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_trans_handle *trans,
  78                               struct btrfs_root *root, u64 bytenr,
  79                               u64 num_bytes, int alloc);
  80 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  81                                 struct btrfs_root *root,
  82                                 struct btrfs_delayed_ref_node *node, u64 parent,
  83                                 u64 root_objectid, u64 owner_objectid,
  84                                 u64 owner_offset, int refs_to_drop,
  85                                 struct btrfs_delayed_extent_op *extra_op);
  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->caching_ctl) {
 319                 spin_unlock(&cache->lock);
 320                 return NULL;
 321         }
 322
 323         ctl = cache->caching_ctl;
 324         atomic_inc(&ctl->count);
 325         spin_unlock(&cache->lock);
 326         return ctl;
 327 }
 328
 329 static void put_caching_control(struct btrfs_caching_control *ctl)
 330 {
 331         if (atomic_dec_and_test(&ctl->count))
 332                 kfree(ctl);
 333 }
 334
 335 /*
 336  * this is only called by cache_block_group, since we could have freed extents
 337  * we need to check the pinned_extents for any extents that can't be used yet
 338  * since their free space will be released as soon as the transaction commits.
 339  */
 340 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 341                               struct btrfs_fs_info *info, u64 start, u64 end)
 342 {
 343         u64 extent_start, extent_end, size, total_added = 0;
 344         int ret;
 345
 346         while (start < end) {
 347                 ret = find_first_extent_bit(info->pinned_extents, start,
 348                                             &extent_start, &extent_end,
 349                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 350                                             NULL);
 351                 if (ret)
 352                         break;
 353
 354                 if (extent_start <= start) {
 355                         start = extent_end + 1;
 356                 } else if (extent_start > start && extent_start < end) {
 357                         size = extent_start - start;
 358                         total_added += size;
 359                         ret = btrfs_add_free_space(block_group, start,
 360                                                    size);
 361                         BUG_ON(ret); /* -ENOMEM or logic error */
 362                         start = extent_end + 1;
 363                 } else {
 364                         break;
 365                 }
 366         }
 367
 368         if (start < end) {
 369                 size = end - start;
 370                 total_added += size;
 371                 ret = btrfs_add_free_space(block_group, start, size);
 372                 BUG_ON(ret); /* -ENOMEM or logic error */
 373         }
 374
 375         return total_added;
 376 }
 377
 378 static noinline void caching_thread(struct btrfs_work *work)
 379 {
 380         struct btrfs_block_group_cache *block_group;
 381         struct btrfs_fs_info *fs_info;
 382         struct btrfs_caching_control *caching_ctl;
 383         struct btrfs_root *extent_root;
 384         struct btrfs_path *path;
 385         struct extent_buffer *leaf;
 386         struct btrfs_key key;
 387         u64 total_found = 0;
 388         u64 last = 0;
 389         u32 nritems;
 390         int ret = -ENOMEM;
 391
 392         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 393         block_group = caching_ctl->block_group;
 394         fs_info = block_group->fs_info;
 395         extent_root = fs_info->extent_root;
 396
 397         path = btrfs_alloc_path();
 398         if (!path)
 399                 goto out;
 400
 401         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 402
 403         /*
 404          * We don't want to deadlock with somebody trying to allocate a new
 405          * extent for the extent root while also trying to search the extent
 406          * root to add free space.  So we skip locking and search the commit
 407          * root, since its read-only
 408          */
 409         path->skip_locking = 1;
 410         path->search_commit_root = 1;
 411         path->reada = 1;
 412
 413         key.objectid = last;
 414         key.offset = 0;
 415         key.type = BTRFS_EXTENT_ITEM_KEY;
 416 again:
 417         mutex_lock(&caching_ctl->mutex);
 418         /* need to make sure the commit_root doesn't disappear */
 419         down_read(&fs_info->commit_root_sem);
 420
 421 next:
 422         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 423         if (ret < 0)
 424                 goto err;
 425
 426         leaf = path->nodes[0];
 427         nritems = btrfs_header_nritems(leaf);
 428
 429         while (1) {
 430                 if (btrfs_fs_closing(fs_info) > 1) {
 431                         last = (u64)-1;
 432                         break;
 433                 }
 434
 435                 if (path->slots[0] < nritems) {
 436                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 437                 } else {
 438                         ret = find_next_key(path, 0, &key);
 439                         if (ret)
 440                                 break;
 441
 442                         if (need_resched() ||
 443                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 444                                 caching_ctl->progress = last;
 445                                 btrfs_release_path(path);
 446                                 up_read(&fs_info->commit_root_sem);
 447                                 mutex_unlock(&caching_ctl->mutex);
 448                                 cond_resched();
 449                                 goto again;
 450                         }
 451
 452                         ret = btrfs_next_leaf(extent_root, path);
 453                         if (ret < 0)
 454                                 goto err;
 455                         if (ret)
 456                                 break;
 457                         leaf = path->nodes[0];
 458                         nritems = btrfs_header_nritems(leaf);
 459                         continue;
 460                 }
 461
 462                 if (key.objectid < last) {
 463                         key.objectid = last;
 464                         key.offset = 0;
 465                         key.type = BTRFS_EXTENT_ITEM_KEY;
 466
 467                         caching_ctl->progress = last;
 468                         btrfs_release_path(path);
 469                         goto next;
 470                 }
 471
 472                 if (key.objectid < block_group->key.objectid) {
 473                         path->slots[0]++;
 474                         continue;
 475                 }
 476
 477                 if (key.objectid >= block_group->key.objectid +
 478                     block_group->key.offset)
 479                         break;
 480
 481                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 482                     key.type == BTRFS_METADATA_ITEM_KEY) {
 483                         total_found += add_new_free_space(block_group,
 484                                                           fs_info, last,
 485                                                           key.objectid);
 486                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 487                                 last = key.objectid +
 488                                         fs_info->tree_root->nodesize;
 489                         else
 490                                 last = key.objectid + key.offset;
 491
 492                         if (total_found > (1024 * 1024 * 2)) {
 493                                 total_found = 0;
 494                                 wake_up(&caching_ctl->wait);
 495                         }
 496                 }
 497                 path->slots[0]++;
 498         }
 499         ret = 0;
 500
 501         total_found += add_new_free_space(block_group, fs_info, last,
 502                                           block_group->key.objectid +
 503                                           block_group->key.offset);
 504         caching_ctl->progress = (u64)-1;
 505
 506         spin_lock(&block_group->lock);
 507         block_group->caching_ctl = NULL;
 508         block_group->cached = BTRFS_CACHE_FINISHED;
 509         spin_unlock(&block_group->lock);
 510
 511 err:
 512         btrfs_free_path(path);
 513         up_read(&fs_info->commit_root_sem);
 514
 515         free_excluded_extents(extent_root, block_group);
 516
 517         mutex_unlock(&caching_ctl->mutex);
 518 out:
 519         if (ret) {
 520                 spin_lock(&block_group->lock);
 521                 block_group->caching_ctl = NULL;
 522                 block_group->cached = BTRFS_CACHE_ERROR;
 523                 spin_unlock(&block_group->lock);
 524         }
 525         wake_up(&caching_ctl->wait);
 526
 527         put_caching_control(caching_ctl);
 528         btrfs_put_block_group(block_group);
 529 }
 530
 531 static int cache_block_group(struct btrfs_block_group_cache *cache,
 532                              int load_cache_only)
 533 {
 534         DEFINE_WAIT(wait);
 535         struct btrfs_fs_info *fs_info = cache->fs_info;
 536         struct btrfs_caching_control *caching_ctl;
 537         int ret = 0;
 538
 539         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 540         if (!caching_ctl)
 541                 return -ENOMEM;
 542
 543         INIT_LIST_HEAD(&caching_ctl->list);
 544         mutex_init(&caching_ctl->mutex);
 545         init_waitqueue_head(&caching_ctl->wait);
 546         caching_ctl->block_group = cache;
 547         caching_ctl->progress = cache->key.objectid;
 548         atomic_set(&caching_ctl->count, 1);
 549         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 550                         caching_thread, NULL, NULL);
 551
 552         spin_lock(&cache->lock);
 553         /*
 554          * This should be a rare occasion, but this could happen I think in the
 555          * case where one thread starts to load the space cache info, and then
 556          * some other thread starts a transaction commit which tries to do an
 557          * allocation while the other thread is still loading the space cache
 558          * info.  The previous loop should have kept us from choosing this block
 559          * group, but if we've moved to the state where we will wait on caching
 560          * block groups we need to first check if we're doing a fast load here,
 561          * so we can wait for it to finish, otherwise we could end up allocating
 562          * from a block group who's cache gets evicted for one reason or
 563          * another.
 564          */
 565         while (cache->cached == BTRFS_CACHE_FAST) {
 566                 struct btrfs_caching_control *ctl;
 567
 568                 ctl = cache->caching_ctl;
 569                 atomic_inc(&ctl->count);
 570                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 571                 spin_unlock(&cache->lock);
 572
 573                 schedule();
 574
 575                 finish_wait(&ctl->wait, &wait);
 576                 put_caching_control(ctl);
 577                 spin_lock(&cache->lock);
 578         }
 579
 580         if (cache->cached != BTRFS_CACHE_NO) {
 581                 spin_unlock(&cache->lock);
 582                 kfree(caching_ctl);
 583                 return 0;
 584         }
 585         WARN_ON(cache->caching_ctl);
 586         cache->caching_ctl = caching_ctl;
 587         cache->cached = BTRFS_CACHE_FAST;
 588         spin_unlock(&cache->lock);
 589
 590         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 591                 mutex_lock(&caching_ctl->mutex);
 592                 ret = load_free_space_cache(fs_info, cache);
 593
 594                 spin_lock(&cache->lock);
 595                 if (ret == 1) {
 596                         cache->caching_ctl = NULL;
 597                         cache->cached = BTRFS_CACHE_FINISHED;
 598                         cache->last_byte_to_unpin = (u64)-1;
 599                         caching_ctl->progress = (u64)-1;
 600                 } else {
 601                         if (load_cache_only) {
 602                                 cache->caching_ctl = NULL;
 603                                 cache->cached = BTRFS_CACHE_NO;
 604                         } else {
 605                                 cache->cached = BTRFS_CACHE_STARTED;
 606                                 cache->has_caching_ctl = 1;
 607                         }
 608                 }
 609                 spin_unlock(&cache->lock);
 610                 mutex_unlock(&caching_ctl->mutex);
 611
 612                 wake_up(&caching_ctl->wait);
 613                 if (ret == 1) {
 614                         put_caching_control(caching_ctl);
 615                         free_excluded_extents(fs_info->extent_root, cache);
 616                         return 0;
 617                 }
 618         } else {
 619                 /*
 620                  * We are not going to do the fast caching, set cached to the
 621                  * appropriate value and wakeup any waiters.
 622                  */
 623                 spin_lock(&cache->lock);
 624                 if (load_cache_only) {
 625                         cache->caching_ctl = NULL;
 626                         cache->cached = BTRFS_CACHE_NO;
 627                 } else {
 628                         cache->cached = BTRFS_CACHE_STARTED;
 629                         cache->has_caching_ctl = 1;
 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 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                 btrfs_put_bbio(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                                   struct btrfs_delayed_ref_node *node,
1970                                   u64 parent, u64 root_objectid,
1971                                   u64 owner, u64 offset, int refs_to_add,
1972                                   struct btrfs_delayed_extent_op *extent_op)
1973 {
1974         struct btrfs_fs_info *fs_info = root->fs_info;
1975         struct btrfs_path *path;
1976         struct extent_buffer *leaf;
1977         struct btrfs_extent_item *item;
1978         struct btrfs_key key;
1979         u64 bytenr = node->bytenr;
1980         u64 num_bytes = node->num_bytes;
1981         u64 refs;
1982         int ret;
1983         int no_quota = node->no_quota;
1984
1985         path = btrfs_alloc_path();
1986         if (!path)
1987                 return -ENOMEM;
1988
1989         if (!is_fstree(root_objectid) || !root->fs_info->quota_enabled)
1990                 no_quota = 1;
1991
1992         path->reada = 1;
1993         path->leave_spinning = 1;
1994         /* this will setup the path even if it fails to insert the back ref */
1995         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
1996                                            bytenr, num_bytes, parent,
1997                                            root_objectid, owner, offset,
1998                                            refs_to_add, extent_op);
1999         if ((ret < 0 && ret != -EAGAIN) || !ret)
2000                 goto out;
2001
2002         /*
2003          * Ok we had -EAGAIN which means we didn't have space to insert and
2004          * inline extent ref, so just update the reference count and add a
2005          * normal backref.
2006          */
2007         leaf = path->nodes[0];
2008         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2009         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2010         refs = btrfs_extent_refs(leaf, item);
2011         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2012         if (extent_op)
2013                 __run_delayed_extent_op(extent_op, leaf, item);
2014
2015         btrfs_mark_buffer_dirty(leaf);
2016         btrfs_release_path(path);
2017
2018         path->reada = 1;
2019         path->leave_spinning = 1;
2020         /* now insert the actual backref */
2021         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2022                                     path, bytenr, parent, root_objectid,
2023                                     owner, offset, refs_to_add);
2024         if (ret)
2025                 btrfs_abort_transaction(trans, root, ret);
2026 out:
2027         btrfs_free_path(path);
2028         return ret;
2029 }
2030
2031 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2032                                 struct btrfs_root *root,
2033                                 struct btrfs_delayed_ref_node *node,
2034                                 struct btrfs_delayed_extent_op *extent_op,
2035                                 int insert_reserved)
2036 {
2037         int ret = 0;
2038         struct btrfs_delayed_data_ref *ref;
2039         struct btrfs_key ins;
2040         u64 parent = 0;
2041         u64 ref_root = 0;
2042         u64 flags = 0;
2043
2044         ins.objectid = node->bytenr;
2045         ins.offset = node->num_bytes;
2046         ins.type = BTRFS_EXTENT_ITEM_KEY;
2047
2048         ref = btrfs_delayed_node_to_data_ref(node);
2049         trace_run_delayed_data_ref(node, ref, node->action);
2050
2051         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2052                 parent = ref->parent;
2053         ref_root = ref->root;
2054
2055         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2056                 if (extent_op)
2057                         flags |= extent_op->flags_to_set;
2058                 ret = alloc_reserved_file_extent(trans, root,
2059                                                  parent, ref_root, flags,
2060                                                  ref->objectid, ref->offset,
2061                                                  &ins, node->ref_mod);
2062         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2063                 ret = __btrfs_inc_extent_ref(trans, root, node, parent,
2064                                              ref_root, ref->objectid,
2065                                              ref->offset, node->ref_mod,
2066                                              extent_op);
2067         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2068                 ret = __btrfs_free_extent(trans, root, node, parent,
2069                                           ref_root, ref->objectid,
2070                                           ref->offset, node->ref_mod,
2071                                           extent_op);
2072         } else {
2073                 BUG();
2074         }
2075         return ret;
2076 }
2077
2078 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2079                                     struct extent_buffer *leaf,
2080                                     struct btrfs_extent_item *ei)
2081 {
2082         u64 flags = btrfs_extent_flags(leaf, ei);
2083         if (extent_op->update_flags) {
2084                 flags |= extent_op->flags_to_set;
2085                 btrfs_set_extent_flags(leaf, ei, flags);
2086         }
2087
2088         if (extent_op->update_key) {
2089                 struct btrfs_tree_block_info *bi;
2090                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2091                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2092                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2093         }
2094 }
2095
2096 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2097                                  struct btrfs_root *root,
2098                                  struct btrfs_delayed_ref_node *node,
2099                                  struct btrfs_delayed_extent_op *extent_op)
2100 {
2101         struct btrfs_key key;
2102         struct btrfs_path *path;
2103         struct btrfs_extent_item *ei;
2104         struct extent_buffer *leaf;
2105         u32 item_size;
2106         int ret;
2107         int err = 0;
2108         int metadata = !extent_op->is_data;
2109
2110         if (trans->aborted)
2111                 return 0;
2112
2113         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2114                 metadata = 0;
2115
2116         path = btrfs_alloc_path();
2117         if (!path)
2118                 return -ENOMEM;
2119
2120         key.objectid = node->bytenr;
2121
2122         if (metadata) {
2123                 key.type = BTRFS_METADATA_ITEM_KEY;
2124                 key.offset = extent_op->level;
2125         } else {
2126                 key.type = BTRFS_EXTENT_ITEM_KEY;
2127                 key.offset = node->num_bytes;
2128         }
2129
2130 again:
2131         path->reada = 1;
2132         path->leave_spinning = 1;
2133         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2134                                 path, 0, 1);
2135         if (ret < 0) {
2136                 err = ret;
2137                 goto out;
2138         }
2139         if (ret > 0) {
2140                 if (metadata) {
2141                         if (path->slots[0] > 0) {
2142                                 path->slots[0]--;
2143                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2144                                                       path->slots[0]);
2145                                 if (key.objectid == node->bytenr &&
2146                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2147                                     key.offset == node->num_bytes)
2148                                         ret = 0;
2149                         }
2150                         if (ret > 0) {
2151                                 btrfs_release_path(path);
2152                                 metadata = 0;
2153
2154                                 key.objectid = node->bytenr;
2155                                 key.offset = node->num_bytes;
2156                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2157                                 goto again;
2158                         }
2159                 } else {
2160                         err = -EIO;
2161                         goto out;
2162                 }
2163         }
2164
2165         leaf = path->nodes[0];
2166         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2167 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2168         if (item_size < sizeof(*ei)) {
2169                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2170                                              path, (u64)-1, 0);
2171                 if (ret < 0) {
2172                         err = ret;
2173                         goto out;
2174                 }
2175                 leaf = path->nodes[0];
2176                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2177         }
2178 #endif
2179         BUG_ON(item_size < sizeof(*ei));
2180         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2181         __run_delayed_extent_op(extent_op, leaf, ei);
2182
2183         btrfs_mark_buffer_dirty(leaf);
2184 out:
2185         btrfs_free_path(path);
2186         return err;
2187 }
2188
2189 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2190                                 struct btrfs_root *root,
2191                                 struct btrfs_delayed_ref_node *node,
2192                                 struct btrfs_delayed_extent_op *extent_op,
2193                                 int insert_reserved)
2194 {
2195         int ret = 0;
2196         struct btrfs_delayed_tree_ref *ref;
2197         struct btrfs_key ins;
2198         u64 parent = 0;
2199         u64 ref_root = 0;
2200         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2201                                                  SKINNY_METADATA);
2202
2203         ref = btrfs_delayed_node_to_tree_ref(node);
2204         trace_run_delayed_tree_ref(node, ref, node->action);
2205
2206         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2207                 parent = ref->parent;
2208         ref_root = ref->root;
2209
2210         ins.objectid = node->bytenr;
2211         if (skinny_metadata) {
2212                 ins.offset = ref->level;
2213                 ins.type = BTRFS_METADATA_ITEM_KEY;
2214         } else {
2215                 ins.offset = node->num_bytes;
2216                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2217         }
2218
2219         BUG_ON(node->ref_mod != 1);
2220         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2221                 BUG_ON(!extent_op || !extent_op->update_flags);
2222                 ret = alloc_reserved_tree_block(trans, root,
2223                                                 parent, ref_root,
2224                                                 extent_op->flags_to_set,
2225                                                 &extent_op->key,
2226                                                 ref->level, &ins,
2227                                                 node->no_quota);
2228         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2229                 ret = __btrfs_inc_extent_ref(trans, root, node,
2230                                              parent, ref_root,
2231                                              ref->level, 0, 1,
2232                                              extent_op);
2233         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2234                 ret = __btrfs_free_extent(trans, root, node,
2235                                           parent, ref_root,
2236                                           ref->level, 0, 1, extent_op);
2237         } else {
2238                 BUG();
2239         }
2240         return ret;
2241 }
2242
2243 /* helper function to actually process a single delayed ref entry */
2244 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2245                                struct btrfs_root *root,
2246                                struct btrfs_delayed_ref_node *node,
2247                                struct btrfs_delayed_extent_op *extent_op,
2248                                int insert_reserved)
2249 {
2250         int ret = 0;
2251
2252         if (trans->aborted) {
2253                 if (insert_reserved)
2254                         btrfs_pin_extent(root, node->bytenr,
2255                                          node->num_bytes, 1);
2256                 return 0;
2257         }
2258
2259         if (btrfs_delayed_ref_is_head(node)) {
2260                 struct btrfs_delayed_ref_head *head;
2261                 /*
2262                  * we've hit the end of the chain and we were supposed
2263                  * to insert this extent into the tree.  But, it got
2264                  * deleted before we ever needed to insert it, so all
2265                  * we have to do is clean up the accounting
2266                  */
2267                 BUG_ON(extent_op);
2268                 head = btrfs_delayed_node_to_head(node);
2269                 trace_run_delayed_ref_head(node, head, node->action);
2270
2271                 if (insert_reserved) {
2272                         btrfs_pin_extent(root, node->bytenr,
2273                                          node->num_bytes, 1);
2274                         if (head->is_data) {
2275                                 ret = btrfs_del_csums(trans, root,
2276                                                       node->bytenr,
2277                                                       node->num_bytes);
2278                         }
2279                 }
2280                 return ret;
2281         }
2282
2283         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2284             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2285                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2286                                            insert_reserved);
2287         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2288                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2289                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2290                                            insert_reserved);
2291         else
2292                 BUG();
2293         return ret;
2294 }
2295
2296 static inline struct btrfs_delayed_ref_node *
2297 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2298 {
2299         if (list_empty(&head->ref_list))
2300                 return NULL;
2301
2302         return list_entry(head->ref_list.next, struct btrfs_delayed_ref_node,
2303                           list);
2304 }
2305
2306 /*
2307  * Returns 0 on success or if called with an already aborted transaction.
2308  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2309  */
2310 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2311                                              struct btrfs_root *root,
2312                                              unsigned long nr)
2313 {
2314         struct btrfs_delayed_ref_root *delayed_refs;
2315         struct btrfs_delayed_ref_node *ref;
2316         struct btrfs_delayed_ref_head *locked_ref = NULL;
2317         struct btrfs_delayed_extent_op *extent_op;
2318         struct btrfs_fs_info *fs_info = root->fs_info;
2319         ktime_t start = ktime_get();
2320         int ret;
2321         unsigned long count = 0;
2322         unsigned long actual_count = 0;
2323         int must_insert_reserved = 0;
2324
2325         delayed_refs = &trans->transaction->delayed_refs;
2326         while (1) {
2327                 if (!locked_ref) {
2328                         if (count >= nr)
2329                                 break;
2330
2331                         spin_lock(&delayed_refs->lock);
2332                         locked_ref = btrfs_select_ref_head(trans);
2333                         if (!locked_ref) {
2334                                 spin_unlock(&delayed_refs->lock);
2335                                 break;
2336                         }
2337
2338                         /* grab the lock that says we are going to process
2339                          * all the refs for this head */
2340                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2341                         spin_unlock(&delayed_refs->lock);
2342                         /*
2343                          * we may have dropped the spin lock to get the head
2344                          * mutex lock, and that might have given someone else
2345                          * time to free the head.  If that's true, it has been
2346                          * removed from our list and we can move on.
2347                          */
2348                         if (ret == -EAGAIN) {
2349                                 locked_ref = NULL;
2350                                 count++;
2351                                 continue;
2352                         }
2353                 }
2354
2355                 spin_lock(&locked_ref->lock);
2356
2357                 /*
2358                  * locked_ref is the head node, so we have to go one
2359                  * node back for any delayed ref updates
2360                  */
2361                 ref = select_delayed_ref(locked_ref);
2362
2363                 if (ref && ref->seq &&
2364                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2365                         spin_unlock(&locked_ref->lock);
2366                         btrfs_delayed_ref_unlock(locked_ref);
2367                         spin_lock(&delayed_refs->lock);
2368                         locked_ref->processing = 0;
2369                         delayed_refs->num_heads_ready++;
2370                         spin_unlock(&delayed_refs->lock);
2371                         locked_ref = NULL;
2372                         cond_resched();
2373                         count++;
2374                         continue;
2375                 }
2376
2377                 /*
2378                  * record the must insert reserved flag before we
2379                  * drop the spin lock.
2380                  */
2381                 must_insert_reserved = locked_ref->must_insert_reserved;
2382                 locked_ref->must_insert_reserved = 0;
2383
2384                 extent_op = locked_ref->extent_op;
2385                 locked_ref->extent_op = NULL;
2386
2387                 if (!ref) {
2388
2389
2390                         /* All delayed refs have been processed, Go ahead
2391                          * and send the head node to run_one_delayed_ref,
2392                          * so that any accounting fixes can happen
2393                          */
2394                         ref = &locked_ref->node;
2395
2396                         if (extent_op && must_insert_reserved) {
2397                                 btrfs_free_delayed_extent_op(extent_op);
2398                                 extent_op = NULL;
2399                         }
2400
2401                         if (extent_op) {
2402                                 spin_unlock(&locked_ref->lock);
2403                                 ret = run_delayed_extent_op(trans, root,
2404                                                             ref, extent_op);
2405                                 btrfs_free_delayed_extent_op(extent_op);
2406
2407                                 if (ret) {
2408                                         /*
2409                                          * Need to reset must_insert_reserved if
2410                                          * there was an error so the abort stuff
2411                                          * can cleanup the reserved space
2412                                          * properly.
2413                                          */
2414                                         if (must_insert_reserved)
2415                                                 locked_ref->must_insert_reserved = 1;
2416                                         locked_ref->processing = 0;
2417                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2418                                         btrfs_delayed_ref_unlock(locked_ref);
2419                                         return ret;
2420                                 }
2421                                 continue;
2422                         }
2423
2424                         /*
2425                          * Need to drop our head ref lock and re-aqcuire the
2426                          * delayed ref lock and then re-check to make sure
2427                          * nobody got added.
2428                          */
2429                         spin_unlock(&locked_ref->lock);
2430                         spin_lock(&delayed_refs->lock);
2431                         spin_lock(&locked_ref->lock);
2432                         if (!list_empty(&locked_ref->ref_list) ||
2433                             locked_ref->extent_op) {
2434                                 spin_unlock(&locked_ref->lock);
2435                                 spin_unlock(&delayed_refs->lock);
2436                                 continue;
2437                         }
2438                         ref->in_tree = 0;
2439                         delayed_refs->num_heads--;
2440                         rb_erase(&locked_ref->href_node,
2441                                  &delayed_refs->href_root);
2442                         spin_unlock(&delayed_refs->lock);
2443                 } else {
2444                         actual_count++;
2445                         ref->in_tree = 0;
2446                         list_del(&ref->list);
2447                 }
2448                 atomic_dec(&delayed_refs->num_entries);
2449
2450                 if (!btrfs_delayed_ref_is_head(ref)) {
2451                         /*
2452                          * when we play the delayed ref, also correct the
2453                          * ref_mod on head
2454                          */
2455                         switch (ref->action) {
2456                         case BTRFS_ADD_DELAYED_REF:
2457                         case BTRFS_ADD_DELAYED_EXTENT:
2458                                 locked_ref->node.ref_mod -= ref->ref_mod;
2459                                 break;
2460                         case BTRFS_DROP_DELAYED_REF:
2461                                 locked_ref->node.ref_mod += ref->ref_mod;
2462                                 break;
2463                         default:
2464                                 WARN_ON(1);
2465                         }
2466                 }
2467                 spin_unlock(&locked_ref->lock);
2468
2469                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2470                                           must_insert_reserved);
2471
2472                 btrfs_free_delayed_extent_op(extent_op);
2473                 if (ret) {
2474                         locked_ref->processing = 0;
2475                         btrfs_delayed_ref_unlock(locked_ref);
2476                         btrfs_put_delayed_ref(ref);
2477                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2478                         return ret;
2479                 }
2480
2481                 /*
2482                  * If this node is a head, that means all the refs in this head
2483                  * have been dealt with, and we will pick the next head to deal
2484                  * with, so we must unlock the head and drop it from the cluster
2485                  * list before we release it.
2486                  */
2487                 if (btrfs_delayed_ref_is_head(ref)) {
2488                         if (locked_ref->is_data &&
2489                             locked_ref->total_ref_mod < 0) {
2490                                 spin_lock(&delayed_refs->lock);
2491                                 delayed_refs->pending_csums -= ref->num_bytes;
2492                                 spin_unlock(&delayed_refs->lock);
2493                         }
2494                         btrfs_delayed_ref_unlock(locked_ref);
2495                         locked_ref = NULL;
2496                 }
2497                 btrfs_put_delayed_ref(ref);
2498                 count++;
2499                 cond_resched();
2500         }
2501
2502         /*
2503          * We don't want to include ref heads since we can have empty ref heads
2504          * and those will drastically skew our runtime down since we just do
2505          * accounting, no actual extent tree updates.
2506          */
2507         if (actual_count > 0) {
2508                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2509                 u64 avg;
2510
2511                 /*
2512                  * We weigh the current average higher than our current runtime
2513                  * to avoid large swings in the average.
2514                  */
2515                 spin_lock(&delayed_refs->lock);
2516                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2517                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2518                 spin_unlock(&delayed_refs->lock);
2519         }
2520         return 0;
2521 }
2522
2523 #ifdef SCRAMBLE_DELAYED_REFS
2524 /*
2525  * Normally delayed refs get processed in ascending bytenr order. This
2526  * correlates in most cases to the order added. To expose dependencies on this
2527  * order, we start to process the tree in the middle instead of the beginning
2528  */
2529 static u64 find_middle(struct rb_root *root)
2530 {
2531         struct rb_node *n = root->rb_node;
2532         struct btrfs_delayed_ref_node *entry;
2533         int alt = 1;
2534         u64 middle;
2535         u64 first = 0, last = 0;
2536
2537         n = rb_first(root);
2538         if (n) {
2539                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2540                 first = entry->bytenr;
2541         }
2542         n = rb_last(root);
2543         if (n) {
2544                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2545                 last = entry->bytenr;
2546         }
2547         n = root->rb_node;
2548
2549         while (n) {
2550                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2551                 WARN_ON(!entry->in_tree);
2552
2553                 middle = entry->bytenr;
2554
2555                 if (alt)
2556                         n = n->rb_left;
2557                 else
2558                         n = n->rb_right;
2559
2560                 alt = 1 - alt;
2561         }
2562         return middle;
2563 }
2564 #endif
2565
2566 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2567 {
2568         u64 num_bytes;
2569
2570         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2571                              sizeof(struct btrfs_extent_inline_ref));
2572         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2573                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2574
2575         /*
2576          * We don't ever fill up leaves all the way so multiply by 2 just to be
2577          * closer to what we're really going to want to ouse.
2578          */
2579         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2580 }
2581
2582 /*
2583  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2584  * would require to store the csums for that many bytes.
2585  */
2586 u64 btrfs_csum_bytes_to_leaves(struct btrfs_root *root, u64 csum_bytes)
2587 {
2588         u64 csum_size;
2589         u64 num_csums_per_leaf;
2590         u64 num_csums;
2591
2592         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
2593         num_csums_per_leaf = div64_u64(csum_size,
2594                         (u64)btrfs_super_csum_size(root->fs_info->super_copy));
2595         num_csums = div64_u64(csum_bytes, root->sectorsize);
2596         num_csums += num_csums_per_leaf - 1;
2597         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2598         return num_csums;
2599 }
2600
2601 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2602                                        struct btrfs_root *root)
2603 {
2604         struct btrfs_block_rsv *global_rsv;
2605         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2606         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2607         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2608         u64 num_bytes, num_dirty_bgs_bytes;
2609         int ret = 0;
2610
2611         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2612         num_heads = heads_to_leaves(root, num_heads);
2613         if (num_heads > 1)
2614                 num_bytes += (num_heads - 1) * root->nodesize;
2615         num_bytes <<= 1;
2616         num_bytes += btrfs_csum_bytes_to_leaves(root, csum_bytes) * root->nodesize;
2617         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(root,
2618                                                              num_dirty_bgs);
2619         global_rsv = &root->fs_info->global_block_rsv;
2620
2621         /*
2622          * If we can't allocate any more chunks lets make sure we have _lots_ of
2623          * wiggle room since running delayed refs can create more delayed refs.
2624          */
2625         if (global_rsv->space_info->full) {
2626                 num_dirty_bgs_bytes <<= 1;
2627                 num_bytes <<= 1;
2628         }
2629
2630         spin_lock(&global_rsv->lock);
2631         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2632                 ret = 1;
2633         spin_unlock(&global_rsv->lock);
2634         return ret;
2635 }
2636
2637 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2638                                        struct btrfs_root *root)
2639 {
2640         struct btrfs_fs_info *fs_info = root->fs_info;
2641         u64 num_entries =
2642                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2643         u64 avg_runtime;
2644         u64 val;
2645
2646         smp_mb();
2647         avg_runtime = fs_info->avg_delayed_ref_runtime;
2648         val = num_entries * avg_runtime;
2649         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2650                 return 1;
2651         if (val >= NSEC_PER_SEC / 2)
2652                 return 2;
2653
2654         return btrfs_check_space_for_delayed_refs(trans, root);
2655 }
2656
2657 struct async_delayed_refs {
2658         struct btrfs_root *root;
2659         int count;
2660         int error;
2661         int sync;
2662         struct completion wait;
2663         struct btrfs_work work;
2664 };
2665
2666 static void delayed_ref_async_start(struct btrfs_work *work)
2667 {
2668         struct async_delayed_refs *async;
2669         struct btrfs_trans_handle *trans;
2670         int ret;
2671
2672         async = container_of(work, struct async_delayed_refs, work);
2673
2674         trans = btrfs_join_transaction(async->root);
2675         if (IS_ERR(trans)) {
2676                 async->error = PTR_ERR(trans);
2677                 goto done;
2678         }
2679
2680         /*
2681          * trans->sync means that when we call end_transaciton, we won't
2682          * wait on delayed refs
2683          */
2684         trans->sync = true;
2685         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2686         if (ret)
2687                 async->error = ret;
2688
2689         ret = btrfs_end_transaction(trans, async->root);
2690         if (ret && !async->error)
2691                 async->error = ret;
2692 done:
2693         if (async->sync)
2694                 complete(&async->wait);
2695         else
2696                 kfree(async);
2697 }
2698
2699 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2700                                  unsigned long count, int wait)
2701 {
2702         struct async_delayed_refs *async;
2703         int ret;
2704
2705         async = kmalloc(sizeof(*async), GFP_NOFS);
2706         if (!async)
2707                 return -ENOMEM;
2708
2709         async->root = root->fs_info->tree_root;
2710         async->count = count;
2711         async->error = 0;
2712         if (wait)
2713                 async->sync = 1;
2714         else
2715                 async->sync = 0;
2716         init_completion(&async->wait);
2717
2718         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2719                         delayed_ref_async_start, NULL, NULL);
2720
2721         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2722
2723         if (wait) {
2724                 wait_for_completion(&async->wait);
2725                 ret = async->error;
2726                 kfree(async);
2727                 return ret;
2728         }
2729         return 0;
2730 }
2731
2732 /*
2733  * this starts processing the delayed reference count updates and
2734  * extent insertions we have queued up so far.  count can be
2735  * 0, which means to process everything in the tree at the start
2736  * of the run (but not newly added entries), or it can be some target
2737  * number you'd like to process.
