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 <linux/percpu_counter.h>
  28 #include "hash.h"
  29 #include "tree-log.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "volumes.h"
  33 #include "raid56.h"
  34 #include "locking.h"
  35 #include "free-space-cache.h"
  36 #include "math.h"
  37 #include "sysfs.h"
  38 #include "qgroup.h"
  39
  40 #undef SCRAMBLE_DELAYED_REFS
  41
  42 /*
  43  * control flags for do_chunk_alloc's force field
  44  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  45  * if we really need one.
  46  *
  47  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  48  * if we have very few chunks already allocated.  This is
  49  * used as part of the clustering code to help make sure
  50  * we have a good pool of storage to cluster in, without
  51  * filling the FS with empty chunks
  52  *
  53  * CHUNK_ALLOC_FORCE means it must try to allocate one
  54  *
  55  */
  56 enum {
  57         CHUNK_ALLOC_NO_FORCE = 0,
  58         CHUNK_ALLOC_LIMITED = 1,
  59         CHUNK_ALLOC_FORCE = 2,
  60 };
  61
  62 /*
  63  * Control how reservations are dealt with.
  64  *
  65  * RESERVE_FREE - freeing a reservation.
  66  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  67  *   ENOSPC accounting
  68  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  69  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  70  */
  71 enum {
  72         RESERVE_FREE = 0,
  73         RESERVE_ALLOC = 1,
  74         RESERVE_ALLOC_NO_ACCOUNT = 2,
  75 };
  76
  77 static int update_block_group(struct btrfs_root *root,
  78                               u64 bytenr, u64 num_bytes, int alloc);
  79 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  80                                 struct btrfs_root *root,
  81                                 u64 bytenr, u64 num_bytes, u64 parent,
  82                                 u64 root_objectid, u64 owner_objectid,
  83                                 u64 owner_offset, int refs_to_drop,
  84                                 struct btrfs_delayed_extent_op *extra_op,
  85                                 int no_quota);
  86 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  87                                     struct extent_buffer *leaf,
  88                                     struct btrfs_extent_item *ei);
  89 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  90                                       struct btrfs_root *root,
  91                                       u64 parent, u64 root_objectid,
  92                                       u64 flags, u64 owner, u64 offset,
  93                                       struct btrfs_key *ins, int ref_mod);
  94 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  95                                      struct btrfs_root *root,
  96                                      u64 parent, u64 root_objectid,
  97                                      u64 flags, struct btrfs_disk_key *key,
  98                                      int level, struct btrfs_key *ins,
  99                                      int no_quota);
 100 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
 101                           struct btrfs_root *extent_root, u64 flags,
 102                           int force);
 103 static int find_next_key(struct btrfs_path *path, int level,
 104                          struct btrfs_key *key);
 105 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 106                             int dump_block_groups);
 107 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 108                                        u64 num_bytes, int reserve,
 109                                        int delalloc);
 110 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 111                                u64 num_bytes);
 112 int btrfs_pin_extent(struct btrfs_root *root,
 113                      u64 bytenr, u64 num_bytes, int reserved);
 114
 115 static noinline int
 116 block_group_cache_done(struct btrfs_block_group_cache *cache)
 117 {
 118         smp_mb();
 119         return cache->cached == BTRFS_CACHE_FINISHED ||
 120                 cache->cached == BTRFS_CACHE_ERROR;
 121 }
 122
 123 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 124 {
 125         return (cache->flags & bits) == bits;
 126 }
 127
 128 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 129 {
 130         atomic_inc(&cache->count);
 131 }
 132
 133 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 134 {
 135         if (atomic_dec_and_test(&cache->count)) {
 136                 WARN_ON(cache->pinned > 0);
 137                 WARN_ON(cache->reserved > 0);
 138                 kfree(cache->free_space_ctl);
 139                 kfree(cache);
 140         }
 141 }
 142
 143 /*
 144  * this adds the block group to the fs_info rb tree for the block group
 145  * cache
 146  */
 147 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 148                                 struct btrfs_block_group_cache *block_group)
 149 {
 150         struct rb_node **p;
 151         struct rb_node *parent = NULL;
 152         struct btrfs_block_group_cache *cache;
 153
 154         spin_lock(&info->block_group_cache_lock);
 155         p = &info->block_group_cache_tree.rb_node;
 156
 157         while (*p) {
 158                 parent = *p;
 159                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 160                                  cache_node);
 161                 if (block_group->key.objectid < cache->key.objectid) {
 162                         p = &(*p)->rb_left;
 163                 } else if (block_group->key.objectid > cache->key.objectid) {
 164                         p = &(*p)->rb_right;
 165                 } else {
 166                         spin_unlock(&info->block_group_cache_lock);
 167                         return -EEXIST;
 168                 }
 169         }
 170
 171         rb_link_node(&block_group->cache_node, parent, p);
 172         rb_insert_color(&block_group->cache_node,
 173                         &info->block_group_cache_tree);
 174
 175         if (info->first_logical_byte > block_group->key.objectid)
 176                 info->first_logical_byte = block_group->key.objectid;
 177
 178         spin_unlock(&info->block_group_cache_lock);
 179
 180         return 0;
 181 }
 182
 183 /*
 184  * This will return the block group at or after bytenr if contains is 0, else
 185  * it will return the block group that contains the bytenr
 186  */
 187 static struct btrfs_block_group_cache *
 188 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 189                               int contains)
 190 {
 191         struct btrfs_block_group_cache *cache, *ret = NULL;
 192         struct rb_node *n;
 193         u64 end, start;
 194
 195         spin_lock(&info->block_group_cache_lock);
 196         n = info->block_group_cache_tree.rb_node;
 197
 198         while (n) {
 199                 cache = rb_entry(n, struct btrfs_block_group_cache,
 200                                  cache_node);
 201                 end = cache->key.objectid + cache->key.offset - 1;
 202                 start = cache->key.objectid;
 203
 204                 if (bytenr < start) {
 205                         if (!contains && (!ret || start < ret->key.objectid))
 206                                 ret = cache;
 207                         n = n->rb_left;
 208                 } else if (bytenr > start) {
 209                         if (contains && bytenr <= end) {
 210                                 ret = cache;
 211                                 break;
 212                         }
 213                         n = n->rb_right;
 214                 } else {
 215                         ret = cache;
 216                         break;
 217                 }
 218         }
 219         if (ret) {
 220                 btrfs_get_block_group(ret);
 221                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 222                         info->first_logical_byte = ret->key.objectid;
 223         }
 224         spin_unlock(&info->block_group_cache_lock);
 225
 226         return ret;
 227 }
 228
 229 static int add_excluded_extent(struct btrfs_root *root,
 230                                u64 start, u64 num_bytes)
 231 {
 232         u64 end = start + num_bytes - 1;
 233         set_extent_bits(&root->fs_info->freed_extents[0],
 234                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 235         set_extent_bits(&root->fs_info->freed_extents[1],
 236                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 237         return 0;
 238 }
 239
 240 static void free_excluded_extents(struct btrfs_root *root,
 241                                   struct btrfs_block_group_cache *cache)
 242 {
 243         u64 start, end;
 244
 245         start = cache->key.objectid;
 246         end = start + cache->key.offset - 1;
 247
 248         clear_extent_bits(&root->fs_info->freed_extents[0],
 249                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 250         clear_extent_bits(&root->fs_info->freed_extents[1],
 251                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 252 }
 253
 254 static int exclude_super_stripes(struct btrfs_root *root,
 255                                  struct btrfs_block_group_cache *cache)
 256 {
 257         u64 bytenr;
 258         u64 *logical;
 259         int stripe_len;
 260         int i, nr, ret;
 261
 262         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 263                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 264                 cache->bytes_super += stripe_len;
 265                 ret = add_excluded_extent(root, cache->key.objectid,
 266                                           stripe_len);
 267                 if (ret)
 268                         return ret;
 269         }
 270
 271         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 272                 bytenr = btrfs_sb_offset(i);
 273                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 274                                        cache->key.objectid, bytenr,
 275                                        0, &logical, &nr, &stripe_len);
 276                 if (ret)
 277                         return ret;
 278
 279                 while (nr--) {
 280                         u64 start, len;
 281
 282                         if (logical[nr] > cache->key.objectid +
 283                             cache->key.offset)
 284                                 continue;
 285
 286                         if (logical[nr] + stripe_len <= cache->key.objectid)
 287                                 continue;
 288
 289                         start = logical[nr];
 290                         if (start < cache->key.objectid) {
 291                                 start = cache->key.objectid;
 292                                 len = (logical[nr] + stripe_len) - start;
 293                         } else {
 294                                 len = min_t(u64, stripe_len,
 295                                             cache->key.objectid +
 296                                             cache->key.offset - start);
 297                         }
 298
 299                         cache->bytes_super += len;
 300                         ret = add_excluded_extent(root, start, len);
 301                         if (ret) {
 302                                 kfree(logical);
 303                                 return ret;
 304                         }
 305                 }
 306
 307                 kfree(logical);
 308         }
 309         return 0;
 310 }
 311
 312 static struct btrfs_caching_control *
 313 get_caching_control(struct btrfs_block_group_cache *cache)
 314 {
 315         struct btrfs_caching_control *ctl;
 316
 317         spin_lock(&cache->lock);
 318         if (cache->cached != BTRFS_CACHE_STARTED) {
 319                 spin_unlock(&cache->lock);
 320                 return NULL;
 321         }
 322
 323         /* We're loading it the fast way, so we don't have a caching_ctl. */
 324         if (!cache->caching_ctl) {
 325                 spin_unlock(&cache->lock);
 326                 return NULL;
 327         }
 328
 329         ctl = cache->caching_ctl;
 330         atomic_inc(&ctl->count);
 331         spin_unlock(&cache->lock);
 332         return ctl;
 333 }
 334
 335 static void put_caching_control(struct btrfs_caching_control *ctl)
 336 {
 337         if (atomic_dec_and_test(&ctl->count))
 338                 kfree(ctl);
 339 }
 340
 341 /*
 342  * this is only called by cache_block_group, since we could have freed extents
 343  * we need to check the pinned_extents for any extents that can't be used yet
 344  * since their free space will be released as soon as the transaction commits.
 345  */
 346 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 347                               struct btrfs_fs_info *info, u64 start, u64 end)
 348 {
 349         u64 extent_start, extent_end, size, total_added = 0;
 350         int ret;
 351
 352         while (start < end) {
 353                 ret = find_first_extent_bit(info->pinned_extents, start,
 354                                             &extent_start, &extent_end,
 355                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 356                                             NULL);
 357                 if (ret)
 358                         break;
 359
 360                 if (extent_start <= start) {
 361                         start = extent_end + 1;
 362                 } else if (extent_start > start && extent_start < end) {
 363                         size = extent_start - start;
 364                         total_added += size;
 365                         ret = btrfs_add_free_space(block_group, start,
 366                                                    size);
 367                         BUG_ON(ret); /* -ENOMEM or logic error */
 368                         start = extent_end + 1;
 369                 } else {
 370                         break;
 371                 }
 372         }
 373
 374         if (start < end) {
 375                 size = end - start;
 376                 total_added += size;
 377                 ret = btrfs_add_free_space(block_group, start, size);
 378                 BUG_ON(ret); /* -ENOMEM or logic error */
 379         }
 380
 381         return total_added;
 382 }
 383
 384 static noinline void caching_thread(struct btrfs_work *work)
 385 {
 386         struct btrfs_block_group_cache *block_group;
 387         struct btrfs_fs_info *fs_info;
 388         struct btrfs_caching_control *caching_ctl;
 389         struct btrfs_root *extent_root;
 390         struct btrfs_path *path;
 391         struct extent_buffer *leaf;
 392         struct btrfs_key key;
 393         u64 total_found = 0;
 394         u64 last = 0;
 395         u32 nritems;
 396         int ret = -ENOMEM;
 397
 398         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 399         block_group = caching_ctl->block_group;
 400         fs_info = block_group->fs_info;
 401         extent_root = fs_info->extent_root;
 402
 403         path = btrfs_alloc_path();
 404         if (!path)
 405                 goto out;
 406
 407         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 408
 409         /*
 410          * We don't want to deadlock with somebody trying to allocate a new
 411          * extent for the extent root while also trying to search the extent
 412          * root to add free space.  So we skip locking and search the commit
 413          * root, since its read-only
 414          */
 415         path->skip_locking = 1;
 416         path->search_commit_root = 1;
 417         path->reada = 1;
 418
 419         key.objectid = last;
 420         key.offset = 0;
 421         key.type = BTRFS_EXTENT_ITEM_KEY;
 422 again:
 423         mutex_lock(&caching_ctl->mutex);
 424         /* need to make sure the commit_root doesn't disappear */
 425         down_read(&fs_info->commit_root_sem);
 426
 427 next:
 428         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 429         if (ret < 0)
 430                 goto err;
 431
 432         leaf = path->nodes[0];
 433         nritems = btrfs_header_nritems(leaf);
 434
 435         while (1) {
 436                 if (btrfs_fs_closing(fs_info) > 1) {
 437                         last = (u64)-1;
 438                         break;
 439                 }
 440
 441                 if (path->slots[0] < nritems) {
 442                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 443                 } else {
 444                         ret = find_next_key(path, 0, &key);
 445                         if (ret)
 446                                 break;
 447
 448                         if (need_resched() ||
 449                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 450                                 caching_ctl->progress = last;
 451                                 btrfs_release_path(path);
 452                                 up_read(&fs_info->commit_root_sem);
 453                                 mutex_unlock(&caching_ctl->mutex);
 454                                 cond_resched();
 455                                 goto again;
 456                         }
 457
 458                         ret = btrfs_next_leaf(extent_root, path);
 459                         if (ret < 0)
 460                                 goto err;
 461                         if (ret)
 462                                 break;
 463                         leaf = path->nodes[0];
 464                         nritems = btrfs_header_nritems(leaf);
 465                         continue;
 466                 }
 467
 468                 if (key.objectid < last) {
 469                         key.objectid = last;
 470                         key.offset = 0;
 471                         key.type = BTRFS_EXTENT_ITEM_KEY;
 472
 473                         caching_ctl->progress = last;
 474                         btrfs_release_path(path);
 475                         goto next;
 476                 }
 477
 478                 if (key.objectid < block_group->key.objectid) {
 479                         path->slots[0]++;
 480                         continue;
 481                 }
 482
 483                 if (key.objectid >= block_group->key.objectid +
 484                     block_group->key.offset)
 485                         break;
 486
 487                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 488                     key.type == BTRFS_METADATA_ITEM_KEY) {
 489                         total_found += add_new_free_space(block_group,
 490                                                           fs_info, last,
 491                                                           key.objectid);
 492                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 493                                 last = key.objectid +
 494                                         fs_info->tree_root->nodesize;
 495                         else
 496                                 last = key.objectid + key.offset;
 497
 498                         if (total_found > (1024 * 1024 * 2)) {
 499                                 total_found = 0;
 500                                 wake_up(&caching_ctl->wait);
 501                         }
 502                 }
 503                 path->slots[0]++;
 504         }
 505         ret = 0;
 506
 507         total_found += add_new_free_space(block_group, fs_info, last,
 508                                           block_group->key.objectid +
 509                                           block_group->key.offset);
 510         caching_ctl->progress = (u64)-1;
 511
 512         spin_lock(&block_group->lock);
 513         block_group->caching_ctl = NULL;
 514         block_group->cached = BTRFS_CACHE_FINISHED;
 515         spin_unlock(&block_group->lock);
 516
 517 err:
 518         btrfs_free_path(path);
 519         up_read(&fs_info->commit_root_sem);
 520
 521         free_excluded_extents(extent_root, block_group);
 522
 523         mutex_unlock(&caching_ctl->mutex);
 524 out:
 525         if (ret) {
 526                 spin_lock(&block_group->lock);
 527                 block_group->caching_ctl = NULL;
 528                 block_group->cached = BTRFS_CACHE_ERROR;
 529                 spin_unlock(&block_group->lock);
 530         }
 531         wake_up(&caching_ctl->wait);
 532
 533         put_caching_control(caching_ctl);
 534         btrfs_put_block_group(block_group);
 535 }
 536
 537 static int cache_block_group(struct btrfs_block_group_cache *cache,
 538                              int load_cache_only)
 539 {
 540         DEFINE_WAIT(wait);
 541         struct btrfs_fs_info *fs_info = cache->fs_info;
 542         struct btrfs_caching_control *caching_ctl;
 543         int ret = 0;
 544
 545         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 546         if (!caching_ctl)
 547                 return -ENOMEM;
 548
 549         INIT_LIST_HEAD(&caching_ctl->list);
 550         mutex_init(&caching_ctl->mutex);
 551         init_waitqueue_head(&caching_ctl->wait);
 552         caching_ctl->block_group = cache;
 553         caching_ctl->progress = cache->key.objectid;
 554         atomic_set(&caching_ctl->count, 1);
 555         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 556                         caching_thread, NULL, NULL);
 557
 558         spin_lock(&cache->lock);
 559         /*
 560          * This should be a rare occasion, but this could happen I think in the
 561          * case where one thread starts to load the space cache info, and then
 562          * some other thread starts a transaction commit which tries to do an
 563          * allocation while the other thread is still loading the space cache
 564          * info.  The previous loop should have kept us from choosing this block
 565          * group, but if we've moved to the state where we will wait on caching
 566          * block groups we need to first check if we're doing a fast load here,
 567          * so we can wait for it to finish, otherwise we could end up allocating
 568          * from a block group who's cache gets evicted for one reason or
 569          * another.
 570          */
 571         while (cache->cached == BTRFS_CACHE_FAST) {
 572                 struct btrfs_caching_control *ctl;
 573
 574                 ctl = cache->caching_ctl;
 575                 atomic_inc(&ctl->count);
 576                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 577                 spin_unlock(&cache->lock);
 578
 579                 schedule();
 580
 581                 finish_wait(&ctl->wait, &wait);
 582                 put_caching_control(ctl);
 583                 spin_lock(&cache->lock);
 584         }
 585
 586         if (cache->cached != BTRFS_CACHE_NO) {
 587                 spin_unlock(&cache->lock);
 588                 kfree(caching_ctl);
 589                 return 0;
 590         }
 591         WARN_ON(cache->caching_ctl);
 592         cache->caching_ctl = caching_ctl;
 593         cache->cached = BTRFS_CACHE_FAST;
 594         spin_unlock(&cache->lock);
 595
 596         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 597                 ret = load_free_space_cache(fs_info, cache);
 598
 599                 spin_lock(&cache->lock);
 600                 if (ret == 1) {
 601                         cache->caching_ctl = NULL;
 602                         cache->cached = BTRFS_CACHE_FINISHED;
 603                         cache->last_byte_to_unpin = (u64)-1;
 604                 } else {
 605                         if (load_cache_only) {
 606                                 cache->caching_ctl = NULL;
 607                                 cache->cached = BTRFS_CACHE_NO;
 608                         } else {
 609                                 cache->cached = BTRFS_CACHE_STARTED;
 610                         }
 611                 }
 612                 spin_unlock(&cache->lock);
 613                 wake_up(&caching_ctl->wait);
 614                 if (ret == 1) {
 615                         put_caching_control(caching_ctl);
 616                         free_excluded_extents(fs_info->extent_root, cache);
 617                         return 0;
 618                 }
 619         } else {
 620                 /*
 621                  * We are not going to do the fast caching, set cached to the
 622                  * appropriate value and wakeup any waiters.
 623                  */
 624                 spin_lock(&cache->lock);
 625                 if (load_cache_only) {
 626                         cache->caching_ctl = NULL;
 627                         cache->cached = BTRFS_CACHE_NO;
 628                 } else {
 629                         cache->cached = BTRFS_CACHE_STARTED;
 630                 }
 631                 spin_unlock(&cache->lock);
 632                 wake_up(&caching_ctl->wait);
 633         }
 634
 635         if (load_cache_only) {
 636                 put_caching_control(caching_ctl);
 637                 return 0;
 638         }
 639
 640         down_write(&fs_info->commit_root_sem);
 641         atomic_inc(&caching_ctl->count);
 642         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 643         up_write(&fs_info->commit_root_sem);
 644
 645         btrfs_get_block_group(cache);
 646
 647         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 648
 649         return ret;
 650 }
 651
 652 /*
 653  * return the block group that starts at or after bytenr
 654  */
 655 static struct btrfs_block_group_cache *
 656 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 657 {
 658         struct btrfs_block_group_cache *cache;
 659
 660         cache = block_group_cache_tree_search(info, bytenr, 0);
 661
 662         return cache;
 663 }
 664
 665 /*
 666  * return the block group that contains the given bytenr
 667  */
 668 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 669                                                  struct btrfs_fs_info *info,
 670                                                  u64 bytenr)
 671 {
 672         struct btrfs_block_group_cache *cache;
 673
 674         cache = block_group_cache_tree_search(info, bytenr, 1);
 675
 676         return cache;
 677 }
 678
 679 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 680                                                   u64 flags)
 681 {
 682         struct list_head *head = &info->space_info;
 683         struct btrfs_space_info *found;
 684
 685         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 686
 687         rcu_read_lock();
 688         list_for_each_entry_rcu(found, head, list) {
 689                 if (found->flags & flags) {
 690                         rcu_read_unlock();
 691                         return found;
 692                 }
 693         }
 694         rcu_read_unlock();
 695         return NULL;
 696 }
 697
 698 /*
 699  * after adding space to the filesystem, we need to clear the full flags
 700  * on all the space infos.
 701  */
 702 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 703 {
 704         struct list_head *head = &info->space_info;
 705         struct btrfs_space_info *found;
 706
 707         rcu_read_lock();
 708         list_for_each_entry_rcu(found, head, list)
 709                 found->full = 0;
 710         rcu_read_unlock();
 711 }
 712
 713 /* simple helper to search for an existing data extent at a given offset */
 714 int btrfs_lookup_data_extent(struct btrfs_root *root, u64 start, u64 len)
 715 {
 716         int ret;
 717         struct btrfs_key key;
 718         struct btrfs_path *path;
 719
 720         path = btrfs_alloc_path();
 721         if (!path)
 722                 return -ENOMEM;
 723
 724         key.objectid = start;
 725         key.offset = len;
 726         key.type = BTRFS_EXTENT_ITEM_KEY;
 727         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 728                                 0, 0);
 729         btrfs_free_path(path);
 730         return ret;
 731 }
 732
 733 /*
 734  * helper function to lookup reference count and flags of a tree block.
 735  *
 736  * the head node for delayed ref is used to store the sum of all the
 737  * reference count modifications queued up in the rbtree. the head
 738  * node may also store the extent flags to set. This way you can check
 739  * to see what the reference count and extent flags would be if all of
 740  * the delayed refs are not processed.
 741  */
 742 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 743                              struct btrfs_root *root, u64 bytenr,
 744                              u64 offset, int metadata, u64 *refs, u64 *flags)
 745 {
 746         struct btrfs_delayed_ref_head *head;
 747         struct btrfs_delayed_ref_root *delayed_refs;
 748         struct btrfs_path *path;
 749         struct btrfs_extent_item *ei;
 750         struct extent_buffer *leaf;
 751         struct btrfs_key key;
 752         u32 item_size;
 753         u64 num_refs;
 754         u64 extent_flags;
 755         int ret;
 756
 757         /*
 758          * If we don't have skinny metadata, don't bother doing anything
 759          * different
 760          */
 761         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 762                 offset = root->nodesize;
 763                 metadata = 0;
 764         }
 765
 766         path = btrfs_alloc_path();
 767         if (!path)
 768                 return -ENOMEM;
 769
 770         if (!trans) {
 771                 path->skip_locking = 1;
 772                 path->search_commit_root = 1;
 773         }
 774
 775 search_again:
 776         key.objectid = bytenr;
 777         key.offset = offset;
 778         if (metadata)
 779                 key.type = BTRFS_METADATA_ITEM_KEY;
 780         else
 781                 key.type = BTRFS_EXTENT_ITEM_KEY;
 782
 783         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 784                                 &key, path, 0, 0);
 785         if (ret < 0)
 786                 goto out_free;
 787
 788         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 789                 if (path->slots[0]) {
 790                         path->slots[0]--;
 791                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 792                                               path->slots[0]);
 793                         if (key.objectid == bytenr &&
 794                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 795                             key.offset == root->nodesize)
 796                                 ret = 0;
 797                 }
 798         }
 799
 800         if (ret == 0) {
 801                 leaf = path->nodes[0];
 802                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 803                 if (item_size >= sizeof(*ei)) {
 804                         ei = btrfs_item_ptr(leaf, path->slots[0],
 805                                             struct btrfs_extent_item);
 806                         num_refs = btrfs_extent_refs(leaf, ei);
 807                         extent_flags = btrfs_extent_flags(leaf, ei);
 808                 } else {
 809 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 810                         struct btrfs_extent_item_v0 *ei0;
 811                         BUG_ON(item_size != sizeof(*ei0));
 812                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 813                                              struct btrfs_extent_item_v0);
 814                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 815                         /* FIXME: this isn't correct for data */
 816                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 817 #else
 818                         BUG();
 819 #endif
 820                 }
 821                 BUG_ON(num_refs == 0);
 822         } else {
 823                 num_refs = 0;
 824                 extent_flags = 0;
 825                 ret = 0;
 826         }
 827
 828         if (!trans)
 829                 goto out;
 830
 831         delayed_refs = &trans->transaction->delayed_refs;
 832         spin_lock(&delayed_refs->lock);
 833         head = btrfs_find_delayed_ref_head(trans, bytenr);
 834         if (head) {
 835                 if (!mutex_trylock(&head->mutex)) {
 836                         atomic_inc(&head->node.refs);
 837                         spin_unlock(&delayed_refs->lock);
 838
 839                         btrfs_release_path(path);
 840
 841                         /*
 842                          * Mutex was contended, block until it's released and try
 843                          * again
 844                          */
 845                         mutex_lock(&head->mutex);
 846                         mutex_unlock(&head->mutex);
 847                         btrfs_put_delayed_ref(&head->node);
 848                         goto search_again;
 849                 }
 850                 spin_lock(&head->lock);
 851                 if (head->extent_op && head->extent_op->update_flags)
 852                         extent_flags |= head->extent_op->flags_to_set;
 853                 else
 854                         BUG_ON(num_refs == 0);
 855
 856                 num_refs += head->node.ref_mod;
 857                 spin_unlock(&head->lock);
 858                 mutex_unlock(&head->mutex);
 859         }
 860         spin_unlock(&delayed_refs->lock);
 861 out:
 862         WARN_ON(num_refs == 0);
 863         if (refs)
 864                 *refs = num_refs;
 865         if (flags)
 866                 *flags = extent_flags;
 867 out_free:
 868         btrfs_free_path(path);
 869         return ret;
 870 }
 871
 872 /*
 873  * Back reference rules.  Back refs have three main goals:
 874  *
 875  * 1) differentiate between all holders of references to an extent so that
 876  *    when a reference is dropped we can make sure it was a valid reference
 877  *    before freeing the extent.
 878  *
 879  * 2) Provide enough information to quickly find the holders of an extent
 880  *    if we notice a given block is corrupted or bad.
 881  *
 882  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 883  *    maintenance.  This is actually the same as #2, but with a slightly
 884  *    different use case.
 885  *
 886  * There are two kinds of back refs. The implicit back refs is optimized
 887  * for pointers in non-shared tree blocks. For a given pointer in a block,
 888  * back refs of this kind provide information about the block's owner tree
 889  * and the pointer's key. These information allow us to find the block by
 890  * b-tree searching. The full back refs is for pointers in tree blocks not
 891  * referenced by their owner trees. The location of tree block is recorded
 892  * in the back refs. Actually the full back refs is generic, and can be
 893  * used in all cases the implicit back refs is used. The major shortcoming
 894  * of the full back refs is its overhead. Every time a tree block gets
 895  * COWed, we have to update back refs entry for all pointers in it.
 896  *
 897  * For a newly allocated tree block, we use implicit back refs for
 898  * pointers in it. This means most tree related operations only involve
 899  * implicit back refs. For a tree block created in old transaction, the
 900  * only way to drop a reference to it is COW it. So we can detect the
 901  * event that tree block loses its owner tree's reference and do the
 902  * back refs conversion.
 903  *
 904  * When a tree block is COW'd through a tree, there are four cases:
 905  *
 906  * The reference count of the block is one and the tree is the block's
 907  * owner tree. Nothing to do in this case.
 908  *
 909  * The reference count of the block is one and the tree is not the
 910  * block's owner tree. In this case, full back refs is used for pointers
 911  * in the block. Remove these full back refs, add implicit back refs for
 912  * every pointers in the new block.
 913  *
 914  * The reference count of the block is greater than one and the tree is
 915  * the block's owner tree. In this case, implicit back refs is used for
 916  * pointers in the block. Add full back refs for every pointers in the
 917  * block, increase lower level extents' reference counts. The original
 918  * implicit back refs are entailed to the new block.
 919  *
 920  * The reference count of the block is greater than one and the tree is
 921  * not the block's owner tree. Add implicit back refs for every pointer in
 922  * the new block, increase lower level extents' reference count.
 923  *
 924  * Back Reference Key composing:
 925  *
 926  * The key objectid corresponds to the first byte in the extent,
 927  * The key type is used to differentiate between types of back refs.
 928  * There are different meanings of the key offset for different types
 929  * of back refs.
 930  *
 931  * File extents can be referenced by:
 932  *
 933  * - multiple snapshots, subvolumes, or different generations in one subvol
 934  * - different files inside a single subvolume
 935  * - different offsets inside a file (bookend extents in file.c)
 936  *
 937  * The extent ref structure for the implicit back refs has fields for:
 938  *
 939  * - Objectid of the subvolume root
 940  * - objectid of the file holding the reference
 941  * - original offset in the file
 942  * - how many bookend extents
 943  *
 944  * The key offset for the implicit back refs is hash of the first
 945  * three fields.
 946  *
 947  * The extent ref structure for the full back refs has field for:
 948  *
 949  * - number of pointers in the tree leaf
 950  *
 951  * The key offset for the implicit back refs is the first byte of
 952  * the tree leaf
 953  *
 954  * When a file extent is allocated, The implicit back refs is used.
 955  * the fields are filled in:
 956  *
 957  *     (root_key.objectid, inode objectid, offset in file, 1)
 958  *
 959  * When a file extent is removed file truncation, we find the
 960  * corresponding implicit back refs and check the following fields:
 961  *
 962  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 963  *
 964  * Btree extents can be referenced by:
 965  *
 966  * - Different subvolumes
 967  *
 968  * Both the implicit back refs and the full back refs for tree blocks
 969  * only consist of key. The key offset for the implicit back refs is
 970  * objectid of block's owner tree. The key offset for the full back refs
 971  * is the first byte of parent block.
 972  *
 973  * When implicit back refs is used, information about the lowest key and
 974  * level of the tree block are required. These information are stored in
 975  * tree block info structure.
 976  */
 977
 978 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 979 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 980                                   struct btrfs_root *root,
 981                                   struct btrfs_path *path,
 982                                   u64 owner, u32 extra_size)
 983 {
 984         struct btrfs_extent_item *item;
 985         struct btrfs_extent_item_v0 *ei0;
 986         struct btrfs_extent_ref_v0 *ref0;
 987         struct btrfs_tree_block_info *bi;
 988         struct extent_buffer *leaf;
 989         struct btrfs_key key;
 990         struct btrfs_key found_key;
 991         u32 new_size = sizeof(*item);
 992         u64 refs;
 993         int ret;
 994
 995         leaf = path->nodes[0];
 996         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
 997
 998         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 999         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1000                              struct btrfs_extent_item_v0);
1001         refs = btrfs_extent_refs_v0(leaf, ei0);
1002
1003         if (owner == (u64)-1) {
1004                 while (1) {
1005                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1006                                 ret = btrfs_next_leaf(root, path);
1007                                 if (ret < 0)
1008                                         return ret;
1009                                 BUG_ON(ret > 0); /* Corruption */
1010                                 leaf = path->nodes[0];
1011                         }
1012                         btrfs_item_key_to_cpu(leaf, &found_key,
1013                                               path->slots[0]);
1014                         BUG_ON(key.objectid != found_key.objectid);
1015                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1016                                 path->slots[0]++;
1017                                 continue;
1018                         }
1019                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1020                                               struct btrfs_extent_ref_v0);
1021                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1022                         break;
1023                 }
1024         }
1025         btrfs_release_path(path);
1026
1027         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1028                 new_size += sizeof(*bi);
1029
1030         new_size -= sizeof(*ei0);
1031         ret = btrfs_search_slot(trans, root, &key, path,
1032                                 new_size + extra_size, 1);
1033         if (ret < 0)
1034                 return ret;
1035         BUG_ON(ret); /* Corruption */
1036
1037         btrfs_extend_item(root, path, new_size);
1038
1039         leaf = path->nodes[0];
1040         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1041         btrfs_set_extent_refs(leaf, item, refs);
1042         /* FIXME: get real generation */
1043         btrfs_set_extent_generation(leaf, item, 0);
1044         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1045                 btrfs_set_extent_flags(leaf, item,
1046                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1047                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1048                 bi = (struct btrfs_tree_block_info *)(item + 1);
1049                 /* FIXME: get first key of the block */
1050                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1051                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1052         } else {
1053                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1054         }
1055         btrfs_mark_buffer_dirty(leaf);
1056         return 0;
1057 }
1058 #endif
1059
1060 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1061 {
1062         u32 high_crc = ~(u32)0;
1063         u32 low_crc = ~(u32)0;
1064         __le64 lenum;
1065
1066         lenum = cpu_to_le64(root_objectid);
1067         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1068         lenum = cpu_to_le64(owner);
1069         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1070         lenum = cpu_to_le64(offset);
1071         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1072
1073         return ((u64)high_crc << 31) ^ (u64)low_crc;
1074 }
1075
1076 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1077                                      struct btrfs_extent_data_ref *ref)
1078 {
1079         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1080                                     btrfs_extent_data_ref_objectid(leaf, ref),
1081                                     btrfs_extent_data_ref_offset(leaf, ref));
1082 }
1083
1084 static int match_extent_data_ref(struct extent_buffer *leaf,
1085                                  struct btrfs_extent_data_ref *ref,
1086                                  u64 root_objectid, u64 owner, u64 offset)
1087 {
1088         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1089             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1090             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1091                 return 0;
1092         return 1;
1093 }
1094
1095 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1096                                            struct btrfs_root *root,
1097                                            struct btrfs_path *path,
1098                                            u64 bytenr, u64 parent,
1099                                            u64 root_objectid,
1100                                            u64 owner, u64 offset)
1101 {
1102         struct btrfs_key key;
1103         struct btrfs_extent_data_ref *ref;
1104         struct extent_buffer *leaf;
1105         u32 nritems;
1106         int ret;
1107         int recow;
1108         int err = -ENOENT;
1109
1110         key.objectid = bytenr;
1111         if (parent) {
1112                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1113                 key.offset = parent;
1114         } else {
1115                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1116                 key.offset = hash_extent_data_ref(root_objectid,
1117                                                   owner, offset);
1118         }
1119 again:
1120         recow = 0;
1121         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1122         if (ret < 0) {
1123                 err = ret;
1124                 goto fail;
1125         }
1126
1127         if (parent) {
1128                 if (!ret)
1129                         return 0;
1130 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1131                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1132                 btrfs_release_path(path);
1133                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1134                 if (ret < 0) {
1135                         err = ret;
1136                         goto fail;
1137                 }
1138                 if (!ret)
1139                         return 0;
1140 #endif
1141                 goto fail;
1142         }
1143
1144         leaf = path->nodes[0];
1145         nritems = btrfs_header_nritems(leaf);
1146         while (1) {
1147                 if (path->slots[0] >= nritems) {
1148                         ret = btrfs_next_leaf(root, path);
1149                         if (ret < 0)
1150                                 err = ret;
1151                         if (ret)
1152                                 goto fail;
1153
1154                         leaf = path->nodes[0];
1155                         nritems = btrfs_header_nritems(leaf);
1156                         recow = 1;
1157                 }
1158
1159                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1160                 if (key.objectid != bytenr ||
1161                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1162                         goto fail;
1163
1164                 ref = btrfs_item_ptr(leaf, path->slots[0],
1165                                      struct btrfs_extent_data_ref);
1166
1167                 if (match_extent_data_ref(leaf, ref, root_objectid,
1168                                           owner, offset)) {
1169                         if (recow) {
1170                                 btrfs_release_path(path);
1171                                 goto again;
1172                         }
1173                         err = 0;
1174                         break;
1175                 }
1176                 path->slots[0]++;
1177         }
1178 fail:
1179         return err;
1180 }
1181
1182 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1183                                            struct btrfs_root *root,
1184                                            struct btrfs_path *path,
1185                                            u64 bytenr, u64 parent,
1186                                            u64 root_objectid, u64 owner,
1187                                            u64 offset, int refs_to_add)
1188 {
1189         struct btrfs_key key;
1190         struct extent_buffer *leaf;
1191         u32 size;
1192         u32 num_refs;
1193         int ret;
1194
1195         key.objectid = bytenr;
1196         if (parent) {
1197                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1198                 key.offset = parent;
1199                 size = sizeof(struct btrfs_shared_data_ref);
1200         } else {
1201                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1202                 key.offset = hash_extent_data_ref(root_objectid,
1203                                                   owner, offset);
1204                 size = sizeof(struct btrfs_extent_data_ref);
1205         }
1206
1207         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1208         if (ret && ret != -EEXIST)
1209                 goto fail;
1210
1211         leaf = path->nodes[0];
1212         if (parent) {
1213                 struct btrfs_shared_data_ref *ref;
1214                 ref = btrfs_item_ptr(leaf, path->slots[0],
1215                                      struct btrfs_shared_data_ref);
1216                 if (ret == 0) {
1217                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1218                 } else {
1219                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1220                         num_refs += refs_to_add;
1221                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1222                 }
1223         } else {
1224                 struct btrfs_extent_data_ref *ref;
1225                 while (ret == -EEXIST) {
1226                         ref = btrfs_item_ptr(leaf, path->slots[0],
1227                                              struct btrfs_extent_data_ref);
1228                         if (match_extent_data_ref(leaf, ref, root_objectid,
1229                                                   owner, offset))
1230                                 break;
1231                         btrfs_release_path(path);
1232                         key.offset++;
1233                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1234                                                       size);
1235                         if (ret && ret != -EEXIST)
1236                                 goto fail;
1237
1238                         leaf = path->nodes[0];
1239                 }
1240                 ref = btrfs_item_ptr(leaf, path->slots[0],
1241                                      struct btrfs_extent_data_ref);
1242                 if (ret == 0) {
1243                         btrfs_set_extent_data_ref_root(leaf, ref,
1244                                                        root_objectid);
1245                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1246                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1247                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1248                 } else {
1249                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1250                         num_refs += refs_to_add;
1251                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1252                 }
1253         }
1254         btrfs_mark_buffer_dirty(leaf);
1255         ret = 0;
1256 fail:
1257         btrfs_release_path(path);
1258         return ret;
1259 }
1260
1261 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1262                                            struct btrfs_root *root,
1263                                            struct btrfs_path *path,
1264                                            int refs_to_drop, int *last_ref)
1265 {
1266         struct btrfs_key key;
1267         struct btrfs_extent_data_ref *ref1 = NULL;
1268         struct btrfs_shared_data_ref *ref2 = NULL;
1269         struct extent_buffer *leaf;
1270         u32 num_refs = 0;
1271         int ret = 0;
1272
1273         leaf = path->nodes[0];
1274         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1275
1276         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1277                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1278                                       struct btrfs_extent_data_ref);
1279                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1280         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1281                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1282                                       struct btrfs_shared_data_ref);
1283                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1284 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1285         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1286                 struct btrfs_extent_ref_v0 *ref0;
1287                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1288                                       struct btrfs_extent_ref_v0);
1289                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1290 #endif
1291         } else {
1292                 BUG();
1293         }
1294
1295         BUG_ON(num_refs < refs_to_drop);
1296         num_refs -= refs_to_drop;
1297
1298         if (num_refs == 0) {
1299                 ret = btrfs_del_item(trans, root, path);
1300                 *last_ref = 1;
1301         } else {
1302                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1303                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1304                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1305                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1306 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1307                 else {
1308                         struct btrfs_extent_ref_v0 *ref0;
1309                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1310                                         struct btrfs_extent_ref_v0);
1311                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1312                 }
1313 #endif
1314                 btrfs_mark_buffer_dirty(leaf);
1315         }
1316         return ret;
1317 }
1318
1319 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1320                                           struct btrfs_path *path,
1321                                           struct btrfs_extent_inline_ref *iref)
1322 {
1323         struct btrfs_key key;
1324         struct extent_buffer *leaf;
1325         struct btrfs_extent_data_ref *ref1;
1326         struct btrfs_shared_data_ref *ref2;
1327         u32 num_refs = 0;
1328
1329         leaf = path->nodes[0];
1330         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1331         if (iref) {
1332                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1333                     BTRFS_EXTENT_DATA_REF_KEY) {
1334                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1335                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1336                 } else {
1337                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1338                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1339                 }
1340         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1341                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1342                                       struct btrfs_extent_data_ref);
1343                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1344         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1345                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1346                                       struct btrfs_shared_data_ref);
1347                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1348 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1349         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1350                 struct btrfs_extent_ref_v0 *ref0;
1351                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1352                                       struct btrfs_extent_ref_v0);
1353                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1354 #endif
1355         } else {
1356                 WARN_ON(1);
1357         }
1358         return num_refs;
1359 }
1360
1361 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1362                                           struct btrfs_root *root,
1363                                           struct btrfs_path *path,
1364                                           u64 bytenr, u64 parent,
1365                                           u64 root_objectid)
1366 {
1367         struct btrfs_key key;
1368         int ret;
1369
1370         key.objectid = bytenr;
1371         if (parent) {
1372                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1373                 key.offset = parent;
1374         } else {
1375                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1376                 key.offset = root_objectid;
1377         }
1378
1379         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1380         if (ret > 0)
1381                 ret = -ENOENT;
1382 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1383         if (ret == -ENOENT && parent) {
1384                 btrfs_release_path(path);
1385                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1386                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1387                 if (ret > 0)
1388                         ret = -ENOENT;
1389         }
1390 #endif
1391         return ret;
1392 }
1393
1394 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1395                                           struct btrfs_root *root,
1396                                           struct btrfs_path *path,
1397                                           u64 bytenr, u64 parent,
1398                                           u64 root_objectid)
1399 {
1400         struct btrfs_key key;
1401         int ret;
1402
1403         key.objectid = bytenr;
1404         if (parent) {
1405                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1406                 key.offset = parent;
1407         } else {
1408                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1409                 key.offset = root_objectid;
1410         }
1411
1412         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1413         btrfs_release_path(path);
1414         return ret;
1415 }
1416
1417 static inline int extent_ref_type(u64 parent, u64 owner)
1418 {
1419         int type;
1420         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1421                 if (parent > 0)
1422                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1423                 else
1424                         type = BTRFS_TREE_BLOCK_REF_KEY;
1425         } else {
1426                 if (parent > 0)
1427                         type = BTRFS_SHARED_DATA_REF_KEY;
1428                 else
1429                         type = BTRFS_EXTENT_DATA_REF_KEY;
1430         }
1431         return type;
1432 }
1433
1434 static int find_next_key(struct btrfs_path *path, int level,
1435                          struct btrfs_key *key)
1436
1437 {
1438         for (; level < BTRFS_MAX_LEVEL; level++) {
1439                 if (!path->nodes[level])
1440                         break;
1441                 if (path->slots[level] + 1 >=
1442                     btrfs_header_nritems(path->nodes[level]))
1443                         continue;
1444                 if (level == 0)
1445                         btrfs_item_key_to_cpu(path->nodes[level], key,
1446                                               path->slots[level] + 1);
1447                 else
1448                         btrfs_node_key_to_cpu(path->nodes[level], key,
1449                                               path->slots[level] + 1);
1450                 return 0;
1451         }
1452         return 1;
1453 }
1454
1455 /*
1456  * look for inline back ref. if back ref is found, *ref_ret is set
1457  * to the address of inline back ref, and 0 is returned.
