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 "raid56.h"
  35 #include "locking.h"
  36 #include "free-space-cache.h"
  37 #include "math.h"
  38
  39 #undef SCRAMBLE_DELAYED_REFS
  40
  41 /*
  42  * control flags for do_chunk_alloc's force field
  43  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  44  * if we really need one.
  45  *
  46  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  47  * if we have very few chunks already allocated.  This is
  48  * used as part of the clustering code to help make sure
  49  * we have a good pool of storage to cluster in, without
  50  * filling the FS with empty chunks
  51  *
  52  * CHUNK_ALLOC_FORCE means it must try to allocate one
  53  *
  54  */
  55 enum {
  56         CHUNK_ALLOC_NO_FORCE = 0,
  57         CHUNK_ALLOC_LIMITED = 1,
  58         CHUNK_ALLOC_FORCE = 2,
  59 };
  60
  61 /*
  62  * Control how reservations are dealt with.
  63  *
  64  * RESERVE_FREE - freeing a reservation.
  65  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  66  *   ENOSPC accounting
  67  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  68  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  69  */
  70 enum {
  71         RESERVE_FREE = 0,
  72         RESERVE_ALLOC = 1,
  73         RESERVE_ALLOC_NO_ACCOUNT = 2,
  74 };
  75
  76 static int update_block_group(struct btrfs_root *root,
  77                               u64 bytenr, u64 num_bytes, int alloc);
  78 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  79                                 struct btrfs_root *root,
  80                                 u64 bytenr, u64 num_bytes, u64 parent,
  81                                 u64 root_objectid, u64 owner_objectid,
  82                                 u64 owner_offset, int refs_to_drop,
  83                                 struct btrfs_delayed_extent_op *extra_op);
  84 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  85                                     struct extent_buffer *leaf,
  86                                     struct btrfs_extent_item *ei);
  87 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  88                                       struct btrfs_root *root,
  89                                       u64 parent, u64 root_objectid,
  90                                       u64 flags, u64 owner, u64 offset,
  91                                       struct btrfs_key *ins, int ref_mod);
  92 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  93                                      struct btrfs_root *root,
  94                                      u64 parent, u64 root_objectid,
  95                                      u64 flags, struct btrfs_disk_key *key,
  96                                      int level, struct btrfs_key *ins);
  97 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
  98                           struct btrfs_root *extent_root, u64 flags,
  99                           int force);
 100 static int find_next_key(struct btrfs_path *path, int level,
 101                          struct btrfs_key *key);
 102 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 103                             int dump_block_groups);
 104 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 105                                        u64 num_bytes, int reserve);
 106 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 107                                u64 num_bytes);
 108
 109 static noinline int
 110 block_group_cache_done(struct btrfs_block_group_cache *cache)
 111 {
 112         smp_mb();
 113         return cache->cached == BTRFS_CACHE_FINISHED;
 114 }
 115
 116 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 117 {
 118         return (cache->flags & bits) == bits;
 119 }
 120
 121 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 122 {
 123         atomic_inc(&cache->count);
 124 }
 125
 126 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 127 {
 128         if (atomic_dec_and_test(&cache->count)) {
 129                 WARN_ON(cache->pinned > 0);
 130                 WARN_ON(cache->reserved > 0);
 131                 kfree(cache->free_space_ctl);
 132                 kfree(cache);
 133         }
 134 }
 135
 136 /*
 137  * this adds the block group to the fs_info rb tree for the block group
 138  * cache
 139  */
 140 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 141                                 struct btrfs_block_group_cache *block_group)
 142 {
 143         struct rb_node **p;
 144         struct rb_node *parent = NULL;
 145         struct btrfs_block_group_cache *cache;
 146
 147         spin_lock(&info->block_group_cache_lock);
 148         p = &info->block_group_cache_tree.rb_node;
 149
 150         while (*p) {
 151                 parent = *p;
 152                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 153                                  cache_node);
 154                 if (block_group->key.objectid < cache->key.objectid) {
 155                         p = &(*p)->rb_left;
 156                 } else if (block_group->key.objectid > cache->key.objectid) {
 157                         p = &(*p)->rb_right;
 158                 } else {
 159                         spin_unlock(&info->block_group_cache_lock);
 160                         return -EEXIST;
 161                 }
 162         }
 163
 164         rb_link_node(&block_group->cache_node, parent, p);
 165         rb_insert_color(&block_group->cache_node,
 166                         &info->block_group_cache_tree);
 167
 168         if (info->first_logical_byte > block_group->key.objectid)
 169                 info->first_logical_byte = block_group->key.objectid;
 170
 171         spin_unlock(&info->block_group_cache_lock);
 172
 173         return 0;
 174 }
 175
 176 /*
 177  * This will return the block group at or after bytenr if contains is 0, else
 178  * it will return the block group that contains the bytenr
 179  */
 180 static struct btrfs_block_group_cache *
 181 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 182                               int contains)
 183 {
 184         struct btrfs_block_group_cache *cache, *ret = NULL;
 185         struct rb_node *n;
 186         u64 end, start;
 187
 188         spin_lock(&info->block_group_cache_lock);
 189         n = info->block_group_cache_tree.rb_node;
 190
 191         while (n) {
 192                 cache = rb_entry(n, struct btrfs_block_group_cache,
 193                                  cache_node);
 194                 end = cache->key.objectid + cache->key.offset - 1;
 195                 start = cache->key.objectid;
 196
 197                 if (bytenr < start) {
 198                         if (!contains && (!ret || start < ret->key.objectid))
 199                                 ret = cache;
 200                         n = n->rb_left;
 201                 } else if (bytenr > start) {
 202                         if (contains && bytenr <= end) {
 203                                 ret = cache;
 204                                 break;
 205                         }
 206                         n = n->rb_right;
 207                 } else {
 208                         ret = cache;
 209                         break;
 210                 }
 211         }
 212         if (ret) {
 213                 btrfs_get_block_group(ret);
 214                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 215                         info->first_logical_byte = ret->key.objectid;
 216         }
 217         spin_unlock(&info->block_group_cache_lock);
 218
 219         return ret;
 220 }
 221
 222 static int add_excluded_extent(struct btrfs_root *root,
 223                                u64 start, u64 num_bytes)
 224 {
 225         u64 end = start + num_bytes - 1;
 226         set_extent_bits(&root->fs_info->freed_extents[0],
 227                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 228         set_extent_bits(&root->fs_info->freed_extents[1],
 229                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 230         return 0;
 231 }
 232
 233 static void free_excluded_extents(struct btrfs_root *root,
 234                                   struct btrfs_block_group_cache *cache)
 235 {
 236         u64 start, end;
 237
 238         start = cache->key.objectid;
 239         end = start + cache->key.offset - 1;
 240
 241         clear_extent_bits(&root->fs_info->freed_extents[0],
 242                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 243         clear_extent_bits(&root->fs_info->freed_extents[1],
 244                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 245 }
 246
 247 static int exclude_super_stripes(struct btrfs_root *root,
 248                                  struct btrfs_block_group_cache *cache)
 249 {
 250         u64 bytenr;
 251         u64 *logical;
 252         int stripe_len;
 253         int i, nr, ret;
 254
 255         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 256                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 257                 cache->bytes_super += stripe_len;
 258                 ret = add_excluded_extent(root, cache->key.objectid,
 259                                           stripe_len);
 260                 if (ret)
 261                         return ret;
 262         }
 263
 264         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 265                 bytenr = btrfs_sb_offset(i);
 266                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 267                                        cache->key.objectid, bytenr,
 268                                        0, &logical, &nr, &stripe_len);
 269                 if (ret)
 270                         return ret;
 271
 272                 while (nr--) {
 273                         u64 start, len;
 274
 275                         if (logical[nr] > cache->key.objectid +
 276                             cache->key.offset)
 277                                 continue;
 278
 279                         if (logical[nr] + stripe_len <= cache->key.objectid)
 280                                 continue;
 281
 282                         start = logical[nr];
 283                         if (start < cache->key.objectid) {
 284                                 start = cache->key.objectid;
 285                                 len = (logical[nr] + stripe_len) - start;
 286                         } else {
 287                                 len = min_t(u64, stripe_len,
 288                                             cache->key.objectid +
 289                                             cache->key.offset - start);
 290                         }
 291
 292                         cache->bytes_super += len;
 293                         ret = add_excluded_extent(root, start, len);
 294                         if (ret) {
 295                                 kfree(logical);
 296                                 return ret;
 297                         }
 298                 }
 299
 300                 kfree(logical);
 301         }
 302         return 0;
 303 }
 304
 305 static struct btrfs_caching_control *
 306 get_caching_control(struct btrfs_block_group_cache *cache)
 307 {
 308         struct btrfs_caching_control *ctl;
 309
 310         spin_lock(&cache->lock);
 311         if (cache->cached != BTRFS_CACHE_STARTED) {
 312                 spin_unlock(&cache->lock);
 313                 return NULL;
 314         }
 315
 316         /* We're loading it the fast way, so we don't have a caching_ctl. */
 317         if (!cache->caching_ctl) {
 318                 spin_unlock(&cache->lock);
 319                 return NULL;
 320         }
 321
 322         ctl = cache->caching_ctl;
 323         atomic_inc(&ctl->count);
 324         spin_unlock(&cache->lock);
 325         return ctl;
 326 }
 327
 328 static void put_caching_control(struct btrfs_caching_control *ctl)
 329 {
 330         if (atomic_dec_and_test(&ctl->count))
 331                 kfree(ctl);
 332 }
 333
 334 /*
 335  * this is only called by cache_block_group, since we could have freed extents
 336  * we need to check the pinned_extents for any extents that can't be used yet
 337  * since their free space will be released as soon as the transaction commits.
 338  */
 339 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 340                               struct btrfs_fs_info *info, u64 start, u64 end)
 341 {
 342         u64 extent_start, extent_end, size, total_added = 0;
 343         int ret;
 344
 345         while (start < end) {
 346                 ret = find_first_extent_bit(info->pinned_extents, start,
 347                                             &extent_start, &extent_end,
 348                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 349                                             NULL);
 350                 if (ret)
 351                         break;
 352
 353                 if (extent_start <= start) {
 354                         start = extent_end + 1;
 355                 } else if (extent_start > start && extent_start < end) {
 356                         size = extent_start - start;
 357                         total_added += size;
 358                         ret = btrfs_add_free_space(block_group, start,
 359                                                    size);
 360                         BUG_ON(ret); /* -ENOMEM or logic error */
 361                         start = extent_end + 1;
 362                 } else {
 363                         break;
 364                 }
 365         }
 366
 367         if (start < end) {
 368                 size = end - start;
 369                 total_added += size;
 370                 ret = btrfs_add_free_space(block_group, start, size);
 371                 BUG_ON(ret); /* -ENOMEM or logic error */
 372         }
 373
 374         return total_added;
 375 }
 376
 377 static noinline void caching_thread(struct btrfs_work *work)
 378 {
 379         struct btrfs_block_group_cache *block_group;
 380         struct btrfs_fs_info *fs_info;
 381         struct btrfs_caching_control *caching_ctl;
 382         struct btrfs_root *extent_root;
 383         struct btrfs_path *path;
 384         struct extent_buffer *leaf;
 385         struct btrfs_key key;
 386         u64 total_found = 0;
 387         u64 last = 0;
 388         u32 nritems;
 389         int ret = 0;
 390
 391         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 392         block_group = caching_ctl->block_group;
 393         fs_info = block_group->fs_info;
 394         extent_root = fs_info->extent_root;
 395
 396         path = btrfs_alloc_path();
 397         if (!path)
 398                 goto out;
 399
 400         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 401
 402         /*
 403          * We don't want to deadlock with somebody trying to allocate a new
 404          * extent for the extent root while also trying to search the extent
 405          * root to add free space.  So we skip locking and search the commit
 406          * root, since its read-only
 407          */
 408         path->skip_locking = 1;
 409         path->search_commit_root = 1;
 410         path->reada = 1;
 411
 412         key.objectid = last;
 413         key.offset = 0;
 414         key.type = BTRFS_EXTENT_ITEM_KEY;
 415 again:
 416         mutex_lock(&caching_ctl->mutex);
 417         /* need to make sure the commit_root doesn't disappear */
 418         down_read(&fs_info->extent_commit_sem);
 419
 420         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 421         if (ret < 0)
 422                 goto err;
 423
 424         leaf = path->nodes[0];
 425         nritems = btrfs_header_nritems(leaf);
 426
 427         while (1) {
 428                 if (btrfs_fs_closing(fs_info) > 1) {
 429                         last = (u64)-1;
 430                         break;
 431                 }
 432
 433                 if (path->slots[0] < nritems) {
 434                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 435                 } else {
 436                         ret = find_next_key(path, 0, &key);
 437                         if (ret)
 438                                 break;
 439
 440                         if (need_resched()) {
 441                                 caching_ctl->progress = last;
 442                                 btrfs_release_path(path);
 443                                 up_read(&fs_info->extent_commit_sem);
 444                                 mutex_unlock(&caching_ctl->mutex);
 445                                 cond_resched();
 446                                 goto again;
 447                         }
 448
 449                         ret = btrfs_next_leaf(extent_root, path);
 450                         if (ret < 0)
 451                                 goto err;
 452                         if (ret)
 453                                 break;
 454                         leaf = path->nodes[0];
 455                         nritems = btrfs_header_nritems(leaf);
 456                         continue;
 457                 }
 458
 459                 if (key.objectid < block_group->key.objectid) {
 460                         path->slots[0]++;
 461                         continue;
 462                 }
 463
 464                 if (key.objectid >= block_group->key.objectid +
 465                     block_group->key.offset)
 466                         break;
 467
 468                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 469                     key.type == BTRFS_METADATA_ITEM_KEY) {
 470                         total_found += add_new_free_space(block_group,
 471                                                           fs_info, last,
 472                                                           key.objectid);
 473                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 474                                 last = key.objectid +
 475                                         fs_info->tree_root->leafsize;
 476                         else
 477                                 last = key.objectid + key.offset;
 478
 479                         if (total_found > (1024 * 1024 * 2)) {
 480                                 total_found = 0;
 481                                 wake_up(&caching_ctl->wait);
 482                         }
 483                 }
 484                 path->slots[0]++;
 485         }
 486         ret = 0;
 487
 488         total_found += add_new_free_space(block_group, fs_info, last,
 489                                           block_group->key.objectid +
 490                                           block_group->key.offset);
 491         caching_ctl->progress = (u64)-1;
 492
 493         spin_lock(&block_group->lock);
 494         block_group->caching_ctl = NULL;
 495         block_group->cached = BTRFS_CACHE_FINISHED;
 496         spin_unlock(&block_group->lock);
 497
 498 err:
 499         btrfs_free_path(path);
 500         up_read(&fs_info->extent_commit_sem);
 501
 502         free_excluded_extents(extent_root, block_group);
 503
 504         mutex_unlock(&caching_ctl->mutex);
 505 out:
 506         wake_up(&caching_ctl->wait);
 507
 508         put_caching_control(caching_ctl);
 509         btrfs_put_block_group(block_group);
 510 }
 511
 512 static int cache_block_group(struct btrfs_block_group_cache *cache,
 513                              int load_cache_only)
 514 {
 515         DEFINE_WAIT(wait);
 516         struct btrfs_fs_info *fs_info = cache->fs_info;
 517         struct btrfs_caching_control *caching_ctl;
 518         int ret = 0;
 519
 520         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 521         if (!caching_ctl)
 522                 return -ENOMEM;
 523
 524         INIT_LIST_HEAD(&caching_ctl->list);
 525         mutex_init(&caching_ctl->mutex);
 526         init_waitqueue_head(&caching_ctl->wait);
 527         caching_ctl->block_group = cache;
 528         caching_ctl->progress = cache->key.objectid;
 529         atomic_set(&caching_ctl->count, 1);
 530         caching_ctl->work.func = caching_thread;
 531
 532         spin_lock(&cache->lock);
 533         /*
 534          * This should be a rare occasion, but this could happen I think in the
 535          * case where one thread starts to load the space cache info, and then
 536          * some other thread starts a transaction commit which tries to do an
 537          * allocation while the other thread is still loading the space cache
 538          * info.  The previous loop should have kept us from choosing this block
 539          * group, but if we've moved to the state where we will wait on caching
 540          * block groups we need to first check if we're doing a fast load here,
 541          * so we can wait for it to finish, otherwise we could end up allocating
 542          * from a block group who's cache gets evicted for one reason or
 543          * another.
 544          */
 545         while (cache->cached == BTRFS_CACHE_FAST) {
 546                 struct btrfs_caching_control *ctl;
 547
 548                 ctl = cache->caching_ctl;
 549                 atomic_inc(&ctl->count);
 550                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 551                 spin_unlock(&cache->lock);
 552
 553                 schedule();
 554
 555                 finish_wait(&ctl->wait, &wait);
 556                 put_caching_control(ctl);
 557                 spin_lock(&cache->lock);
 558         }
 559
 560         if (cache->cached != BTRFS_CACHE_NO) {
 561                 spin_unlock(&cache->lock);
 562                 kfree(caching_ctl);
 563                 return 0;
 564         }
 565         WARN_ON(cache->caching_ctl);
 566         cache->caching_ctl = caching_ctl;
 567         cache->cached = BTRFS_CACHE_FAST;
 568         spin_unlock(&cache->lock);
 569
 570         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 571                 ret = load_free_space_cache(fs_info, cache);
 572
 573                 spin_lock(&cache->lock);
 574                 if (ret == 1) {
 575                         cache->caching_ctl = NULL;
 576                         cache->cached = BTRFS_CACHE_FINISHED;
 577                         cache->last_byte_to_unpin = (u64)-1;
 578                 } else {
 579                         if (load_cache_only) {
 580                                 cache->caching_ctl = NULL;
 581                                 cache->cached = BTRFS_CACHE_NO;
 582                         } else {
 583                                 cache->cached = BTRFS_CACHE_STARTED;
 584                         }
 585                 }
 586                 spin_unlock(&cache->lock);
 587                 wake_up(&caching_ctl->wait);
 588                 if (ret == 1) {
 589                         put_caching_control(caching_ctl);
 590                         free_excluded_extents(fs_info->extent_root, cache);
 591                         return 0;
 592                 }
 593         } else {
 594                 /*
 595                  * We are not going to do the fast caching, set cached to the
 596                  * appropriate value and wakeup any waiters.
 597                  */
 598                 spin_lock(&cache->lock);
 599                 if (load_cache_only) {
 600                         cache->caching_ctl = NULL;
 601                         cache->cached = BTRFS_CACHE_NO;
 602                 } else {
 603                         cache->cached = BTRFS_CACHE_STARTED;
 604                 }
 605                 spin_unlock(&cache->lock);
 606                 wake_up(&caching_ctl->wait);
 607         }
 608
 609         if (load_cache_only) {
 610                 put_caching_control(caching_ctl);
 611                 return 0;
 612         }
 613
 614         down_write(&fs_info->extent_commit_sem);
 615         atomic_inc(&caching_ctl->count);
 616         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 617         up_write(&fs_info->extent_commit_sem);
 618
 619         btrfs_get_block_group(cache);
 620
 621         btrfs_queue_worker(&fs_info->caching_workers, &caching_ctl->work);
 622
 623         return ret;
 624 }
 625
 626 /*
 627  * return the block group that starts at or after bytenr
 628  */
 629 static struct btrfs_block_group_cache *
 630 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 631 {
 632         struct btrfs_block_group_cache *cache;
 633
 634         cache = block_group_cache_tree_search(info, bytenr, 0);
 635
 636         return cache;
 637 }
 638
 639 /*
 640  * return the block group that contains the given bytenr
 641  */
 642 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 643                                                  struct btrfs_fs_info *info,
 644                                                  u64 bytenr)
 645 {
 646         struct btrfs_block_group_cache *cache;
 647
 648         cache = block_group_cache_tree_search(info, bytenr, 1);
 649
 650         return cache;
 651 }
 652
 653 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 654                                                   u64 flags)
 655 {
 656         struct list_head *head = &info->space_info;
 657         struct btrfs_space_info *found;
 658
 659         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 660
 661         rcu_read_lock();
 662         list_for_each_entry_rcu(found, head, list) {
 663                 if (found->flags & flags) {
 664                         rcu_read_unlock();
 665                         return found;
 666                 }
 667         }
 668         rcu_read_unlock();
 669         return NULL;
 670 }
 671
 672 /*
 673  * after adding space to the filesystem, we need to clear the full flags
 674  * on all the space infos.
 675  */
 676 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 677 {
 678         struct list_head *head = &info->space_info;
 679         struct btrfs_space_info *found;
 680
 681         rcu_read_lock();
 682         list_for_each_entry_rcu(found, head, list)
 683                 found->full = 0;
 684         rcu_read_unlock();
 685 }
 686
 687 u64 btrfs_find_block_group(struct btrfs_root *root,
 688                            u64 search_start, u64 search_hint, int owner)
 689 {
 690         struct btrfs_block_group_cache *cache;
 691         u64 used;
 692         u64 last = max(search_hint, search_start);
 693         u64 group_start = 0;
 694         int full_search = 0;
 695         int factor = 9;
 696         int wrapped = 0;
 697 again:
 698         while (1) {
 699                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
 700                 if (!cache)
 701                         break;
 702
 703                 spin_lock(&cache->lock);
 704                 last = cache->key.objectid + cache->key.offset;
 705                 used = btrfs_block_group_used(&cache->item);
 706
 707                 if ((full_search || !cache->ro) &&
 708                     block_group_bits(cache, BTRFS_BLOCK_GROUP_METADATA)) {
 709                         if (used + cache->pinned + cache->reserved <
 710                             div_factor(cache->key.offset, factor)) {
 711                                 group_start = cache->key.objectid;
 712                                 spin_unlock(&cache->lock);
 713                                 btrfs_put_block_group(cache);
 714                                 goto found;
 715                         }
 716                 }
 717                 spin_unlock(&cache->lock);
 718                 btrfs_put_block_group(cache);
 719                 cond_resched();
 720         }
 721         if (!wrapped) {
 722                 last = search_start;
 723                 wrapped = 1;
 724                 goto again;
 725         }
 726         if (!full_search && factor < 10) {
 727                 last = search_start;
 728                 full_search = 1;
 729                 factor = 10;
 730                 goto again;
 731         }
 732 found:
 733         return group_start;
 734 }
 735
 736 /* simple helper to search for an existing extent at a given offset */
 737 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 738 {
 739         int ret;
 740         struct btrfs_key key;
 741         struct btrfs_path *path;
 742
 743         path = btrfs_alloc_path();
 744         if (!path)
 745                 return -ENOMEM;
 746
 747         key.objectid = start;
 748         key.offset = len;
 749         key.type = BTRFS_EXTENT_ITEM_KEY;
 750         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 751                                 0, 0);
 752         if (ret > 0) {
 753                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 754                 if (key.objectid == start &&
 755                     key.type == BTRFS_METADATA_ITEM_KEY)
 756                         ret = 0;
 757         }
 758         btrfs_free_path(path);
 759         return ret;
 760 }
 761
 762 /*
 763  * helper function to lookup reference count and flags of a tree block.
 764  *
 765  * the head node for delayed ref is used to store the sum of all the
 766  * reference count modifications queued up in the rbtree. the head
 767  * node may also store the extent flags to set. This way you can check
 768  * to see what the reference count and extent flags would be if all of
 769  * the delayed refs are not processed.
 770  */
 771 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 772                              struct btrfs_root *root, u64 bytenr,
 773                              u64 offset, int metadata, u64 *refs, u64 *flags)
 774 {
 775         struct btrfs_delayed_ref_head *head;
 776         struct btrfs_delayed_ref_root *delayed_refs;
 777         struct btrfs_path *path;
 778         struct btrfs_extent_item *ei;
 779         struct extent_buffer *leaf;
 780         struct btrfs_key key;
 781         u32 item_size;
 782         u64 num_refs;
 783         u64 extent_flags;
 784         int ret;
 785
 786         /*
 787          * If we don't have skinny metadata, don't bother doing anything
 788          * different
 789          */
 790         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 791                 offset = root->leafsize;
 792                 metadata = 0;
 793         }
 794
 795         path = btrfs_alloc_path();
 796         if (!path)
 797                 return -ENOMEM;
 798
 799         if (metadata) {
 800                 key.objectid = bytenr;
 801                 key.type = BTRFS_METADATA_ITEM_KEY;
 802                 key.offset = offset;
 803         } else {
 804                 key.objectid = bytenr;
 805                 key.type = BTRFS_EXTENT_ITEM_KEY;
 806                 key.offset = offset;
 807         }
 808
 809         if (!trans) {
 810                 path->skip_locking = 1;
 811                 path->search_commit_root = 1;
 812         }
 813 again:
 814         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 815                                 &key, path, 0, 0);
 816         if (ret < 0)
 817                 goto out_free;
 818
 819         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 820                 key.type = BTRFS_EXTENT_ITEM_KEY;
 821                 key.offset = root->leafsize;
 822                 btrfs_release_path(path);
 823                 goto again;
 824         }
 825
 826         if (ret == 0) {
 827                 leaf = path->nodes[0];
 828                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 829                 if (item_size >= sizeof(*ei)) {
 830                         ei = btrfs_item_ptr(leaf, path->slots[0],
 831                                             struct btrfs_extent_item);
 832                         num_refs = btrfs_extent_refs(leaf, ei);
 833                         extent_flags = btrfs_extent_flags(leaf, ei);
 834                 } else {
 835 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 836                         struct btrfs_extent_item_v0 *ei0;
 837                         BUG_ON(item_size != sizeof(*ei0));
 838                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 839                                              struct btrfs_extent_item_v0);
 840                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 841                         /* FIXME: this isn't correct for data */
 842                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 843 #else
 844                         BUG();
 845 #endif
 846                 }
 847                 BUG_ON(num_refs == 0);
 848         } else {
 849                 num_refs = 0;
 850                 extent_flags = 0;
 851                 ret = 0;
 852         }
 853
 854         if (!trans)
 855                 goto out;
 856
 857         delayed_refs = &trans->transaction->delayed_refs;
 858         spin_lock(&delayed_refs->lock);
 859         head = btrfs_find_delayed_ref_head(trans, bytenr);
 860         if (head) {
 861                 if (!mutex_trylock(&head->mutex)) {
 862                         atomic_inc(&head->node.refs);
 863                         spin_unlock(&delayed_refs->lock);
 864
 865                         btrfs_release_path(path);
 866
 867                         /*
 868                          * Mutex was contended, block until it's released and try
 869                          * again
 870                          */
 871                         mutex_lock(&head->mutex);
 872                         mutex_unlock(&head->mutex);
 873                         btrfs_put_delayed_ref(&head->node);
 874                         goto again;
 875                 }
 876                 if (head->extent_op && head->extent_op->update_flags)
 877                         extent_flags |= head->extent_op->flags_to_set;
 878                 else
 879                         BUG_ON(num_refs == 0);
 880
 881                 num_refs += head->node.ref_mod;
 882                 mutex_unlock(&head->mutex);
 883         }
 884         spin_unlock(&delayed_refs->lock);
 885 out:
 886         WARN_ON(num_refs == 0);
 887         if (refs)
 888                 *refs = num_refs;
 889         if (flags)
 890                 *flags = extent_flags;
 891 out_free:
 892         btrfs_free_path(path);
 893         return ret;
 894 }
 895
 896 /*
 897  * Back reference rules.  Back refs have three main goals:
 898  *
 899  * 1) differentiate between all holders of references to an extent so that
 900  *    when a reference is dropped we can make sure it was a valid reference
 901  *    before freeing the extent.
 902  *
 903  * 2) Provide enough information to quickly find the holders of an extent
 904  *    if we notice a given block is corrupted or bad.
 905  *
 906  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 907  *    maintenance.  This is actually the same as #2, but with a slightly
 908  *    different use case.
 909  *
 910  * There are two kinds of back refs. The implicit back refs is optimized
 911  * for pointers in non-shared tree blocks. For a given pointer in a block,
 912  * back refs of this kind provide information about the block's owner tree
 913  * and the pointer's key. These information allow us to find the block by
 914  * b-tree searching. The full back refs is for pointers in tree blocks not
 915  * referenced by their owner trees. The location of tree block is recorded
 916  * in the back refs. Actually the full back refs is generic, and can be
 917  * used in all cases the implicit back refs is used. The major shortcoming
 918  * of the full back refs is its overhead. Every time a tree block gets
 919  * COWed, we have to update back refs entry for all pointers in it.
 920  *
 921  * For a newly allocated tree block, we use implicit back refs for
 922  * pointers in it. This means most tree related operations only involve
 923  * implicit back refs. For a tree block created in old transaction, the
 924  * only way to drop a reference to it is COW it. So we can detect the
 925  * event that tree block loses its owner tree's reference and do the
 926  * back refs conversion.
 927  *
 928  * When a tree block is COW'd through a tree, there are four cases:
 929  *
 930  * The reference count of the block is one and the tree is the block's
 931  * owner tree. Nothing to do in this case.
 932  *
 933  * The reference count of the block is one and the tree is not the
 934  * block's owner tree. In this case, full back refs is used for pointers
 935  * in the block. Remove these full back refs, add implicit back refs for
 936  * every pointers in the new block.
 937  *
 938  * The reference count of the block is greater than one and the tree is
 939  * the block's owner tree. In this case, implicit back refs is used for
 940  * pointers in the block. Add full back refs for every pointers in the
 941  * block, increase lower level extents' reference counts. The original
 942  * implicit back refs are entailed to the new block.
 943  *
 944  * The reference count of the block is greater than one and the tree is
 945  * not the block's owner tree. Add implicit back refs for every pointer in
 946  * the new block, increase lower level extents' reference count.
 947  *
 948  * Back Reference Key composing:
 949  *
 950  * The key objectid corresponds to the first byte in the extent,
 951  * The key type is used to differentiate between types of back refs.
 952  * There are different meanings of the key offset for different types
 953  * of back refs.
 954  *
 955  * File extents can be referenced by:
 956  *
 957  * - multiple snapshots, subvolumes, or different generations in one subvol
 958  * - different files inside a single subvolume
 959  * - different offsets inside a file (bookend extents in file.c)
 960  *
 961  * The extent ref structure for the implicit back refs has fields for:
 962  *
 963  * - Objectid of the subvolume root
 964  * - objectid of the file holding the reference
 965  * - original offset in the file
 966  * - how many bookend extents
 967  *
 968  * The key offset for the implicit back refs is hash of the first
 969  * three fields.
 970  *
 971  * The extent ref structure for the full back refs has field for:
 972  *
 973  * - number of pointers in the tree leaf
 974  *
 975  * The key offset for the implicit back refs is the first byte of
 976  * the tree leaf
 977  *
 978  * When a file extent is allocated, The implicit back refs is used.
 979  * the fields are filled in:
 980  *
 981  *     (root_key.objectid, inode objectid, offset in file, 1)
 982  *
 983  * When a file extent is removed file truncation, we find the
 984  * corresponding implicit back refs and check the following fields:
 985  *
 986  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 987  *
 988  * Btree extents can be referenced by:
 989  *
 990  * - Different subvolumes
 991  *
 992  * Both the implicit back refs and the full back refs for tree blocks
 993  * only consist of key. The key offset for the implicit back refs is
 994  * objectid of block's owner tree. The key offset for the full back refs
 995  * is the first byte of parent block.
 996  *
 997  * When implicit back refs is used, information about the lowest key and
 998  * level of the tree block are required. These information are stored in
 999  * tree block info structure.
