (offset < search_start)
5807                         btrfs_add_free_space(used_block_group, offset,
5808                                              search_start - offset);
5809                 BUG_ON(offset > search_start);
5810
5811                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
5812                                                   alloc_type);
5813                 if (ret == -EAGAIN) {
5814                         btrfs_add_free_space(used_block_group, offset, num_bytes);
5815                         goto loop;
5816                 }
5817
5818                 /* we are all good, lets return */
5819                 ins->objectid = search_start;
5820                 ins->offset = num_bytes;
5821
5822                 trace_btrfs_reserve_extent(orig_root, block_group,
5823                                            search_start, num_bytes);
5824                 if (used_block_group != block_group)
5825                         btrfs_put_block_group(used_block_group);
5826                 btrfs_put_block_group(block_group);
5827                 break;
5828 loop:
5829                 failed_cluster_refill = false;
5830                 failed_alloc = false;
5831                 BUG_ON(index != get_block_group_index(block_group));
5832                 if (used_block_group != block_group)
5833                         btrfs_put_block_group(used_block_group);
5834                 btrfs_put_block_group(block_group);
5835         }
5836         up_read(&space_info->groups_sem);
5837
5838         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
5839                 goto search;
5840
5841         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
5842                 goto search;
5843
5844         /*
5845          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
5846          *                      caching kthreads as we move along
5847          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
5848          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
5849          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
5850          *                      again
5851          */
5852         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
5853                 index = 0;
5854                 loop++;
5855                 if (loop == LOOP_ALLOC_CHUNK) {
5856                         ret = do_chunk_alloc(trans, root, data,
5857                                              CHUNK_ALLOC_FORCE);
5858                         /*
5859                          * Do not bail out on ENOSPC since we
5860                          * can do more things.
5861                          */
5862                         if (ret < 0 && ret != -ENOSPC) {
5863                                 btrfs_abort_transaction(trans,
5864                                                         root, ret);
5865                                 goto out;
5866                         }
5867                 }
5868
5869                 if (loop == LOOP_NO_EMPTY_SIZE) {
5870                         empty_size = 0;
5871                         empty_cluster = 0;
5872                 }
5873
5874                 goto search;
5875         } else if (!ins->objectid) {
5876                 ret = -ENOSPC;
5877         } else if (ins->objectid) {
5878                 ret = 0;
5879         }
5880 out:
5881
5882         return ret;
5883 }
5884
5885 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
5886                             int dump_block_groups)
5887 {
5888         struct btrfs_block_group_cache *cache;
5889         int index = 0;
5890
5891         spin_lock(&info->lock);
5892         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
5893                (unsigned long long)info->flags,
5894                (unsigned long long)(info->total_bytes - info->bytes_used -
5895                                     info->bytes_pinned - info->bytes_reserved -
5896                                     info->bytes_readonly),
5897                (info->full) ? "" : "not ");
5898         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
5899                "reserved=%llu, may_use=%llu, readonly=%llu\n",
5900                (unsigned long long)info->total_bytes,
5901                (unsigned long long)info->bytes_used,
5902                (unsigned long long)info->bytes_pinned,
5903                (unsigned long long)info->bytes_reserved,
5904                (unsigned long long)info->bytes_may_use,
5905                (unsigned long long)info->bytes_readonly);
5906         spin_unlock(&info->lock);
5907
5908         if (!dump_block_groups)
5909                 return;
5910
5911         down_read(&info->groups_sem);
5912 again:
5913         list_for_each_entry(cache, &info->block_groups[index], list) {
5914                 spin_lock(&cache->lock);
5915                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s\n",
5916                        (unsigned long long)cache->key.objectid,
5917                        (unsigned long long)cache->key.offset,
5918                        (unsigned long long)btrfs_block_group_used(&cache->item),
5919                        (unsigned long long)cache->pinned,
5920                        (unsigned long long)cache->reserved,
5921                        cache->ro ? "[readonly]" : "");
5922                 btrfs_dump_free_space(cache, bytes);
5923                 spin_unlock(&cache->lock);
5924         }
5925         if (++index < BTRFS_NR_RAID_TYPES)
5926                 goto again;
5927         up_read(&info->groups_sem);
5928 }
5929
5930 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
5931                          struct btrfs_root *root,
5932                          u64 num_bytes, u64 min_alloc_size,
5933                          u64 empty_size, u64 hint_byte,
5934                          struct btrfs_key *ins, u64 data)
5935 {
5936         bool final_tried = false;
5937         int ret;
5938
5939         data = btrfs_get_alloc_profile(root, data);
5940 again:
5941         WARN_ON(num_bytes < root->sectorsize);
5942         ret = find_free_extent(trans, root, num_bytes, empty_size,
5943                                hint_byte, ins, data);
5944
5945         if (ret == -ENOSPC) {
5946                 if (!final_tried) {
5947                         num_bytes = num_bytes >> 1;
5948                         num_bytes = num_bytes & ~(root->sectorsize - 1);
5949                         num_bytes = max(num_bytes, min_alloc_size);
5950                         if (num_bytes == min_alloc_size)
5951                                 final_tried = true;
5952                         goto again;
5953                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
5954                         struct btrfs_space_info *sinfo;
5955
5956                         sinfo = __find_space_info(root->fs_info, data);
5957                         printk(KERN_ERR "btrfs allocation failed flags %llu, "
5958                                "wanted %llu\n", (unsigned long long)data,
5959                                (unsigned long long)num_bytes);
5960                         if (sinfo)
5961                                 dump_space_info(sinfo, num_bytes, 1);
5962                 }
5963         }
5964
5965         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
5966
5967         return ret;
5968 }
5969
5970 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
5971                                         u64 start, u64 len, int pin)
5972 {
5973         struct btrfs_block_group_cache *cache;
5974         int ret = 0;
5975
5976         cache = btrfs_lookup_block_group(root->fs_info, start);
5977         if (!cache) {
5978                 printk(KERN_ERR "Unable to find block group for %llu\n",
5979                        (unsigned long long)start);
5980                 return -ENOSPC;
5981         }
5982
5983         if (btrfs_test_opt(root, DISCARD))
5984                 ret = btrfs_discard_extent(root, start, len, NULL);
5985
5986         if (pin)
5987                 pin_down_extent(root, cache, start, len, 1);
5988         else {
5989                 btrfs_add_free_space(cache, start, len);
5990                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
5991         }
5992         btrfs_put_block_group(cache);
5993
5994         trace_btrfs_reserved_extent_free(root, start, len);
5995
5996         return ret;
5997 }
5998
5999 int btrfs_free_reserved_extent(struct btrfs_root *root,
6000                                         u64 start, u64 len)
6001 {
6002         return __btrfs_free_reserved_extent(root, start, len, 0);
6003 }
6004
6005 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6006                                        u64 start, u64 len)
6007 {
6008         return __btrfs_free_reserved_extent(root, start, len, 1);
6009 }
6010
6011 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6012                                       struct btrfs_root *root,
6013                                       u64 parent, u64 root_objectid,
6014                                       u64 flags, u64 owner, u64 offset,
6015                                       struct btrfs_key *ins, int ref_mod)
6016 {
6017         int ret;
6018         struct btrfs_fs_info *fs_info = root->fs_info;
6019         struct btrfs_extent_item *extent_item;
6020         struct btrfs_extent_inline_ref *iref;
6021         struct btrfs_path *path;
6022         struct extent_buffer *leaf;
6023         int type;
6024         u32 size;
6025
6026         if (parent > 0)
6027                 type = BTRFS_SHARED_DATA_REF_KEY;
6028         else
6029                 type = BTRFS_EXTENT_DATA_REF_KEY;
6030
6031         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6032
6033         path = btrfs_alloc_path();
6034         if (!path)
6035                 return -ENOMEM;
6036
6037         path->leave_spinning = 1;
6038         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6039                                       ins, size);
6040         if (ret) {
6041                 btrfs_free_path(path);
6042                 return ret;
6043         }
6044
6045         leaf = path->nodes[0];
6046         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6047                                      struct btrfs_extent_item);
6048         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6049         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6050         btrfs_set_extent_flags(leaf, extent_item,
6051                                flags | BTRFS_EXTENT_FLAG_DATA);
6052
6053         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6054         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6055         if (parent > 0) {
6056                 struct btrfs_shared_data_ref *ref;
6057                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6058                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6059                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6060         } else {
6061                 struct btrfs_extent_data_ref *ref;
6062                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6063                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6064                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6065                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6066                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6067         }
6068