1458  *
1459  * if back ref isn't found, *ref_ret is set to the address where it
1460  * should be inserted, and -ENOENT is returned.
1461  *
1462  * if insert is true and there are too many inline back refs, the path
1463  * points to the extent item, and -EAGAIN is returned.
1464  *
1465  * NOTE: inline back refs are ordered in the same way that back ref
1466  *       items in the tree are ordered.
1467  */
1468 static noinline_for_stack
1469 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1470                                  struct btrfs_root *root,
1471                                  struct btrfs_path *path,
1472                                  struct btrfs_extent_inline_ref **ref_ret,
1473                                  u64 bytenr, u64 num_bytes,
1474                                  u64 parent, u64 root_objectid,
1475                                  u64 owner, u64 offset, int insert)
1476 {
1477         struct btrfs_key key;
1478         struct extent_buffer *leaf;
1479         struct btrfs_extent_item *ei;
1480         struct btrfs_extent_inline_ref *iref;
1481         u64 flags;
1482         u64 item_size;
1483         unsigned long ptr;
1484         unsigned long end;
1485         int extra_size;
1486         int type;
1487         int want;
1488         int ret;
1489         int err = 0;
1490         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1491                                                  SKINNY_METADATA);
1492
1493         key.objectid = bytenr;
1494         key.type = BTRFS_EXTENT_ITEM_KEY;
1495         key.offset = num_bytes;
1496
1497         want = extent_ref_type(parent, owner);
1498         if (insert) {
1499                 extra_size = btrfs_extent_inline_ref_size(want);
1500                 path->keep_locks = 1;
1501         } else
1502                 extra_size = -1;
1503
1504         /*
1505          * Owner is our parent level, so we can just add one to get the level
1506          * for the block we are interested in.
1507          */
1508         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1509                 key.type = BTRFS_METADATA_ITEM_KEY;
1510                 key.offset = owner;
1511         }
1512
1513 again:
1514         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1515         if (ret < 0) {
1516                 err = ret;
1517                 goto out;
1518         }
1519
1520         /*
1521          * We may be a newly converted file system which still has the old fat
1522          * extent entries for metadata, so try and see if we have one of those.
1523          */
1524         if (ret > 0 && skinny_metadata) {
1525                 skinny_metadata = false;
1526                 if (path->slots[0]) {
1527                         path->slots[0]--;
1528                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1529                                               path->slots[0]);
1530                         if (key.objectid == bytenr &&
1531                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1532                             key.offset == num_bytes)
1533                                 ret = 0;
1534                 }
1535                 if (ret) {
1536                         key.objectid = bytenr;
1537                         key.type = BTRFS_EXTENT_ITEM_KEY;
1538                         key.offset = num_bytes;
1539                         btrfs_release_path(path);
1540                         goto again;
1541                 }
1542         }
1543
1544         if (ret && !insert) {
1545                 err = -ENOENT;
1546                 goto out;
1547         } else if (WARN_ON(ret)) {
1548                 err = -EIO;
1549                 goto out;
1550         }
1551
1552         leaf = path->nodes[0];
1553         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1554 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1555         if (item_size < sizeof(*ei)) {
1556                 if (!insert) {
1557                         err = -ENOENT;
1558                         goto out;
1559                 }
1560                 ret = convert_extent_item_v0(trans, root, path, owner,
1561                                              extra_size);
1562                 if (ret < 0) {
1563                         err = ret;
1564                         goto out;
1565                 }
1566                 leaf = path->nodes[0];
1567                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1568         }
1569 #endif
1570         BUG_ON(item_size < sizeof(*ei));
1571
1572         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1573         flags = btrfs_extent_flags(leaf, ei);
1574
1575         ptr = (unsigned long)(ei + 1);
1576         end = (unsigned long)ei + item_size;
1577
1578         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1579                 ptr += sizeof(struct btrfs_tree_block_info);
1580                 BUG_ON(ptr > end);
1581         }
1582
1583         err = -ENOENT;
1584         while (1) {
1585                 if (ptr >= end) {
1586                         WARN_ON(ptr > end);
1587                         break;
1588                 }
1589                 iref = (struct btrfs_extent_inline_ref *)ptr;
1590                 type = btrfs_extent_inline_ref_type(leaf, iref);
1591                 if (want < type)
1592                         break;
1593                 if (want > type) {
1594                         ptr += btrfs_extent_inline_ref_size(type);
1595                         continue;
1596                 }
1597
1598                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1599                         struct btrfs_extent_data_ref *dref;
1600                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1601                         if (match_extent_data_ref(leaf, dref, root_objectid,
1602                                                   owner, offset)) {
1603                                 err = 0;
1604                                 break;
1605                         }
1606                         if (hash_extent_data_ref_item(leaf, dref) <
1607                             hash_extent_data_ref(root_objectid, owner, offset))
1608                                 break;
1609                 } else {
1610                         u64 ref_offset;
1611                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1612                         if (parent > 0) {
1613                                 if (parent == ref_offset) {
1614                                         err = 0;
1615                                         break;
1616                                 }
1617                                 if (ref_offset < parent)
1618                                         break;
1619                         } else {
1620                                 if (root_objectid == ref_offset) {
1621                                         err = 0;
1622                                         break;
1623                                 }
1624                                 if (ref_offset < root_objectid)
1625                                         break;
1626                         }
1627                 }
1628                 ptr += btrfs_extent_inline_ref_size(type);
1629         }
1630         if (err == -ENOENT && insert) {
1631                 if (item_size + extra_size >=
1632                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1633                         err = -EAGAIN;
1634                         goto out;
1635                 }
1636                 /*
1637                  * To add new inline back ref, we have to make sure
1638                  * there is no corresponding back ref item.
1639                  * For simplicity, we just do not add new inline back
1640                  * ref if there is any kind of item for this block
1641                  */
1642                 if (find_next_key(path, 0, &key) == 0 &&
1643                     key.objectid == bytenr &&
1644                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1645                         err = -EAGAIN;
1646                         goto out;
1647                 }
1648         }
1649         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1650 out:
1651         if (insert) {
1652                 path->keep_locks = 0;
1653                 btrfs_unlock_up_safe(path, 1);
1654         }
1655         return err;
1656 }
1657
1658 /*
1659  * helper to add new inline back ref
1660  */
1661 static noinline_for_stack
1662 void setup_inline_extent_backref(struct btrfs_root *root,
1663                                  struct btrfs_path *path,
1664                                  struct btrfs_extent_inline_ref *iref,
1665                                  u64 parent, u64 root_objectid,
1666                                  u64 owner, u64 offset, int refs_to_add,
1667                                  struct btrfs_delayed_extent_op *extent_op)
1668 {
1669         struct extent_buffer *leaf;
1670         struct btrfs_extent_item *ei;
1671         unsigned long ptr;
1672         unsigned long end;
1673         unsigned long item_offset;
1674         u64 refs;
1675         int size;
1676         int type;
1677
1678         leaf = path->nodes[0];
1679         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1680         item_offset = (unsigned long)iref - (unsigned long)ei;
1681
1682         type = extent_ref_type(parent, owner);
1683         size = btrfs_extent_inline_ref_size(type);
1684
1685         btrfs_extend_item(root, path, size);
1686
1687         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1688         refs = btrfs_extent_refs(leaf, ei);
1689         refs += refs_to_add;
1690         btrfs_set_extent_refs(leaf, ei, refs);
1691         if (extent_op)
1692                 __run_delayed_extent_op(extent_op, leaf, ei);
1693
1694         ptr = (unsigned long)ei + item_offset;
1695         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1696         if (ptr < end - size)
1697                 memmove_extent_buffer(leaf, ptr + size, ptr,
1698                                       end - size - ptr);
1699
1700         iref = (struct btrfs_extent_inline_ref *)ptr;
1701         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1702         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1703                 struct btrfs_extent_data_ref *dref;
1704                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1705                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1706                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1707                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1708                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1709         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1710                 struct btrfs_shared_data_ref *sref;
1711                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1712                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1713                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1714         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1715                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1716         } else {
1717                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1718         }
1719         btrfs_mark_buffer_dirty(leaf);
1720 }
1721
1722 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1723                                  struct btrfs_root *root,
1724                                  struct btrfs_path *path,
1725                                  struct btrfs_extent_inline_ref **ref_ret,
1726                                  u64 bytenr, u64 num_bytes, u64 parent,
1727                                  u64 root_objectid, u64 owner, u64 offset)
1728 {
1729         int ret;
1730
1731         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1732                                            bytenr, num_bytes, parent,
1733                                            root_objectid, owner, offset, 0);
1734         if (ret != -ENOENT)
1735                 return ret;
1736
1737         btrfs_release_path(path);
1738         *ref_ret = NULL;
1739
1740         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1741                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1742                                             root_objectid);
1743         } else {
1744                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1745                                              root_objectid, owner, offset);
1746         }
1747         return ret;
1748 }
1749
1750 /*
1751  * helper to update/remove inline back ref
1752  */
1753 static noinline_for_stack
1754 void update_inline_extent_backref(struct btrfs_root *root,
1755                                   struct btrfs_path *path,
1756                                   struct btrfs_extent_inline_ref *iref,
1757                                   int refs_to_mod,
1758                                   struct btrfs_delayed_extent_op *extent_op,
1759                                   int *last_ref)
1760 {
1761         struct extent_buffer *leaf;
1762         struct btrfs_extent_item *ei;
1763         struct btrfs_extent_data_ref *dref = NULL;
1764         struct btrfs_shared_data_ref *sref = NULL;
1765         unsigned long ptr;
1766         unsigned long end;
1767         u32 item_size;
1768         int size;
1769         int type;
1770         u64 refs;
1771
1772         leaf = path->nodes[0];
1773         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1774         refs = btrfs_extent_refs(leaf, ei);
1775         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1776         refs += refs_to_mod;
1777         btrfs_set_extent_refs(leaf, ei, refs);
1778         if (extent_op)
1779                 __run_delayed_extent_op(extent_op, leaf, ei);
1780
1781         type = btrfs_extent_inline_ref_type(leaf, iref);
1782
1783         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1784                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1785                 refs = btrfs_extent_data_ref_count(leaf, dref);
1786         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1787                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1788                 refs = btrfs_shared_data_ref_count(leaf, sref);
1789         } else {
1790                 refs = 1;
1791                 BUG_ON(refs_to_mod != -1);
1792         }
1793
1794         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1795         refs += refs_to_mod;
1796
1797         if (refs > 0) {
1798                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1799                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1800                 else
1801                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1802         } else {
1803                 *last_ref = 1;
1804                 size =  btrfs_extent_inline_ref_size(type);
1805                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1806                 ptr = (unsigned long)iref;
1807                 end = (unsigned long)ei + item_size;
1808                 if (ptr + size < end)
1809                         memmove_extent_buffer(leaf, ptr, ptr + size,
1810                                               end - ptr - size);
1811                 item_size -= size;
1812                 btrfs_truncate_item(root, path, item_size, 1);
1813         }
1814         btrfs_mark_buffer_dirty(leaf);
1815 }
1816
1817 static noinline_for_stack
1818 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1819                                  struct btrfs_root *root,
1820                                  struct btrfs_path *path,
1821                                  u64 bytenr, u64 num_bytes, u64 parent,
1822                                  u64 root_objectid, u64 owner,
1823                                  u64 offset, int refs_to_add,
1824                                  struct btrfs_delayed_extent_op *extent_op)
1825 {
1826         struct btrfs_extent_inline_ref *iref;
1827         int ret;
1828
1829         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1830                                            bytenr, num_bytes, parent,
1831                                            root_objectid, owner, offset, 1);
1832         if (ret == 0) {
1833                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1834                 update_inline_extent_backref(root, path, iref,
1835                                              refs_to_add, extent_op, NULL);
1836         } else if (ret == -ENOENT) {
1837                 setup_inline_extent_backref(root, path, iref, parent,
1838                                             root_objectid, owner, offset,
1839                                             refs_to_add, extent_op);
1840                 ret = 0;
1841         }
1842         return ret;
1843 }
1844
1845 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1846                                  struct btrfs_root *root,
1847                                  struct btrfs_path *path,
1848                                  u64 bytenr, u64 parent, u64 root_objectid,
1849                                  u64 owner, u64 offset, int refs_to_add)
1850 {
1851         int ret;
1852         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1853                 BUG_ON(refs_to_add != 1);
1854                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1855                                             parent, root_objectid);
1856         } else {
1857                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1858                                              parent, root_objectid,
1859                                              owner, offset, refs_to_add);
1860         }
1861         return ret;
1862 }
1863
1864 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1865                                  struct btrfs_root *root,
1866                                  struct btrfs_path *path,
1867                                  struct btrfs_extent_inline_ref *iref,
1868                                  int refs_to_drop, int is_data, int *last_ref)
1869 {
1870         int ret = 0;
1871
1872         BUG_ON(!is_data && refs_to_drop != 1);
1873         if (iref) {
1874                 update_inline_extent_backref(root, path, iref,
1875                                              -refs_to_drop, NULL, last_ref);
1876         } else if (is_data) {
1877                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop,
1878                                              last_ref);
1879         } else {
1880                 *last_ref = 1;
1881                 ret = btrfs_del_item(trans, root, path);
1882         }
1883         return ret;
1884 }
1885
1886 static int btrfs_issue_discard(struct block_device *bdev,
1887                                 u64 start, u64 len)
1888 {
1889         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1890 }
1891
1892 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1893                                 u64 num_bytes, u64 *actual_bytes)
1894 {
1895         int ret;
1896         u64 discarded_bytes = 0;
1897         struct btrfs_bio *bbio = NULL;
1898
1899
1900         /* Tell the block device(s) that the sectors can be discarded */
1901         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1902                               bytenr, &num_bytes, &bbio, 0);
1903         /* Error condition is -ENOMEM */
1904         if (!ret) {
1905                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1906                 int i;
1907
1908
1909                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1910                         if (!stripe->dev->can_discard)
1911                                 continue;
1912
1913                         ret = btrfs_issue_discard(stripe->dev->bdev,
1914                                                   stripe->physical,
1915                                                   stripe->length);
1916                         if (!ret)
1917                                 discarded_bytes += stripe->length;
1918                         else if (ret != -EOPNOTSUPP)
1919                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1920
1921                         /*
1922                          * Just in case we get back EOPNOTSUPP for some reason,
1923                          * just ignore the return value so we don't screw up
1924                          * people calling discard_extent.
1925                          */
1926                         ret = 0;
1927                 }
1928                 kfree(bbio);
1929         }
1930
1931         if (actual_bytes)
1932                 *actual_bytes = discarded_bytes;
1933
1934
1935         if (ret == -EOPNOTSUPP)
1936                 ret = 0;
1937         return ret;
1938 }
1939
1940 /* Can return -ENOMEM */
1941 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1942                          struct btrfs_root *root,
1943                          u64 bytenr, u64 num_bytes, u64 parent,
1944                          u64 root_objectid, u64 owner, u64 offset,
1945                          int no_quota)
1946 {
1947         int ret;
1948         struct btrfs_fs_info *fs_info = root->fs_info;
1949
1950         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1951                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1952
1953         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1954                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1955                                         num_bytes,
1956                                         parent, root_objectid, (int)owner,
1957                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
1958         } else {
1959                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1960                                         num_bytes,
1961                                         parent, root_objectid, owner, offset,
1962                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
1963         }
1964         return ret;
1965 }
1966
1967 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1968                                   struct btrfs_root *root,
1969                                   u64 bytenr, u64 num_bytes,
1970                                   u64 parent, u64 root_objectid,
1971                                   u64 owner, u64 offset, int refs_to_add,
1972                                   int no_quota,
1973                                   struct btrfs_delayed_extent_op *extent_op)
1974 {
1975         struct btrfs_fs_info *fs_info = root->fs_info;
1976         struct btrfs_path *path;
1977         struct extent_buffer *leaf;
1978         struct btrfs_extent_item *item;
1979         struct btrfs_key key;
1980         u64 refs;
1981         int ret;
1982         enum btrfs_qgroup_operation_type type = BTRFS_QGROUP_OPER_ADD_EXCL;
1983
1984         path = btrfs_alloc_path();
1985         if (!path)
1986                 return -ENOMEM;
1987
1988         if (!is_fstree(root_objectid) || !root->fs_info->quota_enabled)
1989                 no_quota = 1;
1990
1991         path->reada = 1;
1992         path->leave_spinning = 1;
1993         /* this will setup the path even if it fails to insert the back ref */
1994         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
1995                                            bytenr, num_bytes, parent,
1996                                            root_objectid, owner, offset,
1997                                            refs_to_add, extent_op);
1998         if ((ret < 0 && ret != -EAGAIN) || (!ret && no_quota))
1999                 goto out;
2000         /*
2001          * Ok we were able to insert an inline extent and it appears to be a new
2002          * reference, deal with the qgroup accounting.
2003          */
2004         if (!ret && !no_quota) {
2005                 ASSERT(root->fs_info->quota_enabled);
2006                 leaf = path->nodes[0];
2007                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2008                 item = btrfs_item_ptr(leaf, path->slots[0],
2009                                       struct btrfs_extent_item);
2010                 if (btrfs_extent_refs(leaf, item) > (u64)refs_to_add)
2011                         type = BTRFS_QGROUP_OPER_ADD_SHARED;
2012                 btrfs_release_path(path);
2013
2014                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
2015                                               bytenr, num_bytes, type, 0);
2016                 goto out;
2017         }
2018
2019         /*
2020          * Ok we had -EAGAIN which means we didn't have space to insert and
2021          * inline extent ref, so just update the reference count and add a
2022          * normal backref.
2023          */
2024         leaf = path->nodes[0];
2025         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2026         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2027         refs = btrfs_extent_refs(leaf, item);
2028         if (refs)
2029                 type = BTRFS_QGROUP_OPER_ADD_SHARED;
2030         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2031         if (extent_op)
2032                 __run_delayed_extent_op(extent_op, leaf, item);
2033
2034         btrfs_mark_buffer_dirty(leaf);
2035         btrfs_release_path(path);
2036
2037         if (!no_quota) {
2038                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
2039                                               bytenr, num_bytes, type, 0);
2040                 if (ret)
2041                         goto out;
2042         }
2043
2044         path->reada = 1;
2045         path->leave_spinning = 1;
2046         /* now insert the actual backref */
2047         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2048                                     path, bytenr, parent, root_objectid,
2049                                     owner, offset, refs_to_add);
2050         if (ret)
2051                 btrfs_abort_transaction(trans, root, ret);
2052 out:
2053         btrfs_free_path(path);
2054         return ret;
2055 }
2056
2057 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2058                                 struct btrfs_root *root,
2059                                 struct btrfs_delayed_ref_node *node,
2060                                 struct btrfs_delayed_extent_op *extent_op,
2061                                 int insert_reserved)
2062 {
2063         int ret = 0;
2064         struct btrfs_delayed_data_ref *ref;
2065         struct btrfs_key ins;
2066         u64 parent = 0;
2067         u64 ref_root = 0;
2068         u64 flags = 0;
2069
2070         ins.objectid = node->bytenr;
2071         ins.offset = node->num_bytes;
2072         ins.type = BTRFS_EXTENT_ITEM_KEY;
2073
2074         ref = btrfs_delayed_node_to_data_ref(node);
2075         trace_run_delayed_data_ref(node, ref, node->action);
2076
2077         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2078                 parent = ref->parent;
2079         ref_root = ref->root;
2080
2081         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2082                 if (extent_op)
2083                         flags |= extent_op->flags_to_set;
2084                 ret = alloc_reserved_file_extent(trans, root,
2085                                                  parent, ref_root, flags,
2086                                                  ref->objectid, ref->offset,
2087                                                  &ins, node->ref_mod);
2088         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2089                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2090                                              node->num_bytes, parent,
2091                                              ref_root, ref->objectid,
2092                                              ref->offset, node->ref_mod,
2093                                              node->no_quota, extent_op);
2094         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2095                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2096                                           node->num_bytes, parent,
2097                                           ref_root, ref->objectid,
2098                                           ref->offset, node->ref_mod,
2099                                           extent_op, node->no_quota);
2100         } else {
2101                 BUG();
2102         }
2103         return ret;
2104 }
2105
2106 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2107                                     struct extent_buffer *leaf,
2108                                     struct btrfs_extent_item *ei)
2109 {
2110         u64 flags = btrfs_extent_flags(leaf, ei);
2111         if (extent_op->update_flags) {
2112                 flags |= extent_op->flags_to_set;
2113                 btrfs_set_extent_flags(leaf, ei, flags);
2114         }
2115
2116         if (extent_op->update_key) {
2117                 struct btrfs_tree_block_info *bi;
2118                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2119                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2120                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2121         }
2122 }
2123
2124 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2125                                  struct btrfs_root *root,
2126                                  struct btrfs_delayed_ref_node *node,
2127                                  struct btrfs_delayed_extent_op *extent_op)
2128 {
2129         struct btrfs_key key;
2130         struct btrfs_path *path;
2131         struct btrfs_extent_item *ei;
2132         struct extent_buffer *leaf;
2133         u32 item_size;
2134         int ret;
2135         int err = 0;
2136         int metadata = !extent_op->is_data;
2137
2138         if (trans->aborted)
2139                 return 0;
2140
2141         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2142                 metadata = 0;
2143
2144         path = btrfs_alloc_path();
2145         if (!path)
2146                 return -ENOMEM;
2147
2148         key.objectid = node->bytenr;
2149
2150         if (metadata) {
2151                 key.type = BTRFS_METADATA_ITEM_KEY;
2152                 key.offset = extent_op->level;
2153         } else {
2154                 key.type = BTRFS_EXTENT_ITEM_KEY;
2155                 key.offset = node->num_bytes;
2156         }
2157
2158 again:
2159         path->reada = 1;
2160         path->leave_spinning = 1;
2161         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2162                                 path, 0, 1);
2163         if (ret < 0) {
2164                 err = ret;
2165                 goto out;
2166         }
2167         if (ret > 0) {
2168                 if (metadata) {
2169                         if (path->slots[0] > 0) {
2170                                 path->slots[0]--;
2171                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2172                                                       path->slots[0]);
2173                                 if (key.objectid == node->bytenr &&
2174                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2175                                     key.offset == node->num_bytes)
2176                                         ret = 0;
2177                         }
2178                         if (ret > 0) {
2179                                 btrfs_release_path(path);
2180                                 metadata = 0;
2181
2182                                 key.objectid = node->bytenr;
2183                                 key.offset = node->num_bytes;
2184                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2185                                 goto again;
2186                         }
2187                 } else {
2188                         err = -EIO;
2189                         goto out;
2190                 }
2191         }
2192
2193         leaf = path->nodes[0];
2194         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2195 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2196         if (item_size < sizeof(*ei)) {
2197                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2198                                              path, (u64)-1, 0);
2199                 if (ret < 0) {
2200                         err = ret;
2201                         goto out;
2202                 }
2203                 leaf = path->nodes[0];
2204                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2205         }
2206 #endif
2207         BUG_ON(item_size < sizeof(*ei));
2208         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2209         __run_delayed_extent_op(extent_op, leaf, ei);
2210
2211         btrfs_mark_buffer_dirty(leaf);
2212 out:
2213         btrfs_free_path(path);
2214         return err;
2215 }
2216
2217 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2218                                 struct btrfs_root *root,
2219                                 struct btrfs_delayed_ref_node *node,
2220                                 struct btrfs_delayed_extent_op *extent_op,
2221                                 int insert_reserved)
2222 {
2223         int ret = 0;
2224         struct btrfs_delayed_tree_ref *ref;
2225         struct btrfs_key ins;
2226         u64 parent = 0;
2227         u64 ref_root = 0;
2228         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2229                                                  SKINNY_METADATA);
2230
2231         ref = btrfs_delayed_node_to_tree_ref(node);
2232         trace_run_delayed_tree_ref(node, ref, node->action);
2233
2234         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2235                 parent = ref->parent;
2236         ref_root = ref->root;
2237
2238         ins.objectid = node->bytenr;
2239         if (skinny_metadata) {
2240                 ins.offset = ref->level;
2241                 ins.type = BTRFS_METADATA_ITEM_KEY;
2242         } else {
2243                 ins.offset = node->num_bytes;
2244                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2245         }
2246
2247         BUG_ON(node->ref_mod != 1);
2248         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2249                 BUG_ON(!extent_op || !extent_op->update_flags);
2250                 ret = alloc_reserved_tree_block(trans, root,
2251                                                 parent, ref_root,
2252                                                 extent_op->flags_to_set,
2253                                                 &extent_op->key,
2254                                                 ref->level, &ins,
2255                                                 node->no_quota);
2256         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2257                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2258                                              node->num_bytes, parent, ref_root,
2259                                              ref->level, 0, 1, node->no_quota,
2260                                              extent_op);
2261         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2262                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2263                                           node->num_bytes, parent, ref_root,
2264                                           ref->level, 0, 1, extent_op,
2265                                           node->no_quota);
2266         } else {
2267                 BUG();
2268         }
2269         return ret;
2270 }
2271
2272 /* helper function to actually process a single delayed ref entry */
2273 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2274                                struct btrfs_root *root,
2275                                struct btrfs_delayed_ref_node *node,
2276                                struct btrfs_delayed_extent_op *extent_op,
2277                                int insert_reserved)
2278 {
2279         int ret = 0;
2280
2281         if (trans->aborted) {
2282                 if (insert_reserved)
2283                         btrfs_pin_extent(root, node->bytenr,
2284                                          node->num_bytes, 1);
2285                 return 0;
2286         }
2287
2288         if (btrfs_delayed_ref_is_head(node)) {
2289                 struct btrfs_delayed_ref_head *head;
2290                 /*
2291                  * we've hit the end of the chain and we were supposed
2292                  * to insert this extent into the tree.  But, it got
2293                  * deleted before we ever needed to insert it, so all
2294                  * we have to do is clean up the accounting
2295                  */
2296                 BUG_ON(extent_op);
2297                 head = btrfs_delayed_node_to_head(node);
2298                 trace_run_delayed_ref_head(node, head, node->action);
2299
2300                 if (insert_reserved) {
2301                         btrfs_pin_extent(root, node->bytenr,
2302                                          node->num_bytes, 1);
2303                         if (head->is_data) {
2304                                 ret = btrfs_del_csums(trans, root,
2305                                                       node->bytenr,
2306                                                       node->num_bytes);
2307                         }
2308                 }
2309                 return ret;
2310         }
2311
2312         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2313             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2314                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2315                                            insert_reserved);
2316         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2317                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2318                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2319                                            insert_reserved);
2320         else
2321                 BUG();
2322         return ret;
2323 }
2324
2325 static noinline struct btrfs_delayed_ref_node *
2326 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2327 {
2328         struct rb_node *node;
2329         struct btrfs_delayed_ref_node *ref, *last = NULL;;
2330
2331         /*
2332          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2333          * this prevents ref count from going down to zero when
2334          * there still are pending delayed ref.
2335          */
2336         node = rb_first(&head->ref_root);
2337         while (node) {
2338                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2339                                 rb_node);
2340                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2341                         return ref;
2342                 else if (last == NULL)
2343                         last = ref;
2344                 node = rb_next(node);
2345         }
2346         return last;
2347 }
2348
2349 /*
2350  * Returns 0 on success or if called with an already aborted transaction.
2351  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2352  */
2353 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2354                                              struct btrfs_root *root,
2355                                              unsigned long nr)
2356 {
2357         struct btrfs_delayed_ref_root *delayed_refs;
2358         struct btrfs_delayed_ref_node *ref;
2359         struct btrfs_delayed_ref_head *locked_ref = NULL;
2360         struct btrfs_delayed_extent_op *extent_op;
2361         struct btrfs_fs_info *fs_info = root->fs_info;
2362         ktime_t start = ktime_get();
2363         int ret;
2364         unsigned long count = 0;
2365         unsigned long actual_count = 0;
2366         int must_insert_reserved = 0;
2367
2368         delayed_refs = &trans->transaction->delayed_refs;
2369         while (1) {
2370                 if (!locked_ref) {
2371                         if (count >= nr)
2372                                 break;
2373
2374                         spin_lock(&delayed_refs->lock);
2375                         locked_ref = btrfs_select_ref_head(trans);
2376                         if (!locked_ref) {
2377                                 spin_unlock(&delayed_refs->lock);
2378                                 break;
2379                         }
2380
2381                         /* grab the lock that says we are going to process
2382                          * all the refs for this head */
2383                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2384                         spin_unlock(&delayed_refs->lock);
2385                         /*
2386                          * we may have dropped the spin lock to get the head
2387                          * mutex lock, and that might have given someone else
2388                          * time to free the head.  If that's true, it has been
2389                          * removed from our list and we can move on.
2390                          */
2391                         if (ret == -EAGAIN) {
2392                                 locked_ref = NULL;
2393                                 count++;
2394                                 continue;
2395                         }
2396                 }
2397
2398                 /*
2399                  * We need to try and merge add/drops of the same ref since we
2400                  * can run into issues with relocate dropping the implicit ref
2401                  * and then it being added back again before the drop can
2402                  * finish.  If we merged anything we need to re-loop so we can
2403                  * get a good ref.
2404                  */
2405                 spin_lock(&locked_ref->lock);
2406                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2407                                          locked_ref);
2408
2409                 /*
2410                  * locked_ref is the head node, so we have to go one
2411                  * node back for any delayed ref updates
2412                  */
2413                 ref = select_delayed_ref(locked_ref);
2414
2415                 if (ref && ref->seq &&
2416                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2417                         spin_unlock(&locked_ref->lock);
2418                         btrfs_delayed_ref_unlock(locked_ref);
2419                         spin_lock(&delayed_refs->lock);
2420                         locked_ref->processing = 0;
2421                         delayed_refs->num_heads_ready++;
2422                         spin_unlock(&delayed_refs->lock);
2423                         locked_ref = NULL;
2424                         cond_resched();
2425                         count++;
2426                         continue;
2427                 }
2428
2429                 /*
2430                  * record the must insert reserved flag before we
2431                  * drop the spin lock.
2432                  */
2433                 must_insert_reserved = locked_ref->must_insert_reserved;
2434                 locked_ref->must_insert_reserved = 0;
2435
2436                 extent_op = locked_ref->extent_op;
2437                 locked_ref->extent_op = NULL;
2438
2439                 if (!ref) {
2440
2441
2442                         /* All delayed refs have been processed, Go ahead
2443                          * and send the head node to run_one_delayed_ref,
2444                          * so that any accounting fixes can happen
2445                          */
2446                         ref = &locked_ref->node;
2447
2448                         if (extent_op && must_insert_reserved) {
2449                                 btrfs_free_delayed_extent_op(extent_op);
2450                                 extent_op = NULL;
2451                         }
2452
2453                         if (extent_op) {
2454                                 spin_unlock(&locked_ref->lock);
2455                                 ret = run_delayed_extent_op(trans, root,
2456                                                             ref, extent_op);
2457                                 btrfs_free_delayed_extent_op(extent_op);
2458
2459                                 if (ret) {
2460                                         /*
2461                                          * Need to reset must_insert_reserved if
2462                                          * there was an error so the abort stuff
2463                                          * can cleanup the reserved space
2464                                          * properly.
2465                                          */
2466                                         if (must_insert_reserved)
2467                                                 locked_ref->must_insert_reserved = 1;
2468                                         locked_ref->processing = 0;
2469                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2470                                         btrfs_delayed_ref_unlock(locked_ref);
2471                                         return ret;
2472                                 }
2473                                 continue;
2474                         }
2475
2476                         /*
2477                          * Need to drop our head ref lock and re-aqcuire the
2478                          * delayed ref lock and then re-check to make sure
2479                          * nobody got added.