1000  */
1001
1002 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1003 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1004                                   struct btrfs_root *root,
1005                                   struct btrfs_path *path,
1006                                   u64 owner, u32 extra_size)
1007 {
1008         struct btrfs_extent_item *item;
1009         struct btrfs_extent_item_v0 *ei0;
1010         struct btrfs_extent_ref_v0 *ref0;
1011         struct btrfs_tree_block_info *bi;
1012         struct extent_buffer *leaf;
1013         struct btrfs_key key;
1014         struct btrfs_key found_key;
1015         u32 new_size = sizeof(*item);
1016         u64 refs;
1017         int ret;
1018
1019         leaf = path->nodes[0];
1020         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1021
1022         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1023         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1024                              struct btrfs_extent_item_v0);
1025         refs = btrfs_extent_refs_v0(leaf, ei0);
1026
1027         if (owner == (u64)-1) {
1028                 while (1) {
1029                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1030                                 ret = btrfs_next_leaf(root, path);
1031                                 if (ret < 0)
1032                                         return ret;
1033                                 BUG_ON(ret > 0); /* Corruption */
1034                                 leaf = path->nodes[0];
1035                         }
1036                         btrfs_item_key_to_cpu(leaf, &found_key,
1037                                               path->slots[0]);
1038                         BUG_ON(key.objectid != found_key.objectid);
1039                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1040                                 path->slots[0]++;
1041                                 continue;
1042                         }
1043                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1044                                               struct btrfs_extent_ref_v0);
1045                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1046                         break;
1047                 }
1048         }
1049         btrfs_release_path(path);
1050
1051         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1052                 new_size += sizeof(*bi);
1053
1054         new_size -= sizeof(*ei0);
1055         ret = btrfs_search_slot(trans, root, &key, path,
1056                                 new_size + extra_size, 1);
1057         if (ret < 0)
1058                 return ret;
1059         BUG_ON(ret); /* Corruption */
1060
1061         btrfs_extend_item(root, path, new_size);
1062
1063         leaf = path->nodes[0];
1064         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1065         btrfs_set_extent_refs(leaf, item, refs);
1066         /* FIXME: get real generation */
1067         btrfs_set_extent_generation(leaf, item, 0);
1068         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1069                 btrfs_set_extent_flags(leaf, item,
1070                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1071                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1072                 bi = (struct btrfs_tree_block_info *)(item + 1);
1073                 /* FIXME: get first key of the block */
1074                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1075                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1076         } else {
1077                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1078         }
1079         btrfs_mark_buffer_dirty(leaf);
1080         return 0;
1081 }
1082 #endif
1083
1084 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1085 {
1086         u32 high_crc = ~(u32)0;
1087         u32 low_crc = ~(u32)0;
1088         __le64 lenum;
1089
1090         lenum = cpu_to_le64(root_objectid);
1091         high_crc = crc32c(high_crc, &lenum, sizeof(lenum));
1092         lenum = cpu_to_le64(owner);
1093         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1094         lenum = cpu_to_le64(offset);
1095         low_crc = crc32c(low_crc, &lenum, sizeof(lenum));
1096
1097         return ((u64)high_crc << 31) ^ (u64)low_crc;
1098 }
1099
1100 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1101                                      struct btrfs_extent_data_ref *ref)
1102 {
1103         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1104                                     btrfs_extent_data_ref_objectid(leaf, ref),
1105                                     btrfs_extent_data_ref_offset(leaf, ref));
1106 }
1107
1108 static int match_extent_data_ref(struct extent_buffer *leaf,
1109                                  struct btrfs_extent_data_ref *ref,
1110                                  u64 root_objectid, u64 owner, u64 offset)
1111 {
1112         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1113             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1114             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1115                 return 0;
1116         return 1;
1117 }
1118
1119 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1120                                            struct btrfs_root *root,
1121                                            struct btrfs_path *path,
1122                                            u64 bytenr, u64 parent,
1123                                            u64 root_objectid,
1124                                            u64 owner, u64 offset)
1125 {
1126         struct btrfs_key key;
1127         struct btrfs_extent_data_ref *ref;
1128         struct extent_buffer *leaf;
1129         u32 nritems;
1130         int ret;
1131         int recow;
1132         int err = -ENOENT;
1133
1134         key.objectid = bytenr;
1135         if (parent) {
1136                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1137                 key.offset = parent;
1138         } else {
1139                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1140                 key.offset = hash_extent_data_ref(root_objectid,
1141                                                   owner, offset);
1142         }
1143 again:
1144         recow = 0;
1145         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1146         if (ret < 0) {
1147                 err = ret;
1148                 goto fail;
1149         }
1150
1151         if (parent) {
1152                 if (!ret)
1153                         return 0;
1154 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1155                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1156                 btrfs_release_path(path);
1157                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1158                 if (ret < 0) {
1159                         err = ret;
1160                         goto fail;
1161                 }
1162                 if (!ret)
1163                         return 0;
1164 #endif
1165                 goto fail;
1166         }
1167
1168         leaf = path->nodes[0];
1169         nritems = btrfs_header_nritems(leaf);
1170         while (1) {
1171                 if (path->slots[0] >= nritems) {
1172                         ret = btrfs_next_leaf(root, path);
1173                         if (ret < 0)
1174                                 err = ret;
1175                         if (ret)
1176                                 goto fail;
1177
1178                         leaf = path->nodes[0];
1179                         nritems = btrfs_header_nritems(leaf);
1180                         recow = 1;
1181                 }
1182
1183                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1184                 if (key.objectid != bytenr ||
1185                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1186                         goto fail;
1187
1188                 ref = btrfs_item_ptr(leaf, path->slots[0],
1189                                      struct btrfs_extent_data_ref);
1190
1191                 if (match_extent_data_ref(leaf, ref, root_objectid,
1192                                           owner, offset)) {
1193                         if (recow) {
1194                                 btrfs_release_path(path);
1195                                 goto again;
1196                         }
1197                         err = 0;
1198                         break;
1199                 }
1200                 path->slots[0]++;
1201         }
1202 fail:
1203         return err;
1204 }
1205
1206 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1207                                            struct btrfs_root *root,
1208                                            struct btrfs_path *path,
1209                                            u64 bytenr, u64 parent,
1210                                            u64 root_objectid, u64 owner,
1211                                            u64 offset, int refs_to_add)
1212 {
1213         struct btrfs_key key;
1214         struct extent_buffer *leaf;
1215         u32 size;
1216         u32 num_refs;
1217         int ret;
1218
1219         key.objectid = bytenr;
1220         if (parent) {
1221                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1222                 key.offset = parent;
1223                 size = sizeof(struct btrfs_shared_data_ref);
1224         } else {
1225                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1226                 key.offset = hash_extent_data_ref(root_objectid,
1227                                                   owner, offset);
1228                 size = sizeof(struct btrfs_extent_data_ref);
1229         }
1230
1231         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1232         if (ret && ret != -EEXIST)
1233                 goto fail;
1234
1235         leaf = path->nodes[0];
1236         if (parent) {
1237                 struct btrfs_shared_data_ref *ref;
1238                 ref = btrfs_item_ptr(leaf, path->slots[0],
1239                                      struct btrfs_shared_data_ref);
1240                 if (ret == 0) {
1241                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1242                 } else {
1243                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1244                         num_refs += refs_to_add;
1245                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1246                 }
1247         } else {
1248                 struct btrfs_extent_data_ref *ref;
1249                 while (ret == -EEXIST) {
1250                         ref = btrfs_item_ptr(leaf, path->slots[0],
1251                                              struct btrfs_extent_data_ref);
1252                         if (match_extent_data_ref(leaf, ref, root_objectid,
1253                                                   owner, offset))
1254                                 break;
1255                         btrfs_release_path(path);
1256                         key.offset++;
1257                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1258                                                       size);
1259                         if (ret && ret != -EEXIST)
1260                                 goto fail;
1261
1262                         leaf = path->nodes[0];
1263                 }
1264                 ref = btrfs_item_ptr(leaf, path->slots[0],
1265                                      struct btrfs_extent_data_ref);
1266                 if (ret == 0) {
1267                         btrfs_set_extent_data_ref_root(leaf, ref,
1268                                                        root_objectid);
1269                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1270                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1271                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1272                 } else {
1273                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1274                         num_refs += refs_to_add;
1275                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1276                 }
1277         }
1278         btrfs_mark_buffer_dirty(leaf);
1279         ret = 0;
1280 fail:
1281         btrfs_release_path(path);
1282         return ret;
1283 }
1284
1285 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1286                                            struct btrfs_root *root,
1287                                            struct btrfs_path *path,
1288                                            int refs_to_drop)
1289 {
1290         struct btrfs_key key;
1291         struct btrfs_extent_data_ref *ref1 = NULL;
1292         struct btrfs_shared_data_ref *ref2 = NULL;
1293         struct extent_buffer *leaf;
1294         u32 num_refs = 0;
1295         int ret = 0;
1296
1297         leaf = path->nodes[0];
1298         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1299
1300         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1301                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1302                                       struct btrfs_extent_data_ref);
1303                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1304         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1305                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1306                                       struct btrfs_shared_data_ref);
1307                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1308 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1309         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1310                 struct btrfs_extent_ref_v0 *ref0;
1311                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1312                                       struct btrfs_extent_ref_v0);
1313                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1314 #endif
1315         } else {
1316                 BUG();
1317         }
1318
1319         BUG_ON(num_refs < refs_to_drop);
1320         num_refs -= refs_to_drop;
1321
1322         if (num_refs == 0) {
1323                 ret = btrfs_del_item(trans, root, path);
1324         } else {
1325                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1326                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1327                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1328                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1329 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1330                 else {
1331                         struct btrfs_extent_ref_v0 *ref0;
1332                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1333                                         struct btrfs_extent_ref_v0);
1334                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1335                 }
1336 #endif
1337                 btrfs_mark_buffer_dirty(leaf);
1338         }
1339         return ret;
1340 }
1341
1342 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1343                                           struct btrfs_path *path,
1344                                           struct btrfs_extent_inline_ref *iref)
1345 {
1346         struct btrfs_key key;
1347         struct extent_buffer *leaf;
1348         struct btrfs_extent_data_ref *ref1;
1349         struct btrfs_shared_data_ref *ref2;
1350         u32 num_refs = 0;
1351
1352         leaf = path->nodes[0];
1353         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1354         if (iref) {
1355                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1356                     BTRFS_EXTENT_DATA_REF_KEY) {
1357                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1358                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1359                 } else {
1360                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1361                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1362                 }
1363         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1364                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1365                                       struct btrfs_extent_data_ref);
1366                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1367         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1368                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1369                                       struct btrfs_shared_data_ref);
1370                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1371 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1372         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1373                 struct btrfs_extent_ref_v0 *ref0;
1374                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1375                                       struct btrfs_extent_ref_v0);
1376                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1377 #endif
1378         } else {
1379                 WARN_ON(1);
1380         }
1381         return num_refs;
1382 }
1383
1384 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1385                                           struct btrfs_root *root,
1386                                           struct btrfs_path *path,
1387                                           u64 bytenr, u64 parent,
1388                                           u64 root_objectid)
1389 {
1390         struct btrfs_key key;
1391         int ret;
1392
1393         key.objectid = bytenr;
1394         if (parent) {
1395                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1396                 key.offset = parent;
1397         } else {
1398                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1399                 key.offset = root_objectid;
1400         }
1401
1402         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1403         if (ret > 0)
1404                 ret = -ENOENT;
1405 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1406         if (ret == -ENOENT && parent) {
1407                 btrfs_release_path(path);
1408                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1409                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1410                 if (ret > 0)
1411                         ret = -ENOENT;
1412         }
1413 #endif
1414         return ret;
1415 }
1416
1417 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1418                                           struct btrfs_root *root,
1419                                           struct btrfs_path *path,
1420                                           u64 bytenr, u64 parent,
1421                                           u64 root_objectid)
1422 {
1423         struct btrfs_key key;
1424         int ret;
1425
1426         key.objectid = bytenr;
1427         if (parent) {
1428                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1429                 key.offset = parent;
1430         } else {
1431                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1432                 key.offset = root_objectid;
1433         }
1434
1435         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1436         btrfs_release_path(path);
1437         return ret;
1438 }
1439
1440 static inline int extent_ref_type(u64 parent, u64 owner)
1441 {
1442         int type;
1443         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1444                 if (parent > 0)
1445                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1446                 else
1447                         type = BTRFS_TREE_BLOCK_REF_KEY;
1448         } else {
1449                 if (parent > 0)
1450                         type = BTRFS_SHARED_DATA_REF_KEY;
1451                 else
1452                         type = BTRFS_EXTENT_DATA_REF_KEY;
1453         }
1454         return type;
1455 }
1456
1457 static int find_next_key(struct btrfs_path *path, int level,
1458                          struct btrfs_key *key)
1459
1460 {
1461         for (; level < BTRFS_MAX_LEVEL; level++) {
1462                 if (!path->nodes[level])
1463                         break;
1464                 if (path->slots[level] + 1 >=
1465                     btrfs_header_nritems(path->nodes[level]))
1466                         continue;
1467                 if (level == 0)
1468                         btrfs_item_key_to_cpu(path->nodes[level], key,
1469                                               path->slots[level] + 1);
1470                 else
1471                         btrfs_node_key_to_cpu(path->nodes[level], key,
1472                                               path->slots[level] + 1);
1473                 return 0;
1474         }
1475         return 1;
1476 }
1477
1478 /*
1479  * look for inline back ref. if back ref is found, *ref_ret is set
1480  * to the address of inline back ref, and 0 is returned.
1481  *
1482  * if back ref isn't found, *ref_ret is set to the address where it
1483  * should be inserted, and -ENOENT is returned.
1484  *
1485  * if insert is true and there are too many inline back refs, the path
1486  * points to the extent item, and -EAGAIN is returned.
1487  *
1488  * NOTE: inline back refs are ordered in the same way that back ref
1489  *       items in the tree are ordered.
1490  */
1491 static noinline_for_stack
1492 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1493                                  struct btrfs_root *root,
1494                                  struct btrfs_path *path,
1495                                  struct btrfs_extent_inline_ref **ref_ret,
1496                                  u64 bytenr, u64 num_bytes,
1497                                  u64 parent, u64 root_objectid,
1498                                  u64 owner, u64 offset, int insert)
1499 {
1500         struct btrfs_key key;
1501         struct extent_buffer *leaf;
1502         struct btrfs_extent_item *ei;
1503         struct btrfs_extent_inline_ref *iref;
1504         u64 flags;
1505         u64 item_size;
1506         unsigned long ptr;
1507         unsigned long end;
1508         int extra_size;
1509         int type;
1510         int want;
1511         int ret;
1512         int err = 0;
1513         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1514                                                  SKINNY_METADATA);
1515
1516         key.objectid = bytenr;
1517         key.type = BTRFS_EXTENT_ITEM_KEY;
1518         key.offset = num_bytes;
1519
1520         want = extent_ref_type(parent, owner);
1521         if (insert) {
1522                 extra_size = btrfs_extent_inline_ref_size(want);
1523                 path->keep_locks = 1;
1524         } else
1525                 extra_size = -1;
1526
1527         /*
1528          * Owner is our parent level, so we can just add one to get the level
1529          * for the block we are interested in.
1530          */
1531         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1532                 key.type = BTRFS_METADATA_ITEM_KEY;
1533                 key.offset = owner;
1534         }
1535
1536 again:
1537         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1538         if (ret < 0) {
1539                 err = ret;
1540                 goto out;
1541         }
1542
1543         /*
1544          * We may be a newly converted file system which still has the old fat
1545          * extent entries for metadata, so try and see if we have one of those.
1546          */
1547         if (ret > 0 && skinny_metadata) {
1548                 skinny_metadata = false;
1549                 if (path->slots[0]) {
1550                         path->slots[0]--;
1551                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1552                                               path->slots[0]);
1553                         if (key.objectid == bytenr &&
1554                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1555                             key.offset == num_bytes)
1556                                 ret = 0;
1557                 }
1558                 if (ret) {
1559                         key.type = BTRFS_EXTENT_ITEM_KEY;
1560                         key.offset = num_bytes;
1561                         btrfs_release_path(path);
1562                         goto again;
1563                 }
1564         }
1565
1566         if (ret && !insert) {
1567                 err = -ENOENT;
1568                 goto out;
1569         } else if (ret) {
1570                 err = -EIO;
1571                 WARN_ON(1);
1572                 goto out;
1573         }
1574
1575         leaf = path->nodes[0];
1576         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1577 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1578         if (item_size < sizeof(*ei)) {
1579                 if (!insert) {
1580                         err = -ENOENT;
1581                         goto out;
1582                 }
1583                 ret = convert_extent_item_v0(trans, root, path, owner,
1584                                              extra_size);
1585                 if (ret < 0) {
1586                         err = ret;
1587                         goto out;
1588                 }
1589                 leaf = path->nodes[0];
1590                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1591         }
1592 #endif
1593         BUG_ON(item_size < sizeof(*ei));
1594
1595         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1596         flags = btrfs_extent_flags(leaf, ei);
1597
1598         ptr = (unsigned long)(ei + 1);
1599         end = (unsigned long)ei + item_size;
1600
1601         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1602                 ptr += sizeof(struct btrfs_tree_block_info);
1603                 BUG_ON(ptr > end);
1604         }
1605
1606         err = -ENOENT;
1607         while (1) {
1608                 if (ptr >= end) {
1609                         WARN_ON(ptr > end);
1610                         break;
1611                 }
1612                 iref = (struct btrfs_extent_inline_ref *)ptr;
1613                 type = btrfs_extent_inline_ref_type(leaf, iref);
1614                 if (want < type)
1615                         break;
1616                 if (want > type) {
1617                         ptr += btrfs_extent_inline_ref_size(type);
1618                         continue;
1619                 }
1620
1621                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1622                         struct btrfs_extent_data_ref *dref;
1623                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1624                         if (match_extent_data_ref(leaf, dref, root_objectid,
1625                                                   owner, offset)) {
1626                                 err = 0;
1627                                 break;
1628                         }
1629                         if (hash_extent_data_ref_item(leaf, dref) <
1630                             hash_extent_data_ref(root_objectid, owner, offset))
1631                                 break;
1632                 } else {
1633                         u64 ref_offset;
1634                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1635                         if (parent > 0) {
1636                                 if (parent == ref_offset) {
1637                                         err = 0;
1638                                         break;
1639                                 }
1640                                 if (ref_offset < parent)
1641                                         break;
1642                         } else {
1643                                 if (root_objectid == ref_offset) {
1644                                         err = 0;
1645                                         break;
1646                                 }
1647                                 if (ref_offset < root_objectid)
1648                                         break;
1649                         }
1650                 }
1651                 ptr += btrfs_extent_inline_ref_size(type);
1652         }
1653         if (err == -ENOENT && insert) {
1654                 if (item_size + extra_size >=
1655                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1656                         err = -EAGAIN;
1657                         goto out;
1658                 }
1659                 /*
1660                  * To add new inline back ref, we have to make sure
1661                  * there is no corresponding back ref item.
1662                  * For simplicity, we just do not add new inline back
1663                  * ref if there is any kind of item for this block
1664                  */
1665                 if (find_next_key(path, 0, &key) == 0 &&
1666                     key.objectid == bytenr &&
1667                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1668                         err = -EAGAIN;
1669                         goto out;
1670                 }
1671         }
1672         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1673 out:
1674         if (insert) {
1675                 path->keep_locks = 0;
1676                 btrfs_unlock_up_safe(path, 1);
1677         }
1678         return err;
1679 }
1680
1681 /*
1682  * helper to add new inline back ref
1683  */
1684 static noinline_for_stack
1685 void setup_inline_extent_backref(struct btrfs_root *root,
1686                                  struct btrfs_path *path,
1687                                  struct btrfs_extent_inline_ref *iref,
1688                                  u64 parent, u64 root_objectid,
1689                                  u64 owner, u64 offset, int refs_to_add,
1690                                  struct btrfs_delayed_extent_op *extent_op)
1691 {
1692         struct extent_buffer *leaf;
1693         struct btrfs_extent_item *ei;
1694         unsigned long ptr;
1695         unsigned long end;
1696         unsigned long item_offset;
1697         u64 refs;
1698         int size;
1699         int type;
1700
1701         leaf = path->nodes[0];
1702         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1703         item_offset = (unsigned long)iref - (unsigned long)ei;
1704
1705         type = extent_ref_type(parent, owner);
1706         size = btrfs_extent_inline_ref_size(type);
1707
1708         btrfs_extend_item(root, path, size);
1709
1710         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1711         refs = btrfs_extent_refs(leaf, ei);
1712         refs += refs_to_add;
1713         btrfs_set_extent_refs(leaf, ei, refs);
1714         if (extent_op)
1715                 __run_delayed_extent_op(extent_op, leaf, ei);
1716
1717         ptr = (unsigned long)ei + item_offset;
1718         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1719         if (ptr < end - size)
1720                 memmove_extent_buffer(leaf, ptr + size, ptr,
1721                                       end - size - ptr);
1722
1723         iref = (struct btrfs_extent_inline_ref *)ptr;
1724         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1725         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1726                 struct btrfs_extent_data_ref *dref;
1727                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1728                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1729                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1730                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1731                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1732         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1733                 struct btrfs_shared_data_ref *sref;
1734                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1735                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1736                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1737         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1738                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1739         } else {
1740                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1741         }
1742         btrfs_mark_buffer_dirty(leaf);
1743 }
1744
1745 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1746                                  struct btrfs_root *root,
1747                                  struct btrfs_path *path,
1748                                  struct btrfs_extent_inline_ref **ref_ret,
1749                                  u64 bytenr, u64 num_bytes, u64 parent,
1750                                  u64 root_objectid, u64 owner, u64 offset)
1751 {
1752         int ret;
1753
1754         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1755                                            bytenr, num_bytes, parent,
1756                                            root_objectid, owner, offset, 0);
1757         if (ret != -ENOENT)
1758                 return ret;
1759
1760         btrfs_release_path(path);
1761         *ref_ret = NULL;
1762
1763         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1764                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1765                                             root_objectid);
1766         } else {
1767                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1768                                              root_objectid, owner, offset);
1769         }
1770         return ret;
1771 }
1772
1773 /*
1774  * helper to update/remove inline back ref
1775  */
1776 static noinline_for_stack
1777 void update_inline_extent_backref(struct btrfs_root *root,
1778                                   struct btrfs_path *path,
1779                                   struct btrfs_extent_inline_ref *iref,
1780                                   int refs_to_mod,
1781                                   struct btrfs_delayed_extent_op *extent_op)
1782 {
1783         struct extent_buffer *leaf;
1784         struct btrfs_extent_item *ei;
1785         struct btrfs_extent_data_ref *dref = NULL;
1786         struct btrfs_shared_data_ref *sref = NULL;
1787         unsigned long ptr;
1788         unsigned long end;
1789         u32 item_size;
1790         int size;
1791         int type;
1792         u64 refs;
1793
1794         leaf = path->nodes[0];
1795         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1796         refs = btrfs_extent_refs(leaf, ei);
1797         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1798         refs += refs_to_mod;
1799         btrfs_set_extent_refs(leaf, ei, refs);
1800         if (extent_op)
1801                 __run_delayed_extent_op(extent_op, leaf, ei);
1802
1803         type = btrfs_extent_inline_ref_type(leaf, iref);
1804
1805         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1806                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1807                 refs = btrfs_extent_data_ref_count(leaf, dref);
1808         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1809                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1810                 refs = btrfs_shared_data_ref_count(leaf, sref);
1811         } else {
1812                 refs = 1;
1813                 BUG_ON(refs_to_mod != -1);
1814         }
1815
1816         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1817         refs += refs_to_mod;
1818
1819         if (refs > 0) {
1820                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1821                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1822                 else
1823                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1824         } else {
1825                 size =  btrfs_extent_inline_ref_size(type);
1826                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1827                 ptr = (unsigned long)iref;
1828                 end = (unsigned long)ei + item_size;
1829                 if (ptr + size < end)
1830                         memmove_extent_buffer(leaf, ptr, ptr + size,
1831                                               end - ptr - size);
1832                 item_size -= size;
1833                 btrfs_truncate_item(root, path, item_size, 1);
1834         }
1835         btrfs_mark_buffer_dirty(leaf);
1836 }
1837
1838 static noinline_for_stack
1839 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1840                                  struct btrfs_root *root,
1841                                  struct btrfs_path *path,
1842                                  u64 bytenr, u64 num_bytes, u64 parent,
1843                                  u64 root_objectid, u64 owner,
1844                                  u64 offset, int refs_to_add,
1845                                  struct btrfs_delayed_extent_op *extent_op)
1846 {
1847         struct btrfs_extent_inline_ref *iref;
1848         int ret;
1849
1850         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1851                                            bytenr, num_bytes, parent,
1852                                            root_objectid, owner, offset, 1);
1853         if (ret == 0) {
1854                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1855                 update_inline_extent_backref(root, path, iref,
1856                                              refs_to_add, extent_op);
1857         } else if (ret == -ENOENT) {
1858                 setup_inline_extent_backref(root, path, iref, parent,
1859                                             root_objectid, owner, offset,
1860                                             refs_to_add, extent_op);
1861                 ret = 0;
1862         }
1863         return ret;
1864 }
1865
1866 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1867                                  struct btrfs_root *root,
1868                                  struct btrfs_path *path,
1869                                  u64 bytenr, u64 parent, u64 root_objectid,
1870                                  u64 owner, u64 offset, int refs_to_add)
1871 {
1872         int ret;
1873         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1874                 BUG_ON(refs_to_add != 1);
1875                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1876                                             parent, root_objectid);
1877         } else {
1878                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1879                                              parent, root_objectid,
1880                                              owner, offset, refs_to_add);
1881         }
1882         return ret;
1883 }
1884
1885 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1886                                  struct btrfs_root *root,
1887                                  struct btrfs_path *path,
1888                                  struct btrfs_extent_inline_ref *iref,
1889                                  int refs_to_drop, int is_data)
1890 {
1891         int ret = 0;
1892
1893         BUG_ON(!is_data && refs_to_drop != 1);
1894         if (iref) {
1895                 update_inline_extent_backref(root, path, iref,
1896                                              -refs_to_drop, NULL);
1897         } else if (is_data) {
1898                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop);
1899         } else {
1900                 ret = btrfs_del_item(trans, root, path);
1901         }
1902         return ret;
1903 }
1904
1905 static int btrfs_issue_discard(struct block_device *bdev,
1906                                 u64 start, u64 len)
1907 {
1908         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1909 }
1910
1911 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1912                                 u64 num_bytes, u64 *actual_bytes)
1913 {
1914         int ret;
1915         u64 discarded_bytes = 0;
1916         struct btrfs_bio *bbio = NULL;
1917
1918
1919         /* Tell the block device(s) that the sectors can be discarded */
1920         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1921                               bytenr, &num_bytes, &bbio, 0);
1922         /* Error condition is -ENOMEM */
1923         if (!ret) {
1924                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1925                 int i;
1926
1927
1928                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1929                         if (!stripe->dev->can_discard)
1930                                 continue;
1931
1932                         ret = btrfs_issue_discard(stripe->dev->bdev,
1933                                                   stripe->physical,
1934                                                   stripe->length);
1935                         if (!ret)
1936                                 discarded_bytes += stripe->length;
1937                         else if (ret != -EOPNOTSUPP)
1938                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1939
1940                         /*
1941                          * Just in case we get back EOPNOTSUPP for some reason,
1942                          * just ignore the return value so we don't screw up
1943                          * people calling discard_extent.
1944                          */
1945                         ret = 0;
1946                 }
1947                 kfree(bbio);
1948         }
1949
1950         if (actual_bytes)
1951                 *actual_bytes = discarded_bytes;
1952
1953
1954         if (ret == -EOPNOTSUPP)
1955                 ret = 0;
1956         return ret;
1957 }
1958
1959 /* Can return -ENOMEM */
1960 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1961                          struct btrfs_root *root,
1962                          u64 bytenr, u64 num_bytes, u64 parent,
1963                          u64 root_objectid, u64 owner, u64 offset, int for_cow)
1964 {
1965         int ret;
1966         struct btrfs_fs_info *fs_info = root->fs_info;
1967
1968         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1969                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1970
1971         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1972                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1973                                         num_bytes,
1974                                         parent, root_objectid, (int)owner,
1975                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1976         } else {
1977                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1978                                         num_bytes,
1979                                         parent, root_objectid, owner, offset,
1980                                         BTRFS_ADD_DELAYED_REF, NULL, for_cow);
1981         }
1982         return ret;
1983 }
1984
1985 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1986                                   struct btrfs_root *root,
1987                                   u64 bytenr, u64 num_bytes,
1988                                   u64 parent, u64 root_objectid,
1989                                   u64 owner, u64 offset, int refs_to_add,
1990                                   struct btrfs_delayed_extent_op *extent_op)
1991 {
1992         struct btrfs_path *path;
1993         struct extent_buffer *leaf;
1994         struct btrfs_extent_item *item;
1995         u64 refs;
1996         int ret;
1997         int err = 0;
1998
1999         path = btrfs_alloc_path();
2000         if (!path)
2001                 return -ENOMEM;
2002
2003         path->reada = 1;
2004         path->leave_spinning = 1;
2005         /* this will setup the path even if it fails to insert the back ref */
2006         ret = insert_inline_extent_backref(trans, root->fs_info->extent_root,
2007                                            path, bytenr, num_bytes, parent,
2008                                            root_objectid, owner, offset,
2009                                            refs_to_add, extent_op);
2010         if (ret == 0)
2011                 goto out;
2012
2013         if (ret != -EAGAIN) {
2014                 err = ret;
2015                 goto out;
2016         }
2017
2018         leaf = path->nodes[0];
2019         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2020         refs = btrfs_extent_refs(leaf, item);
2021         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2022         if (extent_op)
2023                 __run_delayed_extent_op(extent_op, leaf, item);
2024
2025         btrfs_mark_buffer_dirty(leaf);
2026         btrfs_release_path(path);
2027
2028         path->reada = 1;
2029         path->leave_spinning = 1;
2030
2031         /* now insert the actual backref */
2032         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2033                                     path, bytenr, parent, root_objectid,
2034                                     owner, offset, refs_to_add);
2035         if (ret)
2036                 btrfs_abort_transaction(trans, root, ret);
2037 out:
2038         btrfs_free_path(path);
2039         return err;
2040 }
2041
2042 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2043                                 struct btrfs_root *root,
2044                                 struct btrfs_delayed_ref_node *node,
2045                                 struct btrfs_delayed_extent_op *extent_op,
2046                                 int insert_reserved)
2047 {
2048         int ret = 0;
2049         struct btrfs_delayed_data_ref *ref;
2050         struct btrfs_key ins;
2051         u64 parent = 0;
2052         u64 ref_root = 0;
2053         u64 flags = 0;
2054
2055         ins.objectid = node->bytenr;
2056         ins.offset = node->num_bytes;
2057         ins.type = BTRFS_EXTENT_ITEM_KEY;
2058
2059         ref = btrfs_delayed_node_to_data_ref(node);
2060         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2061                 parent = ref->parent;
2062         else
2063                 ref_root = ref->root;
2064
2065         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2066                 if (extent_op)
2067                         flags |= extent_op->flags_to_set;
2068                 ret = alloc_reserved_file_extent(trans, root,
2069                                                  parent, ref_root, flags,
2070                                                  ref->objectid, ref->offset,
2071                                                  &ins, node->ref_mod);
2072         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2073                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2074                                              node->num_bytes, parent,
2075                                              ref_root, ref->objectid,
2076                                              ref->offset, node->ref_mod,
2077                                              extent_op);
2078         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2079                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2080                                           node->num_bytes, parent,
2081                                           ref_root, ref->objectid,
2082                                           ref->offset, node->ref_mod,
2083                                           extent_op);
2084         } else {
2085                 BUG();
2086         }
2087         return ret;
2088 }
2089
2090 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2091                                     struct extent_buffer *leaf,
2092                                     struct btrfs_extent_item *ei)
2093 {
2094         u64 flags = btrfs_extent_flags(leaf, ei);
2095         if (extent_op->update_flags) {
2096                 flags |= extent_op->flags_to_set;
2097                 btrfs_set_extent_flags(leaf, ei, flags);
2098         }
2099
2100         if (extent_op->update_key) {
2101                 struct btrfs_tree_block_info *bi;
2102                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2103                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2104                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2105         }
2106 }
2107
2108 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2109                                  struct btrfs_root *root,
2110                                  struct btrfs_delayed_ref_node *node,
2111                                  struct btrfs_delayed_extent_op *extent_op)
2112 {
2113         struct btrfs_key key;
2114         struct btrfs_path *path;
2115         struct btrfs_extent_item *ei;
2116         struct extent_buffer *leaf;
2117         u32 item_size;
2118         int ret;
2119         int err = 0;
2120         int metadata = (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2121                         node->type == BTRFS_SHARED_BLOCK_REF_KEY);
2122
2123         if (trans->aborted)
2124                 return 0;
2125
2126         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2127                 metadata = 0;
2128
2129         path = btrfs_alloc_path();
2130         if (!path)
2131                 return -ENOMEM;
2132
2133         key.objectid = node->bytenr;
2134
2135         if (metadata) {
2136                 struct btrfs_delayed_tree_ref *tree_ref;
2137
2138                 tree_ref = btrfs_delayed_node_to_tree_ref(node);
2139                 key.type = BTRFS_METADATA_ITEM_KEY;
2140                 key.offset = tree_ref->level;
2141         } else {
2142                 key.type = BTRFS_EXTENT_ITEM_KEY;
2143                 key.offset = node->num_bytes;
2144         }
2145
2146 again:
2147         path->reada = 1;
2148         path->leave_spinning = 1;
2149         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2150                                 path, 0, 1);
2151         if (ret < 0) {
2152                 err = ret;
2153                 goto out;
2154         }
2155         if (ret > 0) {
2156                 if (metadata) {
2157                         btrfs_release_path(path);
2158                         metadata = 0;
2159
2160                         key.offset = node->num_bytes;
2161                         key.type = BTRFS_EXTENT_ITEM_KEY;
2162                         goto again;
2163                 }
2164                 err = -EIO;
2165                 goto out;
2166         }
2167
2168         leaf = path->nodes[0];
2169         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2170 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2171         if (item_size < sizeof(*ei)) {
2172                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2173                                              path, (u64)-1, 0);
2174                 if (ret < 0) {
2175                         err = ret;
2176                         goto out;
2177                 }
2178                 leaf = path->nodes[0];
2179                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2180         }
2181 #endif
2182         BUG_ON(item_size < sizeof(*ei));
2183         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2184         __run_delayed_extent_op(extent_op, leaf, ei);
2185
2186         btrfs_mark_buffer_dirty(leaf);
2187 out:
2188         btrfs_free_path(path);
2189         return err;
2190 }
2191
2192 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2193                                 struct btrfs_root *root,
2194                                 struct btrfs_delayed_ref_node *node,
2195                                 struct btrfs_delayed_extent_op *extent_op,
2196                                 int insert_reserved)
2197 {
2198         int ret = 0;
2199         struct btrfs_delayed_tree_ref *ref;
2200         struct btrfs_key ins;
2201         u64 parent = 0;
2202         u64 ref_root = 0;
2203         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2204                                                  SKINNY_METADATA);
2205
2206         ref = btrfs_delayed_node_to_tree_ref(node);
2207         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2208                 parent = ref->parent;
2209         else
2210                 ref_root = ref->root;
2211
2212         ins.objectid = node->bytenr;
2213         if (skinny_metadata) {
2214                 ins.offset = ref->level;
2215                 ins.type = BTRFS_METADATA_ITEM_KEY;
2216         } else {
2217                 ins.offset = node->num_bytes;
2218                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2219         }
2220
2221         BUG_ON(node->ref_mod != 1);
2222         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2223                 BUG_ON(!extent_op || !extent_op->update_flags);
2224                 ret = alloc_reserved_tree_block(trans, root,
2225                                                 parent, ref_root,
2226                                                 extent_op->flags_to_set,
2227                                                 &extent_op->key,
2228                                                 ref->level, &ins);
2229         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2230                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2231                                              node->num_bytes, parent, ref_root,
2232                                              ref->level, 0, 1, extent_op);
2233         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2234                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2235                                           node->num_bytes, parent, ref_root,
2236                                           ref->level, 0, 1, extent_op);
2237         } else {
2238                 BUG();
2239         }
2240         return ret;
2241 }
2242
2243 /* helper function to actually process a single delayed ref entry */
2244 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2245                                struct btrfs_root *root,
2246                                struct btrfs_delayed_ref_node *node,
2247                                struct btrfs_delayed_extent_op *extent_op,
2248                                int insert_reserved)
2249 {
2250         int ret = 0;
2251
2252         if (trans->aborted)
2253                 return 0;
2254
2255         if (btrfs_delayed_ref_is_head(node)) {
2256                 struct btrfs_delayed_ref_head *head;
2257                 /*
2258                  * we've hit the end of the chain and we were supposed
2259                  * to insert this extent into the tree.  But, it got
2260                  * deleted before we ever needed to insert it, so all
2261                  * we have to do is clean up the accounting
2262                  */
2263                 BUG_ON(extent_op);
2264                 head = btrfs_delayed_node_to_head(node);
2265                 if (insert_reserved) {
2266                         btrfs_pin_extent(root, node->bytenr,
2267                                          node->num_bytes, 1);
2268                         if (head->is_data) {
2269                                 ret = btrfs_del_csums(trans, root,
2270                                                       node->bytenr,
2271                                                       node->num_bytes);
2272                         }
2273                 }
2274                 return ret;
2275         }
2276
2277         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2278             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2279                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2280                                            insert_reserved);
2281         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2282                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2283                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2284                                            insert_reserved);
2285         else
2286                 BUG();
2287         return ret;
2288 }
2289
2290 static noinline struct btrfs_delayed_ref_node *
2291 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2292 {
2293         struct rb_node *node;
2294         struct btrfs_delayed_ref_node *ref;
2295         int action = BTRFS_ADD_DELAYED_REF;
2296 again:
2297         /*
2298          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2299          * this prevents ref count from going down to zero when
2300          * there still are pending delayed ref.