6069         btrfs_mark_buffer_dirty(path->nodes[0]);
6070         btrfs_free_path(path);
6071
6072         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
6073         if (ret) { /* -ENOENT, logic error */
6074                 printk(KERN_ERR "btrfs update block group failed for %llu "
6075                        "%llu\n", (unsigned long long)ins->objectid,
6076                        (unsigned long long)ins->offset);
6077                 BUG();
6078         }
6079         return ret;
6080 }
6081
6082 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6083                                      struct btrfs_root *root,
6084                                      u64 parent, u64 root_objectid,
6085                                      u64 flags, struct btrfs_disk_key *key,
6086                                      int level, struct btrfs_key *ins)
6087 {
6088         int ret;
6089         struct btrfs_fs_info *fs_info = root->fs_info;
6090         struct btrfs_extent_item *extent_item;
6091         struct btrfs_tree_block_info *block_info;
6092         struct btrfs_extent_inline_ref *iref;
6093         struct btrfs_path *path;
6094         struct extent_buffer *leaf;
6095         u32 size = sizeof(*extent_item) + sizeof(*block_info) + sizeof(*iref);
6096
6097         path = btrfs_alloc_path();
6098         if (!path)
6099                 return -ENOMEM;
6100
6101         path->leave_spinning = 1;
6102         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6103                                       ins, size);
6104         if (ret) {
6105                 btrfs_free_path(path);
6106                 return ret;
6107         }
6108
6109         leaf = path->nodes[0];
6110         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6111                                      struct btrfs_extent_item);
6112         btrfs_set_extent_refs(leaf, extent_item, 1);
6113         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6114         btrfs_set_extent_flags(leaf, extent_item,
6115                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6116         block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6117
6118         btrfs_set_tree_block_key(leaf, block_info, key);
6119         btrfs_set_tree_block_level(leaf, block_info, level);
6120
6121         iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6122         if (parent > 0) {
6123                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6124                 btrfs_set_extent_inline_ref_type(leaf, iref,
6125                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6126                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6127         } else {
6128                 btrfs_set_extent_inline_ref_type(leaf, iref,
6129                                                  BTRFS_TREE_BLOCK_REF_KEY);
6130                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6131         }
6132
6133         btrfs_mark_buffer_dirty(leaf);
6134         btrfs_free_path(path);
6135
6136         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
6137         if (ret) { /* -ENOENT, logic error */
6138                 printk(KERN_ERR "btrfs update block group failed for %llu "
6139                        "%llu\n", (unsigned long long)ins->objectid,
6140                        (unsigned long long)ins->offset);
6141                 BUG();
6142         }
6143         return ret;
6144 }
6145
6146 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6147                                      struct btrfs_root *root,
6148                                      u64 root_objectid, u64 owner,
6149                                      u64 offset, struct btrfs_key *ins)
6150 {
6151         int ret;
6152
6153         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6154
6155         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6156                                          ins->offset, 0,
6157                                          root_objectid, owner, offset,
6158                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6159         return ret;
6160 }
6161
6162 /*
6163  * this is used by the tree logging recovery code.  It records that
6164  * an extent has been allocated and makes sure to clear the free
6165  * space cache bits as well
6166  */
6167 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6168                                    struct btrfs_root *root,
6169                                    u64 root_objectid, u64 owner, u64 offset,
6170                                    struct btrfs_key *ins)
6171 {
6172         int ret;
6173         struct btrfs_block_group_cache *block_group;
6174         struct btrfs_caching_control *caching_ctl;
6175         u64 start = ins->objectid;
6176         u64 num_bytes = ins->offset;
6177
6178         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6179         cache_block_group(block_group, trans, NULL, 0);
6180         caching_ctl = get_caching_control(block_group);
6181
6182         if (!caching_ctl) {
6183                 BUG_ON(!block_group_cache_done(block_group));
6184                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6185                 BUG_ON(ret); /* -ENOMEM */
6186         } else {
6187                 mutex_lock(&caching_ctl->mutex);
6188
6189                 if (start >= caching_ctl->progress) {
6190                         ret = add_excluded_extent(root, start, num_bytes);
6191                         BUG_ON(ret); /* -ENOMEM */
6192                 } else if (start + num_bytes <= caching_ctl->progress) {
6193                         ret = btrfs_remove_free_space(block_group,
6194                                                       start, num_bytes);
6195                         BUG_ON(ret); /* -ENOMEM */
6196                 } else {
6197                         num_bytes = caching_ctl->progress - start;
6198                         ret = btrfs_remove_free_space(block_group,
6199                                                       start, num_bytes);
6200                         BUG_ON(ret); /* -ENOMEM */
6201
6202                         start = caching_ctl->progress;
6203                         num_bytes = ins->objectid + ins->offset -
6204                                     caching_ctl->progress;
6205                         ret = add_excluded_extent(root, start, num_bytes);
6206                         BUG_ON(ret); /* -ENOMEM */
6207                 }
6208
6209                 mutex_unlock(&caching_ctl->mutex);
6210                 put_caching_control(caching_ctl);
6211         }
6212
6213         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6214                                           RESERVE_ALLOC_NO_ACCOUNT);
6215         BUG_ON(ret); /* logic error */
6216         btrfs_put_block_group(block_group);
6217         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6218                                          0, owner, offset, ins, 1);
6219         return ret;
6220 }
6221
6222 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
6223                                             struct btrfs_root *root,
6224                                             u64 bytenr, u32 blocksize,
6225                                             int level)
6226 {
6227         struct extent_buffer *buf;
6228
6229         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6230         if (!buf)
6231                 return ERR_PTR(-ENOMEM);
6232         btrfs_set_header_generation(buf, trans->transid);
6233         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6234         btrfs_tree_lock(buf);
6235         clean_tree_block(trans, root, buf);
6236         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6237
6238         btrfs_set_lock_blocking(buf);
6239         btrfs_set_buffer_uptodate(buf);
6240
6241         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6242                 /*
6243                  * we allow two log transactions at a time, use different
6244                  * EXENT bit to differentiate dirty pages.
6245                  */
6246                 if (root->log_transid % 2 == 0)
6247                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6248                                         buf->start + buf->len - 1, GFP_NOFS);
6249                 else
6250                         set_extent_new(&root->dirty_log_pages, buf->start,
6251                                         buf->start + buf->len - 1, GFP_NOFS);
6252         } else {
6253                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6254                          buf->start + buf->len - 1, GFP_NOFS);
6255         }
6256         trans->blocks_used++;
6257         /* this returns a buffer locked for blocking */
6258         return buf;
6259 }
6260
6261 static struct btrfs_block_rsv *
6262 use_block_rsv(struct btrfs_trans_handle *trans,
6263               struct btrfs_root *root, u32 blocksize)
6264 {
6265         struct btrfs_block_rsv *block_rsv;
6266         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6267         int ret;
6268
6269         block_rsv = get_block_rsv(trans, root);
6270
6271         if (block_rsv->size == 0) {
6272                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
6273                 /*
6274                  * If we couldn't reserve metadata bytes try and use some from
6275                  * the global reserve.
6276                  */
6277                 if (ret && block_rsv != global_rsv) {
6278                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6279                         if (!ret)
6280                                 return global_rsv;
6281                         return ERR_PTR(ret);
6282                 } else if (ret) {
6283                         return ERR_PTR(ret);
6284                 }
6285                 return block_rsv;
6286         }
6287
6288         ret = block_rsv_use_bytes(block_rsv, blocksize);
6289         if (!ret)
6290                 return block_rsv;
6291         if (ret && !block_rsv->failfast) {
6292                 static DEFINE_RATELIMIT_STATE(_rs,
6293                                 DEFAULT_RATELIMIT_INTERVAL,
6294                                 /*DEFAULT_RATELIMIT_BURST*/ 2);
6295                 if (__ratelimit(&_rs)) {
6296                         printk(KERN_DEBUG "btrfs: block rsv returned %d\n", ret);
6297                         WARN_ON(1);
6298                 }
6299                 ret = reserve_metadata_bytes(root, block_rsv, blocksize, 0);
6300                 if (!ret) {
6301                         return block_rsv;
6302                 } else if (ret && block_rsv != global_rsv) {
6303                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6304                         if (!ret)
6305                                 return global_rsv;
6306                 }
6307         }
6308
6309         return ERR_PTR(-ENOSPC);
6310 }
6311
6312 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6313                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6314 {
6315         block_rsv_add_bytes(block_rsv, blocksize, 0);
6316         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6317 }
6318
6319 /*
6320  * finds a free extent and does all the dirty work required for allocation
6321  * returns the key for the extent through ins, and a tree buffer for
6322  * the first block of the extent through buf.
6323  *
6324  * returns the tree buffer or NULL.