2738  *
2739  * Returns 0 on success or if called with an aborted transaction
2740  * Returns <0 on error and aborts the transaction
2741  */
2742 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2743                            struct btrfs_root *root, unsigned long count)
2744 {
2745         struct rb_node *node;
2746         struct btrfs_delayed_ref_root *delayed_refs;
2747         struct btrfs_delayed_ref_head *head;
2748         int ret;
2749         int run_all = count == (unsigned long)-1;
2750
2751         /* We'll clean this up in btrfs_cleanup_transaction */
2752         if (trans->aborted)
2753                 return 0;
2754
2755         if (root == root->fs_info->extent_root)
2756                 root = root->fs_info->tree_root;
2757
2758         delayed_refs = &trans->transaction->delayed_refs;
2759         if (count == 0)
2760                 count = atomic_read(&delayed_refs->num_entries) * 2;
2761
2762 again:
2763 #ifdef SCRAMBLE_DELAYED_REFS
2764         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2765 #endif
2766         ret = __btrfs_run_delayed_refs(trans, root, count);
2767         if (ret < 0) {
2768                 btrfs_abort_transaction(trans, root, ret);
2769                 return ret;
2770         }
2771
2772         if (run_all) {
2773                 if (!list_empty(&trans->new_bgs))
2774                         btrfs_create_pending_block_groups(trans, root);
2775
2776                 spin_lock(&delayed_refs->lock);
2777                 node = rb_first(&delayed_refs->href_root);
2778                 if (!node) {
2779                         spin_unlock(&delayed_refs->lock);
2780                         goto out;
2781                 }
2782                 count = (unsigned long)-1;
2783
2784                 while (node) {
2785                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2786                                         href_node);
2787                         if (btrfs_delayed_ref_is_head(&head->node)) {
2788                                 struct btrfs_delayed_ref_node *ref;
2789
2790                                 ref = &head->node;
2791                                 atomic_inc(&ref->refs);
2792
2793                                 spin_unlock(&delayed_refs->lock);
2794                                 /*
2795                                  * Mutex was contended, block until it's
2796                                  * released and try again
2797                                  */
2798                                 mutex_lock(&head->mutex);
2799                                 mutex_unlock(&head->mutex);
2800
2801                                 btrfs_put_delayed_ref(ref);
2802                                 cond_resched();
2803                                 goto again;
2804                         } else {
2805                                 WARN_ON(1);
2806                         }
2807                         node = rb_next(node);
2808                 }
2809                 spin_unlock(&delayed_refs->lock);
2810                 cond_resched();
2811                 goto again;
2812         }
2813 out:
2814         assert_qgroups_uptodate(trans);
2815         return 0;
2816 }
2817
2818 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2819                                 struct btrfs_root *root,
2820                                 u64 bytenr, u64 num_bytes, u64 flags,
2821                                 int level, int is_data)
2822 {
2823         struct btrfs_delayed_extent_op *extent_op;
2824         int ret;
2825
2826         extent_op = btrfs_alloc_delayed_extent_op();
2827         if (!extent_op)
2828                 return -ENOMEM;
2829
2830         extent_op->flags_to_set = flags;
2831         extent_op->update_flags = 1;
2832         extent_op->update_key = 0;
2833         extent_op->is_data = is_data ? 1 : 0;
2834         extent_op->level = level;
2835
2836         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2837                                           num_bytes, extent_op);
2838         if (ret)
2839                 btrfs_free_delayed_extent_op(extent_op);
2840         return ret;
2841 }
2842
2843 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2844                                       struct btrfs_root *root,
2845                                       struct btrfs_path *path,
2846                                       u64 objectid, u64 offset, u64 bytenr)
2847 {
2848         struct btrfs_delayed_ref_head *head;
2849         struct btrfs_delayed_ref_node *ref;
2850         struct btrfs_delayed_data_ref *data_ref;
2851         struct btrfs_delayed_ref_root *delayed_refs;
2852         int ret = 0;
2853
2854         delayed_refs = &trans->transaction->delayed_refs;
2855         spin_lock(&delayed_refs->lock);
2856         head = btrfs_find_delayed_ref_head(trans, bytenr);
2857         if (!head) {
2858                 spin_unlock(&delayed_refs->lock);
2859                 return 0;
2860         }
2861
2862         if (!mutex_trylock(&head->mutex)) {
2863                 atomic_inc(&head->node.refs);
2864                 spin_unlock(&delayed_refs->lock);
2865
2866                 btrfs_release_path(path);
2867
2868                 /*
2869                  * Mutex was contended, block until it's released and let
2870                  * caller try again
2871                  */
2872                 mutex_lock(&head->mutex);
2873                 mutex_unlock(&head->mutex);
2874                 btrfs_put_delayed_ref(&head->node);
2875                 return -EAGAIN;
2876         }
2877         spin_unlock(&delayed_refs->lock);
2878
2879         spin_lock(&head->lock);
2880         list_for_each_entry(ref, &head->ref_list, list) {
2881                 /* If it's a shared ref we know a cross reference exists */
2882                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2883                         ret = 1;
2884                         break;
2885                 }
2886
2887                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2888
2889                 /*
2890                  * If our ref doesn't match the one we're currently looking at
2891                  * then we have a cross reference.
2892                  */
2893                 if (data_ref->root != root->root_key.objectid ||
2894                     data_ref->objectid != objectid ||
2895                     data_ref->offset != offset) {
2896                         ret = 1;
2897                         break;
2898                 }
2899         }
2900         spin_unlock(&head->lock);
2901         mutex_unlock(&head->mutex);
2902         return ret;
2903 }
2904
2905 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2906                                         struct btrfs_root *root,
2907                                         struct btrfs_path *path,
2908                                         u64 objectid, u64 offset, u64 bytenr)
2909 {
2910         struct btrfs_root *extent_root = root->fs_info->extent_root;
2911         struct extent_buffer *leaf;
2912         struct btrfs_extent_data_ref *ref;
2913         struct btrfs_extent_inline_ref *iref;
2914         struct btrfs_extent_item *ei;
2915         struct btrfs_key key;
2916         u32 item_size;
2917         int ret;
2918
2919         key.objectid = bytenr;
2920         key.offset = (u64)-1;
2921         key.type = BTRFS_EXTENT_ITEM_KEY;
2922
2923         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2924         if (ret < 0)
2925                 goto out;
2926         BUG_ON(ret == 0); /* Corruption */
2927
2928         ret = -ENOENT;
2929         if (path->slots[0] == 0)
2930                 goto out;
2931
2932         path->slots[0]--;
2933         leaf = path->nodes[0];
2934         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2935
2936         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2937                 goto out;
2938
2939         ret = 1;
2940         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2941 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2942         if (item_size < sizeof(*ei)) {
2943                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2944                 goto out;
2945         }
2946 #endif
2947         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2948
2949         if (item_size != sizeof(*ei) +
2950             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2951                 goto out;
2952
2953         if (btrfs_extent_generation(leaf, ei) <=
2954             btrfs_root_last_snapshot(&root->root_item))
2955                 goto out;
2956
2957         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
2958         if (btrfs_extent_inline_ref_type(leaf, iref) !=
2959             BTRFS_EXTENT_DATA_REF_KEY)
2960                 goto out;
2961
2962         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
2963         if (btrfs_extent_refs(leaf, ei) !=
2964             btrfs_extent_data_ref_count(leaf, ref) ||
2965             btrfs_extent_data_ref_root(leaf, ref) !=
2966             root->root_key.objectid ||
2967             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
2968             btrfs_extent_data_ref_offset(leaf, ref) != offset)
2969                 goto out;
2970
2971         ret = 0;
2972 out:
2973         return ret;
2974 }
2975
2976 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
2977                           struct btrfs_root *root,
2978                           u64 objectid, u64 offset, u64 bytenr)
2979 {
2980         struct btrfs_path *path;
2981         int ret;
2982         int ret2;
2983
2984         path = btrfs_alloc_path();
2985         if (!path)
2986                 return -ENOENT;
2987
2988         do {
2989                 ret = check_committed_ref(trans, root, path, objectid,
2990                                           offset, bytenr);
2991                 if (ret && ret != -ENOENT)
2992                         goto out;
2993
2994                 ret2 = check_delayed_ref(trans, root, path, objectid,
2995                                          offset, bytenr);
2996         } while (ret2 == -EAGAIN);
2997
2998         if (ret2 && ret2 != -ENOENT) {
2999                 ret = ret2;
3000                 goto out;
3001         }
3002
3003         if (ret != -ENOENT || ret2 != -ENOENT)
3004                 ret = 0;
3005 out:
3006         btrfs_free_path(path);
3007         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3008                 WARN_ON(ret > 0);
3009         return ret;
3010 }
3011
3012 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3013                            struct btrfs_root *root,
3014                            struct extent_buffer *buf,
3015                            int full_backref, int inc)
3016 {
3017         u64 bytenr;
3018         u64 num_bytes;
3019         u64 parent;
3020         u64 ref_root;
3021         u32 nritems;
3022         struct btrfs_key key;
3023         struct btrfs_file_extent_item *fi;
3024         int i;
3025         int level;
3026         int ret = 0;
3027         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3028                             u64, u64, u64, u64, u64, u64, int);
3029
3030
3031         if (btrfs_test_is_dummy_root(root))
3032                 return 0;
3033
3034         ref_root = btrfs_header_owner(buf);
3035         nritems = btrfs_header_nritems(buf);
3036         level = btrfs_header_level(buf);
3037
3038         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3039                 return 0;
3040
3041         if (inc)
3042                 process_func = btrfs_inc_extent_ref;
3043         else
3044                 process_func = btrfs_free_extent;
3045
3046         if (full_backref)
3047                 parent = buf->start;
3048         else
3049                 parent = 0;
3050
3051         for (i = 0; i < nritems; i++) {
3052                 if (level == 0) {
3053                         btrfs_item_key_to_cpu(buf, &key, i);
3054                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3055                                 continue;
3056                         fi = btrfs_item_ptr(buf, i,
3057                                             struct btrfs_file_extent_item);
3058                         if (btrfs_file_extent_type(buf, fi) ==
3059                             BTRFS_FILE_EXTENT_INLINE)
3060                                 continue;
3061                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3062                         if (bytenr == 0)
3063                                 continue;
3064
3065                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3066                         key.offset -= btrfs_file_extent_offset(buf, fi);
3067                         ret = process_func(trans, root, bytenr, num_bytes,
3068                                            parent, ref_root, key.objectid,
3069                                            key.offset, 1);
3070                         if (ret)
3071                                 goto fail;
3072                 } else {
3073                         bytenr = btrfs_node_blockptr(buf, i);
3074                         num_bytes = root->nodesize;
3075                         ret = process_func(trans, root, bytenr, num_bytes,
3076                                            parent, ref_root, level - 1, 0,
3077                                            1);
3078                         if (ret)
3079                                 goto fail;
3080                 }
3081         }
3082         return 0;
3083 fail:
3084         return ret;
3085 }
3086
3087 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3088                   struct extent_buffer *buf, int full_backref)
3089 {
3090         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3091 }
3092
3093 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3094                   struct extent_buffer *buf, int full_backref)
3095 {
3096         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3097 }
3098
3099 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3100                                  struct btrfs_root *root,
3101                                  struct btrfs_path *path,
3102                                  struct btrfs_block_group_cache *cache)
3103 {
3104         int ret;
3105         struct btrfs_root *extent_root = root->fs_info->extent_root;
3106         unsigned long bi;
3107         struct extent_buffer *leaf;
3108
3109         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3110         if (ret) {
3111                 if (ret > 0)
3112                         ret = -ENOENT;
3113                 goto fail;
3114         }
3115
3116         leaf = path->nodes[0];
3117         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3118         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3119         btrfs_mark_buffer_dirty(leaf);
3120 fail:
3121         btrfs_release_path(path);
3122         return ret;
3123
3124 }
3125
3126 static struct btrfs_block_group_cache *
3127 next_block_group(struct btrfs_root *root,
3128                  struct btrfs_block_group_cache *cache)
3129 {
3130         struct rb_node *node;
3131
3132         spin_lock(&root->fs_info->block_group_cache_lock);
3133
3134         /* If our block group was removed, we need a full search. */
3135         if (RB_EMPTY_NODE(&cache->cache_node)) {
3136                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3137
3138                 spin_unlock(&root->fs_info->block_group_cache_lock);
3139                 btrfs_put_block_group(cache);
3140                 cache = btrfs_lookup_first_block_group(root->fs_info,
3141                                                        next_bytenr);
3142                 return cache;
3143         }
3144         node = rb_next(&cache->cache_node);
3145         btrfs_put_block_group(cache);
3146         if (node) {
3147                 cache = rb_entry(node, struct btrfs_block_group_cache,
3148                                  cache_node);
3149                 btrfs_get_block_group(cache);
3150         } else
3151                 cache = NULL;
3152         spin_unlock(&root->fs_info->block_group_cache_lock);
3153         return cache;
3154 }
3155
3156 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3157                             struct btrfs_trans_handle *trans,
3158                             struct btrfs_path *path)
3159 {
3160         struct btrfs_root *root = block_group->fs_info->tree_root;
3161         struct inode *inode = NULL;
3162         u64 alloc_hint = 0;
3163         int dcs = BTRFS_DC_ERROR;
3164         u64 num_pages = 0;
3165         int retries = 0;
3166         int ret = 0;
3167
3168         /*
3169          * If this block group is smaller than 100 megs don't bother caching the
3170          * block group.
3171          */
3172         if (block_group->key.offset < (100 * 1024 * 1024)) {
3173                 spin_lock(&block_group->lock);
3174                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3175                 spin_unlock(&block_group->lock);
3176                 return 0;
3177         }
3178
3179         if (trans->aborted)
3180                 return 0;
3181 again:
3182         inode = lookup_free_space_inode(root, block_group, path);
3183         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3184                 ret = PTR_ERR(inode);
3185                 btrfs_release_path(path);
3186                 goto out;
3187         }
3188
3189         if (IS_ERR(inode)) {
3190                 BUG_ON(retries);
3191                 retries++;
3192
3193                 if (block_group->ro)
3194                         goto out_free;
3195
3196                 ret = create_free_space_inode(root, trans, block_group, path);
3197                 if (ret)
3198                         goto out_free;
3199                 goto again;
3200         }
3201
3202         /* We've already setup this transaction, go ahead and exit */
3203         if (block_group->cache_generation == trans->transid &&
3204             i_size_read(inode)) {
3205                 dcs = BTRFS_DC_SETUP;
3206                 goto out_put;
3207         }
3208
3209         /*
3210          * We want to set the generation to 0, that way if anything goes wrong
3211          * from here on out we know not to trust this cache when we load up next
3212          * time.
3213          */
3214         BTRFS_I(inode)->generation = 0;
3215         ret = btrfs_update_inode(trans, root, inode);
3216         if (ret) {
3217                 /*
3218                  * So theoretically we could recover from this, simply set the
3219                  * super cache generation to 0 so we know to invalidate the
3220                  * cache, but then we'd have to keep track of the block groups
3221                  * that fail this way so we know we _have_ to reset this cache
3222                  * before the next commit or risk reading stale cache.  So to
3223                  * limit our exposure to horrible edge cases lets just abort the
3224                  * transaction, this only happens in really bad situations
3225                  * anyway.
3226                  */
3227                 btrfs_abort_transaction(trans, root, ret);
3228                 goto out_put;
3229         }
3230         WARN_ON(ret);
3231
3232         if (i_size_read(inode) > 0) {
3233                 ret = btrfs_check_trunc_cache_free_space(root,
3234                                         &root->fs_info->global_block_rsv);
3235                 if (ret)
3236                         goto out_put;
3237
3238                 ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
3239                 if (ret)
3240                         goto out_put;
3241         }
3242
3243         spin_lock(&block_group->lock);
3244         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3245             !btrfs_test_opt(root, SPACE_CACHE)) {
3246                 /*
3247                  * don't bother trying to write stuff out _if_
3248                  * a) we're not cached,
3249                  * b) we're with nospace_cache mount option.
3250                  */
3251                 dcs = BTRFS_DC_WRITTEN;
3252                 spin_unlock(&block_group->lock);
3253                 goto out_put;
3254         }
3255         spin_unlock(&block_group->lock);
3256
3257         /*
3258          * Try to preallocate enough space based on how big the block group is.
3259          * Keep in mind this has to include any pinned space which could end up
3260          * taking up quite a bit since it's not folded into the other space
3261          * cache.
3262          */
3263         num_pages = div_u64(block_group->key.offset, 256 * 1024 * 1024);
3264         if (!num_pages)
3265                 num_pages = 1;
3266
3267         num_pages *= 16;
3268         num_pages *= PAGE_CACHE_SIZE;
3269
3270         ret = btrfs_check_data_free_space(inode, num_pages, num_pages);
3271         if (ret)
3272                 goto out_put;
3273
3274         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3275                                               num_pages, num_pages,
3276                                               &alloc_hint);
3277         if (!ret)
3278                 dcs = BTRFS_DC_SETUP;
3279         btrfs_free_reserved_data_space(inode, num_pages);
3280
3281 out_put:
3282         iput(inode);
3283 out_free:
3284         btrfs_release_path(path);
3285 out:
3286         spin_lock(&block_group->lock);
3287         if (!ret && dcs == BTRFS_DC_SETUP)
3288                 block_group->cache_generation = trans->transid;
3289         block_group->disk_cache_state = dcs;
3290         spin_unlock(&block_group->lock);
3291
3292         return ret;
3293 }
3294
3295 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3296                             struct btrfs_root *root)
3297 {
3298         struct btrfs_block_group_cache *cache, *tmp;
3299         struct btrfs_transaction *cur_trans = trans->transaction;
3300         struct btrfs_path *path;
3301
3302         if (list_empty(&cur_trans->dirty_bgs) ||
3303             !btrfs_test_opt(root, SPACE_CACHE))
3304                 return 0;
3305
3306         path = btrfs_alloc_path();
3307         if (!path)
3308                 return -ENOMEM;
3309
3310         /* Could add new block groups, use _safe just in case */
3311         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3312                                  dirty_list) {
3313                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3314                         cache_save_setup(cache, trans, path);
3315         }
3316
3317         btrfs_free_path(path);
3318         return 0;
3319 }
3320
3321 /*
3322  * transaction commit does final block group cache writeback during a
3323  * critical section where nothing is allowed to change the FS.  This is
3324  * required in order for the cache to actually match the block group,
3325  * but can introduce a lot of latency into the commit.
3326  *
3327  * So, btrfs_start_dirty_block_groups is here to kick off block group
3328  * cache IO.  There's a chance we'll have to redo some of it if the
3329  * block group changes again during the commit, but it greatly reduces
3330  * the commit latency by getting rid of the easy block groups while
3331  * we're still allowing others to join the commit.
3332  */
3333 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3334                                    struct btrfs_root *root)
3335 {
3336         struct btrfs_block_group_cache *cache;
3337         struct btrfs_transaction *cur_trans = trans->transaction;
3338         int ret = 0;
3339         int should_put;
3340         struct btrfs_path *path = NULL;
3341         LIST_HEAD(dirty);
3342         struct list_head *io = &cur_trans->io_bgs;
3343         int num_started = 0;
3344         int loops = 0;
3345
3346         spin_lock(&cur_trans->dirty_bgs_lock);
3347         if (list_empty(&cur_trans->dirty_bgs)) {
3348                 spin_unlock(&cur_trans->dirty_bgs_lock);
3349                 return 0;
3350         }
3351         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3352         spin_unlock(&cur_trans->dirty_bgs_lock);
3353
3354 again:
3355         /*
3356          * make sure all the block groups on our dirty list actually
3357          * exist
3358          */
3359         btrfs_create_pending_block_groups(trans, root);
3360
3361         if (!path) {
3362                 path = btrfs_alloc_path();
3363                 if (!path)
3364                         return -ENOMEM;
3365         }
3366
3367         /*
3368          * cache_write_mutex is here only to save us from balance or automatic
3369          * removal of empty block groups deleting this block group while we are
3370          * writing out the cache
3371          */
3372         mutex_lock(&trans->transaction->cache_write_mutex);
3373         while (!list_empty(&dirty)) {
3374                 cache = list_first_entry(&dirty,
3375                                          struct btrfs_block_group_cache,
3376                                          dirty_list);
3377                 /*
3378                  * this can happen if something re-dirties a block
3379                  * group that is already under IO.  Just wait for it to
3380                  * finish and then do it all again
3381                  */
3382                 if (!list_empty(&cache->io_list)) {
3383                         list_del_init(&cache->io_list);
3384                         btrfs_wait_cache_io(root, trans, cache,
3385                                             &cache->io_ctl, path,
3386                                             cache->key.objectid);
3387                         btrfs_put_block_group(cache);
3388                 }
3389
3390
3391                 /*
3392                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3393                  * if it should update the cache_state.  Don't delete
3394                  * until after we wait.
3395                  *
3396                  * Since we're not running in the commit critical section
3397                  * we need the dirty_bgs_lock to protect from update_block_group
3398                  */
3399                 spin_lock(&cur_trans->dirty_bgs_lock);
3400                 list_del_init(&cache->dirty_list);
3401                 spin_unlock(&cur_trans->dirty_bgs_lock);
3402
3403                 should_put = 1;
3404
3405                 cache_save_setup(cache, trans, path);
3406
3407                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3408                         cache->io_ctl.inode = NULL;
3409                         ret = btrfs_write_out_cache(root, trans, cache, path);
3410                         if (ret == 0 && cache->io_ctl.inode) {
3411                                 num_started++;
3412                                 should_put = 0;
3413
3414                                 /*
3415                                  * the cache_write_mutex is protecting
3416                                  * the io_list
3417                                  */
3418                                 list_add_tail(&cache->io_list, io);
3419                         } else {
3420                                 /*
3421                                  * if we failed to write the cache, the
3422                                  * generation will be bad and life goes on
3423                                  */
3424                                 ret = 0;
3425                         }
3426                 }
3427                 if (!ret) {
3428                         ret = write_one_cache_group(trans, root, path, cache);
3429                         /*
3430                          * Our block group might still be attached to the list
3431                          * of new block groups in the transaction handle of some
3432                          * other task (struct btrfs_trans_handle->new_bgs). This
3433                          * means its block group item isn't yet in the extent
3434                          * tree. If this happens ignore the error, as we will
3435                          * try again later in the critical section of the
3436                          * transaction commit.
3437                          */
3438                         if (ret == -ENOENT) {
3439                                 ret = 0;
3440                                 spin_lock(&cur_trans->dirty_bgs_lock);
3441                                 if (list_empty(&cache->dirty_list)) {
3442                                         list_add_tail(&cache->dirty_list,
3443                                                       &cur_trans->dirty_bgs);
3444                                         btrfs_get_block_group(cache);
3445                                 }
3446                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3447                         } else if (ret) {
3448                                 btrfs_abort_transaction(trans, root, ret);
3449                         }
3450                 }
3451
3452                 /* if its not on the io list, we need to put the block group */
3453                 if (should_put)
3454                         btrfs_put_block_group(cache);
3455
3456                 if (ret)
3457                         break;
3458
3459                 /*
3460                  * Avoid blocking other tasks for too long. It might even save
3461                  * us from writing caches for block groups that are going to be
3462                  * removed.
3463                  */
3464                 mutex_unlock(&trans->transaction->cache_write_mutex);
3465                 mutex_lock(&trans->transaction->cache_write_mutex);
3466         }
3467         mutex_unlock(&trans->transaction->cache_write_mutex);
3468
3469         /*
3470          * go through delayed refs for all the stuff we've just kicked off
3471          * and then loop back (just once)
3472          */
3473         ret = btrfs_run_delayed_refs(trans, root, 0);
3474         if (!ret && loops == 0) {
3475                 loops++;
3476                 spin_lock(&cur_trans->dirty_bgs_lock);
3477                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3478                 /*
3479                  * dirty_bgs_lock protects us from concurrent block group
3480                  * deletes too (not just cache_write_mutex).
3481                  */
3482                 if (!list_empty(&dirty)) {
3483                         spin_unlock(&cur_trans->dirty_bgs_lock);
3484                         goto again;
3485                 }
3486                 spin_unlock(&cur_trans->dirty_bgs_lock);
3487         }
3488
3489         btrfs_free_path(path);
3490         return ret;
3491 }
3492
3493 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3494                                    struct btrfs_root *root)
3495 {
3496         struct btrfs_block_group_cache *cache;
3497         struct btrfs_transaction *cur_trans = trans->transaction;
3498         int ret = 0;
3499         int should_put;
3500         struct btrfs_path *path;
3501         struct list_head *io = &cur_trans->io_bgs;
3502         int num_started = 0;
3503
3504         path = btrfs_alloc_path();
3505         if (!path)
3506                 return -ENOMEM;
3507
3508         /*
3509          * We don't need the lock here since we are protected by the transaction
3510          * commit.  We want to do the cache_save_setup first and then run the
3511          * delayed refs to make sure we have the best chance at doing this all
3512          * in one shot.
3513          */
3514         while (!list_empty(&cur_trans->dirty_bgs)) {
3515                 cache = list_first_entry(&cur_trans->dirty_bgs,
3516                                          struct btrfs_block_group_cache,
3517                                          dirty_list);
3518
3519                 /*
3520                  * this can happen if cache_save_setup re-dirties a block
3521                  * group that is already under IO.  Just wait for it to
3522                  * finish and then do it all again
3523                  */
3524                 if (!list_empty(&cache->io_list)) {
3525                         list_del_init(&cache->io_list);
3526                         btrfs_wait_cache_io(root, trans, cache,
3527                                             &cache->io_ctl, path,
3528                                             cache->key.objectid);
3529                         btrfs_put_block_group(cache);
3530                 }
3531
3532                 /*
3533                  * don't remove from the dirty list until after we've waited
3534                  * on any pending IO
3535                  */
3536                 list_del_init(&cache->dirty_list);
3537                 should_put = 1;
3538
3539                 cache_save_setup(cache, trans, path);
3540
3541                 if (!ret)
3542                         ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
3543
3544                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3545                         cache->io_ctl.inode = NULL;
3546                         ret = btrfs_write_out_cache(root, trans, cache, path);
3547                         if (ret == 0 && cache->io_ctl.inode) {
3548                                 num_started++;
3549                                 should_put = 0;
3550                                 list_add_tail(&cache->io_list, io);
3551                         } else {
3552                                 /*
3553                                  * if we failed to write the cache, the
3554                                  * generation will be bad and life goes on
3555                                  */
3556                                 ret = 0;
3557                         }
3558                 }
3559                 if (!ret) {
3560                         ret = write_one_cache_group(trans, root, path, cache);
3561                         if (ret)
3562                                 btrfs_abort_transaction(trans, root, ret);
3563                 }
3564
3565                 /* if its not on the io list, we need to put the block group */
3566                 if (should_put)
3567                         btrfs_put_block_group(cache);
3568         }
3569
3570         while (!list_empty(io)) {
3571                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3572                                          io_list);
3573                 list_del_init(&cache->io_list);
3574                 btrfs_wait_cache_io(root, trans, cache,
3575                                     &cache->io_ctl, path, cache->key.objectid);
3576                 btrfs_put_block_group(cache);
3577         }
3578
3579         btrfs_free_path(path);
3580         return ret;
3581 }
3582
3583 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3584 {
3585         struct btrfs_block_group_cache *block_group;
3586         int readonly = 0;
3587
3588         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3589         if (!block_group || block_group->ro)
3590                 readonly = 1;
3591         if (block_group)
3592                 btrfs_put_block_group(block_group);
3593         return readonly;
3594 }
3595
3596 static const char *alloc_name(u64 flags)
3597 {
3598         switch (flags) {
3599         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3600                 return "mixed";
3601         case BTRFS_BLOCK_GROUP_METADATA:
3602                 return "metadata";
3603         case BTRFS_BLOCK_GROUP_DATA:
3604                 return "data";
3605         case BTRFS_BLOCK_GROUP_SYSTEM:
3606                 return "system";
3607         default:
3608                 WARN_ON(1);
3609                 return "invalid-combination";
3610         };
3611 }
3612
3613 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3614                              u64 total_bytes, u64 bytes_used,
3615                              struct btrfs_space_info **space_info)
3616 {
3617         struct btrfs_space_info *found;
3618         int i;
3619         int factor;
3620         int ret;
3621
3622         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3623                      BTRFS_BLOCK_GROUP_RAID10))
3624                 factor = 2;
3625         else
3626                 factor = 1;
3627
3628         found = __find_space_info(info, flags);
3629         if (found) {
3630                 spin_lock(&found->lock);
3631                 found->total_bytes += total_bytes;
3632                 found->disk_total += total_bytes * factor;
3633                 found->bytes_used += bytes_used;
3634                 found->disk_used += bytes_used * factor;
3635                 if (total_bytes > 0)
3636                         found->full = 0;
3637                 spin_unlock(&found->lock);
3638                 *space_info = found;
3639                 return 0;
3640         }
3641         found = kzalloc(sizeof(*found), GFP_NOFS);
3642         if (!found)
3643                 return -ENOMEM;
3644
3645         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3646         if (ret) {
3647                 kfree(found);
3648                 return ret;
3649         }
3650
3651         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3652                 INIT_LIST_HEAD(&found->block_groups[i]);
3653         init_rwsem(&found->groups_sem);
3654         spin_lock_init(&found->lock);
3655         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3656         found->total_bytes = total_bytes;
3657         found->disk_total = total_bytes * factor;
3658         found->bytes_used = bytes_used;
3659         found->disk_used = bytes_used * factor;
3660         found->bytes_pinned = 0;
3661         found->bytes_reserved = 0;
3662         found->bytes_readonly = 0;
3663         found->bytes_may_use = 0;
3664         if (total_bytes > 0)
3665                 found->full = 0;
3666         else
3667                 found->full = 1;
3668         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3669         found->chunk_alloc = 0;
3670         found->flush = 0;
3671         init_waitqueue_head(&found->wait);
3672         INIT_LIST_HEAD(&found->ro_bgs);
3673
3674         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3675                                     info->space_info_kobj, "%s",
3676                                     alloc_name(found->flags));
3677         if (ret) {
3678                 kfree(found);
3679                 return ret;
3680         }
3681
3682         *space_info = found;
3683         list_add_rcu(&found->list, &info->space_info);
3684         if (flags & BTRFS_BLOCK_GROUP_DATA)
3685                 info->data_sinfo = found;
3686
3687         return ret;
3688 }
3689
3690 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3691 {
3692         u64 extra_flags = chunk_to_extended(flags) &
3693                                 BTRFS_EXTENDED_PROFILE_MASK;
3694
3695         write_seqlock(&fs_info->profiles_lock);
3696         if (flags & BTRFS_BLOCK_GROUP_DATA)
3697                 fs_info->avail_data_alloc_bits |= extra_flags;
3698         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3699                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3700         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3701                 fs_info->avail_system_alloc_bits |= extra_flags;
3702         write_sequnlock(&fs_info->profiles_lock);
3703 }
3704
3705 /*
3706  * returns target flags in extended format or 0 if restripe for this
3707  * chunk_type is not in progress
3708  *
3709  * should be called with either volume_mutex or balance_lock held
3710  */
3711 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3712 {
3713         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3714         u64 target = 0;
3715
3716         if (!bctl)
3717                 return 0;
3718
3719         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3720             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3721                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3722         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3723                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3724                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3725         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3726                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3727                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3728         }
3729
3730         return target;
3731 }
3732
3733 /*
3734  * @flags: available profiles in extended format (see ctree.h)
3735  *
3736  * Returns reduced profile in chunk format.  If profile changing is in
3737  * progress (either running or paused) picks the target profile (if it's
3738  * already available), otherwise falls back to plain reducing.
3739  */
3740 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3741 {
3742         u64 num_devices = root->fs_info->fs_devices->rw_devices;
3743         u64 target;
3744         u64 tmp;
3745
3746         /*
3747          * see if restripe for this chunk_type is in progress, if so
3748          * try to reduce to the target profile
3749          */
3750         spin_lock(&root->fs_info->balance_lock);
3751         target = get_restripe_target(root->fs_info, flags);
3752         if (target) {
3753                 /* pick target profile only if it's already available */
3754                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3755                         spin_unlock(&root->fs_info->balance_lock);
3756                         return extended_to_chunk(target);
3757                 }
3758         }
3759         spin_unlock(&root->fs_info->balance_lock);
3760
3761         /* First, mask out the RAID levels which aren't possible */
3762         if (num_devices == 1)
3763                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3764                            BTRFS_BLOCK_GROUP_RAID5);
3765         if (num_devices < 3)
3766                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3767         if (num_devices < 4)
3768                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3769
3770         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3771                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3772                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3773         flags &= ~tmp;
3774
3775         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3776                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3777         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3778                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3779         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3780                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3781         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3782                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3783         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3784                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3785
3786         return extended_to_chunk(flags | tmp);
3787 }
3788
3789 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
3790 {
3791         unsigned seq;
3792         u64 flags;
3793
3794         do {
3795                 flags = orig_flags;
3796                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3797
3798                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3799                         flags |= root->fs_info->avail_data_alloc_bits;
3800                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3801                         flags |= root->fs_info->avail_system_alloc_bits;
3802                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3803                         flags |= root->fs_info->avail_metadata_alloc_bits;
3804         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3805
3806         return btrfs_reduce_alloc_profile(root, flags);
3807 }
3808
3809 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3810 {
3811         u64 flags;
3812         u64 ret;
3813
3814         if (data)
3815                 flags = BTRFS_BLOCK_GROUP_DATA;
3816         else if (root == root->fs_info->chunk_root)
3817                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3818         else
3819                 flags = BTRFS_BLOCK_GROUP_METADATA;
3820
3821         ret = get_alloc_profile(root, flags);
3822         return ret;
3823 }
3824
3825 /*
3826  * This will check the space that the inode allocates from to make sure we have
3827  * enough space for bytes.
