fs/btrfs/extent-tree.c

   1 /*
   2  * Copyright (C) 2007 Oracle.  All rights reserved.
   3  *
   4  * This program is free software; you can redistribute it and/or
   5  * modify it under the terms of the GNU General Public
   6  * License v2 as published by the Free Software Foundation.
   7  *
   8  * This program is distributed in the hope that it will be useful,
   9  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  10  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  11  * General Public License for more details.
  12  *
  13  * You should have received a copy of the GNU General Public
  14  * License along with this program; if not, write to the
  15  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
  16  * Boston, MA 021110-1307, USA.
  17  */
  18 #include <linux/sched.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/writeback.h>
  21 #include <linux/blkdev.h>
  22 #include <linux/sort.h>
  23 #include <linux/rcupdate.h>
  24 #include <linux/kthread.h>
  25 #include <linux/slab.h>
  26 #include <linux/ratelimit.h>
  27 #include <linux/percpu_counter.h>
  28 #include "hash.h"
  29 #include "tree-log.h"
  30 #include "disk-io.h"
  31 #include "print-tree.h"
  32 #include "volumes.h"
  33 #include "raid56.h"
  34 #include "locking.h"
  35 #include "free-space-cache.h"
  36 #include "math.h"
  37 #include "sysfs.h"
  38 #include "qgroup.h"
  39
  40 #undef SCRAMBLE_DELAYED_REFS
  41
  42 /*
  43  * control flags for do_chunk_alloc's force field
  44  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  45  * if we really need one.
  46  *
  47  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  48  * if we have very few chunks already allocated.  This is
  49  * used as part of the clustering code to help make sure
  50  * we have a good pool of storage to cluster in, without
  51  * filling the FS with empty chunks
  52  *
  53  * CHUNK_ALLOC_FORCE means it must try to allocate one
  54  *
  55  */
  56 enum {
  57         CHUNK_ALLOC_NO_FORCE = 0,
  58         CHUNK_ALLOC_LIMITED = 1,
  59         CHUNK_ALLOC_FORCE = 2,
  60 };
  61
  62 /*
  63  * Control how reservations are dealt with.
  64  *
  65  * RESERVE_FREE - freeing a reservation.
  66  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  67  *   ENOSPC accounting
  68  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  69  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  70  */
  71 enum {
  72         RESERVE_FREE = 0,
  73         RESERVE_ALLOC = 1,
  74         RESERVE_ALLOC_NO_ACCOUNT = 2,
  75 };
  76
  77 static int update_block_group(struct btrfs_root *root,
  78                               u64 bytenr, u64 num_bytes, int alloc);
  79 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  80                                 struct btrfs_root *root,
  81                                 u64 bytenr, u64 num_bytes, u64 parent,
  82                                 u64 root_objectid, u64 owner_objectid,
  83                                 u64 owner_offset, int refs_to_drop,
  84                                 struct btrfs_delayed_extent_op *extra_op,
  85                                 int no_quota);
  86 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  87                                     struct extent_buffer *leaf,
  88                                     struct btrfs_extent_item *ei);
  89 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  90                                       struct btrfs_root *root,
  91                                       u64 parent, u64 root_objectid,
  92                                       u64 flags, u64 owner, u64 offset,
  93                                       struct btrfs_key *ins, int ref_mod);
  94 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  95                                      struct btrfs_root *root,
  96                                      u64 parent, u64 root_objectid,
  97                                      u64 flags, struct btrfs_disk_key *key,
  98                                      int level, struct btrfs_key *ins,
  99                                      int no_quota);
 100 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
 101                           struct btrfs_root *extent_root, u64 flags,
 102                           int force);
 103 static int find_next_key(struct btrfs_path *path, int level,
 104                          struct btrfs_key *key);
 105 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 106                             int dump_block_groups);
 107 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 108                                        u64 num_bytes, int reserve,
 109                                        int delalloc);
 110 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 111                                u64 num_bytes);
 112 int btrfs_pin_extent(struct btrfs_root *root,
 113                      u64 bytenr, u64 num_bytes, int reserved);
 114
 115 static noinline int
 116 block_group_cache_done(struct btrfs_block_group_cache *cache)
 117 {
 118         smp_mb();
 119         return cache->cached == BTRFS_CACHE_FINISHED ||
 120                 cache->cached == BTRFS_CACHE_ERROR;
 121 }
 122
 123 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 124 {
 125         return (cache->flags & bits) == bits;
 126 }
 127
 128 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 129 {
 130         atomic_inc(&cache->count);
 131 }
 132
 133 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 134 {
 135         if (atomic_dec_and_test(&cache->count)) {
 136                 WARN_ON(cache->pinned > 0);
 137                 WARN_ON(cache->reserved > 0);
 138                 kfree(cache->free_space_ctl);
 139                 kfree(cache);
 140         }
 141 }
 142
 143 /*
 144  * this adds the block group to the fs_info rb tree for the block group
 145  * cache
 146  */
 147 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 148                                 struct btrfs_block_group_cache *block_group)
 149 {
 150         struct rb_node **p;
 151         struct rb_node *parent = NULL;
 152         struct btrfs_block_group_cache *cache;
 153
 154         spin_lock(&info->block_group_cache_lock);
 155         p = &info->block_group_cache_tree.rb_node;
 156
 157         while (*p) {
 158                 parent = *p;
 159                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 160                                  cache_node);
 161                 if (block_group->key.objectid < cache->key.objectid) {
 162                         p = &(*p)->rb_left;
 163                 } else if (block_group->key.objectid > cache->key.objectid) {
 164                         p = &(*p)->rb_right;
 165                 } else {
 166                         spin_unlock(&info->block_group_cache_lock);
 167                         return -EEXIST;
 168                 }
 169         }
 170
 171         rb_link_node(&block_group->cache_node, parent, p);
 172         rb_insert_color(&block_group->cache_node,
 173                         &info->block_group_cache_tree);
 174
 175         if (info->first_logical_byte > block_group->key.objectid)
 176                 info->first_logical_byte = block_group->key.objectid;
 177
 178         spin_unlock(&info->block_group_cache_lock);
 179
 180         return 0;
 181 }
 182
 183 /*
 184  * This will return the block group at or after bytenr if contains is 0, else
 185  * it will return the block group that contains the bytenr
 186  */
 187 static struct btrfs_block_group_cache *
 188 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 189                               int contains)
 190 {
 191         struct btrfs_block_group_cache *cache, *ret = NULL;
 192         struct rb_node *n;
 193         u64 end, start;
 194
 195         spin_lock(&info->block_group_cache_lock);
 196         n = info->block_group_cache_tree.rb_node;
 197
 198         while (n) {
 199                 cache = rb_entry(n, struct btrfs_block_group_cache,
 200                                  cache_node);
 201                 end = cache->key.objectid + cache->key.offset - 1;
 202                 start = cache->key.objectid;
 203
 204                 if (bytenr < start) {
 205                         if (!contains && (!ret || start < ret->key.objectid))
 206                                 ret = cache;
 207                         n = n->rb_left;
 208                 } else if (bytenr > start) {
 209                         if (contains && bytenr <= end) {
 210                                 ret = cache;
 211                                 break;
 212                         }
 213                         n = n->rb_right;
 214                 } else {
 215                         ret = cache;
 216                         break;
 217                 }
 218         }
 219         if (ret) {
 220                 btrfs_get_block_group(ret);
 221                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 222                         info->first_logical_byte = ret->key.objectid;
 223         }
 224         spin_unlock(&info->block_group_cache_lock);
 225
 226         return ret;
 227 }
 228
 229 static int add_excluded_extent(struct btrfs_root *root,
 230                                u64 start, u64 num_bytes)
 231 {
 232         u64 end = start + num_bytes - 1;
 233         set_extent_bits(&root->fs_info->freed_extents[0],
 234                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 235         set_extent_bits(&root->fs_info->freed_extents[1],
 236                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 237         return 0;
 238 }
 239
 240 static void free_excluded_extents(struct btrfs_root *root,
 241                                   struct btrfs_block_group_cache *cache)
 242 {
 243         u64 start, end;
 244
 245         start = cache->key.objectid;
 246         end = start + cache->key.offset - 1;
 247
 248         clear_extent_bits(&root->fs_info->freed_extents[0],
 249                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 250         clear_extent_bits(&root->fs_info->freed_extents[1],
 251                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 252 }
 253
 254 static int exclude_super_stripes(struct btrfs_root *root,
 255                                  struct btrfs_block_group_cache *cache)
 256 {
 257         u64 bytenr;
 258         u64 *logical;
 259         int stripe_len;
 260         int i, nr, ret;
 261
 262         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 263                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 264                 cache->bytes_super += stripe_len;
 265                 ret = add_excluded_extent(root, cache->key.objectid,
 266                                           stripe_len);
 267                 if (ret)
 268                         return ret;
 269         }
 270
 271         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 272                 bytenr = btrfs_sb_offset(i);
 273                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 274                                        cache->key.objectid, bytenr,
 275                                        0, &logical, &nr, &stripe_len);
 276                 if (ret)
 277                         return ret;
 278
 279                 while (nr--) {
 280                         u64 start, len;
 281
 282                         if (logical[nr] > cache->key.objectid +
 283                             cache->key.offset)
 284                                 continue;
 285
 286                         if (logical[nr] + stripe_len <= cache->key.objectid)
 287                                 continue;
 288
 289                         start = logical[nr];
 290                         if (start < cache->key.objectid) {
 291                                 start = cache->key.objectid;
 292                                 len = (logical[nr] + stripe_len) - start;
 293                         } else {
 294                                 len = min_t(u64, stripe_len,
 295                                             cache->key.objectid +
 296                                             cache->key.offset - start);
 297                         }
 298
 299                         cache->bytes_super += len;
 300                         ret = add_excluded_extent(root, start, len);
 301                         if (ret) {
 302                                 kfree(logical);
 303                                 return ret;
 304                         }
 305                 }
 306
 307                 kfree(logical);
 308         }
 309         return 0;
 310 }
 311
 312 static struct btrfs_caching_control *
 313 get_caching_control(struct btrfs_block_group_cache *cache)
 314 {
 315         struct btrfs_caching_control *ctl;
 316
 317         spin_lock(&cache->lock);
 318         if (cache->cached != BTRFS_CACHE_STARTED) {
 319                 spin_unlock(&cache->lock);
 320                 return NULL;
 321         }
 322
 323         /* We're loading it the fast way, so we don't have a caching_ctl. */
 324         if (!cache->caching_ctl) {
 325                 spin_unlock(&cache->lock);
 326                 return NULL;
 327         }
 328
 329         ctl = cache->caching_ctl;
 330         atomic_inc(&ctl->count);
 331         spin_unlock(&cache->lock);
 332         return ctl;
 333 }
 334
 335 static void put_caching_control(struct btrfs_caching_control *ctl)
 336 {
 337         if (atomic_dec_and_test(&ctl->count))
 338                 kfree(ctl);
 339 }
 340
 341 /*
 342  * this is only called by cache_block_group, since we could have freed extents
 343  * we need to check the pinned_extents for any extents that can't be used yet
 344  * since their free space will be released as soon as the transaction commits.