6325  */
6326 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6327                                         struct btrfs_root *root, u32 blocksize,
6328                                         u64 parent, u64 root_objectid,
6329                                         struct btrfs_disk_key *key, int level,
6330                                         u64 hint, u64 empty_size)
6331 {
6332         struct btrfs_key ins;
6333         struct btrfs_block_rsv *block_rsv;
6334         struct extent_buffer *buf;
6335         u64 flags = 0;
6336         int ret;
6337
6338
6339         block_rsv = use_block_rsv(trans, root, blocksize);
6340         if (IS_ERR(block_rsv))
6341                 return ERR_CAST(block_rsv);
6342
6343         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6344                                    empty_size, hint, &ins, 0);
6345         if (ret) {
6346                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6347                 return ERR_PTR(ret);
6348         }
6349
6350         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6351                                     blocksize, level);
6352         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6353
6354         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6355                 if (parent == 0)
6356                         parent = ins.objectid;
6357                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6358         } else
6359                 BUG_ON(parent > 0);
6360
6361         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6362                 struct btrfs_delayed_extent_op *extent_op;
6363                 extent_op = kmalloc(sizeof(*extent_op), GFP_NOFS);
6364                 BUG_ON(!extent_op); /* -ENOMEM */
6365                 if (key)
6366                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6367                 else
6368                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6369                 extent_op->flags_to_set = flags;
6370                 extent_op->update_key = 1;
6371                 extent_op->update_flags = 1;
6372                 extent_op->is_data = 0;
6373
6374                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6375                                         ins.objectid,
6376                                         ins.offset, parent, root_objectid,
6377                                         level, BTRFS_ADD_DELAYED_EXTENT,
6378                                         extent_op, 0);
6379                 BUG_ON(ret); /* -ENOMEM */
6380         }
6381         return buf;
6382 }
6383
6384 struct walk_control {
6385         u64 refs[BTRFS_MAX_LEVEL];
6386         u64 flags[BTRFS_MAX_LEVEL];
6387         struct btrfs_key update_progress;
6388         int stage;
6389         int level;
6390         int shared_level;
6391         int update_ref;
6392         int keep_locks;
6393         int reada_slot;
6394         int reada_count;
6395         int for_reloc;
6396 };
6397
6398 #define DROP_REFERENCE  1
6399 #define UPDATE_BACKREF  2
6400
6401 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6402                                      struct btrfs_root *root,
6403                                      struct walk_control *wc,
6404                                      struct btrfs_path *path)
6405 {
6406         u64 bytenr;
6407         u64 generation;
6408         u64 refs;
6409         u64 flags;
6410         u32 nritems;
6411         u32 blocksize;
6412         struct btrfs_key key;
6413         struct extent_buffer *eb;
6414         int ret;
6415         int slot;
6416         int nread = 0;
6417
6418         if (path->slots[wc->level] < wc->reada_slot) {
6419                 wc->reada_count = wc->reada_count * 2 / 3;
6420                 wc->reada_count = max(wc->reada_count, 2);
6421         } else {
6422                 wc->reada_count = wc->reada_count * 3 / 2;
6423                 wc->reada_count = min_t(int, wc->reada_count,
6424                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6425         }
6426
6427         eb = path->nodes[wc->level];
6428         nritems = btrfs_header_nritems(eb);
6429         blocksize = btrfs_level_size(root, wc->level - 1);
6430
6431         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6432                 if (nread >= wc->reada_count)
6433                         break;
6434
6435                 cond_resched();
6436                 bytenr = btrfs_node_blockptr(eb, slot);
6437                 generation = btrfs_node_ptr_generation(eb, slot);
6438
6439                 if (slot == path->slots[wc->level])
6440                         goto reada;
6441
6442                 if (wc->stage == UPDATE_BACKREF &&
6443                     generation <= root->root_key.offset)
6444                         continue;
6445
6446                 /* We don't lock the tree block, it's OK to be racy here */
6447                 ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6448                                                &refs, &flags);
6449                 /* We don't care about errors in readahead. */
6450                 if (ret < 0)
6451                         continue;
6452                 BUG_ON(refs == 0);
6453
6454                 if (wc->stage == DROP_REFERENCE) {
6455                         if (refs == 1)
6456                                 goto reada;
6457
6458                         if (wc->level == 1 &&
6459                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6460                                 continue;
6461                         if (!wc->update_ref ||
6462                             generation <= root->root_key.offset)
6463                                 continue;
6464                         btrfs_node_key_to_cpu(eb, &key, slot);
6465                         ret = btrfs_comp_cpu_keys(&key,
6466                                                   &wc->update_progress);
6467                         if (ret < 0)
6468                                 continue;
6469                 } else {
6470                         if (wc->level == 1 &&
6471                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6472                                 continue;
6473                 }
6474 reada:
6475                 ret = readahead_tree_block(root, bytenr, blocksize,
6476                                            generation);
6477                 if (ret)
6478                         break;
6479                 nread++;
6480         }
6481         wc->reada_slot = slot;
6482 }
6483
6484 /*
6485  * hepler to process tree block while walking down the tree.
6486  *
6487  * when wc->stage == UPDATE_BACKREF, this function updates
6488  * back refs for pointers in the block.
6489  *
6490  * NOTE: return value 1 means we should stop walking down.
6491  */
6492 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6493                                    struct btrfs_root *root,
6494                                    struct btrfs_path *path,
6495                                    struct walk_control *wc, int lookup_info)
6496 {
6497         int level = wc->level;
6498         struct extent_buffer *eb = path->nodes[level];
6499         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6500         int ret;
6501
6502         if (wc->stage == UPDATE_BACKREF &&
6503             btrfs_header_owner(eb) != root->root_key.objectid)
6504                 return 1;
6505
6506         /*
6507          * when reference count of tree block is 1, it won't increase
6508          * again. once full backref flag is set, we never clear it.
6509          */
6510         if (lookup_info &&
6511             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6512              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6513                 BUG_ON(!path->locks[level]);
6514                 ret = btrfs_lookup_extent_info(trans, root,
6515                                                eb->start, eb->len,
6516                                                &wc->refs[level],
6517                                                &wc->flags[level]);
6518                 BUG_ON(ret == -ENOMEM);
6519                 if (ret)
6520                         return ret;
6521                 BUG_ON(wc->refs[level] == 0);
6522         }
6523
6524         if (wc->stage == DROP_REFERENCE) {
6525                 if (wc->refs[level] > 1)
6526                         return 1;
6527
6528                 if (path->locks[level] && !wc->keep_locks) {
6529                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6530                         path->locks[level] = 0;
6531                 }
6532                 return 0;
6533         }
6534
6535         /* wc->stage == UPDATE_BACKREF */
6536         if (!(wc->flags[level] & flag)) {
6537                 BUG_ON(!path->locks[level]);
6538                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
6539                 BUG_ON(ret); /* -ENOMEM */
6540                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
6541                 BUG_ON(ret); /* -ENOMEM */
6542                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6543                                                   eb->len, flag, 0);
6544                 BUG_ON(ret); /* -ENOMEM */
6545                 wc->flags[level] |= flag;
6546         }
6547
6548         /*
6549          * the block is shared by multiple trees, so it's not good to
6550          * keep the tree lock
6551          */
6552         if (path->locks[level] && level > 0) {
6553                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6554                 path->locks[level] = 0;
6555         }
6556         return 0;
6557 }
6558
6559 /*
6560  * hepler to process tree block pointer.
6561  *
6562  * when wc->stage == DROP_REFERENCE, this function checks
6563  * reference count of the block pointed to. if the block
6564  * is shared and we need update back refs for the subtree
6565  * rooted at the block, this function changes wc->stage to
6566  * UPDATE_BACKREF. if the block is shared and there is no
6567  * need to update back, this function drops the reference
6568  * to the block.
6569  *
6570  * NOTE: return value 1 means we should stop walking down.
6571  */
6572 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
6573                                  struct btrfs_root *root,
6574                                  struct btrfs_path *path,
6575                                  struct walk_control *wc, int *lookup_info)
6576 {
6577         u64 bytenr;
6578         u64 generation;
6579         u64 parent;
6580         u32 blocksize;
6581         struct btrfs_key key;
6582         struct extent_buffer *next;
6583         int level = wc->level;
6584         int reada = 0;
6585         int ret = 0;
6586
6587         generation = btrfs_node_ptr_generation(path->nodes[level],
6588                                                path->slots[level]);
6589         /*
6590          * if the lower level block was created before the snapshot
6591          * was created, we know there is no need to update back refs
6592          * for the subtree
6593          */
6594         if (wc->stage == UPDATE_BACKREF &&
6595             generation <= root->root_key.offset) {
6596                 *lookup_info = 1;
6597                 return 1;
6598         }
6599
6600         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
6601         blocksize = btrfs_level_size(root, level - 1);
6602
6603         next = btrfs_find_tree_block(root, bytenr, blocksize);
6604         if (!next) {
6605                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
6606                 if (!next)
6607                         return -ENOMEM;
6608                 reada = 1;
6609         }
6610         btrfs_tree_lock(next);
6611         btrfs_set_lock_blocking(next);
6612
6613         ret = btrfs_lookup_extent_info(trans, root, bytenr, blocksize,
6614                                        &wc->refs[level - 1],
6615                                        &wc->flags[level - 1]);
6616         if (ret < 0) {
6617                 btrfs_tree_unlock(next);
6618                 return ret;
6619         }
6620
6621         BUG_ON(wc->refs[level - 1] == 0);
6622         *lookup_info = 0;
6623
6624         if (wc->stage == DROP_REFERENCE) {
6625                 if (wc->refs[level - 1] > 1) {
6626                         if (level == 1 &&
6627                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6628                                 goto skip;
6629
6630                         if (!wc->update_ref ||
6631                             generation <= root->root_key.offset)
6632                                 goto skip;
6633
6634                         btrfs_node_key_to_cpu(path->nodes[level], &key,
6635                                               path->slots[level]);
6636                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
6637                         if (ret < 0)
6638                                 goto skip;
6639
6640                         wc->stage = UPDATE_BACKREF;
6641                         wc->shared_level = level - 1;
6642                 }
6643         } else {
6644                 if (level == 1 &&
6645                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6646                         goto skip;
6647         }
6648
6649         if (!btrfs_buffer_uptodate(next, generation, 0)) {
6650                 btrfs_tree_unlock(next);
6651                 free_extent_buffer(next);
6652                 next = NULL;
6653                 *lookup_info = 1;
6654         }
6655
6656         if (!next) {
6657                 if (reada && level == 1)
6658                         reada_walk_down(trans, root, wc, path);
6659                 next = read_tree_block(root, bytenr, blocksize, generation);
6660                 if (!next)
6661                         return -EIO;
6662                 btrfs_tree_lock(next);
6663                 btrfs_set_lock_blocking(next);
6664         }
6665
6666         level--;
6667         BUG_ON(level != btrfs_header_level(next));
6668         path->nodes[level] = next;
6669         path->slots[level] = 0;
6670         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6671         wc->level = level;
6672         if (wc->level == 1)
6673                 wc->reada_slot = 0;
6674         return 0;
6675 skip:
6676         wc->refs[level - 1] = 0;
6677         wc->flags[level - 1] = 0;
6678         if (wc->stage == DROP_REFERENCE) {
6679                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
6680                         parent = path->nodes[level]->start;
6681                 } else {
6682                         BUG_ON(root->root_key.objectid !=
6683                                btrfs_header_owner(path->nodes[level]));
6684                         parent = 0;
6685                 }
6686
6687                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
6688                                 root->root_key.objectid, level - 1, 0, 0);
6689                 BUG_ON(ret); /* -ENOMEM */
6690         }
6691         btrfs_tree_unlock(next);
6692         free_extent_buffer(next);
6693         *lookup_info = 1;
6694         return 1;
6695 }
6696
6697 /*
6698  * hepler to process tree block while walking up the tree.