3828  */
3829 int btrfs_check_data_free_space(struct inode *inode, u64 bytes, u64 write_bytes)
3830 {
3831         struct btrfs_space_info *data_sinfo;
3832         struct btrfs_root *root = BTRFS_I(inode)->root;
3833         struct btrfs_fs_info *fs_info = root->fs_info;
3834         u64 used;
3835         int ret = 0;
3836         int need_commit = 2;
3837         int have_pinned_space;
3838
3839         /* make sure bytes are sectorsize aligned */
3840         bytes = ALIGN(bytes, root->sectorsize);
3841
3842         if (btrfs_is_free_space_inode(inode)) {
3843                 need_commit = 0;
3844                 ASSERT(current->journal_info);
3845         }
3846
3847         data_sinfo = fs_info->data_sinfo;
3848         if (!data_sinfo)
3849                 goto alloc;
3850
3851 again:
3852         /* make sure we have enough space to handle the data first */
3853         spin_lock(&data_sinfo->lock);
3854         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3855                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3856                 data_sinfo->bytes_may_use;
3857
3858         if (used + bytes > data_sinfo->total_bytes) {
3859                 struct btrfs_trans_handle *trans;
3860
3861                 /*
3862                  * if we don't have enough free bytes in this space then we need
3863                  * to alloc a new chunk.
3864                  */
3865                 if (!data_sinfo->full) {
3866                         u64 alloc_target;
3867
3868                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3869                         spin_unlock(&data_sinfo->lock);
3870 alloc:
3871                         alloc_target = btrfs_get_alloc_profile(root, 1);
3872                         /*
3873                          * It is ugly that we don't call nolock join
3874                          * transaction for the free space inode case here.
3875                          * But it is safe because we only do the data space
3876                          * reservation for the free space cache in the
3877                          * transaction context, the common join transaction
3878                          * just increase the counter of the current transaction
3879                          * handler, doesn't try to acquire the trans_lock of
3880                          * the fs.
3881                          */
3882                         trans = btrfs_join_transaction(root);
3883                         if (IS_ERR(trans))
3884                                 return PTR_ERR(trans);
3885
3886                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3887                                              alloc_target,
3888                                              CHUNK_ALLOC_NO_FORCE);
3889                         btrfs_end_transaction(trans, root);
3890                         if (ret < 0) {
3891                                 if (ret != -ENOSPC)
3892                                         return ret;
3893                                 else {
3894                                         have_pinned_space = 1;
3895                                         goto commit_trans;
3896                                 }
3897                         }
3898
3899                         if (!data_sinfo)
3900                                 data_sinfo = fs_info->data_sinfo;
3901
3902                         goto again;
3903                 }
3904
3905                 /*
3906                  * If we don't have enough pinned space to deal with this
3907                  * allocation, and no removed chunk in current transaction,
3908                  * don't bother committing the transaction.
3909                  */
3910                 have_pinned_space = percpu_counter_compare(
3911                         &data_sinfo->total_bytes_pinned,
3912                         used + bytes - data_sinfo->total_bytes);
3913                 spin_unlock(&data_sinfo->lock);
3914
3915                 /* commit the current transaction and try again */
3916 commit_trans:
3917                 if (need_commit &&
3918                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3919                         need_commit--;
3920
3921                         if (need_commit > 0)
3922                                 btrfs_wait_ordered_roots(fs_info, -1);
3923
3924                         trans = btrfs_join_transaction(root);
3925                         if (IS_ERR(trans))
3926                                 return PTR_ERR(trans);
3927                         if (have_pinned_space >= 0 ||
3928                             trans->transaction->have_free_bgs ||
3929                             need_commit > 0) {
3930                                 ret = btrfs_commit_transaction(trans, root);
3931                                 if (ret)
3932                                         return ret;
3933                                 /*
3934                                  * make sure that all running delayed iput are
3935                                  * done
3936                                  */
3937                                 down_write(&root->fs_info->delayed_iput_sem);
3938                                 up_write(&root->fs_info->delayed_iput_sem);
3939                                 goto again;
3940                         } else {
3941                                 btrfs_end_transaction(trans, root);
3942                         }
3943                 }
3944
3945                 trace_btrfs_space_reservation(root->fs_info,
3946                                               "space_info:enospc",
3947                                               data_sinfo->flags, bytes, 1);
3948                 return -ENOSPC;
3949         }
3950         ret = btrfs_qgroup_reserve(root, write_bytes);
3951         if (ret)
3952                 goto out;
3953         data_sinfo->bytes_may_use += bytes;
3954         trace_btrfs_space_reservation(root->fs_info, "space_info",
3955                                       data_sinfo->flags, bytes, 1);
3956 out:
3957         spin_unlock(&data_sinfo->lock);
3958
3959         return ret;
3960 }
3961
3962 /*
3963  * Called if we need to clear a data reservation for this inode.
3964  */
3965 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3966 {
3967         struct btrfs_root *root = BTRFS_I(inode)->root;
3968         struct btrfs_space_info *data_sinfo;
3969
3970         /* make sure bytes are sectorsize aligned */
3971         bytes = ALIGN(bytes, root->sectorsize);
3972
3973         data_sinfo = root->fs_info->data_sinfo;
3974         spin_lock(&data_sinfo->lock);
3975         WARN_ON(data_sinfo->bytes_may_use < bytes);
3976         data_sinfo->bytes_may_use -= bytes;
3977         trace_btrfs_space_reservation(root->fs_info, "space_info",
3978                                       data_sinfo->flags, bytes, 0);
3979         spin_unlock(&data_sinfo->lock);
3980 }
3981
3982 static void force_metadata_allocation(struct btrfs_fs_info *info)
3983 {
3984         struct list_head *head = &info->space_info;
3985         struct btrfs_space_info *found;
3986
3987         rcu_read_lock();
3988         list_for_each_entry_rcu(found, head, list) {
3989                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3990                         found->force_alloc = CHUNK_ALLOC_FORCE;
3991         }
3992         rcu_read_unlock();
3993 }
3994
3995 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
3996 {
3997         return (global->size << 1);
3998 }
3999
4000 static int should_alloc_chunk(struct btrfs_root *root,
4001                               struct btrfs_space_info *sinfo, int force)
4002 {
4003         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4004         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4005         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4006         u64 thresh;
4007
4008         if (force == CHUNK_ALLOC_FORCE)
4009                 return 1;
4010
4011         /*
4012          * We need to take into account the global rsv because for all intents
4013          * and purposes it's used space.  Don't worry about locking the
4014          * global_rsv, it doesn't change except when the transaction commits.
4015          */
4016         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4017                 num_allocated += calc_global_rsv_need_space(global_rsv);
4018
4019         /*
4020          * in limited mode, we want to have some free space up to
4021          * about 1% of the FS size.
4022          */
4023         if (force == CHUNK_ALLOC_LIMITED) {
4024                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4025                 thresh = max_t(u64, 64 * 1024 * 1024,
4026                                div_factor_fine(thresh, 1));
4027
4028                 if (num_bytes - num_allocated < thresh)
4029                         return 1;
4030         }
4031
4032         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
4033                 return 0;
4034         return 1;
4035 }
4036
4037 static u64 get_profile_num_devs(struct btrfs_root *root, u64 type)
4038 {
4039         u64 num_dev;
4040
4041         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4042                     BTRFS_BLOCK_GROUP_RAID0 |
4043                     BTRFS_BLOCK_GROUP_RAID5 |
4044                     BTRFS_BLOCK_GROUP_RAID6))
4045                 num_dev = root->fs_info->fs_devices->rw_devices;
4046         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4047                 num_dev = 2;
4048         else
4049                 num_dev = 1;    /* DUP or single */
4050
4051         return num_dev;
4052 }
4053
4054 /*
4055  * If @is_allocation is true, reserve space in the system space info necessary
4056  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4057  * removing a chunk.
4058  */
4059 void check_system_chunk(struct btrfs_trans_handle *trans,
4060                         struct btrfs_root *root,
4061                         u64 type)
4062 {
4063         struct btrfs_space_info *info;
4064         u64 left;
4065         u64 thresh;
4066         int ret = 0;
4067         u64 num_devs;
4068
4069         /*
4070          * Needed because we can end up allocating a system chunk and for an
4071          * atomic and race free space reservation in the chunk block reserve.
4072          */
4073         ASSERT(mutex_is_locked(&root->fs_info->chunk_mutex));
4074
4075         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4076         spin_lock(&info->lock);
4077         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4078                 info->bytes_reserved - info->bytes_readonly -
4079                 info->bytes_may_use;
4080         spin_unlock(&info->lock);
4081
4082         num_devs = get_profile_num_devs(root, type);
4083
4084         /* num_devs device items to update and 1 chunk item to add or remove */
4085         thresh = btrfs_calc_trunc_metadata_size(root, num_devs) +
4086                 btrfs_calc_trans_metadata_size(root, 1);
4087
4088         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
4089                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4090                         left, thresh, type);
4091                 dump_space_info(info, 0, 0);
4092         }
4093
4094         if (left < thresh) {
4095                 u64 flags;
4096
4097                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4098                 /*
4099                  * Ignore failure to create system chunk. We might end up not
4100                  * needing it, as we might not need to COW all nodes/leafs from
4101                  * the paths we visit in the chunk tree (they were already COWed
4102                  * or created in the current transaction for example).
4103                  */
4104                 ret = btrfs_alloc_chunk(trans, root, flags);
4105         }
4106
4107         if (!ret) {
4108                 ret = btrfs_block_rsv_add(root->fs_info->chunk_root,
4109                                           &root->fs_info->chunk_block_rsv,
4110                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4111                 if (!ret)
4112                         trans->chunk_bytes_reserved += thresh;
4113         }
4114 }
4115
4116 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4117                           struct btrfs_root *extent_root, u64 flags, int force)
4118 {
4119         struct btrfs_space_info *space_info;
4120         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4121         int wait_for_alloc = 0;
4122         int ret = 0;
4123
4124         /* Don't re-enter if we're already allocating a chunk */
4125         if (trans->allocating_chunk)
4126                 return -ENOSPC;
4127
4128         space_info = __find_space_info(extent_root->fs_info, flags);
4129         if (!space_info) {
4130                 ret = update_space_info(extent_root->fs_info, flags,
4131                                         0, 0, &space_info);
4132                 BUG_ON(ret); /* -ENOMEM */
4133         }
4134         BUG_ON(!space_info); /* Logic error */
4135
4136 again:
4137         spin_lock(&space_info->lock);
4138         if (force < space_info->force_alloc)
4139                 force = space_info->force_alloc;
4140         if (space_info->full) {
4141                 if (should_alloc_chunk(extent_root, space_info, force))
4142                         ret = -ENOSPC;
4143                 else
4144                         ret = 0;
4145                 spin_unlock(&space_info->lock);
4146                 return ret;
4147         }
4148
4149         if (!should_alloc_chunk(extent_root, space_info, force)) {
4150                 spin_unlock(&space_info->lock);
4151                 return 0;
4152         } else if (space_info->chunk_alloc) {
4153                 wait_for_alloc = 1;
4154         } else {
4155                 space_info->chunk_alloc = 1;
4156         }
4157
4158         spin_unlock(&space_info->lock);
4159
4160         mutex_lock(&fs_info->chunk_mutex);
4161
4162         /*
4163          * The chunk_mutex is held throughout the entirety of a chunk
4164          * allocation, so once we've acquired the chunk_mutex we know that the
4165          * other guy is done and we need to recheck and see if we should
4166          * allocate.
4167          */
4168         if (wait_for_alloc) {
4169                 mutex_unlock(&fs_info->chunk_mutex);
4170                 wait_for_alloc = 0;
4171                 goto again;
4172         }
4173
4174         trans->allocating_chunk = true;
4175
4176         /*
4177          * If we have mixed data/metadata chunks we want to make sure we keep
4178          * allocating mixed chunks instead of individual chunks.
4179          */
4180         if (btrfs_mixed_space_info(space_info))
4181                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4182
4183         /*
4184          * if we're doing a data chunk, go ahead and make sure that
4185          * we keep a reasonable number of metadata chunks allocated in the
4186          * FS as well.
4187          */
4188         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4189                 fs_info->data_chunk_allocations++;
4190                 if (!(fs_info->data_chunk_allocations %
4191                       fs_info->metadata_ratio))
4192                         force_metadata_allocation(fs_info);
4193         }
4194
4195         /*
4196          * Check if we have enough space in SYSTEM chunk because we may need
4197          * to update devices.
4198          */
4199         check_system_chunk(trans, extent_root, flags);
4200
4201         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4202         trans->allocating_chunk = false;
4203
4204         spin_lock(&space_info->lock);
4205         if (ret < 0 && ret != -ENOSPC)
4206                 goto out;
4207         if (ret)
4208                 space_info->full = 1;
4209         else
4210                 ret = 1;
4211
4212         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4213 out:
4214         space_info->chunk_alloc = 0;
4215         spin_unlock(&space_info->lock);
4216         mutex_unlock(&fs_info->chunk_mutex);
4217         return ret;
4218 }
4219
4220 static int can_overcommit(struct btrfs_root *root,
4221                           struct btrfs_space_info *space_info, u64 bytes,
4222                           enum btrfs_reserve_flush_enum flush)
4223 {
4224         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4225         u64 profile = btrfs_get_alloc_profile(root, 0);
4226         u64 space_size;
4227         u64 avail;
4228         u64 used;
4229
4230         used = space_info->bytes_used + space_info->bytes_reserved +
4231                 space_info->bytes_pinned + space_info->bytes_readonly;
4232
4233         /*
4234          * We only want to allow over committing if we have lots of actual space
4235          * free, but if we don't have enough space to handle the global reserve
4236          * space then we could end up having a real enospc problem when trying
4237          * to allocate a chunk or some other such important allocation.
4238          */
4239         spin_lock(&global_rsv->lock);
4240         space_size = calc_global_rsv_need_space(global_rsv);
4241         spin_unlock(&global_rsv->lock);
4242         if (used + space_size >= space_info->total_bytes)
4243                 return 0;
4244
4245         used += space_info->bytes_may_use;
4246
4247         spin_lock(&root->fs_info->free_chunk_lock);
4248         avail = root->fs_info->free_chunk_space;
4249         spin_unlock(&root->fs_info->free_chunk_lock);
4250
4251         /*
4252          * If we have dup, raid1 or raid10 then only half of the free
4253          * space is actually useable.  For raid56, the space info used
4254          * doesn't include the parity drive, so we don't have to
4255          * change the math
4256          */
4257         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4258                        BTRFS_BLOCK_GROUP_RAID1 |
4259                        BTRFS_BLOCK_GROUP_RAID10))
4260                 avail >>= 1;
4261
4262         /*
4263          * If we aren't flushing all things, let us overcommit up to
4264          * 1/2th of the space. If we can flush, don't let us overcommit
4265          * too much, let it overcommit up to 1/8 of the space.
4266          */
4267         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4268                 avail >>= 3;
4269         else
4270                 avail >>= 1;
4271
4272         if (used + bytes < space_info->total_bytes + avail)
4273                 return 1;
4274         return 0;
4275 }
4276
4277 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4278                                          unsigned long nr_pages, int nr_items)
4279 {
4280         struct super_block *sb = root->fs_info->sb;
4281
4282         if (down_read_trylock(&sb->s_umount)) {
4283                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4284                 up_read(&sb->s_umount);
4285         } else {
4286                 /*
4287                  * We needn't worry the filesystem going from r/w to r/o though
4288                  * we don't acquire ->s_umount mutex, because the filesystem
4289                  * should guarantee the delalloc inodes list be empty after
4290                  * the filesystem is readonly(all dirty pages are written to
4291                  * the disk).
4292                  */
4293                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4294                 if (!current->journal_info)
4295                         btrfs_wait_ordered_roots(root->fs_info, nr_items);
4296         }
4297 }
4298
4299 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4300 {
4301         u64 bytes;
4302         int nr;
4303
4304         bytes = btrfs_calc_trans_metadata_size(root, 1);
4305         nr = (int)div64_u64(to_reclaim, bytes);
4306         if (!nr)
4307                 nr = 1;
4308         return nr;
4309 }
4310
4311 #define EXTENT_SIZE_PER_ITEM    (256 * 1024)
4312
4313 /*
4314  * shrink metadata reservation for delalloc
4315  */
4316 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4317                             bool wait_ordered)
4318 {
4319         struct btrfs_block_rsv *block_rsv;
4320         struct btrfs_space_info *space_info;
4321         struct btrfs_trans_handle *trans;
4322         u64 delalloc_bytes;
4323         u64 max_reclaim;
4324         long time_left;
4325         unsigned long nr_pages;
4326         int loops;
4327         int items;
4328         enum btrfs_reserve_flush_enum flush;
4329
4330         /* Calc the number of the pages we need flush for space reservation */
4331         items = calc_reclaim_items_nr(root, to_reclaim);
4332         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4333
4334         trans = (struct btrfs_trans_handle *)current->journal_info;
4335         block_rsv = &root->fs_info->delalloc_block_rsv;
4336         space_info = block_rsv->space_info;
4337
4338         delalloc_bytes = percpu_counter_sum_positive(
4339                                                 &root->fs_info->delalloc_bytes);
4340         if (delalloc_bytes == 0) {
4341                 if (trans)
4342                         return;
4343                 if (wait_ordered)
4344                         btrfs_wait_ordered_roots(root->fs_info, items);
4345                 return;
4346         }
4347
4348         loops = 0;
4349         while (delalloc_bytes && loops < 3) {
4350                 max_reclaim = min(delalloc_bytes, to_reclaim);
4351                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4352                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4353                 /*
4354                  * We need to wait for the async pages to actually start before
4355                  * we do anything.
4356                  */
4357                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4358                 if (!max_reclaim)
4359                         goto skip_async;
4360
4361                 if (max_reclaim <= nr_pages)
4362                         max_reclaim = 0;
4363                 else
4364                         max_reclaim -= nr_pages;
4365
4366                 wait_event(root->fs_info->async_submit_wait,
4367                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4368                            (int)max_reclaim);
4369 skip_async:
4370                 if (!trans)
4371                         flush = BTRFS_RESERVE_FLUSH_ALL;
4372                 else
4373                         flush = BTRFS_RESERVE_NO_FLUSH;
4374                 spin_lock(&space_info->lock);
4375                 if (can_overcommit(root, space_info, orig, flush)) {
4376                         spin_unlock(&space_info->lock);
4377                         break;
4378                 }
4379                 spin_unlock(&space_info->lock);
4380
4381                 loops++;
4382                 if (wait_ordered && !trans) {
4383                         btrfs_wait_ordered_roots(root->fs_info, items);
4384                 } else {
4385                         time_left = schedule_timeout_killable(1);
4386                         if (time_left)
4387                                 break;
4388                 }
4389                 delalloc_bytes = percpu_counter_sum_positive(
4390                                                 &root->fs_info->delalloc_bytes);
4391         }
4392 }
4393
4394 /**
4395  * maybe_commit_transaction - possibly commit the transaction if its ok to
4396  * @root - the root we're allocating for
4397  * @bytes - the number of bytes we want to reserve
4398  * @force - force the commit
4399  *
4400  * This will check to make sure that committing the transaction will actually
4401  * get us somewhere and then commit the transaction if it does.  Otherwise it
4402  * will return -ENOSPC.
4403  */
4404 static int may_commit_transaction(struct btrfs_root *root,
4405                                   struct btrfs_space_info *space_info,
4406                                   u64 bytes, int force)
4407 {
4408         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4409         struct btrfs_trans_handle *trans;
4410
4411         trans = (struct btrfs_trans_handle *)current->journal_info;
4412         if (trans)
4413                 return -EAGAIN;
4414
4415         if (force)
4416                 goto commit;
4417
4418         /* See if there is enough pinned space to make this reservation */
4419         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4420                                    bytes) >= 0)
4421                 goto commit;
4422
4423         /*
4424          * See if there is some space in the delayed insertion reservation for
4425          * this reservation.
4426          */
4427         if (space_info != delayed_rsv->space_info)
4428                 return -ENOSPC;
4429
4430         spin_lock(&delayed_rsv->lock);
4431         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4432                                    bytes - delayed_rsv->size) >= 0) {
4433                 spin_unlock(&delayed_rsv->lock);
4434                 return -ENOSPC;
4435         }
4436         spin_unlock(&delayed_rsv->lock);
4437
4438 commit:
4439         trans = btrfs_join_transaction(root);
4440         if (IS_ERR(trans))
4441                 return -ENOSPC;
4442
4443         return btrfs_commit_transaction(trans, root);
4444 }
4445
4446 enum flush_state {
4447         FLUSH_DELAYED_ITEMS_NR  =       1,
4448         FLUSH_DELAYED_ITEMS     =       2,
4449         FLUSH_DELALLOC          =       3,
4450         FLUSH_DELALLOC_WAIT     =       4,
4451         ALLOC_CHUNK             =       5,
4452         COMMIT_TRANS            =       6,
4453 };
4454
4455 static int flush_space(struct btrfs_root *root,
4456                        struct btrfs_space_info *space_info, u64 num_bytes,
4457                        u64 orig_bytes, int state)
4458 {
4459         struct btrfs_trans_handle *trans;
4460         int nr;
4461         int ret = 0;
4462
4463         switch (state) {
4464         case FLUSH_DELAYED_ITEMS_NR:
4465         case FLUSH_DELAYED_ITEMS:
4466                 if (state == FLUSH_DELAYED_ITEMS_NR)
4467                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4468                 else
4469                         nr = -1;
4470
4471                 trans = btrfs_join_transaction(root);
4472                 if (IS_ERR(trans)) {
4473                         ret = PTR_ERR(trans);
4474                         break;
4475                 }
4476                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4477                 btrfs_end_transaction(trans, root);
4478                 break;
4479         case FLUSH_DELALLOC:
4480         case FLUSH_DELALLOC_WAIT:
4481                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4482                                 state == FLUSH_DELALLOC_WAIT);
4483                 break;
4484         case ALLOC_CHUNK:
4485                 trans = btrfs_join_transaction(root);
4486                 if (IS_ERR(trans)) {
4487                         ret = PTR_ERR(trans);
4488                         break;
4489                 }
4490                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4491                                      btrfs_get_alloc_profile(root, 0),
4492                                      CHUNK_ALLOC_NO_FORCE);
4493                 btrfs_end_transaction(trans, root);
4494                 if (ret == -ENOSPC)
4495                         ret = 0;
4496                 break;
4497         case COMMIT_TRANS:
4498                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4499                 break;
4500         default:
4501                 ret = -ENOSPC;
4502                 break;
4503         }
4504
4505         return ret;
4506 }
4507
4508 static inline u64
4509 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4510                                  struct btrfs_space_info *space_info)
4511 {
4512         u64 used;
4513         u64 expected;
4514         u64 to_reclaim;
4515
4516         to_reclaim = min_t(u64, num_online_cpus() * 1024 * 1024,
4517                                 16 * 1024 * 1024);
4518         spin_lock(&space_info->lock);
4519         if (can_overcommit(root, space_info, to_reclaim,
4520                            BTRFS_RESERVE_FLUSH_ALL)) {
4521                 to_reclaim = 0;
4522                 goto out;
4523         }
4524
4525         used = space_info->bytes_used + space_info->bytes_reserved +
4526                space_info->bytes_pinned + space_info->bytes_readonly +
4527                space_info->bytes_may_use;
4528         if (can_overcommit(root, space_info, 1024 * 1024,
4529                            BTRFS_RESERVE_FLUSH_ALL))
4530                 expected = div_factor_fine(space_info->total_bytes, 95);
4531         else
4532                 expected = div_factor_fine(space_info->total_bytes, 90);
4533
4534         if (used > expected)
4535                 to_reclaim = used - expected;
4536         else
4537                 to_reclaim = 0;
4538         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4539                                      space_info->bytes_reserved);
4540 out:
4541         spin_unlock(&space_info->lock);
4542
4543         return to_reclaim;
4544 }
4545
4546 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4547                                         struct btrfs_fs_info *fs_info, u64 used)
4548 {
4549         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4550
4551         /* If we're just plain full then async reclaim just slows us down. */
4552         if (space_info->bytes_used >= thresh)
4553                 return 0;
4554
4555         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4556                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4557 }
4558
4559 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4560                                        struct btrfs_fs_info *fs_info,
4561                                        int flush_state)
4562 {
4563         u64 used;
4564
4565         spin_lock(&space_info->lock);
4566         /*
4567          * We run out of space and have not got any free space via flush_space,
4568          * so don't bother doing async reclaim.
4569          */
4570         if (flush_state > COMMIT_TRANS && space_info->full) {
4571                 spin_unlock(&space_info->lock);
4572                 return 0;
4573         }
4574
4575         used = space_info->bytes_used + space_info->bytes_reserved +
4576                space_info->bytes_pinned + space_info->bytes_readonly +
4577                space_info->bytes_may_use;
4578         if (need_do_async_reclaim(space_info, fs_info, used)) {
4579                 spin_unlock(&space_info->lock);
4580                 return 1;
4581         }
4582         spin_unlock(&space_info->lock);
4583
4584         return 0;
4585 }
4586
4587 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4588 {
4589         struct btrfs_fs_info *fs_info;
4590         struct btrfs_space_info *space_info;
4591         u64 to_reclaim;
4592         int flush_state;
4593
4594         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4595         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4596
4597         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4598                                                       space_info);
4599         if (!to_reclaim)
4600                 return;
4601
4602         flush_state = FLUSH_DELAYED_ITEMS_NR;
4603         do {
4604                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4605                             to_reclaim, flush_state);
4606                 flush_state++;
4607                 if (!btrfs_need_do_async_reclaim(space_info, fs_info,
4608                                                  flush_state))
4609                         return;
4610         } while (flush_state < COMMIT_TRANS);
4611 }
4612
4613 void btrfs_init_async_reclaim_work(struct work_struct *work)
4614 {
4615         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4616 }
4617
4618 /**
4619  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4620  * @root - the root we're allocating for
4621  * @block_rsv - the block_rsv we're allocating for
4622  * @orig_bytes - the number of bytes we want
4623  * @flush - whether or not we can flush to make our reservation
4624  *
4625  * This will reserve orgi_bytes number of bytes from the space info associated
4626  * with the block_rsv.  If there is not enough space it will make an attempt to
4627  * flush out space to make room.  It will do this by flushing delalloc if
4628  * possible or committing the transaction.  If flush is 0 then no attempts to
4629  * regain reservations will be made and this will fail if there is not enough
4630  * space already.
4631  */
4632 static int reserve_metadata_bytes(struct btrfs_root *root,
4633                                   struct btrfs_block_rsv *block_rsv,
4634                                   u64 orig_bytes,
4635                                   enum btrfs_reserve_flush_enum flush)
4636 {
4637         struct btrfs_space_info *space_info = block_rsv->space_info;
4638         u64 used;
4639         u64 num_bytes = orig_bytes;
4640         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4641         int ret = 0;
4642         bool flushing = false;
4643
4644 again:
4645         ret = 0;
4646         spin_lock(&space_info->lock);
4647         /*
4648          * We only want to wait if somebody other than us is flushing and we
4649          * are actually allowed to flush all things.
4650          */
4651         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4652                space_info->flush) {
4653                 spin_unlock(&space_info->lock);
4654                 /*
4655                  * If we have a trans handle we can't wait because the flusher
4656                  * may have to commit the transaction, which would mean we would
4657                  * deadlock since we are waiting for the flusher to finish, but
4658                  * hold the current transaction open.
4659                  */
4660                 if (current->journal_info)
4661                         return -EAGAIN;
4662                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4663                 /* Must have been killed, return */
4664                 if (ret)
4665                         return -EINTR;
4666
4667                 spin_lock(&space_info->lock);
4668         }
4669
4670         ret = -ENOSPC;
4671         used = space_info->bytes_used + space_info->bytes_reserved +
4672                 space_info->bytes_pinned + space_info->bytes_readonly +
4673                 space_info->bytes_may_use;
4674
4675         /*
4676          * The idea here is that we've not already over-reserved the block group
4677          * then we can go ahead and save our reservation first and then start
4678          * flushing if we need to.  Otherwise if we've already overcommitted
4679          * lets start flushing stuff first and then come back and try to make
4680          * our reservation.
4681          */
4682         if (used <= space_info->total_bytes) {
4683                 if (used + orig_bytes <= space_info->total_bytes) {
4684                         space_info->bytes_may_use += orig_bytes;
4685                         trace_btrfs_space_reservation(root->fs_info,
4686                                 "space_info", space_info->flags, orig_bytes, 1);
4687                         ret = 0;
4688                 } else {
4689                         /*
4690                          * Ok set num_bytes to orig_bytes since we aren't
4691                          * overocmmitted, this way we only try and reclaim what
4692                          * we need.
4693                          */
4694                         num_bytes = orig_bytes;
4695                 }
4696         } else {
4697                 /*
4698                  * Ok we're over committed, set num_bytes to the overcommitted
4699                  * amount plus the amount of bytes that we need for this
4700                  * reservation.
4701                  */
4702                 num_bytes = used - space_info->total_bytes +
4703                         (orig_bytes * 2);
4704         }
4705
4706         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4707                 space_info->bytes_may_use += orig_bytes;
4708                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4709                                               space_info->flags, orig_bytes,
4710                                               1);
4711                 ret = 0;
4712         }
4713
4714         /*
4715          * Couldn't make our reservation, save our place so while we're trying
4716          * to reclaim space we can actually use it instead of somebody else
4717          * stealing it from us.
4718          *
4719          * We make the other tasks wait for the flush only when we can flush
4720          * all things.
4721          */
4722         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4723                 flushing = true;
4724                 space_info->flush = 1;
4725         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
4726                 used += orig_bytes;
4727                 /*
4728                  * We will do the space reservation dance during log replay,
4729                  * which means we won't have fs_info->fs_root set, so don't do
4730                  * the async reclaim as we will panic.
4731                  */
4732                 if (!root->fs_info->log_root_recovering &&
4733                     need_do_async_reclaim(space_info, root->fs_info, used) &&
4734                     !work_busy(&root->fs_info->async_reclaim_work))
4735                         queue_work(system_unbound_wq,
4736                                    &root->fs_info->async_reclaim_work);
4737         }
4738         spin_unlock(&space_info->lock);
4739
4740         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4741                 goto out;
4742
4743         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4744                           flush_state);
4745         flush_state++;
4746
4747         /*
4748          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4749          * would happen. So skip delalloc flush.