2301          */
2302         node = rb_prev(&head->node.rb_node);
2303         while (1) {
2304                 if (!node)
2305                         break;
2306                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2307                                 rb_node);
2308                 if (ref->bytenr != head->node.bytenr)
2309                         break;
2310                 if (ref->action == action)
2311                         return ref;
2312                 node = rb_prev(node);
2313         }
2314         if (action == BTRFS_ADD_DELAYED_REF) {
2315                 action = BTRFS_DROP_DELAYED_REF;
2316                 goto again;
2317         }
2318         return NULL;
2319 }
2320
2321 /*
2322  * Returns 0 on success or if called with an already aborted transaction.
2323  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2324  */
2325 static noinline int run_clustered_refs(struct btrfs_trans_handle *trans,
2326                                        struct btrfs_root *root,
2327                                        struct list_head *cluster)
2328 {
2329         struct btrfs_delayed_ref_root *delayed_refs;
2330         struct btrfs_delayed_ref_node *ref;
2331         struct btrfs_delayed_ref_head *locked_ref = NULL;
2332         struct btrfs_delayed_extent_op *extent_op;
2333         struct btrfs_fs_info *fs_info = root->fs_info;
2334         int ret;
2335         int count = 0;
2336         int must_insert_reserved = 0;
2337
2338         delayed_refs = &trans->transaction->delayed_refs;
2339         while (1) {
2340                 if (!locked_ref) {
2341                         /* pick a new head ref from the cluster list */
2342                         if (list_empty(cluster))
2343                                 break;
2344
2345                         locked_ref = list_entry(cluster->next,
2346                                      struct btrfs_delayed_ref_head, cluster);
2347
2348                         /* grab the lock that says we are going to process
2349                          * all the refs for this head */
2350                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2351
2352                         /*
2353                          * we may have dropped the spin lock to get the head
2354                          * mutex lock, and that might have given someone else
2355                          * time to free the head.  If that's true, it has been
2356                          * removed from our list and we can move on.
2357                          */
2358                         if (ret == -EAGAIN) {
2359                                 locked_ref = NULL;
2360                                 count++;
2361                                 continue;
2362                         }
2363                 }
2364
2365                 /*
2366                  * We need to try and merge add/drops of the same ref since we
2367                  * can run into issues with relocate dropping the implicit ref
2368                  * and then it being added back again before the drop can
2369                  * finish.  If we merged anything we need to re-loop so we can
2370                  * get a good ref.
2371                  */
2372                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2373                                          locked_ref);
2374
2375                 /*
2376                  * locked_ref is the head node, so we have to go one
2377                  * node back for any delayed ref updates
2378                  */
2379                 ref = select_delayed_ref(locked_ref);
2380
2381                 if (ref && ref->seq &&
2382                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2383                         /*
2384                          * there are still refs with lower seq numbers in the
2385                          * process of being added. Don't run this ref yet.
2386                          */
2387                         list_del_init(&locked_ref->cluster);
2388                         btrfs_delayed_ref_unlock(locked_ref);
2389                         locked_ref = NULL;
2390                         delayed_refs->num_heads_ready++;
2391                         spin_unlock(&delayed_refs->lock);
2392                         cond_resched();
2393                         spin_lock(&delayed_refs->lock);
2394                         continue;
2395                 }
2396
2397                 /*
2398                  * record the must insert reserved flag before we
2399                  * drop the spin lock.
2400                  */
2401                 must_insert_reserved = locked_ref->must_insert_reserved;
2402                 locked_ref->must_insert_reserved = 0;
2403
2404                 extent_op = locked_ref->extent_op;
2405                 locked_ref->extent_op = NULL;
2406
2407                 if (!ref) {
2408                         /* All delayed refs have been processed, Go ahead
2409                          * and send the head node to run_one_delayed_ref,
2410                          * so that any accounting fixes can happen
2411                          */
2412                         ref = &locked_ref->node;
2413
2414                         if (extent_op && must_insert_reserved) {
2415                                 btrfs_free_delayed_extent_op(extent_op);
2416                                 extent_op = NULL;
2417                         }
2418
2419                         if (extent_op) {
2420                                 spin_unlock(&delayed_refs->lock);
2421
2422                                 ret = run_delayed_extent_op(trans, root,
2423                                                             ref, extent_op);
2424                                 btrfs_free_delayed_extent_op(extent_op);
2425
2426                                 if (ret) {
2427                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2428                                         spin_lock(&delayed_refs->lock);
2429                                         btrfs_delayed_ref_unlock(locked_ref);
2430                                         return ret;
2431                                 }
2432
2433                                 goto next;
2434                         }
2435                 }
2436
2437                 ref->in_tree = 0;
2438                 rb_erase(&ref->rb_node, &delayed_refs->root);
2439                 delayed_refs->num_entries--;
2440                 if (!btrfs_delayed_ref_is_head(ref)) {
2441                         /*
2442                          * when we play the delayed ref, also correct the
2443                          * ref_mod on head
2444                          */
2445                         switch (ref->action) {
2446                         case BTRFS_ADD_DELAYED_REF:
2447                         case BTRFS_ADD_DELAYED_EXTENT:
2448                                 locked_ref->node.ref_mod -= ref->ref_mod;
2449                                 break;
2450                         case BTRFS_DROP_DELAYED_REF:
2451                                 locked_ref->node.ref_mod += ref->ref_mod;
2452                                 break;
2453                         default:
2454                                 WARN_ON(1);
2455                         }
2456                 }
2457                 spin_unlock(&delayed_refs->lock);
2458
2459                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2460                                           must_insert_reserved);
2461
2462                 btrfs_free_delayed_extent_op(extent_op);
2463                 if (ret) {
2464                         btrfs_delayed_ref_unlock(locked_ref);
2465                         btrfs_put_delayed_ref(ref);
2466                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2467                         spin_lock(&delayed_refs->lock);
2468                         return ret;
2469                 }
2470
2471                 /*
2472                  * If this node is a head, that means all the refs in this head
2473                  * have been dealt with, and we will pick the next head to deal
2474                  * with, so we must unlock the head and drop it from the cluster
2475                  * list before we release it.
2476                  */
2477                 if (btrfs_delayed_ref_is_head(ref)) {
2478                         list_del_init(&locked_ref->cluster);
2479                         btrfs_delayed_ref_unlock(locked_ref);
2480                         locked_ref = NULL;
2481                 }
2482                 btrfs_put_delayed_ref(ref);
2483                 count++;
2484 next:
2485                 cond_resched();
2486                 spin_lock(&delayed_refs->lock);
2487         }
2488         return count;
2489 }
2490
2491 #ifdef SCRAMBLE_DELAYED_REFS
2492 /*
2493  * Normally delayed refs get processed in ascending bytenr order. This
2494  * correlates in most cases to the order added. To expose dependencies on this
2495  * order, we start to process the tree in the middle instead of the beginning
2496  */
2497 static u64 find_middle(struct rb_root *root)
2498 {
2499         struct rb_node *n = root->rb_node;
2500         struct btrfs_delayed_ref_node *entry;
2501         int alt = 1;
2502         u64 middle;
2503         u64 first = 0, last = 0;
2504
2505         n = rb_first(root);
2506         if (n) {
2507                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2508                 first = entry->bytenr;
2509         }
2510         n = rb_last(root);
2511         if (n) {
2512                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2513                 last = entry->bytenr;
2514         }
2515         n = root->rb_node;
2516
2517         while (n) {
2518                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2519                 WARN_ON(!entry->in_tree);
2520
2521                 middle = entry->bytenr;
2522
2523                 if (alt)
2524                         n = n->rb_left;
2525                 else
2526                         n = n->rb_right;
2527
2528                 alt = 1 - alt;
2529         }
2530         return middle;
2531 }
2532 #endif
2533
2534 int btrfs_delayed_refs_qgroup_accounting(struct btrfs_trans_handle *trans,
2535                                          struct btrfs_fs_info *fs_info)
2536 {
2537         struct qgroup_update *qgroup_update;
2538         int ret = 0;
2539
2540         if (list_empty(&trans->qgroup_ref_list) !=
2541             !trans->delayed_ref_elem.seq) {
2542                 /* list without seq or seq without list */
2543                 btrfs_err(fs_info,
2544                         "qgroup accounting update error, list is%s empty, seq is %#x.%x",
2545                         list_empty(&trans->qgroup_ref_list) ? "" : " not",
2546                         (u32)(trans->delayed_ref_elem.seq >> 32),
2547                         (u32)trans->delayed_ref_elem.seq);
2548                 BUG();
2549         }
2550
2551         if (!trans->delayed_ref_elem.seq)
2552                 return 0;
2553
2554         while (!list_empty(&trans->qgroup_ref_list)) {
2555                 qgroup_update = list_first_entry(&trans->qgroup_ref_list,
2556                                                  struct qgroup_update, list);
2557                 list_del(&qgroup_update->list);
2558                 if (!ret)
2559                         ret = btrfs_qgroup_account_ref(
2560                                         trans, fs_info, qgroup_update->node,
2561                                         qgroup_update->extent_op);
2562                 kfree(qgroup_update);
2563         }
2564
2565         btrfs_put_tree_mod_seq(fs_info, &trans->delayed_ref_elem);
2566
2567         return ret;
2568 }
2569
2570 static int refs_newer(struct btrfs_delayed_ref_root *delayed_refs, int seq,
2571                       int count)
2572 {
2573         int val = atomic_read(&delayed_refs->ref_seq);
2574
2575         if (val < seq || val >= seq + count)
2576                 return 1;
2577         return 0;
2578 }
2579
2580 /*
2581  * this starts processing the delayed reference count updates and
2582  * extent insertions we have queued up so far.  count can be
2583  * 0, which means to process everything in the tree at the start
2584  * of the run (but not newly added entries), or it can be some target
2585  * number you'd like to process.
2586  *
2587  * Returns 0 on success or if called with an aborted transaction
2588  * Returns <0 on error and aborts the transaction
2589  */
2590 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2591                            struct btrfs_root *root, unsigned long count)
2592 {
2593         struct rb_node *node;
2594         struct btrfs_delayed_ref_root *delayed_refs;
2595         struct btrfs_delayed_ref_node *ref;
2596         struct list_head cluster;
2597         int ret;
2598         u64 delayed_start;
2599         int run_all = count == (unsigned long)-1;
2600         int run_most = 0;
2601         int loops;
2602
2603         /* We'll clean this up in btrfs_cleanup_transaction */
2604         if (trans->aborted)
2605                 return 0;
2606
2607         if (root == root->fs_info->extent_root)
2608                 root = root->fs_info->tree_root;
2609
2610         btrfs_delayed_refs_qgroup_accounting(trans, root->fs_info);
2611
2612         delayed_refs = &trans->transaction->delayed_refs;
2613         INIT_LIST_HEAD(&cluster);
2614         if (count == 0) {
2615                 count = delayed_refs->num_entries * 2;
2616                 run_most = 1;
2617         }
2618
2619         if (!run_all && !run_most) {
2620                 int old;
2621                 int seq = atomic_read(&delayed_refs->ref_seq);
2622
2623 progress:
2624                 old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2625                 if (old) {
2626                         DEFINE_WAIT(__wait);
2627                         if (delayed_refs->num_entries < 16348)
2628                                 return 0;
2629
2630                         prepare_to_wait(&delayed_refs->wait, &__wait,
2631                                         TASK_UNINTERRUPTIBLE);
2632
2633                         old = atomic_cmpxchg(&delayed_refs->procs_running_refs, 0, 1);
2634                         if (old) {
2635                                 schedule();
2636                                 finish_wait(&delayed_refs->wait, &__wait);
2637
2638                                 if (!refs_newer(delayed_refs, seq, 256))
2639                                         goto progress;
2640                                 else
2641                                         return 0;
2642                         } else {
2643                                 finish_wait(&delayed_refs->wait, &__wait);
2644                                 goto again;
2645                         }
2646                 }
2647
2648         } else {
2649                 atomic_inc(&delayed_refs->procs_running_refs);
2650         }
2651
2652 again:
2653         loops = 0;
2654         spin_lock(&delayed_refs->lock);
2655
2656 #ifdef SCRAMBLE_DELAYED_REFS
2657         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2658 #endif
2659
2660         while (1) {
2661                 if (!(run_all || run_most) &&
2662                     delayed_refs->num_heads_ready < 64)
2663                         break;
2664
2665                 /*
2666                  * go find something we can process in the rbtree.  We start at
2667                  * the beginning of the tree, and then build a cluster
2668                  * of refs to process starting at the first one we are able to
2669                  * lock
2670                  */
2671                 delayed_start = delayed_refs->run_delayed_start;
2672                 ret = btrfs_find_ref_cluster(trans, &cluster,
2673                                              delayed_refs->run_delayed_start);
2674                 if (ret)
2675                         break;
2676
2677                 ret = run_clustered_refs(trans, root, &cluster);
2678                 if (ret < 0) {
2679                         btrfs_release_ref_cluster(&cluster);
2680                         spin_unlock(&delayed_refs->lock);
2681                         btrfs_abort_transaction(trans, root, ret);
2682                         atomic_dec(&delayed_refs->procs_running_refs);
2683                         return ret;
2684                 }
2685
2686                 atomic_add(ret, &delayed_refs->ref_seq);
2687
2688                 count -= min_t(unsigned long, ret, count);
2689
2690                 if (count == 0)
2691                         break;
2692
2693                 if (delayed_start >= delayed_refs->run_delayed_start) {
2694                         if (loops == 0) {
2695                                 /*
2696                                  * btrfs_find_ref_cluster looped. let's do one
2697                                  * more cycle. if we don't run any delayed ref
2698                                  * during that cycle (because we can't because
2699                                  * all of them are blocked), bail out.
2700                                  */
2701                                 loops = 1;
2702                         } else {
2703                                 /*
2704                                  * no runnable refs left, stop trying
2705                                  */
2706                                 BUG_ON(run_all);
2707                                 break;
2708                         }
2709                 }
2710                 if (ret) {
2711                         /* refs were run, let's reset staleness detection */
2712                         loops = 0;
2713                 }
2714         }
2715
2716         if (run_all) {
2717                 if (!list_empty(&trans->new_bgs)) {
2718                         spin_unlock(&delayed_refs->lock);
2719                         btrfs_create_pending_block_groups(trans, root);
2720                         spin_lock(&delayed_refs->lock);
2721                 }
2722
2723                 node = rb_first(&delayed_refs->root);
2724                 if (!node)
2725                         goto out;
2726                 count = (unsigned long)-1;
2727
2728                 while (node) {
2729                         ref = rb_entry(node, struct btrfs_delayed_ref_node,
2730                                        rb_node);
2731                         if (btrfs_delayed_ref_is_head(ref)) {
2732                                 struct btrfs_delayed_ref_head *head;
2733
2734                                 head = btrfs_delayed_node_to_head(ref);
2735                                 atomic_inc(&ref->refs);
2736
2737                                 spin_unlock(&delayed_refs->lock);
2738                                 /*
2739                                  * Mutex was contended, block until it's
2740                                  * released and try again
2741                                  */
2742                                 mutex_lock(&head->mutex);
2743                                 mutex_unlock(&head->mutex);
2744
2745                                 btrfs_put_delayed_ref(ref);
2746                                 cond_resched();
2747                                 goto again;
2748                         }
2749                         node = rb_next(node);
2750                 }
2751                 spin_unlock(&delayed_refs->lock);
2752                 schedule_timeout(1);
2753                 goto again;
2754         }
2755 out:
2756         atomic_dec(&delayed_refs->procs_running_refs);
2757         smp_mb();
2758         if (waitqueue_active(&delayed_refs->wait))
2759                 wake_up(&delayed_refs->wait);
2760
2761         spin_unlock(&delayed_refs->lock);
2762         assert_qgroups_uptodate(trans);
2763         return 0;
2764 }
2765
2766 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2767                                 struct btrfs_root *root,
2768                                 u64 bytenr, u64 num_bytes, u64 flags,
2769                                 int is_data)
2770 {
2771         struct btrfs_delayed_extent_op *extent_op;
2772         int ret;
2773
2774         extent_op = btrfs_alloc_delayed_extent_op();
2775         if (!extent_op)
2776                 return -ENOMEM;
2777
2778         extent_op->flags_to_set = flags;
2779         extent_op->update_flags = 1;
2780         extent_op->update_key = 0;
2781         extent_op->is_data = is_data ? 1 : 0;
2782
2783         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2784                                           num_bytes, extent_op);
2785         if (ret)
2786                 btrfs_free_delayed_extent_op(extent_op);
2787         return ret;
2788 }
2789
2790 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2791                                       struct btrfs_root *root,
2792                                       struct btrfs_path *path,
2793                                       u64 objectid, u64 offset, u64 bytenr)
2794 {
2795         struct btrfs_delayed_ref_head *head;
2796         struct btrfs_delayed_ref_node *ref;
2797         struct btrfs_delayed_data_ref *data_ref;
2798         struct btrfs_delayed_ref_root *delayed_refs;
2799         struct rb_node *node;
2800         int ret = 0;
2801
2802         ret = -ENOENT;
2803         delayed_refs = &trans->transaction->delayed_refs;
2804         spin_lock(&delayed_refs->lock);
2805         head = btrfs_find_delayed_ref_head(trans, bytenr);
2806         if (!head)
2807                 goto out;
2808
2809         if (!mutex_trylock(&head->mutex)) {
2810                 atomic_inc(&head->node.refs);
2811                 spin_unlock(&delayed_refs->lock);
2812
2813                 btrfs_release_path(path);
2814
2815                 /*
2816                  * Mutex was contended, block until it's released and let
2817                  * caller try again
2818                  */
2819                 mutex_lock(&head->mutex);
2820                 mutex_unlock(&head->mutex);
2821                 btrfs_put_delayed_ref(&head->node);
2822                 return -EAGAIN;
2823         }
2824
2825         node = rb_prev(&head->node.rb_node);
2826         if (!node)
2827                 goto out_unlock;
2828
2829         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2830
2831         if (ref->bytenr != bytenr)
2832                 goto out_unlock;
2833
2834         ret = 1;
2835         if (ref->type != BTRFS_EXTENT_DATA_REF_KEY)
2836                 goto out_unlock;
2837
2838         data_ref = btrfs_delayed_node_to_data_ref(ref);
2839
2840         node = rb_prev(node);
2841         if (node) {
2842                 int seq = ref->seq;
2843
2844                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2845                 if (ref->bytenr == bytenr && ref->seq == seq)
2846                         goto out_unlock;
2847         }
2848
2849         if (data_ref->root != root->root_key.objectid ||
2850             data_ref->objectid != objectid || data_ref->offset != offset)
2851                 goto out_unlock;
2852
2853         ret = 0;
2854 out_unlock:
2855         mutex_unlock(&head->mutex);
2856 out:
2857         spin_unlock(&delayed_refs->lock);
2858         return ret;
2859 }
2860
2861 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2862                                         struct btrfs_root *root,
2863                                         struct btrfs_path *path,
2864                                         u64 objectid, u64 offset, u64 bytenr)
2865 {
2866         struct btrfs_root *extent_root = root->fs_info->extent_root;
2867         struct extent_buffer *leaf;
2868         struct btrfs_extent_data_ref *ref;
2869         struct btrfs_extent_inline_ref *iref;
2870         struct btrfs_extent_item *ei;
2871         struct btrfs_key key;
2872         u32 item_size;
2873         int ret;
2874
2875         key.objectid = bytenr;
2876         key.offset = (u64)-1;
2877         key.type = BTRFS_EXTENT_ITEM_KEY;
2878
2879         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2880         if (ret < 0)
2881                 goto out;
2882         BUG_ON(ret == 0); /* Corruption */
2883
2884         ret = -ENOENT;
2885         if (path->slots[0] == 0)
2886                 goto out;
2887
2888         path->slots[0]--;
2889         leaf = path->nodes[0];
2890         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2891
2892         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2893                 goto out;
2894
2895         ret = 1;
2896         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2897 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2898         if (item_size < sizeof(*ei)) {
2899                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2900                 goto out;
2901         }
2902 #endif
2903         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2904
2905         if (item_size != sizeof(*ei) +
2906             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2907                 goto out;
2908
2909         if (btrfs_extent_generation(leaf, ei) <=
2910             btrfs_root_last_snapshot(&root->root_item))
2911                 goto out;
2912
2913         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2914         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2915             BTRFS_EXTENT_DATA_REF_KEY)
2916                 goto out;
2917
2918         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2919         if (btrfs_extent_refs(leaf, ei) !=
2920             btrfs_extent_data_ref_count(leaf, ref) ||
2921             btrfs_extent_data_ref_root(leaf, ref) !=
2922             root->root_key.objectid ||
2923             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2924             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2925                 goto out;
2926
2927         ret = 0;
2928 out:
2929         return ret;
2930 }
2931
2932 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2933                           struct btrfs_root *root,
2934                           u64 objectid, u64 offset, u64 bytenr)
2935 {
2936         struct btrfs_path *path;
2937         int ret;
2938         int ret2;
2939
2940         path = btrfs_alloc_path();
2941         if (!path)
2942                 return -ENOENT;
2943
2944         do {
2945                 ret = check_committed_ref(trans, root, path, objectid,
2946                                           offset, bytenr);
2947                 if (ret && ret != -ENOENT)
2948                         goto out;
2949
2950                 ret2 = check_delayed_ref(trans, root, path, objectid,
2951                                          offset, bytenr);
2952         } while (ret2 == -EAGAIN);
2953
2954         if (ret2 && ret2 != -ENOENT) {
2955                 ret = ret2;
2956                 goto out;
2957         }
2958
2959         if (ret != -ENOENT || ret2 != -ENOENT)
2960                 ret = 0;
2961 out:
2962         btrfs_free_path(path);
2963         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
2964                 WARN_ON(ret > 0);
2965         return ret;
2966 }
2967
2968 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
2969                            struct btrfs_root *root,
2970                            struct extent_buffer *buf,
2971                            int full_backref, int inc, int for_cow)
2972 {
2973         u64 bytenr;
2974         u64 num_bytes;
2975         u64 parent;
2976         u64 ref_root;
2977         u32 nritems;
2978         struct btrfs_key key;
2979         struct btrfs_file_extent_item *fi;
2980         int i;
2981         int level;
2982         int ret = 0;
2983         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
2984                             u64, u64, u64, u64, u64, u64, int);
2985
2986         ref_root = btrfs_header_owner(buf);
2987         nritems = btrfs_header_nritems(buf);
2988         level = btrfs_header_level(buf);
2989
2990         if (!root->ref_cows && level == 0)
2991                 return 0;
2992
2993         if (inc)
2994                 process_func = btrfs_inc_extent_ref;
2995         else
2996                 process_func = btrfs_free_extent;
2997
2998         if (full_backref)
2999                 parent = buf->start;
3000         else
3001                 parent = 0;
3002
3003         for (i = 0; i < nritems; i++) {
3004                 if (level == 0) {
3005                         btrfs_item_key_to_cpu(buf, &key, i);
3006                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
3007                                 continue;
3008                         fi = btrfs_item_ptr(buf, i,
3009                                             struct btrfs_file_extent_item);
3010                         if (btrfs_file_extent_type(buf, fi) ==
3011                             BTRFS_FILE_EXTENT_INLINE)
3012                                 continue;
3013                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3014                         if (bytenr == 0)
3015                                 continue;
3016
3017                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3018                         key.offset -= btrfs_file_extent_offset(buf, fi);
3019                         ret = process_func(trans, root, bytenr, num_bytes,
3020                                            parent, ref_root, key.objectid,
3021                                            key.offset, for_cow);
3022                         if (ret)
3023                                 goto fail;
3024                 } else {
3025                         bytenr = btrfs_node_blockptr(buf, i);
3026                         num_bytes = btrfs_level_size(root, level - 1);
3027                         ret = process_func(trans, root, bytenr, num_bytes,
3028                                            parent, ref_root, level - 1, 0,
3029                                            for_cow);
3030                         if (ret)
3031                                 goto fail;
3032                 }
3033         }
3034         return 0;
3035 fail:
3036         return ret;
3037 }
3038
3039 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3040                   struct extent_buffer *buf, int full_backref, int for_cow)
3041 {
3042         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, for_cow);
3043 }
3044
3045 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3046                   struct extent_buffer *buf, int full_backref, int for_cow)
3047 {
3048         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, for_cow);
3049 }
3050
3051 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3052                                  struct btrfs_root *root,
3053                                  struct btrfs_path *path,
3054                                  struct btrfs_block_group_cache *cache)
3055 {
3056         int ret;
3057         struct btrfs_root *extent_root = root->fs_info->extent_root;
3058         unsigned long bi;
3059         struct extent_buffer *leaf;
3060
3061         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3062         if (ret < 0)
3063                 goto fail;
3064         BUG_ON(ret); /* Corruption */
3065
3066         leaf = path->nodes[0];
3067         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3068         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3069         btrfs_mark_buffer_dirty(leaf);
3070         btrfs_release_path(path);
3071 fail:
3072         if (ret) {
3073                 btrfs_abort_transaction(trans, root, ret);
3074                 return ret;
3075         }
3076         return 0;
3077
3078 }
3079
3080 static struct btrfs_block_group_cache *
3081 next_block_group(struct btrfs_root *root,
3082                  struct btrfs_block_group_cache *cache)
3083 {
3084         struct rb_node *node;
3085         spin_lock(&root->fs_info->block_group_cache_lock);
3086         node = rb_next(&cache->cache_node);
3087         btrfs_put_block_group(cache);
3088         if (node) {
3089                 cache = rb_entry(node, struct btrfs_block_group_cache,
3090                                  cache_node);
3091                 btrfs_get_block_group(cache);
3092         } else
3093                 cache = NULL;
3094         spin_unlock(&root->fs_info->block_group_cache_lock);
3095         return cache;
3096 }
3097
3098 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3099                             struct btrfs_trans_handle *trans,
3100                             struct btrfs_path *path)
3101 {
3102         struct btrfs_root *root = block_group->fs_info->tree_root;
3103         struct inode *inode = NULL;
3104         u64 alloc_hint = 0;
3105         int dcs = BTRFS_DC_ERROR;
3106         int num_pages = 0;
3107         int retries = 0;
3108         int ret = 0;
3109
3110         /*
3111          * If this block group is smaller than 100 megs don't bother caching the
3112          * block group.
3113          */
3114         if (block_group->key.offset < (100 * 1024 * 1024)) {
3115                 spin_lock(&block_group->lock);
3116                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3117                 spin_unlock(&block_group->lock);
3118                 return 0;
3119         }
3120
3121 again:
3122         inode = lookup_free_space_inode(root, block_group, path);
3123         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3124                 ret = PTR_ERR(inode);
3125                 btrfs_release_path(path);
3126                 goto out;
3127         }
3128
3129         if (IS_ERR(inode)) {
3130                 BUG_ON(retries);
3131                 retries++;
3132
3133                 if (block_group->ro)
3134                         goto out_free;
3135
3136                 ret = create_free_space_inode(root, trans, block_group, path);
3137                 if (ret)
3138                         goto out_free;
3139                 goto again;
3140         }
3141
3142         /* We've already setup this transaction, go ahead and exit */
3143         if (block_group->cache_generation == trans->transid &&
3144             i_size_read(inode)) {
3145                 dcs = BTRFS_DC_SETUP;
3146                 goto out_put;
3147         }
3148
3149         /*
3150          * We want to set the generation to 0, that way if anything goes wrong
3151          * from here on out we know not to trust this cache when we load up next
3152          * time.
3153          */
3154         BTRFS_I(inode)->generation = 0;
3155         ret = btrfs_update_inode(trans, root, inode);
3156         WARN_ON(ret);
3157
3158         if (i_size_read(inode) > 0) {
3159                 ret = btrfs_truncate_free_space_cache(root, trans, path,
3160                                                       inode);
3161                 if (ret)
3162                         goto out_put;
3163         }
3164
3165         spin_lock(&block_group->lock);
3166         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3167             !btrfs_test_opt(root, SPACE_CACHE)) {
3168                 /*
3169                  * don't bother trying to write stuff out _if_
3170                  * a) we're not cached,
3171                  * b) we're with nospace_cache mount option.
3172                  */
3173                 dcs = BTRFS_DC_WRITTEN;
3174                 spin_unlock(&block_group->lock);
3175                 goto out_put;
3176         }
3177         spin_unlock(&block_group->lock);
3178
3179         /*
3180          * Try to preallocate enough space based on how big the block group is.
3181          * Keep in mind this has to include any pinned space which could end up
3182          * taking up quite a bit since it's not folded into the other space
3183          * cache.
3184          */
3185         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3186         if (!num_pages)
3187                 num_pages = 1;
3188
3189         num_pages *= 16;
3190         num_pages *= PAGE_CACHE_SIZE;
3191
3192         ret = btrfs_check_data_free_space(inode, num_pages);
3193         if (ret)
3194                 goto out_put;
3195
3196         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3197                                               num_pages, num_pages,
3198                                               &alloc_hint);
3199         if (!ret)
3200                 dcs = BTRFS_DC_SETUP;
3201         btrfs_free_reserved_data_space(inode, num_pages);
3202
3203 out_put:
3204         iput(inode);
3205 out_free:
3206         btrfs_release_path(path);
3207 out:
3208         spin_lock(&block_group->lock);
3209         if (!ret && dcs == BTRFS_DC_SETUP)
3210                 block_group->cache_generation = trans->transid;
3211         block_group->disk_cache_state = dcs;
3212         spin_unlock(&block_group->lock);
3213
3214         return ret;
3215 }
3216
3217 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3218                                    struct btrfs_root *root)
3219 {
3220         struct btrfs_block_group_cache *cache;
3221         int err = 0;
3222         struct btrfs_path *path;
3223         u64 last = 0;
3224
3225         path = btrfs_alloc_path();
3226         if (!path)
3227                 return -ENOMEM;
3228
3229 again:
3230         while (1) {
3231                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3232                 while (cache) {
3233                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3234                                 break;
3235                         cache = next_block_group(root, cache);
3236                 }
3237                 if (!cache) {
3238                         if (last == 0)
3239                                 break;
3240                         last = 0;
3241                         continue;
3242                 }
3243                 err = cache_save_setup(cache, trans, path);
3244                 last = cache->key.objectid + cache->key.offset;
3245                 btrfs_put_block_group(cache);
3246         }
3247
3248         while (1) {
3249                 if (last == 0) {
3250                         err = btrfs_run_delayed_refs(trans, root,
3251                                                      (unsigned long)-1);
3252                         if (err) /* File system offline */
3253                                 goto out;
3254                 }
3255
3256                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3257                 while (cache) {
3258                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3259                                 btrfs_put_block_group(cache);
3260                                 goto again;
3261                         }
3262
3263                         if (cache->dirty)
3264                                 break;
3265                         cache = next_block_group(root, cache);
3266                 }
3267                 if (!cache) {
3268                         if (last == 0)
3269                                 break;
3270                         last = 0;
3271                         continue;
3272                 }
3273
3274                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3275                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3276                 cache->dirty = 0;
3277                 last = cache->key.objectid + cache->key.offset;
3278
3279                 err = write_one_cache_group(trans, root, path, cache);
3280                 if (err) /* File system offline */
3281                         goto out;
3282
3283                 btrfs_put_block_group(cache);
3284         }
3285
3286         while (1) {
3287                 /*
3288                  * I don't think this is needed since we're just marking our
3289                  * preallocated extent as written, but just in case it can't
3290                  * hurt.
3291                  */
3292                 if (last == 0) {
3293                         err = btrfs_run_delayed_refs(trans, root,
3294                                                      (unsigned long)-1);
3295                         if (err) /* File system offline */
3296                                 goto out;
3297                 }
3298
3299                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3300                 while (cache) {
3301                         /*
3302                          * Really this shouldn't happen, but it could if we
3303                          * couldn't write the entire preallocated extent and
3304                          * splitting the extent resulted in a new block.
3305                          */
3306                         if (cache->dirty) {
3307                                 btrfs_put_block_group(cache);
3308                                 goto again;
3309                         }
3310                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3311                                 break;
3312                         cache = next_block_group(root, cache);
3313                 }
3314                 if (!cache) {
3315                         if (last == 0)
3316                                 break;
3317                         last = 0;
3318                         continue;
3319                 }
3320
3321                 err = btrfs_write_out_cache(root, trans, cache, path);
3322
3323                 /*
3324                  * If we didn't have an error then the cache state is still
3325                  * NEED_WRITE, so we can set it to WRITTEN.