 345  */
 346 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 347                               struct btrfs_fs_info *info, u64 start, u64 end)
 348 {
 349         u64 extent_start, extent_end, size, total_added = 0;
 350         int ret;
 351
 352         while (start < end) {
 353                 ret = find_first_extent_bit(info->pinned_extents, start,
 354                                             &extent_start, &extent_end,
 355                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 356                                             NULL);
 357                 if (ret)
 358                         break;
 359
 360                 if (extent_start <= start) {
 361                         start = extent_end + 1;
 362                 } else if (extent_start > start && extent_start < end) {
 363                         size = extent_start - start;
 364                         total_added += size;
 365                         ret = btrfs_add_free_space(block_group, start,
 366                                                    size);
 367                         BUG_ON(ret); /* -ENOMEM or logic error */
 368                         start = extent_end + 1;
 369                 } else {
 370                         break;
 371                 }
 372         }
 373
 374         if (start < end) {
 375                 size = end - start;
 376                 total_added += size;
 377                 ret = btrfs_add_free_space(block_group, start, size);
 378                 BUG_ON(ret); /* -ENOMEM or logic error */
 379         }
 380
 381         return total_added;
 382 }
 383
 384 static noinline void caching_thread(struct btrfs_work *work)
 385 {
 386         struct btrfs_block_group_cache *block_group;
 387         struct btrfs_fs_info *fs_info;
 388         struct btrfs_caching_control *caching_ctl;
 389         struct btrfs_root *extent_root;
 390         struct btrfs_path *path;
 391         struct extent_buffer *leaf;
 392         struct btrfs_key key;
 393         u64 total_found = 0;
 394         u64 last = 0;
 395         u32 nritems;
 396         int ret = -ENOMEM;
 397
 398         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 399         block_group = caching_ctl->block_group;
 400         fs_info = block_group->fs_info;
 401         extent_root = fs_info->extent_root;
 402
 403         path = btrfs_alloc_path();
 404         if (!path)
 405                 goto out;
 406
 407         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 408
 409         /*
 410          * We don't want to deadlock with somebody trying to allocate a new
 411          * extent for the extent root while also trying to search the extent
 412          * root to add free space.  So we skip locking and search the commit
 413          * root, since its read-only
 414          */
 415         path->skip_locking = 1;
 416         path->search_commit_root = 1;
 417         path->reada = 1;
 418
 419         key.objectid = last;
 420         key.offset = 0;
 421         key.type = BTRFS_EXTENT_ITEM_KEY;
 422 again:
 423         mutex_lock(&caching_ctl->mutex);
 424         /* need to make sure the commit_root doesn't disappear */
 425         down_read(&fs_info->commit_root_sem);
 426
 427 next:
 428         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 429         if (ret < 0)
 430                 goto err;
 431
 432         leaf = path->nodes[0];
 433         nritems = btrfs_header_nritems(leaf);
 434
 435         while (1) {
 436                 if (btrfs_fs_closing(fs_info) > 1) {
 437                         last = (u64)-1;
 438                         break;
 439                 }
 440
 441                 if (path->slots[0] < nritems) {
 442                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 443                 } else {
 444                         ret = find_next_key(path, 0, &key);
 445                         if (ret)
 446                                 break;
 447
 448                         if (need_resched() ||
 449                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 450                                 caching_ctl->progress = last;
 451                                 btrfs_release_path(path);
 452                                 up_read(&fs_info->commit_root_sem);
 453                                 mutex_unlock(&caching_ctl->mutex);
 454                                 cond_resched();
 455                                 goto again;
 456                         }
 457
 458                         ret = btrfs_next_leaf(extent_root, path);
 459                         if (ret < 0)
 460                                 goto err;
 461                         if (ret)
 462                                 break;
 463                         leaf = path->nodes[0];
 464                         nritems = btrfs_header_nritems(leaf);
 465                         continue;
 466                 }
 467
 468                 if (key.objectid < last) {
 469                         key.objectid = last;
 470                         key.offset = 0;
 471                         key.type = BTRFS_EXTENT_ITEM_KEY;
 472
 473                         caching_ctl->progress = last;
 474                         btrfs_release_path(path);
 475                         goto next;
 476                 }
 477
 478                 if (key.objectid < block_group->key.objectid) {
 479                         path->slots[0]++;
 480                         continue;
 481                 }
 482
 483                 if (key.objectid >= block_group->key.objectid +
 484                     block_group->key.offset)
 485                         break;
 486
 487                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 488                     key.type == BTRFS_METADATA_ITEM_KEY) {
 489                         total_found += add_new_free_space(block_group,
 490                                                           fs_info, last,
 491                                                           key.objectid);
 492                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 493                                 last = key.objectid +
 494                                         fs_info->tree_root->leafsize;
 495                         else
 496                                 last = key.objectid + key.offset;
 497
 498                         if (total_found > (1024 * 1024 * 2)) {
 499                                 total_found = 0;
 500                                 wake_up(&caching_ctl->wait);
 501                         }
 502                 }
 503                 path->slots[0]++;
 504         }
 505         ret = 0;
 506
 507         total_found += add_new_free_space(block_group, fs_info, last,
 508                                           block_group->key.objectid +
 509                                           block_group->key.offset);
 510         caching_ctl->progress = (u64)-1;
 511
 512         spin_lock(&block_group->lock);
 513         block_group->caching_ctl = NULL;
 514         block_group->cached = BTRFS_CACHE_FINISHED;
 515         spin_unlock(&block_group->lock);
 516
 517 err:
 518         btrfs_free_path(path);
 519         up_read(&fs_info->commit_root_sem);
 520
 521         free_excluded_extents(extent_root, block_group);
 522
 523         mutex_unlock(&caching_ctl->mutex);
 524 out:
 525         if (ret) {
 526                 spin_lock(&block_group->lock);
 527                 block_group->caching_ctl = NULL;
 528                 block_group->cached = BTRFS_CACHE_ERROR;
 529                 spin_unlock(&block_group->lock);
 530         }
 531         wake_up(&caching_ctl->wait);
 532
 533         put_caching_control(caching_ctl);
 534         btrfs_put_block_group(block_group);
 535 }
 536
 537 static int cache_block_group(struct btrfs_block_group_cache *cache,
 538                              int load_cache_only)
 539 {
 540         DEFINE_WAIT(wait);
 541         struct btrfs_fs_info *fs_info = cache->fs_info;
 542         struct btrfs_caching_control *caching_ctl;
 543         int ret = 0;
 544
 545         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 546         if (!caching_ctl)
 547                 return -ENOMEM;
 548
 549         INIT_LIST_HEAD(&caching_ctl->list);
 550         mutex_init(&caching_ctl->mutex);
 551         init_waitqueue_head(&caching_ctl->wait);
 552         caching_ctl->block_group = cache;
 553         caching_ctl->progress = cache->key.objectid;
 554         atomic_set(&caching_ctl->count, 1);
 555         btrfs_init_work(&caching_ctl->work, caching_thread, NULL, NULL);
 556
 557         spin_lock(&cache->lock);
 558         /*
 559          * This should be a rare occasion, but this could happen I think in the
 560          * case where one thread starts to load the space cache info, and then
 561          * some other thread starts a transaction commit which tries to do an
 562          * allocation while the other thread is still loading the space cache
 563          * info.  The previous loop should have kept us from choosing this block
 564          * group, but if we've moved to the state where we will wait on caching
 565          * block groups we need to first check if we're doing a fast load here,
 566          * so we can wait for it to finish, otherwise we could end up allocating
 567          * from a block group who's cache gets evicted for one reason or
 568          * another.
 569          */
 570         while (cache->cached == BTRFS_CACHE_FAST) {
 571                 struct btrfs_caching_control *ctl;
 572
 573                 ctl = cache->caching_ctl;
 574                 atomic_inc(&ctl->count);
 575                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 576                 spin_unlock(&cache->lock);
 577
 578                 schedule();
 579
 580                 finish_wait(&ctl->wait, &wait);
 581                 put_caching_control(ctl);
 582                 spin_lock(&cache->lock);
 583         }
 584
 585         if (cache->cached != BTRFS_CACHE_NO) {
 586                 spin_unlock(&cache->lock);
 587                 kfree(caching_ctl);
 588                 return 0;
 589         }
 590         WARN_ON(cache->caching_ctl);
 591         cache->caching_ctl = caching_ctl;
 592         cache->cached = BTRFS_CACHE_FAST;
 593         spin_unlock(&cache->lock);
 594
 595         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 596                 ret = load_free_space_cache(fs_info, cache);
 597
 598                 spin_lock(&cache->lock);
 599                 if (ret == 1) {
 600                         cache->caching_ctl = NULL;
 601                         cache->cached = BTRFS_CACHE_FINISHED;
 602                         cache->last_byte_to_unpin = (u64)-1;
 603                 } else {
 604                         if (load_cache_only) {
 605                                 cache->caching_ctl = NULL;
 606                                 cache->cached = BTRFS_CACHE_NO;
 607                         } else {
 608                                 cache->cached = BTRFS_CACHE_STARTED;
 609                         }
 610                 }
 611                 spin_unlock(&cache->lock);
 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                 }
 630                 spin_unlock(&cache->lock);
 631                 wake_up(&caching_ctl->wait);
 632         }
 633
 634         if (load_cache_only) {
 635                 put_caching_control(caching_ctl);
 636                 return 0;
 637         }
 638
 639         down_write(&fs_info->commit_root_sem);
 640         atomic_inc(&caching_ctl->count);
 641         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 642         up_write(&fs_info->commit_root_sem);
 643
 644         btrfs_get_block_group(cache);
 645
 646         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 647
 648         return ret;
 649 }
 650
 651 /*
 652  * return the block group that starts at or after bytenr
 653  */
 654 static struct btrfs_block_group_cache *
 655 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 656 {
 657         struct btrfs_block_group_cache *cache;
 658
 659         cache = block_group_cache_tree_search(info, bytenr, 0);
 660
 661         return cache;
 662 }
 663
 664 /*
 665  * return the block group that contains the given bytenr
 666  */
 667 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 668                                                  struct btrfs_fs_info *info,
 669                                                  u64 bytenr)
 670 {
 671         struct btrfs_block_group_cache *cache;
 672
 673         cache = block_group_cache_tree_search(info, bytenr, 1);
 674
 675         return cache;
 676 }
 677
 678 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 679                                                   u64 flags)
 680 {
 681         struct list_head *head = &info->space_info;
 682         struct btrfs_space_info *found;
 683
 684         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 685
 686         rcu_read_lock();
 687         list_for_each_entry_rcu(found, head, list) {
 688                 if (found->flags & flags) {
 689                         rcu_read_unlock();
 690                         return found;
 691                 }
 692         }
 693         rcu_read_unlock();
 694         return NULL;
 695 }
 696
 697 /*
 698  * after adding space to the filesystem, we need to clear the full flags
 699  * on all the space infos.
 700  */
 701 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 702 {
 703         struct list_head *head = &info->space_info;
 704         struct btrfs_space_info *found;
 705
 706         rcu_read_lock();
 707         list_for_each_entry_rcu(found, head, list)
 708                 found->full = 0;
 709         rcu_read_unlock();
 710 }
 711
 712 /* simple helper to search for an existing extent at a given offset */
 713 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 714 {
 715         int ret;
 716         struct btrfs_key key;
 717         struct btrfs_path *path;
 718
 719         path = btrfs_alloc_path();
 720         if (!path)
 721                 return -ENOMEM;
 722
 723         key.objectid = start;
 724         key.offset = len;
 725         key.type = BTRFS_EXTENT_ITEM_KEY;
 726         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 727                                 0, 0);
 728         if (ret > 0) {
 729                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 730                 if (key.objectid == start &&
 731                     key.type == BTRFS_METADATA_ITEM_KEY)
 732                         ret = 0;
 733         }
 734         btrfs_free_path(path);
 735         return ret;
 736 }
 737
 738 /*
 739  * helper function to lookup reference count and flags of a tree block.