6699  *
6700  * when wc->stage == DROP_REFERENCE, this function drops
6701  * reference count on the block.
6702  *
6703  * when wc->stage == UPDATE_BACKREF, this function changes
6704  * wc->stage back to DROP_REFERENCE if we changed wc->stage
6705  * to UPDATE_BACKREF previously while processing the block.
6706  *
6707  * NOTE: return value 1 means we should stop walking up.
6708  */
6709 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
6710                                  struct btrfs_root *root,
6711                                  struct btrfs_path *path,
6712                                  struct walk_control *wc)
6713 {
6714         int ret;
6715         int level = wc->level;
6716         struct extent_buffer *eb = path->nodes[level];
6717         u64 parent = 0;
6718
6719         if (wc->stage == UPDATE_BACKREF) {
6720                 BUG_ON(wc->shared_level < level);
6721                 if (level < wc->shared_level)
6722                         goto out;
6723
6724                 ret = find_next_key(path, level + 1, &wc->update_progress);
6725                 if (ret > 0)
6726                         wc->update_ref = 0;
6727
6728                 wc->stage = DROP_REFERENCE;
6729                 wc->shared_level = -1;
6730                 path->slots[level] = 0;
6731
6732                 /*
6733                  * check reference count again if the block isn't locked.
6734                  * we should start walking down the tree again if reference
6735                  * count is one.
6736                  */
6737                 if (!path->locks[level]) {
6738                         BUG_ON(level == 0);
6739                         btrfs_tree_lock(eb);
6740                         btrfs_set_lock_blocking(eb);
6741                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6742
6743                         ret = btrfs_lookup_extent_info(trans, root,
6744                                                        eb->start, eb->len,
6745                                                        &wc->refs[level],
6746                                                        &wc->flags[level]);
6747                         if (ret < 0) {
6748                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6749                                 return ret;
6750                         }
6751                         BUG_ON(wc->refs[level] == 0);
6752                         if (wc->refs[level] == 1) {
6753                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6754                                 return 1;
6755                         }
6756                 }
6757         }
6758
6759         /* wc->stage == DROP_REFERENCE */
6760         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
6761
6762         if (wc->refs[level] == 1) {
6763                 if (level == 0) {
6764                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6765                                 ret = btrfs_dec_ref(trans, root, eb, 1,
6766                                                     wc->for_reloc);
6767                         else
6768                                 ret = btrfs_dec_ref(trans, root, eb, 0,
6769                                                     wc->for_reloc);
6770                         BUG_ON(ret); /* -ENOMEM */
6771                 }
6772                 /* make block locked assertion in clean_tree_block happy */
6773                 if (!path->locks[level] &&
6774                     btrfs_header_generation(eb) == trans->transid) {
6775                         btrfs_tree_lock(eb);
6776                         btrfs_set_lock_blocking(eb);
6777                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6778                 }
6779                 clean_tree_block(trans, root, eb);
6780         }
6781
6782         if (eb == root->node) {
6783                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6784                         parent = eb->start;
6785                 else
6786                         BUG_ON(root->root_key.objectid !=
6787                                btrfs_header_owner(eb));
6788         } else {
6789                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
6790                         parent = path->nodes[level + 1]->start;
6791                 else
6792                         BUG_ON(root->root_key.objectid !=
6793                                btrfs_header_owner(path->nodes[level + 1]));
6794         }
6795
6796         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
6797 out:
6798         wc->refs[level] = 0;
6799         wc->flags[level] = 0;
6800         return 0;
6801 }
6802
6803 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
6804                                    struct btrfs_root *root,
6805                                    struct btrfs_path *path,
6806                                    struct walk_control *wc)
6807 {
6808         int level = wc->level;
6809         int lookup_info = 1;
6810         int ret;
6811
6812         while (level >= 0) {
6813                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
6814                 if (ret > 0)
6815                         break;
6816
6817                 if (level == 0)
6818                         break;
6819
6820                 if (path->slots[level] >=
6821                     btrfs_header_nritems(path->nodes[level]))
6822                         break;
6823
6824                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
6825                 if (ret > 0) {
6826                         path->slots[level]++;
6827                         continue;
6828                 } else if (ret < 0)
6829                         return ret;
6830                 level = wc->level;
6831         }
6832         return 0;
6833 }
6834
6835 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
6836                                  struct btrfs_root *root,
6837                                  struct btrfs_path *path,
6838                                  struct walk_control *wc, int max_level)
6839 {
6840         int level = wc->level;
6841         int ret;
6842
6843         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
6844         while (level < max_level && path->nodes[level]) {
6845                 wc->level = level;
6846                 if (path->slots[level] + 1 <
6847                     btrfs_header_nritems(path->nodes[level])) {
6848                         path->slots[level]++;
6849                         return 0;
6850                 } else {
6851                         ret = walk_up_proc(trans, root, path, wc);
6852                         if (ret > 0)
6853                                 return 0;
6854
6855                         if (path->locks[level]) {
6856                                 btrfs_tree_unlock_rw(path->nodes[level],
6857                                                      path->locks[level]);
6858                                 path->locks[level] = 0;
6859                         }
6860                         free_extent_buffer(path->nodes[level]);
6861                         path->nodes[level] = NULL;
6862                         level++;
6863                 }
6864         }
6865         return 1;
6866 }
6867
6868 /*
6869  * drop a subvolume tree.
6870  *
6871  * this function traverses the tree freeing any blocks that only
6872  * referenced by the tree.
6873  *
6874  * when a shared tree block is found. this function decreases its
6875  * reference count by one. if update_ref is true, this function
6876  * also make sure backrefs for the shared block and all lower level
6877  * blocks are properly updated.
6878  */
6879 int btrfs_drop_snapshot(struct btrfs_root *root,
6880                          struct btrfs_block_rsv *block_rsv, int update_ref,
6881                          int for_reloc)
6882 {
6883         struct btrfs_path *path;
6884         struct btrfs_trans_handle *trans;
6885         struct btrfs_root *tree_root = root->fs_info->tree_root;
6886         struct btrfs_root_item *root_item = &root->root_item;
6887         struct walk_control *wc;
6888         struct btrfs_key key;
6889         int err = 0;
6890         int ret;
6891         int level;
6892
6893         path = btrfs_alloc_path();
6894         if (!path) {
6895                 err = -ENOMEM;
6896                 goto out;
6897         }
6898
6899         wc = kzalloc(sizeof(*wc), GFP_NOFS);
6900         if (!wc) {
6901                 btrfs_free_path(path);
6902                 err = -ENOMEM;
6903                 goto out;
6904         }
6905
6906         trans = btrfs_start_transaction(tree_root, 0);
6907         if (IS_ERR(trans)) {
6908                 err = PTR_ERR(trans);
6909                 goto out_free;
6910         }
6911
6912         if (block_rsv)
6913                 trans->block_rsv = block_rsv;
6914
6915         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
6916                 level = btrfs_header_level(root->node);
6917                 path->nodes[level] = btrfs_lock_root_node(root);
6918                 btrfs_set_lock_blocking(path->nodes[level]);
6919                 path->slots[level] = 0;
6920                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
6921                 memset(&wc->update_progress, 0,
6922                        sizeof(wc->update_progress));
6923         } else {
6924                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
6925                 memcpy(&wc->update_progress, &key,
6926                        sizeof(wc->update_progress));
6927
6928                 level = root_item->drop_level;
6929                 BUG_ON(level == 0);
6930                 path->lowest_level = level;
6931                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
6932                 path->lowest_level = 0;
6933                 if (ret < 0) {
6934                         err = ret;
6935                         goto out_end_trans;
6936                 }
6937                 WARN_ON(ret > 0);
6938
6939                 /*
6940                  * unlock our path, this is safe because only this
6941                  * function is allowed to delete this snapshot
6942                  */
6943                 btrfs_unlock_up_safe(path, 0);
6944
6945                 level = btrfs_header_level(root->node);
6946                 while (1) {
6947                         btrfs_tree_lock(path->nodes[level]);
6948                         btrfs_set_lock_blocking(path->nodes[level]);
6949
6950                         ret = btrfs_lookup_extent_info(trans, root,
6951                                                 path->nodes[level]->start,
6952                                                 path->nodes[level]->len,
6953                                                 &wc->refs[level],
6954                                                 &wc->flags[level]);
6955                         if (ret < 0) {
6956                                 err = ret;
6957                                 goto out_end_trans;
6958                         }
6959                         BUG_ON(wc->refs[level] == 0);
6960
6961                         if (level == root_item->drop_level)
6962                                 break;
6963
6964                         btrfs_tree_unlock(path->nodes[level]);
6965                         WARN_ON(wc->refs[level] != 1);
6966                         level--;
6967                 }
6968         }
6969
6970         wc->level = level;
6971         wc->shared_level = -1;
6972         wc->stage = DROP_REFERENCE;
6973         wc->update_ref = update_ref;
6974         wc->keep_locks = 0;
6975         wc->for_reloc = for_reloc;
6976         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
6977
6978         while (1) {
6979                 ret = walk_down_tree(trans, root, path, wc);
6980                 if (ret < 0) {
6981                         err = ret;
6982                         break;
6983                 }
6984
6985                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
6986                 if (ret < 0) {
6987                         err = ret;
6988                         break;
6989                 }
6990
6991                 if (ret > 0) {
6992                         BUG_ON(wc->stage != DROP_REFERENCE);
6993                         break;
6994                 }
6995
6996                 if (wc->stage == DROP_REFERENCE) {
6997                         level = wc->level;
6998                         btrfs_node_key(path->nodes[level],
6999                                        &root_item->drop_progress,
7000                                        path->slots[level]);
7001                         root_item->drop_level = level;
7002                 }
7003
7004                 BUG_ON(wc->level == 0);
7005                 if (btrfs_should_end_transaction(trans, tree_root)) {
7006                         ret = btrfs_update_root(trans, tree_root,
7007                                                 &root->root_key,
7008                                                 root_item);
7009                         if (ret) {
7010                                 btrfs_abort_transaction(trans, tree_root, ret);
7011                                 err = ret;
7012                                 goto out_end_trans;
7013                         }
7014
7015                         btrfs_end_transaction_throttle(trans, tree_root);
7016                         trans = btrfs_start_transaction(tree_root, 0);
7017                         if (IS_ERR(trans)) {
7018                                 err = PTR_ERR(trans);
7019                                 goto out_free;
7020                         }
7021                         if (block_rsv)
7022                                 trans->block_rsv = block_rsv;
7023                 }
7024         }
7025         btrfs_release_path(path);
7026         if (err)
7027                 goto out_end_trans;
7028
7029         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7030         if (ret) {
7031                 btrfs_abort_transaction(trans, tree_root, ret);
7032                 goto out_end_trans;
7033         }
7034
7035         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7036                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
7037                                            NULL, NULL);
7038                 if (ret < 0) {
7039                         btrfs_abort_transaction(trans, tree_root, ret);
7040                         err = ret;
7041                         goto out_end_trans;
7042                 } else if (ret > 0) {
7043                         /* if we fail to delete the orphan item this time
7044                          * around, it'll get picked up the next time.