3326                  */
3327                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3328                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3329                 last = cache->key.objectid + cache->key.offset;
3330                 btrfs_put_block_group(cache);
3331         }
3332 out:
3333
3334         btrfs_free_path(path);
3335         return err;
3336 }
3337
3338 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3339 {
3340         struct btrfs_block_group_cache *block_group;
3341         int readonly = 0;
3342
3343         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3344         if (!block_group || block_group->ro)
3345                 readonly = 1;
3346         if (block_group)
3347                 btrfs_put_block_group(block_group);
3348         return readonly;
3349 }
3350
3351 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3352                              u64 total_bytes, u64 bytes_used,
3353                              struct btrfs_space_info **space_info)
3354 {
3355         struct btrfs_space_info *found;
3356         int i;
3357         int factor;
3358
3359         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3360                      BTRFS_BLOCK_GROUP_RAID10))
3361                 factor = 2;
3362         else
3363                 factor = 1;
3364
3365         found = __find_space_info(info, flags);
3366         if (found) {
3367                 spin_lock(&found->lock);
3368                 found->total_bytes += total_bytes;
3369                 found->disk_total += total_bytes * factor;
3370                 found->bytes_used += bytes_used;
3371                 found->disk_used += bytes_used * factor;
3372                 found->full = 0;
3373                 spin_unlock(&found->lock);
3374                 *space_info = found;
3375                 return 0;
3376         }
3377         found = kzalloc(sizeof(*found), GFP_NOFS);
3378         if (!found)
3379                 return -ENOMEM;
3380
3381         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3382                 INIT_LIST_HEAD(&found->block_groups[i]);
3383         init_rwsem(&found->groups_sem);
3384         spin_lock_init(&found->lock);
3385         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3386         found->total_bytes = total_bytes;
3387         found->disk_total = total_bytes * factor;
3388         found->bytes_used = bytes_used;
3389         found->disk_used = bytes_used * factor;
3390         found->bytes_pinned = 0;
3391         found->bytes_reserved = 0;
3392         found->bytes_readonly = 0;
3393         found->bytes_may_use = 0;
3394         found->full = 0;
3395         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3396         found->chunk_alloc = 0;
3397         found->flush = 0;
3398         init_waitqueue_head(&found->wait);
3399         *space_info = found;
3400         list_add_rcu(&found->list, &info->space_info);
3401         if (flags & BTRFS_BLOCK_GROUP_DATA)
3402                 info->data_sinfo = found;
3403         return 0;
3404 }
3405
3406 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3407 {
3408         u64 extra_flags = chunk_to_extended(flags) &
3409                                 BTRFS_EXTENDED_PROFILE_MASK;
3410
3411         write_seqlock(&fs_info->profiles_lock);
3412         if (flags & BTRFS_BLOCK_GROUP_DATA)
3413                 fs_info->avail_data_alloc_bits |= extra_flags;
3414         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3415                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3416         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3417                 fs_info->avail_system_alloc_bits |= extra_flags;
3418         write_sequnlock(&fs_info->profiles_lock);
3419 }
3420
3421 /*
3422  * returns target flags in extended format or 0 if restripe for this
3423  * chunk_type is not in progress
3424  *
3425  * should be called with either volume_mutex or balance_lock held
3426  */
3427 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3428 {
3429         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3430         u64 target = 0;
3431
3432         if (!bctl)
3433                 return 0;
3434
3435         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3436             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3437                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3438         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3439                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3440                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3441         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3442                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3443                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3444         }
3445
3446         return target;
3447 }
3448
3449 /*
3450  * @flags: available profiles in extended format (see ctree.h)
3451  *
3452  * Returns reduced profile in chunk format.  If profile changing is in
3453  * progress (either running or paused) picks the target profile (if it's
3454  * already available), otherwise falls back to plain reducing.
3455  */
3456 u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3457 {
3458         /*
3459          * we add in the count of missing devices because we want
3460          * to make sure that any RAID levels on a degraded FS
3461          * continue to be honored.
3462          */
3463         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3464                 root->fs_info->fs_devices->missing_devices;
3465         u64 target;
3466         u64 tmp;
3467
3468         /*
3469          * see if restripe for this chunk_type is in progress, if so
3470          * try to reduce to the target profile
3471          */
3472         spin_lock(&root->fs_info->balance_lock);
3473         target = get_restripe_target(root->fs_info, flags);
3474         if (target) {
3475                 /* pick target profile only if it's already available */
3476                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3477                         spin_unlock(&root->fs_info->balance_lock);
3478                         return extended_to_chunk(target);
3479                 }
3480         }
3481         spin_unlock(&root->fs_info->balance_lock);
3482
3483         /* First, mask out the RAID levels which aren't possible */
3484         if (num_devices == 1)
3485                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3486                            BTRFS_BLOCK_GROUP_RAID5);
3487         if (num_devices < 3)
3488                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3489         if (num_devices < 4)
3490                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3491
3492         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3493                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3494                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3495         flags &= ~tmp;
3496
3497         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3498                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3499         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3500                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3501         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3502                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3503         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3504                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3505         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3506                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3507
3508         return extended_to_chunk(flags | tmp);
3509 }
3510
3511 static u64 get_alloc_profile(struct btrfs_root *root, u64 flags)
3512 {
3513         unsigned seq;
3514
3515         do {
3516                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3517
3518                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3519                         flags |= root->fs_info->avail_data_alloc_bits;
3520                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3521                         flags |= root->fs_info->avail_system_alloc_bits;
3522                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3523                         flags |= root->fs_info->avail_metadata_alloc_bits;
3524         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3525
3526         return btrfs_reduce_alloc_profile(root, flags);
3527 }
3528
3529 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3530 {
3531         u64 flags;
3532         u64 ret;
3533
3534         if (data)
3535                 flags = BTRFS_BLOCK_GROUP_DATA;
3536         else if (root == root->fs_info->chunk_root)
3537                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3538         else
3539                 flags = BTRFS_BLOCK_GROUP_METADATA;
3540
3541         ret = get_alloc_profile(root, flags);
3542         return ret;
3543 }
3544
3545 /*
3546  * This will check the space that the inode allocates from to make sure we have
3547  * enough space for bytes.
3548  */
3549 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3550 {
3551         struct btrfs_space_info *data_sinfo;
3552         struct btrfs_root *root = BTRFS_I(inode)->root;
3553         struct btrfs_fs_info *fs_info = root->fs_info;
3554         u64 used;
3555         int ret = 0, committed = 0, alloc_chunk = 1;
3556
3557         /* make sure bytes are sectorsize aligned */
3558         bytes = ALIGN(bytes, root->sectorsize);
3559
3560         if (root == root->fs_info->tree_root ||
3561             BTRFS_I(inode)->location.objectid == BTRFS_FREE_INO_OBJECTID) {
3562                 alloc_chunk = 0;
3563                 committed = 1;
3564         }
3565
3566         data_sinfo = fs_info->data_sinfo;
3567         if (!data_sinfo)
3568                 goto alloc;
3569
3570 again:
3571         /* make sure we have enough space to handle the data first */
3572         spin_lock(&data_sinfo->lock);
3573         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3574                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3575                 data_sinfo->bytes_may_use;
3576
3577         if (used + bytes > data_sinfo->total_bytes) {
3578                 struct btrfs_trans_handle *trans;
3579
3580                 /*
3581                  * if we don't have enough free bytes in this space then we need
3582                  * to alloc a new chunk.
3583                  */
3584                 if (!data_sinfo->full && alloc_chunk) {
3585                         u64 alloc_target;
3586
3587                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3588                         spin_unlock(&data_sinfo->lock);
3589 alloc:
3590                         alloc_target = btrfs_get_alloc_profile(root, 1);
3591                         trans = btrfs_join_transaction(root);
3592                         if (IS_ERR(trans))
3593                                 return PTR_ERR(trans);
3594
3595                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3596                                              alloc_target,
3597                                              CHUNK_ALLOC_NO_FORCE);
3598                         btrfs_end_transaction(trans, root);
3599                         if (ret < 0) {
3600                                 if (ret != -ENOSPC)
3601                                         return ret;
3602                                 else
3603                                         goto commit_trans;
3604                         }
3605
3606                         if (!data_sinfo)
3607                                 data_sinfo = fs_info->data_sinfo;
3608
3609                         goto again;
3610                 }
3611
3612                 /*
3613                  * If we have less pinned bytes than we want to allocate then
3614                  * don't bother committing the transaction, it won't help us.
3615                  */
3616                 if (data_sinfo->bytes_pinned < bytes)
3617                         committed = 1;
3618                 spin_unlock(&data_sinfo->lock);
3619
3620                 /* commit the current transaction and try again */
3621 commit_trans:
3622                 if (!committed &&
3623                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3624                         committed = 1;
3625                         trans = btrfs_join_transaction(root);
3626                         if (IS_ERR(trans))
3627                                 return PTR_ERR(trans);
3628                         ret = btrfs_commit_transaction(trans, root);
3629                         if (ret)
3630                                 return ret;
3631                         goto again;
3632                 }
3633
3634                 return -ENOSPC;
3635         }
3636         data_sinfo->bytes_may_use += bytes;
3637         trace_btrfs_space_reservation(root->fs_info, "space_info",
3638                                       data_sinfo->flags, bytes, 1);
3639         spin_unlock(&data_sinfo->lock);
3640
3641         return 0;
3642 }
3643
3644 /*
3645  * Called if we need to clear a data reservation for this inode.
3646  */
3647 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3648 {
3649         struct btrfs_root *root = BTRFS_I(inode)->root;
3650         struct btrfs_space_info *data_sinfo;
3651
3652         /* make sure bytes are sectorsize aligned */
3653         bytes = ALIGN(bytes, root->sectorsize);
3654
3655         data_sinfo = root->fs_info->data_sinfo;
3656         spin_lock(&data_sinfo->lock);
3657         data_sinfo->bytes_may_use -= bytes;
3658         trace_btrfs_space_reservation(root->fs_info, "space_info",
3659                                       data_sinfo->flags, bytes, 0);
3660         spin_unlock(&data_sinfo->lock);
3661 }
3662
3663 static void force_metadata_allocation(struct btrfs_fs_info *info)
3664 {
3665         struct list_head *head = &info->space_info;
3666         struct btrfs_space_info *found;
3667
3668         rcu_read_lock();
3669         list_for_each_entry_rcu(found, head, list) {
3670                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3671                         found->force_alloc = CHUNK_ALLOC_FORCE;
3672         }
3673         rcu_read_unlock();
3674 }
3675
3676 static int should_alloc_chunk(struct btrfs_root *root,
3677                               struct btrfs_space_info *sinfo, int force)
3678 {
3679         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3680         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3681         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3682         u64 thresh;
3683
3684         if (force == CHUNK_ALLOC_FORCE)
3685                 return 1;
3686
3687         /*
3688          * We need to take into account the global rsv because for all intents
3689          * and purposes it's used space.  Don't worry about locking the
3690          * global_rsv, it doesn't change except when the transaction commits.
3691          */
3692         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3693                 num_allocated += global_rsv->size;
3694
3695         /*
3696          * in limited mode, we want to have some free space up to
3697          * about 1% of the FS size.
3698          */
3699         if (force == CHUNK_ALLOC_LIMITED) {
3700                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3701                 thresh = max_t(u64, 64 * 1024 * 1024,
3702                                div_factor_fine(thresh, 1));
3703
3704                 if (num_bytes - num_allocated < thresh)
3705                         return 1;
3706         }
3707
3708         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3709                 return 0;
3710         return 1;
3711 }
3712
3713 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3714 {
3715         u64 num_dev;
3716
3717         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
3718                     BTRFS_BLOCK_GROUP_RAID0 |
3719                     BTRFS_BLOCK_GROUP_RAID5 |
3720                     BTRFS_BLOCK_GROUP_RAID6))
3721                 num_dev = root->fs_info->fs_devices->rw_devices;
3722         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3723                 num_dev = 2;
3724         else
3725                 num_dev = 1;    /* DUP or single */
3726
3727         /* metadata for updaing devices and chunk tree */
3728         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3729 }
3730
3731 static void check_system_chunk(struct btrfs_trans_handle *trans,
3732                                struct btrfs_root *root, u64 type)
3733 {
3734         struct btrfs_space_info *info;
3735         u64 left;
3736         u64 thresh;
3737
3738         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3739         spin_lock(&info->lock);
3740         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3741                 info->bytes_reserved - info->bytes_readonly;
3742         spin_unlock(&info->lock);
3743
3744         thresh = get_system_chunk_thresh(root, type);
3745         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3746                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
3747                         left, thresh, type);
3748                 dump_space_info(info, 0, 0);
3749         }
3750
3751         if (left < thresh) {
3752                 u64 flags;
3753
3754                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3755                 btrfs_alloc_chunk(trans, root, flags);
3756         }
3757 }
3758
3759 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3760                           struct btrfs_root *extent_root, u64 flags, int force)
3761 {
3762         struct btrfs_space_info *space_info;
3763         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3764         int wait_for_alloc = 0;
3765         int ret = 0;
3766
3767         /* Don't re-enter if we're already allocating a chunk */
3768         if (trans->allocating_chunk)
3769                 return -ENOSPC;
3770
3771         space_info = __find_space_info(extent_root->fs_info, flags);
3772         if (!space_info) {
3773                 ret = update_space_info(extent_root->fs_info, flags,
3774                                         0, 0, &space_info);
3775                 BUG_ON(ret); /* -ENOMEM */
3776         }
3777         BUG_ON(!space_info); /* Logic error */
3778
3779 again:
3780         spin_lock(&space_info->lock);
3781         if (force < space_info->force_alloc)
3782                 force = space_info->force_alloc;
3783         if (space_info->full) {
3784                 spin_unlock(&space_info->lock);
3785                 return 0;
3786         }
3787
3788         if (!should_alloc_chunk(extent_root, space_info, force)) {
3789                 spin_unlock(&space_info->lock);
3790                 return 0;
3791         } else if (space_info->chunk_alloc) {
3792                 wait_for_alloc = 1;
3793         } else {
3794                 space_info->chunk_alloc = 1;
3795         }
3796
3797         spin_unlock(&space_info->lock);
3798
3799         mutex_lock(&fs_info->chunk_mutex);
3800
3801         /*
3802          * The chunk_mutex is held throughout the entirety of a chunk
3803          * allocation, so once we've acquired the chunk_mutex we know that the
3804          * other guy is done and we need to recheck and see if we should
3805          * allocate.
3806          */
3807         if (wait_for_alloc) {
3808                 mutex_unlock(&fs_info->chunk_mutex);
3809                 wait_for_alloc = 0;
3810                 goto again;
3811         }
3812
3813         trans->allocating_chunk = true;
3814
3815         /*
3816          * If we have mixed data/metadata chunks we want to make sure we keep
3817          * allocating mixed chunks instead of individual chunks.
3818          */
3819         if (btrfs_mixed_space_info(space_info))
3820                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3821
3822         /*
3823          * if we're doing a data chunk, go ahead and make sure that
3824          * we keep a reasonable number of metadata chunks allocated in the
3825          * FS as well.
3826          */
3827         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3828                 fs_info->data_chunk_allocations++;
3829                 if (!(fs_info->data_chunk_allocations %
3830                       fs_info->metadata_ratio))
3831                         force_metadata_allocation(fs_info);
3832         }
3833
3834         /*
3835          * Check if we have enough space in SYSTEM chunk because we may need
3836          * to update devices.
3837          */
3838         check_system_chunk(trans, extent_root, flags);
3839
3840         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3841         trans->allocating_chunk = false;
3842
3843         spin_lock(&space_info->lock);
3844         if (ret < 0 && ret != -ENOSPC)
3845                 goto out;
3846         if (ret)
3847                 space_info->full = 1;
3848         else
3849                 ret = 1;
3850
3851         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3852 out:
3853         space_info->chunk_alloc = 0;
3854         spin_unlock(&space_info->lock);
3855         mutex_unlock(&fs_info->chunk_mutex);
3856         return ret;
3857 }
3858
3859 static int can_overcommit(struct btrfs_root *root,
3860                           struct btrfs_space_info *space_info, u64 bytes,
3861                           enum btrfs_reserve_flush_enum flush)
3862 {
3863         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3864         u64 profile = btrfs_get_alloc_profile(root, 0);
3865         u64 rsv_size = 0;
3866         u64 avail;
3867         u64 used;
3868         u64 to_add;
3869
3870         used = space_info->bytes_used + space_info->bytes_reserved +
3871                 space_info->bytes_pinned + space_info->bytes_readonly;
3872
3873         spin_lock(&global_rsv->lock);
3874         rsv_size = global_rsv->size;
3875         spin_unlock(&global_rsv->lock);
3876
3877         /*
3878          * We only want to allow over committing if we have lots of actual space
3879          * free, but if we don't have enough space to handle the global reserve
3880          * space then we could end up having a real enospc problem when trying
3881          * to allocate a chunk or some other such important allocation.
3882          */
3883         rsv_size <<= 1;
3884         if (used + rsv_size >= space_info->total_bytes)
3885                 return 0;
3886
3887         used += space_info->bytes_may_use;
3888
3889         spin_lock(&root->fs_info->free_chunk_lock);
3890         avail = root->fs_info->free_chunk_space;
3891         spin_unlock(&root->fs_info->free_chunk_lock);
3892
3893         /*
3894          * If we have dup, raid1 or raid10 then only half of the free
3895          * space is actually useable.  For raid56, the space info used
3896          * doesn't include the parity drive, so we don't have to
3897          * change the math
3898          */
3899         if (profile & (BTRFS_BLOCK_GROUP_DUP |
3900                        BTRFS_BLOCK_GROUP_RAID1 |
3901                        BTRFS_BLOCK_GROUP_RAID10))
3902                 avail >>= 1;
3903
3904         to_add = space_info->total_bytes;
3905
3906         /*
3907          * If we aren't flushing all things, let us overcommit up to
3908          * 1/2th of the space. If we can flush, don't let us overcommit
3909          * too much, let it overcommit up to 1/8 of the space.
3910          */
3911         if (flush == BTRFS_RESERVE_FLUSH_ALL)
3912                 to_add >>= 3;
3913         else
3914                 to_add >>= 1;
3915
3916         /*
3917          * Limit the overcommit to the amount of free space we could possibly
3918          * allocate for chunks.
3919          */
3920         to_add = min(avail, to_add);
3921
3922         if (used + bytes < space_info->total_bytes + to_add)
3923                 return 1;
3924         return 0;
3925 }
3926
3927 void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
3928                                   unsigned long nr_pages)
3929 {
3930         struct super_block *sb = root->fs_info->sb;
3931         int started;
3932
3933         /* If we can not start writeback, just sync all the delalloc file. */
3934         started = try_to_writeback_inodes_sb_nr(sb, nr_pages,
3935                                                       WB_REASON_FS_FREE_SPACE);
3936         if (!started) {
3937                 /*
3938                  * We needn't worry the filesystem going from r/w to r/o though
3939                  * we don't acquire ->s_umount mutex, because the filesystem
3940                  * should guarantee the delalloc inodes list be empty after
3941                  * the filesystem is readonly(all dirty pages are written to
3942                  * the disk).
3943                  */
3944                 btrfs_start_delalloc_inodes(root, 0);
3945                 if (!current->journal_info)
3946                         btrfs_wait_ordered_extents(root, 0);
3947         }
3948 }
3949
3950 /*
3951  * shrink metadata reservation for delalloc
3952  */
3953 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
3954                             bool wait_ordered)
3955 {
3956         struct btrfs_block_rsv *block_rsv;
3957         struct btrfs_space_info *space_info;
3958         struct btrfs_trans_handle *trans;
3959         u64 delalloc_bytes;
3960         u64 max_reclaim;
3961         long time_left;
3962         unsigned long nr_pages = (2 * 1024 * 1024) >> PAGE_CACHE_SHIFT;
3963         int loops = 0;
3964         enum btrfs_reserve_flush_enum flush;
3965
3966         trans = (struct btrfs_trans_handle *)current->journal_info;
3967         block_rsv = &root->fs_info->delalloc_block_rsv;
3968         space_info = block_rsv->space_info;
3969
3970         smp_mb();
3971         delalloc_bytes = percpu_counter_sum_positive(
3972                                                 &root->fs_info->delalloc_bytes);
3973         if (delalloc_bytes == 0) {
3974                 if (trans)
3975                         return;
3976                 btrfs_wait_ordered_extents(root, 0);
3977                 return;
3978         }
3979
3980         while (delalloc_bytes && loops < 3) {
3981                 max_reclaim = min(delalloc_bytes, to_reclaim);
3982                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
3983                 btrfs_writeback_inodes_sb_nr(root, nr_pages);
3984                 /*
3985                  * We need to wait for the async pages to actually start before
3986                  * we do anything.
3987                  */
3988                 wait_event(root->fs_info->async_submit_wait,
3989                            !atomic_read(&root->fs_info->async_delalloc_pages));
3990
3991                 if (!trans)
3992                         flush = BTRFS_RESERVE_FLUSH_ALL;
3993                 else
3994                         flush = BTRFS_RESERVE_NO_FLUSH;
3995                 spin_lock(&space_info->lock);
3996                 if (can_overcommit(root, space_info, orig, flush)) {
3997                         spin_unlock(&space_info->lock);
3998                         break;
3999                 }
4000                 spin_unlock(&space_info->lock);
4001
4002                 loops++;
4003                 if (wait_ordered && !trans) {
4004                         btrfs_wait_ordered_extents(root, 0);
4005                 } else {
4006                         time_left = schedule_timeout_killable(1);
4007                         if (time_left)
4008                                 break;
4009                 }
4010                 smp_mb();
4011                 delalloc_bytes = percpu_counter_sum_positive(
4012                                                 &root->fs_info->delalloc_bytes);
4013         }
4014 }
4015
4016 /**
4017  * maybe_commit_transaction - possibly commit the transaction if its ok to
4018  * @root - the root we're allocating for
4019  * @bytes - the number of bytes we want to reserve
4020  * @force - force the commit
4021  *
4022  * This will check to make sure that committing the transaction will actually
4023  * get us somewhere and then commit the transaction if it does.  Otherwise it
4024  * will return -ENOSPC.
4025  */
4026 static int may_commit_transaction(struct btrfs_root *root,
4027                                   struct btrfs_space_info *space_info,
4028                                   u64 bytes, int force)
4029 {
4030         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4031         struct btrfs_trans_handle *trans;
4032
4033         trans = (struct btrfs_trans_handle *)current->journal_info;
4034         if (trans)
4035                 return -EAGAIN;
4036
4037         if (force)
4038                 goto commit;
4039
4040         /* See if there is enough pinned space to make this reservation */
4041         spin_lock(&space_info->lock);
4042         if (space_info->bytes_pinned >= bytes) {
4043                 spin_unlock(&space_info->lock);
4044                 goto commit;
4045         }
4046         spin_unlock(&space_info->lock);
4047
4048         /*
4049          * See if there is some space in the delayed insertion reservation for
4050          * this reservation.
4051          */
4052         if (space_info != delayed_rsv->space_info)
4053                 return -ENOSPC;
4054
4055         spin_lock(&space_info->lock);
4056         spin_lock(&delayed_rsv->lock);
4057         if (space_info->bytes_pinned + delayed_rsv->size < bytes) {
4058                 spin_unlock(&delayed_rsv->lock);
4059                 spin_unlock(&space_info->lock);
4060                 return -ENOSPC;
4061         }
4062         spin_unlock(&delayed_rsv->lock);
4063         spin_unlock(&space_info->lock);
4064
4065 commit:
4066         trans = btrfs_join_transaction(root);
4067         if (IS_ERR(trans))
4068                 return -ENOSPC;
4069
4070         return btrfs_commit_transaction(trans, root);
4071 }
4072
4073 enum flush_state {
4074         FLUSH_DELAYED_ITEMS_NR  =       1,
4075         FLUSH_DELAYED_ITEMS     =       2,
4076         FLUSH_DELALLOC          =       3,
4077         FLUSH_DELALLOC_WAIT     =       4,
4078         ALLOC_CHUNK             =       5,
4079         COMMIT_TRANS            =       6,
4080 };
4081
4082 static int flush_space(struct btrfs_root *root,
4083                        struct btrfs_space_info *space_info, u64 num_bytes,
4084                        u64 orig_bytes, int state)
4085 {
4086         struct btrfs_trans_handle *trans;
4087         int nr;
4088         int ret = 0;
4089
4090         switch (state) {
4091         case FLUSH_DELAYED_ITEMS_NR:
4092         case FLUSH_DELAYED_ITEMS:
4093                 if (state == FLUSH_DELAYED_ITEMS_NR) {
4094                         u64 bytes = btrfs_calc_trans_metadata_size(root, 1);
4095
4096                         nr = (int)div64_u64(num_bytes, bytes);
4097                         if (!nr)
4098                                 nr = 1;
4099                         nr *= 2;
4100                 } else {
4101                         nr = -1;
4102                 }
4103                 trans = btrfs_join_transaction(root);
4104                 if (IS_ERR(trans)) {
4105                         ret = PTR_ERR(trans);
4106                         break;
4107                 }
4108                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4109                 btrfs_end_transaction(trans, root);
4110                 break;
4111         case FLUSH_DELALLOC:
4112         case FLUSH_DELALLOC_WAIT:
4113                 shrink_delalloc(root, num_bytes, orig_bytes,
4114                                 state == FLUSH_DELALLOC_WAIT);
4115                 break;
4116         case ALLOC_CHUNK:
4117                 trans = btrfs_join_transaction(root);
4118                 if (IS_ERR(trans)) {
4119                         ret = PTR_ERR(trans);
4120                         break;
4121                 }
4122                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4123                                      btrfs_get_alloc_profile(root, 0),
4124                                      CHUNK_ALLOC_NO_FORCE);
4125                 btrfs_end_transaction(trans, root);
4126                 if (ret == -ENOSPC)
4127                         ret = 0;
4128                 break;
4129         case COMMIT_TRANS:
4130                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4131                 break;
4132         default:
4133                 ret = -ENOSPC;
4134                 break;
4135         }
4136
4137         return ret;
4138 }
4139 /**
4140  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4141  * @root - the root we're allocating for
4142  * @block_rsv - the block_rsv we're allocating for
4143  * @orig_bytes - the number of bytes we want
4144  * @flush - whether or not we can flush to make our reservation
4145  *
4146  * This will reserve orgi_bytes number of bytes from the space info associated
4147  * with the block_rsv.  If there is not enough space it will make an attempt to
4148  * flush out space to make room.  It will do this by flushing delalloc if
4149  * possible or committing the transaction.  If flush is 0 then no attempts to
4150  * regain reservations will be made and this will fail if there is not enough
4151  * space already.
4152  */
4153 static int reserve_metadata_bytes(struct btrfs_root *root,
4154                                   struct btrfs_block_rsv *block_rsv,
4155                                   u64 orig_bytes,
4156                                   enum btrfs_reserve_flush_enum flush)
4157 {
4158         struct btrfs_space_info *space_info = block_rsv->space_info;
4159         u64 used;
4160         u64 num_bytes = orig_bytes;
4161         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4162         int ret = 0;
4163         bool flushing = false;
4164
4165 again:
4166         ret = 0;
4167         spin_lock(&space_info->lock);
4168         /*
4169          * We only want to wait if somebody other than us is flushing and we
4170          * are actually allowed to flush all things.
4171          */
4172         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4173                space_info->flush) {
4174                 spin_unlock(&space_info->lock);
4175                 /*
4176                  * If we have a trans handle we can't wait because the flusher
4177                  * may have to commit the transaction, which would mean we would
4178                  * deadlock since we are waiting for the flusher to finish, but
4179                  * hold the current transaction open.
4180                  */
4181                 if (current->journal_info)
4182                         return -EAGAIN;
4183                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4184                 /* Must have been killed, return */
4185                 if (ret)
4186                         return -EINTR;
4187
4188                 spin_lock(&space_info->lock);
4189         }
4190
4191         ret = -ENOSPC;
4192         used = space_info->bytes_used + space_info->bytes_reserved +
4193                 space_info->bytes_pinned + space_info->bytes_readonly +
4194                 space_info->bytes_may_use;
4195
4196         /*
4197          * The idea here is that we've not already over-reserved the block group
4198          * then we can go ahead and save our reservation first and then start
4199          * flushing if we need to.  Otherwise if we've already overcommitted
4200          * lets start flushing stuff first and then come back and try to make
4201          * our reservation.
4202          */
4203         if (used <= space_info->total_bytes) {
4204                 if (used + orig_bytes <= space_info->total_bytes) {
4205                         space_info->bytes_may_use += orig_bytes;
4206                         trace_btrfs_space_reservation(root->fs_info,
4207                                 "space_info", space_info->flags, orig_bytes, 1);
4208                         ret = 0;
4209                 } else {
4210                         /*
4211                          * Ok set num_bytes to orig_bytes since we aren't
4212                          * overocmmitted, this way we only try and reclaim what
4213                          * we need.
4214                          */
4215                         num_bytes = orig_bytes;
4216                 }
4217         } else {
4218                 /*
4219                  * Ok we're over committed, set num_bytes to the overcommitted
4220                  * amount plus the amount of bytes that we need for this
4221                  * reservation.
4222                  */
4223                 num_bytes = used - space_info->total_bytes +
4224                         (orig_bytes * 2);
4225         }
4226
4227         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4228                 space_info->bytes_may_use += orig_bytes;
4229                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4230                                               space_info->flags, orig_bytes,
4231                                               1);
4232                 ret = 0;
4233         }
4234
4235         /*
4236          * Couldn't make our reservation, save our place so while we're trying
4237          * to reclaim space we can actually use it instead of somebody else
4238          * stealing it from us.
4239          *
4240          * We make the other tasks wait for the flush only when we can flush
4241          * all things.
4242          */
4243         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4244                 flushing = true;
4245                 space_info->flush = 1;
4246         }
4247
4248         spin_unlock(&space_info->lock);
4249
4250         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4251                 goto out;
4252
4253         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4254                           flush_state);
4255         flush_state++;
4256
4257         /*
4258          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4259          * would happen. So skip delalloc flush.
4260          */
4261         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4262             (flush_state == FLUSH_DELALLOC ||
4263              flush_state == FLUSH_DELALLOC_WAIT))
4264                 flush_state = ALLOC_CHUNK;
4265
4266         if (!ret)
4267                 goto again;
4268         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4269                  flush_state < COMMIT_TRANS)
4270                 goto again;
4271         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4272                  flush_state <= COMMIT_TRANS)
4273                 goto again;
4274
4275 out:
4276         if (ret == -ENOSPC &&
4277             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4278                 struct btrfs_block_rsv *global_rsv =
4279                         &root->fs_info->global_block_rsv;
4280
4281                 if (block_rsv != global_rsv &&
4282                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4283                         ret = 0;
4284         }
4285         if (flushing) {
4286                 spin_lock(&space_info->lock);
4287                 space_info->flush = 0;
4288                 wake_up_all(&space_info->wait);
4289                 spin_unlock(&space_info->lock);
4290         }
4291         return ret;
4292 }
4293
4294 static struct btrfs_block_rsv *get_block_rsv(
4295                                         const struct btrfs_trans_handle *trans,
4296                                         const struct btrfs_root *root)
4297 {
4298         struct btrfs_block_rsv *block_rsv = NULL;
4299
4300         if (root->ref_cows)
4301                 block_rsv = trans->block_rsv;
4302
4303         if (root == root->fs_info->csum_root && trans->adding_csums)
4304                 block_rsv = trans->block_rsv;
4305
4306         if (!block_rsv)
4307                 block_rsv = root->block_rsv;
4308
4309         if (!block_rsv)
4310                 block_rsv = &root->fs_info->empty_block_rsv;
4311
4312         return block_rsv;
4313 }
4314
4315 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4316                                u64 num_bytes)
4317 {
4318         int ret = -ENOSPC;
4319         spin_lock(&block_rsv->lock);
4320         if (block_rsv->reserved >= num_bytes) {
4321                 block_rsv->reserved -= num_bytes;
4322                 if (block_rsv->reserved < block_rsv->size)
4323                         block_rsv->full = 0;
4324                 ret = 0;
4325         }
4326         spin_unlock(&block_rsv->lock);
4327         return ret;
4328 }
4329
4330 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4331                                 u64 num_bytes, int update_size)
4332 {
4333         spin_lock(&block_rsv->lock);
4334         block_rsv->reserved += num_bytes;
4335         if (update_size)
4336                 block_rsv->size += num_bytes;
4337         else if (block_rsv->reserved >= block_rsv->size)
4338                 block_rsv->full = 1;
4339         spin_unlock(&block_rsv->lock);
4340 }
4341
4342 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4343                                     struct btrfs_block_rsv *block_rsv,
4344                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4345 {
4346         struct btrfs_space_info *space_info = block_rsv->space_info;
4347
4348         spin_lock(&block_rsv->lock);
4349         if (num_bytes == (u64)-1)
4350                 num_bytes = block_rsv->size;
4351         block_rsv->size -= num_bytes;
4352         if (block_rsv->reserved >= block_rsv->size) {
4353                 num_bytes = block_rsv->reserved - block_rsv->size;
4354                 block_rsv->reserved = block_rsv->size;
4355                 block_rsv->full = 1;
4356         } else {
4357                 num_bytes = 0;
4358         }
4359         spin_unlock(&block_rsv->lock);
4360
4361         if (num_bytes > 0) {
4362                 if (dest) {
4363                         spin_lock(&dest->lock);
4364                         if (!dest->full) {
4365                                 u64 bytes_to_add;
4366
4367                                 bytes_to_add = dest->size - dest->reserved;
4368                                 bytes_to_add = min(num_bytes, bytes_to_add);
4369                                 dest->reserved += bytes_to_add;
4370                                 if (dest->reserved >= dest->size)
4371                                         dest->full = 1;
4372                                 num_bytes -= bytes_to_add;
4373                         }
4374                         spin_unlock(&dest->lock);
4375                 }
4376                 if (num_bytes) {
4377                         spin_lock(&space_info->lock);
4378                         space_info->bytes_may_use -= num_bytes;
4379                         trace_btrfs_space_reservation(fs_info, "space_info",
4380                                         space_info->flags, num_bytes, 0);
4381                         space_info->reservation_progress++;
4382                         spin_unlock(&space_info->lock);
4383                 }
4384         }
4385 }
4386
4387 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4388                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4389 {
4390         int ret;
4391
4392         ret = block_rsv_use_bytes(src, num_bytes);
4393         if (ret)
4394                 return ret;
4395
4396         block_rsv_add_bytes(dst, num_bytes, 1);
4397         return 0;
4398 }
4399
4400 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4401 {
4402         memset(rsv, 0, sizeof(*rsv));
4403         spin_lock_init(&rsv->lock);
4404         rsv->type = type;
4405 }
4406
4407 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4408                                               unsigned short type)
4409 {
4410         struct btrfs_block_rsv *block_rsv;
4411         struct btrfs_fs_info *fs_info = root->fs_info;
4412
4413         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4414         if (!block_rsv)
4415                 return NULL;
4416
4417         btrfs_init_block_rsv(block_rsv, type);
4418         block_rsv->space_info = __find_space_info(fs_info,
4419                                                   BTRFS_BLOCK_GROUP_METADATA);
4420         return block_rsv;
4421 }
4422
4423 void btrfs_free_block_rsv(struct btrfs_root *root,
4424                           struct btrfs_block_rsv *rsv)
4425 {
4426         if (!rsv)
4427                 return;
4428         btrfs_block_rsv_release(root, rsv, (u64)-1);
4429         kfree(rsv);
4430 }
4431
4432 int btrfs_block_rsv_add(struct btrfs_root *root,
4433                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4434                         enum btrfs_reserve_flush_enum flush)
4435 {
4436         int ret;
4437
4438         if (num_bytes == 0)
4439                 return 0;
4440
4441         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4442         if (!ret) {
4443                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4444                 return 0;
4445         }
4446
4447         return ret;
4448 }
4449
4450 int btrfs_block_rsv_check(struct btrfs_root *root,
4451                           struct btrfs_block_rsv *block_rsv, int min_factor)
4452 {
4453         u64 num_bytes = 0;
4454         int ret = -ENOSPC;
4455
4456         if (!block_rsv)
4457                 return 0;
4458
4459         spin_lock(&block_rsv->lock);
4460         num_bytes = div_factor(block_rsv->size, min_factor);
4461         if (block_rsv->reserved >= num_bytes)
4462                 ret = 0;
4463         spin_unlock(&block_rsv->lock);
4464
4465         return ret;
4466 }
4467
4468 int btrfs_block_rsv_refill(struct btrfs_root *root,
4469                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4470                            enum btrfs_reserve_flush_enum flush)
4471 {
4472         u64 num_bytes = 0;
4473         int ret = -ENOSPC;
4474
4475         if (!block_rsv)
4476                 return 0;
4477
4478         spin_lock(&block_rsv->lock);
4479         num_bytes = min_reserved;
4480         if (block_rsv->reserved >= num_bytes)
4481                 ret = 0;
4482         else
4483                 num_bytes -= block_rsv->reserved;
4484         spin_unlock(&block_rsv->lock);
4485
4486         if (!ret)
4487                 return 0;
4488
4489         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4490         if (!ret) {
4491                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4492                 return 0;
4493         }
4494
4495         return ret;
4496 }
4497
4498 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4499                             struct btrfs_block_rsv *dst_rsv,
4500                             u64 num_bytes)
4501 {
4502         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4503 }
4504
4505 void btrfs_block_rsv_release(struct btrfs_root *root,
4506                              struct btrfs_block_rsv *block_rsv,
4507                              u64 num_bytes)
4508 {
4509         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4510         if (global_rsv->full || global_rsv == block_rsv ||
4511             block_rsv->space_info != global_rsv->space_info)
4512                 global_rsv = NULL;
4513         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4514                                 num_bytes);
4515 }
4516
4517 /*
4518  * helper to calculate size of global block reservation.