4750          */
4751         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4752             (flush_state == FLUSH_DELALLOC ||
4753              flush_state == FLUSH_DELALLOC_WAIT))
4754                 flush_state = ALLOC_CHUNK;
4755
4756         if (!ret)
4757                 goto again;
4758         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4759                  flush_state < COMMIT_TRANS)
4760                 goto again;
4761         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4762                  flush_state <= COMMIT_TRANS)
4763                 goto again;
4764
4765 out:
4766         if (ret == -ENOSPC &&
4767             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4768                 struct btrfs_block_rsv *global_rsv =
4769                         &root->fs_info->global_block_rsv;
4770
4771                 if (block_rsv != global_rsv &&
4772                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4773                         ret = 0;
4774         }
4775         if (ret == -ENOSPC)
4776                 trace_btrfs_space_reservation(root->fs_info,
4777                                               "space_info:enospc",
4778                                               space_info->flags, orig_bytes, 1);
4779         if (flushing) {
4780                 spin_lock(&space_info->lock);
4781                 space_info->flush = 0;
4782                 wake_up_all(&space_info->wait);
4783                 spin_unlock(&space_info->lock);
4784         }
4785         return ret;
4786 }
4787
4788 static struct btrfs_block_rsv *get_block_rsv(
4789                                         const struct btrfs_trans_handle *trans,
4790                                         const struct btrfs_root *root)
4791 {
4792         struct btrfs_block_rsv *block_rsv = NULL;
4793
4794         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4795                 block_rsv = trans->block_rsv;
4796
4797         if (root == root->fs_info->csum_root && trans->adding_csums)
4798                 block_rsv = trans->block_rsv;
4799
4800         if (root == root->fs_info->uuid_root)
4801                 block_rsv = trans->block_rsv;
4802
4803         if (!block_rsv)
4804                 block_rsv = root->block_rsv;
4805
4806         if (!block_rsv)
4807                 block_rsv = &root->fs_info->empty_block_rsv;
4808
4809         return block_rsv;
4810 }
4811
4812 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4813                                u64 num_bytes)
4814 {
4815         int ret = -ENOSPC;
4816         spin_lock(&block_rsv->lock);
4817         if (block_rsv->reserved >= num_bytes) {
4818                 block_rsv->reserved -= num_bytes;
4819                 if (block_rsv->reserved < block_rsv->size)
4820                         block_rsv->full = 0;
4821                 ret = 0;
4822         }
4823         spin_unlock(&block_rsv->lock);
4824         return ret;
4825 }
4826
4827 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4828                                 u64 num_bytes, int update_size)
4829 {
4830         spin_lock(&block_rsv->lock);
4831         block_rsv->reserved += num_bytes;
4832         if (update_size)
4833                 block_rsv->size += num_bytes;
4834         else if (block_rsv->reserved >= block_rsv->size)
4835                 block_rsv->full = 1;
4836         spin_unlock(&block_rsv->lock);
4837 }
4838
4839 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4840                              struct btrfs_block_rsv *dest, u64 num_bytes,
4841                              int min_factor)
4842 {
4843         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4844         u64 min_bytes;
4845
4846         if (global_rsv->space_info != dest->space_info)
4847                 return -ENOSPC;
4848
4849         spin_lock(&global_rsv->lock);
4850         min_bytes = div_factor(global_rsv->size, min_factor);
4851         if (global_rsv->reserved < min_bytes + num_bytes) {
4852                 spin_unlock(&global_rsv->lock);
4853                 return -ENOSPC;
4854         }
4855         global_rsv->reserved -= num_bytes;
4856         if (global_rsv->reserved < global_rsv->size)
4857                 global_rsv->full = 0;
4858         spin_unlock(&global_rsv->lock);
4859
4860         block_rsv_add_bytes(dest, num_bytes, 1);
4861         return 0;
4862 }
4863
4864 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4865                                     struct btrfs_block_rsv *block_rsv,
4866                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4867 {
4868         struct btrfs_space_info *space_info = block_rsv->space_info;
4869
4870         spin_lock(&block_rsv->lock);
4871         if (num_bytes == (u64)-1)
4872                 num_bytes = block_rsv->size;
4873         block_rsv->size -= num_bytes;
4874         if (block_rsv->reserved >= block_rsv->size) {
4875                 num_bytes = block_rsv->reserved - block_rsv->size;
4876                 block_rsv->reserved = block_rsv->size;
4877                 block_rsv->full = 1;
4878         } else {
4879                 num_bytes = 0;
4880         }
4881         spin_unlock(&block_rsv->lock);
4882
4883         if (num_bytes > 0) {
4884                 if (dest) {
4885                         spin_lock(&dest->lock);
4886                         if (!dest->full) {
4887                                 u64 bytes_to_add;
4888
4889                                 bytes_to_add = dest->size - dest->reserved;
4890                                 bytes_to_add = min(num_bytes, bytes_to_add);
4891                                 dest->reserved += bytes_to_add;
4892                                 if (dest->reserved >= dest->size)
4893                                         dest->full = 1;
4894                                 num_bytes -= bytes_to_add;
4895                         }
4896                         spin_unlock(&dest->lock);
4897                 }
4898                 if (num_bytes) {
4899                         spin_lock(&space_info->lock);
4900                         space_info->bytes_may_use -= num_bytes;
4901                         trace_btrfs_space_reservation(fs_info, "space_info",
4902                                         space_info->flags, num_bytes, 0);
4903                         spin_unlock(&space_info->lock);
4904                 }
4905         }
4906 }
4907
4908 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4909                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4910 {
4911         int ret;
4912
4913         ret = block_rsv_use_bytes(src, num_bytes);
4914         if (ret)
4915                 return ret;
4916
4917         block_rsv_add_bytes(dst, num_bytes, 1);
4918         return 0;
4919 }
4920
4921 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4922 {
4923         memset(rsv, 0, sizeof(*rsv));
4924         spin_lock_init(&rsv->lock);
4925         rsv->type = type;
4926 }
4927
4928 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4929                                               unsigned short type)
4930 {
4931         struct btrfs_block_rsv *block_rsv;
4932         struct btrfs_fs_info *fs_info = root->fs_info;
4933
4934         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4935         if (!block_rsv)
4936                 return NULL;
4937
4938         btrfs_init_block_rsv(block_rsv, type);
4939         block_rsv->space_info = __find_space_info(fs_info,
4940                                                   BTRFS_BLOCK_GROUP_METADATA);
4941         return block_rsv;
4942 }
4943
4944 void btrfs_free_block_rsv(struct btrfs_root *root,
4945                           struct btrfs_block_rsv *rsv)
4946 {
4947         if (!rsv)
4948                 return;
4949         btrfs_block_rsv_release(root, rsv, (u64)-1);
4950         kfree(rsv);
4951 }
4952
4953 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
4954 {
4955         kfree(rsv);
4956 }
4957
4958 int btrfs_block_rsv_add(struct btrfs_root *root,
4959                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4960                         enum btrfs_reserve_flush_enum flush)
4961 {
4962         int ret;
4963
4964         if (num_bytes == 0)
4965                 return 0;
4966
4967         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4968         if (!ret) {
4969                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4970                 return 0;
4971         }
4972
4973         return ret;
4974 }
4975
4976 int btrfs_block_rsv_check(struct btrfs_root *root,
4977                           struct btrfs_block_rsv *block_rsv, int min_factor)
4978 {
4979         u64 num_bytes = 0;
4980         int ret = -ENOSPC;
4981
4982         if (!block_rsv)
4983                 return 0;
4984
4985         spin_lock(&block_rsv->lock);
4986         num_bytes = div_factor(block_rsv->size, min_factor);
4987         if (block_rsv->reserved >= num_bytes)
4988                 ret = 0;
4989         spin_unlock(&block_rsv->lock);
4990
4991         return ret;
4992 }
4993
4994 int btrfs_block_rsv_refill(struct btrfs_root *root,
4995                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4996                            enum btrfs_reserve_flush_enum flush)
4997 {
4998         u64 num_bytes = 0;
4999         int ret = -ENOSPC;
5000
5001         if (!block_rsv)
5002                 return 0;
5003
5004         spin_lock(&block_rsv->lock);
5005         num_bytes = min_reserved;
5006         if (block_rsv->reserved >= num_bytes)
5007                 ret = 0;
5008         else
5009                 num_bytes -= block_rsv->reserved;
5010         spin_unlock(&block_rsv->lock);
5011
5012         if (!ret)
5013                 return 0;
5014
5015         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5016         if (!ret) {
5017                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5018                 return 0;
5019         }
5020
5021         return ret;
5022 }
5023
5024 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
5025                             struct btrfs_block_rsv *dst_rsv,
5026                             u64 num_bytes)
5027 {
5028         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5029 }
5030
5031 void btrfs_block_rsv_release(struct btrfs_root *root,
5032                              struct btrfs_block_rsv *block_rsv,
5033                              u64 num_bytes)
5034 {
5035         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5036         if (global_rsv == block_rsv ||
5037             block_rsv->space_info != global_rsv->space_info)
5038                 global_rsv = NULL;
5039         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5040                                 num_bytes);
5041 }
5042
5043 /*
5044  * helper to calculate size of global block reservation.
5045  * the desired value is sum of space used by extent tree,
5046  * checksum tree and root tree
5047  */
5048 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
5049 {
5050         struct btrfs_space_info *sinfo;
5051         u64 num_bytes;
5052         u64 meta_used;
5053         u64 data_used;
5054         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
5055
5056         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
5057         spin_lock(&sinfo->lock);
5058         data_used = sinfo->bytes_used;
5059         spin_unlock(&sinfo->lock);
5060
5061         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5062         spin_lock(&sinfo->lock);
5063         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
5064                 data_used = 0;
5065         meta_used = sinfo->bytes_used;
5066         spin_unlock(&sinfo->lock);
5067
5068         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
5069                     csum_size * 2;
5070         num_bytes += div_u64(data_used + meta_used, 50);
5071
5072         if (num_bytes * 3 > meta_used)
5073                 num_bytes = div_u64(meta_used, 3);
5074
5075         return ALIGN(num_bytes, fs_info->extent_root->nodesize << 10);
5076 }
5077
5078 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5079 {
5080         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5081         struct btrfs_space_info *sinfo = block_rsv->space_info;
5082         u64 num_bytes;
5083
5084         num_bytes = calc_global_metadata_size(fs_info);
5085
5086         spin_lock(&sinfo->lock);
5087         spin_lock(&block_rsv->lock);
5088
5089         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
5090
5091         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5092                     sinfo->bytes_reserved + sinfo->bytes_readonly +
5093                     sinfo->bytes_may_use;
5094
5095         if (sinfo->total_bytes > num_bytes) {
5096                 num_bytes = sinfo->total_bytes - num_bytes;
5097                 block_rsv->reserved += num_bytes;
5098                 sinfo->bytes_may_use += num_bytes;
5099                 trace_btrfs_space_reservation(fs_info, "space_info",
5100                                       sinfo->flags, num_bytes, 1);
5101         }
5102
5103         if (block_rsv->reserved >= block_rsv->size) {
5104                 num_bytes = block_rsv->reserved - block_rsv->size;
5105                 sinfo->bytes_may_use -= num_bytes;
5106                 trace_btrfs_space_reservation(fs_info, "space_info",
5107                                       sinfo->flags, num_bytes, 0);
5108                 block_rsv->reserved = block_rsv->size;
5109                 block_rsv->full = 1;
5110         }
5111
5112         spin_unlock(&block_rsv->lock);
5113         spin_unlock(&sinfo->lock);
5114 }
5115
5116 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5117 {
5118         struct btrfs_space_info *space_info;
5119
5120         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5121         fs_info->chunk_block_rsv.space_info = space_info;
5122
5123         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5124         fs_info->global_block_rsv.space_info = space_info;
5125         fs_info->delalloc_block_rsv.space_info = space_info;
5126         fs_info->trans_block_rsv.space_info = space_info;
5127         fs_info->empty_block_rsv.space_info = space_info;
5128         fs_info->delayed_block_rsv.space_info = space_info;
5129
5130         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5131         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5132         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5133         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5134         if (fs_info->quota_root)
5135                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5136         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5137
5138         update_global_block_rsv(fs_info);
5139 }
5140
5141 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5142 {
5143         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5144                                 (u64)-1);
5145         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5146         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5147         WARN_ON(fs_info->trans_block_rsv.size > 0);
5148         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5149         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5150         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5151         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5152         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5153 }
5154
5155 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5156                                   struct btrfs_root *root)
5157 {
5158         if (!trans->block_rsv)
5159                 return;
5160
5161         if (!trans->bytes_reserved)
5162                 return;
5163
5164         trace_btrfs_space_reservation(root->fs_info, "transaction",
5165                                       trans->transid, trans->bytes_reserved, 0);
5166         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5167         trans->bytes_reserved = 0;
5168 }
5169
5170 /*
5171  * To be called after all the new block groups attached to the transaction
5172  * handle have been created (btrfs_create_pending_block_groups()).
5173  */
5174 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5175 {
5176         struct btrfs_fs_info *fs_info = trans->root->fs_info;
5177
5178         if (!trans->chunk_bytes_reserved)
5179                 return;
5180
5181         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5182
5183         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5184                                 trans->chunk_bytes_reserved);
5185         trans->chunk_bytes_reserved = 0;
5186 }
5187
5188 /* Can only return 0 or -ENOSPC */
5189 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5190                                   struct inode *inode)
5191 {
5192         struct btrfs_root *root = BTRFS_I(inode)->root;
5193         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
5194         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5195
5196         /*
5197          * We need to hold space in order to delete our orphan item once we've
5198          * added it, so this takes the reservation so we can release it later
5199          * when we are truly done with the orphan item.
5200          */
5201         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5202         trace_btrfs_space_reservation(root->fs_info, "orphan",
5203                                       btrfs_ino(inode), num_bytes, 1);
5204         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5205 }
5206
5207 void btrfs_orphan_release_metadata(struct inode *inode)
5208 {
5209         struct btrfs_root *root = BTRFS_I(inode)->root;
5210         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5211         trace_btrfs_space_reservation(root->fs_info, "orphan",
5212                                       btrfs_ino(inode), num_bytes, 0);
5213         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5214 }
5215
5216 /*
5217  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5218  * root: the root of the parent directory
5219  * rsv: block reservation
5220  * items: the number of items that we need do reservation
5221  * qgroup_reserved: used to return the reserved size in qgroup
5222  *
5223  * This function is used to reserve the space for snapshot/subvolume
5224  * creation and deletion. Those operations are different with the
5225  * common file/directory operations, they change two fs/file trees
5226  * and root tree, the number of items that the qgroup reserves is
5227  * different with the free space reservation. So we can not use
5228  * the space reseravtion mechanism in start_transaction().
5229  */
5230 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5231                                      struct btrfs_block_rsv *rsv,
5232                                      int items,
5233                                      u64 *qgroup_reserved,
5234                                      bool use_global_rsv)
5235 {
5236         u64 num_bytes;
5237         int ret;
5238         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5239
5240         if (root->fs_info->quota_enabled) {
5241                 /* One for parent inode, two for dir entries */
5242                 num_bytes = 3 * root->nodesize;
5243                 ret = btrfs_qgroup_reserve(root, num_bytes);
5244                 if (ret)
5245                         return ret;
5246         } else {
5247                 num_bytes = 0;
5248         }
5249
5250         *qgroup_reserved = num_bytes;
5251
5252         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5253         rsv->space_info = __find_space_info(root->fs_info,
5254                                             BTRFS_BLOCK_GROUP_METADATA);
5255         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5256                                   BTRFS_RESERVE_FLUSH_ALL);
5257
5258         if (ret == -ENOSPC && use_global_rsv)
5259                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
5260
5261         if (ret) {
5262                 if (*qgroup_reserved)
5263                         btrfs_qgroup_free(root, *qgroup_reserved);
5264         }
5265
5266         return ret;
5267 }
5268
5269 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5270                                       struct btrfs_block_rsv *rsv,
5271                                       u64 qgroup_reserved)
5272 {
5273         btrfs_block_rsv_release(root, rsv, (u64)-1);
5274 }
5275
5276 /**
5277  * drop_outstanding_extent - drop an outstanding extent
5278  * @inode: the inode we're dropping the extent for
5279  * @num_bytes: the number of bytes we're relaseing.
5280  *
5281  * This is called when we are freeing up an outstanding extent, either called
5282  * after an error or after an extent is written.  This will return the number of
5283  * reserved extents that need to be freed.  This must be called with
5284  * BTRFS_I(inode)->lock held.
5285  */
5286 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5287 {
5288         unsigned drop_inode_space = 0;
5289         unsigned dropped_extents = 0;
5290         unsigned num_extents = 0;
5291
5292         num_extents = (unsigned)div64_u64(num_bytes +
5293                                           BTRFS_MAX_EXTENT_SIZE - 1,
5294                                           BTRFS_MAX_EXTENT_SIZE);
5295         ASSERT(num_extents);
5296         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5297         BTRFS_I(inode)->outstanding_extents -= num_extents;
5298
5299         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5300             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5301                                &BTRFS_I(inode)->runtime_flags))
5302                 drop_inode_space = 1;
5303
5304         /*
5305          * If we have more or the same amount of outsanding extents than we have
5306          * reserved then we need to leave the reserved extents count alone.
5307          */
5308         if (BTRFS_I(inode)->outstanding_extents >=
5309             BTRFS_I(inode)->reserved_extents)
5310                 return drop_inode_space;
5311
5312         dropped_extents = BTRFS_I(inode)->reserved_extents -
5313                 BTRFS_I(inode)->outstanding_extents;
5314         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5315         return dropped_extents + drop_inode_space;
5316 }
5317
5318 /**
5319  * calc_csum_metadata_size - return the amount of metada space that must be
5320  *      reserved/free'd for the given bytes.
5321  * @inode: the inode we're manipulating
5322  * @num_bytes: the number of bytes in question
5323  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5324  *
5325  * This adjusts the number of csum_bytes in the inode and then returns the
5326  * correct amount of metadata that must either be reserved or freed.  We
5327  * calculate how many checksums we can fit into one leaf and then divide the
5328  * number of bytes that will need to be checksumed by this value to figure out
5329  * how many checksums will be required.  If we are adding bytes then the number
5330  * may go up and we will return the number of additional bytes that must be
5331  * reserved.  If it is going down we will return the number of bytes that must
5332  * be freed.
5333  *
5334  * This must be called with BTRFS_I(inode)->lock held.
5335  */
5336 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5337                                    int reserve)
5338 {
5339         struct btrfs_root *root = BTRFS_I(inode)->root;
5340         u64 old_csums, num_csums;
5341
5342         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5343             BTRFS_I(inode)->csum_bytes == 0)
5344                 return 0;
5345
5346         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5347         if (reserve)
5348                 BTRFS_I(inode)->csum_bytes += num_bytes;
5349         else
5350                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5351         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5352
5353         /* No change, no need to reserve more */
5354         if (old_csums == num_csums)
5355                 return 0;
5356
5357         if (reserve)
5358                 return btrfs_calc_trans_metadata_size(root,
5359                                                       num_csums - old_csums);
5360
5361         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5362 }
5363
5364 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5365 {
5366         struct btrfs_root *root = BTRFS_I(inode)->root;
5367         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5368         u64 to_reserve = 0;
5369         u64 csum_bytes;
5370         unsigned nr_extents = 0;
5371         int extra_reserve = 0;
5372         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5373         int ret = 0;
5374         bool delalloc_lock = true;
5375         u64 to_free = 0;
5376         unsigned dropped;
5377
5378         /* If we are a free space inode we need to not flush since we will be in
5379          * the middle of a transaction commit.  We also don't need the delalloc
5380          * mutex since we won't race with anybody.  We need this mostly to make
5381          * lockdep shut its filthy mouth.
5382          */
5383         if (btrfs_is_free_space_inode(inode)) {
5384                 flush = BTRFS_RESERVE_NO_FLUSH;
5385                 delalloc_lock = false;
5386         }
5387
5388         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5389             btrfs_transaction_in_commit(root->fs_info))
5390                 schedule_timeout(1);
5391
5392         if (delalloc_lock)
5393                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5394
5395         num_bytes = ALIGN(num_bytes, root->sectorsize);
5396
5397         spin_lock(&BTRFS_I(inode)->lock);
5398         nr_extents = (unsigned)div64_u64(num_bytes +
5399                                          BTRFS_MAX_EXTENT_SIZE - 1,
5400                                          BTRFS_MAX_EXTENT_SIZE);
5401         BTRFS_I(inode)->outstanding_extents += nr_extents;
5402         nr_extents = 0;
5403
5404         if (BTRFS_I(inode)->outstanding_extents >
5405             BTRFS_I(inode)->reserved_extents)
5406                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5407                         BTRFS_I(inode)->reserved_extents;
5408
5409         /*
5410          * Add an item to reserve for updating the inode when we complete the
5411          * delalloc io.
5412          */
5413         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5414                       &BTRFS_I(inode)->runtime_flags)) {
5415                 nr_extents++;
5416                 extra_reserve = 1;
5417         }
5418
5419         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5420         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5421         csum_bytes = BTRFS_I(inode)->csum_bytes;
5422         spin_unlock(&BTRFS_I(inode)->lock);
5423
5424         if (root->fs_info->quota_enabled) {
5425                 ret = btrfs_qgroup_reserve(root, nr_extents * root->nodesize);
5426                 if (ret)
5427                         goto out_fail;
5428         }
5429
5430         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5431         if (unlikely(ret)) {
5432                 if (root->fs_info->quota_enabled)
5433                         btrfs_qgroup_free(root, nr_extents * root->nodesize);
5434                 goto out_fail;
5435         }
5436
5437         spin_lock(&BTRFS_I(inode)->lock);
5438         if (extra_reserve) {
5439                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5440                         &BTRFS_I(inode)->runtime_flags);
5441                 nr_extents--;
5442         }
5443         BTRFS_I(inode)->reserved_extents += nr_extents;
5444         spin_unlock(&BTRFS_I(inode)->lock);
5445
5446         if (delalloc_lock)
5447                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5448
5449         if (to_reserve)
5450                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5451                                               btrfs_ino(inode), to_reserve, 1);
5452         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5453
5454         return 0;
5455
5456 out_fail:
5457         spin_lock(&BTRFS_I(inode)->lock);
5458         dropped = drop_outstanding_extent(inode, num_bytes);
5459         /*
5460          * If the inodes csum_bytes is the same as the original
5461          * csum_bytes then we know we haven't raced with any free()ers
5462          * so we can just reduce our inodes csum bytes and carry on.
5463          */
5464         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5465                 calc_csum_metadata_size(inode, num_bytes, 0);
5466         } else {
5467                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5468                 u64 bytes;
5469
5470                 /*
5471                  * This is tricky, but first we need to figure out how much we
5472                  * free'd from any free-ers that occured during this
5473                  * reservation, so we reset ->csum_bytes to the csum_bytes
5474                  * before we dropped our lock, and then call the free for the
5475                  * number of bytes that were freed while we were trying our
5476                  * reservation.
5477                  */
5478                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5479                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5480                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5481
5482
5483                 /*
5484                  * Now we need to see how much we would have freed had we not
5485                  * been making this reservation and our ->csum_bytes were not
5486                  * artificially inflated.
5487                  */
5488                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5489                 bytes = csum_bytes - orig_csum_bytes;
5490                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5491
5492                 /*
5493                  * Now reset ->csum_bytes to what it should be.  If bytes is
5494                  * more than to_free then we would have free'd more space had we
5495                  * not had an artificially high ->csum_bytes, so we need to free
5496                  * the remainder.  If bytes is the same or less then we don't
5497                  * need to do anything, the other free-ers did the correct
5498                  * thing.
5499                  */
5500                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5501                 if (bytes > to_free)
5502                         to_free = bytes - to_free;
5503                 else
5504                         to_free = 0;
5505         }
5506         spin_unlock(&BTRFS_I(inode)->lock);
5507         if (dropped)
5508                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5509
5510         if (to_free) {
5511                 btrfs_block_rsv_release(root, block_rsv, to_free);
5512                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5513                                               btrfs_ino(inode), to_free, 0);
5514         }
5515         if (delalloc_lock)
5516                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5517         return ret;
5518 }
5519
5520 /**
5521  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5522  * @inode: the inode to release the reservation for
5523  * @num_bytes: the number of bytes we're releasing
5524  *
5525  * This will release the metadata reservation for an inode.  This can be called
5526  * once we complete IO for a given set of bytes to release their metadata
5527  * reservations.
5528  */
5529 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5530 {
5531         struct btrfs_root *root = BTRFS_I(inode)->root;
5532         u64 to_free = 0;
5533         unsigned dropped;
5534
5535         num_bytes = ALIGN(num_bytes, root->sectorsize);
5536         spin_lock(&BTRFS_I(inode)->lock);
5537         dropped = drop_outstanding_extent(inode, num_bytes);
5538
5539         if (num_bytes)
5540                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5541         spin_unlock(&BTRFS_I(inode)->lock);
5542         if (dropped > 0)
5543                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5544
5545         if (btrfs_test_is_dummy_root(root))
5546                 return;
5547
5548         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5549                                       btrfs_ino(inode), to_free, 0);
5550
5551         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5552                                 to_free);
5553 }
5554
5555 /**
5556  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5557  * @inode: inode we're writing to
5558  * @num_bytes: the number of bytes we want to allocate
5559  *
5560  * This will do the following things
5561  *
5562  * o reserve space in the data space info for num_bytes
5563  * o reserve space in the metadata space info based on number of outstanding
5564  *   extents and how much csums will be needed
5565  * o add to the inodes ->delalloc_bytes
5566  * o add it to the fs_info's delalloc inodes list.
5567  *
5568  * This will return 0 for success and -ENOSPC if there is no space left.
5569  */
5570 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5571 {
5572         int ret;
5573
5574         ret = btrfs_check_data_free_space(inode, num_bytes, num_bytes);
5575         if (ret)
5576                 return ret;
5577
5578         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5579         if (ret) {
5580                 btrfs_free_reserved_data_space(inode, num_bytes);
5581                 return ret;
5582         }
5583
5584         return 0;
5585 }
5586
5587 /**
5588  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5589  * @inode: inode we're releasing space for
5590  * @num_bytes: the number of bytes we want to free up
5591  *
5592  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5593  * called in the case that we don't need the metadata AND data reservations
5594  * anymore.  So if there is an error or we insert an inline extent.
5595  *
5596  * This function will release the metadata space that was not used and will
5597  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5598  * list if there are no delalloc bytes left.
5599  */
5600 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5601 {
5602         btrfs_delalloc_release_metadata(inode, num_bytes);
5603         btrfs_free_reserved_data_space(inode, num_bytes);
5604 }
5605
5606 static int update_block_group(struct btrfs_trans_handle *trans,
5607                               struct btrfs_root *root, u64 bytenr,
5608                               u64 num_bytes, int alloc)
5609 {
5610         struct btrfs_block_group_cache *cache = NULL;
5611         struct btrfs_fs_info *info = root->fs_info;
5612         u64 total = num_bytes;
5613         u64 old_val;
5614         u64 byte_in_group;
5615         int factor;
5616
5617         /* block accounting for super block */
5618         spin_lock(&info->delalloc_root_lock);
5619         old_val = btrfs_super_bytes_used(info->super_copy);
5620         if (alloc)
5621                 old_val += num_bytes;
5622         else
5623                 old_val -= num_bytes;
5624         btrfs_set_super_bytes_used(info->super_copy, old_val);
5625         spin_unlock(&info->delalloc_root_lock);
5626
5627         while (total) {
5628                 cache = btrfs_lookup_block_group(info, bytenr);
5629                 if (!cache)
5630                         return -ENOENT;
5631                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5632                                     BTRFS_BLOCK_GROUP_RAID1 |
5633                                     BTRFS_BLOCK_GROUP_RAID10))
5634                         factor = 2;
5635                 else
5636                         factor = 1;
5637                 /*
5638                  * If this block group has free space cache written out, we
5639                  * need to make sure to load it if we are removing space.  This
5640                  * is because we need the unpinning stage to actually add the
5641                  * space back to the block group, otherwise we will leak space.
5642                  */
5643                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5644                         cache_block_group(cache, 1);
5645
5646                 byte_in_group = bytenr - cache->key.objectid;
5647                 WARN_ON(byte_in_group > cache->key.offset);
5648
5649                 spin_lock(&cache->space_info->lock);
5650                 spin_lock(&cache->lock);
5651
5652                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5653                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5654                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5655
5656                 old_val = btrfs_block_group_used(&cache->item);
5657                 num_bytes = min(total, cache->key.offset - byte_in_group);
5658                 if (alloc) {
5659                         old_val += num_bytes;
5660                         btrfs_set_block_group_used(&cache->item, old_val);
5661                         cache->reserved -= num_bytes;
5662                         cache->space_info->bytes_reserved -= num_bytes;
5663                         cache->space_info->bytes_used += num_bytes;
5664                         cache->space_info->disk_used += num_bytes * factor;
5665                         spin_unlock(&cache->lock);
5666                         spin_unlock(&cache->space_info->lock);
5667                 } else {
5668                         old_val -= num_bytes;
5669                         btrfs_set_block_group_used(&cache->item, old_val);
5670                         cache->pinned += num_bytes;
5671                         cache->space_info->bytes_pinned += num_bytes;
5672                         cache->space_info->bytes_used -= num_bytes;
5673                         cache->space_info->disk_used -= num_bytes * factor;
5674                         spin_unlock(&cache->lock);
5675                         spin_unlock(&cache->space_info->lock);
5676
5677                         set_extent_dirty(info->pinned_extents,
5678                                          bytenr, bytenr + num_bytes - 1,
5679                                          GFP_NOFS | __GFP_NOFAIL);
5680                         /*
5681                          * No longer have used bytes in this block group, queue
5682                          * it for deletion.
5683                          */
5684                         if (old_val == 0) {
5685                                 spin_lock(&info->unused_bgs_lock);
5686                                 if (list_empty(&cache->bg_list)) {
5687                                         btrfs_get_block_group(cache);
5688                                         list_add_tail(&cache->bg_list,
5689                                                       &info->unused_bgs);
5690                                 }
5691                                 spin_unlock(&info->unused_bgs_lock);
5692                         }
5693                 }
5694
5695                 spin_lock(&trans->transaction->dirty_bgs_lock);
5696                 if (list_empty(&cache->dirty_list)) {
5697                         list_add_tail(&cache->dirty_list,
5698                                       &trans->transaction->dirty_bgs);
5699                                 trans->transaction->num_dirty_bgs++;
5700                         btrfs_get_block_group(cache);
5701                 }
5702                 spin_unlock(&trans->transaction->dirty_bgs_lock);
5703
5704                 btrfs_put_block_group(cache);
5705                 total -= num_bytes;
5706                 bytenr += num_bytes;
5707         }
5708         return 0;
5709 }
5710
5711 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5712 {
5713         struct btrfs_block_group_cache *cache;
5714         u64 bytenr;
5715
5716         spin_lock(&root->fs_info->block_group_cache_lock);
5717         bytenr = root->fs_info->first_logical_byte;
5718         spin_unlock(&root->fs_info->block_group_cache_lock);
5719
5720         if (bytenr < (u64)-1)
5721                 return bytenr;
5722
5723         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5724         if (!cache)
5725                 return 0;
5726
5727         bytenr = cache->key.objectid;
5728         btrfs_put_block_group(cache);
5729
5730         return bytenr;
5731 }
5732
5733 static int pin_down_extent(struct btrfs_root *root,
5734                            struct btrfs_block_group_cache *cache,
5735                            u64 bytenr, u64 num_bytes, int reserved)
5736 {
5737         spin_lock(&cache->space_info->lock);
5738         spin_lock(&cache->lock);
5739         cache->pinned += num_bytes;
5740         cache->space_info->bytes_pinned += num_bytes;
5741         if (reserved) {
5742                 cache->reserved -= num_bytes;
5743                 cache->space_info->bytes_reserved -= num_bytes;
5744         }
5745         spin_unlock(&cache->lock);
5746         spin_unlock(&cache->space_info->lock);
5747
5748         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5749                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5750         if (reserved)
5751                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
5752         return 0;
5753 }
5754
5755 /*
5756  * this function must be called within transaction
5757  */
5758 int btrfs_pin_extent(struct btrfs_root *root,
5759                      u64 bytenr, u64 num_bytes, int reserved)
5760 {
5761         struct btrfs_block_group_cache *cache;
5762
5763         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5764         BUG_ON(!cache); /* Logic error */
5765
5766         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5767
5768         btrfs_put_block_group(cache);
5769         return 0;
5770 }
5771
5772 /*
5773  * this function must be called within transaction
5774  */
5775 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5776                                     u64 bytenr, u64 num_bytes)
5777 {
5778         struct btrfs_block_group_cache *cache;
5779         int ret;
5780
5781         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5782         if (!cache)
5783                 return -EINVAL;
5784
5785         /*
5786          * pull in the free space cache (if any) so that our pin
5787          * removes the free space from the cache.  We have load_only set
5788          * to one because the slow code to read in the free extents does check
5789          * the pinned extents.
5790          */
5791         cache_block_group(cache, 1);
5792
5793         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5794
5795         /* remove us from the free space cache (if we're there at all) */
5796         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5797         btrfs_put_block_group(cache);
5798         return ret;
5799 }
5800
5801 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5802 {
5803         int ret;
5804         struct btrfs_block_group_cache *block_group;
5805         struct btrfs_caching_control *caching_ctl;
5806
5807         block_group = btrfs_lookup_block_group(root->fs_info, start);
5808         if (!block_group)
5809                 return -EINVAL;
5810
5811         cache_block_group(block_group, 0);
5812         caching_ctl = get_caching_control(block_group);
5813
5814         if (!caching_ctl) {
5815                 /* Logic error */
5816                 BUG_ON(!block_group_cache_done(block_group));
5817                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5818         } else {
5819                 mutex_lock(&caching_ctl->mutex);
5820
5821                 if (start >= caching_ctl->progress) {
5822                         ret = add_excluded_extent(root, start, num_bytes);
5823                 } else if (start + num_bytes <= caching_ctl->progress) {
5824                         ret = btrfs_remove_free_space(block_group,
5825                                                       start, num_bytes);
5826                 } else {
5827                         num_bytes = caching_ctl->progress - start;
5828                         ret = btrfs_remove_free_space(block_group,
5829                                                       start, num_bytes);
5830                         if (ret)
5831                                 goto out_lock;
5832
5833                         num_bytes = (start + num_bytes) -
5834                                 caching_ctl->progress;
5835                         start = caching_ctl->progress;
5836                         ret = add_excluded_extent(root, start, num_bytes);
5837                 }
5838 out_lock:
5839                 mutex_unlock(&caching_ctl->mutex);
5840                 put_caching_control(caching_ctl);
5841         }
5842         btrfs_put_block_group(block_group);
5843         return ret;
5844 }
5845
5846 int btrfs_exclude_logged_extents(struct btrfs_root *log,
5847                                  struct extent_buffer *eb)
5848 {
5849         struct btrfs_file_extent_item *item;
5850         struct btrfs_key key;
5851         int found_type;
5852         int i;
5853
5854         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
5855                 return 0;
5856
5857         for (i = 0; i < btrfs_header_nritems(eb); i++) {
5858                 btrfs_item_key_to_cpu(eb, &key, i);
5859                 if (key.type != BTRFS_EXTENT_DATA_KEY)
5860                         continue;
5861                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
5862                 found_type = btrfs_file_extent_type(eb, item);
5863                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
5864                         continue;
5865                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
5866                         continue;
5867                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
5868                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
5869                 __exclude_logged_extent(log, key.objectid, key.offset);
5870         }
5871
5872         return 0;
5873 }
5874
5875 /**
5876  * btrfs_update_reserved_bytes - update the block_group and space info counters
5877  * @cache:      The cache we are manipulating
5878  * @num_bytes:  The number of bytes in question
5879  * @reserve:    One of the reservation enums
5880  * @delalloc:   The blocks are allocated for the delalloc write
5881  *
5882  * This is called by the allocator when it reserves space, or by somebody who is
5883  * freeing space that was never actually used on disk.  For example if you
5884  * reserve some space for a new leaf in transaction A and before transaction A
5885  * commits you free that leaf, you call this with reserve set to 0 in order to
5886  * clear the reservation.
5887  *
5888  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5889  * ENOSPC accounting.  For data we handle the reservation through clearing the
5890  * delalloc bits in the io_tree.  We have to do this since we could end up
5891  * allocating less disk space for the amount of data we have reserved in the
5892  * case of compression.