2480                          */
2481                         spin_unlock(&locked_ref->lock);
2482                         spin_lock(&delayed_refs->lock);
2483                         spin_lock(&locked_ref->lock);
2484                         if (rb_first(&locked_ref->ref_root) ||
2485                             locked_ref->extent_op) {
2486                                 spin_unlock(&locked_ref->lock);
2487                                 spin_unlock(&delayed_refs->lock);
2488                                 continue;
2489                         }
2490                         ref->in_tree = 0;
2491                         delayed_refs->num_heads--;
2492                         rb_erase(&locked_ref->href_node,
2493                                  &delayed_refs->href_root);
2494                         spin_unlock(&delayed_refs->lock);
2495                 } else {
2496                         actual_count++;
2497                         ref->in_tree = 0;
2498                         rb_erase(&ref->rb_node, &locked_ref->ref_root);
2499                 }
2500                 atomic_dec(&delayed_refs->num_entries);
2501
2502                 if (!btrfs_delayed_ref_is_head(ref)) {
2503                         /*
2504                          * when we play the delayed ref, also correct the
2505                          * ref_mod on head
2506                          */
2507                         switch (ref->action) {
2508                         case BTRFS_ADD_DELAYED_REF:
2509                         case BTRFS_ADD_DELAYED_EXTENT:
2510                                 locked_ref->node.ref_mod -= ref->ref_mod;
2511                                 break;
2512                         case BTRFS_DROP_DELAYED_REF:
2513                                 locked_ref->node.ref_mod += ref->ref_mod;
2514                                 break;
2515                         default:
2516                                 WARN_ON(1);
2517                         }
2518                 }
2519                 spin_unlock(&locked_ref->lock);
2520
2521                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2522                                           must_insert_reserved);
2523
2524                 btrfs_free_delayed_extent_op(extent_op);
2525                 if (ret) {
2526                         locked_ref->processing = 0;
2527                         btrfs_delayed_ref_unlock(locked_ref);
2528                         btrfs_put_delayed_ref(ref);
2529                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2530                         return ret;
2531                 }
2532
2533                 /*
2534                  * If this node is a head, that means all the refs in this head
2535                  * have been dealt with, and we will pick the next head to deal
2536                  * with, so we must unlock the head and drop it from the cluster
2537                  * list before we release it.
2538                  */
2539                 if (btrfs_delayed_ref_is_head(ref)) {
2540                         btrfs_delayed_ref_unlock(locked_ref);
2541                         locked_ref = NULL;
2542                 }
2543                 btrfs_put_delayed_ref(ref);
2544                 count++;
2545                 cond_resched();
2546         }
2547
2548         /*
2549          * We don't want to include ref heads since we can have empty ref heads
2550          * and those will drastically skew our runtime down since we just do
2551          * accounting, no actual extent tree updates.
2552          */
2553         if (actual_count > 0) {
2554                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2555                 u64 avg;
2556
2557                 /*
2558                  * We weigh the current average higher than our current runtime
2559                  * to avoid large swings in the average.
2560                  */
2561                 spin_lock(&delayed_refs->lock);
2562                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2563                 avg = div64_u64(avg, 4);
2564                 fs_info->avg_delayed_ref_runtime = avg;
2565                 spin_unlock(&delayed_refs->lock);
2566         }
2567         return 0;
2568 }
2569
2570 #ifdef SCRAMBLE_DELAYED_REFS
2571 /*
2572  * Normally delayed refs get processed in ascending bytenr order. This
2573  * correlates in most cases to the order added. To expose dependencies on this
2574  * order, we start to process the tree in the middle instead of the beginning
2575  */
2576 static u64 find_middle(struct rb_root *root)
2577 {
2578         struct rb_node *n = root->rb_node;
2579         struct btrfs_delayed_ref_node *entry;
2580         int alt = 1;
2581         u64 middle;
2582         u64 first = 0, last = 0;
2583
2584         n = rb_first(root);
2585         if (n) {
2586                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2587                 first = entry->bytenr;
2588         }
2589         n = rb_last(root);
2590         if (n) {
2591                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2592                 last = entry->bytenr;
2593         }
2594         n = root->rb_node;
2595
2596         while (n) {
2597                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2598                 WARN_ON(!entry->in_tree);
2599
2600                 middle = entry->bytenr;
2601
2602                 if (alt)
2603                         n = n->rb_left;
2604                 else
2605                         n = n->rb_right;
2606
2607                 alt = 1 - alt;
2608         }
2609         return middle;
2610 }
2611 #endif
2612
2613 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2614 {
2615         u64 num_bytes;
2616
2617         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2618                              sizeof(struct btrfs_extent_inline_ref));
2619         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2620                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2621
2622         /*
2623          * We don't ever fill up leaves all the way so multiply by 2 just to be
2624          * closer to what we're really going to want to ouse.
2625          */
2626         return div64_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2627 }
2628
2629 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2630                                        struct btrfs_root *root)
2631 {
2632         struct btrfs_block_rsv *global_rsv;
2633         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2634         u64 num_bytes;
2635         int ret = 0;
2636
2637         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2638         num_heads = heads_to_leaves(root, num_heads);
2639         if (num_heads > 1)
2640                 num_bytes += (num_heads - 1) * root->nodesize;
2641         num_bytes <<= 1;
2642         global_rsv = &root->fs_info->global_block_rsv;
2643
2644         /*
2645          * If we can't allocate any more chunks lets make sure we have _lots_ of
2646          * wiggle room since running delayed refs can create more delayed refs.
2647          */
2648         if (global_rsv->space_info->full)
2649                 num_bytes <<= 1;
2650
2651         spin_lock(&global_rsv->lock);
2652         if (global_rsv->reserved <= num_bytes)
2653                 ret = 1;
2654         spin_unlock(&global_rsv->lock);
2655         return ret;
2656 }
2657
2658 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2659                                        struct btrfs_root *root)
2660 {
2661         struct btrfs_fs_info *fs_info = root->fs_info;
2662         u64 num_entries =
2663                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2664         u64 avg_runtime;
2665         u64 val;
2666
2667         smp_mb();
2668         avg_runtime = fs_info->avg_delayed_ref_runtime;
2669         val = num_entries * avg_runtime;
2670         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2671                 return 1;
2672         if (val >= NSEC_PER_SEC / 2)
2673                 return 2;
2674
2675         return btrfs_check_space_for_delayed_refs(trans, root);
2676 }
2677
2678 struct async_delayed_refs {
2679         struct btrfs_root *root;
2680         int count;
2681         int error;
2682         int sync;
2683         struct completion wait;
2684         struct btrfs_work work;
2685 };
2686
2687 static void delayed_ref_async_start(struct btrfs_work *work)
2688 {
2689         struct async_delayed_refs *async;
2690         struct btrfs_trans_handle *trans;
2691         int ret;
2692
2693         async = container_of(work, struct async_delayed_refs, work);
2694
2695         trans = btrfs_join_transaction(async->root);
2696         if (IS_ERR(trans)) {
2697                 async->error = PTR_ERR(trans);
2698                 goto done;
2699         }
2700
2701         /*
2702          * trans->sync means that when we call end_transaciton, we won't
2703          * wait on delayed refs
2704          */
2705         trans->sync = true;
2706         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2707         if (ret)
2708                 async->error = ret;
2709
2710         ret = btrfs_end_transaction(trans, async->root);
2711         if (ret && !async->error)
2712                 async->error = ret;
2713 done:
2714         if (async->sync)
2715                 complete(&async->wait);
2716         else
2717                 kfree(async);
2718 }
2719
2720 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2721                                  unsigned long count, int wait)
2722 {
2723         struct async_delayed_refs *async;
2724         int ret;
2725
2726         async = kmalloc(sizeof(*async), GFP_NOFS);
2727         if (!async)
2728                 return -ENOMEM;
2729
2730         async->root = root->fs_info->tree_root;
2731         async->count = count;
2732         async->error = 0;
2733         if (wait)
2734                 async->sync = 1;
2735         else
2736                 async->sync = 0;
2737         init_completion(&async->wait);
2738
2739         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2740                         delayed_ref_async_start, NULL, NULL);
2741
2742         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2743
2744         if (wait) {
2745                 wait_for_completion(&async->wait);
2746                 ret = async->error;
2747                 kfree(async);
2748                 return ret;
2749         }
2750         return 0;
2751 }
2752
2753 /*
2754  * this starts processing the delayed reference count updates and
2755  * extent insertions we have queued up so far.  count can be
2756  * 0, which means to process everything in the tree at the start
2757  * of the run (but not newly added entries), or it can be some target
2758  * number you'd like to process.
2759  *
2760  * Returns 0 on success or if called with an aborted transaction
2761  * Returns <0 on error and aborts the transaction
2762  */
2763 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2764                            struct btrfs_root *root, unsigned long count)
2765 {
2766         struct rb_node *node;
2767         struct btrfs_delayed_ref_root *delayed_refs;
2768         struct btrfs_delayed_ref_head *head;
2769         int ret;
2770         int run_all = count == (unsigned long)-1;
2771         int run_most = 0;
2772
2773         /* We'll clean this up in btrfs_cleanup_transaction */
2774         if (trans->aborted)
2775                 return 0;
2776
2777         if (root == root->fs_info->extent_root)
2778                 root = root->fs_info->tree_root;
2779
2780         delayed_refs = &trans->transaction->delayed_refs;
2781         if (count == 0) {
2782                 count = atomic_read(&delayed_refs->num_entries) * 2;
2783                 run_most = 1;
2784         }
2785
2786 again:
2787 #ifdef SCRAMBLE_DELAYED_REFS
2788         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2789 #endif
2790         ret = __btrfs_run_delayed_refs(trans, root, count);
2791         if (ret < 0) {
2792                 btrfs_abort_transaction(trans, root, ret);
2793                 return ret;
2794         }
2795
2796         if (run_all) {
2797                 if (!list_empty(&trans->new_bgs))
2798                         btrfs_create_pending_block_groups(trans, root);
2799
2800                 spin_lock(&delayed_refs->lock);
2801                 node = rb_first(&delayed_refs->href_root);
2802                 if (!node) {
2803                         spin_unlock(&delayed_refs->lock);
2804                         goto out;
2805                 }
2806                 count = (unsigned long)-1;
2807
2808                 while (node) {
2809                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2810                                         href_node);
2811                         if (btrfs_delayed_ref_is_head(&head->node)) {
2812                                 struct btrfs_delayed_ref_node *ref;
2813
2814                                 ref = &head->node;
2815                                 atomic_inc(&ref->refs);
2816
2817                                 spin_unlock(&delayed_refs->lock);
2818                                 /*
2819                                  * Mutex was contended, block until it's
2820                                  * released and try again
2821                                  */
2822                                 mutex_lock(&head->mutex);
2823                                 mutex_unlock(&head->mutex);
2824
2825                                 btrfs_put_delayed_ref(ref);
2826                                 cond_resched();
2827                                 goto again;
2828                         } else {
2829                                 WARN_ON(1);
2830                         }
2831                         node = rb_next(node);
2832                 }
2833                 spin_unlock(&delayed_refs->lock);
2834                 cond_resched();
2835                 goto again;
2836         }
2837 out:
2838         ret = btrfs_delayed_qgroup_accounting(trans, root->fs_info);
2839         if (ret)
2840                 return ret;
2841         assert_qgroups_uptodate(trans);
2842         return 0;
2843 }
2844
2845 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2846                                 struct btrfs_root *root,
2847                                 u64 bytenr, u64 num_bytes, u64 flags,
2848                                 int level, int is_data)
2849 {
2850         struct btrfs_delayed_extent_op *extent_op;
2851         int ret;
2852
2853         extent_op = btrfs_alloc_delayed_extent_op();
2854         if (!extent_op)
2855                 return -ENOMEM;
2856
2857         extent_op->flags_to_set = flags;
2858         extent_op->update_flags = 1;
2859         extent_op->update_key = 0;
2860         extent_op->is_data = is_data ? 1 : 0;
2861         extent_op->level = level;
2862
2863         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2864                                           num_bytes, extent_op);
2865         if (ret)
2866                 btrfs_free_delayed_extent_op(extent_op);
2867         return ret;
2868 }
2869
2870 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2871                                       struct btrfs_root *root,
2872                                       struct btrfs_path *path,
2873                                       u64 objectid, u64 offset, u64 bytenr)
2874 {
2875         struct btrfs_delayed_ref_head *head;
2876         struct btrfs_delayed_ref_node *ref;
2877         struct btrfs_delayed_data_ref *data_ref;
2878         struct btrfs_delayed_ref_root *delayed_refs;
2879         struct rb_node *node;
2880         int ret = 0;
2881
2882         delayed_refs = &trans->transaction->delayed_refs;
2883         spin_lock(&delayed_refs->lock);
2884         head = btrfs_find_delayed_ref_head(trans, bytenr);
2885         if (!head) {
2886                 spin_unlock(&delayed_refs->lock);
2887                 return 0;
2888         }
2889
2890         if (!mutex_trylock(&head->mutex)) {
2891                 atomic_inc(&head->node.refs);
2892                 spin_unlock(&delayed_refs->lock);
2893
2894                 btrfs_release_path(path);
2895
2896                 /*
2897                  * Mutex was contended, block until it's released and let
2898                  * caller try again
2899                  */
2900                 mutex_lock(&head->mutex);
2901                 mutex_unlock(&head->mutex);
2902                 btrfs_put_delayed_ref(&head->node);
2903                 return -EAGAIN;
2904         }
2905         spin_unlock(&delayed_refs->lock);
2906
2907         spin_lock(&head->lock);
2908         node = rb_first(&head->ref_root);
2909         while (node) {
2910                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2911                 node = rb_next(node);
2912
2913                 /* If it's a shared ref we know a cross reference exists */
2914                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2915                         ret = 1;
2916                         break;
2917                 }
2918
2919                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2920
2921                 /*
2922                  * If our ref doesn't match the one we're currently looking at
2923                  * then we have a cross reference.
2924                  */
2925                 if (data_ref->root != root->root_key.objectid ||
2926                     data_ref->objectid != objectid ||
2927                     data_ref->offset != offset) {
2928                         ret = 1;
2929                         break;
2930                 }
2931         }
2932         spin_unlock(&head->lock);
2933         mutex_unlock(&head->mutex);
2934         return ret;
2935 }
2936
2937 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2938                                         struct btrfs_root *root,
2939                                         struct btrfs_path *path,
2940                                         u64 objectid, u64 offset, u64 bytenr)
2941 {
2942         struct btrfs_root *extent_root = root->fs_info->extent_root;
2943         struct extent_buffer *leaf;
2944         struct btrfs_extent_data_ref *ref;
2945         struct btrfs_extent_inline_ref *iref;
2946         struct btrfs_extent_item *ei;
2947         struct btrfs_key key;
2948         u32 item_size;
2949         int ret;
2950
2951         key.objectid = bytenr;
2952         key.offset = (u64)-1;
2953         key.type = BTRFS_EXTENT_ITEM_KEY;
2954
2955         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2956         if (ret < 0)
2957                 goto out;
2958         BUG_ON(ret == 0); /* Corruption */
2959
2960         ret = -ENOENT;
2961         if (path->slots[0] == 0)
2962                 goto out;
2963
2964         path->slots[0]--;
2965         leaf = path->nodes[0];
2966         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2967
2968         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2969                 goto out;
2970
2971         ret = 1;
2972         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2973 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2974         if (item_size < sizeof(*ei)) {
2975                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2976                 goto out;
2977         }
2978 #endif
2979         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2980
2981         if (item_size != sizeof(*ei) +
2982             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2983                 goto out;
2984
2985         if (btrfs_extent_generation(leaf, ei) <=
2986             btrfs_root_last_snapshot(&root->root_item))
2987                 goto out;
2988
2989         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2990         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2991             BTRFS_EXTENT_DATA_REF_KEY)
2992                 goto out;
2993
2994         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2995         if (btrfs_extent_refs(leaf, ei) !=
2996             btrfs_extent_data_ref_count(leaf, ref) ||
2997             btrfs_extent_data_ref_root(leaf, ref) !=
2998             root->root_key.objectid ||
2999             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3000             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3001                 goto out;
3002
3003         ret = 0;
3004 out:
3005         return ret;
3006 }
3007
3008 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3009                           struct btrfs_root *root,
3010                           u64 objectid, u64 offset, u64 bytenr)
3011 {
3012         struct btrfs_path *path;
3013         int ret;
3014         int ret2;
3015
3016         path = btrfs_alloc_path();
3017         if (!path)
3018                 return -ENOENT;
3019
3020         do {
3021                 ret = check_committed_ref(trans, root, path, objectid,
3022                                           offset, bytenr);
3023                 if (ret && ret != -ENOENT)
3024                         goto out;
3025
3026                 ret2 = check_delayed_ref(trans, root, path, objectid,
3027                                          offset, bytenr);
3028         } while (ret2 == -EAGAIN);
3029
3030         if (ret2 && ret2 != -ENOENT) {
3031                 ret = ret2;
3032                 goto out;
3033         }
3034
3035         if (ret != -ENOENT || ret2 != -ENOENT)
3036                 ret = 0;
3037 out:
3038         btrfs_free_path(path);
3039         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3040                 WARN_ON(ret > 0);
3041         return ret;
3042 }
3043
3044 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3045                            struct btrfs_root *root,
3046                            struct extent_buffer *buf,
3047                            int full_backref, int inc)
3048 {
3049         u64 bytenr;
3050         u64 num_bytes;
3051         u64 parent;
3052         u64 ref_root;
3053         u32 nritems;
3054         struct btrfs_key key;
3055         struct btrfs_file_extent_item *fi;
3056         int i;
3057         int level;
3058         int ret = 0;
3059         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3060                             u64, u64, u64, u64, u64, u64, int);
3061
3062
3063         if (btrfs_test_is_dummy_root(root))
3064                 return 0;
3065
3066         ref_root = btrfs_header_owner(buf);
3067         nritems = btrfs_header_nritems(buf);
3068         level = btrfs_header_level(buf);
3069
3070         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3071                 return 0;
3072
3073         if (inc)
3074                 process_func = btrfs_inc_extent_ref;
3075         else
3076                 process_func = btrfs_free_extent;
3077
3078         if (full_backref)
3079                 parent = buf->start;
3080         else
3081                 parent = 0;
3082
3083         for (i = 0; i < nritems; i++) {
3084                 if (level == 0) {
3085                         btrfs_item_key_to_cpu(buf, &key, i);
3086                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3087                                 continue;
3088                         fi = btrfs_item_ptr(buf, i,
3089                                             struct btrfs_file_extent_item);
3090                         if (btrfs_file_extent_type(buf, fi) ==
3091                             BTRFS_FILE_EXTENT_INLINE)
3092                                 continue;
3093                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3094                         if (bytenr == 0)
3095                                 continue;
3096
3097                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3098                         key.offset -= btrfs_file_extent_offset(buf, fi);
3099                         ret = process_func(trans, root, bytenr, num_bytes,
3100                                            parent, ref_root, key.objectid,
3101                                            key.offset, 1);
3102                         if (ret)
3103                                 goto fail;
3104                 } else {
3105                         bytenr = btrfs_node_blockptr(buf, i);
3106                         num_bytes = root->nodesize;
3107                         ret = process_func(trans, root, bytenr, num_bytes,
3108                                            parent, ref_root, level - 1, 0,
3109                                            1);
3110                         if (ret)
3111                                 goto fail;
3112                 }
3113         }
3114         return 0;
3115 fail:
3116         return ret;
3117 }
3118
3119 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3120                   struct extent_buffer *buf, int full_backref)
3121 {
3122         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3123 }
3124
3125 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3126                   struct extent_buffer *buf, int full_backref)
3127 {
3128         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3129 }
3130
3131 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3132                                  struct btrfs_root *root,
3133                                  struct btrfs_path *path,
3134                                  struct btrfs_block_group_cache *cache)
3135 {
3136         int ret;
3137         struct btrfs_root *extent_root = root->fs_info->extent_root;
3138         unsigned long bi;
3139         struct extent_buffer *leaf;
3140
3141         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3142         if (ret < 0)
3143                 goto fail;
3144         BUG_ON(ret); /* Corruption */
3145
3146         leaf = path->nodes[0];
3147         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3148         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3149         btrfs_mark_buffer_dirty(leaf);
3150         btrfs_release_path(path);
3151 fail:
3152         if (ret) {
3153                 btrfs_abort_transaction(trans, root, ret);
3154                 return ret;
3155         }
3156         return 0;
3157
3158 }
3159
3160 static struct btrfs_block_group_cache *
3161 next_block_group(struct btrfs_root *root,
3162                  struct btrfs_block_group_cache *cache)
3163 {
3164         struct rb_node *node;
3165         spin_lock(&root->fs_info->block_group_cache_lock);
3166         node = rb_next(&cache->cache_node);
3167         btrfs_put_block_group(cache);
3168         if (node) {
3169                 cache = rb_entry(node, struct btrfs_block_group_cache,
3170                                  cache_node);
3171                 btrfs_get_block_group(cache);
3172         } else
3173                 cache = NULL;
3174         spin_unlock(&root->fs_info->block_group_cache_lock);
3175         return cache;
3176 }
3177
3178 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3179                             struct btrfs_trans_handle *trans,
3180                             struct btrfs_path *path)
3181 {
3182         struct btrfs_root *root = block_group->fs_info->tree_root;
3183         struct inode *inode = NULL;
3184         u64 alloc_hint = 0;
3185         int dcs = BTRFS_DC_ERROR;
3186         int num_pages = 0;
3187         int retries = 0;
3188         int ret = 0;
3189
3190         /*
3191          * If this block group is smaller than 100 megs don't bother caching the
3192          * block group.
3193          */
3194         if (block_group->key.offset < (100 * 1024 * 1024)) {
3195                 spin_lock(&block_group->lock);
3196                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3197                 spin_unlock(&block_group->lock);
3198                 return 0;
3199         }
3200
3201 again:
3202         inode = lookup_free_space_inode(root, block_group, path);
3203         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3204                 ret = PTR_ERR(inode);
3205                 btrfs_release_path(path);
3206                 goto out;
3207         }
3208
3209         if (IS_ERR(inode)) {
3210                 BUG_ON(retries);
3211                 retries++;
3212
3213                 if (block_group->ro)
3214                         goto out_free;
3215
3216                 ret = create_free_space_inode(root, trans, block_group, path);
3217                 if (ret)
3218                         goto out_free;
3219                 goto again;
3220         }
3221
3222         /* We've already setup this transaction, go ahead and exit */
3223         if (block_group->cache_generation == trans->transid &&
3224             i_size_read(inode)) {
3225                 dcs = BTRFS_DC_SETUP;
3226                 goto out_put;
3227         }
3228
3229         /*
3230          * We want to set the generation to 0, that way if anything goes wrong
3231          * from here on out we know not to trust this cache when we load up next
3232          * time.
3233          */
3234         BTRFS_I(inode)->generation = 0;
3235         ret = btrfs_update_inode(trans, root, inode);
3236         WARN_ON(ret);
3237
3238         if (i_size_read(inode) > 0) {
3239                 ret = btrfs_check_trunc_cache_free_space(root,
3240                                         &root->fs_info->global_block_rsv);
3241                 if (ret)
3242                         goto out_put;
3243
3244                 ret = btrfs_truncate_free_space_cache(root, trans, inode);
3245                 if (ret)
3246                         goto out_put;
3247         }
3248
3249         spin_lock(&block_group->lock);
3250         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3251             !btrfs_test_opt(root, SPACE_CACHE) ||
3252             block_group->delalloc_bytes) {
3253                 /*
3254                  * don't bother trying to write stuff out _if_
3255                  * a) we're not cached,
3256                  * b) we're with nospace_cache mount option.
3257                  */
3258                 dcs = BTRFS_DC_WRITTEN;
3259                 spin_unlock(&block_group->lock);
3260                 goto out_put;
3261         }
3262         spin_unlock(&block_group->lock);
3263
3264         /*
3265          * Try to preallocate enough space based on how big the block group is.
3266          * Keep in mind this has to include any pinned space which could end up
3267          * taking up quite a bit since it's not folded into the other space
3268          * cache.
3269          */
3270         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3271         if (!num_pages)
3272                 num_pages = 1;
3273
3274         num_pages *= 16;
3275         num_pages *= PAGE_CACHE_SIZE;
3276
3277         ret = btrfs_check_data_free_space(inode, num_pages);
3278         if (ret)
3279                 goto out_put;
3280
3281         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3282                                               num_pages, num_pages,
3283                                               &alloc_hint);
3284         if (!ret)
3285                 dcs = BTRFS_DC_SETUP;
3286         btrfs_free_reserved_data_space(inode, num_pages);
3287
3288 out_put:
3289         iput(inode);
3290 out_free:
3291         btrfs_release_path(path);
3292 out:
3293         spin_lock(&block_group->lock);
3294         if (!ret && dcs == BTRFS_DC_SETUP)
3295                 block_group->cache_generation = trans->transid;
3296         block_group->disk_cache_state = dcs;
3297         spin_unlock(&block_group->lock);
3298
3299         return ret;
3300 }
3301
3302 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3303                                    struct btrfs_root *root)
3304 {
3305         struct btrfs_block_group_cache *cache;
3306         int err = 0;
3307         struct btrfs_path *path;
3308         u64 last = 0;
3309
3310         path = btrfs_alloc_path();
3311         if (!path)
3312                 return -ENOMEM;
3313
3314 again:
3315         while (1) {
3316                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3317                 while (cache) {
3318                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3319                                 break;
3320                         cache = next_block_group(root, cache);
3321                 }
3322                 if (!cache) {
3323                         if (last == 0)
3324                                 break;
3325                         last = 0;
3326                         continue;
3327                 }
3328                 err = cache_save_setup(cache, trans, path);
3329                 last = cache->key.objectid + cache->key.offset;
3330                 btrfs_put_block_group(cache);
3331         }
3332
3333         while (1) {
3334                 if (last == 0) {
3335                         err = btrfs_run_delayed_refs(trans, root,
3336                                                      (unsigned long)-1);
3337                         if (err) /* File system offline */
3338                                 goto out;
3339                 }
3340
3341                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3342                 while (cache) {
3343                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3344                                 btrfs_put_block_group(cache);
3345                                 goto again;
3346                         }
3347
3348                         if (cache->dirty)
3349                                 break;
3350                         cache = next_block_group(root, cache);
3351                 }
3352                 if (!cache) {
3353                         if (last == 0)
3354                                 break;
3355                         last = 0;
3356                         continue;
3357                 }
3358
3359                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3360                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3361                 cache->dirty = 0;
3362                 last = cache->key.objectid + cache->key.offset;
3363
3364                 err = write_one_cache_group(trans, root, path, cache);
3365                 btrfs_put_block_group(cache);
3366                 if (err) /* File system offline */
3367                         goto out;
3368         }
3369
3370         while (1) {
3371                 /*
3372                  * I don't think this is needed since we're just marking our
3373                  * preallocated extent as written, but just in case it can't
3374                  * hurt.
3375                  */
3376                 if (last == 0) {
3377                         err = btrfs_run_delayed_refs(trans, root,
3378                                                      (unsigned long)-1);
3379                         if (err) /* File system offline */
3380                                 goto out;
3381                 }
3382
3383                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3384                 while (cache) {
3385                         /*
3386                          * Really this shouldn't happen, but it could if we
3387                          * couldn't write the entire preallocated extent and
3388                          * splitting the extent resulted in a new block.
3389                          */
3390                         if (cache->dirty) {
3391                                 btrfs_put_block_group(cache);
3392                                 goto again;
3393                         }
3394                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3395                                 break;
3396                         cache = next_block_group(root, cache);
3397                 }
3398                 if (!cache) {
3399                         if (last == 0)
3400                                 break;
3401                         last = 0;
3402                         continue;
3403                 }
3404
3405                 err = btrfs_write_out_cache(root, trans, cache, path);
3406
3407                 /*
3408                  * If we didn't have an error then the cache state is still
3409                  * NEED_WRITE, so we can set it to WRITTEN.
3410                  */
3411                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3412                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3413                 last = cache->key.objectid + cache->key.offset;
3414                 btrfs_put_block_group(cache);
3415         }
3416 out:
3417
3418         btrfs_free_path(path);
3419         return err;
3420 }
3421
3422 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3423 {
3424         struct btrfs_block_group_cache *block_group;
3425         int readonly = 0;
3426
3427         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3428         if (!block_group || block_group->ro)
3429                 readonly = 1;
3430         if (block_group)
3431                 btrfs_put_block_group(block_group);
3432         return readonly;
3433 }
3434
3435 static const char *alloc_name(u64 flags)
3436 {
3437         switch (flags) {
3438         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3439                 return "mixed";
3440         case BTRFS_BLOCK_GROUP_METADATA:
3441                 return "metadata";
3442         case BTRFS_BLOCK_GROUP_DATA:
3443                 return "data";
3444         case BTRFS_BLOCK_GROUP_SYSTEM:
3445                 return "system";
3446         default:
3447                 WARN_ON(1);
3448                 return "invalid-combination";
3449         };
3450 }
3451
3452 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3453                              u64 total_bytes, u64 bytes_used,
3454                              struct btrfs_space_info **space_info)
3455 {
3456         struct btrfs_space_info *found;
3457         int i;
3458         int factor;
3459         int ret;
3460
3461         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3462                      BTRFS_BLOCK_GROUP_RAID10))
3463                 factor = 2;
3464         else
3465                 factor = 1;
3466
3467         found = __find_space_info(info, flags);
3468         if (found) {
3469                 spin_lock(&found->lock);
3470                 found->total_bytes += total_bytes;
3471                 found->disk_total += total_bytes * factor;
3472                 found->bytes_used += bytes_used;
3473                 found->disk_used += bytes_used * factor;
3474                 found->full = 0;
3475                 spin_unlock(&found->lock);
3476                 *space_info = found;
3477                 return 0;
3478         }
3479         found = kzalloc(sizeof(*found), GFP_NOFS);
3480         if (!found)
3481                 return -ENOMEM;
3482
3483         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3484         if (ret) {
3485                 kfree(found);
3486                 return ret;
3487         }
3488
3489         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3490                 INIT_LIST_HEAD(&found->block_groups[i]);
3491         init_rwsem(&found->groups_sem);
3492         spin_lock_init(&found->lock);
3493         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3494         found->total_bytes = total_bytes;
3495         found->disk_total = total_bytes * factor;
3496         found->bytes_used = bytes_used;
3497         found->disk_used = bytes_used * factor;
3498         found->bytes_pinned = 0;
3499         found->bytes_reserved = 0;
3500         found->bytes_readonly = 0;
3501         found->bytes_may_use = 0;
3502         found->full = 0;
3503         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3504         found->chunk_alloc = 0;
3505         found->flush = 0;
3506         init_waitqueue_head(&found->wait);
3507
3508         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3509                                     info->space_info_kobj, "%s",
3510                                     alloc_name(found->flags));
3511         if (ret) {
3512                 kfree(found);
3513                 return ret;
3514         }
3515
3516         *space_info = found;
3517         list_add_rcu(&found->list, &info->space_info);
3518         if (flags & BTRFS_BLOCK_GROUP_DATA)
3519                 info->data_sinfo = found;
3520
3521         return ret;
3522 }
3523
3524 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3525 {
3526         u64 extra_flags = chunk_to_extended(flags) &
3527                                 BTRFS_EXTENDED_PROFILE_MASK;
3528
3529         write_seqlock(&fs_info->profiles_lock);
3530         if (flags & BTRFS_BLOCK_GROUP_DATA)
3531                 fs_info->avail_data_alloc_bits |= extra_flags;
3532         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3533                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3534         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3535                 fs_info->avail_system_alloc_bits |= extra_flags;
3536         write_sequnlock(&fs_info->profiles_lock);
3537 }
3538
3539 /*
3540  * returns target flags in extended format or 0 if restripe for this
3541  * chunk_type is not in progress
3542  *
3543  * should be called with either volume_mutex or balance_lock held
3544  */
3545 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3546 {
3547         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3548         u64 target = 0;
3549
3550         if (!bctl)
3551                 return 0;
3552
3553         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3554             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3555                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3556         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3557                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3558                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3559         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3560                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3561                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3562         }
3563
3564         return target;
3565 }
3566
3567 /*
3568  * @flags: available profiles in extended format (see ctree.h)
3569  *
3570  * Returns reduced profile in chunk format.  If profile changing is in
3571  * progress (either running or paused) picks the target profile (if it's
3572  * already available), otherwise falls back to plain reducing.
3573  */
3574 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3575 {
3576         u64 num_devices = root->fs_info->fs_devices->rw_devices;
3577         u64 target;
3578         u64 tmp;
3579
3580         /*
3581          * see if restripe for this chunk_type is in progress, if so
3582          * try to reduce to the target profile
3583          */
3584         spin_lock(&root->fs_info->balance_lock);
3585         target = get_restripe_target(root->fs_info, flags);
3586         if (target) {
3587                 /* pick target profile only if it's already available */
3588                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3589                         spin_unlock(&root->fs_info->balance_lock);
3590                         return extended_to_chunk(target);
3591                 }
3592         }
3593         spin_unlock(&root->fs_info->balance_lock);
3594
3595         /* First, mask out the RAID levels which aren't possible */
3596         if (num_devices == 1)
3597                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3598                            BTRFS_BLOCK_GROUP_RAID5);
3599         if (num_devices < 3)
3600                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3601         if (num_devices < 4)
3602                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3603
3604         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3605                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3606                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3607         flags &= ~tmp;
3608
3609         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3610                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3611         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3612                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3613         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3614                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3615         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3616                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3617         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3618                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3619
3620         return extended_to_chunk(flags | tmp);
3621 }
3622
3623 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
3624 {
3625         unsigned seq;
3626         u64 flags;
3627
3628         do {
3629                 flags = orig_flags;
3630                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3631
3632                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3633                         flags |= root->fs_info->avail_data_alloc_bits;
3634                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3635                         flags |= root->fs_info->avail_system_alloc_bits;
3636                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3637                         flags |= root->fs_info->avail_metadata_alloc_bits;
3638         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3639
3640         return btrfs_reduce_alloc_profile(root, flags);
3641 }
3642
3643 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3644 {
3645         u64 flags;
3646         u64 ret;
3647
3648         if (data)
3649                 flags = BTRFS_BLOCK_GROUP_DATA;
3650         else if (root == root->fs_info->chunk_root)
3651                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3652         else
3653                 flags = BTRFS_BLOCK_GROUP_METADATA;
3654
3655         ret = get_alloc_profile(root, flags);
3656         return ret;
3657 }
3658
3659 /*
3660  * This will check the space that the inode allocates from to make sure we have
3661  * enough space for bytes.
3662  */
3663 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3664 {
3665         struct btrfs_space_info *data_sinfo;
3666         struct btrfs_root *root = BTRFS_I(inode)->root;
3667         struct btrfs_fs_info *fs_info = root->fs_info;
3668         u64 used;
3669         int ret = 0, committed = 0, alloc_chunk = 1;
3670
3671         /* make sure bytes are sectorsize aligned */
3672         bytes = ALIGN(bytes, root->sectorsize);
3673
3674         if (btrfs_is_free_space_inode(inode)) {
3675                 committed = 1;
3676                 ASSERT(current->journal_info);
3677         }
3678
3679         data_sinfo = fs_info->data_sinfo;
3680         if (!data_sinfo)
3681                 goto alloc;
3682
3683 again:
3684         /* make sure we have enough space to handle the data first */
3685         spin_lock(&data_sinfo->lock);
3686         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3687                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3688                 data_sinfo->bytes_may_use;
3689
3690         if (used + bytes > data_sinfo->total_bytes) {
3691                 struct btrfs_trans_handle *trans;
3692
3693                 /*
3694                  * if we don't have enough free bytes in this space then we need
3695                  * to alloc a new chunk.
3696                  */
3697                 if (!data_sinfo->full && alloc_chunk) {
3698                         u64 alloc_target;
3699
3700                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3701                         spin_unlock(&data_sinfo->lock);
3702 alloc:
3703                         alloc_target = btrfs_get_alloc_profile(root, 1);
3704                         /*
3705                          * It is ugly that we don't call nolock join
3706                          * transaction for the free space inode case here.
3707                          * But it is safe because we only do the data space
3708                          * reservation for the free space cache in the
3709                          * transaction context, the common join transaction
3710                          * just increase the counter of the current transaction
3711                          * handler, doesn't try to acquire the trans_lock of
3712                          * the fs.
3713                          */
3714                         trans = btrfs_join_transaction(root);
3715                         if (IS_ERR(trans))
3716                                 return PTR_ERR(trans);
3717
3718                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3719                                              alloc_target,
3720                                              CHUNK_ALLOC_NO_FORCE);
3721                         btrfs_end_transaction(trans, root);
3722                         if (ret < 0) {
3723                                 if (ret != -ENOSPC)
3724                                         return ret;
3725                                 else
3726                                         goto commit_trans;
3727                         }
3728
3729                         if (!data_sinfo)
3730                                 data_sinfo = fs_info->data_sinfo;
3731
3732                         goto again;
3733                 }
3734
3735                 /*
3736                  * If we don't have enough pinned space to deal with this
3737                  * allocation don't bother committing the transaction.
3738                  */
3739                 if (percpu_counter_compare(&data_sinfo->total_bytes_pinned,
3740                                            bytes) < 0)
3741                         committed = 1;
3742                 spin_unlock(&data_sinfo->lock);
3743
3744                 /* commit the current transaction and try again */
3745 commit_trans:
3746                 if (!committed &&
3747                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3748                         committed = 1;
3749
3750                         trans = btrfs_join_transaction(root);
3751                         if (IS_ERR(trans))
3752                                 return PTR_ERR(trans);
3753                         ret = btrfs_commit_transaction(trans, root);
3754                         if (ret)
3755                                 return ret;
3756                         goto again;
3757                 }
3758
3759                 trace_btrfs_space_reservation(root->fs_info,
3760                                               "space_info:enospc",
3761                                               data_sinfo->flags, bytes, 1);
3762                 return -ENOSPC;
3763         }
3764         data_sinfo->bytes_may_use += bytes;
3765         trace_btrfs_space_reservation(root->fs_info, "space_info",
3766                                       data_sinfo->flags, bytes, 1);
3767         spin_unlock(&data_sinfo->lock);
3768
3769         return 0;
3770 }
3771
3772 /*
3773  * Called if we need to clear a data reservation for this inode.