4519  * the desired value is sum of space used by extent tree,
4520  * checksum tree and root tree
4521  */
4522 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4523 {
4524         struct btrfs_space_info *sinfo;
4525         u64 num_bytes;
4526         u64 meta_used;
4527         u64 data_used;
4528         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4529
4530         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4531         spin_lock(&sinfo->lock);
4532         data_used = sinfo->bytes_used;
4533         spin_unlock(&sinfo->lock);
4534
4535         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4536         spin_lock(&sinfo->lock);
4537         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4538                 data_used = 0;
4539         meta_used = sinfo->bytes_used;
4540         spin_unlock(&sinfo->lock);
4541
4542         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4543                     csum_size * 2;
4544         num_bytes += div64_u64(data_used + meta_used, 50);
4545
4546         if (num_bytes * 3 > meta_used)
4547                 num_bytes = div64_u64(meta_used, 3);
4548
4549         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4550 }
4551
4552 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4553 {
4554         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4555         struct btrfs_space_info *sinfo = block_rsv->space_info;
4556         u64 num_bytes;
4557
4558         num_bytes = calc_global_metadata_size(fs_info);
4559
4560         spin_lock(&sinfo->lock);
4561         spin_lock(&block_rsv->lock);
4562
4563         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
4564
4565         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4566                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4567                     sinfo->bytes_may_use;
4568
4569         if (sinfo->total_bytes > num_bytes) {
4570                 num_bytes = sinfo->total_bytes - num_bytes;
4571                 block_rsv->reserved += num_bytes;
4572                 sinfo->bytes_may_use += num_bytes;
4573                 trace_btrfs_space_reservation(fs_info, "space_info",
4574                                       sinfo->flags, num_bytes, 1);
4575         }
4576
4577         if (block_rsv->reserved >= block_rsv->size) {
4578                 num_bytes = block_rsv->reserved - block_rsv->size;
4579                 sinfo->bytes_may_use -= num_bytes;
4580                 trace_btrfs_space_reservation(fs_info, "space_info",
4581                                       sinfo->flags, num_bytes, 0);
4582                 sinfo->reservation_progress++;
4583                 block_rsv->reserved = block_rsv->size;
4584                 block_rsv->full = 1;
4585         }
4586
4587         spin_unlock(&block_rsv->lock);
4588         spin_unlock(&sinfo->lock);
4589 }
4590
4591 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4592 {
4593         struct btrfs_space_info *space_info;
4594
4595         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4596         fs_info->chunk_block_rsv.space_info = space_info;
4597
4598         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4599         fs_info->global_block_rsv.space_info = space_info;
4600         fs_info->delalloc_block_rsv.space_info = space_info;
4601         fs_info->trans_block_rsv.space_info = space_info;
4602         fs_info->empty_block_rsv.space_info = space_info;
4603         fs_info->delayed_block_rsv.space_info = space_info;
4604
4605         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4606         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4607         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4608         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4609         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4610
4611         update_global_block_rsv(fs_info);
4612 }
4613
4614 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4615 {
4616         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4617                                 (u64)-1);
4618         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4619         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4620         WARN_ON(fs_info->trans_block_rsv.size > 0);
4621         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4622         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4623         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4624         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4625         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4626 }
4627
4628 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4629                                   struct btrfs_root *root)
4630 {
4631         if (!trans->block_rsv)
4632                 return;
4633
4634         if (!trans->bytes_reserved)
4635                 return;
4636
4637         trace_btrfs_space_reservation(root->fs_info, "transaction",
4638                                       trans->transid, trans->bytes_reserved, 0);
4639         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4640         trans->bytes_reserved = 0;
4641 }
4642
4643 /* Can only return 0 or -ENOSPC */
4644 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4645                                   struct inode *inode)
4646 {
4647         struct btrfs_root *root = BTRFS_I(inode)->root;
4648         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4649         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4650
4651         /*
4652          * We need to hold space in order to delete our orphan item once we've
4653          * added it, so this takes the reservation so we can release it later
4654          * when we are truly done with the orphan item.
4655          */
4656         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4657         trace_btrfs_space_reservation(root->fs_info, "orphan",
4658                                       btrfs_ino(inode), num_bytes, 1);
4659         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4660 }
4661
4662 void btrfs_orphan_release_metadata(struct inode *inode)
4663 {
4664         struct btrfs_root *root = BTRFS_I(inode)->root;
4665         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4666         trace_btrfs_space_reservation(root->fs_info, "orphan",
4667                                       btrfs_ino(inode), num_bytes, 0);
4668         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4669 }
4670
4671 /*
4672  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
4673  * root: the root of the parent directory
4674  * rsv: block reservation
4675  * items: the number of items that we need do reservation
4676  * qgroup_reserved: used to return the reserved size in qgroup
4677  *
4678  * This function is used to reserve the space for snapshot/subvolume
4679  * creation and deletion. Those operations are different with the
4680  * common file/directory operations, they change two fs/file trees
4681  * and root tree, the number of items that the qgroup reserves is
4682  * different with the free space reservation. So we can not use
4683  * the space reseravtion mechanism in start_transaction().
4684  */
4685 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
4686                                      struct btrfs_block_rsv *rsv,
4687                                      int items,
4688                                      u64 *qgroup_reserved)
4689 {
4690         u64 num_bytes;
4691         int ret;
4692
4693         if (root->fs_info->quota_enabled) {
4694                 /* One for parent inode, two for dir entries */
4695                 num_bytes = 3 * root->leafsize;
4696                 ret = btrfs_qgroup_reserve(root, num_bytes);
4697                 if (ret)
4698                         return ret;
4699         } else {
4700                 num_bytes = 0;
4701         }
4702
4703         *qgroup_reserved = num_bytes;
4704
4705         num_bytes = btrfs_calc_trans_metadata_size(root, items);
4706         rsv->space_info = __find_space_info(root->fs_info,
4707                                             BTRFS_BLOCK_GROUP_METADATA);
4708         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
4709                                   BTRFS_RESERVE_FLUSH_ALL);
4710         if (ret) {
4711                 if (*qgroup_reserved)
4712                         btrfs_qgroup_free(root, *qgroup_reserved);
4713         }
4714
4715         return ret;
4716 }
4717
4718 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
4719                                       struct btrfs_block_rsv *rsv,
4720                                       u64 qgroup_reserved)
4721 {
4722         btrfs_block_rsv_release(root, rsv, (u64)-1);
4723         if (qgroup_reserved)
4724                 btrfs_qgroup_free(root, qgroup_reserved);
4725 }
4726
4727 /**
4728  * drop_outstanding_extent - drop an outstanding extent
4729  * @inode: the inode we're dropping the extent for
4730  *
4731  * This is called when we are freeing up an outstanding extent, either called
4732  * after an error or after an extent is written.  This will return the number of
4733  * reserved extents that need to be freed.  This must be called with
4734  * BTRFS_I(inode)->lock held.
4735  */
4736 static unsigned drop_outstanding_extent(struct inode *inode)
4737 {
4738         unsigned drop_inode_space = 0;
4739         unsigned dropped_extents = 0;
4740
4741         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
4742         BTRFS_I(inode)->outstanding_extents--;
4743
4744         if (BTRFS_I(inode)->outstanding_extents == 0 &&
4745             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4746                                &BTRFS_I(inode)->runtime_flags))
4747                 drop_inode_space = 1;
4748
4749         /*
4750          * If we have more or the same amount of outsanding extents than we have
4751          * reserved then we need to leave the reserved extents count alone.
4752          */
4753         if (BTRFS_I(inode)->outstanding_extents >=
4754             BTRFS_I(inode)->reserved_extents)
4755                 return drop_inode_space;
4756
4757         dropped_extents = BTRFS_I(inode)->reserved_extents -
4758                 BTRFS_I(inode)->outstanding_extents;
4759         BTRFS_I(inode)->reserved_extents -= dropped_extents;
4760         return dropped_extents + drop_inode_space;
4761 }
4762
4763 /**
4764  * calc_csum_metadata_size - return the amount of metada space that must be
4765  *      reserved/free'd for the given bytes.
4766  * @inode: the inode we're manipulating
4767  * @num_bytes: the number of bytes in question
4768  * @reserve: 1 if we are reserving space, 0 if we are freeing space
4769  *
4770  * This adjusts the number of csum_bytes in the inode and then returns the
4771  * correct amount of metadata that must either be reserved or freed.  We
4772  * calculate how many checksums we can fit into one leaf and then divide the
4773  * number of bytes that will need to be checksumed by this value to figure out
4774  * how many checksums will be required.  If we are adding bytes then the number
4775  * may go up and we will return the number of additional bytes that must be
4776  * reserved.  If it is going down we will return the number of bytes that must
4777  * be freed.
4778  *
4779  * This must be called with BTRFS_I(inode)->lock held.
4780  */
4781 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
4782                                    int reserve)
4783 {
4784         struct btrfs_root *root = BTRFS_I(inode)->root;
4785         u64 csum_size;
4786         int num_csums_per_leaf;
4787         int num_csums;
4788         int old_csums;
4789
4790         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
4791             BTRFS_I(inode)->csum_bytes == 0)
4792                 return 0;
4793
4794         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4795         if (reserve)
4796                 BTRFS_I(inode)->csum_bytes += num_bytes;
4797         else
4798                 BTRFS_I(inode)->csum_bytes -= num_bytes;
4799         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
4800         num_csums_per_leaf = (int)div64_u64(csum_size,
4801                                             sizeof(struct btrfs_csum_item) +
4802                                             sizeof(struct btrfs_disk_key));
4803         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
4804         num_csums = num_csums + num_csums_per_leaf - 1;
4805         num_csums = num_csums / num_csums_per_leaf;
4806
4807         old_csums = old_csums + num_csums_per_leaf - 1;
4808         old_csums = old_csums / num_csums_per_leaf;
4809
4810         /* No change, no need to reserve more */
4811         if (old_csums == num_csums)
4812                 return 0;
4813
4814         if (reserve)
4815                 return btrfs_calc_trans_metadata_size(root,
4816                                                       num_csums - old_csums);
4817
4818         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
4819 }
4820
4821 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
4822 {
4823         struct btrfs_root *root = BTRFS_I(inode)->root;
4824         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
4825         u64 to_reserve = 0;
4826         u64 csum_bytes;
4827         unsigned nr_extents = 0;
4828         int extra_reserve = 0;
4829         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
4830         int ret = 0;
4831         bool delalloc_lock = true;
4832         u64 to_free = 0;
4833         unsigned dropped;
4834
4835         /* If we are a free space inode we need to not flush since we will be in
4836          * the middle of a transaction commit.  We also don't need the delalloc
4837          * mutex since we won't race with anybody.  We need this mostly to make
4838          * lockdep shut its filthy mouth.
4839          */
4840         if (btrfs_is_free_space_inode(inode)) {
4841                 flush = BTRFS_RESERVE_NO_FLUSH;
4842                 delalloc_lock = false;
4843         }
4844
4845         if (flush != BTRFS_RESERVE_NO_FLUSH &&
4846             btrfs_transaction_in_commit(root->fs_info))
4847                 schedule_timeout(1);
4848
4849         if (delalloc_lock)
4850                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
4851
4852         num_bytes = ALIGN(num_bytes, root->sectorsize);
4853
4854         spin_lock(&BTRFS_I(inode)->lock);
4855         BTRFS_I(inode)->outstanding_extents++;
4856
4857         if (BTRFS_I(inode)->outstanding_extents >
4858             BTRFS_I(inode)->reserved_extents)
4859                 nr_extents = BTRFS_I(inode)->outstanding_extents -
4860                         BTRFS_I(inode)->reserved_extents;
4861
4862         /*
4863          * Add an item to reserve for updating the inode when we complete the
4864          * delalloc io.
4865          */
4866         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4867                       &BTRFS_I(inode)->runtime_flags)) {
4868                 nr_extents++;
4869                 extra_reserve = 1;
4870         }
4871
4872         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
4873         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
4874         csum_bytes = BTRFS_I(inode)->csum_bytes;
4875         spin_unlock(&BTRFS_I(inode)->lock);
4876
4877         if (root->fs_info->quota_enabled) {
4878                 ret = btrfs_qgroup_reserve(root, num_bytes +
4879                                            nr_extents * root->leafsize);
4880                 if (ret)
4881                         goto out_fail;
4882         }
4883
4884         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
4885         if (unlikely(ret)) {
4886                 if (root->fs_info->quota_enabled)
4887                         btrfs_qgroup_free(root, num_bytes +
4888                                                 nr_extents * root->leafsize);
4889                 goto out_fail;
4890         }
4891
4892         spin_lock(&BTRFS_I(inode)->lock);
4893         if (extra_reserve) {
4894                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
4895                         &BTRFS_I(inode)->runtime_flags);
4896                 nr_extents--;
4897         }
4898         BTRFS_I(inode)->reserved_extents += nr_extents;
4899         spin_unlock(&BTRFS_I(inode)->lock);
4900
4901         if (delalloc_lock)
4902                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4903
4904         if (to_reserve)
4905                 trace_btrfs_space_reservation(root->fs_info,"delalloc",
4906                                               btrfs_ino(inode), to_reserve, 1);
4907         block_rsv_add_bytes(block_rsv, to_reserve, 1);
4908
4909         return 0;
4910
4911 out_fail:
4912         spin_lock(&BTRFS_I(inode)->lock);
4913         dropped = drop_outstanding_extent(inode);
4914         /*
4915          * If the inodes csum_bytes is the same as the original
4916          * csum_bytes then we know we haven't raced with any free()ers
4917          * so we can just reduce our inodes csum bytes and carry on.
4918          */
4919         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
4920                 calc_csum_metadata_size(inode, num_bytes, 0);
4921         } else {
4922                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
4923                 u64 bytes;
4924
4925                 /*
4926                  * This is tricky, but first we need to figure out how much we
4927                  * free'd from any free-ers that occured during this
4928                  * reservation, so we reset ->csum_bytes to the csum_bytes
4929                  * before we dropped our lock, and then call the free for the
4930                  * number of bytes that were freed while we were trying our
4931                  * reservation.
4932                  */
4933                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
4934                 BTRFS_I(inode)->csum_bytes = csum_bytes;
4935                 to_free = calc_csum_metadata_size(inode, bytes, 0);
4936
4937
4938                 /*
4939                  * Now we need to see how much we would have freed had we not
4940                  * been making this reservation and our ->csum_bytes were not
4941                  * artificially inflated.
4942                  */
4943                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
4944                 bytes = csum_bytes - orig_csum_bytes;
4945                 bytes = calc_csum_metadata_size(inode, bytes, 0);
4946
4947                 /*
4948                  * Now reset ->csum_bytes to what it should be.  If bytes is
4949                  * more than to_free then we would have free'd more space had we
4950                  * not had an artificially high ->csum_bytes, so we need to free
4951                  * the remainder.  If bytes is the same or less then we don't
4952                  * need to do anything, the other free-ers did the correct
4953                  * thing.
4954                  */
4955                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
4956                 if (bytes > to_free)
4957                         to_free = bytes - to_free;
4958                 else
4959                         to_free = 0;
4960         }
4961         spin_unlock(&BTRFS_I(inode)->lock);
4962         if (dropped)
4963                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4964
4965         if (to_free) {
4966                 btrfs_block_rsv_release(root, block_rsv, to_free);
4967                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
4968                                               btrfs_ino(inode), to_free, 0);
4969         }
4970         if (delalloc_lock)
4971                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
4972         return ret;
4973 }
4974
4975 /**
4976  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
4977  * @inode: the inode to release the reservation for
4978  * @num_bytes: the number of bytes we're releasing
4979  *
4980  * This will release the metadata reservation for an inode.  This can be called
4981  * once we complete IO for a given set of bytes to release their metadata
4982  * reservations.
4983  */
4984 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
4985 {
4986         struct btrfs_root *root = BTRFS_I(inode)->root;
4987         u64 to_free = 0;
4988         unsigned dropped;
4989
4990         num_bytes = ALIGN(num_bytes, root->sectorsize);
4991         spin_lock(&BTRFS_I(inode)->lock);
4992         dropped = drop_outstanding_extent(inode);
4993
4994         if (num_bytes)
4995                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
4996         spin_unlock(&BTRFS_I(inode)->lock);
4997         if (dropped > 0)
4998                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
4999
5000         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5001                                       btrfs_ino(inode), to_free, 0);
5002         if (root->fs_info->quota_enabled) {
5003                 btrfs_qgroup_free(root, num_bytes +
5004                                         dropped * root->leafsize);
5005         }
5006
5007         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5008                                 to_free);
5009 }
5010
5011 /**
5012  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5013  * @inode: inode we're writing to
5014  * @num_bytes: the number of bytes we want to allocate
5015  *
5016  * This will do the following things
5017  *
5018  * o reserve space in the data space info for num_bytes
5019  * o reserve space in the metadata space info based on number of outstanding
5020  *   extents and how much csums will be needed
5021  * o add to the inodes ->delalloc_bytes
5022  * o add it to the fs_info's delalloc inodes list.
5023  *
5024  * This will return 0 for success and -ENOSPC if there is no space left.
5025  */
5026 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5027 {
5028         int ret;
5029
5030         ret = btrfs_check_data_free_space(inode, num_bytes);
5031         if (ret)
5032                 return ret;
5033
5034         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5035         if (ret) {
5036                 btrfs_free_reserved_data_space(inode, num_bytes);
5037                 return ret;
5038         }
5039
5040         return 0;
5041 }
5042
5043 /**
5044  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5045  * @inode: inode we're releasing space for
5046  * @num_bytes: the number of bytes we want to free up
5047  *
5048  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5049  * called in the case that we don't need the metadata AND data reservations
5050  * anymore.  So if there is an error or we insert an inline extent.
5051  *
5052  * This function will release the metadata space that was not used and will
5053  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5054  * list if there are no delalloc bytes left.
5055  */
5056 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5057 {
5058         btrfs_delalloc_release_metadata(inode, num_bytes);
5059         btrfs_free_reserved_data_space(inode, num_bytes);
5060 }
5061
5062 static int update_block_group(struct btrfs_root *root,
5063                               u64 bytenr, u64 num_bytes, int alloc)
5064 {
5065         struct btrfs_block_group_cache *cache = NULL;
5066         struct btrfs_fs_info *info = root->fs_info;
5067         u64 total = num_bytes;
5068         u64 old_val;
5069         u64 byte_in_group;
5070         int factor;
5071
5072         /* block accounting for super block */
5073         spin_lock(&info->delalloc_lock);
5074         old_val = btrfs_super_bytes_used(info->super_copy);
5075         if (alloc)
5076                 old_val += num_bytes;
5077         else
5078                 old_val -= num_bytes;
5079         btrfs_set_super_bytes_used(info->super_copy, old_val);
5080         spin_unlock(&info->delalloc_lock);
5081
5082         while (total) {
5083                 cache = btrfs_lookup_block_group(info, bytenr);
5084                 if (!cache)
5085                         return -ENOENT;
5086                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5087                                     BTRFS_BLOCK_GROUP_RAID1 |
5088                                     BTRFS_BLOCK_GROUP_RAID10))
5089                         factor = 2;
5090                 else
5091                         factor = 1;
5092                 /*
5093                  * If this block group has free space cache written out, we
5094                  * need to make sure to load it if we are removing space.  This
5095                  * is because we need the unpinning stage to actually add the
5096                  * space back to the block group, otherwise we will leak space.
5097                  */
5098                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5099                         cache_block_group(cache, 1);
5100
5101                 byte_in_group = bytenr - cache->key.objectid;
5102                 WARN_ON(byte_in_group > cache->key.offset);
5103
5104                 spin_lock(&cache->space_info->lock);
5105                 spin_lock(&cache->lock);
5106
5107                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5108                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5109                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5110
5111                 cache->dirty = 1;
5112                 old_val = btrfs_block_group_used(&cache->item);
5113                 num_bytes = min(total, cache->key.offset - byte_in_group);
5114                 if (alloc) {
5115                         old_val += num_bytes;
5116                         btrfs_set_block_group_used(&cache->item, old_val);
5117                         cache->reserved -= num_bytes;
5118                         cache->space_info->bytes_reserved -= num_bytes;
5119                         cache->space_info->bytes_used += num_bytes;
5120                         cache->space_info->disk_used += num_bytes * factor;
5121                         spin_unlock(&cache->lock);
5122                         spin_unlock(&cache->space_info->lock);
5123                 } else {
5124                         old_val -= num_bytes;
5125                         btrfs_set_block_group_used(&cache->item, old_val);
5126                         cache->pinned += num_bytes;
5127                         cache->space_info->bytes_pinned += num_bytes;
5128                         cache->space_info->bytes_used -= num_bytes;
5129                         cache->space_info->disk_used -= num_bytes * factor;
5130                         spin_unlock(&cache->lock);
5131                         spin_unlock(&cache->space_info->lock);
5132
5133                         set_extent_dirty(info->pinned_extents,
5134                                          bytenr, bytenr + num_bytes - 1,
5135                                          GFP_NOFS | __GFP_NOFAIL);
5136                 }
5137                 btrfs_put_block_group(cache);
5138                 total -= num_bytes;
5139                 bytenr += num_bytes;
5140         }
5141         return 0;
5142 }
5143
5144 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5145 {
5146         struct btrfs_block_group_cache *cache;
5147         u64 bytenr;
5148
5149         spin_lock(&root->fs_info->block_group_cache_lock);
5150         bytenr = root->fs_info->first_logical_byte;
5151         spin_unlock(&root->fs_info->block_group_cache_lock);
5152
5153         if (bytenr < (u64)-1)
5154                 return bytenr;
5155
5156         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5157         if (!cache)
5158                 return 0;
5159
5160         bytenr = cache->key.objectid;
5161         btrfs_put_block_group(cache);
5162
5163         return bytenr;
5164 }
5165
5166 static int pin_down_extent(struct btrfs_root *root,
5167                            struct btrfs_block_group_cache *cache,
5168                            u64 bytenr, u64 num_bytes, int reserved)
5169 {
5170         spin_lock(&cache->space_info->lock);
5171         spin_lock(&cache->lock);
5172         cache->pinned += num_bytes;
5173         cache->space_info->bytes_pinned += num_bytes;
5174         if (reserved) {
5175                 cache->reserved -= num_bytes;
5176                 cache->space_info->bytes_reserved -= num_bytes;
5177         }
5178         spin_unlock(&cache->lock);
5179         spin_unlock(&cache->space_info->lock);
5180
5181         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5182                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5183         return 0;
5184 }
5185
5186 /*
5187  * this function must be called within transaction
5188  */
5189 int btrfs_pin_extent(struct btrfs_root *root,
5190                      u64 bytenr, u64 num_bytes, int reserved)
5191 {
5192         struct btrfs_block_group_cache *cache;
5193
5194         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5195         BUG_ON(!cache); /* Logic error */
5196
5197         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5198
5199         btrfs_put_block_group(cache);
5200         return 0;
5201 }
5202
5203 /*
5204  * this function must be called within transaction
5205  */
5206 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5207                                     u64 bytenr, u64 num_bytes)
5208 {
5209         struct btrfs_block_group_cache *cache;
5210
5211         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5212         BUG_ON(!cache); /* Logic error */
5213
5214         /*
5215          * pull in the free space cache (if any) so that our pin
5216          * removes the free space from the cache.  We have load_only set
5217          * to one because the slow code to read in the free extents does check
5218          * the pinned extents.
5219          */
5220         cache_block_group(cache, 1);
5221
5222         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5223
5224         /* remove us from the free space cache (if we're there at all) */
5225         btrfs_remove_free_space(cache, bytenr, num_bytes);
5226         btrfs_put_block_group(cache);
5227         return 0;
5228 }
5229
5230 /**
5231  * btrfs_update_reserved_bytes - update the block_group and space info counters
5232  * @cache:      The cache we are manipulating
5233  * @num_bytes:  The number of bytes in question
5234  * @reserve:    One of the reservation enums
5235  *
5236  * This is called by the allocator when it reserves space, or by somebody who is
5237  * freeing space that was never actually used on disk.  For example if you
5238  * reserve some space for a new leaf in transaction A and before transaction A
5239  * commits you free that leaf, you call this with reserve set to 0 in order to
5240  * clear the reservation.
5241  *
5242  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5243  * ENOSPC accounting.  For data we handle the reservation through clearing the
5244  * delalloc bits in the io_tree.  We have to do this since we could end up
5245  * allocating less disk space for the amount of data we have reserved in the
5246  * case of compression.
5247  *
5248  * If this is a reservation and the block group has become read only we cannot
5249  * make the reservation and return -EAGAIN, otherwise this function always
5250  * succeeds.
5251  */
5252 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5253                                        u64 num_bytes, int reserve)
5254 {
5255         struct btrfs_space_info *space_info = cache->space_info;
5256         int ret = 0;
5257
5258         spin_lock(&space_info->lock);
5259         spin_lock(&cache->lock);
5260         if (reserve != RESERVE_FREE) {
5261                 if (cache->ro) {
5262                         ret = -EAGAIN;
5263                 } else {
5264                         cache->reserved += num_bytes;
5265                         space_info->bytes_reserved += num_bytes;
5266                         if (reserve == RESERVE_ALLOC) {
5267                                 trace_btrfs_space_reservation(cache->fs_info,
5268                                                 "space_info", space_info->flags,
5269                                                 num_bytes, 0);
5270                                 space_info->bytes_may_use -= num_bytes;
5271                         }
5272                 }
5273         } else {
5274                 if (cache->ro)
5275                         space_info->bytes_readonly += num_bytes;
5276                 cache->reserved -= num_bytes;
5277                 space_info->bytes_reserved -= num_bytes;
5278                 space_info->reservation_progress++;
5279         }
5280         spin_unlock(&cache->lock);
5281         spin_unlock(&space_info->lock);
5282         return ret;
5283 }
5284
5285 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5286                                 struct btrfs_root *root)
5287 {
5288         struct btrfs_fs_info *fs_info = root->fs_info;
5289         struct btrfs_caching_control *next;
5290         struct btrfs_caching_control *caching_ctl;
5291         struct btrfs_block_group_cache *cache;
5292
5293         down_write(&fs_info->extent_commit_sem);
5294
5295         list_for_each_entry_safe(caching_ctl, next,
5296                                  &fs_info->caching_block_groups, list) {
5297                 cache = caching_ctl->block_group;
5298                 if (block_group_cache_done(cache)) {
5299                         cache->last_byte_to_unpin = (u64)-1;
5300                         list_del_init(&caching_ctl->list);
5301                         put_caching_control(caching_ctl);
5302                 } else {
5303                         cache->last_byte_to_unpin = caching_ctl->progress;
5304                 }
5305         }
5306
5307         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5308                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5309         else
5310                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5311
5312         up_write(&fs_info->extent_commit_sem);
5313
5314         update_global_block_rsv(fs_info);
5315 }
5316
5317 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5318 {
5319         struct btrfs_fs_info *fs_info = root->fs_info;
5320         struct btrfs_block_group_cache *cache = NULL;
5321         struct btrfs_space_info *space_info;
5322         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5323         u64 len;
5324         bool readonly;
5325
5326         while (start <= end) {
5327                 readonly = false;
5328                 if (!cache ||
5329                     start >= cache->key.objectid + cache->key.offset) {
5330                         if (cache)
5331                                 btrfs_put_block_group(cache);
5332                         cache = btrfs_lookup_block_group(fs_info, start);
5333                         BUG_ON(!cache); /* Logic error */
5334                 }
5335
5336                 len = cache->key.objectid + cache->key.offset - start;
5337                 len = min(len, end + 1 - start);
5338
5339                 if (start < cache->last_byte_to_unpin) {
5340                         len = min(len, cache->last_byte_to_unpin - start);
5341                         btrfs_add_free_space(cache, start, len);
5342                 }
5343
5344                 start += len;
5345                 space_info = cache->space_info;
5346
5347                 spin_lock(&space_info->lock);
5348                 spin_lock(&cache->lock);
5349                 cache->pinned -= len;
5350                 space_info->bytes_pinned -= len;
5351                 if (cache->ro) {
5352                         space_info->bytes_readonly += len;
5353                         readonly = true;
5354                 }
5355                 spin_unlock(&cache->lock);
5356                 if (!readonly && global_rsv->space_info == space_info) {
5357                         spin_lock(&global_rsv->lock);
5358                         if (!global_rsv->full) {
5359                                 len = min(len, global_rsv->size -
5360                                           global_rsv->reserved);
5361                                 global_rsv->reserved += len;
5362                                 space_info->bytes_may_use += len;
5363                                 if (global_rsv->reserved >= global_rsv->size)
5364                                         global_rsv->full = 1;
5365                         }
5366                         spin_unlock(&global_rsv->lock);
5367                 }
5368                 spin_unlock(&space_info->lock);
5369         }
5370
5371         if (cache)
5372                 btrfs_put_block_group(cache);
5373         return 0;
5374 }
5375
5376 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5377                                struct btrfs_root *root)
5378 {
5379         struct btrfs_fs_info *fs_info = root->fs_info;
5380         struct extent_io_tree *unpin;
5381         u64 start;
5382         u64 end;
5383         int ret;
5384
5385         if (trans->aborted)
5386                 return 0;
5387
5388         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5389                 unpin = &fs_info->freed_extents[1];
5390         else
5391                 unpin = &fs_info->freed_extents[0];
5392
5393         while (1) {
5394                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5395                                             EXTENT_DIRTY, NULL);
5396                 if (ret)
5397                         break;
5398
5399                 if (btrfs_test_opt(root, DISCARD))
5400                         ret = btrfs_discard_extent(root, start,
5401                                                    end + 1 - start, NULL);
5402
5403                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5404                 unpin_extent_range(root, start, end);
5405                 cond_resched();
5406         }
5407
5408         return 0;
5409 }
5410
5411 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5412                                 struct btrfs_root *root,
5413                                 u64 bytenr, u64 num_bytes, u64 parent,
5414                                 u64 root_objectid, u64 owner_objectid,
5415                                 u64 owner_offset, int refs_to_drop,
5416                                 struct btrfs_delayed_extent_op *extent_op)
5417 {
5418         struct btrfs_key key;
5419         struct btrfs_path *path;
5420         struct btrfs_fs_info *info = root->fs_info;
5421         struct btrfs_root *extent_root = info->extent_root;
5422         struct extent_buffer *leaf;
5423         struct btrfs_extent_item *ei;
5424         struct btrfs_extent_inline_ref *iref;
5425         int ret;
5426         int is_data;
5427         int extent_slot = 0;
5428         int found_extent = 0;
5429         int num_to_del = 1;
5430         u32 item_size;
5431         u64 refs;
5432         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
5433                                                  SKINNY_METADATA);
5434
5435         path = btrfs_alloc_path();
5436         if (!path)
5437                 return -ENOMEM;
5438
5439         path->reada = 1;
5440         path->leave_spinning = 1;
5441
5442         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5443         BUG_ON(!is_data && refs_to_drop != 1);
5444
5445         if (is_data)
5446                 skinny_metadata = 0;
5447
5448         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5449                                     bytenr, num_bytes, parent,
5450                                     root_objectid, owner_objectid,
5451                                     owner_offset);
5452         if (ret == 0) {
5453                 extent_slot = path->slots[0];
5454                 while (extent_slot >= 0) {
5455                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5456                                               extent_slot);
5457                         if (key.objectid != bytenr)
5458                                 break;
5459                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5460                             key.offset == num_bytes) {
5461                                 found_extent = 1;
5462                                 break;
5463                         }
5464                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
5465                             key.offset == owner_objectid) {
5466                                 found_extent = 1;
5467                                 break;
5468                         }
5469                         if (path->slots[0] - extent_slot > 5)
5470                                 break;
5471                         extent_slot--;
5472                 }
5473 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5474                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5475                 if (found_extent && item_size < sizeof(*ei))
5476                         found_extent = 0;
5477 #endif
5478                 if (!found_extent) {
5479                         BUG_ON(iref);
5480                         ret = remove_extent_backref(trans, extent_root, path,
5481                                                     NULL, refs_to_drop,
5482                                                     is_data);
5483                         if (ret) {
5484                                 btrfs_abort_transaction(trans, extent_root, ret);
5485                                 goto out;
5486                         }
5487                         btrfs_release_path(path);
5488                         path->leave_spinning = 1;
5489
5490                         key.objectid = bytenr;
5491                         key.type = BTRFS_EXTENT_ITEM_KEY;
5492                         key.offset = num_bytes;
5493
5494                         if (!is_data && skinny_metadata) {
5495                                 key.type = BTRFS_METADATA_ITEM_KEY;
5496                                 key.offset = owner_objectid;
5497                         }
5498
5499                         ret = btrfs_search_slot(trans, extent_root,
5500                                                 &key, path, -1, 1);
5501                         if (ret > 0 && skinny_metadata && path->slots[0]) {
5502                                 /*
5503                                  * Couldn't find our skinny metadata item,
5504                                  * see if we have ye olde extent item.