7045                          *
7046                          * The most common failure here is just -ENOENT.
7047                          */
7048                         btrfs_del_orphan_item(trans, tree_root,
7049                                               root->root_key.objectid);
7050                 }
7051         }
7052
7053         if (root->in_radix) {
7054                 btrfs_free_fs_root(tree_root->fs_info, root);
7055         } else {
7056                 free_extent_buffer(root->node);
7057                 free_extent_buffer(root->commit_root);
7058                 kfree(root);
7059         }
7060 out_end_trans:
7061         btrfs_end_transaction_throttle(trans, tree_root);
7062 out_free:
7063         kfree(wc);
7064         btrfs_free_path(path);
7065 out:
7066         if (err)
7067                 btrfs_std_error(root->fs_info, err);
7068         return err;
7069 }
7070
7071 /*
7072  * drop subtree rooted at tree block 'node'.
7073  *
7074  * NOTE: this function will unlock and release tree block 'node'
7075  * only used by relocation code
7076  */
7077 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7078                         struct btrfs_root *root,
7079                         struct extent_buffer *node,
7080                         struct extent_buffer *parent)
7081 {
7082         struct btrfs_path *path;
7083         struct walk_control *wc;
7084         int level;
7085         int parent_level;
7086         int ret = 0;
7087         int wret;
7088
7089         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7090
7091         path = btrfs_alloc_path();
7092         if (!path)
7093                 return -ENOMEM;
7094
7095         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7096         if (!wc) {
7097                 btrfs_free_path(path);
7098                 return -ENOMEM;
7099         }
7100
7101         btrfs_assert_tree_locked(parent);
7102         parent_level = btrfs_header_level(parent);
7103         extent_buffer_get(parent);
7104         path->nodes[parent_level] = parent;
7105         path->slots[parent_level] = btrfs_header_nritems(parent);
7106
7107         btrfs_assert_tree_locked(node);
7108         level = btrfs_header_level(node);
7109         path->nodes[level] = node;
7110         path->slots[level] = 0;
7111         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7112
7113         wc->refs[parent_level] = 1;
7114         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7115         wc->level = level;
7116         wc->shared_level = -1;
7117         wc->stage = DROP_REFERENCE;
7118         wc->update_ref = 0;
7119         wc->keep_locks = 1;
7120         wc->for_reloc = 1;
7121         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7122
7123         while (1) {
7124                 wret = walk_down_tree(trans, root, path, wc);
7125                 if (wret < 0) {
7126                         ret = wret;
7127                         break;
7128                 }
7129
7130                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7131                 if (wret < 0)
7132                         ret = wret;
7133                 if (wret != 0)
7134                         break;
7135         }
7136
7137         kfree(wc);
7138         btrfs_free_path(path);
7139         return ret;
7140 }
7141
7142 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7143 {
7144         u64 num_devices;
7145         u64 stripped;
7146
7147         /*
7148          * if restripe for this chunk_type is on pick target profile and
7149          * return, otherwise do the usual balance
7150          */
7151         stripped = get_restripe_target(root->fs_info, flags);
7152         if (stripped)
7153                 return extended_to_chunk(stripped);
7154
7155         /*
7156          * we add in the count of missing devices because we want
7157          * to make sure that any RAID levels on a degraded FS
7158          * continue to be honored.
7159          */
7160         num_devices = root->fs_info->fs_devices->rw_devices +
7161                 root->fs_info->fs_devices->missing_devices;
7162
7163         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7164                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7165
7166         if (num_devices == 1) {
7167                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7168                 stripped = flags & ~stripped;
7169
7170                 /* turn raid0 into single device chunks */
7171                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7172                         return stripped;
7173
7174                 /* turn mirroring into duplication */
7175                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7176                              BTRFS_BLOCK_GROUP_RAID10))
7177                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7178         } else {
7179                 /* they already had raid on here, just return */
7180                 if (flags & stripped)
7181                         return flags;
7182
7183                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7184                 stripped = flags & ~stripped;
7185
7186                 /* switch duplicated blocks with raid1 */
7187                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7188                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7189
7190                 /* this is drive concat, leave it alone */
7191         }
7192
7193         return flags;
7194 }
7195
7196 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7197 {
7198         struct btrfs_space_info *sinfo = cache->space_info;
7199         u64 num_bytes;
7200         u64 min_allocable_bytes;
7201         int ret = -ENOSPC;
7202
7203
7204         /*
7205          * We need some metadata space and system metadata space for
7206          * allocating chunks in some corner cases until we force to set
7207          * it to be readonly.
7208          */
7209         if ((sinfo->flags &
7210              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7211             !force)
7212                 min_allocable_bytes = 1 * 1024 * 1024;
7213         else
7214                 min_allocable_bytes = 0;
7215
7216         spin_lock(&sinfo->lock);
7217         spin_lock(&cache->lock);
7218
7219         if (cache->ro) {
7220                 ret = 0;
7221                 goto out;
7222         }
7223
7224         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7225                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7226
7227         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7228             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7229             min_allocable_bytes <= sinfo->total_bytes) {
7230                 sinfo->bytes_readonly += num_bytes;
7231                 cache->ro = 1;
7232                 ret = 0;
7233         }
7234 out:
7235         spin_unlock(&cache->lock);
7236         spin_unlock(&sinfo->lock);
7237         return ret;
7238 }
7239
7240 int btrfs_set_block_group_ro(struct btrfs_root *root,
7241                              struct btrfs_block_group_cache *cache)
7242
7243 {
7244         struct btrfs_trans_handle *trans;
7245         u64 alloc_flags;
7246         int ret;
7247
7248         BUG_ON(cache->ro);
7249
7250         trans = btrfs_join_transaction(root);
7251         if (IS_ERR(trans))
7252                 return PTR_ERR(trans);
7253
7254         alloc_flags = update_block_group_flags(root, cache->flags);
7255         if (alloc_flags != cache->flags) {
7256                 ret = do_chunk_alloc(trans, root, alloc_flags,
7257                                      CHUNK_ALLOC_FORCE);
7258                 if (ret < 0)
7259                         goto out;
7260         }
7261
7262         ret = set_block_group_ro(cache, 0);
7263         if (!ret)
7264                 goto out;
7265         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7266         ret = do_chunk_alloc(trans, root, alloc_flags,
7267                              CHUNK_ALLOC_FORCE);
7268         if (ret < 0)
7269                 goto out;
7270         ret = set_block_group_ro(cache, 0);
7271 out:
7272         btrfs_end_transaction(trans, root);
7273         return ret;
7274 }
7275
7276 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7277                             struct btrfs_root *root, u64 type)
7278 {
7279         u64 alloc_flags = get_alloc_profile(root, type);
7280         return do_chunk_alloc(trans, root, alloc_flags,
7281                               CHUNK_ALLOC_FORCE);
7282 }
7283
7284 /*
7285  * helper to account the unused space of all the readonly block group in the
7286  * list. takes mirrors into account.