5893  *
5894  * If this is a reservation and the block group has become read only we cannot
5895  * make the reservation and return -EAGAIN, otherwise this function always
5896  * succeeds.
5897  */
5898 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5899                                        u64 num_bytes, int reserve, int delalloc)
5900 {
5901         struct btrfs_space_info *space_info = cache->space_info;
5902         int ret = 0;
5903
5904         spin_lock(&space_info->lock);
5905         spin_lock(&cache->lock);
5906         if (reserve != RESERVE_FREE) {
5907                 if (cache->ro) {
5908                         ret = -EAGAIN;
5909                 } else {
5910                         cache->reserved += num_bytes;
5911                         space_info->bytes_reserved += num_bytes;
5912                         if (reserve == RESERVE_ALLOC) {
5913                                 trace_btrfs_space_reservation(cache->fs_info,
5914                                                 "space_info", space_info->flags,
5915                                                 num_bytes, 0);
5916                                 space_info->bytes_may_use -= num_bytes;
5917                         }
5918
5919                         if (delalloc)
5920                                 cache->delalloc_bytes += num_bytes;
5921                 }
5922         } else {
5923                 if (cache->ro)
5924                         space_info->bytes_readonly += num_bytes;
5925                 cache->reserved -= num_bytes;
5926                 space_info->bytes_reserved -= num_bytes;
5927
5928                 if (delalloc)
5929                         cache->delalloc_bytes -= num_bytes;
5930         }
5931         spin_unlock(&cache->lock);
5932         spin_unlock(&space_info->lock);
5933         return ret;
5934 }
5935
5936 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5937                                 struct btrfs_root *root)
5938 {
5939         struct btrfs_fs_info *fs_info = root->fs_info;
5940         struct btrfs_caching_control *next;
5941         struct btrfs_caching_control *caching_ctl;
5942         struct btrfs_block_group_cache *cache;
5943
5944         down_write(&fs_info->commit_root_sem);
5945
5946         list_for_each_entry_safe(caching_ctl, next,
5947                                  &fs_info->caching_block_groups, list) {
5948                 cache = caching_ctl->block_group;
5949                 if (block_group_cache_done(cache)) {
5950                         cache->last_byte_to_unpin = (u64)-1;
5951                         list_del_init(&caching_ctl->list);
5952                         put_caching_control(caching_ctl);
5953                 } else {
5954                         cache->last_byte_to_unpin = caching_ctl->progress;
5955                 }
5956         }
5957
5958         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5959                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5960         else
5961                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5962
5963         up_write(&fs_info->commit_root_sem);
5964
5965         update_global_block_rsv(fs_info);
5966 }
5967
5968 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
5969                               const bool return_free_space)
5970 {
5971         struct btrfs_fs_info *fs_info = root->fs_info;
5972         struct btrfs_block_group_cache *cache = NULL;
5973         struct btrfs_space_info *space_info;
5974         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5975         u64 len;
5976         bool readonly;
5977
5978         while (start <= end) {
5979                 readonly = false;
5980                 if (!cache ||
5981                     start >= cache->key.objectid + cache->key.offset) {
5982                         if (cache)
5983                                 btrfs_put_block_group(cache);
5984                         cache = btrfs_lookup_block_group(fs_info, start);
5985                         BUG_ON(!cache); /* Logic error */
5986                 }
5987
5988                 len = cache->key.objectid + cache->key.offset - start;
5989                 len = min(len, end + 1 - start);
5990
5991                 if (start < cache->last_byte_to_unpin) {
5992                         len = min(len, cache->last_byte_to_unpin - start);
5993                         if (return_free_space)
5994                                 btrfs_add_free_space(cache, start, len);
5995                 }
5996
5997                 start += len;
5998                 space_info = cache->space_info;
5999
6000                 spin_lock(&space_info->lock);
6001                 spin_lock(&cache->lock);
6002                 cache->pinned -= len;
6003                 space_info->bytes_pinned -= len;
6004                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6005                 if (cache->ro) {
6006                         space_info->bytes_readonly += len;
6007                         readonly = true;
6008                 }
6009                 spin_unlock(&cache->lock);
6010                 if (!readonly && global_rsv->space_info == space_info) {
6011                         spin_lock(&global_rsv->lock);
6012                         if (!global_rsv->full) {
6013                                 len = min(len, global_rsv->size -
6014                                           global_rsv->reserved);
6015                                 global_rsv->reserved += len;
6016                                 space_info->bytes_may_use += len;
6017                                 if (global_rsv->reserved >= global_rsv->size)
6018                                         global_rsv->full = 1;
6019                         }
6020                         spin_unlock(&global_rsv->lock);
6021                 }
6022                 spin_unlock(&space_info->lock);
6023         }
6024
6025         if (cache)
6026                 btrfs_put_block_group(cache);
6027         return 0;
6028 }
6029
6030 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6031                                struct btrfs_root *root)
6032 {
6033         struct btrfs_fs_info *fs_info = root->fs_info;
6034         struct extent_io_tree *unpin;
6035         u64 start;
6036         u64 end;
6037         int ret;
6038
6039         if (trans->aborted)
6040                 return 0;
6041
6042         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6043                 unpin = &fs_info->freed_extents[1];
6044         else
6045                 unpin = &fs_info->freed_extents[0];
6046
6047         while (1) {
6048                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6049                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6050                                             EXTENT_DIRTY, NULL);
6051                 if (ret) {
6052                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6053                         break;
6054                 }
6055
6056                 if (btrfs_test_opt(root, DISCARD))
6057                         ret = btrfs_discard_extent(root, start,
6058                                                    end + 1 - start, NULL);
6059
6060                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
6061                 unpin_extent_range(root, start, end, true);
6062                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6063                 cond_resched();
6064         }
6065
6066         return 0;
6067 }
6068
6069 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6070                              u64 owner, u64 root_objectid)
6071 {
6072         struct btrfs_space_info *space_info;
6073         u64 flags;
6074
6075         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6076                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6077                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6078                 else
6079                         flags = BTRFS_BLOCK_GROUP_METADATA;
6080         } else {
6081                 flags = BTRFS_BLOCK_GROUP_DATA;
6082         }
6083
6084         space_info = __find_space_info(fs_info, flags);
6085         BUG_ON(!space_info); /* Logic bug */
6086         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6087 }
6088
6089
6090 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6091                                 struct btrfs_root *root,
6092                                 struct btrfs_delayed_ref_node *node, u64 parent,
6093                                 u64 root_objectid, u64 owner_objectid,
6094                                 u64 owner_offset, int refs_to_drop,
6095                                 struct btrfs_delayed_extent_op *extent_op)
6096 {
6097         struct btrfs_key key;
6098         struct btrfs_path *path;
6099         struct btrfs_fs_info *info = root->fs_info;
6100         struct btrfs_root *extent_root = info->extent_root;
6101         struct extent_buffer *leaf;
6102         struct btrfs_extent_item *ei;
6103         struct btrfs_extent_inline_ref *iref;
6104         int ret;
6105         int is_data;
6106         int extent_slot = 0;
6107         int found_extent = 0;
6108         int num_to_del = 1;
6109         int no_quota = node->no_quota;
6110         u32 item_size;
6111         u64 refs;
6112         u64 bytenr = node->bytenr;
6113         u64 num_bytes = node->num_bytes;
6114         int last_ref = 0;
6115         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6116                                                  SKINNY_METADATA);
6117
6118         if (!info->quota_enabled || !is_fstree(root_objectid))
6119                 no_quota = 1;
6120
6121         path = btrfs_alloc_path();
6122         if (!path)
6123                 return -ENOMEM;
6124
6125         path->reada = 1;
6126         path->leave_spinning = 1;
6127
6128         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6129         BUG_ON(!is_data && refs_to_drop != 1);
6130
6131         if (is_data)
6132                 skinny_metadata = 0;
6133
6134         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6135                                     bytenr, num_bytes, parent,
6136                                     root_objectid, owner_objectid,
6137                                     owner_offset);
6138         if (ret == 0) {
6139                 extent_slot = path->slots[0];
6140                 while (extent_slot >= 0) {
6141                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6142                                               extent_slot);
6143                         if (key.objectid != bytenr)
6144                                 break;
6145                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6146                             key.offset == num_bytes) {
6147                                 found_extent = 1;
6148                                 break;
6149                         }
6150                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6151                             key.offset == owner_objectid) {
6152                                 found_extent = 1;
6153                                 break;
6154                         }
6155                         if (path->slots[0] - extent_slot > 5)
6156                                 break;
6157                         extent_slot--;
6158                 }
6159 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6160                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6161                 if (found_extent && item_size < sizeof(*ei))
6162                         found_extent = 0;
6163 #endif
6164                 if (!found_extent) {
6165                         BUG_ON(iref);
6166                         ret = remove_extent_backref(trans, extent_root, path,
6167                                                     NULL, refs_to_drop,
6168                                                     is_data, &last_ref);
6169                         if (ret) {
6170                                 btrfs_abort_transaction(trans, extent_root, ret);
6171                                 goto out;
6172                         }
6173                         btrfs_release_path(path);
6174                         path->leave_spinning = 1;
6175
6176                         key.objectid = bytenr;
6177                         key.type = BTRFS_EXTENT_ITEM_KEY;
6178                         key.offset = num_bytes;
6179
6180                         if (!is_data && skinny_metadata) {
6181                                 key.type = BTRFS_METADATA_ITEM_KEY;
6182                                 key.offset = owner_objectid;
6183                         }
6184
6185                         ret = btrfs_search_slot(trans, extent_root,
6186                                                 &key, path, -1, 1);
6187                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6188                                 /*
6189                                  * Couldn't find our skinny metadata item,
6190                                  * see if we have ye olde extent item.
6191                                  */
6192                                 path->slots[0]--;
6193                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6194                                                       path->slots[0]);
6195                                 if (key.objectid == bytenr &&
6196                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6197                                     key.offset == num_bytes)
6198                                         ret = 0;
6199                         }
6200
6201                         if (ret > 0 && skinny_metadata) {
6202                                 skinny_metadata = false;
6203                                 key.objectid = bytenr;
6204                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6205                                 key.offset = num_bytes;
6206                                 btrfs_release_path(path);
6207                                 ret = btrfs_search_slot(trans, extent_root,
6208                                                         &key, path, -1, 1);
6209                         }
6210
6211                         if (ret) {
6212                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6213                                         ret, bytenr);
6214                                 if (ret > 0)
6215                                         btrfs_print_leaf(extent_root,
6216                                                          path->nodes[0]);
6217                         }
6218                         if (ret < 0) {
6219                                 btrfs_abort_transaction(trans, extent_root, ret);
6220                                 goto out;
6221                         }
6222                         extent_slot = path->slots[0];
6223                 }
6224         } else if (WARN_ON(ret == -ENOENT)) {
6225                 btrfs_print_leaf(extent_root, path->nodes[0]);
6226                 btrfs_err(info,
6227                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6228                         bytenr, parent, root_objectid, owner_objectid,
6229                         owner_offset);
6230                 btrfs_abort_transaction(trans, extent_root, ret);
6231                 goto out;
6232         } else {
6233                 btrfs_abort_transaction(trans, extent_root, ret);
6234                 goto out;
6235         }
6236
6237         leaf = path->nodes[0];
6238         item_size = btrfs_item_size_nr(leaf, extent_slot);
6239 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6240         if (item_size < sizeof(*ei)) {
6241                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6242                 ret = convert_extent_item_v0(trans, extent_root, path,
6243                                              owner_objectid, 0);
6244                 if (ret < 0) {
6245                         btrfs_abort_transaction(trans, extent_root, ret);
6246                         goto out;
6247                 }
6248
6249                 btrfs_release_path(path);
6250                 path->leave_spinning = 1;
6251
6252                 key.objectid = bytenr;
6253                 key.type = BTRFS_EXTENT_ITEM_KEY;
6254                 key.offset = num_bytes;
6255
6256                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6257                                         -1, 1);
6258                 if (ret) {
6259                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6260                                 ret, bytenr);
6261                         btrfs_print_leaf(extent_root, path->nodes[0]);
6262                 }
6263                 if (ret < 0) {
6264                         btrfs_abort_transaction(trans, extent_root, ret);
6265                         goto out;
6266                 }
6267
6268                 extent_slot = path->slots[0];
6269                 leaf = path->nodes[0];
6270                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6271         }
6272 #endif
6273         BUG_ON(item_size < sizeof(*ei));
6274         ei = btrfs_item_ptr(leaf, extent_slot,
6275                             struct btrfs_extent_item);
6276         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6277             key.type == BTRFS_EXTENT_ITEM_KEY) {
6278                 struct btrfs_tree_block_info *bi;
6279                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6280                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6281                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6282         }
6283
6284         refs = btrfs_extent_refs(leaf, ei);
6285         if (refs < refs_to_drop) {
6286                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6287                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6288                 ret = -EINVAL;
6289                 btrfs_abort_transaction(trans, extent_root, ret);
6290                 goto out;
6291         }
6292         refs -= refs_to_drop;
6293
6294         if (refs > 0) {
6295                 if (extent_op)
6296                         __run_delayed_extent_op(extent_op, leaf, ei);
6297                 /*
6298                  * In the case of inline back ref, reference count will
6299                  * be updated by remove_extent_backref
6300                  */
6301                 if (iref) {
6302                         BUG_ON(!found_extent);
6303                 } else {
6304                         btrfs_set_extent_refs(leaf, ei, refs);
6305                         btrfs_mark_buffer_dirty(leaf);
6306                 }
6307                 if (found_extent) {
6308                         ret = remove_extent_backref(trans, extent_root, path,
6309                                                     iref, refs_to_drop,
6310                                                     is_data, &last_ref);
6311                         if (ret) {
6312                                 btrfs_abort_transaction(trans, extent_root, ret);
6313                                 goto out;
6314                         }
6315                 }
6316                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6317                                  root_objectid);
6318         } else {
6319                 if (found_extent) {
6320                         BUG_ON(is_data && refs_to_drop !=
6321                                extent_data_ref_count(root, path, iref));
6322                         if (iref) {
6323                                 BUG_ON(path->slots[0] != extent_slot);
6324                         } else {
6325                                 BUG_ON(path->slots[0] != extent_slot + 1);
6326                                 path->slots[0] = extent_slot;
6327                                 num_to_del = 2;
6328                         }
6329                 }
6330
6331                 last_ref = 1;
6332                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6333                                       num_to_del);
6334                 if (ret) {
6335                         btrfs_abort_transaction(trans, extent_root, ret);
6336                         goto out;
6337                 }
6338                 btrfs_release_path(path);
6339
6340                 if (is_data) {
6341                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6342                         if (ret) {
6343                                 btrfs_abort_transaction(trans, extent_root, ret);
6344                                 goto out;
6345                         }
6346                 }
6347
6348                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
6349                 if (ret) {
6350                         btrfs_abort_transaction(trans, extent_root, ret);
6351                         goto out;
6352                 }
6353         }
6354         btrfs_release_path(path);
6355
6356 out:
6357         btrfs_free_path(path);
6358         return ret;
6359 }
6360
6361 /*
6362  * when we free an block, it is possible (and likely) that we free the last
6363  * delayed ref for that extent as well.  This searches the delayed ref tree for
6364  * a given extent, and if there are no other delayed refs to be processed, it
6365  * removes it from the tree.
6366  */
6367 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6368                                       struct btrfs_root *root, u64 bytenr)
6369 {
6370         struct btrfs_delayed_ref_head *head;
6371         struct btrfs_delayed_ref_root *delayed_refs;
6372         int ret = 0;
6373
6374         delayed_refs = &trans->transaction->delayed_refs;
6375         spin_lock(&delayed_refs->lock);
6376         head = btrfs_find_delayed_ref_head(trans, bytenr);
6377         if (!head)
6378                 goto out_delayed_unlock;
6379
6380         spin_lock(&head->lock);
6381         if (!list_empty(&head->ref_list))
6382                 goto out;
6383
6384         if (head->extent_op) {
6385                 if (!head->must_insert_reserved)
6386                         goto out;
6387                 btrfs_free_delayed_extent_op(head->extent_op);
6388                 head->extent_op = NULL;
6389         }
6390
6391         /*
6392          * waiting for the lock here would deadlock.  If someone else has it
6393          * locked they are already in the process of dropping it anyway
6394          */
6395         if (!mutex_trylock(&head->mutex))
6396                 goto out;
6397
6398         /*
6399          * at this point we have a head with no other entries.  Go
6400          * ahead and process it.
6401          */
6402         head->node.in_tree = 0;
6403         rb_erase(&head->href_node, &delayed_refs->href_root);
6404
6405         atomic_dec(&delayed_refs->num_entries);
6406
6407         /*
6408          * we don't take a ref on the node because we're removing it from the
6409          * tree, so we just steal the ref the tree was holding.
6410          */
6411         delayed_refs->num_heads--;
6412         if (head->processing == 0)
6413                 delayed_refs->num_heads_ready--;
6414         head->processing = 0;
6415         spin_unlock(&head->lock);
6416         spin_unlock(&delayed_refs->lock);
6417
6418         BUG_ON(head->extent_op);
6419         if (head->must_insert_reserved)
6420                 ret = 1;
6421
6422         mutex_unlock(&head->mutex);
6423         btrfs_put_delayed_ref(&head->node);
6424         return ret;
6425 out:
6426         spin_unlock(&head->lock);
6427
6428 out_delayed_unlock:
6429         spin_unlock(&delayed_refs->lock);
6430         return 0;
6431 }
6432
6433 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6434                            struct btrfs_root *root,
6435                            struct extent_buffer *buf,
6436                            u64 parent, int last_ref)
6437 {
6438         int pin = 1;
6439         int ret;
6440
6441         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6442                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6443                                         buf->start, buf->len,
6444                                         parent, root->root_key.objectid,
6445                                         btrfs_header_level(buf),
6446                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
6447                 BUG_ON(ret); /* -ENOMEM */
6448         }
6449
6450         if (!last_ref)
6451                 return;
6452
6453         if (btrfs_header_generation(buf) == trans->transid) {
6454                 struct btrfs_block_group_cache *cache;
6455
6456                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6457                         ret = check_ref_cleanup(trans, root, buf->start);
6458                         if (!ret)
6459                                 goto out;
6460                 }
6461
6462                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6463
6464                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6465                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6466                         btrfs_put_block_group(cache);
6467                         goto out;
6468                 }
6469
6470                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6471
6472                 btrfs_add_free_space(cache, buf->start, buf->len);
6473                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
6474                 btrfs_put_block_group(cache);
6475                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6476                 pin = 0;
6477         }
6478 out:
6479         if (pin)
6480                 add_pinned_bytes(root->fs_info, buf->len,
6481                                  btrfs_header_level(buf),
6482                                  root->root_key.objectid);
6483
6484         /*
6485          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6486          * anymore.
6487          */
6488         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6489 }
6490
6491 /* Can return -ENOMEM */
6492 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6493                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6494                       u64 owner, u64 offset, int no_quota)
6495 {
6496         int ret;
6497         struct btrfs_fs_info *fs_info = root->fs_info;
6498
6499         if (btrfs_test_is_dummy_root(root))
6500                 return 0;
6501
6502         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6503
6504         /*
6505          * tree log blocks never actually go into the extent allocation
6506          * tree, just update pinning info and exit early.
6507          */
6508         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6509                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6510                 /* unlocks the pinned mutex */
6511                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6512                 ret = 0;
6513         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6514                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6515                                         num_bytes,
6516                                         parent, root_objectid, (int)owner,
6517                                         BTRFS_DROP_DELAYED_REF, NULL, no_quota);
6518         } else {
6519                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6520                                                 num_bytes,
6521                                                 parent, root_objectid, owner,
6522                                                 offset, BTRFS_DROP_DELAYED_REF,
6523                                                 NULL, no_quota);
6524         }
6525         return ret;
6526 }
6527
6528 /*
6529  * when we wait for progress in the block group caching, its because
6530  * our allocation attempt failed at least once.  So, we must sleep
6531  * and let some progress happen before we try again.
6532  *
6533  * This function will sleep at least once waiting for new free space to
6534  * show up, and then it will check the block group free space numbers
6535  * for our min num_bytes.  Another option is to have it go ahead
6536  * and look in the rbtree for a free extent of a given size, but this
6537  * is a good start.
6538  *
6539  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6540  * any of the information in this block group.
6541  */
6542 static noinline void
6543 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6544                                 u64 num_bytes)
6545 {
6546         struct btrfs_caching_control *caching_ctl;
6547
6548         caching_ctl = get_caching_control(cache);
6549         if (!caching_ctl)
6550                 return;
6551
6552         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6553                    (cache->free_space_ctl->free_space >= num_bytes));
6554
6555         put_caching_control(caching_ctl);
6556 }
6557
6558 static noinline int
6559 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6560 {
6561         struct btrfs_caching_control *caching_ctl;
6562         int ret = 0;
6563
6564         caching_ctl = get_caching_control(cache);
6565         if (!caching_ctl)
6566                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6567
6568         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6569         if (cache->cached == BTRFS_CACHE_ERROR)
6570                 ret = -EIO;
6571         put_caching_control(caching_ctl);
6572         return ret;
6573 }
6574
6575 int __get_raid_index(u64 flags)
6576 {
6577         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6578                 return BTRFS_RAID_RAID10;
6579         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6580                 return BTRFS_RAID_RAID1;
6581         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6582                 return BTRFS_RAID_DUP;
6583         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6584                 return BTRFS_RAID_RAID0;
6585         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6586                 return BTRFS_RAID_RAID5;
6587         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6588                 return BTRFS_RAID_RAID6;
6589
6590         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6591 }
6592
6593 int get_block_group_index(struct btrfs_block_group_cache *cache)
6594 {
6595         return __get_raid_index(cache->flags);
6596 }
6597
6598 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6599         [BTRFS_RAID_RAID10]     = "raid10",
6600         [BTRFS_RAID_RAID1]      = "raid1",
6601         [BTRFS_RAID_DUP]        = "dup",
6602         [BTRFS_RAID_RAID0]      = "raid0",
6603         [BTRFS_RAID_SINGLE]     = "single",
6604         [BTRFS_RAID_RAID5]      = "raid5",
6605         [BTRFS_RAID_RAID6]      = "raid6",
6606 };
6607
6608 static const char *get_raid_name(enum btrfs_raid_types type)
6609 {
6610         if (type >= BTRFS_NR_RAID_TYPES)
6611                 return NULL;
6612
6613         return btrfs_raid_type_names[type];
6614 }
6615
6616 enum btrfs_loop_type {
6617         LOOP_CACHING_NOWAIT = 0,
6618         LOOP_CACHING_WAIT = 1,
6619         LOOP_ALLOC_CHUNK = 2,
6620         LOOP_NO_EMPTY_SIZE = 3,
6621 };
6622
6623 static inline void
6624 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
6625                        int delalloc)
6626 {
6627         if (delalloc)
6628                 down_read(&cache->data_rwsem);
6629 }
6630
6631 static inline void
6632 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
6633                        int delalloc)
6634 {
6635         btrfs_get_block_group(cache);
6636         if (delalloc)
6637                 down_read(&cache->data_rwsem);
6638 }
6639
6640 static struct btrfs_block_group_cache *
6641 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
6642                    struct btrfs_free_cluster *cluster,
6643                    int delalloc)
6644 {
6645         struct btrfs_block_group_cache *used_bg;
6646         bool locked = false;
6647 again:
6648         spin_lock(&cluster->refill_lock);
6649         if (locked) {
6650                 if (used_bg == cluster->block_group)
6651                         return used_bg;
6652
6653                 up_read(&used_bg->data_rwsem);
6654                 btrfs_put_block_group(used_bg);
6655         }
6656
6657         used_bg = cluster->block_group;
6658         if (!used_bg)
6659                 return NULL;
6660
6661         if (used_bg == block_group)
6662                 return used_bg;
6663
6664         btrfs_get_block_group(used_bg);
6665
6666         if (!delalloc)
6667                 return used_bg;
6668
6669         if (down_read_trylock(&used_bg->data_rwsem))
6670                 return used_bg;
6671
6672         spin_unlock(&cluster->refill_lock);
6673         down_read(&used_bg->data_rwsem);
6674         locked = true;
6675         goto again;
6676 }
6677
6678 static inline void
6679 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
6680                          int delalloc)
6681 {
6682         if (delalloc)
6683                 up_read(&cache->data_rwsem);
6684         btrfs_put_block_group(cache);
6685 }
6686
6687 /*
6688  * walks the btree of allocated extents and find a hole of a given size.
6689  * The key ins is changed to record the hole:
6690  * ins->objectid == start position
6691  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6692  * ins->offset == the size of the hole.
6693  * Any available blocks before search_start are skipped.
6694  *
6695  * If there is no suitable free space, we will record the max size of
6696  * the free space extent currently.
6697  */
6698 static noinline int find_free_extent(struct btrfs_root *orig_root,
6699                                      u64 num_bytes, u64 empty_size,
6700                                      u64 hint_byte, struct btrfs_key *ins,
6701                                      u64 flags, int delalloc)
6702 {
6703         int ret = 0;
6704         struct btrfs_root *root = orig_root->fs_info->extent_root;
6705         struct btrfs_free_cluster *last_ptr = NULL;
6706         struct btrfs_block_group_cache *block_group = NULL;
6707         u64 search_start = 0;
6708         u64 max_extent_size = 0;
6709         int empty_cluster = 2 * 1024 * 1024;
6710         struct btrfs_space_info *space_info;
6711         int loop = 0;
6712         int index = __get_raid_index(flags);
6713         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6714                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6715         bool failed_cluster_refill = false;
6716         bool failed_alloc = false;
6717         bool use_cluster = true;
6718         bool have_caching_bg = false;
6719
6720         WARN_ON(num_bytes < root->sectorsize);
6721         ins->type = BTRFS_EXTENT_ITEM_KEY;
6722         ins->objectid = 0;
6723         ins->offset = 0;
6724
6725         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6726
6727         space_info = __find_space_info(root->fs_info, flags);
6728         if (!space_info) {
6729                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6730                 return -ENOSPC;
6731         }
6732
6733         /*
6734          * If the space info is for both data and metadata it means we have a
6735          * small filesystem and we can't use the clustering stuff.
6736          */
6737         if (btrfs_mixed_space_info(space_info))
6738                 use_cluster = false;
6739
6740         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6741                 last_ptr = &root->fs_info->meta_alloc_cluster;
6742                 if (!btrfs_test_opt(root, SSD))
6743                         empty_cluster = 64 * 1024;
6744         }
6745
6746         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6747             btrfs_test_opt(root, SSD)) {
6748                 last_ptr = &root->fs_info->data_alloc_cluster;
6749         }
6750
6751         if (last_ptr) {
6752                 spin_lock(&last_ptr->lock);
6753                 if (last_ptr->block_group)
6754                         hint_byte = last_ptr->window_start;
6755                 spin_unlock(&last_ptr->lock);
6756         }
6757
6758         search_start = max(search_start, first_logical_byte(root, 0));
6759         search_start = max(search_start, hint_byte);
6760
6761         if (!last_ptr)
6762                 empty_cluster = 0;
6763
6764         if (search_start == hint_byte) {
6765                 block_group = btrfs_lookup_block_group(root->fs_info,
6766                                                        search_start);
6767                 /*
6768                  * we don't want to use the block group if it doesn't match our
6769                  * allocation bits, or if its not cached.
6770                  *
6771                  * However if we are re-searching with an ideal block group
6772                  * picked out then we don't care that the block group is cached.
6773                  */
6774                 if (block_group && block_group_bits(block_group, flags) &&
6775                     block_group->cached != BTRFS_CACHE_NO) {
6776                         down_read(&space_info->groups_sem);
6777                         if (list_empty(&block_group->list) ||
6778                             block_group->ro) {
6779                                 /*
6780                                  * someone is removing this block group,
6781                                  * we can't jump into the have_block_group
6782                                  * target because our list pointers are not
6783                                  * valid
6784                                  */
6785                                 btrfs_put_block_group(block_group);
6786                                 up_read(&space_info->groups_sem);
6787                         } else {
6788                                 index = get_block_group_index(block_group);
6789                                 btrfs_lock_block_group(block_group, delalloc);
6790                                 goto have_block_group;
6791                         }
6792                 } else if (block_group) {
6793                         btrfs_put_block_group(block_group);
6794                 }
6795         }
6796 search:
6797         have_caching_bg = false;
6798         down_read(&space_info->groups_sem);
6799         list_for_each_entry(block_group, &space_info->block_groups[index],
6800                             list) {
6801                 u64 offset;
6802                 int cached;
6803
6804                 btrfs_grab_block_group(block_group, delalloc);
6805                 search_start = block_group->key.objectid;
6806
6807                 /*
6808                  * this can happen if we end up cycling through all the
6809                  * raid types, but we want to make sure we only allocate
6810                  * for the proper type.
6811                  */
6812                 if (!block_group_bits(block_group, flags)) {
6813                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6814                                 BTRFS_BLOCK_GROUP_RAID1 |
6815                                 BTRFS_BLOCK_GROUP_RAID5 |
6816                                 BTRFS_BLOCK_GROUP_RAID6 |
6817                                 BTRFS_BLOCK_GROUP_RAID10;
6818
6819                         /*
6820                          * if they asked for extra copies and this block group
6821                          * doesn't provide them, bail.  This does allow us to
6822                          * fill raid0 from raid1.
6823                          */
6824                         if ((flags & extra) && !(block_group->flags & extra))
6825                                 goto loop;
6826                 }
6827
6828 have_block_group:
6829                 cached = block_group_cache_done(block_group);
6830                 if (unlikely(!cached)) {
6831                         ret = cache_block_group(block_group, 0);
6832                         BUG_ON(ret < 0);
6833                         ret = 0;
6834                 }
6835
6836                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
6837                         goto loop;
6838                 if (unlikely(block_group->ro))
6839                         goto loop;
6840
6841                 /*
6842                  * Ok we want to try and use the cluster allocator, so
6843                  * lets look there
6844                  */
6845                 if (last_ptr) {
6846                         struct btrfs_block_group_cache *used_block_group;
6847                         unsigned long aligned_cluster;
6848                         /*
6849                          * the refill lock keeps out other
6850                          * people trying to start a new cluster
6851                          */
6852                         used_block_group = btrfs_lock_cluster(block_group,
6853                                                               last_ptr,
6854                                                               delalloc);
6855                         if (!used_block_group)
6856                                 goto refill_cluster;
6857
6858                         if (used_block_group != block_group &&
6859                             (used_block_group->ro ||
6860                              !block_group_bits(used_block_group, flags)))
6861                                 goto release_cluster;
6862
6863                         offset = btrfs_alloc_from_cluster(used_block_group,
6864                                                 last_ptr,
6865                                                 num_bytes,
6866                                                 used_block_group->key.objectid,
6867                                                 &max_extent_size);
6868                         if (offset) {
6869                                 /* we have a block, we're done */
6870                                 spin_unlock(&last_ptr->refill_lock);
6871                                 trace_btrfs_reserve_extent_cluster(root,
6872                                                 used_block_group,
6873                                                 search_start, num_bytes);
6874                                 if (used_block_group != block_group) {
6875                                         btrfs_release_block_group(block_group,
6876                                                                   delalloc);
6877                                         block_group = used_block_group;
6878                                 }
6879                                 goto checks;
6880                         }
6881
6882                         WARN_ON(last_ptr->block_group != used_block_group);
6883 release_cluster:
6884                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6885                          * set up a new clusters, so lets just skip it
6886                          * and let the allocator find whatever block
6887                          * it can find.  If we reach this point, we
6888                          * will have tried the cluster allocator
6889                          * plenty of times and not have found
6890                          * anything, so we are likely way too
6891                          * fragmented for the clustering stuff to find
6892                          * anything.
6893                          *
6894                          * However, if the cluster is taken from the
6895                          * current block group, release the cluster
6896                          * first, so that we stand a better chance of
6897                          * succeeding in the unclustered
6898                          * allocation.  */
6899                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6900                             used_block_group != block_group) {
6901                                 spin_unlock(&last_ptr->refill_lock);
6902                                 btrfs_release_block_group(used_block_group,
6903                                                           delalloc);
6904                                 goto unclustered_alloc;
6905                         }
6906
6907                         /*
6908                          * this cluster didn't work out, free it and
6909                          * start over
6910                          */
6911                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6912
6913                         if (used_block_group != block_group)
6914                                 btrfs_release_block_group(used_block_group,
6915                                                           delalloc);
6916 refill_cluster:
6917                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6918                                 spin_unlock(&last_ptr->refill_lock);
6919                                 goto unclustered_alloc;
6920                         }
6921
6922                         aligned_cluster = max_t(unsigned long,
6923                                                 empty_cluster + empty_size,
6924                                               block_group->full_stripe_len);
6925
6926                         /* allocate a cluster in this block group */
6927                         ret = btrfs_find_space_cluster(root, block_group,
6928                                                        last_ptr, search_start,
6929                                                        num_bytes,
6930                                                        aligned_cluster);
6931                         if (ret == 0) {
6932                                 /*
6933                                  * now pull our allocation out of this
6934                                  * cluster
6935                                  */
6936                                 offset = btrfs_alloc_from_cluster(block_group,
6937                                                         last_ptr,
6938                                                         num_bytes,
6939                                                         search_start,
6940                                                         &max_extent_size);
6941                                 if (offset) {
6942                                         /* we found one, proceed */
6943                                         spin_unlock(&last_ptr->refill_lock);
6944                                         trace_btrfs_reserve_extent_cluster(root,
6945                                                 block_group, search_start,
6946                                                 num_bytes);
6947                                         goto checks;
6948                                 }
6949                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6950                                    && !failed_cluster_refill) {
6951                                 spin_unlock(&last_ptr->refill_lock);
6952
6953                                 failed_cluster_refill = true;
6954                                 wait_block_group_cache_progress(block_group,
6955                                        num_bytes + empty_cluster + empty_size);
6956                                 goto have_block_group;
6957                         }
6958
6959                         /*
6960                          * at this point we either didn't find a cluster
6961                          * or we weren't able to allocate a block from our
6962                          * cluster.  Free the cluster we've been trying
6963                          * to use, and go to the next block group
6964                          */
6965                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6966                         spin_unlock(&last_ptr->refill_lock);
6967                         goto loop;
6968                 }
6969
6970 unclustered_alloc:
6971                 spin_lock(&block_group->free_space_ctl->tree_lock);
6972                 if (cached &&
6973                     block_group->free_space_ctl->free_space <
6974                     num_bytes + empty_cluster + empty_size) {
6975                         if (block_group->free_space_ctl->free_space >
6976                             max_extent_size)
6977                                 max_extent_size =
6978                                         block_group->free_space_ctl->free_space;
6979                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6980                         goto loop;
6981                 }
6982                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6983
6984                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6985                                                     num_bytes, empty_size,
6986                                                     &max_extent_size);
6987                 /*
6988                  * If we didn't find a chunk, and we haven't failed on this
6989                  * block group before, and this block group is in the middle of
6990                  * caching and we are ok with waiting, then go ahead and wait
6991                  * for progress to be made, and set failed_alloc to true.