3774  */
3775 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3776 {
3777         struct btrfs_root *root = BTRFS_I(inode)->root;
3778         struct btrfs_space_info *data_sinfo;
3779
3780         /* make sure bytes are sectorsize aligned */
3781         bytes = ALIGN(bytes, root->sectorsize);
3782
3783         data_sinfo = root->fs_info->data_sinfo;
3784         spin_lock(&data_sinfo->lock);
3785         WARN_ON(data_sinfo->bytes_may_use < bytes);
3786         data_sinfo->bytes_may_use -= bytes;
3787         trace_btrfs_space_reservation(root->fs_info, "space_info",
3788                                       data_sinfo->flags, bytes, 0);
3789         spin_unlock(&data_sinfo->lock);
3790 }
3791
3792 static void force_metadata_allocation(struct btrfs_fs_info *info)
3793 {
3794         struct list_head *head = &info->space_info;
3795         struct btrfs_space_info *found;
3796
3797         rcu_read_lock();
3798         list_for_each_entry_rcu(found, head, list) {
3799                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3800                         found->force_alloc = CHUNK_ALLOC_FORCE;
3801         }
3802         rcu_read_unlock();
3803 }
3804
3805 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
3806 {
3807         return (global->size << 1);
3808 }
3809
3810 static int should_alloc_chunk(struct btrfs_root *root,
3811                               struct btrfs_space_info *sinfo, int force)
3812 {
3813         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3814         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3815         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3816         u64 thresh;
3817
3818         if (force == CHUNK_ALLOC_FORCE)
3819                 return 1;
3820
3821         /*
3822          * We need to take into account the global rsv because for all intents
3823          * and purposes it's used space.  Don't worry about locking the
3824          * global_rsv, it doesn't change except when the transaction commits.
3825          */
3826         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3827                 num_allocated += calc_global_rsv_need_space(global_rsv);
3828
3829         /*
3830          * in limited mode, we want to have some free space up to
3831          * about 1% of the FS size.
3832          */
3833         if (force == CHUNK_ALLOC_LIMITED) {
3834                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3835                 thresh = max_t(u64, 64 * 1024 * 1024,
3836                                div_factor_fine(thresh, 1));
3837
3838                 if (num_bytes - num_allocated < thresh)
3839                         return 1;
3840         }
3841
3842         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3843                 return 0;
3844         return 1;
3845 }
3846
3847 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3848 {
3849         u64 num_dev;
3850
3851         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
3852                     BTRFS_BLOCK_GROUP_RAID0 |
3853                     BTRFS_BLOCK_GROUP_RAID5 |
3854                     BTRFS_BLOCK_GROUP_RAID6))
3855                 num_dev = root->fs_info->fs_devices->rw_devices;
3856         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3857                 num_dev = 2;
3858         else
3859                 num_dev = 1;    /* DUP or single */
3860
3861         /* metadata for updaing devices and chunk tree */
3862         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3863 }
3864
3865 static void check_system_chunk(struct btrfs_trans_handle *trans,
3866                                struct btrfs_root *root, u64 type)
3867 {
3868         struct btrfs_space_info *info;
3869         u64 left;
3870         u64 thresh;
3871
3872         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3873         spin_lock(&info->lock);
3874         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3875                 info->bytes_reserved - info->bytes_readonly;
3876         spin_unlock(&info->lock);
3877
3878         thresh = get_system_chunk_thresh(root, type);
3879         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3880                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
3881                         left, thresh, type);
3882                 dump_space_info(info, 0, 0);
3883         }
3884
3885         if (left < thresh) {
3886                 u64 flags;
3887
3888                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3889                 btrfs_alloc_chunk(trans, root, flags);
3890         }
3891 }
3892
3893 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3894                           struct btrfs_root *extent_root, u64 flags, int force)
3895 {
3896         struct btrfs_space_info *space_info;
3897         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3898         int wait_for_alloc = 0;
3899         int ret = 0;
3900
3901         /* Don't re-enter if we're already allocating a chunk */
3902         if (trans->allocating_chunk)
3903                 return -ENOSPC;
3904
3905         space_info = __find_space_info(extent_root->fs_info, flags);
3906         if (!space_info) {
3907                 ret = update_space_info(extent_root->fs_info, flags,
3908                                         0, 0, &space_info);
3909                 BUG_ON(ret); /* -ENOMEM */
3910         }
3911         BUG_ON(!space_info); /* Logic error */
3912
3913 again:
3914         spin_lock(&space_info->lock);
3915         if (force < space_info->force_alloc)
3916                 force = space_info->force_alloc;
3917         if (space_info->full) {
3918                 if (should_alloc_chunk(extent_root, space_info, force))
3919                         ret = -ENOSPC;
3920                 else
3921                         ret = 0;
3922                 spin_unlock(&space_info->lock);
3923                 return ret;
3924         }
3925
3926         if (!should_alloc_chunk(extent_root, space_info, force)) {
3927                 spin_unlock(&space_info->lock);
3928                 return 0;
3929         } else if (space_info->chunk_alloc) {
3930                 wait_for_alloc = 1;
3931         } else {
3932                 space_info->chunk_alloc = 1;
3933         }
3934
3935         spin_unlock(&space_info->lock);
3936
3937         mutex_lock(&fs_info->chunk_mutex);
3938
3939         /*
3940          * The chunk_mutex is held throughout the entirety of a chunk
3941          * allocation, so once we've acquired the chunk_mutex we know that the
3942          * other guy is done and we need to recheck and see if we should
3943          * allocate.
3944          */
3945         if (wait_for_alloc) {
3946                 mutex_unlock(&fs_info->chunk_mutex);
3947                 wait_for_alloc = 0;
3948                 goto again;
3949         }
3950
3951         trans->allocating_chunk = true;
3952
3953         /*
3954          * If we have mixed data/metadata chunks we want to make sure we keep
3955          * allocating mixed chunks instead of individual chunks.
3956          */
3957         if (btrfs_mixed_space_info(space_info))
3958                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3959
3960         /*
3961          * if we're doing a data chunk, go ahead and make sure that
3962          * we keep a reasonable number of metadata chunks allocated in the
3963          * FS as well.
3964          */
3965         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3966                 fs_info->data_chunk_allocations++;
3967                 if (!(fs_info->data_chunk_allocations %
3968                       fs_info->metadata_ratio))
3969                         force_metadata_allocation(fs_info);
3970         }
3971
3972         /*
3973          * Check if we have enough space in SYSTEM chunk because we may need
3974          * to update devices.
3975          */
3976         check_system_chunk(trans, extent_root, flags);
3977
3978         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3979         trans->allocating_chunk = false;
3980
3981         spin_lock(&space_info->lock);
3982         if (ret < 0 && ret != -ENOSPC)
3983                 goto out;
3984         if (ret)
3985                 space_info->full = 1;
3986         else
3987                 ret = 1;
3988
3989         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
3990 out:
3991         space_info->chunk_alloc = 0;
3992         spin_unlock(&space_info->lock);
3993         mutex_unlock(&fs_info->chunk_mutex);
3994         return ret;
3995 }
3996
3997 static int can_overcommit(struct btrfs_root *root,
3998                           struct btrfs_space_info *space_info, u64 bytes,
3999                           enum btrfs_reserve_flush_enum flush)
4000 {
4001         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4002         u64 profile = btrfs_get_alloc_profile(root, 0);
4003         u64 space_size;
4004         u64 avail;
4005         u64 used;
4006
4007         used = space_info->bytes_used + space_info->bytes_reserved +
4008                 space_info->bytes_pinned + space_info->bytes_readonly;
4009
4010         /*
4011          * We only want to allow over committing if we have lots of actual space
4012          * free, but if we don't have enough space to handle the global reserve
4013          * space then we could end up having a real enospc problem when trying
4014          * to allocate a chunk or some other such important allocation.
4015          */
4016         spin_lock(&global_rsv->lock);
4017         space_size = calc_global_rsv_need_space(global_rsv);
4018         spin_unlock(&global_rsv->lock);
4019         if (used + space_size >= space_info->total_bytes)
4020                 return 0;
4021
4022         used += space_info->bytes_may_use;
4023
4024         spin_lock(&root->fs_info->free_chunk_lock);
4025         avail = root->fs_info->free_chunk_space;
4026         spin_unlock(&root->fs_info->free_chunk_lock);
4027
4028         /*
4029          * If we have dup, raid1 or raid10 then only half of the free
4030          * space is actually useable.  For raid56, the space info used
4031          * doesn't include the parity drive, so we don't have to
4032          * change the math
4033          */
4034         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4035                        BTRFS_BLOCK_GROUP_RAID1 |
4036                        BTRFS_BLOCK_GROUP_RAID10))
4037                 avail >>= 1;
4038
4039         /*
4040          * If we aren't flushing all things, let us overcommit up to
4041          * 1/2th of the space. If we can flush, don't let us overcommit
4042          * too much, let it overcommit up to 1/8 of the space.
4043          */
4044         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4045                 avail >>= 3;
4046         else
4047                 avail >>= 1;
4048
4049         if (used + bytes < space_info->total_bytes + avail)
4050                 return 1;
4051         return 0;
4052 }
4053
4054 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4055                                          unsigned long nr_pages, int nr_items)
4056 {
4057         struct super_block *sb = root->fs_info->sb;
4058
4059         if (down_read_trylock(&sb->s_umount)) {
4060                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4061                 up_read(&sb->s_umount);
4062         } else {
4063                 /*
4064                  * We needn't worry the filesystem going from r/w to r/o though
4065                  * we don't acquire ->s_umount mutex, because the filesystem
4066                  * should guarantee the delalloc inodes list be empty after
4067                  * the filesystem is readonly(all dirty pages are written to
4068                  * the disk).
4069                  */
4070                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4071                 if (!current->journal_info)
4072                         btrfs_wait_ordered_roots(root->fs_info, nr_items);
4073         }
4074 }
4075
4076 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4077 {
4078         u64 bytes;
4079         int nr;
4080
4081         bytes = btrfs_calc_trans_metadata_size(root, 1);
4082         nr = (int)div64_u64(to_reclaim, bytes);
4083         if (!nr)
4084                 nr = 1;
4085         return nr;
4086 }
4087
4088 #define EXTENT_SIZE_PER_ITEM    (256 * 1024)
4089
4090 /*
4091  * shrink metadata reservation for delalloc
4092  */
4093 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4094                             bool wait_ordered)
4095 {
4096         struct btrfs_block_rsv *block_rsv;
4097         struct btrfs_space_info *space_info;
4098         struct btrfs_trans_handle *trans;
4099         u64 delalloc_bytes;
4100         u64 max_reclaim;
4101         long time_left;
4102         unsigned long nr_pages;
4103         int loops;
4104         int items;
4105         enum btrfs_reserve_flush_enum flush;
4106
4107         /* Calc the number of the pages we need flush for space reservation */
4108         items = calc_reclaim_items_nr(root, to_reclaim);
4109         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4110
4111         trans = (struct btrfs_trans_handle *)current->journal_info;
4112         block_rsv = &root->fs_info->delalloc_block_rsv;
4113         space_info = block_rsv->space_info;
4114
4115         delalloc_bytes = percpu_counter_sum_positive(
4116                                                 &root->fs_info->delalloc_bytes);
4117         if (delalloc_bytes == 0) {
4118                 if (trans)
4119                         return;
4120                 if (wait_ordered)
4121                         btrfs_wait_ordered_roots(root->fs_info, items);
4122                 return;
4123         }
4124
4125         loops = 0;
4126         while (delalloc_bytes && loops < 3) {
4127                 max_reclaim = min(delalloc_bytes, to_reclaim);
4128                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4129                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4130                 /*
4131                  * We need to wait for the async pages to actually start before
4132                  * we do anything.
4133                  */
4134                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4135                 if (!max_reclaim)
4136                         goto skip_async;
4137
4138                 if (max_reclaim <= nr_pages)
4139                         max_reclaim = 0;
4140                 else
4141                         max_reclaim -= nr_pages;
4142
4143                 wait_event(root->fs_info->async_submit_wait,
4144                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4145                            (int)max_reclaim);
4146 skip_async:
4147                 if (!trans)
4148                         flush = BTRFS_RESERVE_FLUSH_ALL;
4149                 else
4150                         flush = BTRFS_RESERVE_NO_FLUSH;
4151                 spin_lock(&space_info->lock);
4152                 if (can_overcommit(root, space_info, orig, flush)) {
4153                         spin_unlock(&space_info->lock);
4154                         break;
4155                 }
4156                 spin_unlock(&space_info->lock);
4157
4158                 loops++;
4159                 if (wait_ordered && !trans) {
4160                         btrfs_wait_ordered_roots(root->fs_info, items);
4161                 } else {
4162                         time_left = schedule_timeout_killable(1);
4163                         if (time_left)
4164                                 break;
4165                 }
4166                 delalloc_bytes = percpu_counter_sum_positive(
4167                                                 &root->fs_info->delalloc_bytes);
4168         }
4169 }
4170
4171 /**
4172  * maybe_commit_transaction - possibly commit the transaction if its ok to
4173  * @root - the root we're allocating for
4174  * @bytes - the number of bytes we want to reserve
4175  * @force - force the commit
4176  *
4177  * This will check to make sure that committing the transaction will actually
4178  * get us somewhere and then commit the transaction if it does.  Otherwise it
4179  * will return -ENOSPC.
4180  */
4181 static int may_commit_transaction(struct btrfs_root *root,
4182                                   struct btrfs_space_info *space_info,
4183                                   u64 bytes, int force)
4184 {
4185         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4186         struct btrfs_trans_handle *trans;
4187
4188         trans = (struct btrfs_trans_handle *)current->journal_info;
4189         if (trans)
4190                 return -EAGAIN;
4191
4192         if (force)
4193                 goto commit;
4194
4195         /* See if there is enough pinned space to make this reservation */
4196         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4197                                    bytes) >= 0)
4198                 goto commit;
4199
4200         /*
4201          * See if there is some space in the delayed insertion reservation for
4202          * this reservation.
4203          */
4204         if (space_info != delayed_rsv->space_info)
4205                 return -ENOSPC;
4206
4207         spin_lock(&delayed_rsv->lock);
4208         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4209                                    bytes - delayed_rsv->size) >= 0) {
4210                 spin_unlock(&delayed_rsv->lock);
4211                 return -ENOSPC;
4212         }
4213         spin_unlock(&delayed_rsv->lock);
4214
4215 commit:
4216         trans = btrfs_join_transaction(root);
4217         if (IS_ERR(trans))
4218                 return -ENOSPC;
4219
4220         return btrfs_commit_transaction(trans, root);
4221 }
4222
4223 enum flush_state {
4224         FLUSH_DELAYED_ITEMS_NR  =       1,
4225         FLUSH_DELAYED_ITEMS     =       2,
4226         FLUSH_DELALLOC          =       3,
4227         FLUSH_DELALLOC_WAIT     =       4,
4228         ALLOC_CHUNK             =       5,
4229         COMMIT_TRANS            =       6,
4230 };
4231
4232 static int flush_space(struct btrfs_root *root,
4233                        struct btrfs_space_info *space_info, u64 num_bytes,
4234                        u64 orig_bytes, int state)
4235 {
4236         struct btrfs_trans_handle *trans;
4237         int nr;
4238         int ret = 0;
4239
4240         switch (state) {
4241         case FLUSH_DELAYED_ITEMS_NR:
4242         case FLUSH_DELAYED_ITEMS:
4243                 if (state == FLUSH_DELAYED_ITEMS_NR)
4244                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4245                 else
4246                         nr = -1;
4247
4248                 trans = btrfs_join_transaction(root);
4249                 if (IS_ERR(trans)) {
4250                         ret = PTR_ERR(trans);
4251                         break;
4252                 }
4253                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4254                 btrfs_end_transaction(trans, root);
4255                 break;
4256         case FLUSH_DELALLOC:
4257         case FLUSH_DELALLOC_WAIT:
4258                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4259                                 state == FLUSH_DELALLOC_WAIT);
4260                 break;
4261         case ALLOC_CHUNK:
4262                 trans = btrfs_join_transaction(root);
4263                 if (IS_ERR(trans)) {
4264                         ret = PTR_ERR(trans);
4265                         break;
4266                 }
4267                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4268                                      btrfs_get_alloc_profile(root, 0),
4269                                      CHUNK_ALLOC_NO_FORCE);
4270                 btrfs_end_transaction(trans, root);
4271                 if (ret == -ENOSPC)
4272                         ret = 0;
4273                 break;
4274         case COMMIT_TRANS:
4275                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4276                 break;
4277         default:
4278                 ret = -ENOSPC;
4279                 break;
4280         }
4281
4282         return ret;
4283 }
4284
4285 static inline u64
4286 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4287                                  struct btrfs_space_info *space_info)
4288 {
4289         u64 used;
4290         u64 expected;
4291         u64 to_reclaim;
4292
4293         to_reclaim = min_t(u64, num_online_cpus() * 1024 * 1024,
4294                                 16 * 1024 * 1024);
4295         spin_lock(&space_info->lock);
4296         if (can_overcommit(root, space_info, to_reclaim,
4297                            BTRFS_RESERVE_FLUSH_ALL)) {
4298                 to_reclaim = 0;
4299                 goto out;
4300         }
4301
4302         used = space_info->bytes_used + space_info->bytes_reserved +
4303                space_info->bytes_pinned + space_info->bytes_readonly +
4304                space_info->bytes_may_use;
4305         if (can_overcommit(root, space_info, 1024 * 1024,
4306                            BTRFS_RESERVE_FLUSH_ALL))
4307                 expected = div_factor_fine(space_info->total_bytes, 95);
4308         else
4309                 expected = div_factor_fine(space_info->total_bytes, 90);
4310
4311         if (used > expected)
4312                 to_reclaim = used - expected;
4313         else
4314                 to_reclaim = 0;
4315         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4316                                      space_info->bytes_reserved);
4317 out:
4318         spin_unlock(&space_info->lock);
4319
4320         return to_reclaim;
4321 }
4322
4323 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4324                                         struct btrfs_fs_info *fs_info, u64 used)
4325 {
4326         return (used >= div_factor_fine(space_info->total_bytes, 98) &&
4327                 !btrfs_fs_closing(fs_info) &&
4328                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4329 }
4330
4331 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4332                                        struct btrfs_fs_info *fs_info,
4333                                        int flush_state)
4334 {
4335         u64 used;
4336
4337         spin_lock(&space_info->lock);
4338         /*
4339          * We run out of space and have not got any free space via flush_space,
4340          * so don't bother doing async reclaim.
4341          */
4342         if (flush_state > COMMIT_TRANS && space_info->full) {
4343                 spin_unlock(&space_info->lock);
4344                 return 0;
4345         }
4346
4347         used = space_info->bytes_used + space_info->bytes_reserved +
4348                space_info->bytes_pinned + space_info->bytes_readonly +
4349                space_info->bytes_may_use;
4350         if (need_do_async_reclaim(space_info, fs_info, used)) {
4351                 spin_unlock(&space_info->lock);
4352                 return 1;
4353         }
4354         spin_unlock(&space_info->lock);
4355
4356         return 0;
4357 }
4358
4359 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4360 {
4361         struct btrfs_fs_info *fs_info;
4362         struct btrfs_space_info *space_info;
4363         u64 to_reclaim;
4364         int flush_state;
4365
4366         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4367         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4368
4369         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4370                                                       space_info);
4371         if (!to_reclaim)
4372                 return;
4373
4374         flush_state = FLUSH_DELAYED_ITEMS_NR;
4375         do {
4376                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4377                             to_reclaim, flush_state);
4378                 flush_state++;
4379                 if (!btrfs_need_do_async_reclaim(space_info, fs_info,
4380                                                  flush_state))
4381                         return;
4382         } while (flush_state <= COMMIT_TRANS);
4383
4384         if (btrfs_need_do_async_reclaim(space_info, fs_info, flush_state))
4385                 queue_work(system_unbound_wq, work);
4386 }
4387
4388 void btrfs_init_async_reclaim_work(struct work_struct *work)
4389 {
4390         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4391 }
4392
4393 /**
4394  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4395  * @root - the root we're allocating for
4396  * @block_rsv - the block_rsv we're allocating for
4397  * @orig_bytes - the number of bytes we want
4398  * @flush - whether or not we can flush to make our reservation
4399  *
4400  * This will reserve orgi_bytes number of bytes from the space info associated
4401  * with the block_rsv.  If there is not enough space it will make an attempt to
4402  * flush out space to make room.  It will do this by flushing delalloc if
4403  * possible or committing the transaction.  If flush is 0 then no attempts to
4404  * regain reservations will be made and this will fail if there is not enough
4405  * space already.
4406  */
4407 static int reserve_metadata_bytes(struct btrfs_root *root,
4408                                   struct btrfs_block_rsv *block_rsv,
4409                                   u64 orig_bytes,
4410                                   enum btrfs_reserve_flush_enum flush)
4411 {
4412         struct btrfs_space_info *space_info = block_rsv->space_info;
4413         u64 used;
4414         u64 num_bytes = orig_bytes;
4415         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4416         int ret = 0;
4417         bool flushing = false;
4418
4419 again:
4420         ret = 0;
4421         spin_lock(&space_info->lock);
4422         /*
4423          * We only want to wait if somebody other than us is flushing and we
4424          * are actually allowed to flush all things.
4425          */
4426         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4427                space_info->flush) {
4428                 spin_unlock(&space_info->lock);
4429                 /*
4430                  * If we have a trans handle we can't wait because the flusher
4431                  * may have to commit the transaction, which would mean we would
4432                  * deadlock since we are waiting for the flusher to finish, but
4433                  * hold the current transaction open.
4434                  */
4435                 if (current->journal_info)
4436                         return -EAGAIN;
4437                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4438                 /* Must have been killed, return */
4439                 if (ret)
4440                         return -EINTR;
4441
4442                 spin_lock(&space_info->lock);
4443         }
4444
4445         ret = -ENOSPC;
4446         used = space_info->bytes_used + space_info->bytes_reserved +
4447                 space_info->bytes_pinned + space_info->bytes_readonly +
4448                 space_info->bytes_may_use;
4449
4450         /*
4451          * The idea here is that we've not already over-reserved the block group
4452          * then we can go ahead and save our reservation first and then start
4453          * flushing if we need to.  Otherwise if we've already overcommitted
4454          * lets start flushing stuff first and then come back and try to make
4455          * our reservation.
4456          */
4457         if (used <= space_info->total_bytes) {
4458                 if (used + orig_bytes <= space_info->total_bytes) {
4459                         space_info->bytes_may_use += orig_bytes;
4460                         trace_btrfs_space_reservation(root->fs_info,
4461                                 "space_info", space_info->flags, orig_bytes, 1);
4462                         ret = 0;
4463                 } else {
4464                         /*
4465                          * Ok set num_bytes to orig_bytes since we aren't
4466                          * overocmmitted, this way we only try and reclaim what
4467                          * we need.
4468                          */
4469                         num_bytes = orig_bytes;
4470                 }
4471         } else {
4472                 /*
4473                  * Ok we're over committed, set num_bytes to the overcommitted
4474                  * amount plus the amount of bytes that we need for this
4475                  * reservation.
4476                  */
4477                 num_bytes = used - space_info->total_bytes +
4478                         (orig_bytes * 2);
4479         }
4480
4481         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4482                 space_info->bytes_may_use += orig_bytes;
4483                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4484                                               space_info->flags, orig_bytes,
4485                                               1);
4486                 ret = 0;
4487         }
4488
4489         /*
4490          * Couldn't make our reservation, save our place so while we're trying
4491          * to reclaim space we can actually use it instead of somebody else
4492          * stealing it from us.
4493          *
4494          * We make the other tasks wait for the flush only when we can flush
4495          * all things.
4496          */
4497         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4498                 flushing = true;
4499                 space_info->flush = 1;
4500         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
4501                 used += orig_bytes;
4502                 /*
4503                  * We will do the space reservation dance during log replay,
4504                  * which means we won't have fs_info->fs_root set, so don't do
4505                  * the async reclaim as we will panic.
4506                  */
4507                 if (!root->fs_info->log_root_recovering &&
4508                     need_do_async_reclaim(space_info, root->fs_info, used) &&
4509                     !work_busy(&root->fs_info->async_reclaim_work))
4510                         queue_work(system_unbound_wq,
4511                                    &root->fs_info->async_reclaim_work);
4512         }
4513         spin_unlock(&space_info->lock);
4514
4515         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4516                 goto out;
4517
4518         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4519                           flush_state);
4520         flush_state++;
4521
4522         /*
4523          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4524          * would happen. So skip delalloc flush.
4525          */
4526         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4527             (flush_state == FLUSH_DELALLOC ||
4528              flush_state == FLUSH_DELALLOC_WAIT))
4529                 flush_state = ALLOC_CHUNK;
4530
4531         if (!ret)
4532                 goto again;
4533         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4534                  flush_state < COMMIT_TRANS)
4535                 goto again;
4536         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4537                  flush_state <= COMMIT_TRANS)
4538                 goto again;
4539
4540 out:
4541         if (ret == -ENOSPC &&
4542             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4543                 struct btrfs_block_rsv *global_rsv =
4544                         &root->fs_info->global_block_rsv;
4545
4546                 if (block_rsv != global_rsv &&
4547                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4548                         ret = 0;
4549         }
4550         if (ret == -ENOSPC)
4551                 trace_btrfs_space_reservation(root->fs_info,
4552                                               "space_info:enospc",
4553                                               space_info->flags, orig_bytes, 1);
4554         if (flushing) {
4555                 spin_lock(&space_info->lock);
4556                 space_info->flush = 0;
4557                 wake_up_all(&space_info->wait);
4558                 spin_unlock(&space_info->lock);
4559         }
4560         return ret;
4561 }
4562
4563 static struct btrfs_block_rsv *get_block_rsv(
4564                                         const struct btrfs_trans_handle *trans,
4565                                         const struct btrfs_root *root)
4566 {
4567         struct btrfs_block_rsv *block_rsv = NULL;
4568
4569         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4570                 block_rsv = trans->block_rsv;
4571
4572         if (root == root->fs_info->csum_root && trans->adding_csums)
4573                 block_rsv = trans->block_rsv;
4574
4575         if (root == root->fs_info->uuid_root)
4576                 block_rsv = trans->block_rsv;
4577
4578         if (!block_rsv)
4579                 block_rsv = root->block_rsv;
4580
4581         if (!block_rsv)
4582                 block_rsv = &root->fs_info->empty_block_rsv;
4583
4584         return block_rsv;
4585 }
4586
4587 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4588                                u64 num_bytes)
4589 {
4590         int ret = -ENOSPC;
4591         spin_lock(&block_rsv->lock);
4592         if (block_rsv->reserved >= num_bytes) {
4593                 block_rsv->reserved -= num_bytes;
4594                 if (block_rsv->reserved < block_rsv->size)
4595                         block_rsv->full = 0;
4596                 ret = 0;
4597         }
4598         spin_unlock(&block_rsv->lock);
4599         return ret;
4600 }
4601
4602 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4603                                 u64 num_bytes, int update_size)
4604 {
4605         spin_lock(&block_rsv->lock);
4606         block_rsv->reserved += num_bytes;
4607         if (update_size)
4608                 block_rsv->size += num_bytes;
4609         else if (block_rsv->reserved >= block_rsv->size)
4610                 block_rsv->full = 1;
4611         spin_unlock(&block_rsv->lock);
4612 }
4613
4614 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4615                              struct btrfs_block_rsv *dest, u64 num_bytes,
4616                              int min_factor)
4617 {
4618         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4619         u64 min_bytes;
4620
4621         if (global_rsv->space_info != dest->space_info)
4622                 return -ENOSPC;
4623
4624         spin_lock(&global_rsv->lock);
4625         min_bytes = div_factor(global_rsv->size, min_factor);
4626         if (global_rsv->reserved < min_bytes + num_bytes) {
4627                 spin_unlock(&global_rsv->lock);
4628                 return -ENOSPC;
4629         }
4630         global_rsv->reserved -= num_bytes;
4631         if (global_rsv->reserved < global_rsv->size)
4632                 global_rsv->full = 0;
4633         spin_unlock(&global_rsv->lock);
4634
4635         block_rsv_add_bytes(dest, num_bytes, 1);
4636         return 0;
4637 }
4638
4639 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4640                                     struct btrfs_block_rsv *block_rsv,
4641                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4642 {
4643         struct btrfs_space_info *space_info = block_rsv->space_info;
4644
4645         spin_lock(&block_rsv->lock);
4646         if (num_bytes == (u64)-1)
4647                 num_bytes = block_rsv->size;
4648         block_rsv->size -= num_bytes;
4649         if (block_rsv->reserved >= block_rsv->size) {
4650                 num_bytes = block_rsv->reserved - block_rsv->size;
4651                 block_rsv->reserved = block_rsv->size;
4652                 block_rsv->full = 1;
4653         } else {
4654                 num_bytes = 0;
4655         }
4656         spin_unlock(&block_rsv->lock);
4657
4658         if (num_bytes > 0) {
4659                 if (dest) {
4660                         spin_lock(&dest->lock);
4661                         if (!dest->full) {
4662                                 u64 bytes_to_add;
4663
4664                                 bytes_to_add = dest->size - dest->reserved;
4665                                 bytes_to_add = min(num_bytes, bytes_to_add);
4666                                 dest->reserved += bytes_to_add;
4667                                 if (dest->reserved >= dest->size)
4668                                         dest->full = 1;
4669                                 num_bytes -= bytes_to_add;
4670                         }
4671                         spin_unlock(&dest->lock);
4672                 }
4673                 if (num_bytes) {
4674                         spin_lock(&space_info->lock);
4675                         space_info->bytes_may_use -= num_bytes;
4676                         trace_btrfs_space_reservation(fs_info, "space_info",
4677                                         space_info->flags, num_bytes, 0);
4678                         spin_unlock(&space_info->lock);
4679                 }
4680         }
4681 }
4682
4683 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4684                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4685 {
4686         int ret;
4687
4688         ret = block_rsv_use_bytes(src, num_bytes);
4689         if (ret)
4690                 return ret;
4691
4692         block_rsv_add_bytes(dst, num_bytes, 1);
4693         return 0;
4694 }
4695
4696 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4697 {
4698         memset(rsv, 0, sizeof(*rsv));
4699         spin_lock_init(&rsv->lock);
4700         rsv->type = type;
4701 }
4702
4703 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4704                                               unsigned short type)
4705 {
4706         struct btrfs_block_rsv *block_rsv;
4707         struct btrfs_fs_info *fs_info = root->fs_info;
4708
4709         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4710         if (!block_rsv)
4711                 return NULL;
4712
4713         btrfs_init_block_rsv(block_rsv, type);
4714         block_rsv->space_info = __find_space_info(fs_info,
4715                                                   BTRFS_BLOCK_GROUP_METADATA);
4716         return block_rsv;
4717 }
4718
4719 void btrfs_free_block_rsv(struct btrfs_root *root,
4720                           struct btrfs_block_rsv *rsv)
4721 {
4722         if (!rsv)
4723                 return;
4724         btrfs_block_rsv_release(root, rsv, (u64)-1);
4725         kfree(rsv);
4726 }
4727
4728 int btrfs_block_rsv_add(struct btrfs_root *root,
4729                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4730                         enum btrfs_reserve_flush_enum flush)
4731 {
4732         int ret;
4733
4734         if (num_bytes == 0)
4735                 return 0;
4736
4737         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4738         if (!ret) {
4739                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4740                 return 0;
4741         }
4742
4743         return ret;
4744 }
4745
4746 int btrfs_block_rsv_check(struct btrfs_root *root,
4747                           struct btrfs_block_rsv *block_rsv, int min_factor)
4748 {
4749         u64 num_bytes = 0;
4750         int ret = -ENOSPC;
4751
4752         if (!block_rsv)
4753                 return 0;
4754
4755         spin_lock(&block_rsv->lock);
4756         num_bytes = div_factor(block_rsv->size, min_factor);
4757         if (block_rsv->reserved >= num_bytes)
4758                 ret = 0;
4759         spin_unlock(&block_rsv->lock);
4760
4761         return ret;
4762 }
4763
4764 int btrfs_block_rsv_refill(struct btrfs_root *root,
4765                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4766                            enum btrfs_reserve_flush_enum flush)
4767 {
4768         u64 num_bytes = 0;
4769         int ret = -ENOSPC;
4770
4771         if (!block_rsv)
4772                 return 0;
4773
4774         spin_lock(&block_rsv->lock);
4775         num_bytes = min_reserved;
4776         if (block_rsv->reserved >= num_bytes)
4777                 ret = 0;
4778         else
4779                 num_bytes -= block_rsv->reserved;
4780         spin_unlock(&block_rsv->lock);
4781
4782         if (!ret)
4783                 return 0;
4784
4785         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4786         if (!ret) {
4787                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4788                 return 0;
4789         }
4790
4791         return ret;
4792 }
4793
4794 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4795                             struct btrfs_block_rsv *dst_rsv,
4796                             u64 num_bytes)
4797 {
4798         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4799 }
4800
4801 void btrfs_block_rsv_release(struct btrfs_root *root,
4802                              struct btrfs_block_rsv *block_rsv,
4803                              u64 num_bytes)
4804 {
4805         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4806         if (global_rsv == block_rsv ||
4807             block_rsv->space_info != global_rsv->space_info)
4808                 global_rsv = NULL;
4809         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4810                                 num_bytes);
4811 }
4812
4813 /*
4814  * helper to calculate size of global block reservation.
4815  * the desired value is sum of space used by extent tree,
4816  * checksum tree and root tree
4817  */
4818 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4819 {
4820         struct btrfs_space_info *sinfo;
4821         u64 num_bytes;
4822         u64 meta_used;
4823         u64 data_used;
4824         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4825
4826         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4827         spin_lock(&sinfo->lock);
4828         data_used = sinfo->bytes_used;
4829         spin_unlock(&sinfo->lock);
4830
4831         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4832         spin_lock(&sinfo->lock);
4833         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4834                 data_used = 0;
4835         meta_used = sinfo->bytes_used;
4836         spin_unlock(&sinfo->lock);
4837
4838         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4839                     csum_size * 2;
4840         num_bytes += div64_u64(data_used + meta_used, 50);
4841
4842         if (num_bytes * 3 > meta_used)
4843                 num_bytes = div64_u64(meta_used, 3);
4844
4845         return ALIGN(num_bytes, fs_info->extent_root->nodesize << 10);
4846 }
4847
4848 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4849 {
4850         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4851         struct btrfs_space_info *sinfo = block_rsv->space_info;
4852         u64 num_bytes;
4853
4854         num_bytes = calc_global_metadata_size(fs_info);
4855
4856         spin_lock(&sinfo->lock);
4857         spin_lock(&block_rsv->lock);
4858
4859         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
4860
4861         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4862                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4863                     sinfo->bytes_may_use;
4864
4865         if (sinfo->total_bytes > num_bytes) {
4866                 num_bytes = sinfo->total_bytes - num_bytes;
4867                 block_rsv->reserved += num_bytes;
4868                 sinfo->bytes_may_use += num_bytes;
4869                 trace_btrfs_space_reservation(fs_info, "space_info",
4870                                       sinfo->flags, num_bytes, 1);
4871         }
4872
4873         if (block_rsv->reserved >= block_rsv->size) {
4874                 num_bytes = block_rsv->reserved - block_rsv->size;
4875                 sinfo->bytes_may_use -= num_bytes;
4876                 trace_btrfs_space_reservation(fs_info, "space_info",
4877                                       sinfo->flags, num_bytes, 0);
4878                 block_rsv->reserved = block_rsv->size;
4879                 block_rsv->full = 1;
4880         }
4881
4882         spin_unlock(&block_rsv->lock);
4883         spin_unlock(&sinfo->lock);
4884 }
4885
4886 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4887 {
4888         struct btrfs_space_info *space_info;
4889
4890         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4891         fs_info->chunk_block_rsv.space_info = space_info;
4892
4893         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4894         fs_info->global_block_rsv.space_info = space_info;
4895         fs_info->delalloc_block_rsv.space_info = space_info;
4896         fs_info->trans_block_rsv.space_info = space_info;
4897         fs_info->empty_block_rsv.space_info = space_info;
4898         fs_info->delayed_block_rsv.space_info = space_info;
4899
4900         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4901         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4902         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4903         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4904         if (fs_info->quota_root)
4905                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
4906         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4907
4908         update_global_block_rsv(fs_info);
4909 }
4910
4911 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4912 {
4913         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4914                                 (u64)-1);
4915         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4916         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4917         WARN_ON(fs_info->trans_block_rsv.size > 0);
4918         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4919         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4920         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4921         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4922         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4923 }
4924
4925 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4926                                   struct btrfs_root *root)
4927 {
4928         if (!trans->block_rsv)
4929                 return;
4930
4931         if (!trans->bytes_reserved)
4932                 return;
4933
4934         trace_btrfs_space_reservation(root->fs_info, "transaction",
4935                                       trans->transid, trans->bytes_reserved, 0);
4936         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4937         trans->bytes_reserved = 0;
4938 }
4939
4940 /* Can only return 0 or -ENOSPC */
4941 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4942                                   struct inode *inode)
4943 {
4944         struct btrfs_root *root = BTRFS_I(inode)->root;
4945         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4946         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4947
4948         /*
4949          * We need to hold space in order to delete our orphan item once we've
4950          * added it, so this takes the reservation so we can release it later
4951          * when we are truly done with the orphan item.