 740  *
 741  * the head node for delayed ref is used to store the sum of all the
 742  * reference count modifications queued up in the rbtree. the head
 743  * node may also store the extent flags to set. This way you can check
 744  * to see what the reference count and extent flags would be if all of
 745  * the delayed refs are not processed.
 746  */
 747 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 748                              struct btrfs_root *root, u64 bytenr,
 749                              u64 offset, int metadata, u64 *refs, u64 *flags)
 750 {
 751         struct btrfs_delayed_ref_head *head;
 752         struct btrfs_delayed_ref_root *delayed_refs;
 753         struct btrfs_path *path;
 754         struct btrfs_extent_item *ei;
 755         struct extent_buffer *leaf;
 756         struct btrfs_key key;
 757         u32 item_size;
 758         u64 num_refs;
 759         u64 extent_flags;
 760         int ret;
 761
 762         /*
 763          * If we don't have skinny metadata, don't bother doing anything
 764          * different
 765          */
 766         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 767                 offset = root->leafsize;
 768                 metadata = 0;
 769         }
 770
 771         path = btrfs_alloc_path();
 772         if (!path)
 773                 return -ENOMEM;
 774
 775         if (!trans) {
 776                 path->skip_locking = 1;
 777                 path->search_commit_root = 1;
 778         }
 779
 780 search_again:
 781         key.objectid = bytenr;
 782         key.offset = offset;
 783         if (metadata)
 784                 key.type = BTRFS_METADATA_ITEM_KEY;
 785         else
 786                 key.type = BTRFS_EXTENT_ITEM_KEY;
 787
 788 again:
 789         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 790                                 &key, path, 0, 0);
 791         if (ret < 0)
 792                 goto out_free;
 793
 794         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 795                 if (path->slots[0]) {
 796                         path->slots[0]--;
 797                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 798                                               path->slots[0]);
 799                         if (key.objectid == bytenr &&
 800                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 801                             key.offset == root->leafsize)
 802                                 ret = 0;
 803                 }
 804                 if (ret) {
 805                         key.objectid = bytenr;
 806                         key.type = BTRFS_EXTENT_ITEM_KEY;
 807                         key.offset = root->leafsize;
 808                         btrfs_release_path(path);
 809                         goto again;
 810                 }
 811         }
 812
 813         if (ret == 0) {
 814                 leaf = path->nodes[0];
 815                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 816                 if (item_size >= sizeof(*ei)) {
 817                         ei = btrfs_item_ptr(leaf, path->slots[0],
 818                                             struct btrfs_extent_item);
 819                         num_refs = btrfs_extent_refs(leaf, ei);
 820                         extent_flags = btrfs_extent_flags(leaf, ei);
 821                 } else {
 822 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 823                         struct btrfs_extent_item_v0 *ei0;
 824                         BUG_ON(item_size != sizeof(*ei0));
 825                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 826                                              struct btrfs_extent_item_v0);
 827                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 828                         /* FIXME: this isn't correct for data */
 829                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 830 #else
 831                         BUG();
 832 #endif
 833                 }
 834                 BUG_ON(num_refs == 0);
 835         } else {
 836                 num_refs = 0;
 837                 extent_flags = 0;
 838                 ret = 0;
 839         }
 840
 841         if (!trans)
 842                 goto out;
 843
 844         delayed_refs = &trans->transaction->delayed_refs;
 845         spin_lock(&delayed_refs->lock);
 846         head = btrfs_find_delayed_ref_head(trans, bytenr);
 847         if (head) {
 848                 if (!mutex_trylock(&head->mutex)) {
 849                         atomic_inc(&head->node.refs);
 850                         spin_unlock(&delayed_refs->lock);
 851
 852                         btrfs_release_path(path);
 853
 854                         /*
 855                          * Mutex was contended, block until it's released and try
 856                          * again
 857                          */
 858                         mutex_lock(&head->mutex);
 859                         mutex_unlock(&head->mutex);
 860                         btrfs_put_delayed_ref(&head->node);
 861                         goto search_again;
 862                 }
 863                 spin_lock(&head->lock);
 864                 if (head->extent_op && head->extent_op->update_flags)
 865                         extent_flags |= head->extent_op->flags_to_set;
 866                 else
 867                         BUG_ON(num_refs == 0);
 868
 869                 num_refs += head->node.ref_mod;
 870                 spin_unlock(&head->lock);
 871                 mutex_unlock(&head->mutex);
 872         }
 873         spin_unlock(&delayed_refs->lock);
 874 out:
 875         WARN_ON(num_refs == 0);
 876         if (refs)
 877                 *refs = num_refs;
 878         if (flags)
 879                 *flags = extent_flags;
 880 out_free:
 881         btrfs_free_path(path);
 882         return ret;
 883 }
 884
 885 /*
 886  * Back reference rules.  Back refs have three main goals:
 887  *
 888  * 1) differentiate between all holders of references to an extent so that
 889  *    when a reference is dropped we can make sure it was a valid reference
 890  *    before freeing the extent.
 891  *
 892  * 2) Provide enough information to quickly find the holders of an extent
 893  *    if we notice a given block is corrupted or bad.
 894  *
 895  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 896  *    maintenance.  This is actually the same as #2, but with a slightly
 897  *    different use case.
 898  *
 899  * There are two kinds of back refs. The implicit back refs is optimized
 900  * for pointers in non-shared tree blocks. For a given pointer in a block,
 901  * back refs of this kind provide information about the block's owner tree
 902  * and the pointer's key. These information allow us to find the block by
 903  * b-tree searching. The full back refs is for pointers in tree blocks not
 904  * referenced by their owner trees. The location of tree block is recorded
 905  * in the back refs. Actually the full back refs is generic, and can be
 906  * used in all cases the implicit back refs is used. The major shortcoming
 907  * of the full back refs is its overhead. Every time a tree block gets
 908  * COWed, we have to update back refs entry for all pointers in it.
 909  *
 910  * For a newly allocated tree block, we use implicit back refs for
 911  * pointers in it. This means most tree related operations only involve
 912  * implicit back refs. For a tree block created in old transaction, the
 913  * only way to drop a reference to it is COW it. So we can detect the
 914  * event that tree block loses its owner tree's reference and do the
 915  * back refs conversion.
 916  *
 917  * When a tree block is COW'd through a tree, there are four cases:
 918  *
 919  * The reference count of the block is one and the tree is the block's
 920  * owner tree. Nothing to do in this case.
 921  *
 922  * The reference count of the block is one and the tree is not the
 923  * block's owner tree. In this case, full back refs is used for pointers
 924  * in the block. Remove these full back refs, add implicit back refs for
 925  * every pointers in the new block.
 926  *
 927  * The reference count of the block is greater than one and the tree is
 928  * the block's owner tree. In this case, implicit back refs is used for
 929  * pointers in the block. Add full back refs for every pointers in the
 930  * block, increase lower level extents' reference counts. The original
 931  * implicit back refs are entailed to the new block.
 932  *
 933  * The reference count of the block is greater than one and the tree is
 934  * not the block's owner tree. Add implicit back refs for every pointer in
 935  * the new block, increase lower level extents' reference count.
 936  *
 937  * Back Reference Key composing:
 938  *
 939  * The key objectid corresponds to the first byte in the extent,
 940  * The key type is used to differentiate between types of back refs.
 941  * There are different meanings of the key offset for different types
 942  * of back refs.
 943  *
 944  * File extents can be referenced by:
 945  *
 946  * - multiple snapshots, subvolumes, or different generations in one subvol
 947  * - different files inside a single subvolume
 948  * - different offsets inside a file (bookend extents in file.c)
 949  *
 950  * The extent ref structure for the implicit back refs has fields for:
 951  *
 952  * - Objectid of the subvolume root
 953  * - objectid of the file holding the reference
 954  * - original offset in the file
 955  * - how many bookend extents
 956  *
 957  * The key offset for the implicit back refs is hash of the first
 958  * three fields.
 959  *
 960  * The extent ref structure for the full back refs has field for:
 961  *
 962  * - number of pointers in the tree leaf
 963  *
 964  * The key offset for the implicit back refs is the first byte of
 965  * the tree leaf
 966  *
 967  * When a file extent is allocated, The implicit back refs is used.
 968  * the fields are filled in:
 969  *
 970  *     (root_key.objectid, inode objectid, offset in file, 1)
 971  *
 972  * When a file extent is removed file truncation, we find the
 973  * corresponding implicit back refs and check the following fields:
 974  *
 975  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 976  *
 977  * Btree extents can be referenced by:
 978  *
 979  * - Different subvolumes
 980  *
 981  * Both the implicit back refs and the full back refs for tree blocks
 982  * only consist of key. The key offset for the implicit back refs is
 983  * objectid of block's owner tree. The key offset for the full back refs
 984  * is the first byte of parent block.
 985  *
 986  * When implicit back refs is used, information about the lowest key and
 987  * level of the tree block are required. These information are stored in
 988  * tree block info structure.