6992                  *
6993                  * If failed_alloc is true then we've already waited on this
6994                  * block group once and should move on to the next block group.
6995                  */
6996                 if (!offset && !failed_alloc && !cached &&
6997                     loop > LOOP_CACHING_NOWAIT) {
6998                         wait_block_group_cache_progress(block_group,
6999                                                 num_bytes + empty_size);
7000                         failed_alloc = true;
7001                         goto have_block_group;
7002                 } else if (!offset) {
7003                         if (!cached)
7004                                 have_caching_bg = true;
7005                         goto loop;
7006                 }
7007 checks:
7008                 search_start = ALIGN(offset, root->stripesize);
7009
7010                 /* move on to the next group */
7011                 if (search_start + num_bytes >
7012                     block_group->key.objectid + block_group->key.offset) {
7013                         btrfs_add_free_space(block_group, offset, num_bytes);
7014                         goto loop;
7015                 }
7016
7017                 if (offset < search_start)
7018                         btrfs_add_free_space(block_group, offset,
7019                                              search_start - offset);
7020                 BUG_ON(offset > search_start);
7021
7022                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
7023                                                   alloc_type, delalloc);
7024                 if (ret == -EAGAIN) {
7025                         btrfs_add_free_space(block_group, offset, num_bytes);
7026                         goto loop;
7027                 }
7028
7029                 /* we are all good, lets return */
7030                 ins->objectid = search_start;
7031                 ins->offset = num_bytes;
7032
7033                 trace_btrfs_reserve_extent(orig_root, block_group,
7034                                            search_start, num_bytes);
7035                 btrfs_release_block_group(block_group, delalloc);
7036                 break;
7037 loop:
7038                 failed_cluster_refill = false;
7039                 failed_alloc = false;
7040                 BUG_ON(index != get_block_group_index(block_group));
7041                 btrfs_release_block_group(block_group, delalloc);
7042         }
7043         up_read(&space_info->groups_sem);
7044
7045         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7046                 goto search;
7047
7048         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7049                 goto search;
7050
7051         /*
7052          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7053          *                      caching kthreads as we move along
7054          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7055          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7056          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7057          *                      again
7058          */
7059         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7060                 index = 0;
7061                 loop++;
7062                 if (loop == LOOP_ALLOC_CHUNK) {
7063                         struct btrfs_trans_handle *trans;
7064                         int exist = 0;
7065
7066                         trans = current->journal_info;
7067                         if (trans)
7068                                 exist = 1;
7069                         else
7070                                 trans = btrfs_join_transaction(root);
7071
7072                         if (IS_ERR(trans)) {
7073                                 ret = PTR_ERR(trans);
7074                                 goto out;
7075                         }
7076
7077                         ret = do_chunk_alloc(trans, root, flags,
7078                                              CHUNK_ALLOC_FORCE);
7079                         /*
7080                          * Do not bail out on ENOSPC since we
7081                          * can do more things.
7082                          */
7083                         if (ret < 0 && ret != -ENOSPC)
7084                                 btrfs_abort_transaction(trans,
7085                                                         root, ret);
7086                         else
7087                                 ret = 0;
7088                         if (!exist)
7089                                 btrfs_end_transaction(trans, root);
7090                         if (ret)
7091                                 goto out;
7092                 }
7093
7094                 if (loop == LOOP_NO_EMPTY_SIZE) {
7095                         empty_size = 0;
7096                         empty_cluster = 0;
7097                 }
7098
7099                 goto search;
7100         } else if (!ins->objectid) {
7101                 ret = -ENOSPC;
7102         } else if (ins->objectid) {
7103                 ret = 0;
7104         }
7105 out:
7106         if (ret == -ENOSPC)
7107                 ins->offset = max_extent_size;
7108         return ret;
7109 }
7110
7111 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
7112                             int dump_block_groups)
7113 {
7114         struct btrfs_block_group_cache *cache;
7115         int index = 0;
7116
7117         spin_lock(&info->lock);
7118         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
7119                info->flags,
7120                info->total_bytes - info->bytes_used - info->bytes_pinned -
7121                info->bytes_reserved - info->bytes_readonly,
7122                (info->full) ? "" : "not ");
7123         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
7124                "reserved=%llu, may_use=%llu, readonly=%llu\n",
7125                info->total_bytes, info->bytes_used, info->bytes_pinned,
7126                info->bytes_reserved, info->bytes_may_use,
7127                info->bytes_readonly);
7128         spin_unlock(&info->lock);
7129
7130         if (!dump_block_groups)
7131                 return;
7132
7133         down_read(&info->groups_sem);
7134 again:
7135         list_for_each_entry(cache, &info->block_groups[index], list) {
7136                 spin_lock(&cache->lock);
7137                 printk(KERN_INFO "BTRFS: "
7138                            "block group %llu has %llu bytes, "
7139                            "%llu used %llu pinned %llu reserved %s\n",
7140                        cache->key.objectid, cache->key.offset,
7141                        btrfs_block_group_used(&cache->item), cache->pinned,
7142                        cache->reserved, cache->ro ? "[readonly]" : "");
7143                 btrfs_dump_free_space(cache, bytes);
7144                 spin_unlock(&cache->lock);
7145         }
7146         if (++index < BTRFS_NR_RAID_TYPES)
7147                 goto again;
7148         up_read(&info->groups_sem);
7149 }
7150
7151 int btrfs_reserve_extent(struct btrfs_root *root,
7152                          u64 num_bytes, u64 min_alloc_size,
7153                          u64 empty_size, u64 hint_byte,
7154                          struct btrfs_key *ins, int is_data, int delalloc)
7155 {
7156         bool final_tried = false;
7157         u64 flags;
7158         int ret;
7159
7160         flags = btrfs_get_alloc_profile(root, is_data);
7161 again:
7162         WARN_ON(num_bytes < root->sectorsize);
7163         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
7164                                flags, delalloc);
7165
7166         if (ret == -ENOSPC) {
7167                 if (!final_tried && ins->offset) {
7168                         num_bytes = min(num_bytes >> 1, ins->offset);
7169                         num_bytes = round_down(num_bytes, root->sectorsize);
7170                         num_bytes = max(num_bytes, min_alloc_size);
7171                         if (num_bytes == min_alloc_size)
7172                                 final_tried = true;
7173                         goto again;
7174                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7175                         struct btrfs_space_info *sinfo;
7176
7177                         sinfo = __find_space_info(root->fs_info, flags);
7178                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
7179                                 flags, num_bytes);
7180                         if (sinfo)
7181                                 dump_space_info(sinfo, num_bytes, 1);
7182                 }
7183         }
7184
7185         return ret;
7186 }
7187
7188 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7189                                         u64 start, u64 len,
7190                                         int pin, int delalloc)
7191 {
7192         struct btrfs_block_group_cache *cache;
7193         int ret = 0;
7194
7195         cache = btrfs_lookup_block_group(root->fs_info, start);
7196         if (!cache) {
7197                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
7198                         start);
7199                 return -ENOSPC;
7200         }
7201
7202         if (pin)
7203                 pin_down_extent(root, cache, start, len, 1);
7204         else {
7205                 if (btrfs_test_opt(root, DISCARD))
7206                         ret = btrfs_discard_extent(root, start, len, NULL);
7207                 btrfs_add_free_space(cache, start, len);
7208                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
7209         }
7210
7211         btrfs_put_block_group(cache);
7212
7213         trace_btrfs_reserved_extent_free(root, start, len);
7214
7215         return ret;
7216 }
7217
7218 int btrfs_free_reserved_extent(struct btrfs_root *root,
7219                                u64 start, u64 len, int delalloc)
7220 {
7221         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
7222 }
7223
7224 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
7225                                        u64 start, u64 len)
7226 {
7227         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
7228 }
7229
7230 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7231                                       struct btrfs_root *root,
7232                                       u64 parent, u64 root_objectid,
7233                                       u64 flags, u64 owner, u64 offset,
7234                                       struct btrfs_key *ins, int ref_mod)
7235 {
7236         int ret;
7237         struct btrfs_fs_info *fs_info = root->fs_info;
7238         struct btrfs_extent_item *extent_item;
7239         struct btrfs_extent_inline_ref *iref;
7240         struct btrfs_path *path;
7241         struct extent_buffer *leaf;
7242         int type;
7243         u32 size;
7244
7245         if (parent > 0)
7246                 type = BTRFS_SHARED_DATA_REF_KEY;
7247         else
7248                 type = BTRFS_EXTENT_DATA_REF_KEY;
7249
7250         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7251
7252         path = btrfs_alloc_path();
7253         if (!path)
7254                 return -ENOMEM;
7255
7256         path->leave_spinning = 1;
7257         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7258                                       ins, size);
7259         if (ret) {
7260                 btrfs_free_path(path);
7261                 return ret;
7262         }
7263
7264         leaf = path->nodes[0];
7265         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7266                                      struct btrfs_extent_item);
7267         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7268         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7269         btrfs_set_extent_flags(leaf, extent_item,
7270                                flags | BTRFS_EXTENT_FLAG_DATA);
7271
7272         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7273         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7274         if (parent > 0) {
7275                 struct btrfs_shared_data_ref *ref;
7276                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7277                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7278                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7279         } else {
7280                 struct btrfs_extent_data_ref *ref;
7281                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7282                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7283                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7284                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7285                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7286         }
7287
7288         btrfs_mark_buffer_dirty(path->nodes[0]);
7289         btrfs_free_path(path);
7290
7291         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
7292         if (ret) { /* -ENOENT, logic error */
7293                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7294                         ins->objectid, ins->offset);
7295                 BUG();
7296         }
7297         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
7298         return ret;
7299 }
7300
7301 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7302                                      struct btrfs_root *root,
7303                                      u64 parent, u64 root_objectid,
7304                                      u64 flags, struct btrfs_disk_key *key,
7305                                      int level, struct btrfs_key *ins,
7306                                      int no_quota)
7307 {
7308         int ret;
7309         struct btrfs_fs_info *fs_info = root->fs_info;
7310         struct btrfs_extent_item *extent_item;
7311         struct btrfs_tree_block_info *block_info;
7312         struct btrfs_extent_inline_ref *iref;
7313         struct btrfs_path *path;
7314         struct extent_buffer *leaf;
7315         u32 size = sizeof(*extent_item) + sizeof(*iref);
7316         u64 num_bytes = ins->offset;
7317         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7318                                                  SKINNY_METADATA);
7319
7320         if (!skinny_metadata)
7321                 size += sizeof(*block_info);
7322
7323         path = btrfs_alloc_path();
7324         if (!path) {
7325                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7326                                                    root->nodesize);
7327                 return -ENOMEM;
7328         }
7329
7330         path->leave_spinning = 1;
7331         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7332                                       ins, size);
7333         if (ret) {
7334                 btrfs_free_path(path);
7335                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7336                                                    root->nodesize);
7337                 return ret;
7338         }
7339
7340         leaf = path->nodes[0];
7341         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7342                                      struct btrfs_extent_item);
7343         btrfs_set_extent_refs(leaf, extent_item, 1);
7344         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7345         btrfs_set_extent_flags(leaf, extent_item,
7346                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7347
7348         if (skinny_metadata) {
7349                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7350                 num_bytes = root->nodesize;
7351         } else {
7352                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7353                 btrfs_set_tree_block_key(leaf, block_info, key);
7354                 btrfs_set_tree_block_level(leaf, block_info, level);
7355                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7356         }
7357
7358         if (parent > 0) {
7359                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7360                 btrfs_set_extent_inline_ref_type(leaf, iref,
7361                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7362                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7363         } else {
7364                 btrfs_set_extent_inline_ref_type(leaf, iref,
7365                                                  BTRFS_TREE_BLOCK_REF_KEY);
7366                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7367         }
7368
7369         btrfs_mark_buffer_dirty(leaf);
7370         btrfs_free_path(path);
7371
7372         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
7373                                  1);
7374         if (ret) { /* -ENOENT, logic error */
7375                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7376                         ins->objectid, ins->offset);
7377                 BUG();
7378         }
7379
7380         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
7381         return ret;
7382 }
7383
7384 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7385                                      struct btrfs_root *root,
7386                                      u64 root_objectid, u64 owner,
7387                                      u64 offset, struct btrfs_key *ins)
7388 {
7389         int ret;
7390
7391         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7392
7393         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7394                                          ins->offset, 0,
7395                                          root_objectid, owner, offset,
7396                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
7397         return ret;
7398 }
7399
7400 /*
7401  * this is used by the tree logging recovery code.  It records that
7402  * an extent has been allocated and makes sure to clear the free
7403  * space cache bits as well
7404  */
7405 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7406                                    struct btrfs_root *root,
7407                                    u64 root_objectid, u64 owner, u64 offset,
7408                                    struct btrfs_key *ins)
7409 {
7410         int ret;
7411         struct btrfs_block_group_cache *block_group;
7412
7413         /*
7414          * Mixed block groups will exclude before processing the log so we only
7415          * need to do the exlude dance if this fs isn't mixed.
7416          */
7417         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7418                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7419                 if (ret)
7420                         return ret;
7421         }
7422
7423         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
7424         if (!block_group)
7425                 return -EINVAL;
7426
7427         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
7428                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
7429         BUG_ON(ret); /* logic error */
7430         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
7431                                          0, owner, offset, ins, 1);
7432         btrfs_put_block_group(block_group);
7433         return ret;
7434 }
7435
7436 static struct extent_buffer *
7437 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7438                       u64 bytenr, int level)
7439 {
7440         struct extent_buffer *buf;
7441
7442         buf = btrfs_find_create_tree_block(root, bytenr);
7443         if (!buf)
7444                 return ERR_PTR(-ENOMEM);
7445         btrfs_set_header_generation(buf, trans->transid);
7446         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
7447         btrfs_tree_lock(buf);
7448         clean_tree_block(trans, root->fs_info, buf);
7449         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
7450
7451         btrfs_set_lock_blocking(buf);
7452         btrfs_set_buffer_uptodate(buf);
7453
7454         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
7455                 buf->log_index = root->log_transid % 2;
7456                 /*
7457                  * we allow two log transactions at a time, use different
7458                  * EXENT bit to differentiate dirty pages.
7459                  */
7460                 if (buf->log_index == 0)
7461                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7462                                         buf->start + buf->len - 1, GFP_NOFS);
7463                 else
7464                         set_extent_new(&root->dirty_log_pages, buf->start,
7465                                         buf->start + buf->len - 1, GFP_NOFS);
7466         } else {
7467                 buf->log_index = -1;
7468                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7469                          buf->start + buf->len - 1, GFP_NOFS);
7470         }
7471         trans->blocks_used++;
7472         /* this returns a buffer locked for blocking */
7473         return buf;
7474 }
7475
7476 static struct btrfs_block_rsv *
7477 use_block_rsv(struct btrfs_trans_handle *trans,
7478               struct btrfs_root *root, u32 blocksize)
7479 {
7480         struct btrfs_block_rsv *block_rsv;
7481         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7482         int ret;
7483         bool global_updated = false;
7484
7485         block_rsv = get_block_rsv(trans, root);
7486
7487         if (unlikely(block_rsv->size == 0))
7488                 goto try_reserve;
7489 again:
7490         ret = block_rsv_use_bytes(block_rsv, blocksize);
7491         if (!ret)
7492                 return block_rsv;
7493
7494         if (block_rsv->failfast)
7495                 return ERR_PTR(ret);
7496
7497         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7498                 global_updated = true;
7499                 update_global_block_rsv(root->fs_info);
7500                 goto again;
7501         }
7502
7503         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7504                 static DEFINE_RATELIMIT_STATE(_rs,
7505                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7506                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7507                 if (__ratelimit(&_rs))
7508                         WARN(1, KERN_DEBUG
7509                                 "BTRFS: block rsv returned %d\n", ret);
7510         }
7511 try_reserve:
7512         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7513                                      BTRFS_RESERVE_NO_FLUSH);
7514         if (!ret)
7515                 return block_rsv;
7516         /*
7517          * If we couldn't reserve metadata bytes try and use some from
7518          * the global reserve if its space type is the same as the global
7519          * reservation.
7520          */
7521         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7522             block_rsv->space_info == global_rsv->space_info) {
7523                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7524                 if (!ret)
7525                         return global_rsv;
7526         }
7527         return ERR_PTR(ret);
7528 }
7529
7530 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7531                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7532 {
7533         block_rsv_add_bytes(block_rsv, blocksize, 0);
7534         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7535 }
7536
7537 /*
7538  * finds a free extent and does all the dirty work required for allocation
7539  * returns the key for the extent through ins, and a tree buffer for
7540  * the first block of the extent through buf.
7541  *
7542  * returns the tree buffer or an ERR_PTR on error.
7543  */
7544 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
7545                                         struct btrfs_root *root,
7546                                         u64 parent, u64 root_objectid,
7547                                         struct btrfs_disk_key *key, int level,
7548                                         u64 hint, u64 empty_size)
7549 {
7550         struct btrfs_key ins;
7551         struct btrfs_block_rsv *block_rsv;
7552         struct extent_buffer *buf;
7553         struct btrfs_delayed_extent_op *extent_op;
7554         u64 flags = 0;
7555         int ret;
7556         u32 blocksize = root->nodesize;
7557         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7558                                                  SKINNY_METADATA);
7559
7560         if (btrfs_test_is_dummy_root(root)) {
7561                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
7562                                             level);
7563                 if (!IS_ERR(buf))
7564                         root->alloc_bytenr += blocksize;
7565                 return buf;
7566         }
7567
7568         block_rsv = use_block_rsv(trans, root, blocksize);
7569         if (IS_ERR(block_rsv))
7570                 return ERR_CAST(block_rsv);
7571
7572         ret = btrfs_reserve_extent(root, blocksize, blocksize,
7573                                    empty_size, hint, &ins, 0, 0);
7574         if (ret)
7575                 goto out_unuse;
7576
7577         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
7578         if (IS_ERR(buf)) {
7579                 ret = PTR_ERR(buf);
7580                 goto out_free_reserved;
7581         }
7582
7583         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
7584                 if (parent == 0)
7585                         parent = ins.objectid;
7586                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
7587         } else
7588                 BUG_ON(parent > 0);
7589
7590         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
7591                 extent_op = btrfs_alloc_delayed_extent_op();
7592                 if (!extent_op) {
7593                         ret = -ENOMEM;
7594                         goto out_free_buf;
7595                 }
7596                 if (key)
7597                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
7598                 else
7599                         memset(&extent_op->key, 0, sizeof(extent_op->key));
7600                 extent_op->flags_to_set = flags;
7601                 if (skinny_metadata)
7602                         extent_op->update_key = 0;
7603                 else
7604                         extent_op->update_key = 1;
7605                 extent_op->update_flags = 1;
7606                 extent_op->is_data = 0;
7607                 extent_op->level = level;
7608
7609                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7610                                                  ins.objectid, ins.offset,
7611                                                  parent, root_objectid, level,
7612                                                  BTRFS_ADD_DELAYED_EXTENT,
7613                                                  extent_op, 0);
7614                 if (ret)
7615                         goto out_free_delayed;
7616         }
7617         return buf;
7618
7619 out_free_delayed:
7620         btrfs_free_delayed_extent_op(extent_op);
7621 out_free_buf:
7622         free_extent_buffer(buf);
7623 out_free_reserved:
7624         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 0);
7625 out_unuse:
7626         unuse_block_rsv(root->fs_info, block_rsv, blocksize);
7627         return ERR_PTR(ret);
7628 }
7629
7630 struct walk_control {
7631         u64 refs[BTRFS_MAX_LEVEL];
7632         u64 flags[BTRFS_MAX_LEVEL];
7633         struct btrfs_key update_progress;
7634         int stage;
7635         int level;
7636         int shared_level;
7637         int update_ref;
7638         int keep_locks;
7639         int reada_slot;
7640         int reada_count;
7641         int for_reloc;
7642 };
7643
7644 #define DROP_REFERENCE  1
7645 #define UPDATE_BACKREF  2
7646
7647 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
7648                                      struct btrfs_root *root,
7649                                      struct walk_control *wc,
7650                                      struct btrfs_path *path)
7651 {
7652         u64 bytenr;
7653         u64 generation;
7654         u64 refs;
7655         u64 flags;
7656         u32 nritems;
7657         u32 blocksize;
7658         struct btrfs_key key;
7659         struct extent_buffer *eb;
7660         int ret;
7661         int slot;
7662         int nread = 0;
7663
7664         if (path->slots[wc->level] < wc->reada_slot) {
7665                 wc->reada_count = wc->reada_count * 2 / 3;
7666                 wc->reada_count = max(wc->reada_count, 2);
7667         } else {
7668                 wc->reada_count = wc->reada_count * 3 / 2;
7669                 wc->reada_count = min_t(int, wc->reada_count,
7670                                         BTRFS_NODEPTRS_PER_BLOCK(root));
7671         }
7672
7673         eb = path->nodes[wc->level];
7674         nritems = btrfs_header_nritems(eb);
7675         blocksize = root->nodesize;
7676
7677         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
7678                 if (nread >= wc->reada_count)
7679                         break;
7680
7681                 cond_resched();
7682                 bytenr = btrfs_node_blockptr(eb, slot);
7683                 generation = btrfs_node_ptr_generation(eb, slot);
7684
7685                 if (slot == path->slots[wc->level])
7686                         goto reada;
7687
7688                 if (wc->stage == UPDATE_BACKREF &&
7689                     generation <= root->root_key.offset)
7690                         continue;
7691
7692                 /* We don't lock the tree block, it's OK to be racy here */
7693                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7694                                                wc->level - 1, 1, &refs,
7695                                                &flags);
7696                 /* We don't care about errors in readahead. */
7697                 if (ret < 0)
7698                         continue;
7699                 BUG_ON(refs == 0);
7700
7701                 if (wc->stage == DROP_REFERENCE) {
7702                         if (refs == 1)
7703                                 goto reada;
7704
7705                         if (wc->level == 1 &&
7706                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7707                                 continue;
7708                         if (!wc->update_ref ||
7709                             generation <= root->root_key.offset)
7710                                 continue;
7711                         btrfs_node_key_to_cpu(eb, &key, slot);
7712                         ret = btrfs_comp_cpu_keys(&key,
7713                                                   &wc->update_progress);
7714                         if (ret < 0)
7715                                 continue;
7716                 } else {
7717                         if (wc->level == 1 &&
7718                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7719                                 continue;
7720                 }
7721 reada:
7722                 readahead_tree_block(root, bytenr);
7723                 nread++;
7724         }
7725         wc->reada_slot = slot;
7726 }
7727
7728 /*
7729  * TODO: Modify related function to add related node/leaf to dirty_extent_root,
7730  * for later qgroup accounting.
7731  *
7732  * Current, this function does nothing.
7733  */
7734 static int account_leaf_items(struct btrfs_trans_handle *trans,
7735                               struct btrfs_root *root,
7736                               struct extent_buffer *eb)
7737 {
7738         int nr = btrfs_header_nritems(eb);
7739         int i, extent_type;
7740         struct btrfs_key key;
7741         struct btrfs_file_extent_item *fi;
7742         u64 bytenr, num_bytes;
7743
7744         for (i = 0; i < nr; i++) {
7745                 btrfs_item_key_to_cpu(eb, &key, i);
7746
7747                 if (key.type != BTRFS_EXTENT_DATA_KEY)
7748                         continue;
7749
7750                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
7751                 /* filter out non qgroup-accountable extents  */
7752                 extent_type = btrfs_file_extent_type(eb, fi);
7753
7754                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
7755                         continue;
7756
7757                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
7758                 if (!bytenr)
7759                         continue;
7760
7761                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
7762         }
7763         return 0;
7764 }
7765
7766 /*
7767  * Walk up the tree from the bottom, freeing leaves and any interior
7768  * nodes which have had all slots visited. If a node (leaf or
7769  * interior) is freed, the node above it will have it's slot
7770  * incremented. The root node will never be freed.
7771  *
7772  * At the end of this function, we should have a path which has all
7773  * slots incremented to the next position for a search. If we need to
7774  * read a new node it will be NULL and the node above it will have the
7775  * correct slot selected for a later read.
7776  *
7777  * If we increment the root nodes slot counter past the number of
7778  * elements, 1 is returned to signal completion of the search.
7779  */
7780 static int adjust_slots_upwards(struct btrfs_root *root,
7781                                 struct btrfs_path *path, int root_level)
7782 {
7783         int level = 0;
7784         int nr, slot;
7785         struct extent_buffer *eb;
7786
7787         if (root_level == 0)
7788                 return 1;
7789
7790         while (level <= root_level) {
7791                 eb = path->nodes[level];
7792                 nr = btrfs_header_nritems(eb);
7793                 path->slots[level]++;
7794                 slot = path->slots[level];
7795                 if (slot >= nr || level == 0) {
7796                         /*
7797                          * Don't free the root -  we will detect this
7798                          * condition after our loop and return a
7799                          * positive value for caller to stop walking the tree.
7800                          */
7801                         if (level != root_level) {
7802                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7803                                 path->locks[level] = 0;
7804
7805                                 free_extent_buffer(eb);
7806                                 path->nodes[level] = NULL;
7807                                 path->slots[level] = 0;
7808                         }
7809                 } else {
7810                         /*
7811                          * We have a valid slot to walk back down
7812                          * from. Stop here so caller can process these
7813                          * new nodes.
7814                          */
7815                         break;
7816                 }
7817
7818                 level++;
7819         }
7820
7821         eb = path->nodes[root_level];
7822         if (path->slots[root_level] >= btrfs_header_nritems(eb))
7823                 return 1;
7824
7825         return 0;
7826 }
7827
7828 /*
7829  * root_eb is the subtree root and is locked before this function is called.
7830  * TODO: Modify this function to mark all (including complete shared node)
7831  * to dirty_extent_root to allow it get accounted in qgroup.
7832  */
7833 static int account_shared_subtree(struct btrfs_trans_handle *trans,
7834                                   struct btrfs_root *root,
7835                                   struct extent_buffer *root_eb,
7836                                   u64 root_gen,
7837                                   int root_level)
7838 {
7839         int ret = 0;
7840         int level;
7841         struct extent_buffer *eb = root_eb;
7842         struct btrfs_path *path = NULL;
7843
7844         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
7845         BUG_ON(root_eb == NULL);
7846
7847         if (!root->fs_info->quota_enabled)
7848                 return 0;
7849
7850         if (!extent_buffer_uptodate(root_eb)) {
7851                 ret = btrfs_read_buffer(root_eb, root_gen);
7852                 if (ret)
7853                         goto out;
7854         }
7855
7856         if (root_level == 0) {
7857                 ret = account_leaf_items(trans, root, root_eb);
7858                 goto out;
7859         }
7860
7861         path = btrfs_alloc_path();
7862         if (!path)
7863                 return -ENOMEM;
7864
7865         /*
7866          * Walk down the tree.  Missing extent blocks are filled in as
7867          * we go. Metadata is accounted every time we read a new
7868          * extent block.
7869          *
7870          * When we reach a leaf, we account for file extent items in it,
7871          * walk back up the tree (adjusting slot pointers as we go)
7872          * and restart the search process.
7873          */
7874         extent_buffer_get(root_eb); /* For path */
7875         path->nodes[root_level] = root_eb;
7876         path->slots[root_level] = 0;
7877         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
7878 walk_down:
7879         level = root_level;
7880         while (level >= 0) {
7881                 if (path->nodes[level] == NULL) {
7882                         int parent_slot;
7883                         u64 child_gen;
7884                         u64 child_bytenr;
7885
7886                         /* We need to get child blockptr/gen from
7887                          * parent before we can read it. */
7888                         eb = path->nodes[level + 1];
7889                         parent_slot = path->slots[level + 1];
7890                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
7891                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
7892
7893                         eb = read_tree_block(root, child_bytenr, child_gen);
7894                         if (IS_ERR(eb)) {
7895                                 ret = PTR_ERR(eb);
7896                                 goto out;
7897                         } else if (!extent_buffer_uptodate(eb)) {
7898                                 free_extent_buffer(eb);
7899                                 ret = -EIO;
7900                                 goto out;
7901                         }
7902
7903                         path->nodes[level] = eb;
7904                         path->slots[level] = 0;
7905
7906                         btrfs_tree_read_lock(eb);
7907                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
7908                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
7909                 }
7910
7911                 if (level == 0) {
7912                         ret = account_leaf_items(trans, root, path->nodes[level]);
7913                         if (ret)
7914                                 goto out;
7915
7916                         /* Nonzero return here means we completed our search */
7917                         ret = adjust_slots_upwards(root, path, root_level);
7918                         if (ret)
7919                                 break;
7920
7921                         /* Restart search with new slots */
7922                         goto walk_down;
7923                 }
7924
7925                 level--;
7926         }
7927
7928         ret = 0;
7929 out:
7930         btrfs_free_path(path);
7931
7932         return ret;
7933 }
7934
7935 /*
7936  * helper to process tree block while walking down the tree.
7937  *
7938  * when wc->stage == UPDATE_BACKREF, this function updates
7939  * back refs for pointers in the block.
7940  *
7941  * NOTE: return value 1 means we should stop walking down.
7942  */
7943 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
7944                                    struct btrfs_root *root,
7945                                    struct btrfs_path *path,
7946                                    struct walk_control *wc, int lookup_info)
7947 {
7948         int level = wc->level;
7949         struct extent_buffer *eb = path->nodes[level];
7950         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7951         int ret;
7952
7953         if (wc->stage == UPDATE_BACKREF &&
7954             btrfs_header_owner(eb) != root->root_key.objectid)
7955                 return 1;
7956
7957         /*
7958          * when reference count of tree block is 1, it won't increase
7959          * again. once full backref flag is set, we never clear it.
7960          */
7961         if (lookup_info &&
7962             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
7963              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
7964                 BUG_ON(!path->locks[level]);
7965                 ret = btrfs_lookup_extent_info(trans, root,
7966                                                eb->start, level, 1,
7967                                                &wc->refs[level],
7968                                                &wc->flags[level]);
7969                 BUG_ON(ret == -ENOMEM);
7970                 if (ret)
7971                         return ret;
7972                 BUG_ON(wc->refs[level] == 0);
7973         }
7974
7975         if (wc->stage == DROP_REFERENCE) {
7976                 if (wc->refs[level] > 1)
7977                         return 1;
7978
7979                 if (path->locks[level] && !wc->keep_locks) {
7980                         btrfs_tree_unlock_rw(eb, path->locks[level]);
7981                         path->locks[level] = 0;
7982                 }
7983                 return 0;
7984         }
7985
7986         /* wc->stage == UPDATE_BACKREF */
7987         if (!(wc->flags[level] & flag)) {
7988                 BUG_ON(!path->locks[level]);
7989                 ret = btrfs_inc_ref(trans, root, eb, 1);
7990                 BUG_ON(ret); /* -ENOMEM */
7991                 ret = btrfs_dec_ref(trans, root, eb, 0);
7992                 BUG_ON(ret); /* -ENOMEM */
7993                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
7994                                                   eb->len, flag,
7995                                                   btrfs_header_level(eb), 0);
7996                 BUG_ON(ret); /* -ENOMEM */
7997                 wc->flags[level] |= flag;
7998         }
7999
8000         /*
8001          * the block is shared by multiple trees, so it's not good to
8002          * keep the tree lock
8003          */
8004         if (path->locks[level] && level > 0) {
8005                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8006                 path->locks[level] = 0;
8007         }
8008         return 0;
8009 }
8010
8011 /*
8012  * helper to process tree block pointer.
8013  *
8014  * when wc->stage == DROP_REFERENCE, this function checks
8015  * reference count of the block pointed to. if the block
8016  * is shared and we need update back refs for the subtree
8017  * rooted at the block, this function changes wc->stage to
8018  * UPDATE_BACKREF. if the block is shared and there is no
8019  * need to update back, this function drops the reference
8020  * to the block.
8021  *
8022  * NOTE: return value 1 means we should stop walking down.