5505                                  */
5506                                 path->slots[0]--;
5507                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
5508                                                       path->slots[0]);
5509                                 if (key.objectid == bytenr &&
5510                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
5511                                     key.offset == num_bytes)
5512                                         ret = 0;
5513                         }
5514
5515                         if (ret > 0 && skinny_metadata) {
5516                                 skinny_metadata = false;
5517                                 key.type = BTRFS_EXTENT_ITEM_KEY;
5518                                 key.offset = num_bytes;
5519                                 btrfs_release_path(path);
5520                                 ret = btrfs_search_slot(trans, extent_root,
5521                                                         &key, path, -1, 1);
5522                         }
5523
5524                         if (ret) {
5525                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5526                                         ret, (unsigned long long)bytenr);
5527                                 if (ret > 0)
5528                                         btrfs_print_leaf(extent_root,
5529                                                          path->nodes[0]);
5530                         }
5531                         if (ret < 0) {
5532                                 btrfs_abort_transaction(trans, extent_root, ret);
5533                                 goto out;
5534                         }
5535                         extent_slot = path->slots[0];
5536                 }
5537         } else if (ret == -ENOENT) {
5538                 btrfs_print_leaf(extent_root, path->nodes[0]);
5539                 WARN_ON(1);
5540                 btrfs_err(info,
5541                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
5542                         (unsigned long long)bytenr,
5543                         (unsigned long long)parent,
5544                         (unsigned long long)root_objectid,
5545                         (unsigned long long)owner_objectid,
5546                         (unsigned long long)owner_offset);
5547         } else {
5548                 btrfs_abort_transaction(trans, extent_root, ret);
5549                 goto out;
5550         }
5551
5552         leaf = path->nodes[0];
5553         item_size = btrfs_item_size_nr(leaf, extent_slot);
5554 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5555         if (item_size < sizeof(*ei)) {
5556                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5557                 ret = convert_extent_item_v0(trans, extent_root, path,
5558                                              owner_objectid, 0);
5559                 if (ret < 0) {
5560                         btrfs_abort_transaction(trans, extent_root, ret);
5561                         goto out;
5562                 }
5563
5564                 btrfs_release_path(path);
5565                 path->leave_spinning = 1;
5566
5567                 key.objectid = bytenr;
5568                 key.type = BTRFS_EXTENT_ITEM_KEY;
5569                 key.offset = num_bytes;
5570
5571                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5572                                         -1, 1);
5573                 if (ret) {
5574                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5575                                 ret, (unsigned long long)bytenr);
5576                         btrfs_print_leaf(extent_root, path->nodes[0]);
5577                 }
5578                 if (ret < 0) {
5579                         btrfs_abort_transaction(trans, extent_root, ret);
5580                         goto out;
5581                 }
5582
5583                 extent_slot = path->slots[0];
5584                 leaf = path->nodes[0];
5585                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5586         }
5587 #endif
5588         BUG_ON(item_size < sizeof(*ei));
5589         ei = btrfs_item_ptr(leaf, extent_slot,
5590                             struct btrfs_extent_item);
5591         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
5592             key.type == BTRFS_EXTENT_ITEM_KEY) {
5593                 struct btrfs_tree_block_info *bi;
5594                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
5595                 bi = (struct btrfs_tree_block_info *)(ei + 1);
5596                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
5597         }
5598
5599         refs = btrfs_extent_refs(leaf, ei);
5600         if (refs < refs_to_drop) {
5601                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
5602                           "for bytenr %Lu\n", refs_to_drop, refs, bytenr);
5603                 ret = -EINVAL;
5604                 btrfs_abort_transaction(trans, extent_root, ret);
5605                 goto out;
5606         }
5607         refs -= refs_to_drop;
5608
5609         if (refs > 0) {
5610                 if (extent_op)
5611                         __run_delayed_extent_op(extent_op, leaf, ei);
5612                 /*
5613                  * In the case of inline back ref, reference count will
5614                  * be updated by remove_extent_backref
5615                  */
5616                 if (iref) {
5617                         BUG_ON(!found_extent);
5618                 } else {
5619                         btrfs_set_extent_refs(leaf, ei, refs);
5620                         btrfs_mark_buffer_dirty(leaf);
5621                 }
5622                 if (found_extent) {
5623                         ret = remove_extent_backref(trans, extent_root, path,
5624                                                     iref, refs_to_drop,
5625                                                     is_data);
5626                         if (ret) {
5627                                 btrfs_abort_transaction(trans, extent_root, ret);
5628                                 goto out;
5629                         }
5630                 }
5631         } else {
5632                 if (found_extent) {
5633                         BUG_ON(is_data && refs_to_drop !=
5634                                extent_data_ref_count(root, path, iref));
5635                         if (iref) {
5636                                 BUG_ON(path->slots[0] != extent_slot);
5637                         } else {
5638                                 BUG_ON(path->slots[0] != extent_slot + 1);
5639                                 path->slots[0] = extent_slot;
5640                                 num_to_del = 2;
5641                         }
5642                 }
5643
5644                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
5645                                       num_to_del);
5646                 if (ret) {
5647                         btrfs_abort_transaction(trans, extent_root, ret);
5648                         goto out;
5649                 }
5650                 btrfs_release_path(path);
5651
5652                 if (is_data) {
5653                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
5654                         if (ret) {
5655                                 btrfs_abort_transaction(trans, extent_root, ret);
5656                                 goto out;
5657                         }
5658                 }
5659
5660                 ret = update_block_group(root, bytenr, num_bytes, 0);
5661                 if (ret) {
5662                         btrfs_abort_transaction(trans, extent_root, ret);
5663                         goto out;
5664                 }
5665         }
5666 out:
5667         btrfs_free_path(path);
5668         return ret;
5669 }
5670
5671 /*
5672  * when we free an block, it is possible (and likely) that we free the last
5673  * delayed ref for that extent as well.  This searches the delayed ref tree for
5674  * a given extent, and if there are no other delayed refs to be processed, it
5675  * removes it from the tree.
5676  */
5677 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
5678                                       struct btrfs_root *root, u64 bytenr)
5679 {
5680         struct btrfs_delayed_ref_head *head;
5681         struct btrfs_delayed_ref_root *delayed_refs;
5682         struct btrfs_delayed_ref_node *ref;
5683         struct rb_node *node;
5684         int ret = 0;
5685
5686         delayed_refs = &trans->transaction->delayed_refs;
5687         spin_lock(&delayed_refs->lock);
5688         head = btrfs_find_delayed_ref_head(trans, bytenr);
5689         if (!head)
5690                 goto out;
5691
5692         node = rb_prev(&head->node.rb_node);
5693         if (!node)
5694                 goto out;
5695
5696         ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
5697
5698         /* there are still entries for this ref, we can't drop it */
5699         if (ref->bytenr == bytenr)
5700                 goto out;
5701
5702         if (head->extent_op) {
5703                 if (!head->must_insert_reserved)
5704                         goto out;
5705                 btrfs_free_delayed_extent_op(head->extent_op);
5706                 head->extent_op = NULL;
5707         }
5708
5709         /*
5710          * waiting for the lock here would deadlock.  If someone else has it
5711          * locked they are already in the process of dropping it anyway
5712          */
5713         if (!mutex_trylock(&head->mutex))
5714                 goto out;
5715
5716         /*
5717          * at this point we have a head with no other entries.  Go
5718          * ahead and process it.
5719          */
5720         head->node.in_tree = 0;
5721         rb_erase(&head->node.rb_node, &delayed_refs->root);
5722
5723         delayed_refs->num_entries--;
5724
5725         /*
5726          * we don't take a ref on the node because we're removing it from the
5727          * tree, so we just steal the ref the tree was holding.
5728          */
5729         delayed_refs->num_heads--;
5730         if (list_empty(&head->cluster))
5731                 delayed_refs->num_heads_ready--;
5732
5733         list_del_init(&head->cluster);
5734         spin_unlock(&delayed_refs->lock);
5735
5736         BUG_ON(head->extent_op);
5737         if (head->must_insert_reserved)
5738                 ret = 1;
5739
5740         mutex_unlock(&head->mutex);
5741         btrfs_put_delayed_ref(&head->node);
5742         return ret;
5743 out:
5744         spin_unlock(&delayed_refs->lock);
5745         return 0;
5746 }
5747
5748 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
5749                            struct btrfs_root *root,
5750                            struct extent_buffer *buf,
5751                            u64 parent, int last_ref)
5752 {
5753         struct btrfs_block_group_cache *cache = NULL;
5754         int ret;
5755
5756         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5757                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
5758                                         buf->start, buf->len,
5759                                         parent, root->root_key.objectid,
5760                                         btrfs_header_level(buf),
5761                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
5762                 BUG_ON(ret); /* -ENOMEM */
5763         }
5764
5765         if (!last_ref)
5766                 return;
5767
5768         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
5769
5770         if (btrfs_header_generation(buf) == trans->transid) {
5771                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
5772                         ret = check_ref_cleanup(trans, root, buf->start);
5773                         if (!ret)
5774                                 goto out;
5775                 }
5776
5777                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
5778                         pin_down_extent(root, cache, buf->start, buf->len, 1);
5779                         goto out;
5780                 }
5781
5782                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
5783
5784                 btrfs_add_free_space(cache, buf->start, buf->len);
5785                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
5786         }
5787 out:
5788         /*
5789          * Deleting the buffer, clear the corrupt flag since it doesn't matter
5790          * anymore.
5791          */
5792         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
5793         btrfs_put_block_group(cache);
5794 }
5795
5796 /* Can return -ENOMEM */
5797 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
5798                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
5799                       u64 owner, u64 offset, int for_cow)
5800 {
5801         int ret;
5802         struct btrfs_fs_info *fs_info = root->fs_info;
5803
5804         /*
5805          * tree log blocks never actually go into the extent allocation
5806          * tree, just update pinning info and exit early.
5807          */
5808         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
5809                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
5810                 /* unlocks the pinned mutex */
5811                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
5812                 ret = 0;
5813         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5814                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
5815                                         num_bytes,
5816                                         parent, root_objectid, (int)owner,
5817                                         BTRFS_DROP_DELAYED_REF, NULL, for_cow);
5818         } else {
5819                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
5820                                                 num_bytes,
5821                                                 parent, root_objectid, owner,
5822                                                 offset, BTRFS_DROP_DELAYED_REF,
5823                                                 NULL, for_cow);
5824         }
5825         return ret;
5826 }
5827
5828 static u64 stripe_align(struct btrfs_root *root,
5829                         struct btrfs_block_group_cache *cache,
5830                         u64 val, u64 num_bytes)
5831 {
5832         u64 ret = ALIGN(val, root->stripesize);
5833         return ret;
5834 }
5835
5836 /*
5837  * when we wait for progress in the block group caching, its because
5838  * our allocation attempt failed at least once.  So, we must sleep
5839  * and let some progress happen before we try again.
5840  *
5841  * This function will sleep at least once waiting for new free space to
5842  * show up, and then it will check the block group free space numbers
5843  * for our min num_bytes.  Another option is to have it go ahead
5844  * and look in the rbtree for a free extent of a given size, but this
5845  * is a good start.
5846  */
5847 static noinline int
5848 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
5849                                 u64 num_bytes)
5850 {
5851         struct btrfs_caching_control *caching_ctl;
5852
5853         caching_ctl = get_caching_control(cache);
5854         if (!caching_ctl)
5855                 return 0;
5856
5857         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
5858                    (cache->free_space_ctl->free_space >= num_bytes));
5859
5860         put_caching_control(caching_ctl);
5861         return 0;
5862 }
5863
5864 static noinline int
5865 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
5866 {
5867         struct btrfs_caching_control *caching_ctl;
5868
5869         caching_ctl = get_caching_control(cache);
5870         if (!caching_ctl)
5871                 return 0;
5872
5873         wait_event(caching_ctl->wait, block_group_cache_done(cache));
5874
5875         put_caching_control(caching_ctl);
5876         return 0;
5877 }
5878
5879 int __get_raid_index(u64 flags)
5880 {
5881         if (flags & BTRFS_BLOCK_GROUP_RAID10)
5882                 return BTRFS_RAID_RAID10;
5883         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
5884                 return BTRFS_RAID_RAID1;
5885         else if (flags & BTRFS_BLOCK_GROUP_DUP)
5886                 return BTRFS_RAID_DUP;
5887         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
5888                 return BTRFS_RAID_RAID0;
5889         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
5890                 return BTRFS_RAID_RAID5;
5891         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
5892                 return BTRFS_RAID_RAID6;
5893
5894         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
5895 }
5896
5897 static int get_block_group_index(struct btrfs_block_group_cache *cache)
5898 {
5899         return __get_raid_index(cache->flags);
5900 }
5901
5902 enum btrfs_loop_type {
5903         LOOP_CACHING_NOWAIT = 0,
5904         LOOP_CACHING_WAIT = 1,
5905         LOOP_ALLOC_CHUNK = 2,
5906         LOOP_NO_EMPTY_SIZE = 3,
5907 };
5908
5909 /*
5910  * walks the btree of allocated extents and find a hole of a given size.
5911  * The key ins is changed to record the hole:
5912  * ins->objectid == block start
5913  * ins->flags = BTRFS_EXTENT_ITEM_KEY
5914  * ins->offset == number of blocks
5915  * Any available blocks before search_start are skipped.
5916  */
5917 static noinline int find_free_extent(struct btrfs_trans_handle *trans,
5918                                      struct btrfs_root *orig_root,
5919                                      u64 num_bytes, u64 empty_size,
5920                                      u64 hint_byte, struct btrfs_key *ins,
5921                                      u64 data)
5922 {
5923         int ret = 0;
5924         struct btrfs_root *root = orig_root->fs_info->extent_root;
5925         struct btrfs_free_cluster *last_ptr = NULL;
5926         struct btrfs_block_group_cache *block_group = NULL;
5927         struct btrfs_block_group_cache *used_block_group;
5928         u64 search_start = 0;
5929         int empty_cluster = 2 * 1024 * 1024;
5930         struct btrfs_space_info *space_info;
5931         int loop = 0;
5932         int index = __get_raid_index(data);
5933         int alloc_type = (data & BTRFS_BLOCK_GROUP_DATA) ?
5934                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
5935         bool found_uncached_bg = false;
5936         bool failed_cluster_refill = false;
5937         bool failed_alloc = false;
5938         bool use_cluster = true;
5939         bool have_caching_bg = false;
5940
5941         WARN_ON(num_bytes < root->sectorsize);
5942         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
5943         ins->objectid = 0;
5944         ins->offset = 0;
5945
5946         trace_find_free_extent(orig_root, num_bytes, empty_size, data);
5947
5948         space_info = __find_space_info(root->fs_info, data);
5949         if (!space_info) {
5950                 btrfs_err(root->fs_info, "No space info for %llu", data);
5951                 return -ENOSPC;
5952         }
5953
5954         /*
5955          * If the space info is for both data and metadata it means we have a
5956          * small filesystem and we can't use the clustering stuff.
5957          */
5958         if (btrfs_mixed_space_info(space_info))
5959                 use_cluster = false;
5960
5961         if (data & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
5962                 last_ptr = &root->fs_info->meta_alloc_cluster;
5963                 if (!btrfs_test_opt(root, SSD))
5964                         empty_cluster = 64 * 1024;
5965         }
5966
5967         if ((data & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
5968             btrfs_test_opt(root, SSD)) {
5969                 last_ptr = &root->fs_info->data_alloc_cluster;
5970         }
5971
5972         if (last_ptr) {
5973                 spin_lock(&last_ptr->lock);
5974                 if (last_ptr->block_group)
5975                         hint_byte = last_ptr->window_start;
5976                 spin_unlock(&last_ptr->lock);
5977         }
5978
5979         search_start = max(search_start, first_logical_byte(root, 0));
5980         search_start = max(search_start, hint_byte);
5981
5982         if (!last_ptr)
5983                 empty_cluster = 0;
5984
5985         if (search_start == hint_byte) {
5986                 block_group = btrfs_lookup_block_group(root->fs_info,
5987                                                        search_start);
5988                 used_block_group = block_group;
5989                 /*
5990                  * we don't want to use the block group if it doesn't match our
5991                  * allocation bits, or if its not cached.
5992                  *
5993                  * However if we are re-searching with an ideal block group
5994                  * picked out then we don't care that the block group is cached.
5995                  */
5996                 if (block_group && block_group_bits(block_group, data) &&
5997                     block_group->cached != BTRFS_CACHE_NO) {
5998                         down_read(&space_info->groups_sem);
5999                         if (list_empty(&block_group->list) ||
6000                             block_group->ro) {
6001                                 /*
6002                                  * someone is removing this block group,
6003                                  * we can't jump into the have_block_group
6004                                  * target because our list pointers are not
6005                                  * valid
6006                                  */
6007                                 btrfs_put_block_group(block_group);
6008                                 up_read(&space_info->groups_sem);
6009                         } else {
6010                                 index = get_block_group_index(block_group);
6011                                 goto have_block_group;
6012                         }
6013                 } else if (block_group) {
6014                         btrfs_put_block_group(block_group);
6015                 }
6016         }
6017 search:
6018         have_caching_bg = false;
6019         down_read(&space_info->groups_sem);
6020         list_for_each_entry(block_group, &space_info->block_groups[index],
6021                             list) {
6022                 u64 offset;
6023                 int cached;
6024
6025                 used_block_group = block_group;
6026                 btrfs_get_block_group(block_group);
6027                 search_start = block_group->key.objectid;
6028
6029                 /*
6030                  * this can happen if we end up cycling through all the
6031                  * raid types, but we want to make sure we only allocate
6032                  * for the proper type.
6033                  */
6034                 if (!block_group_bits(block_group, data)) {
6035                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6036                                 BTRFS_BLOCK_GROUP_RAID1 |
6037                                 BTRFS_BLOCK_GROUP_RAID5 |
6038                                 BTRFS_BLOCK_GROUP_RAID6 |
6039                                 BTRFS_BLOCK_GROUP_RAID10;
6040
6041                         /*
6042                          * if they asked for extra copies and this block group
6043                          * doesn't provide them, bail.  This does allow us to
6044                          * fill raid0 from raid1.
6045                          */
6046                         if ((data & extra) && !(block_group->flags & extra))
6047                                 goto loop;
6048                 }
6049
6050 have_block_group:
6051                 cached = block_group_cache_done(block_group);
6052                 if (unlikely(!cached)) {
6053                         found_uncached_bg = true;
6054                         ret = cache_block_group(block_group, 0);
6055                         BUG_ON(ret < 0);
6056                         ret = 0;
6057                 }
6058
6059                 if (unlikely(block_group->ro))
6060                         goto loop;
6061
6062                 /*
6063                  * Ok we want to try and use the cluster allocator, so
6064                  * lets look there
6065                  */
6066                 if (last_ptr) {
6067                         unsigned long aligned_cluster;
6068                         /*
6069                          * the refill lock keeps out other
6070                          * people trying to start a new cluster
6071                          */
6072                         spin_lock(&last_ptr->refill_lock);
6073                         used_block_group = last_ptr->block_group;
6074                         if (used_block_group != block_group &&
6075                             (!used_block_group ||
6076                              used_block_group->ro ||
6077                              !block_group_bits(used_block_group, data))) {
6078                                 used_block_group = block_group;
6079                                 goto refill_cluster;
6080                         }
6081
6082                         if (used_block_group != block_group)
6083                                 btrfs_get_block_group(used_block_group);
6084
6085                         offset = btrfs_alloc_from_cluster(used_block_group,
6086                           last_ptr, num_bytes, used_block_group->key.objectid);
6087                         if (offset) {
6088                                 /* we have a block, we're done */
6089                                 spin_unlock(&last_ptr->refill_lock);
6090                                 trace_btrfs_reserve_extent_cluster(root,
6091                                         block_group, search_start, num_bytes);
6092                                 goto checks;
6093                         }
6094
6095                         WARN_ON(last_ptr->block_group != used_block_group);
6096                         if (used_block_group != block_group) {
6097                                 btrfs_put_block_group(used_block_group);
6098                                 used_block_group = block_group;
6099                         }
6100 refill_cluster:
6101                         BUG_ON(used_block_group != block_group);
6102                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6103                          * set up a new clusters, so lets just skip it
6104                          * and let the allocator find whatever block
6105                          * it can find.  If we reach this point, we
6106                          * will have tried the cluster allocator
6107                          * plenty of times and not have found
6108                          * anything, so we are likely way too
6109                          * fragmented for the clustering stuff to find
6110                          * anything.
6111                          *
6112                          * However, if the cluster is taken from the
6113                          * current block group, release the cluster
6114                          * first, so that we stand a better chance of
6115                          * succeeding in the unclustered
6116                          * allocation.  */
6117                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6118                             last_ptr->block_group != block_group) {
6119                                 spin_unlock(&last_ptr->refill_lock);
6120                                 goto unclustered_alloc;
6121                         }
6122
6123                         /*
6124                          * this cluster didn't work out, free it and
6125                          * start over
6126                          */
6127                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6128
6129                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6130                                 spin_unlock(&last_ptr->refill_lock);
6131                                 goto unclustered_alloc;
6132                         }
6133
6134                         aligned_cluster = max_t(unsigned long,
6135                                                 empty_cluster + empty_size,
6136                                               block_group->full_stripe_len);
6137
6138                         /* allocate a cluster in this block group */
6139                         ret = btrfs_find_space_cluster(trans, root,
6140                                                block_group, last_ptr,
6141                                                search_start, num_bytes,
6142                                                aligned_cluster);
6143                         if (ret == 0) {
6144                                 /*
6145                                  * now pull our allocation out of this
6146                                  * cluster
6147                                  */
6148                                 offset = btrfs_alloc_from_cluster(block_group,
6149                                                   last_ptr, num_bytes,
6150                                                   search_start);
6151                                 if (offset) {
6152                                         /* we found one, proceed */
6153                                         spin_unlock(&last_ptr->refill_lock);
6154                                         trace_btrfs_reserve_extent_cluster(root,
6155                                                 block_group, search_start,
6156                                                 num_bytes);
6157                                         goto checks;
6158                                 }
6159                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6160                                    && !failed_cluster_refill) {
6161                                 spin_unlock(&last_ptr->refill_lock);
6162
6163                                 failed_cluster_refill = true;
6164                                 wait_block_group_cache_progress(block_group,
6165                                        num_bytes + empty_cluster + empty_size);
6166                                 goto have_block_group;
6167                         }
6168
6169                         /*
6170                          * at this point we either didn't find a cluster
6171                          * or we weren't able to allocate a block from our
6172                          * cluster.  Free the cluster we've been trying
6173                          * to use, and go to the next block group
6174                          */
6175                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6176                         spin_unlock(&last_ptr->refill_lock);
6177                         goto loop;
6178                 }
6179
6180 unclustered_alloc:
6181                 spin_lock(&block_group->free_space_ctl->tree_lock);
6182                 if (cached &&
6183                     block_group->free_space_ctl->free_space <
6184                     num_bytes + empty_cluster + empty_size) {
6185                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6186                         goto loop;
6187                 }
6188                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6189
6190                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6191                                                     num_bytes, empty_size);
6192                 /*
6193                  * If we didn't find a chunk, and we haven't failed on this
6194                  * block group before, and this block group is in the middle of
6195                  * caching and we are ok with waiting, then go ahead and wait
6196                  * for progress to be made, and set failed_alloc to true.
6197                  *
6198                  * If failed_alloc is true then we've already waited on this
6199                  * block group once and should move on to the next block group.
6200                  */
6201                 if (!offset && !failed_alloc && !cached &&
6202                     loop > LOOP_CACHING_NOWAIT) {
6203                         wait_block_group_cache_progress(block_group,
6204                                                 num_bytes + empty_size);
6205                         failed_alloc = true;
6206                         goto have_block_group;
6207                 } else if (!offset) {
6208                         if (!cached)
6209                                 have_caching_bg = true;
6210                         goto loop;
6211                 }
6212 checks:
6213                 search_start = stripe_align(root, used_block_group,
6214                                             offset, num_bytes);
6215
6216                 /* move on to the next group */
6217                 if (search_start + num_bytes >
6218                     used_block_group->key.objectid + used_block_group->key.offset) {
6219                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6220                         goto loop;
6221                 }
6222
6223                 if (offset < search_start)
6224                         btrfs_add_free_space(used_block_group, offset,
6225                                              search_start - offset);
6226                 BUG_ON(offset > search_start);
6227
6228                 ret = btrfs_update_reserved_bytes(used_block_group, num_bytes,
6229                                                   alloc_type);
6230                 if (ret == -EAGAIN) {
6231                         btrfs_add_free_space(used_block_group, offset, num_bytes);
6232                         goto loop;
6233                 }
6234
6235                 /* we are all good, lets return */
6236                 ins->objectid = search_start;
6237                 ins->offset = num_bytes;
6238
6239                 trace_btrfs_reserve_extent(orig_root, block_group,
6240                                            search_start, num_bytes);
6241                 if (used_block_group != block_group)
6242                         btrfs_put_block_group(used_block_group);
6243                 btrfs_put_block_group(block_group);
6244                 break;
6245 loop:
6246                 failed_cluster_refill = false;
6247                 failed_alloc = false;
6248                 BUG_ON(index != get_block_group_index(block_group));
6249                 if (used_block_group != block_group)
6250                         btrfs_put_block_group(used_block_group);
6251                 btrfs_put_block_group(block_group);
6252         }
6253         up_read(&space_info->groups_sem);
6254
6255         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6256                 goto search;
6257
6258         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6259                 goto search;
6260
6261         /*
6262          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6263          *                      caching kthreads as we move along
6264          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6265          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6266          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6267          *                      again
6268          */
6269         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6270                 index = 0;
6271                 loop++;
6272                 if (loop == LOOP_ALLOC_CHUNK) {
6273                         ret = do_chunk_alloc(trans, root, data,
6274                                              CHUNK_ALLOC_FORCE);
6275                         /*
6276                          * Do not bail out on ENOSPC since we
6277                          * can do more things.
6278                          */
6279                         if (ret < 0 && ret != -ENOSPC) {
6280                                 btrfs_abort_transaction(trans,
6281                                                         root, ret);
6282                                 goto out;
6283                         }
6284                 }
6285
6286                 if (loop == LOOP_NO_EMPTY_SIZE) {
6287                         empty_size = 0;
6288                         empty_cluster = 0;
6289                 }
6290
6291                 goto search;
6292         } else if (!ins->objectid) {
6293                 ret = -ENOSPC;
6294         } else if (ins->objectid) {
6295                 ret = 0;
6296         }
6297 out:
6298
6299         return ret;
6300 }
6301
6302 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6303                             int dump_block_groups)
6304 {
6305         struct btrfs_block_group_cache *cache;
6306         int index = 0;
6307
6308         spin_lock(&info->lock);
6309         printk(KERN_INFO "space_info %llu has %llu free, is %sfull\n",
6310                (unsigned long long)info->flags,
6311                (unsigned long long)(info->total_bytes - info->bytes_used -
6312                                     info->bytes_pinned - info->bytes_reserved -
6313                                     info->bytes_readonly),
6314                (info->full) ? "" : "not ");
6315         printk(KERN_INFO "space_info total=%llu, used=%llu, pinned=%llu, "
6316                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6317                (unsigned long long)info->total_bytes,
6318                (unsigned long long)info->bytes_used,
6319                (unsigned long long)info->bytes_pinned,
6320                (unsigned long long)info->bytes_reserved,
6321                (unsigned long long)info->bytes_may_use,
6322                (unsigned long long)info->bytes_readonly);
6323         spin_unlock(&info->lock);
6324
6325         if (!dump_block_groups)
6326                 return;
6327
6328         down_read(&info->groups_sem);
6329 again:
6330         list_for_each_entry(cache, &info->block_groups[index], list) {
6331                 spin_lock(&cache->lock);
6332                 printk(KERN_INFO "block group %llu has %llu bytes, %llu used %llu pinned %llu reserved %s\n",
6333                        (unsigned long long)cache->key.objectid,
6334                        (unsigned long long)cache->key.offset,
6335                        (unsigned long long)btrfs_block_group_used(&cache->item),
6336                        (unsigned long long)cache->pinned,
6337                        (unsigned long long)cache->reserved,
6338                        cache->ro ? "[readonly]" : "");
6339                 btrfs_dump_free_space(cache, bytes);
6340                 spin_unlock(&cache->lock);
6341         }
6342         if (++index < BTRFS_NR_RAID_TYPES)
6343                 goto again;
6344         up_read(&info->groups_sem);
6345 }
6346
6347 int btrfs_reserve_extent(struct btrfs_trans_handle *trans,
6348                          struct btrfs_root *root,
6349                          u64 num_bytes, u64 min_alloc_size,
6350                          u64 empty_size, u64 hint_byte,
6351                          struct btrfs_key *ins, u64 data)
6352 {
6353         bool final_tried = false;
6354         int ret;
6355
6356         data = btrfs_get_alloc_profile(root, data);
6357 again:
6358         WARN_ON(num_bytes < root->sectorsize);
6359         ret = find_free_extent(trans, root, num_bytes, empty_size,
6360                                hint_byte, ins, data);
6361
6362         if (ret == -ENOSPC) {
6363                 if (!final_tried) {
6364                         num_bytes = num_bytes >> 1;
6365                         num_bytes = round_down(num_bytes, root->sectorsize);
6366                         num_bytes = max(num_bytes, min_alloc_size);
6367                         if (num_bytes == min_alloc_size)
6368                                 final_tried = true;
6369                         goto again;
6370                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6371                         struct btrfs_space_info *sinfo;
6372
6373                         sinfo = __find_space_info(root->fs_info, data);
6374                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
6375                                 (unsigned long long)data,
6376                                 (unsigned long long)num_bytes);
6377                         if (sinfo)
6378                                 dump_space_info(sinfo, num_bytes, 1);
6379                 }
6380         }
6381
6382         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
6383
6384         return ret;
6385 }
6386
6387 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6388                                         u64 start, u64 len, int pin)
6389 {
6390         struct btrfs_block_group_cache *cache;
6391         int ret = 0;
6392
6393         cache = btrfs_lookup_block_group(root->fs_info, start);
6394         if (!cache) {
6395                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
6396                         (unsigned long long)start);
6397                 return -ENOSPC;
6398         }
6399
6400         if (btrfs_test_opt(root, DISCARD))
6401                 ret = btrfs_discard_extent(root, start, len, NULL);
6402
6403         if (pin)
6404                 pin_down_extent(root, cache, start, len, 1);
6405         else {
6406                 btrfs_add_free_space(cache, start, len);
6407                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6408         }
6409         btrfs_put_block_group(cache);
6410
6411         trace_btrfs_reserved_extent_free(root, start, len);
6412
6413         return ret;
6414 }
6415
6416 int btrfs_free_reserved_extent(struct btrfs_root *root,
6417                                         u64 start, u64 len)
6418 {
6419         return __btrfs_free_reserved_extent(root, start, len, 0);
6420 }
6421
6422 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6423                                        u64 start, u64 len)
6424 {
6425         return __btrfs_free_reserved_extent(root, start, len, 1);
6426 }
6427
6428 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6429                                       struct btrfs_root *root,
6430                                       u64 parent, u64 root_objectid,
6431                                       u64 flags, u64 owner, u64 offset,
6432                                       struct btrfs_key *ins, int ref_mod)
6433 {
6434         int ret;
6435         struct btrfs_fs_info *fs_info = root->fs_info;
6436         struct btrfs_extent_item *extent_item;
6437         struct btrfs_extent_inline_ref *iref;
6438         struct btrfs_path *path;
6439         struct extent_buffer *leaf;
6440         int type;
6441         u32 size;
6442
6443         if (parent > 0)
6444                 type = BTRFS_SHARED_DATA_REF_KEY;
6445         else
6446                 type = BTRFS_EXTENT_DATA_REF_KEY;
6447
6448         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6449
6450         path = btrfs_alloc_path();
6451         if (!path)
6452                 return -ENOMEM;
6453
6454         path->leave_spinning = 1;
6455         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6456                                       ins, size);
6457         if (ret) {
6458                 btrfs_free_path(path);
6459                 return ret;
6460         }
6461
6462         leaf = path->nodes[0];
6463         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6464                                      struct btrfs_extent_item);
6465         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6466         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6467         btrfs_set_extent_flags(leaf, extent_item,
6468                                flags | BTRFS_EXTENT_FLAG_DATA);
6469
6470         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6471         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6472         if (parent > 0) {
6473                 struct btrfs_shared_data_ref *ref;
6474                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6475                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6476                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6477         } else {
6478                 struct btrfs_extent_data_ref *ref;
6479                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6480                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6481                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6482                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6483                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6484         }
6485
6486         btrfs_mark_buffer_dirty(path->nodes[0]);
6487         btrfs_free_path(path);
6488
6489         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6490         if (ret) { /* -ENOENT, logic error */
6491                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6492                         (unsigned long long)ins->objectid,
6493                         (unsigned long long)ins->offset);
6494                 BUG();
6495         }
6496         return ret;
6497 }
6498
6499 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6500                                      struct btrfs_root *root,
6501                                      u64 parent, u64 root_objectid,
6502                                      u64 flags, struct btrfs_disk_key *key,
6503                                      int level, struct btrfs_key *ins)
6504 {
6505         int ret;
6506         struct btrfs_fs_info *fs_info = root->fs_info;
6507         struct btrfs_extent_item *extent_item;
6508         struct btrfs_tree_block_info *block_info;
6509         struct btrfs_extent_inline_ref *iref;
6510         struct btrfs_path *path;
6511         struct extent_buffer *leaf;
6512         u32 size = sizeof(*extent_item) + sizeof(*iref);
6513         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6514                                                  SKINNY_METADATA);
6515
6516         if (!skinny_metadata)
6517                 size += sizeof(*block_info);
6518
6519         path = btrfs_alloc_path();
6520         if (!path)
6521                 return -ENOMEM;
6522
6523         path->leave_spinning = 1;
6524         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6525                                       ins, size);
6526         if (ret) {
6527                 btrfs_free_path(path);
6528                 return ret;
6529         }
6530
6531         leaf = path->nodes[0];
6532         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6533                                      struct btrfs_extent_item);
6534         btrfs_set_extent_refs(leaf, extent_item, 1);
6535         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6536         btrfs_set_extent_flags(leaf, extent_item,
6537                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
6538
6539         if (skinny_metadata) {
6540                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6541         } else {
6542                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
6543                 btrfs_set_tree_block_key(leaf, block_info, key);
6544                 btrfs_set_tree_block_level(leaf, block_info, level);
6545                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
6546         }
6547
6548         if (parent > 0) {
6549                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
6550                 btrfs_set_extent_inline_ref_type(leaf, iref,
6551                                                  BTRFS_SHARED_BLOCK_REF_KEY);
6552                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6553         } else {
6554                 btrfs_set_extent_inline_ref_type(leaf, iref,
6555                                                  BTRFS_TREE_BLOCK_REF_KEY);
6556                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
6557         }
6558
6559         btrfs_mark_buffer_dirty(leaf);
6560         btrfs_free_path(path);
6561
6562         ret = update_block_group(root, ins->objectid, root->leafsize, 1);
6563         if (ret) { /* -ENOENT, logic error */
6564                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6565                         (unsigned long long)ins->objectid,
6566                         (unsigned long long)ins->offset);
6567                 BUG();
6568         }
6569         return ret;
6570 }
6571
6572 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6573                                      struct btrfs_root *root,
6574                                      u64 root_objectid, u64 owner,
6575                                      u64 offset, struct btrfs_key *ins)
6576 {
6577         int ret;
6578
6579         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
6580
6581         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
6582                                          ins->offset, 0,
6583                                          root_objectid, owner, offset,
6584                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
6585         return ret;
6586 }
6587
6588 /*
6589  * this is used by the tree logging recovery code.  It records that
6590  * an extent has been allocated and makes sure to clear the free
6591  * space cache bits as well
6592  */
6593 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
6594                                    struct btrfs_root *root,
6595                                    u64 root_objectid, u64 owner, u64 offset,
6596                                    struct btrfs_key *ins)
6597 {
6598         int ret;
6599         struct btrfs_block_group_cache *block_group;
6600         struct btrfs_caching_control *caching_ctl;
6601         u64 start = ins->objectid;
6602         u64 num_bytes = ins->offset;
6603
6604         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
6605         cache_block_group(block_group, 0);
6606         caching_ctl = get_caching_control(block_group);
6607
6608         if (!caching_ctl) {
6609                 BUG_ON(!block_group_cache_done(block_group));
6610                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6611                 BUG_ON(ret); /* -ENOMEM */
6612         } else {
6613                 mutex_lock(&caching_ctl->mutex);
6614
6615                 if (start >= caching_ctl->progress) {
6616                         ret = add_excluded_extent(root, start, num_bytes);
6617                         BUG_ON(ret); /* -ENOMEM */
6618                 } else if (start + num_bytes <= caching_ctl->progress) {
6619                         ret = btrfs_remove_free_space(block_group,
6620                                                       start, num_bytes);
6621                         BUG_ON(ret); /* -ENOMEM */
6622                 } else {
6623                         num_bytes = caching_ctl->progress - start;
6624                         ret = btrfs_remove_free_space(block_group,
6625                                                       start, num_bytes);
6626                         BUG_ON(ret); /* -ENOMEM */
6627
6628                         start = caching_ctl->progress;
6629                         num_bytes = ins->objectid + ins->offset -
6630                                     caching_ctl->progress;
6631                         ret = add_excluded_extent(root, start, num_bytes);
6632                         BUG_ON(ret); /* -ENOMEM */
6633                 }
6634
6635                 mutex_unlock(&caching_ctl->mutex);
6636                 put_caching_control(caching_ctl);
6637         }
6638
6639         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
6640                                           RESERVE_ALLOC_NO_ACCOUNT);
6641         BUG_ON(ret); /* logic error */
6642         btrfs_put_block_group(block_group);
6643         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
6644                                          0, owner, offset, ins, 1);
6645         return ret;
6646 }
6647
6648 struct extent_buffer *btrfs_init_new_buffer(struct btrfs_trans_handle *trans,
6649                                             struct btrfs_root *root,
6650                                             u64 bytenr, u32 blocksize,
6651                                             int level)
6652 {
6653         struct extent_buffer *buf;
6654
6655         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
6656         if (!buf)
6657                 return ERR_PTR(-ENOMEM);
6658         btrfs_set_header_generation(buf, trans->transid);
6659         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
6660         btrfs_tree_lock(buf);
6661         clean_tree_block(trans, root, buf);
6662         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
6663
6664         btrfs_set_lock_blocking(buf);
6665         btrfs_set_buffer_uptodate(buf);
6666
6667         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
6668                 /*
6669                  * we allow two log transactions at a time, use different
6670                  * EXENT bit to differentiate dirty pages.