 989  */
 990
 991 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 992 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 993                                   struct btrfs_root *root,
 994                                   struct btrfs_path *path,
 995                                   u64 owner, u32 extra_size)
 996 {
 997         struct btrfs_extent_item *item;
 998         struct btrfs_extent_item_v0 *ei0;
 999         struct btrfs_extent_ref_v0 *ref0;
1000         struct btrfs_tree_block_info *bi;
1001         struct extent_buffer *leaf;
1002         struct btrfs_key key;
1003         struct btrfs_key found_key;
1004         u32 new_size = sizeof(*item);
1005         u64 refs;
1006         int ret;
1007
1008         leaf = path->nodes[0];
1009         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1010
1011         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1012         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1013                              struct btrfs_extent_item_v0);
1014         refs = btrfs_extent_refs_v0(leaf, ei0);
1015
1016         if (owner == (u64)-1) {
1017                 while (1) {
1018                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1019                                 ret = btrfs_next_leaf(root, path);
1020                                 if (ret < 0)
1021                                         return ret;
1022                                 BUG_ON(ret > 0); /* Corruption */
1023                                 leaf = path->nodes[0];
1024                         }
1025                         btrfs_item_key_to_cpu(leaf, &found_key,
1026                                               path->slots[0]);
1027                         BUG_ON(key.objectid != found_key.objectid);
1028                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1029                                 path->slots[0]++;
1030                                 continue;
1031                         }
1032                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1033                                               struct btrfs_extent_ref_v0);
1034                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1035                         break;
1036                 }
1037         }
1038         btrfs_release_path(path);
1039
1040         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1041                 new_size += sizeof(*bi);
1042
1043         new_size -= sizeof(*ei0);
1044         ret = btrfs_search_slot(trans, root, &key, path,
1045                                 new_size + extra_size, 1);
1046         if (ret < 0)
1047                 return ret;
1048         BUG_ON(ret); /* Corruption */
1049
1050         btrfs_extend_item(root, path, new_size);
1051
1052         leaf = path->nodes[0];
1053         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1054         btrfs_set_extent_refs(leaf, item, refs);
1055         /* FIXME: get real generation */
1056         btrfs_set_extent_generation(leaf, item, 0);
1057         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1058                 btrfs_set_extent_flags(leaf, item,
1059                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1060                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1061                 bi = (struct btrfs_tree_block_info *)(item + 1);
1062                 /* FIXME: get first key of the block */
1063                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1064                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1065         } else {
1066                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1067         }
1068         btrfs_mark_buffer_dirty(leaf);
1069         return 0;
1070 }
1071 #endif
1072
1073 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1074 {
1075         u32 high_crc = ~(u32)0;
1076         u32 low_crc = ~(u32)0;
1077         __le64 lenum;
1078
1079         lenum = cpu_to_le64(root_objectid);
1080         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1081         lenum = cpu_to_le64(owner);
1082         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1083         lenum = cpu_to_le64(offset);
1084         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1085
1086         return ((u64)high_crc << 31) ^ (u64)low_crc;
1087 }
1088
1089 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1090                                      struct btrfs_extent_data_ref *ref)
1091 {
1092         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1093                                     btrfs_extent_data_ref_objectid(leaf, ref),
1094                                     btrfs_extent_data_ref_offset(leaf, ref));
1095 }
1096
1097 static int match_extent_data_ref(struct extent_buffer *leaf,
1098                                  struct btrfs_extent_data_ref *ref,
1099                                  u64 root_objectid, u64 owner, u64 offset)
1100 {
1101         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1102             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1103             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1104                 return 0;
1105         return 1;
1106 }
1107
1108 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1109                                            struct btrfs_root *root,
1110                                            struct btrfs_path *path,
1111                                            u64 bytenr, u64 parent,
1112                                            u64 root_objectid,
1113                                            u64 owner, u64 offset)
1114 {
1115         struct btrfs_key key;
1116         struct btrfs_extent_data_ref *ref;
1117         struct extent_buffer *leaf;
1118         u32 nritems;
1119         int ret;
1120         int recow;
1121         int err = -ENOENT;
1122
1123         key.objectid = bytenr;
1124         if (parent) {
1125                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1126                 key.offset = parent;
1127         } else {
1128                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1129                 key.offset = hash_extent_data_ref(root_objectid,
1130                                                   owner, offset);
1131         }
1132 again:
1133         recow = 0;
1134         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1135         if (ret < 0) {
1136                 err = ret;
1137                 goto fail;
1138         }
1139
1140         if (parent) {
1141                 if (!ret)
1142                         return 0;
1143 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1144                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1145                 btrfs_release_path(path);
1146                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1147                 if (ret < 0) {
1148                         err = ret;
1149                         goto fail;
1150                 }
1151                 if (!ret)
1152                         return 0;
1153 #endif
1154                 goto fail;
1155         }
1156
1157         leaf = path->nodes[0];
1158         nritems = btrfs_header_nritems(leaf);
1159         while (1) {
1160                 if (path->slots[0] >= nritems) {
1161                         ret = btrfs_next_leaf(root, path);
1162                         if (ret < 0)
1163                                 err = ret;
1164                         if (ret)
1165                                 goto fail;
1166
1167                         leaf = path->nodes[0];
1168                         nritems = btrfs_header_nritems(leaf);
1169                         recow = 1;
1170                 }
1171
1172                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1173                 if (key.objectid != bytenr ||
1174                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1175                         goto fail;
1176
1177                 ref = btrfs_item_ptr(leaf, path->slots[0],
1178                                      struct btrfs_extent_data_ref);
1179
1180                 if (match_extent_data_ref(leaf, ref, root_objectid,
1181                                           owner, offset)) {
1182                         if (recow) {
1183                                 btrfs_release_path(path);
1184                                 goto again;
1185                         }
1186                         err = 0;
1187                         break;
1188                 }
1189                 path->slots[0]++;
1190         }
1191 fail:
1192         return err;
1193 }
1194
1195 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1196                                            struct btrfs_root *root,
1197                                            struct btrfs_path *path,
1198                                            u64 bytenr, u64 parent,
1199                                            u64 root_objectid, u64 owner,
1200                                            u64 offset, int refs_to_add)
1201 {
1202         struct btrfs_key key;
1203         struct extent_buffer *leaf;
1204         u32 size;
1205         u32 num_refs;
1206         int ret;
1207
1208         key.objectid = bytenr;
1209         if (parent) {
1210                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1211                 key.offset = parent;
1212                 size = sizeof(struct btrfs_shared_data_ref);
1213         } else {
1214                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1215                 key.offset = hash_extent_data_ref(root_objectid,
1216                                                   owner, offset);
1217                 size = sizeof(struct btrfs_extent_data_ref);
1218         }
1219
1220         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1221         if (ret && ret != -EEXIST)
1222                 goto fail;
1223
1224         leaf = path->nodes[0];
1225         if (parent) {
1226                 struct btrfs_shared_data_ref *ref;
1227                 ref = btrfs_item_ptr(leaf, path->slots[0],
1228                                      struct btrfs_shared_data_ref);
1229                 if (ret == 0) {
1230                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1231                 } else {
1232                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1233                         num_refs += refs_to_add;
1234                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1235                 }
1236         } else {
1237                 struct btrfs_extent_data_ref *ref;
1238                 while (ret == -EEXIST) {
1239                         ref = btrfs_item_ptr(leaf, path->slots[0],
1240                                              struct btrfs_extent_data_ref);
1241                         if (match_extent_data_ref(leaf, ref, root_objectid,
1242                                                   owner, offset))
1243                                 break;
1244                         btrfs_release_path(path);
1245                         key.offset++;
1246                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1247                                                       size);
1248                         if (ret && ret != -EEXIST)
1249                                 goto fail;
1250
1251                         leaf = path->nodes[0];
1252                 }
1253                 ref = btrfs_item_ptr(leaf, path->slots[0],
1254                                      struct btrfs_extent_data_ref);
1255                 if (ret == 0) {
1256                         btrfs_set_extent_data_ref_root(leaf, ref,
1257                                                        root_objectid);
1258                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1259                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1260                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1261                 } else {
1262                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1263                         num_refs += refs_to_add;
1264                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1265                 }
1266         }
1267         btrfs_mark_buffer_dirty(leaf);
1268         ret = 0;
1269 fail:
1270         btrfs_release_path(path);
1271         return ret;
1272 }
1273
1274 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1275                                            struct btrfs_root *root,
1276                                            struct btrfs_path *path,
1277                                            int refs_to_drop, int *last_ref)
1278 {
1279         struct btrfs_key key;
1280         struct btrfs_extent_data_ref *ref1 = NULL;
1281         struct btrfs_shared_data_ref *ref2 = NULL;
1282         struct extent_buffer *leaf;
1283         u32 num_refs = 0;
1284         int ret = 0;
1285
1286         leaf = path->nodes[0];
1287         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1288
1289         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1290                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1291                                       struct btrfs_extent_data_ref);
1292                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1293         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1294                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1295                                       struct btrfs_shared_data_ref);
1296                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1297 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1298         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1299                 struct btrfs_extent_ref_v0 *ref0;
1300                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1301                                       struct btrfs_extent_ref_v0);
1302                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1303 #endif
1304         } else {
1305                 BUG();
1306         }
1307
1308         BUG_ON(num_refs < refs_to_drop);
1309         num_refs -= refs_to_drop;
1310
1311         if (num_refs == 0) {
1312                 ret = btrfs_del_item(trans, root, path);
1313                 *last_ref = 1;
1314         } else {
1315                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1316                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1317                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1318                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1319 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1320                 else {
1321                         struct btrfs_extent_ref_v0 *ref0;
1322                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1323                                         struct btrfs_extent_ref_v0);
1324                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1325                 }
1326 #endif
1327                 btrfs_mark_buffer_dirty(leaf);
1328         }
1329         return ret;
1330 }
1331
1332 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1333                                           struct btrfs_path *path,
1334                                           struct btrfs_extent_inline_ref *iref)
1335 {
1336         struct btrfs_key key;
1337         struct extent_buffer *leaf;
1338         struct btrfs_extent_data_ref *ref1;
1339         struct btrfs_shared_data_ref *ref2;
1340         u32 num_refs = 0;
1341
1342         leaf = path->nodes[0];
1343         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1344         if (iref) {
1345                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1346                     BTRFS_EXTENT_DATA_REF_KEY) {
1347                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1348                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1349                 } else {
1350                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1351                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1352                 }
1353         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1354                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1355                                       struct btrfs_extent_data_ref);
1356                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1357         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1358                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1359                                       struct btrfs_shared_data_ref);
1360                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1361 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1362         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1363                 struct btrfs_extent_ref_v0 *ref0;
1364                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1365                                       struct btrfs_extent_ref_v0);
1366                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1367 #endif
1368         } else {
1369                 WARN_ON(1);
1370         }
1371         return num_refs;
1372 }
1373
1374 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1375                                           struct btrfs_root *root,
1376                                           struct btrfs_path *path,
1377                                           u64 bytenr, u64 parent,
1378                                           u64 root_objectid)
1379 {
1380         struct btrfs_key key;
1381         int ret;
1382
1383         key.objectid = bytenr;
1384         if (parent) {
1385                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1386                 key.offset = parent;
1387         } else {
1388                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1389                 key.offset = root_objectid;
1390         }
1391
1392         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1393         if (ret > 0)
1394                 ret = -ENOENT;
1395 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1396         if (ret == -ENOENT && parent) {
1397                 btrfs_release_path(path);
1398                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1399                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1400                 if (ret > 0)
1401                         ret = -ENOENT;
1402         }
1403 #endif
1404         return ret;
1405 }
1406
1407 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1408                                           struct btrfs_root *root,
1409                                           struct btrfs_path *path,
1410                                           u64 bytenr, u64 parent,
1411                                           u64 root_objectid)
1412 {
1413         struct btrfs_key key;
1414         int ret;
1415
1416         key.objectid = bytenr;
1417         if (parent) {
1418                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1419                 key.offset = parent;
1420         } else {
1421                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1422                 key.offset = root_objectid;
1423         }
1424
1425         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1426         btrfs_release_path(path);
1427         return ret;
1428 }
1429
1430 static inline int extent_ref_type(u64 parent, u64 owner)
1431 {
1432         int type;
1433         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1434                 if (parent > 0)
1435                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1436                 else
1437                         type = BTRFS_TREE_BLOCK_REF_KEY;
1438         } else {
1439                 if (parent > 0)
1440                         type = BTRFS_SHARED_DATA_REF_KEY;
1441                 else
1442                         type = BTRFS_EXTENT_DATA_REF_KEY;
1443         }
1444         return type;
1445 }
1446
1447 static int find_next_key(struct btrfs_path *path, int level,
1448                          struct btrfs_key *key)
1449
1450 {
1451         for (; level < BTRFS_MAX_LEVEL; level++) {
1452                 if (!path->nodes[level])
1453                         break;
1454                 if (path->slots[level] + 1 >=
1455                     btrfs_header_nritems(path->nodes[level]))
1456                         continue;
1457                 if (level == 0)
1458                         btrfs_item_key_to_cpu(path->nodes[level], key,
1459                                               path->slots[level] + 1);
1460                 else
1461                         btrfs_node_key_to_cpu(path->nodes[level], key,
1462                                               path->slots[level] + 1);
1463                 return 0;
1464         }
1465         return 1;
1466 }
1467
1468 /*
1469  * look for inline back ref. if back ref is found, *ref_ret is set
1470  * to the address of inline back ref, and 0 is returned.
1471  *
1472  * if back ref isn't found, *ref_ret is set to the address where it
1473  * should be inserted, and -ENOENT is returned.
1474  *
1475  * if insert is true and there are too many inline back refs, the path
1476  * points to the extent item, and -EAGAIN is returned.
1477  *
1478  * NOTE: inline back refs are ordered in the same way that back ref
1479  *       items in the tree are ordered.