7287  */
7288 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7289 {
7290         struct btrfs_block_group_cache *block_group;
7291         u64 free_bytes = 0;
7292         int factor;
7293
7294         list_for_each_entry(block_group, groups_list, list) {
7295                 spin_lock(&block_group->lock);
7296
7297                 if (!block_group->ro) {
7298                         spin_unlock(&block_group->lock);
7299                         continue;
7300                 }
7301
7302                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7303                                           BTRFS_BLOCK_GROUP_RAID10 |
7304                                           BTRFS_BLOCK_GROUP_DUP))
7305                         factor = 2;
7306                 else
7307                         factor = 1;
7308
7309                 free_bytes += (block_group->key.offset -
7310                                btrfs_block_group_used(&block_group->item)) *
7311                                factor;
7312
7313                 spin_unlock(&block_group->lock);
7314         }
7315
7316         return free_bytes;
7317 }
7318
7319 /*
7320  * helper to account the unused space of all the readonly block group in the
7321  * space_info. takes mirrors into account.
7322  */
7323 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7324 {
7325         int i;
7326         u64 free_bytes = 0;
7327
7328         spin_lock(&sinfo->lock);
7329
7330         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7331                 if (!list_empty(&sinfo->block_groups[i]))
7332                         free_bytes += __btrfs_get_ro_block_group_free_space(
7333                                                 &sinfo->block_groups[i]);
7334
7335         spin_unlock(&sinfo->lock);
7336
7337         return free_bytes;
7338 }
7339
7340 void btrfs_set_block_group_rw(struct btrfs_root *root,
7341                               struct btrfs_block_group_cache *cache)
7342 {
7343         struct btrfs_space_info *sinfo = cache->space_info;
7344         u64 num_bytes;
7345
7346         BUG_ON(!cache->ro);
7347
7348         spin_lock(&sinfo->lock);
7349         spin_lock(&cache->lock);
7350         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7351                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7352         sinfo->bytes_readonly -= num_bytes;
7353         cache->ro = 0;
7354         spin_unlock(&cache->lock);
7355         spin_unlock(&sinfo->lock);
7356 }
7357
7358 /*
7359  * checks to see if its even possible to relocate this block group.
7360  *
7361  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7362  * ok to go ahead and try.
7363  */
7364 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7365 {
7366         struct btrfs_block_group_cache *block_group;
7367         struct btrfs_space_info *space_info;
7368         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7369         struct btrfs_device *device;
7370         u64 min_free;
7371         u64 dev_min = 1;
7372         u64 dev_nr = 0;
7373         u64 target;
7374         int index;
7375         int full = 0;
7376         int ret = 0;
7377
7378         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7379
7380         /* odd, couldn't find the block group, leave it alone */
7381         if (!block_group)
7382                 return -1;
7383
7384         min_free = btrfs_block_group_used(&block_group->item);
7385
7386         /* no bytes used, we're good */
7387         if (!min_free)
7388                 goto out;
7389
7390         space_info = block_group->space_info;
7391         spin_lock(&space_info->lock);
7392
7393         full = space_info->full;
7394
7395         /*
7396          * if this is the last block group we have in this space, we can't
7397          * relocate it unless we're able to allocate a new chunk below.
7398          *
7399          * Otherwise, we need to make sure we have room in the space to handle
7400          * all of the extents from this block group.  If we can, we're good
7401          */
7402         if ((space_info->total_bytes != block_group->key.offset) &&
7403             (space_info->bytes_used + space_info->bytes_reserved +
7404              space_info->bytes_pinned + space_info->bytes_readonly +
7405              min_free < space_info->total_bytes)) {
7406                 spin_unlock(&space_info->lock);
7407                 goto out;
7408         }
7409         spin_unlock(&space_info->lock);
7410
7411         /*
7412          * ok we don't have enough space, but maybe we have free space on our
7413          * devices to allocate new chunks for relocation, so loop through our
7414          * alloc devices and guess if we have enough space.  if this block
7415          * group is going to be restriped, run checks against the target
7416          * profile instead of the current one.
7417          */
7418         ret = -1;
7419
7420         /*
7421          * index:
7422          *      0: raid10
7423          *      1: raid1
7424          *      2: dup
7425          *      3: raid0
7426          *      4: single
7427          */
7428         target = get_restripe_target(root->fs_info, block_group->flags);
7429         if (target) {
7430                 index = __get_block_group_index(extended_to_chunk(target));
7431         } else {
7432                 /*
7433                  * this is just a balance, so if we were marked as full
7434                  * we know there is no space for a new chunk
7435                  */
7436                 if (full)
7437                         goto out;
7438
7439                 index = get_block_group_index(block_group);
7440         }
7441
7442         if (index == 0) {
7443                 dev_min = 4;
7444                 /* Divide by 2 */
7445                 min_free >>= 1;
7446         } else if (index == 1) {
7447                 dev_min = 2;
7448         } else if (index == 2) {
7449                 /* Multiply by 2 */
7450                 min_free <<= 1;
7451         } else if (index == 3) {
7452                 dev_min = fs_devices->rw_devices;
7453                 do_div(min_free, dev_min);
7454         }
7455
7456         mutex_lock(&root->fs_info->chunk_mutex);
7457         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7458                 u64 dev_offset;
7459
7460                 /*
7461                  * check to make sure we can actually find a chunk with enough
7462                  * space to fit our block group in.
7463                  */
7464                 if (device->total_bytes > device->bytes_used + min_free) {
7465                         ret = find_free_dev_extent(device, min_free,
7466                                                    &dev_offset, NULL);
7467                         if (!ret)
7468                                 dev_nr++;
7469
7470                         if (dev_nr >= dev_min)
7471                                 break;
7472
7473                         ret = -1;
7474                 }
7475         }
7476         mutex_unlock(&root->fs_info->chunk_mutex);
7477 out:
7478         btrfs_put_block_group(block_group);
7479         return ret;
7480 }
7481
7482 static int find_first_block_group(struct btrfs_root *root,
7483                 struct btrfs_path *path, struct btrfs_key *key)
7484 {
7485         int ret = 0;
7486         struct btrfs_key found_key;
7487         struct extent_buffer *leaf;
7488         int slot;
7489
7490         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7491         if (ret < 0)
7492                 goto out;
7493
7494         while (1) {
7495                 slot = path->slots[0];
7496                 leaf = path->nodes[0];
7497                 if (slot >= btrfs_header_nritems(leaf)) {
7498                         ret = btrfs_next_leaf(root, path);
7499                         if (ret == 0)
7500                                 continue;
7501                         if (ret < 0)
7502                                 goto out;
7503                         break;
7504                 }
7505                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7506
7507                 if (found_key.objectid >= key->objectid &&
7508                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7509                         ret = 0;
7510                         goto out;
7511                 }
7512                 path->slots[0]++;
7513         }
7514 out:
7515         return ret;
7516 }
7517
7518 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7519 {
7520         struct btrfs_block_group_cache *block_group;
7521         u64 last = 0;
7522
7523         while (1) {
7524                 struct inode *inode;
7525
7526                 block_group = btrfs_lookup_first_block_group(info, last);
7527                 while (block_group) {
7528                         spin_lock(&block_group->lock);
7529                         if (block_group->iref)
7530                                 break;
7531                         spin_unlock(&block_group->lock);
7532                         block_group = next_block_group(info->tree_root,
7533                                                        block_group);
7534                 }
7535                 if (!block_group) {
7536                         if (last == 0)
7537                                 break;
7538                         last = 0;
7539                         continue;
7540                 }
7541
7542                 inode = block_group->inode;
7543                 block_group->iref = 0;
7544                 block_group->inode = NULL;
7545                 spin_unlock(&block_group->lock);
7546                 iput(inode);
7547                 last = block_group->key.objectid + block_group->key.offset;
7548                 btrfs_put_block_group(block_group);
7549         }
7550 }
7551
7552 int btrfs_free_block_groups(struct btrfs_fs_info *info)
7553 {
7554         struct btrfs_block_group_cache *block_group;
7555         struct btrfs_space_info *space_info;
7556         struct btrfs_caching_control *caching_ctl;
7557         struct rb_node *n;
7558
7559         down_write(&info->extent_commit_sem);
7560         while (!list_empty(&info->caching_block_groups)) {
7561                 caching_ctl = list_entry(info->caching_block_groups.next,
7562                                          struct btrfs_caching_control, list);
7563                 list_del(&caching_ctl->list);
7564                 put_caching_control(caching_ctl);
7565         }
7566         up_write(&info->extent_commit_sem);
7567
7568         spin_lock(&info->block_group_cache_lock);
7569         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
7570                 block_group = rb_entry(n, struct btrfs_block_group_cache,
7571                                        cache_node);
7572                 rb_erase(&block_group->cache_node,
7573                          &info->block_group_cache_tree);
7574                 spin_unlock(&info->block_group_cache_lock);
7575
7576                 down_write(&block_group->space_info->groups_sem);
7577                 list_del(&block_group->list);
7578                 up_write(&block_group->space_info->groups_sem);
7579
7580                 if (block_group->cached == BTRFS_CACHE_STARTED)
7581                         wait_block_group_cache_done(block_group);
7582
7583                 /*
7584                  * We haven't cached this block group, which means we could
7585                  * possibly have excluded extents on this block group.