6671                  */
6672                 if (root->log_transid % 2 == 0)
6673                         set_extent_dirty(&root->dirty_log_pages, buf->start,
6674                                         buf->start + buf->len - 1, GFP_NOFS);
6675                 else
6676                         set_extent_new(&root->dirty_log_pages, buf->start,
6677                                         buf->start + buf->len - 1, GFP_NOFS);
6678         } else {
6679                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
6680                          buf->start + buf->len - 1, GFP_NOFS);
6681         }
6682         trans->blocks_used++;
6683         /* this returns a buffer locked for blocking */
6684         return buf;
6685 }
6686
6687 static struct btrfs_block_rsv *
6688 use_block_rsv(struct btrfs_trans_handle *trans,
6689               struct btrfs_root *root, u32 blocksize)
6690 {
6691         struct btrfs_block_rsv *block_rsv;
6692         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
6693         int ret;
6694
6695         block_rsv = get_block_rsv(trans, root);
6696
6697         if (block_rsv->size == 0) {
6698                 ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6699                                              BTRFS_RESERVE_NO_FLUSH);
6700                 /*
6701                  * If we couldn't reserve metadata bytes try and use some from
6702                  * the global reserve.
6703                  */
6704                 if (ret && block_rsv != global_rsv) {
6705                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6706                         if (!ret)
6707                                 return global_rsv;
6708                         return ERR_PTR(ret);
6709                 } else if (ret) {
6710                         return ERR_PTR(ret);
6711                 }
6712                 return block_rsv;
6713         }
6714
6715         ret = block_rsv_use_bytes(block_rsv, blocksize);
6716         if (!ret)
6717                 return block_rsv;
6718         if (ret && !block_rsv->failfast) {
6719                 if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6720                         static DEFINE_RATELIMIT_STATE(_rs,
6721                                         DEFAULT_RATELIMIT_INTERVAL * 10,
6722                                         /*DEFAULT_RATELIMIT_BURST*/ 1);
6723                         if (__ratelimit(&_rs))
6724                                 WARN(1, KERN_DEBUG
6725                                         "btrfs: block rsv returned %d\n", ret);
6726                 }
6727                 ret = reserve_metadata_bytes(root, block_rsv, blocksize,
6728                                              BTRFS_RESERVE_NO_FLUSH);
6729                 if (!ret) {
6730                         return block_rsv;
6731                 } else if (ret && block_rsv != global_rsv) {
6732                         ret = block_rsv_use_bytes(global_rsv, blocksize);
6733                         if (!ret)
6734                                 return global_rsv;
6735                 }
6736         }
6737
6738         return ERR_PTR(-ENOSPC);
6739 }
6740
6741 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
6742                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
6743 {
6744         block_rsv_add_bytes(block_rsv, blocksize, 0);
6745         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
6746 }
6747
6748 /*
6749  * finds a free extent and does all the dirty work required for allocation
6750  * returns the key for the extent through ins, and a tree buffer for
6751  * the first block of the extent through buf.
6752  *
6753  * returns the tree buffer or NULL.
6754  */
6755 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
6756                                         struct btrfs_root *root, u32 blocksize,
6757                                         u64 parent, u64 root_objectid,
6758                                         struct btrfs_disk_key *key, int level,
6759                                         u64 hint, u64 empty_size)
6760 {
6761         struct btrfs_key ins;
6762         struct btrfs_block_rsv *block_rsv;
6763         struct extent_buffer *buf;
6764         u64 flags = 0;
6765         int ret;
6766         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6767                                                  SKINNY_METADATA);
6768
6769         block_rsv = use_block_rsv(trans, root, blocksize);
6770         if (IS_ERR(block_rsv))
6771                 return ERR_CAST(block_rsv);
6772
6773         ret = btrfs_reserve_extent(trans, root, blocksize, blocksize,
6774                                    empty_size, hint, &ins, 0);
6775         if (ret) {
6776                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
6777                 return ERR_PTR(ret);
6778         }
6779
6780         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
6781                                     blocksize, level);
6782         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
6783
6784         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
6785                 if (parent == 0)
6786                         parent = ins.objectid;
6787                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
6788         } else
6789                 BUG_ON(parent > 0);
6790
6791         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
6792                 struct btrfs_delayed_extent_op *extent_op;
6793                 extent_op = btrfs_alloc_delayed_extent_op();
6794                 BUG_ON(!extent_op); /* -ENOMEM */
6795                 if (key)
6796                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
6797                 else
6798                         memset(&extent_op->key, 0, sizeof(extent_op->key));
6799                 extent_op->flags_to_set = flags;
6800                 if (skinny_metadata)
6801                         extent_op->update_key = 0;
6802                 else
6803                         extent_op->update_key = 1;
6804                 extent_op->update_flags = 1;
6805                 extent_op->is_data = 0;
6806
6807                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6808                                         ins.objectid,
6809                                         ins.offset, parent, root_objectid,
6810                                         level, BTRFS_ADD_DELAYED_EXTENT,
6811                                         extent_op, 0);
6812                 BUG_ON(ret); /* -ENOMEM */
6813         }
6814         return buf;
6815 }
6816
6817 struct walk_control {
6818         u64 refs[BTRFS_MAX_LEVEL];
6819         u64 flags[BTRFS_MAX_LEVEL];
6820         struct btrfs_key update_progress;
6821         int stage;
6822         int level;
6823         int shared_level;
6824         int update_ref;
6825         int keep_locks;
6826         int reada_slot;
6827         int reada_count;
6828         int for_reloc;
6829 };
6830
6831 #define DROP_REFERENCE  1
6832 #define UPDATE_BACKREF  2
6833
6834 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
6835                                      struct btrfs_root *root,
6836                                      struct walk_control *wc,
6837                                      struct btrfs_path *path)
6838 {
6839         u64 bytenr;
6840         u64 generation;
6841         u64 refs;
6842         u64 flags;
6843         u32 nritems;
6844         u32 blocksize;
6845         struct btrfs_key key;
6846         struct extent_buffer *eb;
6847         int ret;
6848         int slot;
6849         int nread = 0;
6850
6851         if (path->slots[wc->level] < wc->reada_slot) {
6852                 wc->reada_count = wc->reada_count * 2 / 3;
6853                 wc->reada_count = max(wc->reada_count, 2);
6854         } else {
6855                 wc->reada_count = wc->reada_count * 3 / 2;
6856                 wc->reada_count = min_t(int, wc->reada_count,
6857                                         BTRFS_NODEPTRS_PER_BLOCK(root));
6858         }
6859
6860         eb = path->nodes[wc->level];
6861         nritems = btrfs_header_nritems(eb);
6862         blocksize = btrfs_level_size(root, wc->level - 1);
6863
6864         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
6865                 if (nread >= wc->reada_count)
6866                         break;
6867
6868                 cond_resched();
6869                 bytenr = btrfs_node_blockptr(eb, slot);
6870                 generation = btrfs_node_ptr_generation(eb, slot);
6871
6872                 if (slot == path->slots[wc->level])
6873                         goto reada;
6874
6875                 if (wc->stage == UPDATE_BACKREF &&
6876                     generation <= root->root_key.offset)
6877                         continue;
6878
6879                 /* We don't lock the tree block, it's OK to be racy here */
6880                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
6881                                                wc->level - 1, 1, &refs,
6882                                                &flags);
6883                 /* We don't care about errors in readahead. */
6884                 if (ret < 0)
6885                         continue;
6886                 BUG_ON(refs == 0);
6887
6888                 if (wc->stage == DROP_REFERENCE) {
6889                         if (refs == 1)
6890                                 goto reada;
6891
6892                         if (wc->level == 1 &&
6893                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6894                                 continue;
6895                         if (!wc->update_ref ||
6896                             generation <= root->root_key.offset)
6897                                 continue;
6898                         btrfs_node_key_to_cpu(eb, &key, slot);
6899                         ret = btrfs_comp_cpu_keys(&key,
6900                                                   &wc->update_progress);
6901                         if (ret < 0)
6902                                 continue;
6903                 } else {
6904                         if (wc->level == 1 &&
6905                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
6906                                 continue;
6907                 }
6908 reada:
6909                 ret = readahead_tree_block(root, bytenr, blocksize,
6910                                            generation);
6911                 if (ret)
6912                         break;
6913                 nread++;
6914         }
6915         wc->reada_slot = slot;
6916 }
6917
6918 /*
6919  * helper to process tree block while walking down the tree.
6920  *
6921  * when wc->stage == UPDATE_BACKREF, this function updates
6922  * back refs for pointers in the block.
6923  *
6924  * NOTE: return value 1 means we should stop walking down.
6925  */
6926 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
6927                                    struct btrfs_root *root,
6928                                    struct btrfs_path *path,
6929                                    struct walk_control *wc, int lookup_info)
6930 {
6931         int level = wc->level;
6932         struct extent_buffer *eb = path->nodes[level];
6933         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
6934         int ret;
6935
6936         if (wc->stage == UPDATE_BACKREF &&
6937             btrfs_header_owner(eb) != root->root_key.objectid)
6938                 return 1;
6939
6940         /*
6941          * when reference count of tree block is 1, it won't increase
6942          * again. once full backref flag is set, we never clear it.
6943          */
6944         if (lookup_info &&
6945             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
6946              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
6947                 BUG_ON(!path->locks[level]);
6948                 ret = btrfs_lookup_extent_info(trans, root,
6949                                                eb->start, level, 1,
6950                                                &wc->refs[level],
6951                                                &wc->flags[level]);
6952                 BUG_ON(ret == -ENOMEM);
6953                 if (ret)
6954                         return ret;
6955                 BUG_ON(wc->refs[level] == 0);
6956         }
6957
6958         if (wc->stage == DROP_REFERENCE) {
6959                 if (wc->refs[level] > 1)
6960                         return 1;
6961
6962                 if (path->locks[level] && !wc->keep_locks) {
6963                         btrfs_tree_unlock_rw(eb, path->locks[level]);
6964                         path->locks[level] = 0;
6965                 }
6966                 return 0;
6967         }
6968
6969         /* wc->stage == UPDATE_BACKREF */
6970         if (!(wc->flags[level] & flag)) {
6971                 BUG_ON(!path->locks[level]);
6972                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
6973                 BUG_ON(ret); /* -ENOMEM */
6974                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
6975                 BUG_ON(ret); /* -ENOMEM */
6976                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
6977                                                   eb->len, flag, 0);
6978                 BUG_ON(ret); /* -ENOMEM */
6979                 wc->flags[level] |= flag;
6980         }
6981
6982         /*
6983          * the block is shared by multiple trees, so it's not good to
6984          * keep the tree lock
6985          */
6986         if (path->locks[level] && level > 0) {
6987                 btrfs_tree_unlock_rw(eb, path->locks[level]);
6988                 path->locks[level] = 0;
6989         }
6990         return 0;
6991 }
6992
6993 /*
6994  * helper to process tree block pointer.
6995  *
6996  * when wc->stage == DROP_REFERENCE, this function checks
6997  * reference count of the block pointed to. if the block
6998  * is shared and we need update back refs for the subtree
6999  * rooted at the block, this function changes wc->stage to
7000  * UPDATE_BACKREF. if the block is shared and there is no
7001  * need to update back, this function drops the reference
7002  * to the block.
7003  *
7004  * NOTE: return value 1 means we should stop walking down.
7005  */
7006 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
7007                                  struct btrfs_root *root,
7008                                  struct btrfs_path *path,
7009                                  struct walk_control *wc, int *lookup_info)
7010 {
7011         u64 bytenr;
7012         u64 generation;
7013         u64 parent;
7014         u32 blocksize;
7015         struct btrfs_key key;
7016         struct extent_buffer *next;
7017         int level = wc->level;
7018         int reada = 0;
7019         int ret = 0;
7020
7021         generation = btrfs_node_ptr_generation(path->nodes[level],
7022                                                path->slots[level]);
7023         /*
7024          * if the lower level block was created before the snapshot
7025          * was created, we know there is no need to update back refs
7026          * for the subtree
7027          */
7028         if (wc->stage == UPDATE_BACKREF &&
7029             generation <= root->root_key.offset) {
7030                 *lookup_info = 1;
7031                 return 1;
7032         }
7033
7034         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
7035         blocksize = btrfs_level_size(root, level - 1);
7036
7037         next = btrfs_find_tree_block(root, bytenr, blocksize);
7038         if (!next) {
7039                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
7040                 if (!next)
7041                         return -ENOMEM;
7042                 reada = 1;
7043         }
7044         btrfs_tree_lock(next);
7045         btrfs_set_lock_blocking(next);
7046
7047         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
7048                                        &wc->refs[level - 1],
7049                                        &wc->flags[level - 1]);
7050         if (ret < 0) {
7051                 btrfs_tree_unlock(next);
7052                 return ret;
7053         }
7054
7055         if (unlikely(wc->refs[level - 1] == 0)) {
7056                 btrfs_err(root->fs_info, "Missing references.");
7057                 BUG();
7058         }
7059         *lookup_info = 0;
7060
7061         if (wc->stage == DROP_REFERENCE) {
7062                 if (wc->refs[level - 1] > 1) {
7063                         if (level == 1 &&
7064                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7065                                 goto skip;
7066
7067                         if (!wc->update_ref ||
7068                             generation <= root->root_key.offset)
7069                                 goto skip;
7070
7071                         btrfs_node_key_to_cpu(path->nodes[level], &key,
7072                                               path->slots[level]);
7073                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
7074                         if (ret < 0)
7075                                 goto skip;
7076
7077                         wc->stage = UPDATE_BACKREF;
7078                         wc->shared_level = level - 1;
7079                 }
7080         } else {
7081                 if (level == 1 &&
7082                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7083                         goto skip;
7084         }
7085
7086         if (!btrfs_buffer_uptodate(next, generation, 0)) {
7087                 btrfs_tree_unlock(next);
7088                 free_extent_buffer(next);
7089                 next = NULL;
7090                 *lookup_info = 1;
7091         }
7092
7093         if (!next) {
7094                 if (reada && level == 1)
7095                         reada_walk_down(trans, root, wc, path);
7096                 next = read_tree_block(root, bytenr, blocksize, generation);
7097                 if (!next || !extent_buffer_uptodate(next)) {
7098                         free_extent_buffer(next);
7099                         return -EIO;
7100                 }
7101                 btrfs_tree_lock(next);
7102                 btrfs_set_lock_blocking(next);
7103         }
7104
7105         level--;
7106         BUG_ON(level != btrfs_header_level(next));
7107         path->nodes[level] = next;
7108         path->slots[level] = 0;
7109         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7110         wc->level = level;
7111         if (wc->level == 1)
7112                 wc->reada_slot = 0;
7113         return 0;
7114 skip:
7115         wc->refs[level - 1] = 0;
7116         wc->flags[level - 1] = 0;
7117         if (wc->stage == DROP_REFERENCE) {
7118                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
7119                         parent = path->nodes[level]->start;
7120                 } else {
7121                         BUG_ON(root->root_key.objectid !=
7122                                btrfs_header_owner(path->nodes[level]));
7123                         parent = 0;
7124                 }
7125
7126                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
7127                                 root->root_key.objectid, level - 1, 0, 0);
7128                 BUG_ON(ret); /* -ENOMEM */
7129         }
7130         btrfs_tree_unlock(next);
7131         free_extent_buffer(next);
7132         *lookup_info = 1;
7133         return 1;
7134 }
7135
7136 /*
7137  * helper to process tree block while walking up the tree.
7138  *
7139  * when wc->stage == DROP_REFERENCE, this function drops
7140  * reference count on the block.
7141  *
7142  * when wc->stage == UPDATE_BACKREF, this function changes
7143  * wc->stage back to DROP_REFERENCE if we changed wc->stage
7144  * to UPDATE_BACKREF previously while processing the block.
7145  *
7146  * NOTE: return value 1 means we should stop walking up.
7147  */
7148 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
7149                                  struct btrfs_root *root,
7150                                  struct btrfs_path *path,
7151                                  struct walk_control *wc)
7152 {
7153         int ret;
7154         int level = wc->level;
7155         struct extent_buffer *eb = path->nodes[level];
7156         u64 parent = 0;
7157
7158         if (wc->stage == UPDATE_BACKREF) {
7159                 BUG_ON(wc->shared_level < level);
7160                 if (level < wc->shared_level)
7161                         goto out;
7162
7163                 ret = find_next_key(path, level + 1, &wc->update_progress);
7164                 if (ret > 0)
7165                         wc->update_ref = 0;
7166
7167                 wc->stage = DROP_REFERENCE;
7168                 wc->shared_level = -1;
7169                 path->slots[level] = 0;
7170
7171                 /*
7172                  * check reference count again if the block isn't locked.
7173                  * we should start walking down the tree again if reference
7174                  * count is one.
7175                  */
7176                 if (!path->locks[level]) {
7177                         BUG_ON(level == 0);
7178                         btrfs_tree_lock(eb);
7179                         btrfs_set_lock_blocking(eb);
7180                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7181
7182                         ret = btrfs_lookup_extent_info(trans, root,
7183                                                        eb->start, level, 1,
7184                                                        &wc->refs[level],
7185                                                        &wc->flags[level]);
7186                         if (ret < 0) {
7187                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7188                                 path->locks[level] = 0;
7189                                 return ret;
7190                         }
7191                         BUG_ON(wc->refs[level] == 0);
7192                         if (wc->refs[level] == 1) {
7193                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7194                                 path->locks[level] = 0;
7195                                 return 1;
7196                         }
7197                 }
7198         }
7199
7200         /* wc->stage == DROP_REFERENCE */
7201         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
7202
7203         if (wc->refs[level] == 1) {
7204                 if (level == 0) {
7205                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7206                                 ret = btrfs_dec_ref(trans, root, eb, 1,
7207                                                     wc->for_reloc);
7208                         else
7209                                 ret = btrfs_dec_ref(trans, root, eb, 0,
7210                                                     wc->for_reloc);
7211                         BUG_ON(ret); /* -ENOMEM */
7212                 }
7213                 /* make block locked assertion in clean_tree_block happy */
7214                 if (!path->locks[level] &&
7215                     btrfs_header_generation(eb) == trans->transid) {
7216                         btrfs_tree_lock(eb);
7217                         btrfs_set_lock_blocking(eb);
7218                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7219                 }
7220                 clean_tree_block(trans, root, eb);
7221         }
7222
7223         if (eb == root->node) {
7224                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7225                         parent = eb->start;
7226                 else
7227                         BUG_ON(root->root_key.objectid !=
7228                                btrfs_header_owner(eb));
7229         } else {
7230                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7231                         parent = path->nodes[level + 1]->start;
7232                 else
7233                         BUG_ON(root->root_key.objectid !=
7234                                btrfs_header_owner(path->nodes[level + 1]));
7235         }
7236
7237         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
7238 out:
7239         wc->refs[level] = 0;
7240         wc->flags[level] = 0;
7241         return 0;
7242 }
7243
7244 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
7245                                    struct btrfs_root *root,
7246                                    struct btrfs_path *path,
7247                                    struct walk_control *wc)
7248 {
7249         int level = wc->level;
7250         int lookup_info = 1;
7251         int ret;
7252
7253         while (level >= 0) {
7254                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
7255                 if (ret > 0)
7256                         break;
7257
7258                 if (level == 0)
7259                         break;
7260
7261                 if (path->slots[level] >=
7262                     btrfs_header_nritems(path->nodes[level]))
7263                         break;
7264
7265                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
7266                 if (ret > 0) {
7267                         path->slots[level]++;
7268                         continue;
7269                 } else if (ret < 0)
7270                         return ret;
7271                 level = wc->level;
7272         }
7273         return 0;
7274 }
7275
7276 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
7277                                  struct btrfs_root *root,
7278                                  struct btrfs_path *path,
7279                                  struct walk_control *wc, int max_level)
7280 {
7281         int level = wc->level;
7282         int ret;
7283
7284         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
7285         while (level < max_level && path->nodes[level]) {
7286                 wc->level = level;
7287                 if (path->slots[level] + 1 <
7288                     btrfs_header_nritems(path->nodes[level])) {
7289                         path->slots[level]++;
7290                         return 0;
7291                 } else {
7292                         ret = walk_up_proc(trans, root, path, wc);
7293                         if (ret > 0)
7294                                 return 0;
7295
7296                         if (path->locks[level]) {
7297                                 btrfs_tree_unlock_rw(path->nodes[level],
7298                                                      path->locks[level]);
7299                                 path->locks[level] = 0;
7300                         }
7301                         free_extent_buffer(path->nodes[level]);
7302                         path->nodes[level] = NULL;
7303                         level++;
7304                 }
7305         }
7306         return 1;
7307 }
7308
7309 /*
7310  * drop a subvolume tree.
7311  *
7312  * this function traverses the tree freeing any blocks that only
7313  * referenced by the tree.
7314  *
7315  * when a shared tree block is found. this function decreases its
7316  * reference count by one. if update_ref is true, this function
7317  * also make sure backrefs for the shared block and all lower level
7318  * blocks are properly updated.
7319  *
7320  * If called with for_reloc == 0, may exit early with -EAGAIN
7321  */
7322 int btrfs_drop_snapshot(struct btrfs_root *root,
7323                          struct btrfs_block_rsv *block_rsv, int update_ref,
7324                          int for_reloc)
7325 {
7326         struct btrfs_path *path;
7327         struct btrfs_trans_handle *trans;
7328         struct btrfs_root *tree_root = root->fs_info->tree_root;
7329         struct btrfs_root_item *root_item = &root->root_item;
7330         struct walk_control *wc;
7331         struct btrfs_key key;
7332         int err = 0;
7333         int ret;
7334         int level;
7335
7336         path = btrfs_alloc_path();
7337         if (!path) {
7338                 err = -ENOMEM;
7339                 goto out;
7340         }
7341
7342         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7343         if (!wc) {
7344                 btrfs_free_path(path);
7345                 err = -ENOMEM;
7346                 goto out;
7347         }
7348
7349         trans = btrfs_start_transaction(tree_root, 0);
7350         if (IS_ERR(trans)) {
7351                 err = PTR_ERR(trans);
7352                 goto out_free;
7353         }
7354
7355         if (block_rsv)
7356                 trans->block_rsv = block_rsv;
7357
7358         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
7359                 level = btrfs_header_level(root->node);
7360                 path->nodes[level] = btrfs_lock_root_node(root);
7361                 btrfs_set_lock_blocking(path->nodes[level]);
7362                 path->slots[level] = 0;
7363                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7364                 memset(&wc->update_progress, 0,
7365                        sizeof(wc->update_progress));
7366         } else {
7367                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
7368                 memcpy(&wc->update_progress, &key,
7369                        sizeof(wc->update_progress));
7370
7371                 level = root_item->drop_level;
7372                 BUG_ON(level == 0);
7373                 path->lowest_level = level;
7374                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
7375                 path->lowest_level = 0;
7376                 if (ret < 0) {
7377                         err = ret;
7378                         goto out_end_trans;
7379                 }
7380                 WARN_ON(ret > 0);
7381
7382                 /*
7383                  * unlock our path, this is safe because only this
7384                  * function is allowed to delete this snapshot
7385                  */
7386                 btrfs_unlock_up_safe(path, 0);
7387
7388                 level = btrfs_header_level(root->node);
7389                 while (1) {
7390                         btrfs_tree_lock(path->nodes[level]);
7391                         btrfs_set_lock_blocking(path->nodes[level]);
7392
7393                         ret = btrfs_lookup_extent_info(trans, root,
7394                                                 path->nodes[level]->start,
7395                                                 level, 1, &wc->refs[level],
7396                                                 &wc->flags[level]);
7397                         if (ret < 0) {
7398                                 err = ret;
7399                                 goto out_end_trans;
7400                         }
7401                         BUG_ON(wc->refs[level] == 0);
7402
7403                         if (level == root_item->drop_level)
7404                                 break;
7405
7406                         btrfs_tree_unlock(path->nodes[level]);
7407                         WARN_ON(wc->refs[level] != 1);
7408                         level--;
7409                 }
7410         }
7411
7412         wc->level = level;
7413         wc->shared_level = -1;
7414         wc->stage = DROP_REFERENCE;
7415         wc->update_ref = update_ref;
7416         wc->keep_locks = 0;
7417         wc->for_reloc = for_reloc;
7418         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7419
7420         while (1) {
7421                 if (!for_reloc && btrfs_fs_closing(root->fs_info)) {
7422                         pr_debug("btrfs: drop snapshot early exit\n");
7423                         err = -EAGAIN;
7424                         goto out_end_trans;
7425                 }
7426
7427                 ret = walk_down_tree(trans, root, path, wc);
7428                 if (ret < 0) {
7429                         err = ret;
7430                         break;
7431                 }
7432
7433                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7434                 if (ret < 0) {
7435                         err = ret;
7436                         break;
7437                 }
7438
7439                 if (ret > 0) {
7440                         BUG_ON(wc->stage != DROP_REFERENCE);
7441                         break;
7442                 }
7443
7444                 if (wc->stage == DROP_REFERENCE) {
7445                         level = wc->level;
7446                         btrfs_node_key(path->nodes[level],
7447                                        &root_item->drop_progress,
7448                                        path->slots[level]);
7449                         root_item->drop_level = level;
7450                 }
7451
7452                 BUG_ON(wc->level == 0);
7453                 if (btrfs_should_end_transaction(trans, tree_root)) {
7454                         ret = btrfs_update_root(trans, tree_root,
7455                                                 &root->root_key,
7456                                                 root_item);
7457                         if (ret) {
7458                                 btrfs_abort_transaction(trans, tree_root, ret);
7459                                 err = ret;
7460                                 goto out_end_trans;
7461                         }
7462
7463                         btrfs_end_transaction_throttle(trans, tree_root);
7464                         trans = btrfs_start_transaction(tree_root, 0);
7465                         if (IS_ERR(trans)) {
7466                                 err = PTR_ERR(trans);
7467                                 goto out_free;
7468                         }
7469                         if (block_rsv)
7470                                 trans->block_rsv = block_rsv;
7471                 }
7472         }
7473         btrfs_release_path(path);
7474         if (err)
7475                 goto out_end_trans;
7476
7477         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7478         if (ret) {
7479                 btrfs_abort_transaction(trans, tree_root, ret);
7480                 goto out_end_trans;
7481         }
7482
7483         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7484                 ret = btrfs_find_last_root(tree_root, root->root_key.objectid,
7485                                            NULL, NULL);
7486                 if (ret < 0) {
7487                         btrfs_abort_transaction(trans, tree_root, ret);
7488                         err = ret;
7489                         goto out_end_trans;
7490                 } else if (ret > 0) {
7491                         /* if we fail to delete the orphan item this time
7492                          * around, it'll get picked up the next time.
7493                          *
7494                          * The most common failure here is just -ENOENT.
7495                          */
7496                         btrfs_del_orphan_item(trans, tree_root,
7497                                               root->root_key.objectid);
7498                 }
7499         }
7500
7501         if (root->in_radix) {
7502                 btrfs_free_fs_root(tree_root->fs_info, root);
7503         } else {
7504                 free_extent_buffer(root->node);
7505                 free_extent_buffer(root->commit_root);
7506                 kfree(root);
7507         }
7508 out_end_trans:
7509         btrfs_end_transaction_throttle(trans, tree_root);
7510 out_free:
7511         kfree(wc);
7512         btrfs_free_path(path);
7513 out:
7514         if (err)
7515                 btrfs_std_error(root->fs_info, err);
7516         return err;
7517 }
7518
7519 /*
7520  * drop subtree rooted at tree block 'node'.
7521  *
7522  * NOTE: this function will unlock and release tree block 'node'
7523  * only used by relocation code
7524  */
7525 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
7526                         struct btrfs_root *root,
7527                         struct extent_buffer *node,
7528                         struct extent_buffer *parent)
7529 {
7530         struct btrfs_path *path;
7531         struct walk_control *wc;
7532         int level;
7533         int parent_level;
7534         int ret = 0;
7535         int wret;
7536
7537         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
7538
7539         path = btrfs_alloc_path();
7540         if (!path)
7541                 return -ENOMEM;
7542
7543         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7544         if (!wc) {
7545                 btrfs_free_path(path);
7546                 return -ENOMEM;
7547         }
7548
7549         btrfs_assert_tree_locked(parent);
7550         parent_level = btrfs_header_level(parent);
7551         extent_buffer_get(parent);
7552         path->nodes[parent_level] = parent;
7553         path->slots[parent_level] = btrfs_header_nritems(parent);
7554
7555         btrfs_assert_tree_locked(node);
7556         level = btrfs_header_level(node);
7557         path->nodes[level] = node;
7558         path->slots[level] = 0;
7559         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7560
7561         wc->refs[parent_level] = 1;
7562         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7563         wc->level = level;
7564         wc->shared_level = -1;
7565         wc->stage = DROP_REFERENCE;
7566         wc->update_ref = 0;
7567         wc->keep_locks = 1;
7568         wc->for_reloc = 1;
7569         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7570
7571         while (1) {
7572                 wret = walk_down_tree(trans, root, path, wc);
7573                 if (wret < 0) {
7574                         ret = wret;
7575                         break;
7576                 }
7577
7578                 wret = walk_up_tree(trans, root, path, wc, parent_level);
7579                 if (wret < 0)
7580                         ret = wret;
7581                 if (wret != 0)
7582                         break;
7583         }
7584
7585         kfree(wc);
7586         btrfs_free_path(path);
7587         return ret;
7588 }
7589
7590 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
7591 {
7592         u64 num_devices;
7593         u64 stripped;
7594
7595         /*
7596          * if restripe for this chunk_type is on pick target profile and
7597          * return, otherwise do the usual balance
7598          */
7599         stripped = get_restripe_target(root->fs_info, flags);
7600         if (stripped)
7601                 return extended_to_chunk(stripped);
7602
7603         /*
7604          * we add in the count of missing devices because we want
7605          * to make sure that any RAID levels on a degraded FS
7606          * continue to be honored.