1480  */
1481 static noinline_for_stack
1482 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1483                                  struct btrfs_root *root,
1484                                  struct btrfs_path *path,
1485                                  struct btrfs_extent_inline_ref **ref_ret,
1486                                  u64 bytenr, u64 num_bytes,
1487                                  u64 parent, u64 root_objectid,
1488                                  u64 owner, u64 offset, int insert)
1489 {
1490         struct btrfs_key key;
1491         struct extent_buffer *leaf;
1492         struct btrfs_extent_item *ei;
1493         struct btrfs_extent_inline_ref *iref;
1494         u64 flags;
1495         u64 item_size;
1496         unsigned long ptr;
1497         unsigned long end;
1498         int extra_size;
1499         int type;
1500         int want;
1501         int ret;
1502         int err = 0;
1503         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1504                                                  SKINNY_METADATA);
1505
1506         key.objectid = bytenr;
1507         key.type = BTRFS_EXTENT_ITEM_KEY;
1508         key.offset = num_bytes;
1509
1510         want = extent_ref_type(parent, owner);
1511         if (insert) {
1512                 extra_size = btrfs_extent_inline_ref_size(want);
1513                 path->keep_locks = 1;
1514         } else
1515                 extra_size = -1;
1516
1517         /*
1518          * Owner is our parent level, so we can just add one to get the level
1519          * for the block we are interested in.
1520          */
1521         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1522                 key.type = BTRFS_METADATA_ITEM_KEY;
1523                 key.offset = owner;
1524         }
1525
1526 again:
1527         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1528         if (ret < 0) {
1529                 err = ret;
1530                 goto out;
1531         }
1532
1533         /*
1534          * We may be a newly converted file system which still has the old fat
1535          * extent entries for metadata, so try and see if we have one of those.
1536          */
1537         if (ret > 0 && skinny_metadata) {
1538                 skinny_metadata = false;
1539                 if (path->slots[0]) {
1540                         path->slots[0]--;
1541                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1542                                               path->slots[0]);
1543                         if (key.objectid == bytenr &&
1544                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1545                             key.offset == num_bytes)
1546                                 ret = 0;
1547                 }
1548                 if (ret) {
1549                         key.objectid = bytenr;
1550                         key.type = BTRFS_EXTENT_ITEM_KEY;
1551                         key.offset = num_bytes;
1552                         btrfs_release_path(path);
1553                         goto again;
1554                 }
1555         }
1556
1557         if (ret && !insert) {
1558                 err = -ENOENT;
1559                 goto out;
1560         } else if (WARN_ON(ret)) {
1561                 err = -EIO;
1562                 goto out;
1563         }
1564
1565         leaf = path->nodes[0];
1566         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1567 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1568         if (item_size < sizeof(*ei)) {
1569                 if (!insert) {
1570                         err = -ENOENT;
1571                         goto out;
1572                 }
1573                 ret = convert_extent_item_v0(trans, root, path, owner,
1574                                              extra_size);
1575                 if (ret < 0) {
1576                         err = ret;
1577                         goto out;
1578                 }
1579                 leaf = path->nodes[0];
1580                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1581         }
1582 #endif
1583         BUG_ON(item_size < sizeof(*ei));
1584
1585         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1586         flags = btrfs_extent_flags(leaf, ei);
1587
1588         ptr = (unsigned long)(ei + 1);
1589         end = (unsigned long)ei + item_size;
1590
1591         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1592                 ptr += sizeof(struct btrfs_tree_block_info);
1593                 BUG_ON(ptr > end);
1594         }
1595
1596         err = -ENOENT;
1597         while (1) {
1598                 if (ptr >= end) {
1599                         WARN_ON(ptr > end);
1600                         break;
1601                 }
1602                 iref = (struct btrfs_extent_inline_ref *)ptr;
1603                 type = btrfs_extent_inline_ref_type(leaf, iref);
1604                 if (want < type)
1605                         break;
1606                 if (want > type) {
1607                         ptr += btrfs_extent_inline_ref_size(type);
1608                         continue;
1609                 }
1610
1611                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1612                         struct btrfs_extent_data_ref *dref;
1613                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1614                         if (match_extent_data_ref(leaf, dref, root_objectid,
1615                                                   owner, offset)) {
1616                                 err = 0;
1617                                 break;
1618                         }
1619                         if (hash_extent_data_ref_item(leaf, dref) <
1620                             hash_extent_data_ref(root_objectid, owner, offset))
1621                                 break;
1622                 } else {
1623                         u64 ref_offset;
1624                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1625                         if (parent > 0) {
1626                                 if (parent == ref_offset) {
1627                                         err = 0;
1628                                         break;
1629                                 }
1630                                 if (ref_offset < parent)
1631                                         break;
1632                         } else {
1633                                 if (root_objectid == ref_offset) {
1634                                         err = 0;
1635                                         break;
1636                                 }
1637                                 if (ref_offset < root_objectid)
1638                                         break;
1639                         }
1640                 }
1641                 ptr += btrfs_extent_inline_ref_size(type);
1642         }
1643         if (err == -ENOENT && insert) {
1644                 if (item_size + extra_size >=
1645                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1646                         err = -EAGAIN;
1647                         goto out;
1648                 }
1649                 /*
1650                  * To add new inline back ref, we have to make sure
1651                  * there is no corresponding back ref item.
1652                  * For simplicity, we just do not add new inline back
1653                  * ref if there is any kind of item for this block
1654                  */
1655                 if (find_next_key(path, 0, &key) == 0 &&
1656                     key.objectid == bytenr &&
1657                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1658                         err = -EAGAIN;
1659                         goto out;
1660                 }
1661         }
1662         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1663 out:
1664         if (insert) {
1665                 path->keep_locks = 0;
1666                 btrfs_unlock_up_safe(path, 1);
1667         }
1668         return err;
1669 }
1670
1671 /*
1672  * helper to add new inline back ref
1673  */
1674 static noinline_for_stack
1675 void setup_inline_extent_backref(struct btrfs_root *root,
1676                                  struct btrfs_path *path,
1677                                  struct btrfs_extent_inline_ref *iref,
1678                                  u64 parent, u64 root_objectid,
1679                                  u64 owner, u64 offset, int refs_to_add,
1680                                  struct btrfs_delayed_extent_op *extent_op)
1681 {
1682         struct extent_buffer *leaf;
1683         struct btrfs_extent_item *ei;
1684         unsigned long ptr;
1685         unsigned long end;
1686         unsigned long item_offset;
1687         u64 refs;
1688         int size;
1689         int type;
1690
1691         leaf = path->nodes[0];
1692         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1693         item_offset = (unsigned long)iref - (unsigned long)ei;
1694
1695         type = extent_ref_type(parent, owner);
1696         size = btrfs_extent_inline_ref_size(type);
1697
1698         btrfs_extend_item(root, path, size);
1699
1700         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1701         refs = btrfs_extent_refs(leaf, ei);
1702         refs += refs_to_add;
1703         btrfs_set_extent_refs(leaf, ei, refs);
1704         if (extent_op)
1705                 __run_delayed_extent_op(extent_op, leaf, ei);
1706
1707         ptr = (unsigned long)ei + item_offset;
1708         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1709         if (ptr < end - size)
1710                 memmove_extent_buffer(leaf, ptr + size, ptr,
1711                                       end - size - ptr);
1712
1713         iref = (struct btrfs_extent_inline_ref *)ptr;
1714         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1715         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1716                 struct btrfs_extent_data_ref *dref;
1717                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1718                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1719                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1720                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1721                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1722         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1723                 struct btrfs_shared_data_ref *sref;
1724                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1725                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1726                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1727         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1728                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1729         } else {
1730                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1731         }
1732         btrfs_mark_buffer_dirty(leaf);
1733 }
1734
1735 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1736                                  struct btrfs_root *root,
1737                                  struct btrfs_path *path,
1738                                  struct btrfs_extent_inline_ref **ref_ret,
1739                                  u64 bytenr, u64 num_bytes, u64 parent,
1740                                  u64 root_objectid, u64 owner, u64 offset)
1741 {
1742         int ret;
1743
1744         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1745                                            bytenr, num_bytes, parent,
1746                                            root_objectid, owner, offset, 0);
1747         if (ret != -ENOENT)
1748                 return ret;
1749
1750         btrfs_release_path(path);
1751         *ref_ret = NULL;
1752
1753         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1754                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1755                                             root_objectid);
1756         } else {
1757                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1758                                              root_objectid, owner, offset);
1759         }
1760         return ret;
1761 }
1762
1763 /*
1764  * helper to update/remove inline back ref
1765  */
1766 static noinline_for_stack
1767 void update_inline_extent_backref(struct btrfs_root *root,
1768                                   struct btrfs_path *path,
1769                                   struct btrfs_extent_inline_ref *iref,
1770                                   int refs_to_mod,
1771                                   struct btrfs_delayed_extent_op *extent_op,
1772                                   int *last_ref)
1773 {
1774         struct extent_buffer *leaf;
1775         struct btrfs_extent_item *ei;
1776         struct btrfs_extent_data_ref *dref = NULL;
1777         struct btrfs_shared_data_ref *sref = NULL;
1778         unsigned long ptr;
1779         unsigned long end;
1780         u32 item_size;
1781         int size;
1782         int type;
1783         u64 refs;
1784
1785         leaf = path->nodes[0];
1786         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1787         refs = btrfs_extent_refs(leaf, ei);
1788         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1789         refs += refs_to_mod;
1790         btrfs_set_extent_refs(leaf, ei, refs);
1791         if (extent_op)
1792                 __run_delayed_extent_op(extent_op, leaf, ei);
1793
1794         type = btrfs_extent_inline_ref_type(leaf, iref);
1795
1796         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1797                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1798                 refs = btrfs_extent_data_ref_count(leaf, dref);
1799         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1800                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1801                 refs = btrfs_shared_data_ref_count(leaf, sref);
1802         } else {
1803                 refs = 1;
1804                 BUG_ON(refs_to_mod != -1);
1805         }
1806
1807         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1808         refs += refs_to_mod;
1809
1810         if (refs > 0) {
1811                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1812                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1813                 else
1814                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1815         } else {
1816                 *last_ref = 1;
1817                 size =  btrfs_extent_inline_ref_size(type);
1818                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1819                 ptr = (unsigned long)iref;
1820                 end = (unsigned long)ei + item_size;
1821                 if (ptr + size < end)
1822                         memmove_extent_buffer(leaf, ptr, ptr + size,
1823                                               end - ptr - size);
1824                 item_size -= size;
1825                 btrfs_truncate_item(root, path, item_size, 1);
1826         }
1827         btrfs_mark_buffer_dirty(leaf);
1828 }
1829
1830 static noinline_for_stack
1831 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1832                                  struct btrfs_root *root,
1833                                  struct btrfs_path *path,
1834                                  u64 bytenr, u64 num_bytes, u64 parent,
1835                                  u64 root_objectid, u64 owner,
1836                                  u64 offset, int refs_to_add,
1837                                  struct btrfs_delayed_extent_op *extent_op)
1838 {
1839         struct btrfs_extent_inline_ref *iref;
1840         int ret;
1841
1842         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1843                                            bytenr, num_bytes, parent,
1844                                            root_objectid, owner, offset, 1);
1845         if (ret == 0) {
1846                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1847                 update_inline_extent_backref(root, path, iref,
1848                                              refs_to_add, extent_op, NULL);
1849         } else if (ret == -ENOENT) {
1850                 setup_inline_extent_backref(root, path, iref, parent,
1851                                             root_objectid, owner, offset,
1852                                             refs_to_add, extent_op);
1853                 ret = 0;
1854         }
1855         return ret;
1856 }
1857
1858 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1859                                  struct btrfs_root *root,
1860                                  struct btrfs_path *path,
1861                                  u64 bytenr, u64 parent, u64 root_objectid,
1862                                  u64 owner, u64 offset, int refs_to_add)
1863 {
1864         int ret;
1865         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1866                 BUG_ON(refs_to_add != 1);
1867                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1868                                             parent, root_objectid);
1869         } else {
1870                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1871                                              parent, root_objectid,
1872                                              owner, offset, refs_to_add);
1873         }
1874         return ret;
1875 }
1876
1877 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1878                                  struct btrfs_root *root,
1879                                  struct btrfs_path *path,
1880                                  struct btrfs_extent_inline_ref *iref,
1881                                  int refs_to_drop, int is_data, int *last_ref)
1882 {
1883         int ret = 0;
1884
1885         BUG_ON(!is_data && refs_to_drop != 1);
1886         if (iref) {
1887                 update_inline_extent_backref(root, path, iref,
1888                                              -refs_to_drop, NULL, last_ref);
1889         } else if (is_data) {
1890                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop,
1891                                              last_ref);
1892         } else {
1893                 *last_ref = 1;
1894                 ret = btrfs_del_item(trans, root, path);
1895         }
1896         return ret;
1897 }
1898
1899 static int btrfs_issue_discard(struct block_device *bdev,
1900                                 u64 start, u64 len)
1901 {
1902         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1903 }
1904
1905 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1906                                 u64 num_bytes, u64 *actual_bytes)
1907 {
1908         int ret;
1909         u64 discarded_bytes = 0;
1910         struct btrfs_bio *bbio = NULL;
1911
1912
1913         /* Tell the block device(s) that the sectors can be discarded */
1914         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1915                               bytenr, &num_bytes, &bbio, 0);
1916         /* Error condition is -ENOMEM */
1917         if (!ret) {
1918                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1919                 int i;
1920
1921
1922                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1923                         if (!stripe->dev->can_discard)
1924                                 continue;
1925
1926                         ret = btrfs_issue_discard(stripe->dev->bdev,
1927                                                   stripe->physical,
1928                                                   stripe->length);
1929                         if (!ret)
1930                                 discarded_bytes += stripe->length;
1931                         else if (ret != -EOPNOTSUPP)
1932                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1933
1934                         /*
1935                          * Just in case we get back EOPNOTSUPP for some reason,
1936                          * just ignore the return value so we don't screw up
1937                          * people calling discard_extent.