7586                  */
7587                 if (block_group->cached == BTRFS_CACHE_NO)
7588                         free_excluded_extents(info->extent_root, block_group);
7589
7590                 btrfs_remove_free_space_cache(block_group);
7591                 btrfs_put_block_group(block_group);
7592
7593                 spin_lock(&info->block_group_cache_lock);
7594         }
7595         spin_unlock(&info->block_group_cache_lock);
7596
7597         /* now that all the block groups are freed, go through and
7598          * free all the space_info structs.  This is only called during
7599          * the final stages of unmount, and so we know nobody is
7600          * using them.  We call synchronize_rcu() once before we start,
7601          * just to be on the safe side.
7602          */
7603         synchronize_rcu();
7604
7605         release_global_block_rsv(info);
7606
7607         while(!list_empty(&info->space_info)) {
7608                 space_info = list_entry(info->space_info.next,
7609                                         struct btrfs_space_info,
7610                                         list);
7611                 if (space_info->bytes_pinned > 0 ||
7612                     space_info->bytes_reserved > 0 ||
7613                     space_info->bytes_may_use > 0) {
7614                         WARN_ON(1);
7615                         dump_space_info(space_info, 0, 0);
7616                 }
7617                 list_del(&space_info->list);
7618                 kfree(space_info);
7619         }
7620         return 0;
7621 }
7622
7623 static void __link_block_group(struct btrfs_space_info *space_info,
7624                                struct btrfs_block_group_cache *cache)
7625 {
7626         int index = get_block_group_index(cache);
7627
7628         down_write(&space_info->groups_sem);
7629         list_add_tail(&cache->list, &space_info->block_groups[index]);
7630         up_write(&space_info->groups_sem);
7631 }
7632
7633 int btrfs_read_block_groups(struct btrfs_root *root)
7634 {
7635         struct btrfs_path *path;
7636         int ret;
7637         struct btrfs_block_group_cache *cache;
7638         struct btrfs_fs_info *info = root->fs_info;
7639         struct btrfs_space_info *space_info;
7640         struct btrfs_key key;
7641         struct btrfs_key found_key;
7642         struct extent_buffer *leaf;
7643         int need_clear = 0;
7644         u64 cache_gen;
7645
7646         root = info->extent_root;
7647         key.objectid = 0;
7648         key.offset = 0;
7649         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
7650         path = btrfs_alloc_path();
7651         if (!path)
7652                 return -ENOMEM;
7653         path->reada = 1;
7654
7655         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
7656         if (btrfs_test_opt(root, SPACE_CACHE) &&
7657             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
7658                 need_clear = 1;
7659         if (btrfs_test_opt(root, CLEAR_CACHE))
7660                 need_clear = 1;
7661
7662         while (1) {
7663                 ret = find_first_block_group(root, path, &key);
7664                 if (ret > 0)
7665                         break;
7666                 if (ret != 0)
7667                         goto error;
7668                 leaf = path->nodes[0];
7669                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
7670                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
7671                 if (!cache) {
7672                         ret = -ENOMEM;
7673                         goto error;
7674                 }
7675                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7676                                                 GFP_NOFS);
7677                 if (!cache->free_space_ctl) {
7678                         kfree(cache);
7679                         ret = -ENOMEM;
7680                         goto error;
7681                 }
7682
7683                 atomic_set(&cache->count, 1);
7684                 spin_lock_init(&cache->lock);
7685                 cache->fs_info = info;
7686                 INIT_LIST_HEAD(&cache->list);
7687                 INIT_LIST_HEAD(&cache->cluster_list);
7688
7689                 if (need_clear) {
7690                         /*
7691                          * When we mount with old space cache, we need to
7692                          * set BTRFS_DC_CLEAR and set dirty flag.
7693                          *
7694                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
7695                          *    truncate the old free space cache inode and
7696                          *    setup a new one.
7697                          * b) Setting 'dirty flag' makes sure that we flush
7698                          *    the new space cache info onto disk.
7699                          */
7700                         cache->disk_cache_state = BTRFS_DC_CLEAR;
7701                         if (btrfs_test_opt(root, SPACE_CACHE))
7702                                 cache->dirty = 1;
7703                 }
7704
7705                 read_extent_buffer(leaf, &cache->item,
7706                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
7707                                    sizeof(cache->item));
7708                 memcpy(&cache->key, &found_key, sizeof(found_key));
7709
7710                 key.objectid = found_key.objectid + found_key.offset;
7711                 btrfs_release_path(path);
7712                 cache->flags = btrfs_block_group_flags(&cache->item);
7713                 cache->sectorsize = root->sectorsize;
7714
7715                 btrfs_init_free_space_ctl(cache);
7716
7717                 /*
7718                  * We need to exclude the super stripes now so that the space
7719                  * info has super bytes accounted for, otherwise we'll think
7720                  * we have more space than we actually do.
7721                  */
7722                 exclude_super_stripes(root, cache);
7723
7724                 /*
7725                  * check for two cases, either we are full, and therefore
7726                  * don't need to bother with the caching work since we won't
7727                  * find any space, or we are empty, and we can just add all
7728                  * the space in and be done with it.  This saves us _alot_ of
7729                  * time, particularly in the full case.
7730                  */
7731                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
7732                         cache->last_byte_to_unpin = (u64)-1;
7733                         cache->cached = BTRFS_CACHE_FINISHED;
7734                         free_excluded_extents(root, cache);
7735                 } else if (btrfs_block_group_used(&cache->item) == 0) {
7736                         cache->last_byte_to_unpin = (u64)-1;
7737                         cache->cached = BTRFS_CACHE_FINISHED;
7738                         add_new_free_space(cache, root->fs_info,
7739                                            found_key.objectid,
7740                                            found_key.objectid +
7741                                            found_key.offset);
7742                         free_excluded_extents(root, cache);
7743                 }
7744
7745                 ret = update_space_info(info, cache->flags, found_key.offset,
7746                                         btrfs_block_group_used(&cache->item),
7747                                         &space_info);
7748                 BUG_ON(ret); /* -ENOMEM */
7749                 cache->space_info = space_info;
7750                 spin_lock(&cache->space_info->lock);
7751                 cache->space_info->bytes_readonly += cache->bytes_super;
7752                 spin_unlock(&cache->space_info->lock);
7753
7754                 __link_block_group(space_info, cache);
7755
7756                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
7757                 BUG_ON(ret); /* Logic error */
7758
7759                 set_avail_alloc_bits(root->fs_info, cache->flags);
7760                 if (btrfs_chunk_readonly(root, cache->key.objectid))
7761                         set_block_group_ro(cache, 1);
7762         }
7763
7764         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
7765                 if (!(get_alloc_profile(root, space_info->flags) &
7766                       (BTRFS_BLOCK_GROUP_RAID10 |
7767                        BTRFS_BLOCK_GROUP_RAID1 |
7768                        BTRFS_BLOCK_GROUP_DUP)))
7769                         continue;
7770                 /*
7771                  * avoid allocating from un-mirrored block group if there are
7772                  * mirrored block groups.
7773                  */
7774                 list_for_each_entry(cache, &space_info->block_groups[3], list)
7775                         set_block_group_ro(cache, 1);
7776                 list_for_each_entry(cache, &space_info->block_groups[4], list)
7777                         set_block_group_ro(cache, 1);
7778         }
7779
7780         init_global_block_rsv(info);
7781         ret = 0;
7782 error:
7783         btrfs_free_path(path);
7784         return ret;
7785 }
7786
7787 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
7788                                        struct btrfs_root *root)
7789 {
7790         struct btrfs_block_group_cache *block_group, *tmp;
7791         struct btrfs_root *extent_root = root->fs_info->extent_root;
7792         struct btrfs_block_group_item item;
7793         struct btrfs_key key;
7794         int ret = 0;
7795
7796         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs,
7797                                  new_bg_list) {
7798                 list_del_init(&block_group->new_bg_list);
7799
7800                 if (ret)
7801                         continue;
7802
7803                 spin_lock(&block_group->lock);
7804                 memcpy(&item, &block_group->item, sizeof(item));
7805                 memcpy(&key, &block_group->key, sizeof(key));
7806                 spin_unlock(&block_group->lock);
7807
7808                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
7809                                         sizeof(item));
7810                 if (ret)
7811                         btrfs_abort_transaction(trans, extent_root, ret);
7812         }
7813 }
7814
7815 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
7816                            struct btrfs_root *root, u64 bytes_used,
7817                            u64 type, u64 chunk_objectid, u64 chunk_offset,
7818                            u64 size)
7819 {
7820         int ret;
7821         struct btrfs_root *extent_root;
7822         struct btrfs_block_group_cache *cache;
7823
7824         extent_root = root->fs_info->extent_root;
7825
7826         root->fs_info->last_trans_log_full_commit = trans->transid;
7827
7828         cache = kzalloc(sizeof(*cache), GFP_NOFS);
7829         if (!cache)
7830                 return -ENOMEM;
7831         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
7832                                         GFP_NOFS);
7833         if (!cache->free_space_ctl) {
7834                 kfree(cache);
7835                 return -ENOMEM;
7836         }
7837
7838         cache->key.objectid = chunk_offset;
7839         cache->key.offset = size;
7840         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
7841         cache->sectorsize = root->sectorsize;
7842         cache->fs_info = root->fs_info;
7843
7844         atomic_set(&cache->count, 1);
7845         spin_lock_init(&cache->lock);
7846         INIT_LIST_HEAD(&cache->list);
7847         INIT_LIST_HEAD(&cache->cluster_list);
7848         INIT_LIST_HEAD(&cache->new_bg_list);
7849
7850         btrfs_init_free_space_ctl(cache);
7851
7852         btrfs_set_block_group_used(&cache->item, bytes_used);
7853         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
7854         cache->flags = type;
7855         btrfs_set_block_group_flags(&cache->item, type);
7856
7857         cache->last_byte_to_unpin = (u64)-1;
7858         cache->cached = BTRFS_CACHE_FINISHED;
7859         exclude_super_stripes(root, cache);
7860
7861         add_new_free_space(cache, root->fs_info, chunk_offset,
7862                            chunk_offset + size);
7863
7864         free_excluded_extents(root, cache);
7865
7866         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
7867                                 &cache->space_info);
7868         BUG_ON(ret); /* -ENOMEM */
7869         update_global_block_rsv(root->fs_info);
7870
7871         spin_lock(&cache->space_info->lock);
7872         cache->space_info->bytes_readonly += cache->bytes_super;
7873         spin_unlock(&cache->space_info->lock);
7874
7875         __link_block_group(cache->space_info, cache);
7876
7877         ret = btrfs_add_block_group_cache(root->fs_info, cache);
7878         BUG_ON(ret); /* Logic error */
7879
7880         list_add_tail(&cache->new_bg_list, &trans->new_bgs);
7881
7882         set_avail_alloc_bits(extent_root->fs_info, type);
7883
7884         return 0;
7885 }
7886
7887 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
7888 {
7889         u64 extra_flags = chunk_to_extended(flags) &
7890                                 BTRFS_EXTENDED_PROFILE_MASK;
7891
7892         if (flags & BTRFS_BLOCK_GROUP_DATA)
7893                 fs_info->avail_data_alloc_bits &= ~extra_flags;
7894         if (flags & BTRFS_BLOCK_GROUP_METADATA)
7895                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
7896         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
7897                 fs_info->avail_system_alloc_bits &= ~extra_flags;
7898 }
7899
7900 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
7901                              struct btrfs_root *root, u64 group_start)
7902 {
7903         struct btrfs_path *path;
7904         struct btrfs_block_group_cache *block_group;
7905         struct btrfs_free_cluster *cluster;
7906         struct btrfs_root *tree_root = root->fs_info->tree_root;
7907         struct btrfs_key key;
7908         struct inode *inode;
7909         int ret;
7910         int index;
7911         int factor;
7912
7913         root = root->fs_info->extent_root;
7914
7915         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
7916         BUG_ON(!block_group);
7917         BUG_ON(!block_group->ro);
7918
7919         /*
7920          * Free the reserved super bytes from this block group before
7921          * remove it.