8023  */
8024 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8025                                  struct btrfs_root *root,
8026                                  struct btrfs_path *path,
8027                                  struct walk_control *wc, int *lookup_info)
8028 {
8029         u64 bytenr;
8030         u64 generation;
8031         u64 parent;
8032         u32 blocksize;
8033         struct btrfs_key key;
8034         struct extent_buffer *next;
8035         int level = wc->level;
8036         int reada = 0;
8037         int ret = 0;
8038         bool need_account = false;
8039
8040         generation = btrfs_node_ptr_generation(path->nodes[level],
8041                                                path->slots[level]);
8042         /*
8043          * if the lower level block was created before the snapshot
8044          * was created, we know there is no need to update back refs
8045          * for the subtree
8046          */
8047         if (wc->stage == UPDATE_BACKREF &&
8048             generation <= root->root_key.offset) {
8049                 *lookup_info = 1;
8050                 return 1;
8051         }
8052
8053         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8054         blocksize = root->nodesize;
8055
8056         next = btrfs_find_tree_block(root->fs_info, bytenr);
8057         if (!next) {
8058                 next = btrfs_find_create_tree_block(root, bytenr);
8059                 if (!next)
8060                         return -ENOMEM;
8061                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8062                                                level - 1);
8063                 reada = 1;
8064         }
8065         btrfs_tree_lock(next);
8066         btrfs_set_lock_blocking(next);
8067
8068         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8069                                        &wc->refs[level - 1],
8070                                        &wc->flags[level - 1]);
8071         if (ret < 0) {
8072                 btrfs_tree_unlock(next);
8073                 return ret;
8074         }
8075
8076         if (unlikely(wc->refs[level - 1] == 0)) {
8077                 btrfs_err(root->fs_info, "Missing references.");
8078                 BUG();
8079         }
8080         *lookup_info = 0;
8081
8082         if (wc->stage == DROP_REFERENCE) {
8083                 if (wc->refs[level - 1] > 1) {
8084                         need_account = true;
8085                         if (level == 1 &&
8086                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8087                                 goto skip;
8088
8089                         if (!wc->update_ref ||
8090                             generation <= root->root_key.offset)
8091                                 goto skip;
8092
8093                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8094                                               path->slots[level]);
8095                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8096                         if (ret < 0)
8097                                 goto skip;
8098
8099                         wc->stage = UPDATE_BACKREF;
8100                         wc->shared_level = level - 1;
8101                 }
8102         } else {
8103                 if (level == 1 &&
8104                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8105                         goto skip;
8106         }
8107
8108         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8109                 btrfs_tree_unlock(next);
8110                 free_extent_buffer(next);
8111                 next = NULL;
8112                 *lookup_info = 1;
8113         }
8114
8115         if (!next) {
8116                 if (reada && level == 1)
8117                         reada_walk_down(trans, root, wc, path);
8118                 next = read_tree_block(root, bytenr, generation);
8119                 if (IS_ERR(next)) {
8120                         return PTR_ERR(next);
8121                 } else if (!extent_buffer_uptodate(next)) {
8122                         free_extent_buffer(next);
8123                         return -EIO;
8124                 }
8125                 btrfs_tree_lock(next);
8126                 btrfs_set_lock_blocking(next);
8127         }
8128
8129         level--;
8130         BUG_ON(level != btrfs_header_level(next));
8131         path->nodes[level] = next;
8132         path->slots[level] = 0;
8133         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8134         wc->level = level;
8135         if (wc->level == 1)
8136                 wc->reada_slot = 0;
8137         return 0;
8138 skip:
8139         wc->refs[level - 1] = 0;
8140         wc->flags[level - 1] = 0;
8141         if (wc->stage == DROP_REFERENCE) {
8142                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8143                         parent = path->nodes[level]->start;
8144                 } else {
8145                         BUG_ON(root->root_key.objectid !=
8146                                btrfs_header_owner(path->nodes[level]));
8147                         parent = 0;
8148                 }
8149
8150                 if (need_account) {
8151                         ret = account_shared_subtree(trans, root, next,
8152                                                      generation, level - 1);
8153                         if (ret) {
8154                                 printk_ratelimited(KERN_ERR "BTRFS: %s Error "
8155                                         "%d accounting shared subtree. Quota "
8156                                         "is out of sync, rescan required.\n",
8157                                         root->fs_info->sb->s_id, ret);
8158                         }
8159                 }
8160                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8161                                 root->root_key.objectid, level - 1, 0, 0);
8162                 BUG_ON(ret); /* -ENOMEM */
8163         }
8164         btrfs_tree_unlock(next);
8165         free_extent_buffer(next);
8166         *lookup_info = 1;
8167         return 1;
8168 }
8169
8170 /*
8171  * helper to process tree block while walking up the tree.
8172  *
8173  * when wc->stage == DROP_REFERENCE, this function drops
8174  * reference count on the block.
8175  *
8176  * when wc->stage == UPDATE_BACKREF, this function changes
8177  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8178  * to UPDATE_BACKREF previously while processing the block.
8179  *
8180  * NOTE: return value 1 means we should stop walking up.
8181  */
8182 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8183                                  struct btrfs_root *root,
8184                                  struct btrfs_path *path,
8185                                  struct walk_control *wc)
8186 {
8187         int ret;
8188         int level = wc->level;
8189         struct extent_buffer *eb = path->nodes[level];
8190         u64 parent = 0;
8191
8192         if (wc->stage == UPDATE_BACKREF) {
8193                 BUG_ON(wc->shared_level < level);
8194                 if (level < wc->shared_level)
8195                         goto out;
8196
8197                 ret = find_next_key(path, level + 1, &wc->update_progress);
8198                 if (ret > 0)
8199                         wc->update_ref = 0;
8200
8201                 wc->stage = DROP_REFERENCE;
8202                 wc->shared_level = -1;
8203                 path->slots[level] = 0;
8204
8205                 /*
8206                  * check reference count again if the block isn't locked.
8207                  * we should start walking down the tree again if reference
8208                  * count is one.
8209                  */
8210                 if (!path->locks[level]) {
8211                         BUG_ON(level == 0);
8212                         btrfs_tree_lock(eb);
8213                         btrfs_set_lock_blocking(eb);
8214                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8215
8216                         ret = btrfs_lookup_extent_info(trans, root,
8217                                                        eb->start, level, 1,
8218                                                        &wc->refs[level],
8219                                                        &wc->flags[level]);
8220                         if (ret < 0) {
8221                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8222                                 path->locks[level] = 0;
8223                                 return ret;
8224                         }
8225                         BUG_ON(wc->refs[level] == 0);
8226                         if (wc->refs[level] == 1) {
8227                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8228                                 path->locks[level] = 0;
8229                                 return 1;
8230                         }
8231                 }
8232         }
8233
8234         /* wc->stage == DROP_REFERENCE */
8235         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8236
8237         if (wc->refs[level] == 1) {
8238                 if (level == 0) {
8239                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8240                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8241                         else
8242                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8243                         BUG_ON(ret); /* -ENOMEM */
8244                         ret = account_leaf_items(trans, root, eb);
8245                         if (ret) {
8246                                 printk_ratelimited(KERN_ERR "BTRFS: %s Error "
8247                                         "%d accounting leaf items. Quota "
8248                                         "is out of sync, rescan required.\n",
8249                                         root->fs_info->sb->s_id, ret);
8250                         }
8251                 }
8252                 /* make block locked assertion in clean_tree_block happy */
8253                 if (!path->locks[level] &&
8254                     btrfs_header_generation(eb) == trans->transid) {
8255                         btrfs_tree_lock(eb);
8256                         btrfs_set_lock_blocking(eb);
8257                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8258                 }
8259                 clean_tree_block(trans, root->fs_info, eb);
8260         }
8261
8262         if (eb == root->node) {
8263                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8264                         parent = eb->start;
8265                 else
8266                         BUG_ON(root->root_key.objectid !=
8267                                btrfs_header_owner(eb));
8268         } else {
8269                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8270                         parent = path->nodes[level + 1]->start;
8271                 else
8272                         BUG_ON(root->root_key.objectid !=
8273                                btrfs_header_owner(path->nodes[level + 1]));
8274         }
8275
8276         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8277 out:
8278         wc->refs[level] = 0;
8279         wc->flags[level] = 0;
8280         return 0;
8281 }
8282
8283 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8284                                    struct btrfs_root *root,
8285                                    struct btrfs_path *path,
8286                                    struct walk_control *wc)
8287 {
8288         int level = wc->level;
8289         int lookup_info = 1;
8290         int ret;
8291
8292         while (level >= 0) {
8293                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8294                 if (ret > 0)
8295                         break;
8296
8297                 if (level == 0)
8298                         break;
8299
8300                 if (path->slots[level] >=
8301                     btrfs_header_nritems(path->nodes[level]))
8302                         break;
8303
8304                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8305                 if (ret > 0) {
8306                         path->slots[level]++;
8307                         continue;
8308                 } else if (ret < 0)
8309                         return ret;
8310                 level = wc->level;
8311         }
8312         return 0;
8313 }
8314
8315 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8316                                  struct btrfs_root *root,
8317                                  struct btrfs_path *path,
8318                                  struct walk_control *wc, int max_level)
8319 {
8320         int level = wc->level;
8321         int ret;
8322
8323         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8324         while (level < max_level && path->nodes[level]) {
8325                 wc->level = level;
8326                 if (path->slots[level] + 1 <
8327                     btrfs_header_nritems(path->nodes[level])) {
8328                         path->slots[level]++;
8329                         return 0;
8330                 } else {
8331                         ret = walk_up_proc(trans, root, path, wc);
8332                         if (ret > 0)
8333                                 return 0;
8334
8335                         if (path->locks[level]) {
8336                                 btrfs_tree_unlock_rw(path->nodes[level],
8337                                                      path->locks[level]);
8338                                 path->locks[level] = 0;
8339                         }
8340                         free_extent_buffer(path->nodes[level]);
8341                         path->nodes[level] = NULL;
8342                         level++;
8343                 }
8344         }
8345         return 1;
8346 }
8347
8348 /*
8349  * drop a subvolume tree.
8350  *
8351  * this function traverses the tree freeing any blocks that only
8352  * referenced by the tree.
8353  *
8354  * when a shared tree block is found. this function decreases its
8355  * reference count by one. if update_ref is true, this function
8356  * also make sure backrefs for the shared block and all lower level
8357  * blocks are properly updated.
8358  *
8359  * If called with for_reloc == 0, may exit early with -EAGAIN
8360  */
8361 int btrfs_drop_snapshot(struct btrfs_root *root,
8362                          struct btrfs_block_rsv *block_rsv, int update_ref,
8363                          int for_reloc)
8364 {
8365         struct btrfs_path *path;
8366         struct btrfs_trans_handle *trans;
8367         struct btrfs_root *tree_root = root->fs_info->tree_root;
8368         struct btrfs_root_item *root_item = &root->root_item;
8369         struct walk_control *wc;
8370         struct btrfs_key key;
8371         int err = 0;
8372         int ret;
8373         int level;
8374         bool root_dropped = false;
8375
8376         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
8377
8378         path = btrfs_alloc_path();
8379         if (!path) {
8380                 err = -ENOMEM;
8381                 goto out;
8382         }
8383
8384         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8385         if (!wc) {
8386                 btrfs_free_path(path);
8387                 err = -ENOMEM;
8388                 goto out;
8389         }
8390
8391         trans = btrfs_start_transaction(tree_root, 0);
8392         if (IS_ERR(trans)) {
8393                 err = PTR_ERR(trans);
8394                 goto out_free;
8395         }
8396
8397         if (block_rsv)
8398                 trans->block_rsv = block_rsv;
8399
8400         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8401                 level = btrfs_header_level(root->node);
8402                 path->nodes[level] = btrfs_lock_root_node(root);
8403                 btrfs_set_lock_blocking(path->nodes[level]);
8404                 path->slots[level] = 0;
8405                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8406                 memset(&wc->update_progress, 0,
8407                        sizeof(wc->update_progress));
8408         } else {
8409                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8410                 memcpy(&wc->update_progress, &key,
8411                        sizeof(wc->update_progress));
8412
8413                 level = root_item->drop_level;
8414                 BUG_ON(level == 0);
8415                 path->lowest_level = level;
8416                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8417                 path->lowest_level = 0;
8418                 if (ret < 0) {
8419                         err = ret;
8420                         goto out_end_trans;
8421                 }
8422                 WARN_ON(ret > 0);
8423
8424                 /*
8425                  * unlock our path, this is safe because only this
8426                  * function is allowed to delete this snapshot
8427                  */
8428                 btrfs_unlock_up_safe(path, 0);
8429
8430                 level = btrfs_header_level(root->node);
8431                 while (1) {
8432                         btrfs_tree_lock(path->nodes[level]);
8433                         btrfs_set_lock_blocking(path->nodes[level]);
8434                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8435
8436                         ret = btrfs_lookup_extent_info(trans, root,
8437                                                 path->nodes[level]->start,
8438                                                 level, 1, &wc->refs[level],
8439                                                 &wc->flags[level]);
8440                         if (ret < 0) {
8441                                 err = ret;
8442                                 goto out_end_trans;
8443                         }
8444                         BUG_ON(wc->refs[level] == 0);
8445
8446                         if (level == root_item->drop_level)
8447                                 break;
8448
8449                         btrfs_tree_unlock(path->nodes[level]);
8450                         path->locks[level] = 0;
8451                         WARN_ON(wc->refs[level] != 1);
8452                         level--;
8453                 }
8454         }
8455
8456         wc->level = level;
8457         wc->shared_level = -1;
8458         wc->stage = DROP_REFERENCE;
8459         wc->update_ref = update_ref;
8460         wc->keep_locks = 0;
8461         wc->for_reloc = for_reloc;
8462         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8463
8464         while (1) {
8465
8466                 ret = walk_down_tree(trans, root, path, wc);
8467                 if (ret < 0) {
8468                         err = ret;
8469                         break;
8470                 }
8471
8472                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8473                 if (ret < 0) {
8474                         err = ret;
8475                         break;
8476                 }
8477
8478                 if (ret > 0) {
8479                         BUG_ON(wc->stage != DROP_REFERENCE);
8480                         break;
8481                 }
8482
8483                 if (wc->stage == DROP_REFERENCE) {
8484                         level = wc->level;
8485                         btrfs_node_key(path->nodes[level],
8486                                        &root_item->drop_progress,
8487                                        path->slots[level]);
8488                         root_item->drop_level = level;
8489                 }
8490
8491                 BUG_ON(wc->level == 0);
8492                 if (btrfs_should_end_transaction(trans, tree_root) ||
8493                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
8494                         ret = btrfs_update_root(trans, tree_root,
8495                                                 &root->root_key,
8496                                                 root_item);
8497                         if (ret) {
8498                                 btrfs_abort_transaction(trans, tree_root, ret);
8499                                 err = ret;
8500                                 goto out_end_trans;
8501                         }
8502
8503                         btrfs_end_transaction_throttle(trans, tree_root);
8504                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
8505                                 pr_debug("BTRFS: drop snapshot early exit\n");
8506                                 err = -EAGAIN;
8507                                 goto out_free;
8508                         }
8509
8510                         trans = btrfs_start_transaction(tree_root, 0);
8511                         if (IS_ERR(trans)) {
8512                                 err = PTR_ERR(trans);
8513                                 goto out_free;
8514                         }
8515                         if (block_rsv)
8516                                 trans->block_rsv = block_rsv;
8517                 }
8518         }
8519         btrfs_release_path(path);
8520         if (err)
8521                 goto out_end_trans;
8522
8523         ret = btrfs_del_root(trans, tree_root, &root->root_key);
8524         if (ret) {
8525                 btrfs_abort_transaction(trans, tree_root, ret);
8526                 goto out_end_trans;
8527         }
8528
8529         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
8530                 ret = btrfs_find_root(tree_root, &root->root_key, path,
8531                                       NULL, NULL);
8532                 if (ret < 0) {
8533                         btrfs_abort_transaction(trans, tree_root, ret);
8534                         err = ret;
8535                         goto out_end_trans;
8536                 } else if (ret > 0) {
8537                         /* if we fail to delete the orphan item this time
8538                          * around, it'll get picked up the next time.
8539                          *
8540                          * The most common failure here is just -ENOENT.
8541                          */
8542                         btrfs_del_orphan_item(trans, tree_root,
8543                                               root->root_key.objectid);
8544                 }
8545         }
8546
8547         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
8548                 btrfs_drop_and_free_fs_root(tree_root->fs_info, root);
8549         } else {
8550                 free_extent_buffer(root->node);
8551                 free_extent_buffer(root->commit_root);
8552                 btrfs_put_fs_root(root);
8553         }
8554         root_dropped = true;
8555 out_end_trans:
8556         btrfs_end_transaction_throttle(trans, tree_root);
8557 out_free:
8558         kfree(wc);
8559         btrfs_free_path(path);
8560 out:
8561         /*
8562          * So if we need to stop dropping the snapshot for whatever reason we
8563          * need to make sure to add it back to the dead root list so that we
8564          * keep trying to do the work later.  This also cleans up roots if we
8565          * don't have it in the radix (like when we recover after a power fail
8566          * or unmount) so we don't leak memory.
8567          */
8568         if (!for_reloc && root_dropped == false)
8569                 btrfs_add_dead_root(root);
8570         if (err && err != -EAGAIN)
8571                 btrfs_std_error(root->fs_info, err);
8572         return err;
8573 }
8574
8575 /*
8576  * drop subtree rooted at tree block 'node'.
8577  *
8578  * NOTE: this function will unlock and release tree block 'node'
8579  * only used by relocation code
8580  */
8581 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
8582                         struct btrfs_root *root,
8583                         struct extent_buffer *node,
8584                         struct extent_buffer *parent)
8585 {
8586         struct btrfs_path *path;
8587         struct walk_control *wc;
8588         int level;
8589         int parent_level;
8590         int ret = 0;
8591         int wret;
8592
8593         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
8594
8595         path = btrfs_alloc_path();
8596         if (!path)
8597                 return -ENOMEM;
8598
8599         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8600         if (!wc) {
8601                 btrfs_free_path(path);
8602                 return -ENOMEM;
8603         }
8604
8605         btrfs_assert_tree_locked(parent);
8606         parent_level = btrfs_header_level(parent);
8607         extent_buffer_get(parent);
8608         path->nodes[parent_level] = parent;
8609         path->slots[parent_level] = btrfs_header_nritems(parent);
8610
8611         btrfs_assert_tree_locked(node);
8612         level = btrfs_header_level(node);
8613         path->nodes[level] = node;
8614         path->slots[level] = 0;
8615         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8616
8617         wc->refs[parent_level] = 1;
8618         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8619         wc->level = level;
8620         wc->shared_level = -1;
8621         wc->stage = DROP_REFERENCE;
8622         wc->update_ref = 0;
8623         wc->keep_locks = 1;
8624         wc->for_reloc = 1;
8625         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8626
8627         while (1) {
8628                 wret = walk_down_tree(trans, root, path, wc);
8629                 if (wret < 0) {
8630                         ret = wret;
8631                         break;
8632                 }
8633
8634                 wret = walk_up_tree(trans, root, path, wc, parent_level);
8635                 if (wret < 0)
8636                         ret = wret;
8637                 if (wret != 0)
8638                         break;
8639         }
8640
8641         kfree(wc);
8642         btrfs_free_path(path);
8643         return ret;
8644 }
8645
8646 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
8647 {
8648         u64 num_devices;
8649         u64 stripped;
8650
8651         /*
8652          * if restripe for this chunk_type is on pick target profile and
8653          * return, otherwise do the usual balance
8654          */
8655         stripped = get_restripe_target(root->fs_info, flags);
8656         if (stripped)
8657                 return extended_to_chunk(stripped);
8658
8659         num_devices = root->fs_info->fs_devices->rw_devices;
8660
8661         stripped = BTRFS_BLOCK_GROUP_RAID0 |
8662                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
8663                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
8664
8665         if (num_devices == 1) {
8666                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8667                 stripped = flags & ~stripped;
8668
8669                 /* turn raid0 into single device chunks */
8670                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
8671                         return stripped;
8672
8673                 /* turn mirroring into duplication */
8674                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
8675                              BTRFS_BLOCK_GROUP_RAID10))
8676                         return stripped | BTRFS_BLOCK_GROUP_DUP;
8677         } else {
8678                 /* they already had raid on here, just return */
8679                 if (flags & stripped)
8680                         return flags;
8681
8682                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8683                 stripped = flags & ~stripped;
8684
8685                 /* switch duplicated blocks with raid1 */
8686                 if (flags & BTRFS_BLOCK_GROUP_DUP)
8687                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
8688
8689                 /* this is drive concat, leave it alone */
8690         }
8691
8692         return flags;
8693 }
8694
8695 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
8696 {
8697         struct btrfs_space_info *sinfo = cache->space_info;
8698         u64 num_bytes;
8699         u64 min_allocable_bytes;
8700         int ret = -ENOSPC;
8701
8702
8703         /*
8704          * We need some metadata space and system metadata space for
8705          * allocating chunks in some corner cases until we force to set
8706          * it to be readonly.
8707          */
8708         if ((sinfo->flags &
8709              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
8710             !force)
8711                 min_allocable_bytes = 1 * 1024 * 1024;
8712         else
8713                 min_allocable_bytes = 0;
8714
8715         spin_lock(&sinfo->lock);
8716         spin_lock(&cache->lock);
8717
8718         if (cache->ro) {
8719                 ret = 0;
8720                 goto out;
8721         }
8722
8723         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8724                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8725
8726         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8727             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
8728             min_allocable_bytes <= sinfo->total_bytes) {
8729                 sinfo->bytes_readonly += num_bytes;
8730                 cache->ro = 1;
8731                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
8732                 ret = 0;
8733         }
8734 out:
8735         spin_unlock(&cache->lock);
8736         spin_unlock(&sinfo->lock);
8737         return ret;
8738 }
8739
8740 int btrfs_set_block_group_ro(struct btrfs_root *root,
8741                              struct btrfs_block_group_cache *cache)
8742
8743 {
8744         struct btrfs_trans_handle *trans;
8745         u64 alloc_flags;
8746         int ret;
8747
8748         BUG_ON(cache->ro);
8749
8750 again:
8751         trans = btrfs_join_transaction(root);
8752         if (IS_ERR(trans))
8753                 return PTR_ERR(trans);
8754
8755         /*
8756          * we're not allowed to set block groups readonly after the dirty
8757          * block groups cache has started writing.  If it already started,
8758          * back off and let this transaction commit
8759          */
8760         mutex_lock(&root->fs_info->ro_block_group_mutex);
8761         if (trans->transaction->dirty_bg_run) {
8762                 u64 transid = trans->transid;
8763
8764                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
8765                 btrfs_end_transaction(trans, root);
8766
8767                 ret = btrfs_wait_for_commit(root, transid);
8768                 if (ret)
8769                         return ret;
8770                 goto again;
8771         }
8772
8773         /*
8774          * if we are changing raid levels, try to allocate a corresponding
8775          * block group with the new raid level.
8776          */
8777         alloc_flags = update_block_group_flags(root, cache->flags);
8778         if (alloc_flags != cache->flags) {
8779                 ret = do_chunk_alloc(trans, root, alloc_flags,
8780                                      CHUNK_ALLOC_FORCE);
8781                 /*
8782                  * ENOSPC is allowed here, we may have enough space
8783                  * already allocated at the new raid level to
8784                  * carry on
8785                  */
8786                 if (ret == -ENOSPC)
8787                         ret = 0;
8788                 if (ret < 0)
8789                         goto out;
8790         }
8791
8792         ret = set_block_group_ro(cache, 0);
8793         if (!ret)
8794                 goto out;
8795         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
8796         ret = do_chunk_alloc(trans, root, alloc_flags,
8797                              CHUNK_ALLOC_FORCE);
8798         if (ret < 0)
8799                 goto out;
8800         ret = set_block_group_ro(cache, 0);
8801 out:
8802         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
8803                 alloc_flags = update_block_group_flags(root, cache->flags);
8804                 lock_chunks(root->fs_info->chunk_root);
8805                 check_system_chunk(trans, root, alloc_flags);
8806                 unlock_chunks(root->fs_info->chunk_root);
8807         }
8808         mutex_unlock(&root->fs_info->ro_block_group_mutex);
8809
8810         btrfs_end_transaction(trans, root);
8811         return ret;
8812 }
8813
8814 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
8815                             struct btrfs_root *root, u64 type)
8816 {
8817         u64 alloc_flags = get_alloc_profile(root, type);
8818         return do_chunk_alloc(trans, root, alloc_flags,
8819                               CHUNK_ALLOC_FORCE);
8820 }
8821
8822 /*
8823  * helper to account the unused space of all the readonly block group in the
8824  * space_info. takes mirrors into account.
8825  */
8826 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
8827 {
8828         struct btrfs_block_group_cache *block_group;
8829         u64 free_bytes = 0;
8830         int factor;
8831
8832         /* It's df, we don't care if it's racey */
8833         if (list_empty(&sinfo->ro_bgs))
8834                 return 0;
8835
8836         spin_lock(&sinfo->lock);
8837         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
8838                 spin_lock(&block_group->lock);
8839
8840                 if (!block_group->ro) {
8841                         spin_unlock(&block_group->lock);
8842                         continue;
8843                 }
8844
8845                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
8846                                           BTRFS_BLOCK_GROUP_RAID10 |
8847                                           BTRFS_BLOCK_GROUP_DUP))
8848                         factor = 2;
8849                 else
8850                         factor = 1;
8851
8852                 free_bytes += (block_group->key.offset -
8853                                btrfs_block_group_used(&block_group->item)) *
8854                                factor;
8855
8856                 spin_unlock(&block_group->lock);
8857         }
8858         spin_unlock(&sinfo->lock);
8859
8860         return free_bytes;
8861 }
8862
8863 void btrfs_set_block_group_rw(struct btrfs_root *root,
8864                               struct btrfs_block_group_cache *cache)
8865 {
8866         struct btrfs_space_info *sinfo = cache->space_info;
8867         u64 num_bytes;
8868
8869         BUG_ON(!cache->ro);
8870
8871         spin_lock(&sinfo->lock);
8872         spin_lock(&cache->lock);
8873         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8874                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8875         sinfo->bytes_readonly -= num_bytes;
8876         cache->ro = 0;
8877         list_del_init(&cache->ro_list);
8878         spin_unlock(&cache->lock);
8879         spin_unlock(&sinfo->lock);
8880 }
8881
8882 /*
8883  * checks to see if its even possible to relocate this block group.
8884  *
8885  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
8886  * ok to go ahead and try.
8887  */
8888 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
8889 {
8890         struct btrfs_block_group_cache *block_group;
8891         struct btrfs_space_info *space_info;
8892         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
8893         struct btrfs_device *device;
8894         struct btrfs_trans_handle *trans;
8895         u64 min_free;
8896         u64 dev_min = 1;
8897         u64 dev_nr = 0;
8898         u64 target;
8899         int index;
8900         int full = 0;
8901         int ret = 0;
8902
8903         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
8904
8905         /* odd, couldn't find the block group, leave it alone */
8906         if (!block_group)
8907                 return -1;
8908
8909         min_free = btrfs_block_group_used(&block_group->item);
8910
8911         /* no bytes used, we're good */
8912         if (!min_free)
8913                 goto out;
8914
8915         space_info = block_group->space_info;
8916         spin_lock(&space_info->lock);
8917
8918         full = space_info->full;
8919
8920         /*
8921          * if this is the last block group we have in this space, we can't
8922          * relocate it unless we're able to allocate a new chunk below.
8923          *
8924          * Otherwise, we need to make sure we have room in the space to handle
8925          * all of the extents from this block group.  If we can, we're good
8926          */
8927         if ((space_info->total_bytes != block_group->key.offset) &&
8928             (space_info->bytes_used + space_info->bytes_reserved +
8929              space_info->bytes_pinned + space_info->bytes_readonly +
8930              min_free < space_info->total_bytes)) {
8931                 spin_unlock(&space_info->lock);
8932                 goto out;
8933         }
8934         spin_unlock(&space_info->lock);
8935
8936         /*
8937          * ok we don't have enough space, but maybe we have free space on our
8938          * devices to allocate new chunks for relocation, so loop through our
8939          * alloc devices and guess if we have enough space.  if this block
8940          * group is going to be restriped, run checks against the target
8941          * profile instead of the current one.
8942          */
8943         ret = -1;
8944
8945         /*
8946          * index:
8947          *      0: raid10
8948          *      1: raid1
8949          *      2: dup
8950          *      3: raid0
8951          *      4: single
8952          */
8953         target = get_restripe_target(root->fs_info, block_group->flags);
8954         if (target) {
8955                 index = __get_raid_index(extended_to_chunk(target));
8956         } else {
8957                 /*
8958                  * this is just a balance, so if we were marked as full
8959                  * we know there is no space for a new chunk
8960                  */
8961                 if (full)
8962                         goto out;
8963
8964                 index = get_block_group_index(block_group);
8965         }
8966
8967         if (index == BTRFS_RAID_RAID10) {
8968                 dev_min = 4;
8969                 /* Divide by 2 */
8970                 min_free >>= 1;
8971         } else if (index == BTRFS_RAID_RAID1) {
8972                 dev_min = 2;
8973         } else if (index == BTRFS_RAID_DUP) {
8974                 /* Multiply by 2 */
8975                 min_free <<= 1;
8976         } else if (index == BTRFS_RAID_RAID0) {
8977                 dev_min = fs_devices->rw_devices;
8978                 min_free = div64_u64(min_free, dev_min);
8979         }
8980
8981         /* We need to do this so that we can look at pending chunks */
8982         trans = btrfs_join_transaction(root);
8983         if (IS_ERR(trans)) {
8984                 ret = PTR_ERR(trans);
8985                 goto out;
8986         }
8987
8988         mutex_lock(&root->fs_info->chunk_mutex);
8989         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
8990                 u64 dev_offset;
8991
8992                 /*
8993                  * check to make sure we can actually find a chunk with enough
8994                  * space to fit our block group in.
8995                  */
8996                 if (device->total_bytes > device->bytes_used + min_free &&
8997                     !device->is_tgtdev_for_dev_replace) {
8998                         ret = find_free_dev_extent(trans, device, min_free,
8999                                                    &dev_offset, NULL);
9000                         if (!ret)
9001                                 dev_nr++;
9002
9003                         if (dev_nr >= dev_min)
9004                                 break;
9005
9006                         ret = -1;
9007                 }
9008         }
9009         mutex_unlock(&root->fs_info->chunk_mutex);
9010         btrfs_end_transaction(trans, root);
9011 out:
9012         btrfs_put_block_group(block_group);
9013         return ret;
9014 }
9015
9016 static int find_first_block_group(struct btrfs_root *root,
9017                 struct btrfs_path *path, struct btrfs_key *key)
9018 {
9019         int ret = 0;
9020         struct btrfs_key found_key;
9021         struct extent_buffer *leaf;
9022         int slot;
9023
9024         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9025         if (ret < 0)
9026                 goto out;
9027
9028         while (1) {
9029                 slot = path->slots[0];
9030                 leaf = path->nodes[0];
9031                 if (slot >= btrfs_header_nritems(leaf)) {
9032                         ret = btrfs_next_leaf(root, path);
9033                         if (ret == 0)
9034                                 continue;
9035                         if (ret < 0)
9036                                 goto out;
9037                         break;
9038                 }
9039                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9040
9041                 if (found_key.objectid >= key->objectid &&
9042                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9043                         ret = 0;
9044                         goto out;
9045                 }
9046                 path->slots[0]++;
9047         }
9048 out:
9049         return ret;
9050 }
9051
9052 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9053 {
9054         struct btrfs_block_group_cache *block_group;
9055         u64 last = 0;
9056
9057         while (1) {
9058                 struct inode *inode;
9059
9060                 block_group = btrfs_lookup_first_block_group(info, last);
9061                 while (block_group) {
9062                         spin_lock(&block_group->lock);
9063                         if (block_group->iref)
9064                                 break;
9065                         spin_unlock(&block_group->lock);
9066                         block_group = next_block_group(info->tree_root,
9067                                                        block_group);
9068                 }
9069                 if (!block_group) {
9070                         if (last == 0)
9071                                 break;
9072                         last = 0;
9073                         continue;
9074                 }
9075
9076                 inode = block_group->inode;
9077                 block_group->iref = 0;
9078                 block_group->inode = NULL;
9079                 spin_unlock(&block_group->lock);
9080                 iput(inode);
9081                 last = block_group->key.objectid + block_group->key.offset;
9082                 btrfs_put_block_group(block_group);
9083         }
9084 }
9085
9086 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9087 {
9088         struct btrfs_block_group_cache *block_group;
9089         struct btrfs_space_info *space_info;
9090         struct btrfs_caching_control *caching_ctl;
9091         struct rb_node *n;
9092
9093         down_write(&info->commit_root_sem);
9094         while (!list_empty(&info->caching_block_groups)) {
9095                 caching_ctl = list_entry(info->caching_block_groups.next,
9096                                          struct btrfs_caching_control, list);
9097                 list_del(&caching_ctl->list);
9098                 put_caching_control(caching_ctl);
9099         }
9100         up_write(&info->commit_root_sem);
9101
9102         spin_lock(&info->unused_bgs_lock);
9103         while (!list_empty(&info->unused_bgs)) {
9104                 block_group = list_first_entry(&info->unused_bgs,
9105                                                struct btrfs_block_group_cache,
9106                                                bg_list);
9107                 list_del_init(&block_group->bg_list);
9108                 btrfs_put_block_group(block_group);
9109         }
9110         spin_unlock(&info->unused_bgs_lock);
9111
9112         spin_lock(&info->block_group_cache_lock);
9113         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9114                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9115                                        cache_node);
9116                 rb_erase(&block_group->cache_node,
9117                          &info->block_group_cache_tree);
9118                 RB_CLEAR_NODE(&block_group->cache_node);
9119                 spin_unlock(&info->block_group_cache_lock);
9120
9121                 down_write(&block_group->space_info->groups_sem);
9122                 list_del(&block_group->list);
9123                 up_write(&block_group->space_info->groups_sem);
9124
9125                 if (block_group->cached == BTRFS_CACHE_STARTED)
9126                         wait_block_group_cache_done(block_group);
9127
9128                 /*
9129                  * We haven't cached this block group, which means we could
9130                  * possibly have excluded extents on this block group.
9131                  */
9132                 if (block_group->cached == BTRFS_CACHE_NO ||
9133                     block_group->cached == BTRFS_CACHE_ERROR)
9134                         free_excluded_extents(info->extent_root, block_group);
9135
9136                 btrfs_remove_free_space_cache(block_group);
9137                 btrfs_put_block_group(block_group);
9138
9139                 spin_lock(&info->block_group_cache_lock);
9140         }
9141         spin_unlock(&info->block_group_cache_lock);
9142
9143         /* now that all the block groups are freed, go through and
9144          * free all the space_info structs.  This is only called during
9145          * the final stages of unmount, and so we know nobody is
9146          * using them.  We call synchronize_rcu() once before we start,
9147          * just to be on the safe side.