1938                          */
1939                         ret = 0;
1940                 }
1941                 kfree(bbio);
1942         }
1943
1944         if (actual_bytes)
1945                 *actual_bytes = discarded_bytes;
1946
1947
1948         if (ret == -EOPNOTSUPP)
1949                 ret = 0;
1950         return ret;
1951 }
1952
1953 /* Can return -ENOMEM */
1954 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1955                          struct btrfs_root *root,
1956                          u64 bytenr, u64 num_bytes, u64 parent,
1957                          u64 root_objectid, u64 owner, u64 offset,
1958                          int no_quota)
1959 {
1960         int ret;
1961         struct btrfs_fs_info *fs_info = root->fs_info;
1962
1963         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1964                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1965
1966         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1967                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1968                                         num_bytes,
1969                                         parent, root_objectid, (int)owner,
1970                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
1971         } else {
1972                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1973                                         num_bytes,
1974                                         parent, root_objectid, owner, offset,
1975                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
1976         }
1977         return ret;
1978 }
1979
1980 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1981                                   struct btrfs_root *root,
1982                                   u64 bytenr, u64 num_bytes,
1983                                   u64 parent, u64 root_objectid,
1984                                   u64 owner, u64 offset, int refs_to_add,
1985                                   int no_quota,
1986                                   struct btrfs_delayed_extent_op *extent_op)
1987 {
1988         struct btrfs_fs_info *fs_info = root->fs_info;
1989         struct btrfs_path *path;
1990         struct extent_buffer *leaf;
1991         struct btrfs_extent_item *item;
1992         struct btrfs_key key;
1993         u64 refs;
1994         int ret;
1995         enum btrfs_qgroup_operation_type type = BTRFS_QGROUP_OPER_ADD_EXCL;
1996
1997         path = btrfs_alloc_path();
1998         if (!path)
1999                 return -ENOMEM;
2000
2001         if (!is_fstree(root_objectid) || !root->fs_info->quota_enabled)
2002                 no_quota = 1;
2003
2004         path->reada = 1;
2005         path->leave_spinning = 1;
2006         /* this will setup the path even if it fails to insert the back ref */
2007         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
2008                                            bytenr, num_bytes, parent,
2009                                            root_objectid, owner, offset,
2010                                            refs_to_add, extent_op);
2011         if ((ret < 0 && ret != -EAGAIN) || (!ret && no_quota))
2012                 goto out;
2013         /*
2014          * Ok we were able to insert an inline extent and it appears to be a new
2015          * reference, deal with the qgroup accounting.
2016          */
2017         if (!ret && !no_quota) {
2018                 ASSERT(root->fs_info->quota_enabled);
2019                 leaf = path->nodes[0];
2020                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2021                 item = btrfs_item_ptr(leaf, path->slots[0],
2022                                       struct btrfs_extent_item);
2023                 if (btrfs_extent_refs(leaf, item) > (u64)refs_to_add)
2024                         type = BTRFS_QGROUP_OPER_ADD_SHARED;
2025                 btrfs_release_path(path);
2026
2027                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
2028                                               bytenr, num_bytes, type, 0);
2029                 goto out;
2030         }
2031
2032         /*
2033          * Ok we had -EAGAIN which means we didn't have space to insert and
2034          * inline extent ref, so just update the reference count and add a
2035          * normal backref.
2036          */
2037         leaf = path->nodes[0];
2038         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2039         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2040         refs = btrfs_extent_refs(leaf, item);
2041         if (refs)
2042                 type = BTRFS_QGROUP_OPER_ADD_SHARED;
2043         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2044         if (extent_op)
2045                 __run_delayed_extent_op(extent_op, leaf, item);
2046
2047         btrfs_mark_buffer_dirty(leaf);
2048         btrfs_release_path(path);
2049
2050         if (!no_quota) {
2051                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
2052                                               bytenr, num_bytes, type, 0);
2053                 if (ret)
2054                         goto out;
2055         }
2056
2057         path->reada = 1;
2058         path->leave_spinning = 1;
2059         /* now insert the actual backref */
2060         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2061                                     path, bytenr, parent, root_objectid,
2062                                     owner, offset, refs_to_add);
2063         if (ret)
2064                 btrfs_abort_transaction(trans, root, ret);
2065 out:
2066         btrfs_free_path(path);
2067         return ret;
2068 }
2069
2070 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2071                                 struct btrfs_root *root,
2072                                 struct btrfs_delayed_ref_node *node,
2073                                 struct btrfs_delayed_extent_op *extent_op,
2074                                 int insert_reserved)
2075 {
2076         int ret = 0;
2077         struct btrfs_delayed_data_ref *ref;
2078         struct btrfs_key ins;
2079         u64 parent = 0;
2080         u64 ref_root = 0;
2081         u64 flags = 0;
2082
2083         ins.objectid = node->bytenr;
2084         ins.offset = node->num_bytes;
2085         ins.type = BTRFS_EXTENT_ITEM_KEY;
2086
2087         ref = btrfs_delayed_node_to_data_ref(node);
2088         trace_run_delayed_data_ref(node, ref, node->action);
2089
2090         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2091                 parent = ref->parent;
2092         ref_root = ref->root;
2093
2094         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2095                 if (extent_op)
2096                         flags |= extent_op->flags_to_set;
2097                 ret = alloc_reserved_file_extent(trans, root,
2098                                                  parent, ref_root, flags,
2099                                                  ref->objectid, ref->offset,
2100                                                  &ins, node->ref_mod);
2101         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2102                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2103                                              node->num_bytes, parent,
2104                                              ref_root, ref->objectid,
2105                                              ref->offset, node->ref_mod,
2106                                              node->no_quota, extent_op);
2107         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2108                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2109                                           node->num_bytes, parent,
2110                                           ref_root, ref->objectid,
2111                                           ref->offset, node->ref_mod,
2112                                           extent_op, node->no_quota);
2113         } else {
2114                 BUG();
2115         }
2116         return ret;
2117 }
2118
2119 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2120                                     struct extent_buffer *leaf,
2121                                     struct btrfs_extent_item *ei)
2122 {
2123         u64 flags = btrfs_extent_flags(leaf, ei);
2124         if (extent_op->update_flags) {
2125                 flags |= extent_op->flags_to_set;
2126                 btrfs_set_extent_flags(leaf, ei, flags);
2127         }
2128
2129         if (extent_op->update_key) {
2130                 struct btrfs_tree_block_info *bi;
2131                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2132                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2133                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2134         }
2135 }
2136
2137 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2138                                  struct btrfs_root *root,
2139                                  struct btrfs_delayed_ref_node *node,
2140                                  struct btrfs_delayed_extent_op *extent_op)
2141 {
2142         struct btrfs_key key;
2143         struct btrfs_path *path;
2144         struct btrfs_extent_item *ei;
2145         struct extent_buffer *leaf;
2146         u32 item_size;
2147         int ret;
2148         int err = 0;
2149         int metadata = !extent_op->is_data;
2150
2151         if (trans->aborted)
2152                 return 0;
2153
2154         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2155                 metadata = 0;
2156
2157         path = btrfs_alloc_path();
2158         if (!path)
2159                 return -ENOMEM;
2160
2161         key.objectid = node->bytenr;
2162
2163         if (metadata) {
2164                 key.type = BTRFS_METADATA_ITEM_KEY;
2165                 key.offset = extent_op->level;
2166         } else {
2167                 key.type = BTRFS_EXTENT_ITEM_KEY;
2168                 key.offset = node->num_bytes;
2169         }
2170
2171 again:
2172         path->reada = 1;
2173         path->leave_spinning = 1;
2174         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2175                                 path, 0, 1);
2176         if (ret < 0) {
2177                 err = ret;
2178                 goto out;
2179         }
2180         if (ret > 0) {
2181                 if (metadata) {
2182                         if (path->slots[0] > 0) {
2183                                 path->slots[0]--;
2184                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2185                                                       path->slots[0]);
2186                                 if (key.objectid == node->bytenr &&
2187                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2188                                     key.offset == node->num_bytes)
2189                                         ret = 0;
2190                         }
2191                         if (ret > 0) {
2192                                 btrfs_release_path(path);
2193                                 metadata = 0;
2194
2195                                 key.objectid = node->bytenr;
2196                                 key.offset = node->num_bytes;
2197                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2198                                 goto again;
2199                         }
2200                 } else {
2201                         err = -EIO;
2202                         goto out;
2203                 }
2204         }
2205
2206         leaf = path->nodes[0];
2207         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2208 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2209         if (item_size < sizeof(*ei)) {
2210                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2211                                              path, (u64)-1, 0);
2212                 if (ret < 0) {
2213                         err = ret;
2214                         goto out;
2215                 }
2216                 leaf = path->nodes[0];
2217                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2218         }
2219 #endif
2220         BUG_ON(item_size < sizeof(*ei));
2221         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2222         __run_delayed_extent_op(extent_op, leaf, ei);
2223
2224         btrfs_mark_buffer_dirty(leaf);
2225 out:
2226         btrfs_free_path(path);
2227         return err;
2228 }
2229
2230 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2231                                 struct btrfs_root *root,
2232                                 struct btrfs_delayed_ref_node *node,
2233                                 struct btrfs_delayed_extent_op *extent_op,
2234                                 int insert_reserved)
2235 {
2236         int ret = 0;
2237         struct btrfs_delayed_tree_ref *ref;
2238         struct btrfs_key ins;
2239         u64 parent = 0;
2240         u64 ref_root = 0;
2241         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2242                                                  SKINNY_METADATA);
2243
2244         ref = btrfs_delayed_node_to_tree_ref(node);
2245         trace_run_delayed_tree_ref(node, ref, node->action);
2246
2247         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2248                 parent = ref->parent;
2249         ref_root = ref->root;
2250
2251         ins.objectid = node->bytenr;
2252         if (skinny_metadata) {
2253                 ins.offset = ref->level;
2254                 ins.type = BTRFS_METADATA_ITEM_KEY;
2255         } else {
2256                 ins.offset = node->num_bytes;
2257                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2258         }
2259
2260         BUG_ON(node->ref_mod != 1);
2261         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2262                 BUG_ON(!extent_op || !extent_op->update_flags);
2263                 ret = alloc_reserved_tree_block(trans, root,
2264                                                 parent, ref_root,
2265                                                 extent_op->flags_to_set,
2266                                                 &extent_op->key,
2267                                                 ref->level, &ins,
2268                                                 node->no_quota);
2269         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2270                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2271                                              node->num_bytes, parent, ref_root,
2272                                              ref->level, 0, 1, node->no_quota,
2273                                              extent_op);
2274         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2275                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2276                                           node->num_bytes, parent, ref_root,
2277                                           ref->level, 0, 1, extent_op,
2278                                           node->no_quota);
2279         } else {
2280                 BUG();
2281         }
2282         return ret;
2283 }
2284
2285 /* helper function to actually process a single delayed ref entry */
2286 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2287                                struct btrfs_root *root,
2288                                struct btrfs_delayed_ref_node *node,
2289                                struct btrfs_delayed_extent_op *extent_op,
2290                                int insert_reserved)
2291 {
2292         int ret = 0;
2293
2294         if (trans->aborted) {
2295                 if (insert_reserved)
2296                         btrfs_pin_extent(root, node->bytenr,
2297                                          node->num_bytes, 1);
2298                 return 0;
2299         }
2300
2301         if (btrfs_delayed_ref_is_head(node)) {
2302                 struct btrfs_delayed_ref_head *head;
2303                 /*
2304                  * we've hit the end of the chain and we were supposed
2305                  * to insert this extent into the tree.  But, it got
2306                  * deleted before we ever needed to insert it, so all
2307                  * we have to do is clean up the accounting
2308                  */
2309                 BUG_ON(extent_op);
2310                 head = btrfs_delayed_node_to_head(node);
2311                 trace_run_delayed_ref_head(node, head, node->action);
2312
2313                 if (insert_reserved) {
2314                         btrfs_pin_extent(root, node->bytenr,
2315                                          node->num_bytes, 1);
2316                         if (head->is_data) {
2317                                 ret = btrfs_del_csums(trans, root,
2318                                                       node->bytenr,
2319                                                       node->num_bytes);
2320                         }
2321                 }
2322                 return ret;
2323         }
2324
2325         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2326             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2327                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2328                                            insert_reserved);
2329         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2330                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2331                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2332                                            insert_reserved);
2333         else
2334                 BUG();
2335         return ret;
2336 }
2337
2338 static noinline struct btrfs_delayed_ref_node *
2339 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2340 {
2341         struct rb_node *node;
2342         struct btrfs_delayed_ref_node *ref, *last = NULL;;
2343
2344         /*
2345          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2346          * this prevents ref count from going down to zero when
2347          * there still are pending delayed ref.