4952          */
4953         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4954         trace_btrfs_space_reservation(root->fs_info, "orphan",
4955                                       btrfs_ino(inode), num_bytes, 1);
4956         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4957 }
4958
4959 void btrfs_orphan_release_metadata(struct inode *inode)
4960 {
4961         struct btrfs_root *root = BTRFS_I(inode)->root;
4962         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4963         trace_btrfs_space_reservation(root->fs_info, "orphan",
4964                                       btrfs_ino(inode), num_bytes, 0);
4965         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4966 }
4967
4968 /*
4969  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
4970  * root: the root of the parent directory
4971  * rsv: block reservation
4972  * items: the number of items that we need do reservation
4973  * qgroup_reserved: used to return the reserved size in qgroup
4974  *
4975  * This function is used to reserve the space for snapshot/subvolume
4976  * creation and deletion. Those operations are different with the
4977  * common file/directory operations, they change two fs/file trees
4978  * and root tree, the number of items that the qgroup reserves is
4979  * different with the free space reservation. So we can not use
4980  * the space reseravtion mechanism in start_transaction().
4981  */
4982 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
4983                                      struct btrfs_block_rsv *rsv,
4984                                      int items,
4985                                      u64 *qgroup_reserved,
4986                                      bool use_global_rsv)
4987 {
4988         u64 num_bytes;
4989         int ret;
4990         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4991
4992         if (root->fs_info->quota_enabled) {
4993                 /* One for parent inode, two for dir entries */
4994                 num_bytes = 3 * root->nodesize;
4995                 ret = btrfs_qgroup_reserve(root, num_bytes);
4996                 if (ret)
4997                         return ret;
4998         } else {
4999                 num_bytes = 0;
5000         }
5001
5002         *qgroup_reserved = num_bytes;
5003
5004         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5005         rsv->space_info = __find_space_info(root->fs_info,
5006                                             BTRFS_BLOCK_GROUP_METADATA);
5007         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5008                                   BTRFS_RESERVE_FLUSH_ALL);
5009
5010         if (ret == -ENOSPC && use_global_rsv)
5011                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
5012
5013         if (ret) {
5014                 if (*qgroup_reserved)
5015                         btrfs_qgroup_free(root, *qgroup_reserved);
5016         }
5017
5018         return ret;
5019 }
5020
5021 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5022                                       struct btrfs_block_rsv *rsv,
5023                                       u64 qgroup_reserved)
5024 {
5025         btrfs_block_rsv_release(root, rsv, (u64)-1);
5026         if (qgroup_reserved)
5027                 btrfs_qgroup_free(root, qgroup_reserved);
5028 }
5029
5030 /**
5031  * drop_outstanding_extent - drop an outstanding extent
5032  * @inode: the inode we're dropping the extent for
5033  *
5034  * This is called when we are freeing up an outstanding extent, either called
5035  * after an error or after an extent is written.  This will return the number of
5036  * reserved extents that need to be freed.  This must be called with
5037  * BTRFS_I(inode)->lock held.
5038  */
5039 static unsigned drop_outstanding_extent(struct inode *inode)
5040 {
5041         unsigned drop_inode_space = 0;
5042         unsigned dropped_extents = 0;
5043
5044         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
5045         BTRFS_I(inode)->outstanding_extents--;
5046
5047         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5048             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5049                                &BTRFS_I(inode)->runtime_flags))
5050                 drop_inode_space = 1;
5051
5052         /*
5053          * If we have more or the same amount of outsanding extents than we have
5054          * reserved then we need to leave the reserved extents count alone.
5055          */
5056         if (BTRFS_I(inode)->outstanding_extents >=
5057             BTRFS_I(inode)->reserved_extents)
5058                 return drop_inode_space;
5059
5060         dropped_extents = BTRFS_I(inode)->reserved_extents -
5061                 BTRFS_I(inode)->outstanding_extents;
5062         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5063         return dropped_extents + drop_inode_space;
5064 }
5065
5066 /**
5067  * calc_csum_metadata_size - return the amount of metada space that must be
5068  *      reserved/free'd for the given bytes.
5069  * @inode: the inode we're manipulating
5070  * @num_bytes: the number of bytes in question
5071  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5072  *
5073  * This adjusts the number of csum_bytes in the inode and then returns the
5074  * correct amount of metadata that must either be reserved or freed.  We
5075  * calculate how many checksums we can fit into one leaf and then divide the
5076  * number of bytes that will need to be checksumed by this value to figure out
5077  * how many checksums will be required.  If we are adding bytes then the number
5078  * may go up and we will return the number of additional bytes that must be
5079  * reserved.  If it is going down we will return the number of bytes that must
5080  * be freed.
5081  *
5082  * This must be called with BTRFS_I(inode)->lock held.
5083  */
5084 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5085                                    int reserve)
5086 {
5087         struct btrfs_root *root = BTRFS_I(inode)->root;
5088         u64 csum_size;
5089         int num_csums_per_leaf;
5090         int num_csums;
5091         int old_csums;
5092
5093         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5094             BTRFS_I(inode)->csum_bytes == 0)
5095                 return 0;
5096
5097         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
5098         if (reserve)
5099                 BTRFS_I(inode)->csum_bytes += num_bytes;
5100         else
5101                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5102         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
5103         num_csums_per_leaf = (int)div64_u64(csum_size,
5104                                             sizeof(struct btrfs_csum_item) +
5105                                             sizeof(struct btrfs_disk_key));
5106         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
5107         num_csums = num_csums + num_csums_per_leaf - 1;
5108         num_csums = num_csums / num_csums_per_leaf;
5109
5110         old_csums = old_csums + num_csums_per_leaf - 1;
5111         old_csums = old_csums / num_csums_per_leaf;
5112
5113         /* No change, no need to reserve more */
5114         if (old_csums == num_csums)
5115                 return 0;
5116
5117         if (reserve)
5118                 return btrfs_calc_trans_metadata_size(root,
5119                                                       num_csums - old_csums);
5120
5121         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5122 }
5123
5124 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5125 {
5126         struct btrfs_root *root = BTRFS_I(inode)->root;
5127         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5128         u64 to_reserve = 0;
5129         u64 csum_bytes;
5130         unsigned nr_extents = 0;
5131         int extra_reserve = 0;
5132         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5133         int ret = 0;
5134         bool delalloc_lock = true;
5135         u64 to_free = 0;
5136         unsigned dropped;
5137
5138         /* If we are a free space inode we need to not flush since we will be in
5139          * the middle of a transaction commit.  We also don't need the delalloc
5140          * mutex since we won't race with anybody.  We need this mostly to make
5141          * lockdep shut its filthy mouth.
5142          */
5143         if (btrfs_is_free_space_inode(inode)) {
5144                 flush = BTRFS_RESERVE_NO_FLUSH;
5145                 delalloc_lock = false;
5146         }
5147
5148         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5149             btrfs_transaction_in_commit(root->fs_info))
5150                 schedule_timeout(1);
5151
5152         if (delalloc_lock)
5153                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5154
5155         num_bytes = ALIGN(num_bytes, root->sectorsize);
5156
5157         spin_lock(&BTRFS_I(inode)->lock);
5158         BTRFS_I(inode)->outstanding_extents++;
5159
5160         if (BTRFS_I(inode)->outstanding_extents >
5161             BTRFS_I(inode)->reserved_extents)
5162                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5163                         BTRFS_I(inode)->reserved_extents;
5164
5165         /*
5166          * Add an item to reserve for updating the inode when we complete the
5167          * delalloc io.
5168          */
5169         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5170                       &BTRFS_I(inode)->runtime_flags)) {
5171                 nr_extents++;
5172                 extra_reserve = 1;
5173         }
5174
5175         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5176         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5177         csum_bytes = BTRFS_I(inode)->csum_bytes;
5178         spin_unlock(&BTRFS_I(inode)->lock);
5179
5180         if (root->fs_info->quota_enabled) {
5181                 ret = btrfs_qgroup_reserve(root, num_bytes +
5182                                            nr_extents * root->nodesize);
5183                 if (ret)
5184                         goto out_fail;
5185         }
5186
5187         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5188         if (unlikely(ret)) {
5189                 if (root->fs_info->quota_enabled)
5190                         btrfs_qgroup_free(root, num_bytes +
5191                                                 nr_extents * root->nodesize);
5192                 goto out_fail;
5193         }
5194
5195         spin_lock(&BTRFS_I(inode)->lock);
5196         if (extra_reserve) {
5197                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5198                         &BTRFS_I(inode)->runtime_flags);
5199                 nr_extents--;
5200         }
5201         BTRFS_I(inode)->reserved_extents += nr_extents;
5202         spin_unlock(&BTRFS_I(inode)->lock);
5203
5204         if (delalloc_lock)
5205                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5206
5207         if (to_reserve)
5208                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5209                                               btrfs_ino(inode), to_reserve, 1);
5210         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5211
5212         return 0;
5213
5214 out_fail:
5215         spin_lock(&BTRFS_I(inode)->lock);
5216         dropped = drop_outstanding_extent(inode);
5217         /*
5218          * If the inodes csum_bytes is the same as the original
5219          * csum_bytes then we know we haven't raced with any free()ers
5220          * so we can just reduce our inodes csum bytes and carry on.
5221          */
5222         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5223                 calc_csum_metadata_size(inode, num_bytes, 0);
5224         } else {
5225                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5226                 u64 bytes;
5227
5228                 /*
5229                  * This is tricky, but first we need to figure out how much we
5230                  * free'd from any free-ers that occured during this
5231                  * reservation, so we reset ->csum_bytes to the csum_bytes
5232                  * before we dropped our lock, and then call the free for the
5233                  * number of bytes that were freed while we were trying our
5234                  * reservation.
5235                  */
5236                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5237                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5238                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5239
5240
5241                 /*
5242                  * Now we need to see how much we would have freed had we not
5243                  * been making this reservation and our ->csum_bytes were not
5244                  * artificially inflated.
5245                  */
5246                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5247                 bytes = csum_bytes - orig_csum_bytes;
5248                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5249
5250                 /*
5251                  * Now reset ->csum_bytes to what it should be.  If bytes is
5252                  * more than to_free then we would have free'd more space had we
5253                  * not had an artificially high ->csum_bytes, so we need to free
5254                  * the remainder.  If bytes is the same or less then we don't
5255                  * need to do anything, the other free-ers did the correct
5256                  * thing.
5257                  */
5258                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5259                 if (bytes > to_free)
5260                         to_free = bytes - to_free;
5261                 else
5262                         to_free = 0;
5263         }
5264         spin_unlock(&BTRFS_I(inode)->lock);
5265         if (dropped)
5266                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5267
5268         if (to_free) {
5269                 btrfs_block_rsv_release(root, block_rsv, to_free);
5270                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5271                                               btrfs_ino(inode), to_free, 0);
5272         }
5273         if (delalloc_lock)
5274                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5275         return ret;
5276 }
5277
5278 /**
5279  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5280  * @inode: the inode to release the reservation for
5281  * @num_bytes: the number of bytes we're releasing
5282  *
5283  * This will release the metadata reservation for an inode.  This can be called
5284  * once we complete IO for a given set of bytes to release their metadata
5285  * reservations.
5286  */
5287 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5288 {
5289         struct btrfs_root *root = BTRFS_I(inode)->root;
5290         u64 to_free = 0;
5291         unsigned dropped;
5292
5293         num_bytes = ALIGN(num_bytes, root->sectorsize);
5294         spin_lock(&BTRFS_I(inode)->lock);
5295         dropped = drop_outstanding_extent(inode);
5296
5297         if (num_bytes)
5298                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5299         spin_unlock(&BTRFS_I(inode)->lock);
5300         if (dropped > 0)
5301                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5302
5303         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5304                                       btrfs_ino(inode), to_free, 0);
5305         if (root->fs_info->quota_enabled) {
5306                 btrfs_qgroup_free(root, num_bytes +
5307                                         dropped * root->nodesize);
5308         }
5309
5310         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5311                                 to_free);
5312 }
5313
5314 /**
5315  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5316  * @inode: inode we're writing to
5317  * @num_bytes: the number of bytes we want to allocate
5318  *
5319  * This will do the following things
5320  *
5321  * o reserve space in the data space info for num_bytes
5322  * o reserve space in the metadata space info based on number of outstanding
5323  *   extents and how much csums will be needed
5324  * o add to the inodes ->delalloc_bytes
5325  * o add it to the fs_info's delalloc inodes list.
5326  *
5327  * This will return 0 for success and -ENOSPC if there is no space left.
5328  */
5329 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5330 {
5331         int ret;
5332
5333         ret = btrfs_check_data_free_space(inode, num_bytes);
5334         if (ret)
5335                 return ret;
5336
5337         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5338         if (ret) {
5339                 btrfs_free_reserved_data_space(inode, num_bytes);
5340                 return ret;
5341         }
5342
5343         return 0;
5344 }
5345
5346 /**
5347  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5348  * @inode: inode we're releasing space for
5349  * @num_bytes: the number of bytes we want to free up
5350  *
5351  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5352  * called in the case that we don't need the metadata AND data reservations
5353  * anymore.  So if there is an error or we insert an inline extent.
5354  *
5355  * This function will release the metadata space that was not used and will
5356  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5357  * list if there are no delalloc bytes left.
5358  */
5359 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5360 {
5361         btrfs_delalloc_release_metadata(inode, num_bytes);
5362         btrfs_free_reserved_data_space(inode, num_bytes);
5363 }
5364
5365 static int update_block_group(struct btrfs_root *root,
5366                               u64 bytenr, u64 num_bytes, int alloc)
5367 {
5368         struct btrfs_block_group_cache *cache = NULL;
5369         struct btrfs_fs_info *info = root->fs_info;
5370         u64 total = num_bytes;
5371         u64 old_val;
5372         u64 byte_in_group;
5373         int factor;
5374
5375         /* block accounting for super block */
5376         spin_lock(&info->delalloc_root_lock);
5377         old_val = btrfs_super_bytes_used(info->super_copy);
5378         if (alloc)
5379                 old_val += num_bytes;
5380         else
5381                 old_val -= num_bytes;
5382         btrfs_set_super_bytes_used(info->super_copy, old_val);
5383         spin_unlock(&info->delalloc_root_lock);
5384
5385         while (total) {
5386                 cache = btrfs_lookup_block_group(info, bytenr);
5387                 if (!cache)
5388                         return -ENOENT;
5389                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5390                                     BTRFS_BLOCK_GROUP_RAID1 |
5391                                     BTRFS_BLOCK_GROUP_RAID10))
5392                         factor = 2;
5393                 else
5394                         factor = 1;
5395                 /*
5396                  * If this block group has free space cache written out, we
5397                  * need to make sure to load it if we are removing space.  This
5398                  * is because we need the unpinning stage to actually add the
5399                  * space back to the block group, otherwise we will leak space.
5400                  */
5401                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5402                         cache_block_group(cache, 1);
5403
5404                 byte_in_group = bytenr - cache->key.objectid;
5405                 WARN_ON(byte_in_group > cache->key.offset);
5406
5407                 spin_lock(&cache->space_info->lock);
5408                 spin_lock(&cache->lock);
5409
5410                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5411                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5412                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5413
5414                 cache->dirty = 1;
5415                 old_val = btrfs_block_group_used(&cache->item);
5416                 num_bytes = min(total, cache->key.offset - byte_in_group);
5417                 if (alloc) {
5418                         old_val += num_bytes;
5419                         btrfs_set_block_group_used(&cache->item, old_val);
5420                         cache->reserved -= num_bytes;
5421                         cache->space_info->bytes_reserved -= num_bytes;
5422                         cache->space_info->bytes_used += num_bytes;
5423                         cache->space_info->disk_used += num_bytes * factor;
5424                         spin_unlock(&cache->lock);
5425                         spin_unlock(&cache->space_info->lock);
5426                 } else {
5427                         old_val -= num_bytes;
5428
5429                         /*
5430                          * No longer have used bytes in this block group, queue
5431                          * it for deletion.
5432                          */
5433                         if (old_val == 0) {
5434                                 spin_lock(&info->unused_bgs_lock);
5435                                 if (list_empty(&cache->bg_list)) {
5436                                         btrfs_get_block_group(cache);
5437                                         list_add_tail(&cache->bg_list,
5438                                                       &info->unused_bgs);
5439                                 }
5440                                 spin_unlock(&info->unused_bgs_lock);
5441                         }
5442                         btrfs_set_block_group_used(&cache->item, old_val);
5443                         cache->pinned += num_bytes;
5444                         cache->space_info->bytes_pinned += num_bytes;
5445                         cache->space_info->bytes_used -= num_bytes;
5446                         cache->space_info->disk_used -= num_bytes * factor;
5447                         spin_unlock(&cache->lock);
5448                         spin_unlock(&cache->space_info->lock);
5449
5450                         set_extent_dirty(info->pinned_extents,
5451                                          bytenr, bytenr + num_bytes - 1,
5452                                          GFP_NOFS | __GFP_NOFAIL);
5453                 }
5454                 btrfs_put_block_group(cache);
5455                 total -= num_bytes;
5456                 bytenr += num_bytes;
5457         }
5458         return 0;
5459 }
5460
5461 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5462 {
5463         struct btrfs_block_group_cache *cache;
5464         u64 bytenr;
5465
5466         spin_lock(&root->fs_info->block_group_cache_lock);
5467         bytenr = root->fs_info->first_logical_byte;
5468         spin_unlock(&root->fs_info->block_group_cache_lock);
5469
5470         if (bytenr < (u64)-1)
5471                 return bytenr;
5472
5473         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5474         if (!cache)
5475                 return 0;
5476
5477         bytenr = cache->key.objectid;
5478         btrfs_put_block_group(cache);
5479
5480         return bytenr;
5481 }
5482
5483 static int pin_down_extent(struct btrfs_root *root,
5484                            struct btrfs_block_group_cache *cache,
5485                            u64 bytenr, u64 num_bytes, int reserved)
5486 {
5487         spin_lock(&cache->space_info->lock);
5488         spin_lock(&cache->lock);
5489         cache->pinned += num_bytes;
5490         cache->space_info->bytes_pinned += num_bytes;
5491         if (reserved) {
5492                 cache->reserved -= num_bytes;
5493                 cache->space_info->bytes_reserved -= num_bytes;
5494         }
5495         spin_unlock(&cache->lock);
5496         spin_unlock(&cache->space_info->lock);
5497
5498         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5499                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5500         if (reserved)
5501                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
5502         return 0;
5503 }
5504
5505 /*
5506  * this function must be called within transaction
5507  */
5508 int btrfs_pin_extent(struct btrfs_root *root,
5509                      u64 bytenr, u64 num_bytes, int reserved)
5510 {
5511         struct btrfs_block_group_cache *cache;
5512
5513         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5514         BUG_ON(!cache); /* Logic error */
5515
5516         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5517
5518         btrfs_put_block_group(cache);
5519         return 0;
5520 }
5521
5522 /*
5523  * this function must be called within transaction
5524  */
5525 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5526                                     u64 bytenr, u64 num_bytes)
5527 {
5528         struct btrfs_block_group_cache *cache;
5529         int ret;
5530
5531         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5532         if (!cache)
5533                 return -EINVAL;
5534
5535         /*
5536          * pull in the free space cache (if any) so that our pin
5537          * removes the free space from the cache.  We have load_only set
5538          * to one because the slow code to read in the free extents does check
5539          * the pinned extents.
5540          */
5541         cache_block_group(cache, 1);
5542
5543         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5544
5545         /* remove us from the free space cache (if we're there at all) */
5546         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5547         btrfs_put_block_group(cache);
5548         return ret;
5549 }
5550
5551 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5552 {
5553         int ret;
5554         struct btrfs_block_group_cache *block_group;
5555         struct btrfs_caching_control *caching_ctl;
5556
5557         block_group = btrfs_lookup_block_group(root->fs_info, start);
5558         if (!block_group)
5559                 return -EINVAL;
5560
5561         cache_block_group(block_group, 0);
5562         caching_ctl = get_caching_control(block_group);
5563
5564         if (!caching_ctl) {
5565                 /* Logic error */
5566                 BUG_ON(!block_group_cache_done(block_group));
5567                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5568         } else {
5569                 mutex_lock(&caching_ctl->mutex);
5570
5571                 if (start >= caching_ctl->progress) {
5572                         ret = add_excluded_extent(root, start, num_bytes);
5573                 } else if (start + num_bytes <= caching_ctl->progress) {
5574                         ret = btrfs_remove_free_space(block_group,
5575                                                       start, num_bytes);
5576                 } else {
5577                         num_bytes = caching_ctl->progress - start;
5578                         ret = btrfs_remove_free_space(block_group,
5579                                                       start, num_bytes);
5580                         if (ret)
5581                                 goto out_lock;
5582
5583                         num_bytes = (start + num_bytes) -
5584                                 caching_ctl->progress;
5585                         start = caching_ctl->progress;
5586                         ret = add_excluded_extent(root, start, num_bytes);
5587                 }
5588 out_lock:
5589                 mutex_unlock(&caching_ctl->mutex);
5590                 put_caching_control(caching_ctl);
5591         }
5592         btrfs_put_block_group(block_group);
5593         return ret;
5594 }
5595
5596 int btrfs_exclude_logged_extents(struct btrfs_root *log,
5597                                  struct extent_buffer *eb)
5598 {
5599         struct btrfs_file_extent_item *item;
5600         struct btrfs_key key;
5601         int found_type;
5602         int i;
5603
5604         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
5605                 return 0;
5606
5607         for (i = 0; i < btrfs_header_nritems(eb); i++) {
5608                 btrfs_item_key_to_cpu(eb, &key, i);
5609                 if (key.type != BTRFS_EXTENT_DATA_KEY)
5610                         continue;
5611                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
5612                 found_type = btrfs_file_extent_type(eb, item);
5613                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
5614                         continue;
5615                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
5616                         continue;
5617                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
5618                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
5619                 __exclude_logged_extent(log, key.objectid, key.offset);
5620         }
5621
5622         return 0;
5623 }
5624
5625 /**
5626  * btrfs_update_reserved_bytes - update the block_group and space info counters
5627  * @cache:      The cache we are manipulating
5628  * @num_bytes:  The number of bytes in question
5629  * @reserve:    One of the reservation enums
5630  * @delalloc:   The blocks are allocated for the delalloc write
5631  *
5632  * This is called by the allocator when it reserves space, or by somebody who is
5633  * freeing space that was never actually used on disk.  For example if you
5634  * reserve some space for a new leaf in transaction A and before transaction A
5635  * commits you free that leaf, you call this with reserve set to 0 in order to
5636  * clear the reservation.
5637  *
5638  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5639  * ENOSPC accounting.  For data we handle the reservation through clearing the
5640  * delalloc bits in the io_tree.  We have to do this since we could end up
5641  * allocating less disk space for the amount of data we have reserved in the
5642  * case of compression.
5643  *
5644  * If this is a reservation and the block group has become read only we cannot
5645  * make the reservation and return -EAGAIN, otherwise this function always
5646  * succeeds.
5647  */
5648 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5649                                        u64 num_bytes, int reserve, int delalloc)
5650 {
5651         struct btrfs_space_info *space_info = cache->space_info;
5652         int ret = 0;
5653
5654         spin_lock(&space_info->lock);
5655         spin_lock(&cache->lock);
5656         if (reserve != RESERVE_FREE) {
5657                 if (cache->ro) {
5658                         ret = -EAGAIN;
5659                 } else {
5660                         cache->reserved += num_bytes;
5661                         space_info->bytes_reserved += num_bytes;
5662                         if (reserve == RESERVE_ALLOC) {
5663                                 trace_btrfs_space_reservation(cache->fs_info,
5664                                                 "space_info", space_info->flags,
5665                                                 num_bytes, 0);
5666                                 space_info->bytes_may_use -= num_bytes;
5667                         }
5668
5669                         if (delalloc)
5670                                 cache->delalloc_bytes += num_bytes;
5671                 }
5672         } else {
5673                 if (cache->ro)
5674                         space_info->bytes_readonly += num_bytes;
5675                 cache->reserved -= num_bytes;
5676                 space_info->bytes_reserved -= num_bytes;
5677
5678                 if (delalloc)
5679                         cache->delalloc_bytes -= num_bytes;
5680         }
5681         spin_unlock(&cache->lock);
5682         spin_unlock(&space_info->lock);
5683         return ret;
5684 }
5685
5686 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5687                                 struct btrfs_root *root)
5688 {
5689         struct btrfs_fs_info *fs_info = root->fs_info;
5690         struct btrfs_caching_control *next;
5691         struct btrfs_caching_control *caching_ctl;
5692         struct btrfs_block_group_cache *cache;
5693
5694         down_write(&fs_info->commit_root_sem);
5695
5696         list_for_each_entry_safe(caching_ctl, next,
5697                                  &fs_info->caching_block_groups, list) {
5698                 cache = caching_ctl->block_group;
5699                 if (block_group_cache_done(cache)) {
5700                         cache->last_byte_to_unpin = (u64)-1;
5701                         list_del_init(&caching_ctl->list);
5702                         put_caching_control(caching_ctl);
5703                 } else {
5704                         cache->last_byte_to_unpin = caching_ctl->progress;
5705                 }
5706         }
5707
5708         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5709                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5710         else
5711                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5712
5713         up_write(&fs_info->commit_root_sem);
5714
5715         update_global_block_rsv(fs_info);
5716 }
5717
5718 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5719 {
5720         struct btrfs_fs_info *fs_info = root->fs_info;
5721         struct btrfs_block_group_cache *cache = NULL;
5722         struct btrfs_space_info *space_info;
5723         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5724         u64 len;
5725         bool readonly;
5726
5727         while (start <= end) {
5728                 readonly = false;
5729                 if (!cache ||
5730                     start >= cache->key.objectid + cache->key.offset) {
5731                         if (cache)
5732                                 btrfs_put_block_group(cache);
5733                         cache = btrfs_lookup_block_group(fs_info, start);
5734                         BUG_ON(!cache); /* Logic error */
5735                 }
5736
5737                 len = cache->key.objectid + cache->key.offset - start;
5738                 len = min(len, end + 1 - start);
5739
5740                 if (start < cache->last_byte_to_unpin) {
5741                         len = min(len, cache->last_byte_to_unpin - start);
5742                         btrfs_add_free_space(cache, start, len);
5743                 }
5744
5745                 start += len;
5746                 space_info = cache->space_info;
5747
5748                 spin_lock(&space_info->lock);
5749                 spin_lock(&cache->lock);
5750                 cache->pinned -= len;
5751                 space_info->bytes_pinned -= len;
5752                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
5753                 if (cache->ro) {
5754                         space_info->bytes_readonly += len;
5755                         readonly = true;
5756                 }
5757                 spin_unlock(&cache->lock);
5758                 if (!readonly && global_rsv->space_info == space_info) {
5759                         spin_lock(&global_rsv->lock);
5760                         if (!global_rsv->full) {
5761                                 len = min(len, global_rsv->size -
5762                                           global_rsv->reserved);
5763                                 global_rsv->reserved += len;
5764                                 space_info->bytes_may_use += len;
5765                                 if (global_rsv->reserved >= global_rsv->size)
5766                                         global_rsv->full = 1;
5767                         }
5768                         spin_unlock(&global_rsv->lock);
5769                 }
5770                 spin_unlock(&space_info->lock);
5771         }
5772
5773         if (cache)
5774                 btrfs_put_block_group(cache);
5775         return 0;
5776 }
5777
5778 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5779                                struct btrfs_root *root)
5780 {
5781         struct btrfs_fs_info *fs_info = root->fs_info;
5782         struct extent_io_tree *unpin;
5783         u64 start;
5784         u64 end;
5785         int ret;
5786
5787         if (trans->aborted)
5788                 return 0;
5789
5790         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5791                 unpin = &fs_info->freed_extents[1];
5792         else
5793                 unpin = &fs_info->freed_extents[0];
5794
5795         while (1) {
5796                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5797                                             EXTENT_DIRTY, NULL);
5798                 if (ret)
5799                         break;
5800
5801                 if (btrfs_test_opt(root, DISCARD))
5802                         ret = btrfs_discard_extent(root, start,
5803                                                    end + 1 - start, NULL);
5804
5805                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5806                 unpin_extent_range(root, start, end);
5807                 cond_resched();
5808         }
5809
5810         return 0;
5811 }
5812
5813 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
5814                              u64 owner, u64 root_objectid)
5815 {
5816         struct btrfs_space_info *space_info;
5817         u64 flags;
5818
5819         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5820                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
5821                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
5822                 else
5823                         flags = BTRFS_BLOCK_GROUP_METADATA;
5824         } else {
5825                 flags = BTRFS_BLOCK_GROUP_DATA;
5826         }
5827
5828         space_info = __find_space_info(fs_info, flags);
5829         BUG_ON(!space_info); /* Logic bug */
5830         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
5831 }
5832
5833
5834 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5835                                 struct btrfs_root *root,
5836                                 u64 bytenr, u64 num_bytes, u64 parent,
5837                                 u64 root_objectid, u64 owner_objectid,
5838                                 u64 owner_offset, int refs_to_drop,
5839                                 struct btrfs_delayed_extent_op *extent_op,
5840                                 int no_quota)
5841 {
5842         struct btrfs_key key;
5843         struct btrfs_path *path;
5844         struct btrfs_fs_info *info = root->fs_info;
5845         struct btrfs_root *extent_root = info->extent_root;
5846         struct extent_buffer *leaf;
5847         struct btrfs_extent_item *ei;
5848         struct btrfs_extent_inline_ref *iref;
5849         int ret;
5850         int is_data;
5851         int extent_slot = 0;
5852         int found_extent = 0;
5853         int num_to_del = 1;
5854         u32 item_size;
5855         u64 refs;
5856         int last_ref = 0;
5857         enum btrfs_qgroup_operation_type type = BTRFS_QGROUP_OPER_SUB_EXCL;
5858         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
5859                                                  SKINNY_METADATA);
5860
5861         if (!info->quota_enabled || !is_fstree(root_objectid))
5862                 no_quota = 1;
5863
5864         path = btrfs_alloc_path();
5865         if (!path)
5866                 return -ENOMEM;
5867
5868         path->reada = 1;
5869         path->leave_spinning = 1;
5870
5871         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5872         BUG_ON(!is_data && refs_to_drop != 1);
5873
5874         if (is_data)
5875                 skinny_metadata = 0;
5876
5877         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5878                                     bytenr, num_bytes, parent,
5879                                     root_objectid, owner_objectid,
5880                                     owner_offset);
5881         if (ret == 0) {
5882                 extent_slot = path->slots[0];
5883                 while (extent_slot >= 0) {
5884                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5885                                               extent_slot);
5886                         if (key.objectid != bytenr)
5887                                 break;
5888                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5889                             key.offset == num_bytes) {
5890                                 found_extent = 1;
5891                                 break;
5892                         }
5893                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
5894                             key.offset == owner_objectid) {
5895                                 found_extent = 1;
5896                                 break;
5897                         }
5898                         if (path->slots[0] - extent_slot > 5)
5899                                 break;
5900                         extent_slot--;
5901                 }
5902 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5903                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5904                 if (found_extent && item_size < sizeof(*ei))
5905                         found_extent = 0;
5906 #endif
5907                 if (!found_extent) {
5908                         BUG_ON(iref);
5909                         ret = remove_extent_backref(trans, extent_root, path,
5910                                                     NULL, refs_to_drop,
5911                                                     is_data, &last_ref);
5912                         if (ret) {
5913                                 btrfs_abort_transaction(trans, extent_root, ret);
5914                                 goto out;
5915                         }
5916                         btrfs_release_path(path);
5917                         path->leave_spinning = 1;
5918
5919                         key.objectid = bytenr;
5920                         key.type = BTRFS_EXTENT_ITEM_KEY;
5921                         key.offset = num_bytes;
5922
5923                         if (!is_data && skinny_metadata) {
5924                                 key.type = BTRFS_METADATA_ITEM_KEY;
5925                                 key.offset = owner_objectid;
5926                         }
5927
5928                         ret = btrfs_search_slot(trans, extent_root,
5929                                                 &key, path, -1, 1);
5930                         if (ret > 0 && skinny_metadata && path->slots[0]) {
5931                                 /*
5932                                  * Couldn't find our skinny metadata item,
5933                                  * see if we have ye olde extent item.
5934                                  */
5935                                 path->slots[0]--;
5936                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
5937                                                       path->slots[0]);
5938                                 if (key.objectid == bytenr &&
5939                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
5940                                     key.offset == num_bytes)
5941                                         ret = 0;
5942                         }
5943
5944                         if (ret > 0 && skinny_metadata) {
5945                                 skinny_metadata = false;
5946                                 key.objectid = bytenr;
5947                                 key.type = BTRFS_EXTENT_ITEM_KEY;
5948                                 key.offset = num_bytes;
5949                                 btrfs_release_path(path);
5950                                 ret = btrfs_search_slot(trans, extent_root,
5951                                                         &key, path, -1, 1);
5952                         }
5953
5954                         if (ret) {
5955                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5956                                         ret, bytenr);
5957                                 if (ret > 0)
5958                                         btrfs_print_leaf(extent_root,
5959                                                          path->nodes[0]);
5960                         }
5961                         if (ret < 0) {
5962                                 btrfs_abort_transaction(trans, extent_root, ret);
5963                                 goto out;
5964                         }
5965                         extent_slot = path->slots[0];
5966                 }
5967         } else if (WARN_ON(ret == -ENOENT)) {
5968                 btrfs_print_leaf(extent_root, path->nodes[0]);
5969                 btrfs_err(info,
5970                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
5971                         bytenr, parent, root_objectid, owner_objectid,
5972                         owner_offset);
5973                 btrfs_abort_transaction(trans, extent_root, ret);
5974                 goto out;
5975         } else {
5976                 btrfs_abort_transaction(trans, extent_root, ret);
5977                 goto out;
5978         }
5979
5980         leaf = path->nodes[0];
5981         item_size = btrfs_item_size_nr(leaf, extent_slot);
5982 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5983         if (item_size < sizeof(*ei)) {
5984                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5985                 ret = convert_extent_item_v0(trans, extent_root, path,
5986                                              owner_objectid, 0);
5987                 if (ret < 0) {
5988                         btrfs_abort_transaction(trans, extent_root, ret);
5989                         goto out;
5990                 }
5991
5992                 btrfs_release_path(path);
5993                 path->leave_spinning = 1;
5994
5995                 key.objectid = bytenr;
5996                 key.type = BTRFS_EXTENT_ITEM_KEY;
5997                 key.offset = num_bytes;
5998
5999                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6000                                         -1, 1);
6001                 if (ret) {
6002                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6003                                 ret, bytenr);
6004                         btrfs_print_leaf(extent_root, path->nodes[0]);
6005                 }
6006                 if (ret < 0) {
6007                         btrfs_abort_transaction(trans, extent_root, ret);
6008                         goto out;
6009                 }
6010
6011                 extent_slot = path->slots[0];
6012                 leaf = path->nodes[0];
6013                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6014         }
6015 #endif
6016         BUG_ON(item_size < sizeof(*ei));
6017         ei = btrfs_item_ptr(leaf, extent_slot,
6018                             struct btrfs_extent_item);
6019         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6020             key.type == BTRFS_EXTENT_ITEM_KEY) {
6021                 struct btrfs_tree_block_info *bi;
6022                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6023                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6024                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6025         }
6026
6027         refs = btrfs_extent_refs(leaf, ei);
6028         if (refs < refs_to_drop) {
6029                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6030                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6031                 ret = -EINVAL;
6032                 btrfs_abort_transaction(trans, extent_root, ret);
6033                 goto out;
6034         }
6035         refs -= refs_to_drop;
6036
6037         if (refs > 0) {
6038                 type = BTRFS_QGROUP_OPER_SUB_SHARED;
6039                 if (extent_op)
6040                         __run_delayed_extent_op(extent_op, leaf, ei);
6041                 /*
6042                  * In the case of inline back ref, reference count will
6043                  * be updated by remove_extent_backref
6044                  */
6045                 if (iref) {
6046                         BUG_ON(!found_extent);
6047                 } else {
6048                         btrfs_set_extent_refs(leaf, ei, refs);
6049                         btrfs_mark_buffer_dirty(leaf);
6050                 }
6051                 if (found_extent) {
6052                         ret = remove_extent_backref(trans, extent_root, path,
6053                                                     iref, refs_to_drop,
6054                                                     is_data, &last_ref);
6055                         if (ret) {
6056                                 btrfs_abort_transaction(trans, extent_root, ret);
6057                                 goto out;
6058                         }
6059                 }
6060                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6061                                  root_objectid);
6062         } else {
6063                 if (found_extent) {
6064                         BUG_ON(is_data && refs_to_drop !=
6065                                extent_data_ref_count(root, path, iref));
6066                         if (iref) {
6067                                 BUG_ON(path->slots[0] != extent_slot);
6068                         } else {
6069                                 BUG_ON(path->slots[0] != extent_slot + 1);
6070                                 path->slots[0] = extent_slot;
6071                                 num_to_del = 2;
6072                         }
6073                 }
6074
6075                 last_ref = 1;
6076                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6077                                       num_to_del);
6078                 if (ret) {
6079                         btrfs_abort_transaction(trans, extent_root, ret);
6080                         goto out;
6081                 }
6082                 btrfs_release_path(path);
6083
6084                 if (is_data) {
6085                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6086                         if (ret) {
6087                                 btrfs_abort_transaction(trans, extent_root, ret);
6088                                 goto out;
6089                         }
6090                 }
6091
6092                 ret = update_block_group(root, bytenr, num_bytes, 0);
6093                 if (ret) {
6094                         btrfs_abort_transaction(trans, extent_root, ret);
6095                         goto out;
6096                 }
6097         }
6098         btrfs_release_path(path);
6099
6100         /* Deal with the quota accounting */
6101         if (!ret && last_ref && !no_quota) {
6102                 int mod_seq = 0;
6103
6104                 if (owner_objectid >= BTRFS_FIRST_FREE_OBJECTID &&
6105                     type == BTRFS_QGROUP_OPER_SUB_SHARED)
6106                         mod_seq = 1;
6107
6108                 ret = btrfs_qgroup_record_ref(trans, info, root_objectid,
6109                                               bytenr, num_bytes, type,
6110                                               mod_seq);
6111         }
6112 out:
6113         btrfs_free_path(path);
6114         return ret;
6115 }
6116
6117 /*
6118  * when we free an block, it is possible (and likely) that we free the last
6119  * delayed ref for that extent as well.  This searches the delayed ref tree for
6120  * a given extent, and if there are no other delayed refs to be processed, it
6121  * removes it from the tree.