7922          */
7923         free_excluded_extents(root, block_group);
7924
7925         memcpy(&key, &block_group->key, sizeof(key));
7926         index = get_block_group_index(block_group);
7927         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
7928                                   BTRFS_BLOCK_GROUP_RAID1 |
7929                                   BTRFS_BLOCK_GROUP_RAID10))
7930                 factor = 2;
7931         else
7932                 factor = 1;
7933
7934         /* make sure this block group isn't part of an allocation cluster */
7935         cluster = &root->fs_info->data_alloc_cluster;
7936         spin_lock(&cluster->refill_lock);
7937         btrfs_return_cluster_to_free_space(block_group, cluster);
7938         spin_unlock(&cluster->refill_lock);
7939
7940         /*
7941          * make sure this block group isn't part of a metadata
7942          * allocation cluster
7943          */
7944         cluster = &root->fs_info->meta_alloc_cluster;
7945         spin_lock(&cluster->refill_lock);
7946         btrfs_return_cluster_to_free_space(block_group, cluster);
7947         spin_unlock(&cluster->refill_lock);
7948
7949         path = btrfs_alloc_path();
7950         if (!path) {
7951                 ret = -ENOMEM;
7952                 goto out;
7953         }
7954
7955         inode = lookup_free_space_inode(tree_root, block_group, path);
7956         if (!IS_ERR(inode)) {
7957                 ret = btrfs_orphan_add(trans, inode);
7958                 if (ret) {
7959                         btrfs_add_delayed_iput(inode);
7960                         goto out;
7961                 }
7962                 clear_nlink(inode);
7963                 /* One for the block groups ref */
7964                 spin_lock(&block_group->lock);
7965                 if (block_group->iref) {
7966                         block_group->iref = 0;
7967                         block_group->inode = NULL;
7968                         spin_unlock(&block_group->lock);
7969                         iput(inode);
7970                 } else {
7971                         spin_unlock(&block_group->lock);
7972                 }
7973                 /* One for our lookup ref */
7974                 btrfs_add_delayed_iput(inode);
7975         }
7976
7977         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
7978         key.offset = block_group->key.objectid;
7979         key.type = 0;
7980
7981         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
7982         if (ret < 0)
7983                 goto out;
7984         if (ret > 0)
7985                 btrfs_release_path(path);
7986         if (ret == 0) {
7987                 ret = btrfs_del_item(trans, tree_root, path);
7988                 if (ret)
7989                         goto out;
7990                 btrfs_release_path(path);
7991         }
7992
7993         spin_lock(&root->fs_info->block_group_cache_lock);
7994         rb_erase(&block_group->cache_node,
7995                  &root->fs_info->block_group_cache_tree);
7996         spin_unlock(&root->fs_info->block_group_cache_lock);
7997
7998         down_write(&block_group->space_info->groups_sem);
7999         /*
8000          * we must use list_del_init so people can check to see if they
8001          * are still on the list after taking the semaphore
8002          */
8003         list_del_init(&block_group->list);
8004         if (list_empty(&block_group->space_info->block_groups[index]))
8005                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
8006         up_write(&block_group->space_info->groups_sem);
8007
8008         if (block_group->cached == BTRFS_CACHE_STARTED)
8009                 wait_block_group_cache_done(block_group);
8010
8011         btrfs_remove_free_space_cache(block_group);
8012
8013         spin_lock(&block_group->space_info->lock);
8014         block_group->space_info->total_bytes -= block_group->key.offset;
8015         block_group->space_info->bytes_readonly -= block_group->key.offset;
8016         block_group->space_info->disk_total -= block_group->key.offset * factor;
8017         spin_unlock(&block_group->space_info->lock);
8018
8019         memcpy(&key, &block_group->key, sizeof(key));
8020
8021         btrfs_clear_space_info_full(root->fs_info);
8022
8023         btrfs_put_block_group(block_group);
8024         btrfs_put_block_group(block_group);
8025
8026         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8027         if (ret > 0)
8028                 ret = -EIO;
8029         if (ret < 0)
8030                 goto out;
8031
8032         ret = btrfs_del_item(trans, root, path);
8033 out:
8034         btrfs_free_path(path);
8035         return ret;
8036 }
8037
8038 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
8039 {
8040         struct btrfs_space_info *space_info;
8041         struct btrfs_super_block *disk_super;
8042         u64 features;
8043         u64 flags;
8044         int mixed = 0;
8045         int ret;
8046
8047         disk_super = fs_info->super_copy;
8048         if (!btrfs_super_root(disk_super))
8049                 return 1;
8050
8051         features = btrfs_super_incompat_flags(disk_super);
8052         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
8053                 mixed = 1;
8054
8055         flags = BTRFS_BLOCK_GROUP_SYSTEM;
8056         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8057         if (ret)
8058                 goto out;
8059
8060         if (mixed) {
8061                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
8062                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8063         } else {
8064                 flags = BTRFS_BLOCK_GROUP_METADATA;
8065                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8066                 if (ret)
8067                         goto out;
8068
8069                 flags = BTRFS_BLOCK_GROUP_DATA;
8070                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8071         }
8072 out:
8073         return ret;
8074 }
8075
8076 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8077 {
8078         return unpin_extent_range(root, start, end);
8079 }
8080
8081 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8082                                u64 num_bytes, u64 *actual_bytes)
8083 {
8084         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8085 }
8086
8087 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8088 {
8089         struct btrfs_fs_info *fs_info = root->fs_info;
8090         struct btrfs_block_group_cache *cache = NULL;
8091         u64 group_trimmed;
8092         u64 start;
8093         u64 end;
8094         u64 trimmed = 0;
8095         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8096         int ret = 0;
8097
8098         /*
8099          * try to trim all FS space, our block group may start from non-zero.
8100          */
8101         if (range->len == total_bytes)
8102                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8103         else
8104                 cache = btrfs_lookup_block_group(fs_info, range->start);
8105
8106         while (cache) {
8107                 if (cache->key.objectid >= (range->start + range->len)) {
8108                         btrfs_put_block_group(cache);
8109                         break;
8110                 }
8111
8112                 start = max(range->start, cache->key.objectid);
8113                 end = min(range->start + range->len,
8114                                 cache->key.objectid + cache->key.offset);
8115
8116                 if (end - start >= range->minlen) {
8117                         if (!block_group_cache_done(cache)) {
8118                                 ret = cache_block_group(cache, NULL, root, 0);
8119                                 if (!ret)
8120                                         wait_block_group_cache_done(cache);
8121                         }
8122                         ret = btrfs_trim_block_group(cache,
8123                                                      &group_trimmed,
8124                                                      start,
8125                                                      end,
8126                                                      range->minlen);
8127
8128                         trimmed += group_trimmed;
8129                         if (ret) {
8130                                 btrfs_put_block_group(cache);
8131                                 break;
8132                         }
8133                 }
8134
8135                 cache = next_block_group(fs_info->tree_root, cache);
8136         }
8137
8138         range->len = trimmed;
8139         return ret;
8140 }