9148          */
9149         synchronize_rcu();
9150
9151         release_global_block_rsv(info);
9152
9153         while (!list_empty(&info->space_info)) {
9154                 int i;
9155
9156                 space_info = list_entry(info->space_info.next,
9157                                         struct btrfs_space_info,
9158                                         list);
9159                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
9160                         if (WARN_ON(space_info->bytes_pinned > 0 ||
9161                             space_info->bytes_reserved > 0 ||
9162                             space_info->bytes_may_use > 0)) {
9163                                 dump_space_info(space_info, 0, 0);
9164                         }
9165                 }
9166                 list_del(&space_info->list);
9167                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9168                         struct kobject *kobj;
9169                         kobj = space_info->block_group_kobjs[i];
9170                         space_info->block_group_kobjs[i] = NULL;
9171                         if (kobj) {
9172                                 kobject_del(kobj);
9173                                 kobject_put(kobj);
9174                         }
9175                 }
9176                 kobject_del(&space_info->kobj);
9177                 kobject_put(&space_info->kobj);
9178         }
9179         return 0;
9180 }
9181
9182 static void __link_block_group(struct btrfs_space_info *space_info,
9183                                struct btrfs_block_group_cache *cache)
9184 {
9185         int index = get_block_group_index(cache);
9186         bool first = false;
9187
9188         down_write(&space_info->groups_sem);
9189         if (list_empty(&space_info->block_groups[index]))
9190                 first = true;
9191         list_add_tail(&cache->list, &space_info->block_groups[index]);
9192         up_write(&space_info->groups_sem);
9193
9194         if (first) {
9195                 struct raid_kobject *rkobj;
9196                 int ret;
9197
9198                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9199                 if (!rkobj)
9200                         goto out_err;
9201                 rkobj->raid_type = index;
9202                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9203                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9204                                   "%s", get_raid_name(index));
9205                 if (ret) {
9206                         kobject_put(&rkobj->kobj);
9207                         goto out_err;
9208                 }
9209                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9210         }
9211
9212         return;
9213 out_err:
9214         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
9215 }
9216
9217 static struct btrfs_block_group_cache *
9218 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
9219 {
9220         struct btrfs_block_group_cache *cache;
9221
9222         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9223         if (!cache)
9224                 return NULL;
9225
9226         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9227                                         GFP_NOFS);
9228         if (!cache->free_space_ctl) {
9229                 kfree(cache);
9230                 return NULL;
9231         }
9232
9233         cache->key.objectid = start;
9234         cache->key.offset = size;
9235         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9236
9237         cache->sectorsize = root->sectorsize;
9238         cache->fs_info = root->fs_info;
9239         cache->full_stripe_len = btrfs_full_stripe_len(root,
9240                                                &root->fs_info->mapping_tree,
9241                                                start);
9242         atomic_set(&cache->count, 1);
9243         spin_lock_init(&cache->lock);
9244         init_rwsem(&cache->data_rwsem);
9245         INIT_LIST_HEAD(&cache->list);
9246         INIT_LIST_HEAD(&cache->cluster_list);
9247         INIT_LIST_HEAD(&cache->bg_list);
9248         INIT_LIST_HEAD(&cache->ro_list);
9249         INIT_LIST_HEAD(&cache->dirty_list);
9250         INIT_LIST_HEAD(&cache->io_list);
9251         btrfs_init_free_space_ctl(cache);
9252         atomic_set(&cache->trimming, 0);
9253
9254         return cache;
9255 }
9256
9257 int btrfs_read_block_groups(struct btrfs_root *root)
9258 {
9259         struct btrfs_path *path;
9260         int ret;
9261         struct btrfs_block_group_cache *cache;
9262         struct btrfs_fs_info *info = root->fs_info;
9263         struct btrfs_space_info *space_info;
9264         struct btrfs_key key;
9265         struct btrfs_key found_key;
9266         struct extent_buffer *leaf;
9267         int need_clear = 0;
9268         u64 cache_gen;
9269
9270         root = info->extent_root;
9271         key.objectid = 0;
9272         key.offset = 0;
9273         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9274         path = btrfs_alloc_path();
9275         if (!path)
9276                 return -ENOMEM;
9277         path->reada = 1;
9278
9279         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
9280         if (btrfs_test_opt(root, SPACE_CACHE) &&
9281             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
9282                 need_clear = 1;
9283         if (btrfs_test_opt(root, CLEAR_CACHE))
9284                 need_clear = 1;
9285
9286         while (1) {
9287                 ret = find_first_block_group(root, path, &key);
9288                 if (ret > 0)
9289                         break;
9290                 if (ret != 0)
9291                         goto error;
9292
9293                 leaf = path->nodes[0];
9294                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9295
9296                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
9297                                                        found_key.offset);
9298                 if (!cache) {
9299                         ret = -ENOMEM;
9300                         goto error;
9301                 }
9302
9303                 if (need_clear) {
9304                         /*
9305                          * When we mount with old space cache, we need to
9306                          * set BTRFS_DC_CLEAR and set dirty flag.
9307                          *
9308                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9309                          *    truncate the old free space cache inode and
9310                          *    setup a new one.
9311                          * b) Setting 'dirty flag' makes sure that we flush
9312                          *    the new space cache info onto disk.
9313                          */
9314                         if (btrfs_test_opt(root, SPACE_CACHE))
9315                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9316                 }
9317
9318                 read_extent_buffer(leaf, &cache->item,
9319                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9320                                    sizeof(cache->item));
9321                 cache->flags = btrfs_block_group_flags(&cache->item);
9322
9323                 key.objectid = found_key.objectid + found_key.offset;
9324                 btrfs_release_path(path);
9325
9326                 /*
9327                  * We need to exclude the super stripes now so that the space
9328                  * info has super bytes accounted for, otherwise we'll think
9329                  * we have more space than we actually do.
9330                  */
9331                 ret = exclude_super_stripes(root, cache);
9332                 if (ret) {
9333                         /*
9334                          * We may have excluded something, so call this just in
9335                          * case.
9336                          */
9337                         free_excluded_extents(root, cache);
9338                         btrfs_put_block_group(cache);
9339                         goto error;
9340                 }
9341
9342                 /*
9343                  * check for two cases, either we are full, and therefore
9344                  * don't need to bother with the caching work since we won't
9345                  * find any space, or we are empty, and we can just add all
9346                  * the space in and be done with it.  This saves us _alot_ of
9347                  * time, particularly in the full case.
9348                  */
9349                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
9350                         cache->last_byte_to_unpin = (u64)-1;
9351                         cache->cached = BTRFS_CACHE_FINISHED;
9352                         free_excluded_extents(root, cache);
9353                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9354                         cache->last_byte_to_unpin = (u64)-1;
9355                         cache->cached = BTRFS_CACHE_FINISHED;
9356                         add_new_free_space(cache, root->fs_info,
9357                                            found_key.objectid,
9358                                            found_key.objectid +
9359                                            found_key.offset);
9360                         free_excluded_extents(root, cache);
9361                 }
9362
9363                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
9364                 if (ret) {
9365                         btrfs_remove_free_space_cache(cache);
9366                         btrfs_put_block_group(cache);
9367                         goto error;
9368                 }
9369
9370                 ret = update_space_info(info, cache->flags, found_key.offset,
9371                                         btrfs_block_group_used(&cache->item),
9372                                         &space_info);
9373                 if (ret) {
9374                         btrfs_remove_free_space_cache(cache);
9375                         spin_lock(&info->block_group_cache_lock);
9376                         rb_erase(&cache->cache_node,
9377                                  &info->block_group_cache_tree);
9378                         RB_CLEAR_NODE(&cache->cache_node);
9379                         spin_unlock(&info->block_group_cache_lock);
9380                         btrfs_put_block_group(cache);
9381                         goto error;
9382                 }
9383
9384                 cache->space_info = space_info;
9385                 spin_lock(&cache->space_info->lock);
9386                 cache->space_info->bytes_readonly += cache->bytes_super;
9387                 spin_unlock(&cache->space_info->lock);
9388
9389                 __link_block_group(space_info, cache);
9390
9391                 set_avail_alloc_bits(root->fs_info, cache->flags);
9392                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
9393                         set_block_group_ro(cache, 1);
9394                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9395                         spin_lock(&info->unused_bgs_lock);
9396                         /* Should always be true but just in case. */
9397                         if (list_empty(&cache->bg_list)) {
9398                                 btrfs_get_block_group(cache);
9399                                 list_add_tail(&cache->bg_list,
9400                                               &info->unused_bgs);
9401                         }
9402                         spin_unlock(&info->unused_bgs_lock);
9403                 }
9404         }
9405
9406         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
9407                 if (!(get_alloc_profile(root, space_info->flags) &
9408                       (BTRFS_BLOCK_GROUP_RAID10 |
9409                        BTRFS_BLOCK_GROUP_RAID1 |
9410                        BTRFS_BLOCK_GROUP_RAID5 |
9411                        BTRFS_BLOCK_GROUP_RAID6 |
9412                        BTRFS_BLOCK_GROUP_DUP)))
9413                         continue;
9414                 /*
9415                  * avoid allocating from un-mirrored block group if there are
9416                  * mirrored block groups.
9417                  */
9418                 list_for_each_entry(cache,
9419                                 &space_info->block_groups[BTRFS_RAID_RAID0],
9420                                 list)
9421                         set_block_group_ro(cache, 1);
9422                 list_for_each_entry(cache,
9423                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
9424                                 list)
9425                         set_block_group_ro(cache, 1);
9426         }
9427
9428         init_global_block_rsv(info);
9429         ret = 0;
9430 error:
9431         btrfs_free_path(path);
9432         return ret;
9433 }
9434
9435 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
9436                                        struct btrfs_root *root)
9437 {
9438         struct btrfs_block_group_cache *block_group, *tmp;
9439         struct btrfs_root *extent_root = root->fs_info->extent_root;
9440         struct btrfs_block_group_item item;
9441         struct btrfs_key key;
9442         int ret = 0;
9443
9444         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
9445                 if (ret)
9446                         goto next;
9447
9448                 spin_lock(&block_group->lock);
9449                 memcpy(&item, &block_group->item, sizeof(item));
9450                 memcpy(&key, &block_group->key, sizeof(key));
9451                 spin_unlock(&block_group->lock);
9452
9453                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
9454                                         sizeof(item));
9455                 if (ret)
9456                         btrfs_abort_transaction(trans, extent_root, ret);
9457                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
9458                                                key.objectid, key.offset);
9459                 if (ret)
9460                         btrfs_abort_transaction(trans, extent_root, ret);
9461 next:
9462                 list_del_init(&block_group->bg_list);
9463         }
9464 }
9465
9466 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
9467                            struct btrfs_root *root, u64 bytes_used,
9468                            u64 type, u64 chunk_objectid, u64 chunk_offset,
9469                            u64 size)
9470 {
9471         int ret;
9472         struct btrfs_root *extent_root;
9473         struct btrfs_block_group_cache *cache;
9474
9475         extent_root = root->fs_info->extent_root;
9476
9477         btrfs_set_log_full_commit(root->fs_info, trans);
9478
9479         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
9480         if (!cache)
9481                 return -ENOMEM;
9482
9483         btrfs_set_block_group_used(&cache->item, bytes_used);
9484         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
9485         btrfs_set_block_group_flags(&cache->item, type);
9486
9487         cache->flags = type;
9488         cache->last_byte_to_unpin = (u64)-1;
9489         cache->cached = BTRFS_CACHE_FINISHED;
9490         ret = exclude_super_stripes(root, cache);
9491         if (ret) {
9492                 /*
9493                  * We may have excluded something, so call this just in
9494                  * case.
9495                  */
9496                 free_excluded_extents(root, cache);
9497                 btrfs_put_block_group(cache);
9498                 return ret;
9499         }
9500
9501         add_new_free_space(cache, root->fs_info, chunk_offset,
9502                            chunk_offset + size);
9503
9504         free_excluded_extents(root, cache);
9505
9506         /*
9507          * Call to ensure the corresponding space_info object is created and
9508          * assigned to our block group, but don't update its counters just yet.
9509          * We want our bg to be added to the rbtree with its ->space_info set.
9510          */
9511         ret = update_space_info(root->fs_info, cache->flags, 0, 0,
9512                                 &cache->space_info);
9513         if (ret) {
9514                 btrfs_remove_free_space_cache(cache);
9515                 btrfs_put_block_group(cache);
9516                 return ret;
9517         }
9518
9519         ret = btrfs_add_block_group_cache(root->fs_info, cache);
9520         if (ret) {
9521                 btrfs_remove_free_space_cache(cache);
9522                 btrfs_put_block_group(cache);
9523                 return ret;
9524         }
9525
9526         /*
9527          * Now that our block group has its ->space_info set and is inserted in
9528          * the rbtree, update the space info's counters.
9529          */
9530         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
9531                                 &cache->space_info);
9532         if (ret) {
9533                 btrfs_remove_free_space_cache(cache);
9534                 spin_lock(&root->fs_info->block_group_cache_lock);
9535                 rb_erase(&cache->cache_node,
9536                          &root->fs_info->block_group_cache_tree);
9537                 RB_CLEAR_NODE(&cache->cache_node);
9538                 spin_unlock(&root->fs_info->block_group_cache_lock);
9539                 btrfs_put_block_group(cache);
9540                 return ret;
9541         }
9542         update_global_block_rsv(root->fs_info);
9543
9544         spin_lock(&cache->space_info->lock);
9545         cache->space_info->bytes_readonly += cache->bytes_super;
9546         spin_unlock(&cache->space_info->lock);
9547
9548         __link_block_group(cache->space_info, cache);
9549
9550         list_add_tail(&cache->bg_list, &trans->new_bgs);
9551
9552         set_avail_alloc_bits(extent_root->fs_info, type);
9553
9554         return 0;
9555 }
9556
9557 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
9558 {
9559         u64 extra_flags = chunk_to_extended(flags) &
9560                                 BTRFS_EXTENDED_PROFILE_MASK;
9561
9562         write_seqlock(&fs_info->profiles_lock);
9563         if (flags & BTRFS_BLOCK_GROUP_DATA)
9564                 fs_info->avail_data_alloc_bits &= ~extra_flags;
9565         if (flags & BTRFS_BLOCK_GROUP_METADATA)
9566                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
9567         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
9568                 fs_info->avail_system_alloc_bits &= ~extra_flags;
9569         write_sequnlock(&fs_info->profiles_lock);
9570 }
9571
9572 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
9573                              struct btrfs_root *root, u64 group_start,
9574                              struct extent_map *em)
9575 {
9576         struct btrfs_path *path;
9577         struct btrfs_block_group_cache *block_group;
9578         struct btrfs_free_cluster *cluster;
9579         struct btrfs_root *tree_root = root->fs_info->tree_root;
9580         struct btrfs_key key;
9581         struct inode *inode;
9582         struct kobject *kobj = NULL;
9583         int ret;
9584         int index;
9585         int factor;
9586         struct btrfs_caching_control *caching_ctl = NULL;
9587         bool remove_em;
9588
9589         root = root->fs_info->extent_root;
9590
9591         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
9592         BUG_ON(!block_group);
9593         BUG_ON(!block_group->ro);
9594
9595         /*
9596          * Free the reserved super bytes from this block group before
9597          * remove it.
9598          */
9599         free_excluded_extents(root, block_group);
9600
9601         memcpy(&key, &block_group->key, sizeof(key));
9602         index = get_block_group_index(block_group);
9603         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
9604                                   BTRFS_BLOCK_GROUP_RAID1 |
9605                                   BTRFS_BLOCK_GROUP_RAID10))
9606                 factor = 2;
9607         else
9608                 factor = 1;
9609
9610         /* make sure this block group isn't part of an allocation cluster */
9611         cluster = &root->fs_info->data_alloc_cluster;
9612         spin_lock(&cluster->refill_lock);
9613         btrfs_return_cluster_to_free_space(block_group, cluster);
9614         spin_unlock(&cluster->refill_lock);
9615
9616         /*
9617          * make sure this block group isn't part of a metadata
9618          * allocation cluster
9619          */
9620         cluster = &root->fs_info->meta_alloc_cluster;
9621         spin_lock(&cluster->refill_lock);
9622         btrfs_return_cluster_to_free_space(block_group, cluster);
9623         spin_unlock(&cluster->refill_lock);
9624
9625         path = btrfs_alloc_path();
9626         if (!path) {
9627                 ret = -ENOMEM;
9628                 goto out;
9629         }
9630
9631         /*
9632          * get the inode first so any iput calls done for the io_list
9633          * aren't the final iput (no unlinks allowed now)
9634          */
9635         inode = lookup_free_space_inode(tree_root, block_group, path);
9636
9637         mutex_lock(&trans->transaction->cache_write_mutex);
9638         /*
9639          * make sure our free spache cache IO is done before remove the
9640          * free space inode
9641          */
9642         spin_lock(&trans->transaction->dirty_bgs_lock);
9643         if (!list_empty(&block_group->io_list)) {
9644                 list_del_init(&block_group->io_list);
9645
9646                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
9647
9648                 spin_unlock(&trans->transaction->dirty_bgs_lock);
9649                 btrfs_wait_cache_io(root, trans, block_group,
9650                                     &block_group->io_ctl, path,
9651                                     block_group->key.objectid);
9652                 btrfs_put_block_group(block_group);
9653                 spin_lock(&trans->transaction->dirty_bgs_lock);
9654         }
9655
9656         if (!list_empty(&block_group->dirty_list)) {
9657                 list_del_init(&block_group->dirty_list);
9658                 btrfs_put_block_group(block_group);
9659         }
9660         spin_unlock(&trans->transaction->dirty_bgs_lock);
9661         mutex_unlock(&trans->transaction->cache_write_mutex);
9662
9663         if (!IS_ERR(inode)) {
9664                 ret = btrfs_orphan_add(trans, inode);
9665                 if (ret) {
9666                         btrfs_add_delayed_iput(inode);
9667                         goto out;
9668                 }
9669                 clear_nlink(inode);
9670                 /* One for the block groups ref */
9671                 spin_lock(&block_group->lock);
9672                 if (block_group->iref) {
9673                         block_group->iref = 0;
9674                         block_group->inode = NULL;
9675                         spin_unlock(&block_group->lock);
9676                         iput(inode);
9677                 } else {
9678                         spin_unlock(&block_group->lock);
9679                 }
9680                 /* One for our lookup ref */
9681                 btrfs_add_delayed_iput(inode);
9682         }
9683
9684         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
9685         key.offset = block_group->key.objectid;
9686         key.type = 0;
9687
9688         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
9689         if (ret < 0)
9690                 goto out;
9691         if (ret > 0)
9692                 btrfs_release_path(path);
9693         if (ret == 0) {
9694                 ret = btrfs_del_item(trans, tree_root, path);
9695                 if (ret)
9696                         goto out;
9697                 btrfs_release_path(path);
9698         }
9699
9700         spin_lock(&root->fs_info->block_group_cache_lock);
9701         rb_erase(&block_group->cache_node,
9702                  &root->fs_info->block_group_cache_tree);
9703         RB_CLEAR_NODE(&block_group->cache_node);
9704
9705         if (root->fs_info->first_logical_byte == block_group->key.objectid)
9706                 root->fs_info->first_logical_byte = (u64)-1;
9707         spin_unlock(&root->fs_info->block_group_cache_lock);
9708
9709         down_write(&block_group->space_info->groups_sem);
9710         /*
9711          * we must use list_del_init so people can check to see if they
9712          * are still on the list after taking the semaphore
9713          */
9714         list_del_init(&block_group->list);
9715         if (list_empty(&block_group->space_info->block_groups[index])) {
9716                 kobj = block_group->space_info->block_group_kobjs[index];
9717                 block_group->space_info->block_group_kobjs[index] = NULL;
9718                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
9719         }
9720         up_write(&block_group->space_info->groups_sem);
9721         if (kobj) {
9722                 kobject_del(kobj);
9723                 kobject_put(kobj);
9724         }
9725
9726         if (block_group->has_caching_ctl)
9727                 caching_ctl = get_caching_control(block_group);
9728         if (block_group->cached == BTRFS_CACHE_STARTED)
9729                 wait_block_group_cache_done(block_group);
9730         if (block_group->has_caching_ctl) {
9731                 down_write(&root->fs_info->commit_root_sem);
9732                 if (!caching_ctl) {
9733                         struct btrfs_caching_control *ctl;
9734
9735                         list_for_each_entry(ctl,
9736                                     &root->fs_info->caching_block_groups, list)
9737                                 if (ctl->block_group == block_group) {
9738                                         caching_ctl = ctl;
9739                                         atomic_inc(&caching_ctl->count);
9740                                         break;
9741                                 }
9742                 }
9743                 if (caching_ctl)
9744                         list_del_init(&caching_ctl->list);
9745                 up_write(&root->fs_info->commit_root_sem);
9746                 if (caching_ctl) {
9747                         /* Once for the caching bgs list and once for us. */
9748                         put_caching_control(caching_ctl);
9749                         put_caching_control(caching_ctl);
9750                 }
9751         }
9752
9753         spin_lock(&trans->transaction->dirty_bgs_lock);
9754         if (!list_empty(&block_group->dirty_list)) {
9755                 WARN_ON(1);
9756         }
9757         if (!list_empty(&block_group->io_list)) {
9758                 WARN_ON(1);
9759         }
9760         spin_unlock(&trans->transaction->dirty_bgs_lock);
9761         btrfs_remove_free_space_cache(block_group);
9762
9763         spin_lock(&block_group->space_info->lock);
9764         list_del_init(&block_group->ro_list);
9765
9766         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
9767                 WARN_ON(block_group->space_info->total_bytes
9768                         < block_group->key.offset);
9769                 WARN_ON(block_group->space_info->bytes_readonly
9770                         < block_group->key.offset);
9771                 WARN_ON(block_group->space_info->disk_total
9772                         < block_group->key.offset * factor);
9773         }
9774         block_group->space_info->total_bytes -= block_group->key.offset;
9775         block_group->space_info->bytes_readonly -= block_group->key.offset;
9776         block_group->space_info->disk_total -= block_group->key.offset * factor;
9777
9778         spin_unlock(&block_group->space_info->lock);
9779
9780         memcpy(&key, &block_group->key, sizeof(key));
9781
9782         lock_chunks(root);
9783         if (!list_empty(&em->list)) {
9784                 /* We're in the transaction->pending_chunks list. */
9785                 free_extent_map(em);
9786         }
9787         spin_lock(&block_group->lock);
9788         block_group->removed = 1;
9789         /*
9790          * At this point trimming can't start on this block group, because we
9791          * removed the block group from the tree fs_info->block_group_cache_tree
9792          * so no one can't find it anymore and even if someone already got this
9793          * block group before we removed it from the rbtree, they have already
9794          * incremented block_group->trimming - if they didn't, they won't find
9795          * any free space entries because we already removed them all when we
9796          * called btrfs_remove_free_space_cache().
9797          *
9798          * And we must not remove the extent map from the fs_info->mapping_tree
9799          * to prevent the same logical address range and physical device space
9800          * ranges from being reused for a new block group. This is because our
9801          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
9802          * completely transactionless, so while it is trimming a range the
9803          * currently running transaction might finish and a new one start,
9804          * allowing for new block groups to be created that can reuse the same
9805          * physical device locations unless we take this special care.
9806          */
9807         remove_em = (atomic_read(&block_group->trimming) == 0);
9808         /*
9809          * Make sure a trimmer task always sees the em in the pinned_chunks list
9810          * if it sees block_group->removed == 1 (needs to lock block_group->lock
9811          * before checking block_group->removed).
9812          */
9813         if (!remove_em) {
9814                 /*
9815                  * Our em might be in trans->transaction->pending_chunks which
9816                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
9817                  * and so is the fs_info->pinned_chunks list.
9818                  *
9819                  * So at this point we must be holding the chunk_mutex to avoid
9820                  * any races with chunk allocation (more specifically at
9821                  * volumes.c:contains_pending_extent()), to ensure it always
9822                  * sees the em, either in the pending_chunks list or in the
9823                  * pinned_chunks list.
9824                  */
9825                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
9826         }
9827         spin_unlock(&block_group->lock);
9828
9829         if (remove_em) {
9830                 struct extent_map_tree *em_tree;
9831
9832                 em_tree = &root->fs_info->mapping_tree.map_tree;
9833                 write_lock(&em_tree->lock);
9834                 /*
9835                  * The em might be in the pending_chunks list, so make sure the
9836                  * chunk mutex is locked, since remove_extent_mapping() will
9837                  * delete us from that list.
9838                  */
9839                 remove_extent_mapping(em_tree, em);
9840                 write_unlock(&em_tree->lock);
9841                 /* once for the tree */
9842                 free_extent_map(em);
9843         }
9844
9845         unlock_chunks(root);
9846
9847         btrfs_put_block_group(block_group);
9848         btrfs_put_block_group(block_group);
9849
9850         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
9851         if (ret > 0)
9852                 ret = -EIO;
9853         if (ret < 0)
9854                 goto out;
9855
9856         ret = btrfs_del_item(trans, root, path);
9857 out:
9858         btrfs_free_path(path);
9859         return ret;
9860 }
9861
9862 /*
9863  * Process the unused_bgs list and remove any that don't have any allocated
9864  * space inside of them.
9865  */
9866 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
9867 {
9868         struct btrfs_block_group_cache *block_group;
9869         struct btrfs_space_info *space_info;
9870         struct btrfs_root *root = fs_info->extent_root;
9871         struct btrfs_trans_handle *trans;
9872         int ret = 0;
9873
9874         if (!fs_info->open)
9875                 return;
9876
9877         spin_lock(&fs_info->unused_bgs_lock);
9878         while (!list_empty(&fs_info->unused_bgs)) {
9879                 u64 start, end;
9880
9881                 block_group = list_first_entry(&fs_info->unused_bgs,
9882                                                struct btrfs_block_group_cache,
9883                                                bg_list);
9884                 space_info = block_group->space_info;
9885                 list_del_init(&block_group->bg_list);
9886                 if (ret || btrfs_mixed_space_info(space_info)) {
9887                         btrfs_put_block_group(block_group);
9888                         continue;
9889                 }
9890                 spin_unlock(&fs_info->unused_bgs_lock);
9891
9892                 mutex_lock(&root->fs_info->delete_unused_bgs_mutex);
9893
9894                 /* Don't want to race with allocators so take the groups_sem */
9895                 down_write(&space_info->groups_sem);
9896                 spin_lock(&block_group->lock);
9897                 if (block_group->reserved ||
9898                     btrfs_block_group_used(&block_group->item) ||
9899                     block_group->ro) {
9900                         /*
9901                          * We want to bail if we made new allocations or have
9902                          * outstanding allocations in this block group.  We do
9903                          * the ro check in case balance is currently acting on
9904                          * this block group.
9905                          */
9906                         spin_unlock(&block_group->lock);
9907                         up_write(&space_info->groups_sem);
9908                         goto next;
9909                 }
9910                 spin_unlock(&block_group->lock);
9911
9912                 /* We don't want to force the issue, only flip if it's ok. */
9913                 ret = set_block_group_ro(block_group, 0);
9914                 up_write(&space_info->groups_sem);
9915                 if (ret < 0) {
9916                         ret = 0;
9917                         goto next;
9918                 }
9919
9920                 /*
9921                  * Want to do this before we do anything else so we can recover
9922                  * properly if we fail to join the transaction.
9923                  */
9924                 /* 1 for btrfs_orphan_reserve_metadata() */
9925                 trans = btrfs_start_transaction(root, 1);
9926                 if (IS_ERR(trans)) {
9927                         btrfs_set_block_group_rw(root, block_group);
9928                         ret = PTR_ERR(trans);
9929                         goto next;
9930                 }
9931
9932                 /*
9933                  * We could have pending pinned extents for this block group,
9934                  * just delete them, we don't care about them anymore.
9935                  */
9936                 start = block_group->key.objectid;
9937                 end = start + block_group->key.offset - 1;
9938                 /*
9939                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
9940                  * btrfs_finish_extent_commit(). If we are at transaction N,
9941                  * another task might be running finish_extent_commit() for the
9942                  * previous transaction N - 1, and have seen a range belonging
9943                  * to the block group in freed_extents[] before we were able to
9944                  * clear the whole block group range from freed_extents[]. This
9945                  * means that task can lookup for the block group after we
9946                  * unpinned it from freed_extents[] and removed it, leading to
9947                  * a BUG_ON() at btrfs_unpin_extent_range().
9948                  */
9949                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
9950                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
9951                                   EXTENT_DIRTY, GFP_NOFS);
9952                 if (ret) {
9953                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9954                         btrfs_set_block_group_rw(root, block_group);
9955                         goto end_trans;
9956                 }
9957                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
9958                                   EXTENT_DIRTY, GFP_NOFS);
9959                 if (ret) {
9960                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9961                         btrfs_set_block_group_rw(root, block_group);
9962                         goto end_trans;
9963                 }
9964                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
9965
9966                 /* Reset pinned so btrfs_put_block_group doesn't complain */
9967                 spin_lock(&space_info->lock);
9968                 spin_lock(&block_group->lock);
9969
9970                 space_info->bytes_pinned -= block_group->pinned;
9971                 space_info->bytes_readonly += block_group->pinned;
9972                 percpu_counter_add(&space_info->total_bytes_pinned,
9973                                    -block_group->pinned);
9974                 block_group->pinned = 0;
9975
9976                 spin_unlock(&block_group->lock);
9977                 spin_unlock(&space_info->lock);
9978
9979                 /*
9980                  * Btrfs_remove_chunk will abort the transaction if things go
9981                  * horribly wrong.
9982                  */
9983                 ret = btrfs_remove_chunk(trans, root,
9984                                          block_group->key.objectid);
9985 end_trans:
9986                 btrfs_end_transaction(trans, root);
9987 next:
9988                 mutex_unlock(&root->fs_info->delete_unused_bgs_mutex);
9989                 btrfs_put_block_group(block_group);
9990                 spin_lock(&fs_info->unused_bgs_lock);
9991         }
9992         spin_unlock(&fs_info->unused_bgs_lock);
9993 }
9994
9995 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
9996 {
9997         struct btrfs_space_info *space_info;
9998         struct btrfs_super_block *disk_super;
9999         u64 features;
10000         u64 flags;
10001         int mixed = 0;
10002         int ret;
10003
10004         disk_super = fs_info->super_copy;
10005         if (!btrfs_super_root(disk_super))
10006                 return 1;
10007
10008         features = btrfs_super_incompat_flags(disk_super);
10009         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10010                 mixed = 1;
10011
10012         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10013         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10014         if (ret)
10015                 goto out;
10016
10017         if (mixed) {
10018                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10019                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10020         } else {
10021                 flags = BTRFS_BLOCK_GROUP_METADATA;
10022                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10023                 if (ret)
10024                         goto out;
10025
10026                 flags = BTRFS_BLOCK_GROUP_DATA;
10027                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10028         }
10029 out:
10030         return ret;
10031 }
10032
10033 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
10034 {
10035         return unpin_extent_range(root, start, end, false);
10036 }
10037
10038 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
10039 {
10040         struct btrfs_fs_info *fs_info = root->fs_info;
10041         struct btrfs_block_group_cache *cache = NULL;
10042         u64 group_trimmed;
10043         u64 start;
10044         u64 end;
10045         u64 trimmed = 0;
10046         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10047         int ret = 0;
10048
10049         /*
10050          * try to trim all FS space, our block group may start from non-zero.
10051          */
10052         if (range->len == total_bytes)
10053                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10054         else
10055                 cache = btrfs_lookup_block_group(fs_info, range->start);
10056
10057         while (cache) {
10058                 if (cache->key.objectid >= (range->start + range->len)) {
10059                         btrfs_put_block_group(cache);
10060                         break;
10061                 }
10062
10063                 start = max(range->start, cache->key.objectid);
10064                 end = min(range->start + range->len,
10065                                 cache->key.objectid + cache->key.offset);
10066
10067                 if (end - start >= range->minlen) {
10068                         if (!block_group_cache_done(cache)) {
10069                                 ret = cache_block_group(cache, 0);
10070                                 if (ret) {
10071                                         btrfs_put_block_group(cache);
10072                                         break;
10073                                 }
10074                                 ret = wait_block_group_cache_done(cache);
10075                                 if (ret) {
10076                                         btrfs_put_block_group(cache);
10077                                         break;
10078                                 }
10079                         }
10080                         ret = btrfs_trim_block_group(cache,
10081                                                      &group_trimmed,
10082                                                      start,
10083                                                      end,
10084                                                      range->minlen);
10085
10086                         trimmed += group_trimmed;
10087                         if (ret) {
10088                                 btrfs_put_block_group(cache);
10089                                 break;
10090                         }
10091                 }
10092
10093                 cache = next_block_group(fs_info->tree_root, cache);
10094         }
10095
10096         range->len = trimmed;
10097         return ret;
10098 }
10099
10100 /*
10101  * btrfs_{start,end}_write_no_snapshoting() are similar to
10102  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10103  * data into the page cache through nocow before the subvolume is snapshoted,
10104  * but flush the data into disk after the snapshot creation, or to prevent
10105  * operations while snapshoting is ongoing and that cause the snapshot to be
10106  * inconsistent (writes followed by expanding truncates for example).
10107  */
10108 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
10109 {
10110         percpu_counter_dec(&root->subv_writers->counter);
10111         /*
10112          * Make sure counter is updated before we wake up
10113          * waiters.
10114          */
10115         smp_mb();
10116         if (waitqueue_active(&root->subv_writers->wait))
10117                 wake_up(&root->subv_writers->wait);
10118 }
10119
10120 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
10121 {
10122         if (atomic_read(&root->will_be_snapshoted))
10123                 return 0;
10124
10125         percpu_counter_inc(&root->subv_writers->counter);
10126         /*
10127          * Make sure counter is updated before we check for snapshot creation.
10128          */
10129         smp_mb();
10130         if (atomic_read(&root->will_be_snapshoted)) {
10131                 btrfs_end_write_no_snapshoting(root);
10132                 return 0;
10133         }
10134         return 1;
10135 }