6122  */
6123 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6124                                       struct btrfs_root *root, u64 bytenr)
6125 {
6126         struct btrfs_delayed_ref_head *head;
6127         struct btrfs_delayed_ref_root *delayed_refs;
6128         int ret = 0;
6129
6130         delayed_refs = &trans->transaction->delayed_refs;
6131         spin_lock(&delayed_refs->lock);
6132         head = btrfs_find_delayed_ref_head(trans, bytenr);
6133         if (!head)
6134                 goto out_delayed_unlock;
6135
6136         spin_lock(&head->lock);
6137         if (rb_first(&head->ref_root))
6138                 goto out;
6139
6140         if (head->extent_op) {
6141                 if (!head->must_insert_reserved)
6142                         goto out;
6143                 btrfs_free_delayed_extent_op(head->extent_op);
6144                 head->extent_op = NULL;
6145         }
6146
6147         /*
6148          * waiting for the lock here would deadlock.  If someone else has it
6149          * locked they are already in the process of dropping it anyway
6150          */
6151         if (!mutex_trylock(&head->mutex))
6152                 goto out;
6153
6154         /*
6155          * at this point we have a head with no other entries.  Go
6156          * ahead and process it.
6157          */
6158         head->node.in_tree = 0;
6159         rb_erase(&head->href_node, &delayed_refs->href_root);
6160
6161         atomic_dec(&delayed_refs->num_entries);
6162
6163         /*
6164          * we don't take a ref on the node because we're removing it from the
6165          * tree, so we just steal the ref the tree was holding.
6166          */
6167         delayed_refs->num_heads--;
6168         if (head->processing == 0)
6169                 delayed_refs->num_heads_ready--;
6170         head->processing = 0;
6171         spin_unlock(&head->lock);
6172         spin_unlock(&delayed_refs->lock);
6173
6174         BUG_ON(head->extent_op);
6175         if (head->must_insert_reserved)
6176                 ret = 1;
6177
6178         mutex_unlock(&head->mutex);
6179         btrfs_put_delayed_ref(&head->node);
6180         return ret;
6181 out:
6182         spin_unlock(&head->lock);
6183
6184 out_delayed_unlock:
6185         spin_unlock(&delayed_refs->lock);
6186         return 0;
6187 }
6188
6189 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6190                            struct btrfs_root *root,
6191                            struct extent_buffer *buf,
6192                            u64 parent, int last_ref)
6193 {
6194         struct btrfs_block_group_cache *cache = NULL;
6195         int pin = 1;
6196         int ret;
6197
6198         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6199                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6200                                         buf->start, buf->len,
6201                                         parent, root->root_key.objectid,
6202                                         btrfs_header_level(buf),
6203                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
6204                 BUG_ON(ret); /* -ENOMEM */
6205         }
6206
6207         if (!last_ref)
6208                 return;
6209
6210         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6211
6212         if (btrfs_header_generation(buf) == trans->transid) {
6213                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6214                         ret = check_ref_cleanup(trans, root, buf->start);
6215                         if (!ret)
6216                                 goto out;
6217                 }
6218
6219                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6220                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6221                         goto out;
6222                 }
6223
6224                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6225
6226                 btrfs_add_free_space(cache, buf->start, buf->len);
6227                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
6228                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6229                 pin = 0;
6230         }
6231 out:
6232         if (pin)
6233                 add_pinned_bytes(root->fs_info, buf->len,
6234                                  btrfs_header_level(buf),
6235                                  root->root_key.objectid);
6236
6237         /*
6238          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6239          * anymore.
6240          */
6241         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6242         btrfs_put_block_group(cache);
6243 }
6244
6245 /* Can return -ENOMEM */
6246 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6247                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6248                       u64 owner, u64 offset, int no_quota)
6249 {
6250         int ret;
6251         struct btrfs_fs_info *fs_info = root->fs_info;
6252
6253         if (btrfs_test_is_dummy_root(root))
6254                 return 0;
6255
6256         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6257
6258         /*
6259          * tree log blocks never actually go into the extent allocation
6260          * tree, just update pinning info and exit early.
6261          */
6262         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6263                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6264                 /* unlocks the pinned mutex */
6265                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6266                 ret = 0;
6267         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6268                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6269                                         num_bytes,
6270                                         parent, root_objectid, (int)owner,
6271                                         BTRFS_DROP_DELAYED_REF, NULL, no_quota);
6272         } else {
6273                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6274                                                 num_bytes,
6275                                                 parent, root_objectid, owner,
6276                                                 offset, BTRFS_DROP_DELAYED_REF,
6277                                                 NULL, no_quota);
6278         }
6279         return ret;
6280 }
6281
6282 /*
6283  * when we wait for progress in the block group caching, its because
6284  * our allocation attempt failed at least once.  So, we must sleep
6285  * and let some progress happen before we try again.
6286  *
6287  * This function will sleep at least once waiting for new free space to
6288  * show up, and then it will check the block group free space numbers
6289  * for our min num_bytes.  Another option is to have it go ahead
6290  * and look in the rbtree for a free extent of a given size, but this
6291  * is a good start.
6292  *
6293  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6294  * any of the information in this block group.
6295  */
6296 static noinline void
6297 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6298                                 u64 num_bytes)
6299 {
6300         struct btrfs_caching_control *caching_ctl;
6301
6302         caching_ctl = get_caching_control(cache);
6303         if (!caching_ctl)
6304                 return;
6305
6306         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6307                    (cache->free_space_ctl->free_space >= num_bytes));
6308
6309         put_caching_control(caching_ctl);
6310 }
6311
6312 static noinline int
6313 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6314 {
6315         struct btrfs_caching_control *caching_ctl;
6316         int ret = 0;
6317
6318         caching_ctl = get_caching_control(cache);
6319         if (!caching_ctl)
6320                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6321
6322         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6323         if (cache->cached == BTRFS_CACHE_ERROR)
6324                 ret = -EIO;
6325         put_caching_control(caching_ctl);
6326         return ret;
6327 }
6328
6329 int __get_raid_index(u64 flags)
6330 {
6331         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6332                 return BTRFS_RAID_RAID10;
6333         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6334                 return BTRFS_RAID_RAID1;
6335         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6336                 return BTRFS_RAID_DUP;
6337         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6338                 return BTRFS_RAID_RAID0;
6339         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6340                 return BTRFS_RAID_RAID5;
6341         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6342                 return BTRFS_RAID_RAID6;
6343
6344         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6345 }
6346
6347 int get_block_group_index(struct btrfs_block_group_cache *cache)
6348 {
6349         return __get_raid_index(cache->flags);
6350 }
6351
6352 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6353         [BTRFS_RAID_RAID10]     = "raid10",
6354         [BTRFS_RAID_RAID1]      = "raid1",
6355         [BTRFS_RAID_DUP]        = "dup",
6356         [BTRFS_RAID_RAID0]      = "raid0",
6357         [BTRFS_RAID_SINGLE]     = "single",
6358         [BTRFS_RAID_RAID5]      = "raid5",
6359         [BTRFS_RAID_RAID6]      = "raid6",
6360 };
6361
6362 static const char *get_raid_name(enum btrfs_raid_types type)
6363 {
6364         if (type >= BTRFS_NR_RAID_TYPES)
6365                 return NULL;
6366
6367         return btrfs_raid_type_names[type];
6368 }
6369
6370 enum btrfs_loop_type {
6371         LOOP_CACHING_NOWAIT = 0,
6372         LOOP_CACHING_WAIT = 1,
6373         LOOP_ALLOC_CHUNK = 2,
6374         LOOP_NO_EMPTY_SIZE = 3,
6375 };
6376
6377 static inline void
6378 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
6379                        int delalloc)
6380 {
6381         if (delalloc)
6382                 down_read(&cache->data_rwsem);
6383 }
6384
6385 static inline void
6386 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
6387                        int delalloc)
6388 {
6389         btrfs_get_block_group(cache);
6390         if (delalloc)
6391                 down_read(&cache->data_rwsem);
6392 }
6393
6394 static struct btrfs_block_group_cache *
6395 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
6396                    struct btrfs_free_cluster *cluster,
6397                    int delalloc)
6398 {
6399         struct btrfs_block_group_cache *used_bg;
6400         bool locked = false;
6401 again:
6402         spin_lock(&cluster->refill_lock);
6403         if (locked) {
6404                 if (used_bg == cluster->block_group)
6405                         return used_bg;
6406
6407                 up_read(&used_bg->data_rwsem);
6408                 btrfs_put_block_group(used_bg);
6409         }
6410
6411         used_bg = cluster->block_group;
6412         if (!used_bg)
6413                 return NULL;
6414
6415         if (used_bg == block_group)
6416                 return used_bg;
6417
6418         btrfs_get_block_group(used_bg);
6419
6420         if (!delalloc)
6421                 return used_bg;
6422
6423         if (down_read_trylock(&used_bg->data_rwsem))
6424                 return used_bg;
6425
6426         spin_unlock(&cluster->refill_lock);
6427         down_read(&used_bg->data_rwsem);
6428         locked = true;
6429         goto again;
6430 }
6431
6432 static inline void
6433 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
6434                          int delalloc)
6435 {
6436         if (delalloc)
6437                 up_read(&cache->data_rwsem);
6438         btrfs_put_block_group(cache);
6439 }
6440
6441 /*
6442  * walks the btree of allocated extents and find a hole of a given size.
6443  * The key ins is changed to record the hole:
6444  * ins->objectid == start position
6445  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6446  * ins->offset == the size of the hole.
6447  * Any available blocks before search_start are skipped.
6448  *
6449  * If there is no suitable free space, we will record the max size of
6450  * the free space extent currently.
6451  */
6452 static noinline int find_free_extent(struct btrfs_root *orig_root,
6453                                      u64 num_bytes, u64 empty_size,
6454                                      u64 hint_byte, struct btrfs_key *ins,
6455                                      u64 flags, int delalloc)
6456 {
6457         int ret = 0;
6458         struct btrfs_root *root = orig_root->fs_info->extent_root;
6459         struct btrfs_free_cluster *last_ptr = NULL;
6460         struct btrfs_block_group_cache *block_group = NULL;
6461         u64 search_start = 0;
6462         u64 max_extent_size = 0;
6463         int empty_cluster = 2 * 1024 * 1024;
6464         struct btrfs_space_info *space_info;
6465         int loop = 0;
6466         int index = __get_raid_index(flags);
6467         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6468                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6469         bool failed_cluster_refill = false;
6470         bool failed_alloc = false;
6471         bool use_cluster = true;
6472         bool have_caching_bg = false;
6473
6474         WARN_ON(num_bytes < root->sectorsize);
6475         ins->type = BTRFS_EXTENT_ITEM_KEY;
6476         ins->objectid = 0;
6477         ins->offset = 0;
6478
6479         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6480
6481         space_info = __find_space_info(root->fs_info, flags);
6482         if (!space_info) {
6483                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6484                 return -ENOSPC;
6485         }
6486
6487         /*
6488          * If the space info is for both data and metadata it means we have a
6489          * small filesystem and we can't use the clustering stuff.
6490          */
6491         if (btrfs_mixed_space_info(space_info))
6492                 use_cluster = false;
6493
6494         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6495                 last_ptr = &root->fs_info->meta_alloc_cluster;
6496                 if (!btrfs_test_opt(root, SSD))
6497                         empty_cluster = 64 * 1024;
6498         }
6499
6500         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6501             btrfs_test_opt(root, SSD)) {
6502                 last_ptr = &root->fs_info->data_alloc_cluster;
6503         }
6504
6505         if (last_ptr) {
6506                 spin_lock(&last_ptr->lock);
6507                 if (last_ptr->block_group)
6508                         hint_byte = last_ptr->window_start;
6509                 spin_unlock(&last_ptr->lock);
6510         }
6511
6512         search_start = max(search_start, first_logical_byte(root, 0));
6513         search_start = max(search_start, hint_byte);
6514
6515         if (!last_ptr)
6516                 empty_cluster = 0;
6517
6518         if (search_start == hint_byte) {
6519                 block_group = btrfs_lookup_block_group(root->fs_info,
6520                                                        search_start);
6521                 /*
6522                  * we don't want to use the block group if it doesn't match our
6523                  * allocation bits, or if its not cached.
6524                  *
6525                  * However if we are re-searching with an ideal block group
6526                  * picked out then we don't care that the block group is cached.
6527                  */
6528                 if (block_group && block_group_bits(block_group, flags) &&
6529                     block_group->cached != BTRFS_CACHE_NO) {
6530                         down_read(&space_info->groups_sem);
6531                         if (list_empty(&block_group->list) ||
6532                             block_group->ro) {
6533                                 /*
6534                                  * someone is removing this block group,
6535                                  * we can't jump into the have_block_group
6536                                  * target because our list pointers are not
6537                                  * valid
6538                                  */
6539                                 btrfs_put_block_group(block_group);
6540                                 up_read(&space_info->groups_sem);
6541                         } else {
6542                                 index = get_block_group_index(block_group);
6543                                 btrfs_lock_block_group(block_group, delalloc);
6544                                 goto have_block_group;
6545                         }
6546                 } else if (block_group) {
6547                         btrfs_put_block_group(block_group);
6548                 }
6549         }
6550 search:
6551         have_caching_bg = false;
6552         down_read(&space_info->groups_sem);
6553         list_for_each_entry(block_group, &space_info->block_groups[index],
6554                             list) {
6555                 u64 offset;
6556                 int cached;
6557
6558                 btrfs_grab_block_group(block_group, delalloc);
6559                 search_start = block_group->key.objectid;
6560
6561                 /*
6562                  * this can happen if we end up cycling through all the
6563                  * raid types, but we want to make sure we only allocate
6564                  * for the proper type.
6565                  */
6566                 if (!block_group_bits(block_group, flags)) {
6567                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6568                                 BTRFS_BLOCK_GROUP_RAID1 |
6569                                 BTRFS_BLOCK_GROUP_RAID5 |
6570                                 BTRFS_BLOCK_GROUP_RAID6 |
6571                                 BTRFS_BLOCK_GROUP_RAID10;
6572
6573                         /*
6574                          * if they asked for extra copies and this block group
6575                          * doesn't provide them, bail.  This does allow us to
6576                          * fill raid0 from raid1.
6577                          */
6578                         if ((flags & extra) && !(block_group->flags & extra))
6579                                 goto loop;
6580                 }
6581
6582 have_block_group:
6583                 cached = block_group_cache_done(block_group);
6584                 if (unlikely(!cached)) {
6585                         ret = cache_block_group(block_group, 0);
6586                         BUG_ON(ret < 0);
6587                         ret = 0;
6588                 }
6589
6590                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
6591                         goto loop;
6592                 if (unlikely(block_group->ro))
6593                         goto loop;
6594
6595                 /*
6596                  * Ok we want to try and use the cluster allocator, so
6597                  * lets look there
6598                  */
6599                 if (last_ptr) {
6600                         struct btrfs_block_group_cache *used_block_group;
6601                         unsigned long aligned_cluster;
6602                         /*
6603                          * the refill lock keeps out other
6604                          * people trying to start a new cluster
6605                          */
6606                         used_block_group = btrfs_lock_cluster(block_group,
6607                                                               last_ptr,
6608                                                               delalloc);
6609                         if (!used_block_group)
6610                                 goto refill_cluster;
6611
6612                         if (used_block_group != block_group &&
6613                             (used_block_group->ro ||
6614                              !block_group_bits(used_block_group, flags)))
6615                                 goto release_cluster;
6616
6617                         offset = btrfs_alloc_from_cluster(used_block_group,
6618                                                 last_ptr,
6619                                                 num_bytes,
6620                                                 used_block_group->key.objectid,
6621                                                 &max_extent_size);
6622                         if (offset) {
6623                                 /* we have a block, we're done */
6624                                 spin_unlock(&last_ptr->refill_lock);
6625                                 trace_btrfs_reserve_extent_cluster(root,
6626                                                 used_block_group,
6627                                                 search_start, num_bytes);
6628                                 if (used_block_group != block_group) {
6629                                         btrfs_release_block_group(block_group,
6630                                                                   delalloc);
6631                                         block_group = used_block_group;
6632                                 }
6633                                 goto checks;
6634                         }
6635
6636                         WARN_ON(last_ptr->block_group != used_block_group);
6637 release_cluster:
6638                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6639                          * set up a new clusters, so lets just skip it
6640                          * and let the allocator find whatever block
6641                          * it can find.  If we reach this point, we
6642                          * will have tried the cluster allocator
6643                          * plenty of times and not have found
6644                          * anything, so we are likely way too
6645                          * fragmented for the clustering stuff to find
6646                          * anything.
6647                          *
6648                          * However, if the cluster is taken from the
6649                          * current block group, release the cluster
6650                          * first, so that we stand a better chance of
6651                          * succeeding in the unclustered
6652                          * allocation.  */
6653                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6654                             used_block_group != block_group) {
6655                                 spin_unlock(&last_ptr->refill_lock);
6656                                 btrfs_release_block_group(used_block_group,
6657                                                           delalloc);
6658                                 goto unclustered_alloc;
6659                         }
6660
6661                         /*
6662                          * this cluster didn't work out, free it and
6663                          * start over
6664                          */
6665                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6666
6667                         if (used_block_group != block_group)
6668                                 btrfs_release_block_group(used_block_group,
6669                                                           delalloc);
6670 refill_cluster:
6671                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6672                                 spin_unlock(&last_ptr->refill_lock);
6673                                 goto unclustered_alloc;
6674                         }
6675
6676                         aligned_cluster = max_t(unsigned long,
6677                                                 empty_cluster + empty_size,
6678                                               block_group->full_stripe_len);
6679
6680                         /* allocate a cluster in this block group */
6681                         ret = btrfs_find_space_cluster(root, block_group,
6682                                                        last_ptr, search_start,
6683                                                        num_bytes,
6684                                                        aligned_cluster);
6685                         if (ret == 0) {
6686                                 /*
6687                                  * now pull our allocation out of this
6688                                  * cluster
6689                                  */
6690                                 offset = btrfs_alloc_from_cluster(block_group,
6691                                                         last_ptr,
6692                                                         num_bytes,
6693                                                         search_start,
6694                                                         &max_extent_size);
6695                                 if (offset) {
6696                                         /* we found one, proceed */
6697                                         spin_unlock(&last_ptr->refill_lock);
6698                                         trace_btrfs_reserve_extent_cluster(root,
6699                                                 block_group, search_start,
6700                                                 num_bytes);
6701                                         goto checks;
6702                                 }
6703                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6704                                    && !failed_cluster_refill) {
6705                                 spin_unlock(&last_ptr->refill_lock);
6706
6707                                 failed_cluster_refill = true;
6708                                 wait_block_group_cache_progress(block_group,
6709                                        num_bytes + empty_cluster + empty_size);
6710                                 goto have_block_group;
6711                         }
6712
6713                         /*
6714                          * at this point we either didn't find a cluster
6715                          * or we weren't able to allocate a block from our
6716                          * cluster.  Free the cluster we've been trying
6717                          * to use, and go to the next block group
6718                          */
6719                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6720                         spin_unlock(&last_ptr->refill_lock);
6721                         goto loop;
6722                 }
6723
6724 unclustered_alloc:
6725                 spin_lock(&block_group->free_space_ctl->tree_lock);
6726                 if (cached &&
6727                     block_group->free_space_ctl->free_space <
6728                     num_bytes + empty_cluster + empty_size) {
6729                         if (block_group->free_space_ctl->free_space >
6730                             max_extent_size)
6731                                 max_extent_size =
6732                                         block_group->free_space_ctl->free_space;
6733                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6734                         goto loop;
6735                 }
6736                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6737
6738                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6739                                                     num_bytes, empty_size,
6740                                                     &max_extent_size);
6741                 /*
6742                  * If we didn't find a chunk, and we haven't failed on this
6743                  * block group before, and this block group is in the middle of
6744                  * caching and we are ok with waiting, then go ahead and wait
6745                  * for progress to be made, and set failed_alloc to true.
6746                  *
6747                  * If failed_alloc is true then we've already waited on this
6748                  * block group once and should move on to the next block group.
6749                  */
6750                 if (!offset && !failed_alloc && !cached &&
6751                     loop > LOOP_CACHING_NOWAIT) {
6752                         wait_block_group_cache_progress(block_group,
6753                                                 num_bytes + empty_size);
6754                         failed_alloc = true;
6755                         goto have_block_group;
6756                 } else if (!offset) {
6757                         if (!cached)
6758                                 have_caching_bg = true;
6759                         goto loop;
6760                 }
6761 checks:
6762                 search_start = ALIGN(offset, root->stripesize);
6763
6764                 /* move on to the next group */
6765                 if (search_start + num_bytes >
6766                     block_group->key.objectid + block_group->key.offset) {
6767                         btrfs_add_free_space(block_group, offset, num_bytes);
6768                         goto loop;
6769                 }
6770
6771                 if (offset < search_start)
6772                         btrfs_add_free_space(block_group, offset,
6773                                              search_start - offset);
6774                 BUG_ON(offset > search_start);
6775
6776                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
6777                                                   alloc_type, delalloc);
6778                 if (ret == -EAGAIN) {
6779                         btrfs_add_free_space(block_group, offset, num_bytes);
6780                         goto loop;
6781                 }
6782
6783                 /* we are all good, lets return */
6784                 ins->objectid = search_start;
6785                 ins->offset = num_bytes;
6786
6787                 trace_btrfs_reserve_extent(orig_root, block_group,
6788                                            search_start, num_bytes);
6789                 btrfs_release_block_group(block_group, delalloc);
6790                 break;
6791 loop:
6792                 failed_cluster_refill = false;
6793                 failed_alloc = false;
6794                 BUG_ON(index != get_block_group_index(block_group));
6795                 btrfs_release_block_group(block_group, delalloc);
6796         }
6797         up_read(&space_info->groups_sem);
6798
6799         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6800                 goto search;
6801
6802         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6803                 goto search;
6804
6805         /*
6806          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6807          *                      caching kthreads as we move along
6808          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6809          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6810          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6811          *                      again
6812          */
6813         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6814                 index = 0;
6815                 loop++;
6816                 if (loop == LOOP_ALLOC_CHUNK) {
6817                         struct btrfs_trans_handle *trans;
6818                         int exist = 0;
6819
6820                         trans = current->journal_info;
6821                         if (trans)
6822                                 exist = 1;
6823                         else
6824                                 trans = btrfs_join_transaction(root);
6825
6826                         if (IS_ERR(trans)) {
6827                                 ret = PTR_ERR(trans);
6828                                 goto out;
6829                         }
6830
6831                         ret = do_chunk_alloc(trans, root, flags,
6832                                              CHUNK_ALLOC_FORCE);
6833                         /*
6834                          * Do not bail out on ENOSPC since we
6835                          * can do more things.
6836                          */
6837                         if (ret < 0 && ret != -ENOSPC)
6838                                 btrfs_abort_transaction(trans,
6839                                                         root, ret);
6840                         else
6841                                 ret = 0;
6842                         if (!exist)
6843                                 btrfs_end_transaction(trans, root);
6844                         if (ret)
6845                                 goto out;
6846                 }
6847
6848                 if (loop == LOOP_NO_EMPTY_SIZE) {
6849                         empty_size = 0;
6850                         empty_cluster = 0;
6851                 }
6852
6853                 goto search;
6854         } else if (!ins->objectid) {
6855                 ret = -ENOSPC;
6856         } else if (ins->objectid) {
6857                 ret = 0;
6858         }
6859 out:
6860         if (ret == -ENOSPC)
6861                 ins->offset = max_extent_size;
6862         return ret;
6863 }
6864
6865 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6866                             int dump_block_groups)
6867 {
6868         struct btrfs_block_group_cache *cache;
6869         int index = 0;
6870
6871         spin_lock(&info->lock);
6872         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
6873                info->flags,
6874                info->total_bytes - info->bytes_used - info->bytes_pinned -
6875                info->bytes_reserved - info->bytes_readonly,
6876                (info->full) ? "" : "not ");
6877         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
6878                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6879                info->total_bytes, info->bytes_used, info->bytes_pinned,
6880                info->bytes_reserved, info->bytes_may_use,
6881                info->bytes_readonly);
6882         spin_unlock(&info->lock);
6883
6884         if (!dump_block_groups)
6885                 return;
6886
6887         down_read(&info->groups_sem);
6888 again:
6889         list_for_each_entry(cache, &info->block_groups[index], list) {
6890                 spin_lock(&cache->lock);
6891                 printk(KERN_INFO "BTRFS: "
6892                            "block group %llu has %llu bytes, "
6893                            "%llu used %llu pinned %llu reserved %s\n",
6894                        cache->key.objectid, cache->key.offset,
6895                        btrfs_block_group_used(&cache->item), cache->pinned,
6896                        cache->reserved, cache->ro ? "[readonly]" : "");
6897                 btrfs_dump_free_space(cache, bytes);
6898                 spin_unlock(&cache->lock);
6899         }
6900         if (++index < BTRFS_NR_RAID_TYPES)
6901                 goto again;
6902         up_read(&info->groups_sem);
6903 }
6904
6905 int btrfs_reserve_extent(struct btrfs_root *root,
6906                          u64 num_bytes, u64 min_alloc_size,
6907                          u64 empty_size, u64 hint_byte,
6908                          struct btrfs_key *ins, int is_data, int delalloc)
6909 {
6910         bool final_tried = false;
6911         u64 flags;
6912         int ret;
6913
6914         flags = btrfs_get_alloc_profile(root, is_data);
6915 again:
6916         WARN_ON(num_bytes < root->sectorsize);
6917         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
6918                                flags, delalloc);
6919
6920         if (ret == -ENOSPC) {
6921                 if (!final_tried && ins->offset) {
6922                         num_bytes = min(num_bytes >> 1, ins->offset);
6923                         num_bytes = round_down(num_bytes, root->sectorsize);
6924                         num_bytes = max(num_bytes, min_alloc_size);
6925                         if (num_bytes == min_alloc_size)
6926                                 final_tried = true;
6927                         goto again;
6928                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6929                         struct btrfs_space_info *sinfo;
6930
6931                         sinfo = __find_space_info(root->fs_info, flags);
6932                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
6933                                 flags, num_bytes);
6934                         if (sinfo)
6935                                 dump_space_info(sinfo, num_bytes, 1);
6936                 }
6937         }
6938
6939         return ret;
6940 }
6941
6942 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6943                                         u64 start, u64 len,
6944                                         int pin, int delalloc)
6945 {
6946         struct btrfs_block_group_cache *cache;
6947         int ret = 0;
6948
6949         cache = btrfs_lookup_block_group(root->fs_info, start);
6950         if (!cache) {
6951                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
6952                         start);
6953                 return -ENOSPC;
6954         }
6955
6956         if (btrfs_test_opt(root, DISCARD))
6957                 ret = btrfs_discard_extent(root, start, len, NULL);
6958
6959         if (pin)
6960                 pin_down_extent(root, cache, start, len, 1);
6961         else {
6962                 btrfs_add_free_space(cache, start, len);
6963                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
6964         }
6965         btrfs_put_block_group(cache);
6966
6967         trace_btrfs_reserved_extent_free(root, start, len);
6968
6969         return ret;
6970 }
6971
6972 int btrfs_free_reserved_extent(struct btrfs_root *root,
6973                                u64 start, u64 len, int delalloc)
6974 {
6975         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
6976 }
6977
6978 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6979                                        u64 start, u64 len)
6980 {
6981         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
6982 }
6983
6984 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6985                                       struct btrfs_root *root,
6986                                       u64 parent, u64 root_objectid,
6987                                       u64 flags, u64 owner, u64 offset,
6988                                       struct btrfs_key *ins, int ref_mod)
6989 {
6990         int ret;
6991         struct btrfs_fs_info *fs_info = root->fs_info;
6992         struct btrfs_extent_item *extent_item;
6993         struct btrfs_extent_inline_ref *iref;
6994         struct btrfs_path *path;
6995         struct extent_buffer *leaf;
6996         int type;
6997         u32 size;
6998
6999         if (parent > 0)
7000                 type = BTRFS_SHARED_DATA_REF_KEY;
7001         else
7002                 type = BTRFS_EXTENT_DATA_REF_KEY;
7003
7004         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7005
7006         path = btrfs_alloc_path();
7007         if (!path)
7008                 return -ENOMEM;
7009
7010         path->leave_spinning = 1;
7011         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7012                                       ins, size);
7013         if (ret) {
7014                 btrfs_free_path(path);
7015                 return ret;
7016         }
7017
7018         leaf = path->nodes[0];
7019         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7020                                      struct btrfs_extent_item);
7021         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7022         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7023         btrfs_set_extent_flags(leaf, extent_item,
7024                                flags | BTRFS_EXTENT_FLAG_DATA);
7025
7026         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7027         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7028         if (parent > 0) {
7029                 struct btrfs_shared_data_ref *ref;
7030                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7031                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7032                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7033         } else {
7034                 struct btrfs_extent_data_ref *ref;
7035                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7036                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7037                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7038                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7039                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7040         }
7041
7042         btrfs_mark_buffer_dirty(path->nodes[0]);
7043         btrfs_free_path(path);
7044
7045         /* Always set parent to 0 here since its exclusive anyway. */
7046         ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
7047                                       ins->objectid, ins->offset,
7048                                       BTRFS_QGROUP_OPER_ADD_EXCL, 0);
7049         if (ret)
7050                 return ret;
7051
7052         ret = update_block_group(root, ins->objectid, ins->offset, 1);
7053         if (ret) { /* -ENOENT, logic error */
7054                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7055                         ins->objectid, ins->offset);
7056                 BUG();
7057         }
7058         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
7059         return ret;
7060 }
7061
7062 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7063                                      struct btrfs_root *root,
7064                                      u64 parent, u64 root_objectid,
7065                                      u64 flags, struct btrfs_disk_key *key,
7066                                      int level, struct btrfs_key *ins,
7067                                      int no_quota)
7068 {
7069         int ret;
7070         struct btrfs_fs_info *fs_info = root->fs_info;
7071         struct btrfs_extent_item *extent_item;
7072         struct btrfs_tree_block_info *block_info;
7073         struct btrfs_extent_inline_ref *iref;
7074         struct btrfs_path *path;
7075         struct extent_buffer *leaf;
7076         u32 size = sizeof(*extent_item) + sizeof(*iref);
7077         u64 num_bytes = ins->offset;
7078         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7079                                                  SKINNY_METADATA);
7080
7081         if (!skinny_metadata)
7082                 size += sizeof(*block_info);
7083
7084         path = btrfs_alloc_path();
7085         if (!path) {
7086                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7087                                                    root->nodesize);
7088                 return -ENOMEM;
7089         }
7090
7091         path->leave_spinning = 1;
7092         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7093                                       ins, size);
7094         if (ret) {
7095                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7096                                                    root->nodesize);
7097                 btrfs_free_path(path);
7098                 return ret;
7099         }
7100
7101         leaf = path->nodes[0];
7102         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7103                                      struct btrfs_extent_item);
7104         btrfs_set_extent_refs(leaf, extent_item, 1);
7105         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7106         btrfs_set_extent_flags(leaf, extent_item,
7107                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7108
7109         if (skinny_metadata) {
7110                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7111                 num_bytes = root->nodesize;
7112         } else {
7113                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7114                 btrfs_set_tree_block_key(leaf, block_info, key);
7115                 btrfs_set_tree_block_level(leaf, block_info, level);
7116                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7117         }
7118
7119         if (parent > 0) {
7120                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7121                 btrfs_set_extent_inline_ref_type(leaf, iref,
7122                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7123                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7124         } else {
7125                 btrfs_set_extent_inline_ref_type(leaf, iref,
7126                                                  BTRFS_TREE_BLOCK_REF_KEY);
7127                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7128         }
7129
7130         btrfs_mark_buffer_dirty(leaf);
7131         btrfs_free_path(path);
7132
7133         if (!no_quota) {
7134                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
7135                                               ins->objectid, num_bytes,
7136                                               BTRFS_QGROUP_OPER_ADD_EXCL, 0);
7137                 if (ret)
7138                         return ret;
7139         }
7140
7141         ret = update_block_group(root, ins->objectid, root->nodesize, 1);
7142         if (ret) { /* -ENOENT, logic error */
7143                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7144                         ins->objectid, ins->offset);
7145                 BUG();
7146         }
7147
7148         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
7149         return ret;
7150 }
7151
7152 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7153                                      struct btrfs_root *root,
7154                                      u64 root_objectid, u64 owner,
7155                                      u64 offset, struct btrfs_key *ins)
7156 {
7157         int ret;
7158
7159         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7160
7161         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7162                                          ins->offset, 0,
7163                                          root_objectid, owner, offset,
7164                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
7165         return ret;
7166 }
7167
7168 /*
7169  * this is used by the tree logging recovery code.  It records that
7170  * an extent has been allocated and makes sure to clear the free
7171  * space cache bits as well
7172  */
7173 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7174                                    struct btrfs_root *root,
7175                                    u64 root_objectid, u64 owner, u64 offset,
7176                                    struct btrfs_key *ins)
7177 {
7178         int ret;
7179         struct btrfs_block_group_cache *block_group;
7180
7181         /*
7182          * Mixed block groups will exclude before processing the log so we only
7183          * need to do the exlude dance if this fs isn't mixed.
7184          */
7185         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7186                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7187                 if (ret)
7188                         return ret;
7189         }
7190
7191         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
7192         if (!block_group)
7193                 return -EINVAL;
7194
7195         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
7196                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
7197         BUG_ON(ret); /* logic error */
7198         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
7199                                          0, owner, offset, ins, 1);
7200         btrfs_put_block_group(block_group);
7201         return ret;
7202 }
7203
7204 static struct extent_buffer *
7205 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7206                       u64 bytenr, u32 blocksize, int level)
7207 {
7208         struct extent_buffer *buf;
7209
7210         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
7211         if (!buf)
7212                 return ERR_PTR(-ENOMEM);
7213         btrfs_set_header_generation(buf, trans->transid);
7214         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
7215         btrfs_tree_lock(buf);
7216         clean_tree_block(trans, root, buf);
7217         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
7218
7219         btrfs_set_lock_blocking(buf);
7220         btrfs_set_buffer_uptodate(buf);
7221
7222         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
7223                 buf->log_index = root->log_transid % 2;
7224                 /*
7225                  * we allow two log transactions at a time, use different
7226                  * EXENT bit to differentiate dirty pages.
7227                  */
7228                 if (buf->log_index == 0)
7229                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7230                                         buf->start + buf->len - 1, GFP_NOFS);
7231                 else
7232                         set_extent_new(&root->dirty_log_pages, buf->start,
7233                                         buf->start + buf->len - 1, GFP_NOFS);
7234         } else {
7235                 buf->log_index = -1;
7236                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7237                          buf->start + buf->len - 1, GFP_NOFS);
7238         }
7239         trans->blocks_used++;
7240         /* this returns a buffer locked for blocking */
7241         return buf;
7242 }
7243
7244 static struct btrfs_block_rsv *
7245 use_block_rsv(struct btrfs_trans_handle *trans,
7246               struct btrfs_root *root, u32 blocksize)
7247 {
7248         struct btrfs_block_rsv *block_rsv;
7249         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7250         int ret;
7251         bool global_updated = false;
7252
7253         block_rsv = get_block_rsv(trans, root);
7254
7255         if (unlikely(block_rsv->size == 0))
7256                 goto try_reserve;
7257 again:
7258         ret = block_rsv_use_bytes(block_rsv, blocksize);
7259         if (!ret)
7260                 return block_rsv;
7261
7262         if (block_rsv->failfast)
7263                 return ERR_PTR(ret);
7264
7265         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7266                 global_updated = true;
7267                 update_global_block_rsv(root->fs_info);
7268                 goto again;
7269         }
7270
7271         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7272                 static DEFINE_RATELIMIT_STATE(_rs,
7273                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7274                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7275                 if (__ratelimit(&_rs))
7276                         WARN(1, KERN_DEBUG
7277                                 "BTRFS: block rsv returned %d\n", ret);
7278         }
7279 try_reserve:
7280         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7281                                      BTRFS_RESERVE_NO_FLUSH);
7282         if (!ret)
7283                 return block_rsv;
7284         /*
7285          * If we couldn't reserve metadata bytes try and use some from
7286          * the global reserve if its space type is the same as the global
7287          * reservation.
7288          */
7289         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7290             block_rsv->space_info == global_rsv->space_info) {
7291                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7292                 if (!ret)
7293                         return global_rsv;
7294         }
7295         return ERR_PTR(ret);
7296 }
7297
7298 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7299                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7300 {
7301         block_rsv_add_bytes(block_rsv, blocksize, 0);
7302         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7303 }
7304
7305 /*
7306  * finds a free extent and does all the dirty work required for allocation
7307  * returns the key for the extent through ins, and a tree buffer for
7308  * the first block of the extent through buf.
7309  *
7310  * returns the tree buffer or NULL.