7607          */
7608         num_devices = root->fs_info->fs_devices->rw_devices +
7609                 root->fs_info->fs_devices->missing_devices;
7610
7611         stripped = BTRFS_BLOCK_GROUP_RAID0 |
7612                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
7613                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
7614
7615         if (num_devices == 1) {
7616                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7617                 stripped = flags & ~stripped;
7618
7619                 /* turn raid0 into single device chunks */
7620                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
7621                         return stripped;
7622
7623                 /* turn mirroring into duplication */
7624                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
7625                              BTRFS_BLOCK_GROUP_RAID10))
7626                         return stripped | BTRFS_BLOCK_GROUP_DUP;
7627         } else {
7628                 /* they already had raid on here, just return */
7629                 if (flags & stripped)
7630                         return flags;
7631
7632                 stripped |= BTRFS_BLOCK_GROUP_DUP;
7633                 stripped = flags & ~stripped;
7634
7635                 /* switch duplicated blocks with raid1 */
7636                 if (flags & BTRFS_BLOCK_GROUP_DUP)
7637                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
7638
7639                 /* this is drive concat, leave it alone */
7640         }
7641
7642         return flags;
7643 }
7644
7645 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
7646 {
7647         struct btrfs_space_info *sinfo = cache->space_info;
7648         u64 num_bytes;
7649         u64 min_allocable_bytes;
7650         int ret = -ENOSPC;
7651
7652
7653         /*
7654          * We need some metadata space and system metadata space for
7655          * allocating chunks in some corner cases until we force to set
7656          * it to be readonly.
7657          */
7658         if ((sinfo->flags &
7659              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
7660             !force)
7661                 min_allocable_bytes = 1 * 1024 * 1024;
7662         else
7663                 min_allocable_bytes = 0;
7664
7665         spin_lock(&sinfo->lock);
7666         spin_lock(&cache->lock);
7667
7668         if (cache->ro) {
7669                 ret = 0;
7670                 goto out;
7671         }
7672
7673         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7674                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7675
7676         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
7677             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
7678             min_allocable_bytes <= sinfo->total_bytes) {
7679                 sinfo->bytes_readonly += num_bytes;
7680                 cache->ro = 1;
7681                 ret = 0;
7682         }
7683 out:
7684         spin_unlock(&cache->lock);
7685         spin_unlock(&sinfo->lock);
7686         return ret;
7687 }
7688
7689 int btrfs_set_block_group_ro(struct btrfs_root *root,
7690                              struct btrfs_block_group_cache *cache)
7691
7692 {
7693         struct btrfs_trans_handle *trans;
7694         u64 alloc_flags;
7695         int ret;
7696
7697         BUG_ON(cache->ro);
7698
7699         trans = btrfs_join_transaction(root);
7700         if (IS_ERR(trans))
7701                 return PTR_ERR(trans);
7702
7703         alloc_flags = update_block_group_flags(root, cache->flags);
7704         if (alloc_flags != cache->flags) {
7705                 ret = do_chunk_alloc(trans, root, alloc_flags,
7706                                      CHUNK_ALLOC_FORCE);
7707                 if (ret < 0)
7708                         goto out;
7709         }
7710
7711         ret = set_block_group_ro(cache, 0);
7712         if (!ret)
7713                 goto out;
7714         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
7715         ret = do_chunk_alloc(trans, root, alloc_flags,
7716                              CHUNK_ALLOC_FORCE);
7717         if (ret < 0)
7718                 goto out;
7719         ret = set_block_group_ro(cache, 0);
7720 out:
7721         btrfs_end_transaction(trans, root);
7722         return ret;
7723 }
7724
7725 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
7726                             struct btrfs_root *root, u64 type)
7727 {
7728         u64 alloc_flags = get_alloc_profile(root, type);
7729         return do_chunk_alloc(trans, root, alloc_flags,
7730                               CHUNK_ALLOC_FORCE);
7731 }
7732
7733 /*
7734  * helper to account the unused space of all the readonly block group in the
7735  * list. takes mirrors into account.
7736  */
7737 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
7738 {
7739         struct btrfs_block_group_cache *block_group;
7740         u64 free_bytes = 0;
7741         int factor;
7742
7743         list_for_each_entry(block_group, groups_list, list) {
7744                 spin_lock(&block_group->lock);
7745
7746                 if (!block_group->ro) {
7747                         spin_unlock(&block_group->lock);
7748                         continue;
7749                 }
7750
7751                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
7752                                           BTRFS_BLOCK_GROUP_RAID10 |
7753                                           BTRFS_BLOCK_GROUP_DUP))
7754                         factor = 2;
7755                 else
7756                         factor = 1;
7757
7758                 free_bytes += (block_group->key.offset -
7759                                btrfs_block_group_used(&block_group->item)) *
7760                                factor;
7761
7762                 spin_unlock(&block_group->lock);
7763         }
7764
7765         return free_bytes;
7766 }
7767
7768 /*
7769  * helper to account the unused space of all the readonly block group in the
7770  * space_info. takes mirrors into account.
7771  */
7772 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
7773 {
7774         int i;
7775         u64 free_bytes = 0;
7776
7777         spin_lock(&sinfo->lock);
7778
7779         for(i = 0; i < BTRFS_NR_RAID_TYPES; i++)
7780                 if (!list_empty(&sinfo->block_groups[i]))
7781                         free_bytes += __btrfs_get_ro_block_group_free_space(
7782                                                 &sinfo->block_groups[i]);
7783
7784         spin_unlock(&sinfo->lock);
7785
7786         return free_bytes;
7787 }
7788
7789 void btrfs_set_block_group_rw(struct btrfs_root *root,
7790                               struct btrfs_block_group_cache *cache)
7791 {
7792         struct btrfs_space_info *sinfo = cache->space_info;
7793         u64 num_bytes;
7794
7795         BUG_ON(!cache->ro);
7796
7797         spin_lock(&sinfo->lock);
7798         spin_lock(&cache->lock);
7799         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
7800                     cache->bytes_super - btrfs_block_group_used(&cache->item);
7801         sinfo->bytes_readonly -= num_bytes;
7802         cache->ro = 0;
7803         spin_unlock(&cache->lock);
7804         spin_unlock(&sinfo->lock);
7805 }
7806
7807 /*
7808  * checks to see if its even possible to relocate this block group.
7809  *
7810  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
7811  * ok to go ahead and try.
7812  */
7813 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
7814 {
7815         struct btrfs_block_group_cache *block_group;
7816         struct btrfs_space_info *space_info;
7817         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
7818         struct btrfs_device *device;
7819         u64 min_free;
7820         u64 dev_min = 1;
7821         u64 dev_nr = 0;
7822         u64 target;
7823         int index;
7824         int full = 0;
7825         int ret = 0;
7826
7827         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
7828
7829         /* odd, couldn't find the block group, leave it alone */
7830         if (!block_group)
7831                 return -1;
7832
7833         min_free = btrfs_block_group_used(&block_group->item);
7834
7835         /* no bytes used, we're good */
7836         if (!min_free)
7837                 goto out;
7838
7839         space_info = block_group->space_info;
7840         spin_lock(&space_info->lock);
7841
7842         full = space_info->full;
7843
7844         /*
7845          * if this is the last block group we have in this space, we can't
7846          * relocate it unless we're able to allocate a new chunk below.
7847          *
7848          * Otherwise, we need to make sure we have room in the space to handle
7849          * all of the extents from this block group.  If we can, we're good
7850          */
7851         if ((space_info->total_bytes != block_group->key.offset) &&
7852             (space_info->bytes_used + space_info->bytes_reserved +
7853              space_info->bytes_pinned + space_info->bytes_readonly +
7854              min_free < space_info->total_bytes)) {
7855                 spin_unlock(&space_info->lock);
7856                 goto out;
7857         }
7858         spin_unlock(&space_info->lock);
7859
7860         /*
7861          * ok we don't have enough space, but maybe we have free space on our
7862          * devices to allocate new chunks for relocation, so loop through our
7863          * alloc devices and guess if we have enough space.  if this block
7864          * group is going to be restriped, run checks against the target
7865          * profile instead of the current one.
7866          */
7867         ret = -1;
7868
7869         /*
7870          * index:
7871          *      0: raid10
7872          *      1: raid1
7873          *      2: dup
7874          *      3: raid0
7875          *      4: single
7876          */
7877         target = get_restripe_target(root->fs_info, block_group->flags);
7878         if (target) {
7879                 index = __get_raid_index(extended_to_chunk(target));
7880         } else {
7881                 /*
7882                  * this is just a balance, so if we were marked as full
7883                  * we know there is no space for a new chunk
7884                  */
7885                 if (full)
7886                         goto out;
7887
7888                 index = get_block_group_index(block_group);
7889         }
7890
7891         if (index == BTRFS_RAID_RAID10) {
7892                 dev_min = 4;
7893                 /* Divide by 2 */
7894                 min_free >>= 1;
7895         } else if (index == BTRFS_RAID_RAID1) {
7896                 dev_min = 2;
7897         } else if (index == BTRFS_RAID_DUP) {
7898                 /* Multiply by 2 */
7899                 min_free <<= 1;
7900         } else if (index == BTRFS_RAID_RAID0) {
7901                 dev_min = fs_devices->rw_devices;
7902                 do_div(min_free, dev_min);
7903         }
7904
7905         mutex_lock(&root->fs_info->chunk_mutex);
7906         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
7907                 u64 dev_offset;
7908
7909                 /*
7910                  * check to make sure we can actually find a chunk with enough
7911                  * space to fit our block group in.
7912                  */
7913                 if (device->total_bytes > device->bytes_used + min_free &&
7914                     !device->is_tgtdev_for_dev_replace) {
7915                         ret = find_free_dev_extent(device, min_free,
7916                                                    &dev_offset, NULL);
7917                         if (!ret)
7918                                 dev_nr++;
7919
7920                         if (dev_nr >= dev_min)
7921                                 break;
7922
7923                         ret = -1;
7924                 }
7925         }
7926         mutex_unlock(&root->fs_info->chunk_mutex);
7927 out:
7928         btrfs_put_block_group(block_group);
7929         return ret;
7930 }
7931
7932 static int find_first_block_group(struct btrfs_root *root,
7933                 struct btrfs_path *path, struct btrfs_key *key)
7934 {
7935         int ret = 0;
7936         struct btrfs_key found_key;
7937         struct extent_buffer *leaf;
7938         int slot;
7939
7940         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
7941         if (ret < 0)
7942                 goto out;
7943
7944         while (1) {
7945                 slot = path->slots[0];
7946                 leaf = path->nodes[0];
7947                 if (slot >= btrfs_header_nritems(leaf)) {
7948                         ret = btrfs_next_leaf(root, path);
7949                         if (ret == 0)
7950                                 continue;
7951                         if (ret < 0)
7952                                 goto out;
7953                         break;
7954                 }
7955                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
7956
7957                 if (found_key.objectid >= key->objectid &&
7958                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
7959                         ret = 0;
7960                         goto out;
7961                 }
7962                 path->slots[0]++;
7963         }
7964 out:
7965         return ret;
7966 }
7967
7968 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
7969 {
7970         struct btrfs_block_group_cache *block_group;
7971         u64 last = 0;
7972
7973         while (1) {
7974                 struct inode *inode;
7975
7976                 block_group = btrfs_lookup_first_block_group(info, last);
7977                 while (block_group) {
7978                         spin_lock(&block_group->lock);
7979                         if (block_group->iref)
7980                                 break;
7981                         spin_unlock(&block_group->lock);
7982                         block_group = next_block_group(info->tree_root,
7983                                                        block_group);
7984                 }
7985                 if (!block_group) {
7986                         if (last == 0)
7987                                 break;
7988                         last = 0;
7989                         continue;
7990                 }
7991
7992                 inode = block_group->inode;
7993                 block_group->iref = 0;
7994                 block_group->inode = NULL;
7995                 spin_unlock(&block_group->lock);
7996                 iput(inode);
7997                 last = block_group->key.objectid + block_group->key.offset;
7998                 btrfs_put_block_group(block_group);
7999         }
8000 }
8001
8002 int btrfs_free_block_groups(struct btrfs_fs_info *info)
8003 {
8004         struct btrfs_block_group_cache *block_group;
8005         struct btrfs_space_info *space_info;
8006         struct btrfs_caching_control *caching_ctl;
8007         struct rb_node *n;
8008
8009         down_write(&info->extent_commit_sem);
8010         while (!list_empty(&info->caching_block_groups)) {
8011                 caching_ctl = list_entry(info->caching_block_groups.next,
8012                                          struct btrfs_caching_control, list);
8013                 list_del(&caching_ctl->list);
8014                 put_caching_control(caching_ctl);
8015         }
8016         up_write(&info->extent_commit_sem);
8017
8018         spin_lock(&info->block_group_cache_lock);
8019         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
8020                 block_group = rb_entry(n, struct btrfs_block_group_cache,
8021                                        cache_node);
8022                 rb_erase(&block_group->cache_node,
8023                          &info->block_group_cache_tree);
8024                 spin_unlock(&info->block_group_cache_lock);
8025
8026                 down_write(&block_group->space_info->groups_sem);
8027                 list_del(&block_group->list);
8028                 up_write(&block_group->space_info->groups_sem);
8029
8030                 if (block_group->cached == BTRFS_CACHE_STARTED)
8031                         wait_block_group_cache_done(block_group);
8032
8033                 /*
8034                  * We haven't cached this block group, which means we could
8035                  * possibly have excluded extents on this block group.
8036                  */
8037                 if (block_group->cached == BTRFS_CACHE_NO)
8038                         free_excluded_extents(info->extent_root, block_group);
8039
8040                 btrfs_remove_free_space_cache(block_group);
8041                 btrfs_put_block_group(block_group);
8042
8043                 spin_lock(&info->block_group_cache_lock);
8044         }
8045         spin_unlock(&info->block_group_cache_lock);
8046
8047         /* now that all the block groups are freed, go through and
8048          * free all the space_info structs.  This is only called during
8049          * the final stages of unmount, and so we know nobody is
8050          * using them.  We call synchronize_rcu() once before we start,
8051          * just to be on the safe side.
8052          */
8053         synchronize_rcu();
8054
8055         release_global_block_rsv(info);
8056
8057         while(!list_empty(&info->space_info)) {
8058                 space_info = list_entry(info->space_info.next,
8059                                         struct btrfs_space_info,
8060                                         list);
8061                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
8062                         if (space_info->bytes_pinned > 0 ||
8063                             space_info->bytes_reserved > 0 ||
8064                             space_info->bytes_may_use > 0) {
8065                                 WARN_ON(1);
8066                                 dump_space_info(space_info, 0, 0);
8067                         }
8068                 }
8069                 list_del(&space_info->list);
8070                 kfree(space_info);
8071         }
8072         return 0;
8073 }
8074
8075 static void __link_block_group(struct btrfs_space_info *space_info,
8076                                struct btrfs_block_group_cache *cache)
8077 {
8078         int index = get_block_group_index(cache);
8079
8080         down_write(&space_info->groups_sem);
8081         list_add_tail(&cache->list, &space_info->block_groups[index]);
8082         up_write(&space_info->groups_sem);
8083 }
8084
8085 int btrfs_read_block_groups(struct btrfs_root *root)
8086 {
8087         struct btrfs_path *path;
8088         int ret;
8089         struct btrfs_block_group_cache *cache;
8090         struct btrfs_fs_info *info = root->fs_info;
8091         struct btrfs_space_info *space_info;
8092         struct btrfs_key key;
8093         struct btrfs_key found_key;
8094         struct extent_buffer *leaf;
8095         int need_clear = 0;
8096         u64 cache_gen;
8097
8098         root = info->extent_root;
8099         key.objectid = 0;
8100         key.offset = 0;
8101         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
8102         path = btrfs_alloc_path();
8103         if (!path)
8104                 return -ENOMEM;
8105         path->reada = 1;
8106
8107         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
8108         if (btrfs_test_opt(root, SPACE_CACHE) &&
8109             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
8110                 need_clear = 1;
8111         if (btrfs_test_opt(root, CLEAR_CACHE))
8112                 need_clear = 1;
8113
8114         while (1) {
8115                 ret = find_first_block_group(root, path, &key);
8116                 if (ret > 0)
8117                         break;
8118                 if (ret != 0)
8119                         goto error;
8120                 leaf = path->nodes[0];
8121                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
8122                 cache = kzalloc(sizeof(*cache), GFP_NOFS);
8123                 if (!cache) {
8124                         ret = -ENOMEM;
8125                         goto error;
8126                 }
8127                 cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8128                                                 GFP_NOFS);
8129                 if (!cache->free_space_ctl) {
8130                         kfree(cache);
8131                         ret = -ENOMEM;
8132                         goto error;
8133                 }
8134
8135                 atomic_set(&cache->count, 1);
8136                 spin_lock_init(&cache->lock);
8137                 cache->fs_info = info;
8138                 INIT_LIST_HEAD(&cache->list);
8139                 INIT_LIST_HEAD(&cache->cluster_list);
8140
8141                 if (need_clear) {
8142                         /*
8143                          * When we mount with old space cache, we need to
8144                          * set BTRFS_DC_CLEAR and set dirty flag.
8145                          *
8146                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
8147                          *    truncate the old free space cache inode and
8148                          *    setup a new one.
8149                          * b) Setting 'dirty flag' makes sure that we flush
8150                          *    the new space cache info onto disk.
8151                          */
8152                         cache->disk_cache_state = BTRFS_DC_CLEAR;
8153                         if (btrfs_test_opt(root, SPACE_CACHE))
8154                                 cache->dirty = 1;
8155                 }
8156
8157                 read_extent_buffer(leaf, &cache->item,
8158                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
8159                                    sizeof(cache->item));
8160                 memcpy(&cache->key, &found_key, sizeof(found_key));
8161
8162                 key.objectid = found_key.objectid + found_key.offset;
8163                 btrfs_release_path(path);
8164                 cache->flags = btrfs_block_group_flags(&cache->item);
8165                 cache->sectorsize = root->sectorsize;
8166                 cache->full_stripe_len = btrfs_full_stripe_len(root,
8167                                                &root->fs_info->mapping_tree,
8168                                                found_key.objectid);
8169                 btrfs_init_free_space_ctl(cache);
8170
8171                 /*
8172                  * We need to exclude the super stripes now so that the space
8173                  * info has super bytes accounted for, otherwise we'll think
8174                  * we have more space than we actually do.
8175                  */
8176                 ret = exclude_super_stripes(root, cache);
8177                 if (ret) {
8178                         /*
8179                          * We may have excluded something, so call this just in
8180                          * case.
8181                          */
8182                         free_excluded_extents(root, cache);
8183                         kfree(cache->free_space_ctl);
8184                         kfree(cache);
8185                         goto error;
8186                 }
8187
8188                 /*
8189                  * check for two cases, either we are full, and therefore
8190                  * don't need to bother with the caching work since we won't
8191                  * find any space, or we are empty, and we can just add all
8192                  * the space in and be done with it.  This saves us _alot_ of
8193                  * time, particularly in the full case.
8194                  */
8195                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
8196                         cache->last_byte_to_unpin = (u64)-1;
8197                         cache->cached = BTRFS_CACHE_FINISHED;
8198                         free_excluded_extents(root, cache);
8199                 } else if (btrfs_block_group_used(&cache->item) == 0) {
8200                         cache->last_byte_to_unpin = (u64)-1;
8201                         cache->cached = BTRFS_CACHE_FINISHED;
8202                         add_new_free_space(cache, root->fs_info,
8203                                            found_key.objectid,
8204                                            found_key.objectid +
8205                                            found_key.offset);
8206                         free_excluded_extents(root, cache);
8207                 }
8208
8209                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
8210                 if (ret) {
8211                         btrfs_remove_free_space_cache(cache);
8212                         btrfs_put_block_group(cache);
8213                         goto error;
8214                 }
8215
8216                 ret = update_space_info(info, cache->flags, found_key.offset,
8217                                         btrfs_block_group_used(&cache->item),
8218                                         &space_info);
8219                 if (ret) {
8220                         btrfs_remove_free_space_cache(cache);
8221                         spin_lock(&info->block_group_cache_lock);
8222                         rb_erase(&cache->cache_node,
8223                                  &info->block_group_cache_tree);
8224                         spin_unlock(&info->block_group_cache_lock);
8225                         btrfs_put_block_group(cache);
8226                         goto error;
8227                 }
8228
8229                 cache->space_info = space_info;
8230                 spin_lock(&cache->space_info->lock);
8231                 cache->space_info->bytes_readonly += cache->bytes_super;
8232                 spin_unlock(&cache->space_info->lock);
8233
8234                 __link_block_group(space_info, cache);
8235
8236                 set_avail_alloc_bits(root->fs_info, cache->flags);
8237                 if (btrfs_chunk_readonly(root, cache->key.objectid))
8238                         set_block_group_ro(cache, 1);
8239         }
8240
8241         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
8242                 if (!(get_alloc_profile(root, space_info->flags) &
8243                       (BTRFS_BLOCK_GROUP_RAID10 |
8244                        BTRFS_BLOCK_GROUP_RAID1 |
8245                        BTRFS_BLOCK_GROUP_RAID5 |
8246                        BTRFS_BLOCK_GROUP_RAID6 |
8247                        BTRFS_BLOCK_GROUP_DUP)))
8248                         continue;
8249                 /*
8250                  * avoid allocating from un-mirrored block group if there are
8251                  * mirrored block groups.
8252                  */
8253                 list_for_each_entry(cache, &space_info->block_groups[3], list)
8254                         set_block_group_ro(cache, 1);
8255                 list_for_each_entry(cache, &space_info->block_groups[4], list)
8256                         set_block_group_ro(cache, 1);
8257         }
8258
8259         init_global_block_rsv(info);
8260         ret = 0;
8261 error:
8262         btrfs_free_path(path);
8263         return ret;
8264 }
8265
8266 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
8267                                        struct btrfs_root *root)
8268 {
8269         struct btrfs_block_group_cache *block_group, *tmp;
8270         struct btrfs_root *extent_root = root->fs_info->extent_root;
8271         struct btrfs_block_group_item item;
8272         struct btrfs_key key;
8273         int ret = 0;
8274
8275         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs,
8276                                  new_bg_list) {
8277                 list_del_init(&block_group->new_bg_list);
8278
8279                 if (ret)
8280                         continue;
8281
8282                 spin_lock(&block_group->lock);
8283                 memcpy(&item, &block_group->item, sizeof(item));
8284                 memcpy(&key, &block_group->key, sizeof(key));
8285                 spin_unlock(&block_group->lock);
8286
8287                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
8288                                         sizeof(item));
8289                 if (ret)
8290                         btrfs_abort_transaction(trans, extent_root, ret);
8291         }
8292 }
8293
8294 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
8295                            struct btrfs_root *root, u64 bytes_used,
8296                            u64 type, u64 chunk_objectid, u64 chunk_offset,
8297                            u64 size)
8298 {
8299         int ret;
8300         struct btrfs_root *extent_root;
8301         struct btrfs_block_group_cache *cache;
8302
8303         extent_root = root->fs_info->extent_root;
8304
8305         root->fs_info->last_trans_log_full_commit = trans->transid;
8306
8307         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8308         if (!cache)
8309                 return -ENOMEM;
8310         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8311                                         GFP_NOFS);
8312         if (!cache->free_space_ctl) {
8313                 kfree(cache);
8314                 return -ENOMEM;
8315         }
8316
8317         cache->key.objectid = chunk_offset;
8318         cache->key.offset = size;
8319         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8320         cache->sectorsize = root->sectorsize;
8321         cache->fs_info = root->fs_info;
8322         cache->full_stripe_len = btrfs_full_stripe_len(root,
8323                                                &root->fs_info->mapping_tree,
8324                                                chunk_offset);
8325
8326         atomic_set(&cache->count, 1);
8327         spin_lock_init(&cache->lock);
8328         INIT_LIST_HEAD(&cache->list);
8329         INIT_LIST_HEAD(&cache->cluster_list);
8330         INIT_LIST_HEAD(&cache->new_bg_list);
8331
8332         btrfs_init_free_space_ctl(cache);
8333
8334         btrfs_set_block_group_used(&cache->item, bytes_used);
8335         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
8336         cache->flags = type;
8337         btrfs_set_block_group_flags(&cache->item, type);
8338
8339         cache->last_byte_to_unpin = (u64)-1;
8340         cache->cached = BTRFS_CACHE_FINISHED;
8341         ret = exclude_super_stripes(root, cache);
8342         if (ret) {
8343                 /*
8344                  * We may have excluded something, so call this just in
8345                  * case.
8346                  */
8347                 free_excluded_extents(root, cache);
8348                 kfree(cache->free_space_ctl);
8349                 kfree(cache);
8350                 return ret;
8351         }
8352
8353         add_new_free_space(cache, root->fs_info, chunk_offset,
8354                            chunk_offset + size);
8355
8356         free_excluded_extents(root, cache);
8357
8358         ret = btrfs_add_block_group_cache(root->fs_info, cache);
8359         if (ret) {
8360                 btrfs_remove_free_space_cache(cache);
8361                 btrfs_put_block_group(cache);
8362                 return ret;
8363         }
8364
8365         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
8366                                 &cache->space_info);
8367         if (ret) {
8368                 btrfs_remove_free_space_cache(cache);
8369                 spin_lock(&root->fs_info->block_group_cache_lock);
8370                 rb_erase(&cache->cache_node,
8371                          &root->fs_info->block_group_cache_tree);
8372                 spin_unlock(&root->fs_info->block_group_cache_lock);
8373                 btrfs_put_block_group(cache);
8374                 return ret;
8375         }
8376         update_global_block_rsv(root->fs_info);
8377
8378         spin_lock(&cache->space_info->lock);
8379         cache->space_info->bytes_readonly += cache->bytes_super;
8380         spin_unlock(&cache->space_info->lock);
8381
8382         __link_block_group(cache->space_info, cache);
8383
8384         list_add_tail(&cache->new_bg_list, &trans->new_bgs);
8385
8386         set_avail_alloc_bits(extent_root->fs_info, type);
8387
8388         return 0;
8389 }
8390
8391 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
8392 {
8393         u64 extra_flags = chunk_to_extended(flags) &
8394                                 BTRFS_EXTENDED_PROFILE_MASK;
8395
8396         write_seqlock(&fs_info->profiles_lock);
8397         if (flags & BTRFS_BLOCK_GROUP_DATA)
8398                 fs_info->avail_data_alloc_bits &= ~extra_flags;
8399         if (flags & BTRFS_BLOCK_GROUP_METADATA)
8400                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
8401         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
8402                 fs_info->avail_system_alloc_bits &= ~extra_flags;
8403         write_sequnlock(&fs_info->profiles_lock);
8404 }
8405
8406 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
8407                              struct btrfs_root *root, u64 group_start)
8408 {
8409         struct btrfs_path *path;
8410         struct btrfs_block_group_cache *block_group;
8411         struct btrfs_free_cluster *cluster;
8412         struct btrfs_root *tree_root = root->fs_info->tree_root;
8413         struct btrfs_key key;
8414         struct inode *inode;
8415         int ret;
8416         int index;
8417         int factor;
8418
8419         root = root->fs_info->extent_root;
8420
8421         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
8422         BUG_ON(!block_group);
8423         BUG_ON(!block_group->ro);
8424
8425         /*
8426          * Free the reserved super bytes from this block group before
8427          * remove it.
8428          */
8429         free_excluded_extents(root, block_group);
8430
8431         memcpy(&key, &block_group->key, sizeof(key));
8432         index = get_block_group_index(block_group);
8433         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
8434                                   BTRFS_BLOCK_GROUP_RAID1 |
8435                                   BTRFS_BLOCK_GROUP_RAID10))
8436                 factor = 2;
8437         else
8438                 factor = 1;
8439
8440         /* make sure this block group isn't part of an allocation cluster */
8441         cluster = &root->fs_info->data_alloc_cluster;
8442         spin_lock(&cluster->refill_lock);
8443         btrfs_return_cluster_to_free_space(block_group, cluster);
8444         spin_unlock(&cluster->refill_lock);
8445
8446         /*
8447          * make sure this block group isn't part of a metadata
8448          * allocation cluster
8449          */
8450         cluster = &root->fs_info->meta_alloc_cluster;
8451         spin_lock(&cluster->refill_lock);
8452         btrfs_return_cluster_to_free_space(block_group, cluster);
8453         spin_unlock(&cluster->refill_lock);
8454
8455         path = btrfs_alloc_path();
8456         if (!path) {
8457                 ret = -ENOMEM;
8458                 goto out;
8459         }
8460
8461         inode = lookup_free_space_inode(tree_root, block_group, path);
8462         if (!IS_ERR(inode)) {
8463                 ret = btrfs_orphan_add(trans, inode);
8464                 if (ret) {
8465                         btrfs_add_delayed_iput(inode);
8466                         goto out;
8467                 }
8468                 clear_nlink(inode);
8469                 /* One for the block groups ref */
8470                 spin_lock(&block_group->lock);
8471                 if (block_group->iref) {
8472                         block_group->iref = 0;
8473                         block_group->inode = NULL;
8474                         spin_unlock(&block_group->lock);
8475                         iput(inode);
8476                 } else {
8477                         spin_unlock(&block_group->lock);
8478                 }
8479                 /* One for our lookup ref */
8480                 btrfs_add_delayed_iput(inode);
8481         }
8482
8483         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
8484         key.offset = block_group->key.objectid;
8485         key.type = 0;
8486
8487         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
8488         if (ret < 0)
8489                 goto out;
8490         if (ret > 0)
8491                 btrfs_release_path(path);
8492         if (ret == 0) {
8493                 ret = btrfs_del_item(trans, tree_root, path);
8494                 if (ret)
8495                         goto out;
8496                 btrfs_release_path(path);
8497         }
8498
8499         spin_lock(&root->fs_info->block_group_cache_lock);
8500         rb_erase(&block_group->cache_node,
8501                  &root->fs_info->block_group_cache_tree);
8502
8503         if (root->fs_info->first_logical_byte == block_group->key.objectid)
8504                 root->fs_info->first_logical_byte = (u64)-1;
8505         spin_unlock(&root->fs_info->block_group_cache_lock);
8506
8507         down_write(&block_group->space_info->groups_sem);
8508         /*
8509          * we must use list_del_init so people can check to see if they
8510          * are still on the list after taking the semaphore
8511          */
8512         list_del_init(&block_group->list);
8513         if (list_empty(&block_group->space_info->block_groups[index]))
8514                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
8515         up_write(&block_group->space_info->groups_sem);
8516
8517         if (block_group->cached == BTRFS_CACHE_STARTED)
8518                 wait_block_group_cache_done(block_group);
8519
8520         btrfs_remove_free_space_cache(block_group);
8521
8522         spin_lock(&block_group->space_info->lock);
8523         block_group->space_info->total_bytes -= block_group->key.offset;
8524         block_group->space_info->bytes_readonly -= block_group->key.offset;
8525         block_group->space_info->disk_total -= block_group->key.offset * factor;
8526         spin_unlock(&block_group->space_info->lock);
8527
8528         memcpy(&key, &block_group->key, sizeof(key));
8529
8530         btrfs_clear_space_info_full(root->fs_info);
8531
8532         btrfs_put_block_group(block_group);
8533         btrfs_put_block_group(block_group);
8534
8535         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
8536         if (ret > 0)
8537                 ret = -EIO;
8538         if (ret < 0)
8539                 goto out;
8540
8541         ret = btrfs_del_item(trans, root, path);
8542 out:
8543         btrfs_free_path(path);
8544         return ret;
8545 }
8546
8547 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
8548 {
8549         struct btrfs_space_info *space_info;
8550         struct btrfs_super_block *disk_super;
8551         u64 features;
8552         u64 flags;
8553         int mixed = 0;
8554         int ret;
8555
8556         disk_super = fs_info->super_copy;
8557         if (!btrfs_super_root(disk_super))
8558                 return 1;
8559
8560         features = btrfs_super_incompat_flags(disk_super);
8561         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
8562                 mixed = 1;
8563
8564         flags = BTRFS_BLOCK_GROUP_SYSTEM;
8565         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8566         if (ret)
8567                 goto out;
8568
8569         if (mixed) {
8570                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
8571                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8572         } else {
8573                 flags = BTRFS_BLOCK_GROUP_METADATA;
8574                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8575                 if (ret)
8576                         goto out;
8577
8578                 flags = BTRFS_BLOCK_GROUP_DATA;
8579                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
8580         }
8581 out:
8582         return ret;
8583 }
8584
8585 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
8586 {
8587         return unpin_extent_range(root, start, end);
8588 }
8589
8590 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
8591                                u64 num_bytes, u64 *actual_bytes)
8592 {
8593         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
8594 }
8595
8596 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
8597 {
8598         struct btrfs_fs_info *fs_info = root->fs_info;
8599         struct btrfs_block_group_cache *cache = NULL;
8600         u64 group_trimmed;
8601         u64 start;
8602         u64 end;
8603         u64 trimmed = 0;
8604         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
8605         int ret = 0;
8606
8607         /*
8608          * try to trim all FS space, our block group may start from non-zero.
8609          */
8610         if (range->len == total_bytes)
8611                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
8612         else
8613                 cache = btrfs_lookup_block_group(fs_info, range->start);
8614
8615         while (cache) {
8616                 if (cache->key.objectid >= (range->start + range->len)) {
8617                         btrfs_put_block_group(cache);
8618                         break;
8619                 }
8620
8621                 start = max(range->start, cache->key.objectid);
8622                 end = min(range->start + range->len,
8623                                 cache->key.objectid + cache->key.offset);
8624
8625                 if (end - start >= range->minlen) {
8626                         if (!block_group_cache_done(cache)) {
8627                                 ret = cache_block_group(cache, 0);
8628                                 if (!ret)
8629                                         wait_block_group_cache_done(cache);
8630                         }
8631                         ret = btrfs_trim_block_group(cache,
8632                                                      &group_trimmed,
8633                                                      start,
8634                                                      end,
8635                                                      range->minlen);
8636
8637                         trimmed += group_trimmed;
8638                         if (ret) {
8639                                 btrfs_put_block_group(cache);
8640                                 break;
8641                         }
8642                 }
8643
8644                 cache = next_block_group(fs_info->tree_root, cache);
8645         }
8646
8647         range->len = trimmed;
8648         return ret;
8649 }