2348          */
2349         node = rb_first(&head->ref_root);
2350         while (node) {
2351                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2352                                 rb_node);
2353                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2354                         return ref;
2355                 else if (last == NULL)
2356                         last = ref;
2357                 node = rb_next(node);
2358         }
2359         return last;
2360 }
2361
2362 /*
2363  * Returns 0 on success or if called with an already aborted transaction.
2364  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2365  */
2366 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2367                                              struct btrfs_root *root,
2368                                              unsigned long nr)
2369 {
2370         struct btrfs_delayed_ref_root *delayed_refs;
2371         struct btrfs_delayed_ref_node *ref;
2372         struct btrfs_delayed_ref_head *locked_ref = NULL;
2373         struct btrfs_delayed_extent_op *extent_op;
2374         struct btrfs_fs_info *fs_info = root->fs_info;
2375         ktime_t start = ktime_get();
2376         int ret;
2377         unsigned long count = 0;
2378         unsigned long actual_count = 0;
2379         int must_insert_reserved = 0;
2380
2381         delayed_refs = &trans->transaction->delayed_refs;
2382         while (1) {
2383                 if (!locked_ref) {
2384                         if (count >= nr)
2385                                 break;
2386
2387                         spin_lock(&delayed_refs->lock);
2388                         locked_ref = btrfs_select_ref_head(trans);
2389                         if (!locked_ref) {
2390                                 spin_unlock(&delayed_refs->lock);
2391                                 break;
2392                         }
2393
2394                         /* grab the lock that says we are going to process
2395                          * all the refs for this head */
2396                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2397                         spin_unlock(&delayed_refs->lock);
2398                         /*
2399                          * we may have dropped the spin lock to get the head
2400                          * mutex lock, and that might have given someone else
2401                          * time to free the head.  If that's true, it has been
2402                          * removed from our list and we can move on.
2403                          */
2404                         if (ret == -EAGAIN) {
2405                                 locked_ref = NULL;
2406                                 count++;
2407                                 continue;
2408                         }
2409                 }
2410
2411                 /*
2412                  * We need to try and merge add/drops of the same ref since we
2413                  * can run into issues with relocate dropping the implicit ref
2414                  * and then it being added back again before the drop can
2415                  * finish.  If we merged anything we need to re-loop so we can
2416                  * get a good ref.
2417                  */
2418                 spin_lock(&locked_ref->lock);
2419                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2420                                          locked_ref);
2421
2422                 /*
2423                  * locked_ref is the head node, so we have to go one
2424                  * node back for any delayed ref updates
2425                  */
2426                 ref = select_delayed_ref(locked_ref);
2427
2428                 if (ref && ref->seq &&
2429                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2430                         spin_unlock(&locked_ref->lock);
2431                         btrfs_delayed_ref_unlock(locked_ref);
2432                         spin_lock(&delayed_refs->lock);
2433                         locked_ref->processing = 0;
2434                         delayed_refs->num_heads_ready++;
2435                         spin_unlock(&delayed_refs->lock);
2436                         locked_ref = NULL;
2437                         cond_resched();
2438                         count++;
2439                         continue;
2440                 }
2441
2442                 /*
2443                  * record the must insert reserved flag before we
2444                  * drop the spin lock.
2445                  */
2446                 must_insert_reserved = locked_ref->must_insert_reserved;
2447                 locked_ref->must_insert_reserved = 0;
2448
2449                 extent_op = locked_ref->extent_op;
2450                 locked_ref->extent_op = NULL;
2451
2452                 if (!ref) {
2453
2454
2455                         /* All delayed refs have been processed, Go ahead
2456                          * and send the head node to run_one_delayed_ref,
2457                          * so that any accounting fixes can happen
2458                          */
2459                         ref = &locked_ref->node;
2460
2461                         if (extent_op && must_insert_reserved) {
2462                                 btrfs_free_delayed_extent_op(extent_op);
2463                                 extent_op = NULL;
2464                         }
2465
2466                         if (extent_op) {
2467                                 spin_unlock(&locked_ref->lock);
2468                                 ret = run_delayed_extent_op(trans, root,
2469                                                             ref, extent_op);
2470                                 btrfs_free_delayed_extent_op(extent_op);
2471
2472                                 if (ret) {
2473                                         /*
2474                                          * Need to reset must_insert_reserved if
2475                                          * there was an error so the abort stuff
2476                                          * can cleanup the reserved space
2477                                          * properly.
2478                                          */
2479                                         if (must_insert_reserved)
2480                                                 locked_ref->must_insert_reserved = 1;
2481                                         locked_ref->processing = 0;
2482                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2483                                         btrfs_delayed_ref_unlock(locked_ref);
2484                                         return ret;
2485                                 }
2486                                 continue;
2487                         }
2488
2489                         /*
2490                          * Need to drop our head ref lock and re-aqcuire the
2491                          * delayed ref lock and then re-check to make sure
2492                          * nobody got added.
2493                          */
2494                         spin_unlock(&locked_ref->lock);
2495                         spin_lock(&delayed_refs->lock);
2496                         spin_lock(&locked_ref->lock);
2497                         if (rb_first(&locked_ref->ref_root) ||
2498                             locked_ref->extent_op) {
2499                                 spin_unlock(&locked_ref->lock);
2500                                 spin_unlock(&delayed_refs->lock);
2501                                 continue;
2502                         }
2503                         ref->in_tree = 0;
2504                         delayed_refs->num_heads--;
2505                         rb_erase(&locked_ref->href_node,
2506                                  &delayed_refs->href_root);
2507                         spin_unlock(&delayed_refs->lock);
2508                 } else {
2509                         actual_count++;
2510                         ref->in_tree = 0;
2511                         rb_erase(&ref->rb_node, &locked_ref->ref_root);
2512                 }
2513                 atomic_dec(&delayed_refs->num_entries);
2514
2515                 if (!btrfs_delayed_ref_is_head(ref)) {
2516                         /*
2517                          * when we play the delayed ref, also correct the
2518                          * ref_mod on head
2519                          */
2520                         switch (ref->action) {
2521                         case BTRFS_ADD_DELAYED_REF:
2522                         case BTRFS_ADD_DELAYED_EXTENT:
2523                                 locked_ref->node.ref_mod -= ref->ref_mod;
2524                                 break;
2525                         case BTRFS_DROP_DELAYED_REF:
2526                                 locked_ref->node.ref_mod += ref->ref_mod;
2527                                 break;
2528                         default:
2529                                 WARN_ON(1);
2530                         }
2531                 }
2532                 spin_unlock(&locked_ref->lock);
2533
2534                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2535                                           must_insert_reserved);
2536
2537                 btrfs_free_delayed_extent_op(extent_op);
2538                 if (ret) {
2539                         locked_ref->processing = 0;
2540                         btrfs_delayed_ref_unlock(locked_ref);
2541                         btrfs_put_delayed_ref(ref);
2542                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2543                         return ret;
2544                 }
2545
2546                 /*
2547                  * If this node is a head, that means all the refs in this head
2548                  * have been dealt with, and we will pick the next head to deal
2549                  * with, so we must unlock the head and drop it from the cluster
2550                  * list before we release it.
2551                  */
2552                 if (btrfs_delayed_ref_is_head(ref)) {
2553                         btrfs_delayed_ref_unlock(locked_ref);
2554                         locked_ref = NULL;
2555                 }
2556                 btrfs_put_delayed_ref(ref);
2557                 count++;
2558                 cond_resched();
2559         }
2560
2561         /*
2562          * We don't want to include ref heads since we can have empty ref heads
2563          * and those will drastically skew our runtime down since we just do
2564          * accounting, no actual extent tree updates.
2565          */
2566         if (actual_count > 0) {
2567                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2568                 u64 avg;
2569
2570                 /*
2571                  * We weigh the current average higher than our current runtime
2572                  * to avoid large swings in the average.