7311  */
7312 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
7313                                         struct btrfs_root *root,
7314                                         u64 parent, u64 root_objectid,
7315                                         struct btrfs_disk_key *key, int level,
7316                                         u64 hint, u64 empty_size)
7317 {
7318         struct btrfs_key ins;
7319         struct btrfs_block_rsv *block_rsv;
7320         struct extent_buffer *buf;
7321         u64 flags = 0;
7322         int ret;
7323         u32 blocksize = root->nodesize;
7324         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7325                                                  SKINNY_METADATA);
7326
7327         if (btrfs_test_is_dummy_root(root)) {
7328                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
7329                                             blocksize, level);
7330                 if (!IS_ERR(buf))
7331                         root->alloc_bytenr += blocksize;
7332                 return buf;
7333         }
7334
7335         block_rsv = use_block_rsv(trans, root, blocksize);
7336         if (IS_ERR(block_rsv))
7337                 return ERR_CAST(block_rsv);
7338
7339         ret = btrfs_reserve_extent(root, blocksize, blocksize,
7340                                    empty_size, hint, &ins, 0, 0);
7341         if (ret) {
7342                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
7343                 return ERR_PTR(ret);
7344         }
7345
7346         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
7347                                     blocksize, level);
7348         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
7349
7350         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
7351                 if (parent == 0)
7352                         parent = ins.objectid;
7353                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
7354         } else
7355                 BUG_ON(parent > 0);
7356
7357         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
7358                 struct btrfs_delayed_extent_op *extent_op;
7359                 extent_op = btrfs_alloc_delayed_extent_op();
7360                 BUG_ON(!extent_op); /* -ENOMEM */
7361                 if (key)
7362                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
7363                 else
7364                         memset(&extent_op->key, 0, sizeof(extent_op->key));
7365                 extent_op->flags_to_set = flags;
7366                 if (skinny_metadata)
7367                         extent_op->update_key = 0;
7368                 else
7369                         extent_op->update_key = 1;
7370                 extent_op->update_flags = 1;
7371                 extent_op->is_data = 0;
7372                 extent_op->level = level;
7373
7374                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7375                                         ins.objectid,
7376                                         ins.offset, parent, root_objectid,
7377                                         level, BTRFS_ADD_DELAYED_EXTENT,
7378                                         extent_op, 0);
7379                 BUG_ON(ret); /* -ENOMEM */
7380         }
7381         return buf;
7382 }
7383
7384 struct walk_control {
7385         u64 refs[BTRFS_MAX_LEVEL];
7386         u64 flags[BTRFS_MAX_LEVEL];
7387         struct btrfs_key update_progress;
7388         int stage;
7389         int level;
7390         int shared_level;
7391         int update_ref;
7392         int keep_locks;
7393         int reada_slot;
7394         int reada_count;
7395         int for_reloc;
7396 };
7397
7398 #define DROP_REFERENCE  1
7399 #define UPDATE_BACKREF  2
7400
7401 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
7402                                      struct btrfs_root *root,
7403                                      struct walk_control *wc,
7404                                      struct btrfs_path *path)
7405 {
7406         u64 bytenr;
7407         u64 generation;
7408         u64 refs;
7409         u64 flags;
7410         u32 nritems;
7411         u32 blocksize;
7412         struct btrfs_key key;
7413         struct extent_buffer *eb;
7414         int ret;
7415         int slot;
7416         int nread = 0;
7417
7418         if (path->slots[wc->level] < wc->reada_slot) {
7419                 wc->reada_count = wc->reada_count * 2 / 3;
7420                 wc->reada_count = max(wc->reada_count, 2);
7421         } else {
7422                 wc->reada_count = wc->reada_count * 3 / 2;
7423                 wc->reada_count = min_t(int, wc->reada_count,
7424                                         BTRFS_NODEPTRS_PER_BLOCK(root));
7425         }
7426
7427         eb = path->nodes[wc->level];
7428         nritems = btrfs_header_nritems(eb);
7429         blocksize = root->nodesize;
7430
7431         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
7432                 if (nread >= wc->reada_count)
7433                         break;
7434
7435                 cond_resched();
7436                 bytenr = btrfs_node_blockptr(eb, slot);
7437                 generation = btrfs_node_ptr_generation(eb, slot);
7438
7439                 if (slot == path->slots[wc->level])
7440                         goto reada;
7441
7442                 if (wc->stage == UPDATE_BACKREF &&
7443                     generation <= root->root_key.offset)
7444                         continue;
7445
7446                 /* We don't lock the tree block, it's OK to be racy here */
7447                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7448                                                wc->level - 1, 1, &refs,
7449                                                &flags);
7450                 /* We don't care about errors in readahead. */
7451                 if (ret < 0)
7452                         continue;
7453                 BUG_ON(refs == 0);
7454
7455                 if (wc->stage == DROP_REFERENCE) {
7456                         if (refs == 1)
7457                                 goto reada;
7458
7459                         if (wc->level == 1 &&
7460                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7461                                 continue;
7462                         if (!wc->update_ref ||
7463                             generation <= root->root_key.offset)
7464                                 continue;
7465                         btrfs_node_key_to_cpu(eb, &key, slot);
7466                         ret = btrfs_comp_cpu_keys(&key,
7467                                                   &wc->update_progress);
7468                         if (ret < 0)
7469                                 continue;
7470                 } else {
7471                         if (wc->level == 1 &&
7472                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7473                                 continue;
7474                 }
7475 reada:
7476                 readahead_tree_block(root, bytenr, blocksize);
7477                 nread++;
7478         }
7479         wc->reada_slot = slot;
7480 }
7481
7482 static int account_leaf_items(struct btrfs_trans_handle *trans,
7483                               struct btrfs_root *root,
7484                               struct extent_buffer *eb)
7485 {
7486         int nr = btrfs_header_nritems(eb);
7487         int i, extent_type, ret;
7488         struct btrfs_key key;
7489         struct btrfs_file_extent_item *fi;
7490         u64 bytenr, num_bytes;
7491
7492         for (i = 0; i < nr; i++) {
7493                 btrfs_item_key_to_cpu(eb, &key, i);
7494
7495                 if (key.type != BTRFS_EXTENT_DATA_KEY)
7496                         continue;
7497
7498                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
7499                 /* filter out non qgroup-accountable extents  */
7500                 extent_type = btrfs_file_extent_type(eb, fi);
7501
7502                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
7503                         continue;
7504
7505                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
7506                 if (!bytenr)
7507                         continue;
7508
7509                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
7510
7511                 ret = btrfs_qgroup_record_ref(trans, root->fs_info,
7512                                               root->objectid,
7513                                               bytenr, num_bytes,
7514                                               BTRFS_QGROUP_OPER_SUB_SUBTREE, 0);
7515                 if (ret)
7516                         return ret;
7517         }
7518         return 0;
7519 }
7520
7521 /*
7522  * Walk up the tree from the bottom, freeing leaves and any interior
7523  * nodes which have had all slots visited. If a node (leaf or
7524  * interior) is freed, the node above it will have it's slot
7525  * incremented. The root node will never be freed.
7526  *
7527  * At the end of this function, we should have a path which has all
7528  * slots incremented to the next position for a search. If we need to
7529  * read a new node it will be NULL and the node above it will have the
7530  * correct slot selected for a later read.
7531  *
7532  * If we increment the root nodes slot counter past the number of
7533  * elements, 1 is returned to signal completion of the search.
7534  */
7535 static int adjust_slots_upwards(struct btrfs_root *root,
7536                                 struct btrfs_path *path, int root_level)
7537 {
7538         int level = 0;
7539         int nr, slot;
7540         struct extent_buffer *eb;
7541
7542         if (root_level == 0)
7543                 return 1;
7544
7545         while (level <= root_level) {
7546                 eb = path->nodes[level];
7547                 nr = btrfs_header_nritems(eb);
7548                 path->slots[level]++;
7549                 slot = path->slots[level];
7550                 if (slot >= nr || level == 0) {
7551                         /*
7552                          * Don't free the root -  we will detect this
7553                          * condition after our loop and return a
7554                          * positive value for caller to stop walking the tree.
7555                          */
7556                         if (level != root_level) {
7557                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7558                                 path->locks[level] = 0;
7559
7560                                 free_extent_buffer(eb);
7561                                 path->nodes[level] = NULL;
7562                                 path->slots[level] = 0;
7563                         }
7564                 } else {
7565                         /*
7566                          * We have a valid slot to walk back down
7567                          * from. Stop here so caller can process these
7568                          * new nodes.
7569                          */
7570                         break;
7571                 }
7572
7573                 level++;
7574         }
7575
7576         eb = path->nodes[root_level];
7577         if (path->slots[root_level] >= btrfs_header_nritems(eb))
7578                 return 1;
7579
7580         return 0;
7581 }
7582
7583 /*
7584  * root_eb is the subtree root and is locked before this function is called.
7585  */
7586 static int account_shared_subtree(struct btrfs_trans_handle *trans,
7587                                   struct btrfs_root *root,
7588                                   struct extent_buffer *root_eb,
7589                                   u64 root_gen,
7590                                   int root_level)
7591 {
7592         int ret = 0;
7593         int level;
7594         struct extent_buffer *eb = root_eb;
7595         struct btrfs_path *path = NULL;
7596
7597         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
7598         BUG_ON(root_eb == NULL);
7599
7600         if (!root->fs_info->quota_enabled)
7601                 return 0;
7602
7603         if (!extent_buffer_uptodate(root_eb)) {
7604                 ret = btrfs_read_buffer(root_eb, root_gen);
7605                 if (ret)
7606                         goto out;
7607         }
7608
7609         if (root_level == 0) {
7610                 ret = account_leaf_items(trans, root, root_eb);
7611                 goto out;
7612         }
7613
7614         path = btrfs_alloc_path();
7615         if (!path)
7616                 return -ENOMEM;
7617
7618         /*
7619          * Walk down the tree.  Missing extent blocks are filled in as
7620          * we go. Metadata is accounted every time we read a new
7621          * extent block.
7622          *
7623          * When we reach a leaf, we account for file extent items in it,
7624          * walk back up the tree (adjusting slot pointers as we go)
7625          * and restart the search process.
7626          */
7627         extent_buffer_get(root_eb); /* For path */
7628         path->nodes[root_level] = root_eb;
7629         path->slots[root_level] = 0;
7630         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
7631 walk_down:
7632         level = root_level;
7633         while (level >= 0) {
7634                 if (path->nodes[level] == NULL) {
7635                         int parent_slot;
7636                         u64 child_gen;
7637                         u64 child_bytenr;
7638
7639                         /* We need to get child blockptr/gen from
7640                          * parent before we can read it. */
7641                         eb = path->nodes[level + 1];
7642                         parent_slot = path->slots[level + 1];
7643                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
7644                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
7645
7646                         eb = read_tree_block(root, child_bytenr, child_gen);
7647                         if (!eb || !extent_buffer_uptodate(eb)) {
7648                                 ret = -EIO;
7649                                 goto out;
7650                         }
7651
7652                         path->nodes[level] = eb;
7653                         path->slots[level] = 0;
7654
7655                         btrfs_tree_read_lock(eb);
7656                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
7657                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
7658
7659                         ret = btrfs_qgroup_record_ref(trans, root->fs_info,
7660                                                 root->objectid,
7661                                                 child_bytenr,
7662                                                 root->nodesize,
7663                                                 BTRFS_QGROUP_OPER_SUB_SUBTREE,
7664                                                 0);
7665                         if (ret)
7666                                 goto out;
7667
7668                 }
7669
7670                 if (level == 0) {
7671                         ret = account_leaf_items(trans, root, path->nodes[level]);
7672                         if (ret)
7673                                 goto out;
7674
7675                         /* Nonzero return here means we completed our search */
7676                         ret = adjust_slots_upwards(root, path, root_level);
7677                         if (ret)
7678                                 break;
7679
7680                         /* Restart search with new slots */
7681                         goto walk_down;
7682                 }
7683
7684                 level--;
7685         }
7686
7687         ret = 0;
7688 out:
7689         btrfs_free_path(path);
7690
7691         return ret;
7692 }
7693
7694 /*
7695  * helper to process tree block while walking down the tree.
7696  *
7697  * when wc->stage == UPDATE_BACKREF, this function updates
7698  * back refs for pointers in the block.
7699  *
7700  * NOTE: return value 1 means we should stop walking down.
7701  */
7702 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
7703                                    struct btrfs_root *root,
7704                                    struct btrfs_path *path,
7705                                    struct walk_control *wc, int lookup_info)
7706 {
7707         int level = wc->level;
7708         struct extent_buffer *eb = path->nodes[level];
7709         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7710         int ret;
7711
7712         if (wc->stage == UPDATE_BACKREF &&
7713             btrfs_header_owner(eb) != root->root_key.objectid)
7714                 return 1;
7715
7716         /*
7717          * when reference count of tree block is 1, it won't increase
7718          * again. once full backref flag is set, we never clear it.
7719          */
7720         if (lookup_info &&
7721             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
7722              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
7723                 BUG_ON(!path->locks[level]);
7724                 ret = btrfs_lookup_extent_info(trans, root,
7725                                                eb->start, level, 1,
7726                                                &wc->refs[level],
7727                                                &wc->flags[level]);
7728                 BUG_ON(ret == -ENOMEM);
7729                 if (ret)
7730                         return ret;
7731                 BUG_ON(wc->refs[level] == 0);
7732         }
7733
7734         if (wc->stage == DROP_REFERENCE) {
7735                 if (wc->refs[level] > 1)
7736                         return 1;
7737
7738                 if (path->locks[level] && !wc->keep_locks) {
7739                         btrfs_tree_unlock_rw(eb, path->locks[level]);
7740                         path->locks[level] = 0;
7741                 }
7742                 return 0;
7743         }
7744
7745         /* wc->stage == UPDATE_BACKREF */
7746         if (!(wc->flags[level] & flag)) {
7747                 BUG_ON(!path->locks[level]);
7748                 ret = btrfs_inc_ref(trans, root, eb, 1);
7749                 BUG_ON(ret); /* -ENOMEM */
7750                 ret = btrfs_dec_ref(trans, root, eb, 0);
7751                 BUG_ON(ret); /* -ENOMEM */
7752                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
7753                                                   eb->len, flag,
7754                                                   btrfs_header_level(eb), 0);
7755                 BUG_ON(ret); /* -ENOMEM */
7756                 wc->flags[level] |= flag;
7757         }
7758
7759         /*
7760          * the block is shared by multiple trees, so it's not good to
7761          * keep the tree lock
7762          */
7763         if (path->locks[level] && level > 0) {
7764                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7765                 path->locks[level] = 0;
7766         }
7767         return 0;
7768 }
7769
7770 /*
7771  * helper to process tree block pointer.
7772  *
7773  * when wc->stage == DROP_REFERENCE, this function checks
7774  * reference count of the block pointed to. if the block
7775  * is shared and we need update back refs for the subtree
7776  * rooted at the block, this function changes wc->stage to
7777  * UPDATE_BACKREF. if the block is shared and there is no
7778  * need to update back, this function drops the reference
7779  * to the block.
7780  *
7781  * NOTE: return value 1 means we should stop walking down.
7782  */
7783 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
7784                                  struct btrfs_root *root,
7785                                  struct btrfs_path *path,
7786                                  struct walk_control *wc, int *lookup_info)
7787 {
7788         u64 bytenr;
7789         u64 generation;
7790         u64 parent;
7791         u32 blocksize;
7792         struct btrfs_key key;
7793         struct extent_buffer *next;
7794         int level = wc->level;
7795         int reada = 0;
7796         int ret = 0;
7797         bool need_account = false;
7798
7799         generation = btrfs_node_ptr_generation(path->nodes[level],
7800                                                path->slots[level]);
7801         /*
7802          * if the lower level block was created before the snapshot
7803          * was created, we know there is no need to update back refs
7804          * for the subtree
7805          */
7806         if (wc->stage == UPDATE_BACKREF &&
7807             generation <= root->root_key.offset) {
7808                 *lookup_info = 1;
7809                 return 1;
7810         }
7811
7812         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
7813         blocksize = root->nodesize;
7814
7815         next = btrfs_find_tree_block(root, bytenr);
7816         if (!next) {
7817                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
7818                 if (!next)
7819                         return -ENOMEM;
7820                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
7821                                                level - 1);
7822                 reada = 1;
7823         }
7824         btrfs_tree_lock(next);
7825         btrfs_set_lock_blocking(next);
7826
7827         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
7828                                        &wc->refs[level - 1],
7829                                        &wc->flags[level - 1]);
7830         if (ret < 0) {
7831                 btrfs_tree_unlock(next);
7832                 return ret;
7833         }
7834
7835         if (unlikely(wc->refs[level - 1] == 0)) {
7836                 btrfs_err(root->fs_info, "Missing references.");
7837                 BUG();
7838         }
7839         *lookup_info = 0;
7840
7841         if (wc->stage == DROP_REFERENCE) {
7842                 if (wc->refs[level - 1] > 1) {
7843                         need_account = true;
7844                         if (level == 1 &&
7845                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7846                                 goto skip;
7847
7848                         if (!wc->update_ref ||
7849                             generation <= root->root_key.offset)
7850                                 goto skip;
7851
7852                         btrfs_node_key_to_cpu(path->nodes[level], &key,
7853                                               path->slots[level]);
7854                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
7855                         if (ret < 0)
7856                                 goto skip;
7857
7858                         wc->stage = UPDATE_BACKREF;
7859                         wc->shared_level = level - 1;
7860                 }
7861         } else {
7862                 if (level == 1 &&
7863                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7864                         goto skip;
7865         }
7866
7867         if (!btrfs_buffer_uptodate(next, generation, 0)) {
7868                 btrfs_tree_unlock(next);
7869                 free_extent_buffer(next);
7870                 next = NULL;
7871                 *lookup_info = 1;
7872         }
7873
7874         if (!next) {
7875                 if (reada && level == 1)
7876                         reada_walk_down(trans, root, wc, path);
7877                 next = read_tree_block(root, bytenr, generation);
7878                 if (!next || !extent_buffer_uptodate(next)) {
7879                         free_extent_buffer(next);
7880                         return -EIO;
7881                 }
7882                 btrfs_tree_lock(next);
7883                 btrfs_set_lock_blocking(next);
7884         }
7885
7886         level--;
7887         BUG_ON(level != btrfs_header_level(next));
7888         path->nodes[level] = next;
7889         path->slots[level] = 0;
7890         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7891         wc->level = level;
7892         if (wc->level == 1)
7893                 wc->reada_slot = 0;
7894         return 0;
7895 skip:
7896         wc->refs[level - 1] = 0;
7897         wc->flags[level - 1] = 0;
7898         if (wc->stage == DROP_REFERENCE) {
7899                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
7900                         parent = path->nodes[level]->start;
7901                 } else {
7902                         BUG_ON(root->root_key.objectid !=
7903                                btrfs_header_owner(path->nodes[level]));
7904                         parent = 0;
7905                 }
7906
7907                 if (need_account) {
7908                         ret = account_shared_subtree(trans, root, next,
7909                                                      generation, level - 1);
7910                         if (ret) {
7911                                 printk_ratelimited(KERN_ERR "BTRFS: %s Error "
7912                                         "%d accounting shared subtree. Quota "
7913                                         "is out of sync, rescan required.\n",
7914                                         root->fs_info->sb->s_id, ret);
7915                         }
7916                 }
7917                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
7918                                 root->root_key.objectid, level - 1, 0, 0);
7919                 BUG_ON(ret); /* -ENOMEM */
7920         }
7921         btrfs_tree_unlock(next);
7922         free_extent_buffer(next);
7923         *lookup_info = 1;
7924         return 1;
7925 }
7926
7927 /*
7928  * helper to process tree block while walking up the tree.
7929  *
7930  * when wc->stage == DROP_REFERENCE, this function drops
7931  * reference count on the block.
7932  *
7933  * when wc->stage == UPDATE_BACKREF, this function changes
7934  * wc->stage back to DROP_REFERENCE if we changed wc->stage
7935  * to UPDATE_BACKREF previously while processing the block.
7936  *
7937  * NOTE: return value 1 means we should stop walking up.
7938  */
7939 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
7940                                  struct btrfs_root *root,
7941                                  struct btrfs_path *path,
7942                                  struct walk_control *wc)
7943 {
7944         int ret;
7945         int level = wc->level;
7946         struct extent_buffer *eb = path->nodes[level];
7947         u64 parent = 0;
7948
7949         if (wc->stage == UPDATE_BACKREF) {
7950                 BUG_ON(wc->shared_level < level);
7951                 if (level < wc->shared_level)
7952                         goto out;
7953
7954                 ret = find_next_key(path, level + 1, &wc->update_progress);
7955                 if (ret > 0)
7956                         wc->update_ref = 0;
7957
7958                 wc->stage = DROP_REFERENCE;
7959                 wc->shared_level = -1;
7960                 path->slots[level] = 0;
7961
7962                 /*
7963                  * check reference count again if the block isn't locked.
7964                  * we should start walking down the tree again if reference
7965                  * count is one.
7966                  */
7967                 if (!path->locks[level]) {
7968                         BUG_ON(level == 0);
7969                         btrfs_tree_lock(eb);
7970                         btrfs_set_lock_blocking(eb);
7971                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7972
7973                         ret = btrfs_lookup_extent_info(trans, root,
7974                                                        eb->start, level, 1,
7975                                                        &wc->refs[level],
7976                                                        &wc->flags[level]);
7977                         if (ret < 0) {
7978                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7979                                 path->locks[level] = 0;
7980                                 return ret;
7981                         }
7982                         BUG_ON(wc->refs[level] == 0);
7983                         if (wc->refs[level] == 1) {
7984                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7985                                 path->locks[level] = 0;
7986                                 return 1;
7987                         }
7988                 }
7989         }
7990
7991         /* wc->stage == DROP_REFERENCE */
7992         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
7993
7994         if (wc->refs[level] == 1) {
7995                 if (level == 0) {
7996                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7997                                 ret = btrfs_dec_ref(trans, root, eb, 1);
7998                         else
7999                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8000                         BUG_ON(ret); /* -ENOMEM */
8001                         ret = account_leaf_items(trans, root, eb);
8002                         if (ret) {
8003                                 printk_ratelimited(KERN_ERR "BTRFS: %s Error "
8004                                         "%d accounting leaf items. Quota "
8005                                         "is out of sync, rescan required.\n",
8006                                         root->fs_info->sb->s_id, ret);
8007                         }
8008                 }
8009                 /* make block locked assertion in clean_tree_block happy */
8010                 if (!path->locks[level] &&
8011                     btrfs_header_generation(eb) == trans->transid) {
8012                         btrfs_tree_lock(eb);
8013                         btrfs_set_lock_blocking(eb);
8014                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8015                 }
8016                 clean_tree_block(trans, root, eb);
8017         }
8018
8019         if (eb == root->node) {
8020                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8021                         parent = eb->start;
8022                 else
8023                         BUG_ON(root->root_key.objectid !=
8024                                btrfs_header_owner(eb));
8025         } else {
8026                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8027                         parent = path->nodes[level + 1]->start;
8028                 else
8029                         BUG_ON(root->root_key.objectid !=
8030                                btrfs_header_owner(path->nodes[level + 1]));
8031         }
8032
8033         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8034 out:
8035         wc->refs[level] = 0;
8036         wc->flags[level] = 0;
8037         return 0;
8038 }
8039
8040 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8041                                    struct btrfs_root *root,
8042                                    struct btrfs_path *path,
8043                                    struct walk_control *wc)
8044 {
8045         int level = wc->level;
8046         int lookup_info = 1;
8047         int ret;
8048
8049         while (level >= 0) {
8050                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8051                 if (ret > 0)
8052                         break;
8053
8054                 if (level == 0)
8055                         break;
8056
8057                 if (path->slots[level] >=
8058                     btrfs_header_nritems(path->nodes[level]))
8059                         break;
8060
8061                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8062                 if (ret > 0) {
8063                         path->slots[level]++;
8064                         continue;
8065                 } else if (ret < 0)
8066                         return ret;
8067                 level = wc->level;
8068         }
8069         return 0;
8070 }
8071
8072 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8073                                  struct btrfs_root *root,
8074                                  struct btrfs_path *path,
8075                                  struct walk_control *wc, int max_level)
8076 {
8077         int level = wc->level;
8078         int ret;
8079
8080         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8081         while (level < max_level && path->nodes[level]) {
8082                 wc->level = level;
8083                 if (path->slots[level] + 1 <
8084                     btrfs_header_nritems(path->nodes[level])) {
8085                         path->slots[level]++;
8086                         return 0;
8087                 } else {
8088                         ret = walk_up_proc(trans, root, path, wc);
8089                         if (ret > 0)
8090                                 return 0;
8091
8092                         if (path->locks[level]) {
8093                                 btrfs_tree_unlock_rw(path->nodes[level],
8094                                                      path->locks[level]);
8095                                 path->locks[level] = 0;
8096                         }
8097                         free_extent_buffer(path->nodes[level]);
8098                         path->nodes[level] = NULL;
8099                         level++;
8100                 }
8101         }
8102         return 1;
8103 }
8104
8105 /*
8106  * drop a subvolume tree.
8107  *
8108  * this function traverses the tree freeing any blocks that only
8109  * referenced by the tree.
8110  *
8111  * when a shared tree block is found. this function decreases its
8112  * reference count by one. if update_ref is true, this function
8113  * also make sure backrefs for the shared block and all lower level
8114  * blocks are properly updated.
8115  *
8116  * If called with for_reloc == 0, may exit early with -EAGAIN
8117  */
8118 int btrfs_drop_snapshot(struct btrfs_root *root,
8119                          struct btrfs_block_rsv *block_rsv, int update_ref,
8120                          int for_reloc)
8121 {
8122         struct btrfs_path *path;
8123         struct btrfs_trans_handle *trans;
8124         struct btrfs_root *tree_root = root->fs_info->tree_root;
8125         struct btrfs_root_item *root_item = &root->root_item;
8126         struct walk_control *wc;
8127         struct btrfs_key key;
8128         int err = 0;
8129         int ret;
8130         int level;
8131         bool root_dropped = false;
8132
8133         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
8134
8135         path = btrfs_alloc_path();
8136         if (!path) {
8137                 err = -ENOMEM;
8138                 goto out;
8139         }
8140
8141         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8142         if (!wc) {
8143                 btrfs_free_path(path);
8144                 err = -ENOMEM;
8145                 goto out;
8146         }
8147
8148         trans = btrfs_start_transaction(tree_root, 0);
8149         if (IS_ERR(trans)) {
8150                 err = PTR_ERR(trans);
8151                 goto out_free;
8152         }
8153
8154         if (block_rsv)
8155                 trans->block_rsv = block_rsv;
8156
8157         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8158                 level = btrfs_header_level(root->node);
8159                 path->nodes[level] = btrfs_lock_root_node(root);
8160                 btrfs_set_lock_blocking(path->nodes[level]);
8161                 path->slots[level] = 0;
8162                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8163                 memset(&wc->update_progress, 0,
8164                        sizeof(wc->update_progress));
8165         } else {
8166                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8167                 memcpy(&wc->update_progress, &key,
8168                        sizeof(wc->update_progress));
8169
8170                 level = root_item->drop_level;
8171                 BUG_ON(level == 0);
8172                 path->lowest_level = level;
8173                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8174                 path->lowest_level = 0;
8175                 if (ret < 0) {
8176                         err = ret;
8177                         goto out_end_trans;
8178                 }
8179                 WARN_ON(ret > 0);
8180
8181                 /*
8182                  * unlock our path, this is safe because only this
8183                  * function is allowed to delete this snapshot
8184                  */
8185                 btrfs_unlock_up_safe(path, 0);
8186
8187                 level = btrfs_header_level(root->node);
8188                 while (1) {
8189                         btrfs_tree_lock(path->nodes[level]);
8190                         btrfs_set_lock_blocking(path->nodes[level]);
8191                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8192
8193                         ret = btrfs_lookup_extent_info(trans, root,
8194                                                 path->nodes[level]->start,
8195                                                 level, 1, &wc->refs[level],
8196                                                 &wc->flags[level]);
8197                         if (ret < 0) {
8198                                 err = ret;
8199                                 goto out_end_trans;
8200                         }
8201                         BUG_ON(wc->refs[level] == 0);
8202
8203                         if (level == root_item->drop_level)
8204                                 break;
8205
8206                         btrfs_tree_unlock(path->nodes[level]);
8207                         path->locks[level] = 0;
8208                         WARN_ON(wc->refs[level] != 1);
8209                         level--;
8210                 }
8211         }
8212
8213         wc->level = level;
8214         wc->shared_level = -1;
8215         wc->stage = DROP_REFERENCE;
8216         wc->update_ref = update_ref;
8217         wc->keep_locks = 0;
8218         wc->for_reloc = for_reloc;
8219         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8220
8221         while (1) {
8222
8223                 ret = walk_down_tree(trans, root, path, wc);
8224                 if (ret < 0) {
8225                         err = ret;
8226                         break;
8227                 }
8228
8229                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8230                 if (ret < 0) {
8231                         err = ret;
8232                         break;
8233                 }
8234
8235                 if (ret > 0) {
8236                         BUG_ON(wc->stage != DROP_REFERENCE);
8237                         break;
8238                 }
8239
8240                 if (wc->stage == DROP_REFERENCE) {
8241                         level = wc->level;
8242                         btrfs_node_key(path->nodes[level],
8243                                        &root_item->drop_progress,
8244                                        path->slots[level]);
8245                         root_item->drop_level = level;
8246                 }
8247
8248                 BUG_ON(wc->level == 0);
8249                 if (btrfs_should_end_transaction(trans, tree_root) ||
8250                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
8251                         ret = btrfs_update_root(trans, tree_root,
8252                                                 &root->root_key,
8253                                                 root_item);
8254                         if (ret) {
8255                                 btrfs_abort_transaction(trans, tree_root, ret);
8256                                 err = ret;
8257                                 goto out_end_trans;
8258                         }
8259
8260                         /*
8261                          * Qgroup update accounting is run from
8262                          * delayed ref handling. This usually works
8263                          * out because delayed refs are normally the
8264                          * only way qgroup updates are added. However,
8265                          * we may have added updates during our tree
8266                          * walk so run qgroups here to make sure we
8267                          * don't lose any updates.
8268                          */
8269                         ret = btrfs_delayed_qgroup_accounting(trans,
8270                                                               root->fs_info);
8271                         if (ret)
8272                                 printk_ratelimited(KERN_ERR "BTRFS: Failure %d "
8273                                                    "running qgroup updates "
8274                                                    "during snapshot delete. "
8275                                                    "Quota is out of sync, "
8276                                                    "rescan required.\n", ret);
8277
8278                         btrfs_end_transaction_throttle(trans, tree_root);
8279                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
8280                                 pr_debug("BTRFS: drop snapshot early exit\n");
8281                                 err = -EAGAIN;
8282                                 goto out_free;
8283                         }
8284
8285                         trans = btrfs_start_transaction(tree_root, 0);
8286                         if (IS_ERR(trans)) {
8287                                 err = PTR_ERR(trans);
8288                                 goto out_free;
8289                         }
8290                         if (block_rsv)
8291                                 trans->block_rsv = block_rsv;
8292                 }
8293         }
8294         btrfs_release_path(path);
8295         if (err)
8296                 goto out_end_trans;
8297
8298         ret = btrfs_del_root(trans, tree_root, &root->root_key);
8299         if (ret) {
8300                 btrfs_abort_transaction(trans, tree_root, ret);
8301                 goto out_end_trans;
8302         }
8303
8304         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
8305                 ret = btrfs_find_root(tree_root, &root->root_key, path,
8306                                       NULL, NULL);
8307                 if (ret < 0) {
8308                         btrfs_abort_transaction(trans, tree_root, ret);
8309                         err = ret;
8310                         goto out_end_trans;
8311                 } else if (ret > 0) {
8312                         /* if we fail to delete the orphan item this time
8313                          * around, it'll get picked up the next time.
8314                          *
8315                          * The most common failure here is just -ENOENT.
8316                          */
8317                         btrfs_del_orphan_item(trans, tree_root,
8318                                               root->root_key.objectid);
8319                 }
8320         }
8321
8322         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
8323                 btrfs_drop_and_free_fs_root(tree_root->fs_info, root);
8324         } else {
8325                 free_extent_buffer(root->node);
8326                 free_extent_buffer(root->commit_root);
8327                 btrfs_put_fs_root(root);
8328         }
8329         root_dropped = true;
8330 out_end_trans:
8331         ret = btrfs_delayed_qgroup_accounting(trans, tree_root->fs_info);
8332         if (ret)
8333                 printk_ratelimited(KERN_ERR "BTRFS: Failure %d "
8334                                    "running qgroup updates "
8335                                    "during snapshot delete. "
8336                                    "Quota is out of sync, "
8337                                    "rescan required.\n", ret);
8338
8339         btrfs_end_transaction_throttle(trans, tree_root);
8340 out_free:
8341         kfree(wc);
8342         btrfs_free_path(path);
8343 out:
8344         /*
8345          * So if we need to stop dropping the snapshot for whatever reason we
8346          * need to make sure to add it back to the dead root list so that we
8347          * keep trying to do the work later.  This also cleans up roots if we
8348          * don't have it in the radix (like when we recover after a power fail
8349          * or unmount) so we don't leak memory.
8350          */
8351         if (!for_reloc && root_dropped == false)
8352                 btrfs_add_dead_root(root);
8353         if (err && err != -EAGAIN)
8354                 btrfs_std_error(root->fs_info, err);
8355         return err;
8356 }
8357
8358 /*
8359  * drop subtree rooted at tree block 'node'.
8360  *
8361  * NOTE: this function will unlock and release tree block 'node'
8362  * only used by relocation code
8363  */
8364 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
8365                         struct btrfs_root *root,
8366                         struct extent_buffer *node,
8367                         struct extent_buffer *parent)
8368 {
8369         struct btrfs_path *path;
8370         struct walk_control *wc;
8371         int level;
8372         int parent_level;
8373         int ret = 0;
8374         int wret;
8375
8376         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
8377
8378         path = btrfs_alloc_path();
8379         if (!path)
8380                 return -ENOMEM;
8381
8382         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8383         if (!wc) {
8384                 btrfs_free_path(path);
8385                 return -ENOMEM;
8386         }
8387
8388         btrfs_assert_tree_locked(parent);
8389         parent_level = btrfs_header_level(parent);
8390         extent_buffer_get(parent);
8391         path->nodes[parent_level] = parent;
8392         path->slots[parent_level] = btrfs_header_nritems(parent);
8393
8394         btrfs_assert_tree_locked(node);
8395         level = btrfs_header_level(node);
8396         path->nodes[level] = node;
8397         path->slots[level] = 0;
8398         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8399
8400         wc->refs[parent_level] = 1;
8401         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8402         wc->level = level;
8403         wc->shared_level = -1;
8404         wc->stage = DROP_REFERENCE;
8405         wc->update_ref = 0;
8406         wc->keep_locks = 1;
8407         wc->for_reloc = 1;
8408         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8409
8410         while (1) {
8411                 wret = walk_down_tree(trans, root, path, wc);
8412                 if (wret < 0) {
8413                         ret = wret;
8414                         break;
8415                 }
8416
8417                 wret = walk_up_tree(trans, root, path, wc, parent_level);
8418                 if (wret < 0)
8419                         ret = wret;
8420                 if (wret != 0)
8421                         break;
8422         }
8423
8424         kfree(wc);
8425         btrfs_free_path(path);
8426         return ret;
8427 }
8428
8429 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
8430 {
8431         u64 num_devices;
8432         u64 stripped;
8433
8434         /*
8435          * if restripe for this chunk_type is on pick target profile and
8436          * return, otherwise do the usual balance
8437          */
8438         stripped = get_restripe_target(root->fs_info, flags);
8439         if (stripped)
8440                 return extended_to_chunk(stripped);
8441
8442         num_devices = root->fs_info->fs_devices->rw_devices;
8443
8444         stripped = BTRFS_BLOCK_GROUP_RAID0 |
8445                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
8446                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
8447
8448         if (num_devices == 1) {
8449                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8450                 stripped = flags & ~stripped;
8451
8452                 /* turn raid0 into single device chunks */
8453                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
8454                         return stripped;
8455
8456                 /* turn mirroring into duplication */
8457                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
8458                              BTRFS_BLOCK_GROUP_RAID10))
8459                         return stripped | BTRFS_BLOCK_GROUP_DUP;
8460         } else {
8461                 /* they already had raid on here, just return */
8462                 if (flags & stripped)
8463                         return flags;
8464
8465                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8466                 stripped = flags & ~stripped;
8467
8468                 /* switch duplicated blocks with raid1 */
8469                 if (flags & BTRFS_BLOCK_GROUP_DUP)
8470                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
8471
8472                 /* this is drive concat, leave it alone */
8473         }
8474
8475         return flags;
8476 }
8477
8478 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
8479 {
8480         struct btrfs_space_info *sinfo = cache->space_info;
8481         u64 num_bytes;
8482         u64 min_allocable_bytes;
8483         int ret = -ENOSPC;
8484
8485
8486         /*
8487          * We need some metadata space and system metadata space for
8488          * allocating chunks in some corner cases until we force to set
8489          * it to be readonly.
8490          */
8491         if ((sinfo->flags &
8492              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
8493             !force)
8494                 min_allocable_bytes = 1 * 1024 * 1024;
8495         else
8496                 min_allocable_bytes = 0;
8497
8498         spin_lock(&sinfo->lock);
8499         spin_lock(&cache->lock);
8500
8501         if (cache->ro) {
8502                 ret = 0;
8503                 goto out;
8504         }
8505
8506         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8507                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8508
8509         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8510             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
8511             min_allocable_bytes <= sinfo->total_bytes) {
8512                 sinfo->bytes_readonly += num_bytes;
8513                 cache->ro = 1;
8514                 ret = 0;
8515         }
8516 out:
8517         spin_unlock(&cache->lock);
8518         spin_unlock(&sinfo->lock);
8519         return ret;
8520 }
8521
8522 int btrfs_set_block_group_ro(struct btrfs_root *root,
8523                              struct btrfs_block_group_cache *cache)
8524
8525 {
8526         struct btrfs_trans_handle *trans;
8527         u64 alloc_flags;
8528         int ret;
8529
8530         BUG_ON(cache->ro);
8531
8532         trans = btrfs_join_transaction(root);
8533         if (IS_ERR(trans))
8534                 return PTR_ERR(trans);
8535
8536         alloc_flags = update_block_group_flags(root, cache->flags);
8537         if (alloc_flags != cache->flags) {
8538                 ret = do_chunk_alloc(trans, root, alloc_flags,
8539                                      CHUNK_ALLOC_FORCE);
8540                 if (ret < 0)
8541                         goto out;
8542         }
8543
8544         ret = set_block_group_ro(cache, 0);
8545         if (!ret)
8546                 goto out;
8547         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
8548         ret = do_chunk_alloc(trans, root, alloc_flags,
8549                              CHUNK_ALLOC_FORCE);
8550         if (ret < 0)
8551                 goto out;
8552         ret = set_block_group_ro(cache, 0);
8553 out:
8554         btrfs_end_transaction(trans, root);
8555         return ret;
8556 }
8557
8558 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
8559                             struct btrfs_root *root, u64 type)
8560 {
8561         u64 alloc_flags = get_alloc_profile(root, type);
8562         return do_chunk_alloc(trans, root, alloc_flags,
8563                               CHUNK_ALLOC_FORCE);
8564 }
8565
8566 /*
8567  * helper to account the unused space of all the readonly block group in the
8568  * list. takes mirrors into account.
8569  */
8570 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
8571 {
8572         struct btrfs_block_group_cache *block_group;
8573         u64 free_bytes = 0;
8574         int factor;
8575
8576         list_for_each_entry(block_group, groups_list, list) {
8577                 spin_lock(&block_group->lock);
8578
8579                 if (!block_group->ro) {
8580                         spin_unlock(&block_group->lock);
8581                         continue;
8582                 }
8583
8584                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
8585                                           BTRFS_BLOCK_GROUP_RAID10 |
8586                                           BTRFS_BLOCK_GROUP_DUP))
8587                         factor = 2;
8588                 else
8589                         factor = 1;
8590
8591                 free_bytes += (block_group->key.offset -
8592                                btrfs_block_group_used(&block_group->item)) *
8593                                factor;
8594
8595                 spin_unlock(&block_group->lock);
8596         }
8597
8598         return free_bytes;
8599 }
8600
8601 /*
8602  * helper to account the unused space of all the readonly block group in the
8603  * space_info. takes mirrors into account.
8604  */
8605 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
8606 {
8607         int i;
8608         u64 free_bytes = 0;
8609
8610         spin_lock(&sinfo->lock);
8611
8612         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
8613                 if (!list_empty(&sinfo->block_groups[i]))
8614                         free_bytes += __btrfs_get_ro_block_group_free_space(
8615                                                 &sinfo->block_groups[i]);
8616
8617         spin_unlock(&sinfo->lock);
8618
8619         return free_bytes;
8620 }
8621
8622 void btrfs_set_block_group_rw(struct btrfs_root *root,
8623                               struct btrfs_block_group_cache *cache)
8624 {
8625         struct btrfs_space_info *sinfo = cache->space_info;
8626         u64 num_bytes;
8627
8628         BUG_ON(!cache->ro);
8629
8630         spin_lock(&sinfo->lock);
8631         spin_lock(&cache->lock);
8632         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8633                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8634         sinfo->bytes_readonly -= num_bytes;
8635         cache->ro = 0;
8636         spin_unlock(&cache->lock);
8637         spin_unlock(&sinfo->lock);
8638 }
8639
8640 /*
8641  * checks to see if its even possible to relocate this block group.
8642  *
8643  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
8644  * ok to go ahead and try.
8645  */
8646 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
8647 {
8648         struct btrfs_block_group_cache *block_group;
8649         struct btrfs_space_info *space_info;
8650         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
8651         struct btrfs_device *device;
8652         struct btrfs_trans_handle *trans;
8653         u64 min_free;
8654         u64 dev_min = 1;
8655         u64 dev_nr = 0;
8656         u64 target;
8657         int index;
8658         int full = 0;
8659         int ret = 0;
8660
8661         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
8662
8663         /* odd, couldn't find the block group, leave it alone */
8664         if (!block_group)
8665                 return -1;
8666
8667         min_free = btrfs_block_group_used(&block_group->item);
8668
8669         /* no bytes used, we're good */
8670         if (!min_free)
8671                 goto out;
8672
8673         space_info = block_group->space_info;
8674         spin_lock(&space_info->lock);
8675
8676         full = space_info->full;
8677
8678         /*
8679          * if this is the last block group we have in this space, we can't
8680          * relocate it unless we're able to allocate a new chunk below.
8681          *
8682          * Otherwise, we need to make sure we have room in the space to handle
8683          * all of the extents from this block group.  If we can, we're good
8684          */
8685         if ((space_info->total_bytes != block_group->key.offset) &&
8686             (space_info->bytes_used + space_info->bytes_reserved +
8687              space_info->bytes_pinned + space_info->bytes_readonly +
8688              min_free < space_info->total_bytes)) {
8689                 spin_unlock(&space_info->lock);
8690                 goto out;
8691         }
8692         spin_unlock(&space_info->lock);
8693
8694         /*
8695          * ok we don't have enough space, but maybe we have free space on our
8696          * devices to allocate new chunks for relocation, so loop through our
8697          * alloc devices and guess if we have enough space.  if this block
8698          * group is going to be restriped, run checks against the target
8699          * profile instead of the current one.
8700          */
8701         ret = -1;
8702
8703         /*
8704          * index:
8705          *      0: raid10
8706          *      1: raid1
8707          *      2: dup
8708          *      3: raid0
8709          *      4: single
8710          */
8711         target = get_restripe_target(root->fs_info, block_group->flags);
8712         if (target) {
8713                 index = __get_raid_index(extended_to_chunk(target));
8714         } else {
8715                 /*
8716                  * this is just a balance, so if we were marked as full
8717                  * we know there is no space for a new chunk
8718                  */
8719                 if (full)
8720                         goto out;
8721
8722                 index = get_block_group_index(block_group);
8723         }
8724
8725         if (index == BTRFS_RAID_RAID10) {
8726                 dev_min = 4;
8727                 /* Divide by 2 */
8728                 min_free >>= 1;
8729         } else if (index == BTRFS_RAID_RAID1) {
8730                 dev_min = 2;
8731         } else if (index == BTRFS_RAID_DUP) {
8732                 /* Multiply by 2 */
8733                 min_free <<= 1;
8734         } else if (index == BTRFS_RAID_RAID0) {
8735                 dev_min = fs_devices->rw_devices;
8736                 do_div(min_free, dev_min);
8737         }
8738
8739         /* We need to do this so that we can look at pending chunks */
8740         trans = btrfs_join_transaction(root);
8741         if (IS_ERR(trans)) {
8742                 ret = PTR_ERR(trans);
8743                 goto out;
8744         }
8745
8746         mutex_lock(&root->fs_info->chunk_mutex);
8747         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
8748                 u64 dev_offset;
8749
8750                 /*
8751                  * check to make sure we can actually find a chunk with enough
8752                  * space to fit our block group in.
8753                  */
8754                 if (device->total_bytes > device->bytes_used + min_free &&
8755                     !device->is_tgtdev_for_dev_replace) {
8756                         ret = find_free_dev_extent(trans, device, min_free,
8757                                                    &dev_offset, NULL);
8758                         if (!ret)
8759                                 dev_nr++;
8760
8761                         if (dev_nr >= dev_min)
8762                                 break;
8763
8764                         ret = -1;
8765                 }
8766         }
8767         mutex_unlock(&root->fs_info->chunk_mutex);
8768         btrfs_end_transaction(trans, root);
8769 out:
8770         btrfs_put_block_group(block_group);
8771         return ret;
8772 }
8773
8774 static int find_first_block_group(struct btrfs_root *root,
8775                 struct btrfs_path *path, struct btrfs_key *key)
8776 {
8777         int ret = 0;
8778         struct btrfs_key found_key;
8779         struct extent_buffer *leaf;
8780         int slot;
8781
8782         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
8783         if (ret < 0)
8784                 goto out;
8785
8786         while (1) {
8787                 slot = path->slots[0];
8788                 leaf = path->nodes[0];
8789                 if (slot >= btrfs_header_nritems(leaf)) {
8790                         ret = btrfs_next_leaf(root, path);
8791                         if (ret == 0)
8792                                 continue;
8793                         if (ret < 0)
8794                                 goto out;
8795                         break;
8796                 }
8797                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
8798
8799                 if (found_key.objectid >= key->objectid &&
8800                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
8801                         ret = 0;
8802                         goto out;
8803                 }
8804                 path->slots[0]++;
8805         }
8806 out:
8807         return ret;
8808 }
8809
8810 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
8811 {
8812         struct btrfs_block_group_cache *block_group;
8813         u64 last = 0;
8814
8815         while (1) {
8816                 struct inode *inode;
8817
8818                 block_group = btrfs_lookup_first_block_group(info, last);
8819                 while (block_group) {
8820                         spin_lock(&block_group->lock);
8821                         if (block_group->iref)
8822                                 break;
8823                         spin_unlock(&block_group->lock);
8824                         block_group = next_block_group(info->tree_root,
8825                                                        block_group);
8826                 }
8827                 if (!block_group) {
8828                         if (last == 0)
8829                                 break;
8830                         last = 0;
8831                         continue;
8832                 }
8833
8834                 inode = block_group->inode;
8835                 block_group->iref = 0;
8836                 block_group->inode = NULL;
8837                 spin_unlock(&block_group->lock);
8838                 iput(inode);
8839                 last = block_group->key.objectid + block_group->key.offset;
8840                 btrfs_put_block_group(block_group);
8841         }
8842 }
8843
8844 int btrfs_free_block_groups(struct btrfs_fs_info *info)
8845 {
8846         struct btrfs_block_group_cache *block_group;
8847         struct btrfs_space_info *space_info;
8848         struct btrfs_caching_control *caching_ctl;
8849         struct rb_node *n;
8850
8851         down_write(&info->commit_root_sem);
8852         while (!list_empty(&info->caching_block_groups)) {
8853                 caching_ctl = list_entry(info->caching_block_groups.next,
8854                                          struct btrfs_caching_control, list);
8855                 list_del(&caching_ctl->list);
8856                 put_caching_control(caching_ctl);
8857         }
8858         up_write(&info->commit_root_sem);
8859
8860         spin_lock(&info->unused_bgs_lock);
8861         while (!list_empty(&info->unused_bgs)) {
8862                 block_group = list_first_entry(&info->unused_bgs,
8863                                                struct btrfs_block_group_cache,
8864                                                bg_list);
8865                 list_del_init(&block_group->bg_list);
8866                 btrfs_put_block_group(block_group);
8867         }
8868         spin_unlock(&info->unused_bgs_lock);
8869
8870         spin_lock(&info->block_group_cache_lock);
8871         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
8872                 block_group = rb_entry(n, struct btrfs_block_group_cache,
8873                                        cache_node);
8874                 rb_erase(&block_group->cache_node,
8875                          &info->block_group_cache_tree);
8876                 spin_unlock(&info->block_group_cache_lock);
8877
8878                 down_write(&block_group->space_info->groups_sem);
8879                 list_del(&block_group->list);
8880                 up_write(&block_group->space_info->groups_sem);
8881
8882                 if (block_group->cached == BTRFS_CACHE_STARTED)
8883                         wait_block_group_cache_done(block_group);
8884
8885                 /*
8886                  * We haven't cached this block group, which means we could
8887                  * possibly have excluded extents on this block group.
8888                  */
8889                 if (block_group->cached == BTRFS_CACHE_NO ||
8890                     block_group->cached == BTRFS_CACHE_ERROR)
8891                         free_excluded_extents(info->extent_root, block_group);
8892
8893                 btrfs_remove_free_space_cache(block_group);
8894                 btrfs_put_block_group(block_group);
8895
8896                 spin_lock(&info->block_group_cache_lock);
8897         }
8898         spin_unlock(&info->block_group_cache_lock);
8899
8900         /* now that all the block groups are freed, go through and
8901          * free all the space_info structs.  This is only called during
8902          * the final stages of unmount, and so we know nobody is
8903          * using them.  We call synchronize_rcu() once before we start,
8904          * just to be on the safe side.
8905          */
8906         synchronize_rcu();
8907
8908         release_global_block_rsv(info);
8909
8910         while (!list_empty(&info->space_info)) {
8911                 int i;
8912
8913                 space_info = list_entry(info->space_info.next,
8914                                         struct btrfs_space_info,
8915                                         list);
8916                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
8917                         if (WARN_ON(space_info->bytes_pinned > 0 ||
8918                             space_info->bytes_reserved > 0 ||
8919                             space_info->bytes_may_use > 0)) {
8920                                 dump_space_info(space_info, 0, 0);
8921                         }
8922                 }
8923                 list_del(&space_info->list);
8924                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
8925                         struct kobject *kobj;
8926                         kobj = space_info->block_group_kobjs[i];
8927                         space_info->block_group_kobjs[i] = NULL;
8928                         if (kobj) {
8929                                 kobject_del(kobj);
8930                                 kobject_put(kobj);
8931                         }
8932                 }
8933                 kobject_del(&space_info->kobj);
8934                 kobject_put(&space_info->kobj);
8935         }
8936         return 0;
8937 }
8938
8939 static void __link_block_group(struct btrfs_space_info *space_info,
8940                                struct btrfs_block_group_cache *cache)
8941 {
8942         int index = get_block_group_index(cache);
8943         bool first = false;
8944
8945         down_write(&space_info->groups_sem);
8946         if (list_empty(&space_info->block_groups[index]))
8947                 first = true;
8948         list_add_tail(&cache->list, &space_info->block_groups[index]);
8949         up_write(&space_info->groups_sem);
8950
8951         if (first) {
8952                 struct raid_kobject *rkobj;
8953                 int ret;
8954
8955                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
8956                 if (!rkobj)
8957                         goto out_err;
8958                 rkobj->raid_type = index;
8959                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
8960                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
8961                                   "%s", get_raid_name(index));
8962                 if (ret) {
8963                         kobject_put(&rkobj->kobj);
8964                         goto out_err;
8965                 }
8966                 space_info->block_group_kobjs[index] = &rkobj->kobj;
8967         }
8968
8969         return;
8970 out_err:
8971         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
8972 }
8973
8974 static struct btrfs_block_group_cache *
8975 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
8976 {
8977         struct btrfs_block_group_cache *cache;
8978
8979         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8980         if (!cache)
8981                 return NULL;
8982
8983         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8984                                         GFP_NOFS);
8985         if (!cache->free_space_ctl) {
8986                 kfree(cache);
8987                 return NULL;
8988         }
8989
8990         cache->key.objectid = start;
8991         cache->key.offset = size;
8992         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8993
8994         cache->sectorsize = root->sectorsize;
8995         cache->fs_info = root->fs_info;
8996         cache->full_stripe_len = btrfs_full_stripe_len(root,
8997                                                &root->fs_info->mapping_tree,
8998                                                start);
8999         atomic_set(&cache->count, 1);
9000         spin_lock_init(&cache->lock);
9001         init_rwsem(&cache->data_rwsem);
9002         INIT_LIST_HEAD(&cache->list);
9003         INIT_LIST_HEAD(&cache->cluster_list);
9004         INIT_LIST_HEAD(&cache->bg_list);
9005         btrfs_init_free_space_ctl(cache);
9006
9007         return cache;
9008 }
9009
9010 int btrfs_read_block_groups(struct btrfs_root *root)
9011 {
9012         struct btrfs_path *path;
9013         int ret;
9014         struct btrfs_block_group_cache *cache;
9015         struct btrfs_fs_info *info = root->fs_info;
9016         struct btrfs_space_info *space_info;
9017         struct btrfs_key key;
9018         struct btrfs_key found_key;
9019         struct extent_buffer *leaf;
9020         int need_clear = 0;
9021         u64 cache_gen;
9022
9023         root = info->extent_root;
9024         key.objectid = 0;
9025         key.offset = 0;
9026         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9027         path = btrfs_alloc_path();
9028         if (!path)
9029                 return -ENOMEM;
9030         path->reada = 1;
9031
9032         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
9033         if (btrfs_test_opt(root, SPACE_CACHE) &&
9034             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
9035                 need_clear = 1;
9036         if (btrfs_test_opt(root, CLEAR_CACHE))
9037                 need_clear = 1;
9038
9039         while (1) {
9040                 ret = find_first_block_group(root, path, &key);
9041                 if (ret > 0)
9042                         break;
9043                 if (ret != 0)
9044                         goto error;
9045
9046                 leaf = path->nodes[0];
9047                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9048
9049                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
9050                                                        found_key.offset);
9051                 if (!cache) {
9052                         ret = -ENOMEM;
9053                         goto error;
9054                 }
9055
9056                 if (need_clear) {
9057                         /*
9058                          * When we mount with old space cache, we need to
9059                          * set BTRFS_DC_CLEAR and set dirty flag.
9060                          *
9061                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9062                          *    truncate the old free space cache inode and
9063                          *    setup a new one.
9064                          * b) Setting 'dirty flag' makes sure that we flush
9065                          *    the new space cache info onto disk.
9066                          */
9067                         cache->disk_cache_state = BTRFS_DC_CLEAR;
9068                         if (btrfs_test_opt(root, SPACE_CACHE))
9069                                 cache->dirty = 1;
9070                 }
9071
9072                 read_extent_buffer(leaf, &cache->item,
9073                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9074                                    sizeof(cache->item));
9075                 cache->flags = btrfs_block_group_flags(&cache->item);
9076
9077                 key.objectid = found_key.objectid + found_key.offset;
9078                 btrfs_release_path(path);
9079
9080                 /*
9081                  * We need to exclude the super stripes now so that the space
9082                  * info has super bytes accounted for, otherwise we'll think
9083                  * we have more space than we actually do.
9084                  */
9085                 ret = exclude_super_stripes(root, cache);
9086                 if (ret) {
9087                         /*
9088                          * We may have excluded something, so call this just in
9089                          * case.
9090                          */
9091                         free_excluded_extents(root, cache);
9092                         btrfs_put_block_group(cache);
9093                         goto error;
9094                 }
9095
9096                 /*
9097                  * check for two cases, either we are full, and therefore
9098                  * don't need to bother with the caching work since we won't
9099                  * find any space, or we are empty, and we can just add all
9100                  * the space in and be done with it.  This saves us _alot_ of
9101                  * time, particularly in the full case.
9102                  */
9103                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
9104                         cache->last_byte_to_unpin = (u64)-1;
9105                         cache->cached = BTRFS_CACHE_FINISHED;
9106                         free_excluded_extents(root, cache);
9107                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9108                         cache->last_byte_to_unpin = (u64)-1;
9109                         cache->cached = BTRFS_CACHE_FINISHED;
9110                         add_new_free_space(cache, root->fs_info,
9111                                            found_key.objectid,
9112                                            found_key.objectid +
9113                                            found_key.offset);
9114                         free_excluded_extents(root, cache);
9115                 }
9116
9117                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
9118                 if (ret) {
9119                         btrfs_remove_free_space_cache(cache);
9120                         btrfs_put_block_group(cache);
9121                         goto error;
9122                 }
9123
9124                 ret = update_space_info(info, cache->flags, found_key.offset,
9125                                         btrfs_block_group_used(&cache->item),
9126                                         &space_info);
9127                 if (ret) {
9128                         btrfs_remove_free_space_cache(cache);
9129                         spin_lock(&info->block_group_cache_lock);
9130                         rb_erase(&cache->cache_node,
9131                                  &info->block_group_cache_tree);
9132                         spin_unlock(&info->block_group_cache_lock);
9133                         btrfs_put_block_group(cache);
9134                         goto error;
9135                 }
9136
9137                 cache->space_info = space_info;
9138                 spin_lock(&cache->space_info->lock);
9139                 cache->space_info->bytes_readonly += cache->bytes_super;
9140                 spin_unlock(&cache->space_info->lock);
9141
9142                 __link_block_group(space_info, cache);
9143
9144                 set_avail_alloc_bits(root->fs_info, cache->flags);
9145                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
9146                         set_block_group_ro(cache, 1);
9147                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9148                         spin_lock(&info->unused_bgs_lock);
9149                         /* Should always be true but just in case. */
9150                         if (list_empty(&cache->bg_list)) {
9151                                 btrfs_get_block_group(cache);
9152                                 list_add_tail(&cache->bg_list,
9153                                               &info->unused_bgs);
9154                         }
9155                         spin_unlock(&info->unused_bgs_lock);
9156                 }
9157         }
9158
9159         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
9160                 if (!(get_alloc_profile(root, space_info->flags) &
9161                       (BTRFS_BLOCK_GROUP_RAID10 |
9162                        BTRFS_BLOCK_GROUP_RAID1 |
9163                        BTRFS_BLOCK_GROUP_RAID5 |
9164                        BTRFS_BLOCK_GROUP_RAID6 |
9165                        BTRFS_BLOCK_GROUP_DUP)))
9166                         continue;
9167                 /*
9168                  * avoid allocating from un-mirrored block group if there are
9169                  * mirrored block groups.
9170                  */
9171                 list_for_each_entry(cache,
9172                                 &space_info->block_groups[BTRFS_RAID_RAID0],
9173                                 list)
9174                         set_block_group_ro(cache, 1);
9175                 list_for_each_entry(cache,
9176                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
9177                                 list)
9178                         set_block_group_ro(cache, 1);
9179         }
9180
9181         init_global_block_rsv(info);
9182         ret = 0;
9183 error:
9184         btrfs_free_path(path);
9185         return ret;
9186 }
9187
9188 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
9189                                        struct btrfs_root *root)
9190 {
9191         struct btrfs_block_group_cache *block_group, *tmp;
9192         struct btrfs_root *extent_root = root->fs_info->extent_root;
9193         struct btrfs_block_group_item item;
9194         struct btrfs_key key;
9195         int ret = 0;
9196
9197         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
9198                 list_del_init(&block_group->bg_list);
9199                 if (ret)
9200                         continue;
9201
9202                 spin_lock(&block_group->lock);
9203                 memcpy(&item, &block_group->item, sizeof(item));
9204                 memcpy(&key, &block_group->key, sizeof(key));
9205                 spin_unlock(&block_group->lock);
9206
9207                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
9208                                         sizeof(item));
9209                 if (ret)
9210                         btrfs_abort_transaction(trans, extent_root, ret);
9211                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
9212                                                key.objectid, key.offset);
9213                 if (ret)
9214                         btrfs_abort_transaction(trans, extent_root, ret);
9215         }
9216 }
9217
9218 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
9219                            struct btrfs_root *root, u64 bytes_used,
9220                            u64 type, u64 chunk_objectid, u64 chunk_offset,
9221                            u64 size)
9222 {
9223         int ret;
9224         struct btrfs_root *extent_root;
9225         struct btrfs_block_group_cache *cache;
9226
9227         extent_root = root->fs_info->extent_root;
9228
9229         btrfs_set_log_full_commit(root->fs_info, trans);
9230
9231         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
9232         if (!cache)
9233                 return -ENOMEM;
9234
9235         btrfs_set_block_group_used(&cache->item, bytes_used);
9236         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
9237         btrfs_set_block_group_flags(&cache->item, type);
9238
9239         cache->flags = type;
9240         cache->last_byte_to_unpin = (u64)-1;
9241         cache->cached = BTRFS_CACHE_FINISHED;
9242         ret = exclude_super_stripes(root, cache);
9243         if (ret) {
9244                 /*
9245                  * We may have excluded something, so call this just in
9246                  * case.
9247                  */
9248                 free_excluded_extents(root, cache);
9249                 btrfs_put_block_group(cache);
9250                 return ret;
9251         }
9252
9253         add_new_free_space(cache, root->fs_info, chunk_offset,
9254                            chunk_offset + size);
9255
9256         free_excluded_extents(root, cache);
9257
9258         ret = btrfs_add_block_group_cache(root->fs_info, cache);
9259         if (ret) {
9260                 btrfs_remove_free_space_cache(cache);
9261                 btrfs_put_block_group(cache);
9262                 return ret;
9263         }
9264
9265         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
9266                                 &cache->space_info);
9267         if (ret) {
9268                 btrfs_remove_free_space_cache(cache);
9269                 spin_lock(&root->fs_info->block_group_cache_lock);
9270                 rb_erase(&cache->cache_node,
9271                          &root->fs_info->block_group_cache_tree);
9272                 spin_unlock(&root->fs_info->block_group_cache_lock);
9273                 btrfs_put_block_group(cache);
9274                 return ret;
9275         }
9276         update_global_block_rsv(root->fs_info);
9277
9278         spin_lock(&cache->space_info->lock);
9279         cache->space_info->bytes_readonly += cache->bytes_super;
9280         spin_unlock(&cache->space_info->lock);
9281
9282         __link_block_group(cache->space_info, cache);
9283
9284         list_add_tail(&cache->bg_list, &trans->new_bgs);
9285
9286         set_avail_alloc_bits(extent_root->fs_info, type);
9287
9288         return 0;
9289 }
9290
9291 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
9292 {
9293         u64 extra_flags = chunk_to_extended(flags) &
9294                                 BTRFS_EXTENDED_PROFILE_MASK;
9295
9296         write_seqlock(&fs_info->profiles_lock);
9297         if (flags & BTRFS_BLOCK_GROUP_DATA)
9298                 fs_info->avail_data_alloc_bits &= ~extra_flags;
9299         if (flags & BTRFS_BLOCK_GROUP_METADATA)
9300                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
9301         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
9302                 fs_info->avail_system_alloc_bits &= ~extra_flags;
9303         write_sequnlock(&fs_info->profiles_lock);
9304 }
9305
9306 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
9307                              struct btrfs_root *root, u64 group_start)
9308 {
9309         struct btrfs_path *path;
9310         struct btrfs_block_group_cache *block_group;
9311         struct btrfs_free_cluster *cluster;
9312         struct btrfs_root *tree_root = root->fs_info->tree_root;
9313         struct btrfs_key key;
9314         struct inode *inode;
9315         struct kobject *kobj = NULL;
9316         int ret;
9317         int index;
9318         int factor;
9319
9320         root = root->fs_info->extent_root;
9321
9322         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
9323         BUG_ON(!block_group);
9324         BUG_ON(!block_group->ro);
9325
9326         /*
9327          * Free the reserved super bytes from this block group before
9328          * remove it.
9329          */
9330         free_excluded_extents(root, block_group);
9331
9332         memcpy(&key, &block_group->key, sizeof(key));
9333         index = get_block_group_index(block_group);
9334         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
9335                                   BTRFS_BLOCK_GROUP_RAID1 |
9336                                   BTRFS_BLOCK_GROUP_RAID10))
9337                 factor = 2;
9338         else
9339                 factor = 1;
9340
9341         /* make sure this block group isn't part of an allocation cluster */
9342         cluster = &root->fs_info->data_alloc_cluster;
9343         spin_lock(&cluster->refill_lock);
9344         btrfs_return_cluster_to_free_space(block_group, cluster);
9345         spin_unlock(&cluster->refill_lock);
9346
9347         /*
9348          * make sure this block group isn't part of a metadata
9349          * allocation cluster
9350          */
9351         cluster = &root->fs_info->meta_alloc_cluster;
9352         spin_lock(&cluster->refill_lock);
9353         btrfs_return_cluster_to_free_space(block_group, cluster);
9354         spin_unlock(&cluster->refill_lock);
9355
9356         path = btrfs_alloc_path();
9357         if (!path) {
9358                 ret = -ENOMEM;
9359                 goto out;
9360         }
9361
9362         inode = lookup_free_space_inode(tree_root, block_group, path);
9363         if (!IS_ERR(inode)) {
9364                 ret = btrfs_orphan_add(trans, inode);
9365                 if (ret) {
9366                         btrfs_add_delayed_iput(inode);
9367                         goto out;
9368                 }
9369                 clear_nlink(inode);
9370                 /* One for the block groups ref */
9371                 spin_lock(&block_group->lock);
9372                 if (block_group->iref) {
9373                         block_group->iref = 0;
9374                         block_group->inode = NULL;
9375                         spin_unlock(&block_group->lock);
9376                         iput(inode);
9377                 } else {
9378                         spin_unlock(&block_group->lock);
9379                 }
9380                 /* One for our lookup ref */
9381                 btrfs_add_delayed_iput(inode);
9382         }
9383
9384         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
9385         key.offset = block_group->key.objectid;
9386         key.type = 0;
9387
9388         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
9389         if (ret < 0)
9390                 goto out;
9391         if (ret > 0)
9392                 btrfs_release_path(path);
9393         if (ret == 0) {
9394                 ret = btrfs_del_item(trans, tree_root, path);
9395                 if (ret)
9396                         goto out;
9397                 btrfs_release_path(path);
9398         }
9399
9400         spin_lock(&root->fs_info->block_group_cache_lock);
9401         rb_erase(&block_group->cache_node,
9402                  &root->fs_info->block_group_cache_tree);
9403
9404         if (root->fs_info->first_logical_byte == block_group->key.objectid)
9405                 root->fs_info->first_logical_byte = (u64)-1;
9406         spin_unlock(&root->fs_info->block_group_cache_lock);
9407
9408         down_write(&block_group->space_info->groups_sem);
9409         /*
9410          * we must use list_del_init so people can check to see if they
9411          * are still on the list after taking the semaphore
9412          */
9413         list_del_init(&block_group->list);
9414         if (list_empty(&block_group->space_info->block_groups[index])) {
9415                 kobj = block_group->space_info->block_group_kobjs[index];
9416                 block_group->space_info->block_group_kobjs[index] = NULL;
9417                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
9418         }
9419         up_write(&block_group->space_info->groups_sem);
9420         if (kobj) {
9421                 kobject_del(kobj);
9422                 kobject_put(kobj);
9423         }
9424
9425         if (block_group->cached == BTRFS_CACHE_STARTED)
9426                 wait_block_group_cache_done(block_group);
9427
9428         btrfs_remove_free_space_cache(block_group);
9429
9430         spin_lock(&block_group->space_info->lock);
9431         block_group->space_info->total_bytes -= block_group->key.offset;
9432         block_group->space_info->bytes_readonly -= block_group->key.offset;
9433         block_group->space_info->disk_total -= block_group->key.offset * factor;
9434         spin_unlock(&block_group->space_info->lock);
9435
9436         memcpy(&key, &block_group->key, sizeof(key));
9437
9438         btrfs_put_block_group(block_group);
9439         btrfs_put_block_group(block_group);
9440
9441         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
9442         if (ret > 0)
9443                 ret = -EIO;
9444         if (ret < 0)
9445                 goto out;
9446
9447         ret = btrfs_del_item(trans, root, path);
9448 out:
9449         btrfs_free_path(path);
9450         return ret;
9451 }
9452
9453 /*
9454  * Process the unused_bgs list and remove any that don't have any allocated
9455  * space inside of them.
9456  */
9457 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
9458 {
9459         struct btrfs_block_group_cache *block_group;
9460         struct btrfs_space_info *space_info;
9461         struct btrfs_root *root = fs_info->extent_root;
9462         struct btrfs_trans_handle *trans;
9463         int ret = 0;
9464
9465         if (!fs_info->open)
9466                 return;
9467
9468         spin_lock(&fs_info->unused_bgs_lock);
9469         while (!list_empty(&fs_info->unused_bgs)) {
9470                 u64 start, end;
9471
9472                 block_group = list_first_entry(&fs_info->unused_bgs,
9473                                                struct btrfs_block_group_cache,
9474                                                bg_list);
9475                 space_info = block_group->space_info;
9476                 list_del_init(&block_group->bg_list);
9477                 if (ret || btrfs_mixed_space_info(space_info)) {
9478                         btrfs_put_block_group(block_group);
9479                         continue;
9480                 }
9481                 spin_unlock(&fs_info->unused_bgs_lock);
9482
9483                 /* Don't want to race with allocators so take the groups_sem */
9484                 down_write(&space_info->groups_sem);
9485                 spin_lock(&block_group->lock);
9486                 if (block_group->reserved ||
9487                     btrfs_block_group_used(&block_group->item) ||
9488                     block_group->ro) {
9489                         /*
9490                          * We want to bail if we made new allocations or have
9491                          * outstanding allocations in this block group.  We do
9492                          * the ro check in case balance is currently acting on
9493                          * this block group.
9494                          */
9495                         spin_unlock(&block_group->lock);
9496                         up_write(&space_info->groups_sem);
9497                         goto next;
9498                 }
9499                 spin_unlock(&block_group->lock);
9500
9501                 /* We don't want to force the issue, only flip if it's ok. */
9502                 ret = set_block_group_ro(block_group, 0);
9503                 up_write(&space_info->groups_sem);
9504                 if (ret < 0) {
9505                         ret = 0;
9506                         goto next;
9507                 }
9508
9509                 /*
9510                  * Want to do this before we do anything else so we can recover
9511                  * properly if we fail to join the transaction.
9512                  */
9513                 trans = btrfs_join_transaction(root);
9514                 if (IS_ERR(trans)) {
9515                         btrfs_set_block_group_rw(root, block_group);
9516                         ret = PTR_ERR(trans);
9517                         goto next;
9518                 }
9519
9520                 /*
9521                  * We could have pending pinned extents for this block group,
9522                  * just delete them, we don't care about them anymore.
9523                  */
9524                 start = block_group->key.objectid;
9525                 end = start + block_group->key.offset - 1;
9526                 clear_extent_bits(&fs_info->freed_extents[0], start, end,
9527                                   EXTENT_DIRTY, GFP_NOFS);
9528                 clear_extent_bits(&fs_info->freed_extents[1], start, end,
9529                                   EXTENT_DIRTY, GFP_NOFS);
9530
9531                 /* Reset pinned so btrfs_put_block_group doesn't complain */
9532                 block_group->pinned = 0;
9533
9534                 /*
9535                  * Btrfs_remove_chunk will abort the transaction if things go
9536                  * horribly wrong.
9537                  */
9538                 ret = btrfs_remove_chunk(trans, root,
9539                                          block_group->key.objectid);
9540                 btrfs_end_transaction(trans, root);
9541 next:
9542                 btrfs_put_block_group(block_group);
9543                 spin_lock(&fs_info->unused_bgs_lock);
9544         }
9545         spin_unlock(&fs_info->unused_bgs_lock);
9546 }
9547
9548 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
9549 {
9550         struct btrfs_space_info *space_info;
9551         struct btrfs_super_block *disk_super;
9552         u64 features;
9553         u64 flags;
9554         int mixed = 0;
9555         int ret;
9556
9557         disk_super = fs_info->super_copy;
9558         if (!btrfs_super_root(disk_super))
9559                 return 1;
9560
9561         features = btrfs_super_incompat_flags(disk_super);
9562         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
9563                 mixed = 1;
9564
9565         flags = BTRFS_BLOCK_GROUP_SYSTEM;
9566         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9567         if (ret)
9568                 goto out;
9569
9570         if (mixed) {
9571                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
9572                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9573         } else {
9574                 flags = BTRFS_BLOCK_GROUP_METADATA;
9575                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9576                 if (ret)
9577                         goto out;
9578
9579                 flags = BTRFS_BLOCK_GROUP_DATA;
9580                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9581         }
9582 out:
9583         return ret;
9584 }
9585
9586 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
9587 {
9588         return unpin_extent_range(root, start, end);
9589 }
9590
9591 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
9592                                u64 num_bytes, u64 *actual_bytes)
9593 {
9594         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
9595 }
9596
9597 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
9598 {
9599         struct btrfs_fs_info *fs_info = root->fs_info;
9600         struct btrfs_block_group_cache *cache = NULL;
9601         u64 group_trimmed;
9602         u64 start;
9603         u64 end;
9604         u64 trimmed = 0;
9605         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
9606         int ret = 0;
9607
9608         /*
9609          * try to trim all FS space, our block group may start from non-zero.
9610          */
9611         if (range->len == total_bytes)
9612                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
9613         else
9614                 cache = btrfs_lookup_block_group(fs_info, range->start);
9615
9616         while (cache) {
9617                 if (cache->key.objectid >= (range->start + range->len)) {
9618                         btrfs_put_block_group(cache);
9619                         break;
9620                 }
9621
9622                 start = max(range->start, cache->key.objectid);
9623                 end = min(range->start + range->len,
9624                                 cache->key.objectid + cache->key.offset);
9625
9626                 if (end - start >= range->minlen) {
9627                         if (!block_group_cache_done(cache)) {
9628                                 ret = cache_block_group(cache, 0);
9629                                 if (ret) {
9630                                         btrfs_put_block_group(cache);
9631                                         break;
9632                                 }
9633                                 ret = wait_block_group_cache_done(cache);
9634                                 if (ret) {
9635                                         btrfs_put_block_group(cache);
9636                                         break;
9637                                 }
9638                         }
9639                         ret = btrfs_trim_block_group(cache,
9640                                                      &group_trimmed,
9641                                                      start,
9642                                                      end,
9643                                                      range->minlen);
9644
9645                         trimmed += group_trimmed;
9646                         if (ret) {
9647                                 btrfs_put_block_group(cache);
9648                                 break;
9649                         }
9650                 }
9651
9652                 cache = next_block_group(fs_info->tree_root, cache);
9653         }
9654
9655         range->len = trimmed;
9656         return ret;
9657 }
9658
9659 /*
9660  * btrfs_{start,end}_write() is similar to mnt_{want, drop}_write(),
9661  * they are used to prevent the some tasks writing data into the page cache
9662  * by nocow before the subvolume is snapshoted, but flush the data into
9663  * the disk after the snapshot creation.
9664  */
9665 void btrfs_end_nocow_write(struct btrfs_root *root)
9666 {
9667         percpu_counter_dec(&root->subv_writers->counter);
9668         /*
9669          * Make sure counter is updated before we wake up
9670          * waiters.
9671          */
9672         smp_mb();
9673         if (waitqueue_active(&root->subv_writers->wait))
9674                 wake_up(&root->subv_writers->wait);
9675 }
9676
9677 int btrfs_start_nocow_write(struct btrfs_root *root)
9678 {
9679         if (atomic_read(&root->will_be_snapshoted))
9680                 return 0;
9681
9682         percpu_counter_inc(&root->subv_writers->counter);
9683         /*
9684          * Make sure counter is updated before we check for snapshot creation.
9685          */
9686         smp_mb();
9687         if (atomic_read(&root->will_be_snapshoted)) {
9688                 btrfs_end_nocow_write(root);
9689                 return 0;
9690         }
9691         return 1;
9692 }