2573                  */
2574                 spin_lock(&delayed_refs->lock);
2575                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2576                 avg = div64_u64(avg, 4);
2577                 fs_info->avg_delayed_ref_runtime = avg;
2578                 spin_unlock(&delayed_refs->lock);
2579         }
2580         return 0;
2581 }
2582
2583 #ifdef SCRAMBLE_DELAYED_REFS
2584 /*
2585  * Normally delayed refs get processed in ascending bytenr order. This
2586  * correlates in most cases to the order added. To expose dependencies on this
2587  * order, we start to process the tree in the middle instead of the beginning
2588  */
2589 static u64 find_middle(struct rb_root *root)
2590 {
2591         struct rb_node *n = root->rb_node;
2592         struct btrfs_delayed_ref_node *entry;
2593         int alt = 1;
2594         u64 middle;
2595         u64 first = 0, last = 0;
2596
2597         n = rb_first(root);
2598         if (n) {
2599                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2600                 first = entry->bytenr;
2601         }
2602         n = rb_last(root);
2603         if (n) {
2604                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2605                 last = entry->bytenr;
2606         }
2607         n = root->rb_node;
2608
2609         while (n) {
2610                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2611                 WARN_ON(!entry->in_tree);
2612
2613                 middle = entry->bytenr;
2614
2615                 if (alt)
2616                         n = n->rb_left;
2617                 else
2618                         n = n->rb_right;
2619
2620                 alt = 1 - alt;
2621         }
2622         return middle;
2623 }
2624 #endif
2625
2626 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2627 {
2628         u64 num_bytes;
2629
2630         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2631                              sizeof(struct btrfs_extent_inline_ref));
2632         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2633                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2634
2635         /*
2636          * We don't ever fill up leaves all the way so multiply by 2 just to be
2637          * closer to what we're really going to want to ouse.
2638          */
2639         return div64_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2640 }
2641
2642 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2643                                        struct btrfs_root *root)
2644 {
2645         struct btrfs_block_rsv *global_rsv;
2646         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2647         u64 num_bytes;
2648         int ret = 0;
2649
2650         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2651         num_heads = heads_to_leaves(root, num_heads);
2652         if (num_heads > 1)
2653                 num_bytes += (num_heads - 1) * root->leafsize;
2654         num_bytes <<= 1;
2655         global_rsv = &root->fs_info->global_block_rsv;
2656
2657         /*
2658          * If we can't allocate any more chunks lets make sure we have _lots_ of
2659          * wiggle room since running delayed refs can create more delayed refs.
2660          */
2661         if (global_rsv->space_info->full)
2662                 num_bytes <<= 1;
2663
2664         spin_lock(&global_rsv->lock);
2665         if (global_rsv->reserved <= num_bytes)
2666                 ret = 1;
2667         spin_unlock(&global_rsv->lock);
2668         return ret;
2669 }
2670
2671 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2672                                        struct btrfs_root *root)
2673 {
2674         struct btrfs_fs_info *fs_info = root->fs_info;
2675         u64 num_entries =
2676                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2677         u64 avg_runtime;
2678         u64 val;
2679
2680         smp_mb();
2681         avg_runtime = fs_info->avg_delayed_ref_runtime;
2682         val = num_entries * avg_runtime;
2683         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2684                 return 1;
2685         if (val >= NSEC_PER_SEC / 2)
2686                 return 2;
2687
2688         return btrfs_check_space_for_delayed_refs(trans, root);
2689 }
2690
2691 struct async_delayed_refs {
2692         struct btrfs_root *root;
2693         int count;
2694         int error;
2695         int sync;
2696         struct completion wait;
2697         struct btrfs_work work;
2698 };
2699
2700 static void delayed_ref_async_start(struct btrfs_work *work)
2701 {
2702         struct async_delayed_refs *async;
2703         struct btrfs_trans_handle *trans;
2704         int ret;
2705
2706         async = container_of(work, struct async_delayed_refs, work);
2707
2708         trans = btrfs_join_transaction(async->root);
2709         if (IS_ERR(trans)) {
2710                 async->error = PTR_ERR(trans);
2711                 goto done;
2712         }
2713
2714         /*
2715          * trans->sync means that when we call end_transaciton, we won't
2716          * wait on delayed refs
2717          */
2718         trans->sync = true;
2719         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2720         if (ret)
2721                 async->error = ret;
2722
2723         ret = btrfs_end_transaction(trans, async->root);
2724         if (ret && !async->error)
2725                 async->error = ret;
2726 done:
2727         if (async->sync)
2728                 complete(&async->wait);
2729         else
2730                 kfree(async);
2731 }
2732
2733 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2734                                  unsigned long count, int wait)
2735 {
2736         struct async_delayed_refs *async;
2737         int ret;
2738
2739         async = kmalloc(sizeof(*async), GFP_NOFS);
2740         if (!async)
2741                 return -ENOMEM;
2742
2743         async->root = root->fs_info->tree_root;
2744         async->count = count;
2745         async->error = 0;
2746         if (wait)
2747                 async->sync = 1;
2748         else
2749                 async->sync = 0;
2750         init_completion(&async->wait);
2751
2752         btrfs_init_work(&async->work, delayed_ref_async_start,
2753                         NULL, NULL);
2754
2755         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2756
2757         if (wait) {
2758                 wait_for_completion(&async->wait);
2759                 ret = async->error;
2760                 kfree(async);
2761                 return ret;
2762         }
2763         return 0;
2764 }
2765
2766 /*
2767  * this starts processing the delayed reference count updates and
2768  * extent insertions we have queued up so far.  count can be
2769  * 0, which means to process everything in the tree at the start
2770  * of the run (but not newly added entries), or it can be some target
2771  * number you'd like to process.
2772  *
2773  * Returns 0 on success or if called with an aborted transaction
2774  * Returns <0 on error and aborts the transaction
2775  */
2776 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2777                            struct btrfs_root *root, unsigned long count)
2778 {
2779         struct rb_node *node;
2780         struct btrfs_delayed_ref_root *delayed_refs;
2781         struct btrfs_delayed_ref_head *head;
2782         int ret;
2783         int run_all = count == (unsigned long)-1;
2784         int run_most = 0;
2785
2786         /* We'll clean this up in btrfs_cleanup_transaction */
2787         if (trans->aborted)
2788                 return 0;
2789
2790         if (root == root->fs_info->extent_root)
2791                 root = root->fs_info->tree_root;
2792
2793         delayed_refs = &trans->transaction->delayed_refs;
2794         if (count == 0) {
2795                 count = atomic_read(&delayed_refs->num_entries) * 2;
2796                 run_most = 1;
2797         }
2798
2799 again:
2800 #ifdef SCRAMBLE_DELAYED_REFS
2801         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2802 #endif
2803         ret = __btrfs_run_delayed_refs(trans, root, count);
2804         if (ret < 0) {
2805                 btrfs_abort_transaction(trans, root, ret);
2806                 return ret;
2807         }
2808
2809         if (run_all) {
2810                 if (!list_empty(&trans->new_bgs))
2811                         btrfs_create_pending_block_groups(trans, root);
2812
2813                 spin_lock(&delayed_refs->lock);
2814                 node = rb_first(&delayed_refs->href_root);
2815                 if (!node) {
2816                         spin_unlock(&delayed_refs->lock);
2817                         goto out;
2818                 }
2819                 count = (unsigned long)-1;
2820
2821                 while (node) {
2822                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2823                                         href_node);
2824                         if (btrfs_delayed_ref_is_head(&head->node)) {
2825                                 struct btrfs_delayed_ref_node *ref;
2826
2827                                 ref = &head->node;
2828                                 atomic_inc(&ref->refs);
2829
2830                                 spin_unlock(&delayed_refs->lock);
2831                                 /*
2832                                  * Mutex was contended, block until it's
2833                                  * released and try again
2834                                  */
2835                                 mutex_lock(&head->mutex);
2836                                 mutex_unlock(&head->mutex);
2837
2838                                 btrfs_put_delayed_ref(ref);
2839                                 cond_resched();
2840                                 goto again;
2841                         } else {
2842                                 WARN_ON(1);
2843                         }
2844                         node = rb_next(node);
2845                 }
2846                 spin_unlock(&delayed_refs->lock);
2847                 cond_resched();
2848                 goto again;
2849         }
2850 out:
2851         ret = btrfs_delayed_qgroup_accounting(trans, root->fs_info);
2852         if (ret)
2853                 return ret;
2854         assert_qgroups_uptodate(trans);
2855         return 0;
2856 }
2857
2858 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2859                                 struct btrfs_root *root,
2860                                 u64 bytenr, u64 num_bytes, u64 flags,
2861                                 int level, int is_data)
2862 {
2863         struct btrfs_delayed_extent_op *extent_op;
2864         int ret;
2865
2866         extent_op = btrfs_alloc_delayed_extent_op();
2867         if (!extent_op)
2868                 return -ENOMEM;
2869
2870         extent_op->flags_to_set = flags;
2871         extent_op->update_flags = 1;
2872         extent_op->update_key = 0;
2873         extent_op->is_data = is_data ? 1 : 0;
2874         extent_op->level = level;
2875
2876         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2877                                           num_bytes, extent_op);
2878         if (ret)
2879                 btrfs_free_delayed_extent_op(extent_op);
2880         return ret;
2881 }
2882
2883 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2884                                       struct btrfs_root *root,
2885                                       struct btrfs_path *path,
2886                                       u64 objectid, u64 offset, u64 bytenr)
2887 {
2888         struct btrfs_delayed_ref_head *head;
2889         struct btrfs_delayed_ref_node *ref;
2890         struct btrfs_delayed_data_ref *data_ref;
2891         struct btrfs_delayed_ref_root *delayed_refs;
2892         struct rb_node *node;
2893         int ret = 0;
2894
2895         delayed_refs = &trans->transaction->delayed_refs;
2896         spin_lock(&delayed_refs->lock);
2897         head = btrfs_find_delayed_ref_head(trans, bytenr);
2898         if (!head) {
2899                 spin_unlock(&delayed_refs->lock);
2900                 return 0;
2901         }
2902
2903         if (!mutex_trylock(&head->mutex)) {
2904                 atomic_inc(&head->node.refs);
2905                 spin_unlock(&delayed_refs->lock);
2906
2907                 btrfs_release_path(path);
2908
2909                 /*
2910                  * Mutex was contended, block until it's released and let
2911                  * caller try again
2912                  */
2913                 mutex_lock(&head->mutex);
2914                 mutex_unlock(&head->mutex);
2915                 btrfs_put_delayed_ref(&head->node);
2916                 return -EAGAIN;
2917         }
2918         spin_unlock(&delayed_refs->lock);
2919
2920         spin_lock(&head->lock);
2921         node = rb_first(&head->ref_root);
2922         while (node) {
2923                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2924                 node = rb_next(node);
2925
2926                 /* If it's a shared ref we know a cross reference exists */
2927                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2928                         ret = 1;
2929                         break;
2930                 }
2931
2932                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2933
2934                 /*
2935                  * If our ref doesn't match the one we're currently looking at
2936                  * then we have a cross reference.
2937                  */
2938                 if (data_ref->root != root->root_key.objectid ||
2939                     data_ref->objectid != objectid ||
2940                     data_ref->offset != offset) {
2941                         ret = 1;
2942                         break;
2943                 }
2944         }
2945         spin_unlock(&head->lock);
2946         mutex_unlock(&head->mutex);
2947         return ret;
2948 }
2949
2950 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2951                                         struct btrfs_root *root,
2952                                         struct btrfs_path *path,
2953                                         u64 objectid, u64 offset, u64 bytenr)
2954 {
2955         struct btrfs_root *extent_root = root->fs_info->extent_root;
2956         struct extent_buffer *leaf;
2957         struct btrfs_extent_data_ref *ref;
2958         struct btrfs_extent_inline_ref *iref;
2959         struct btrfs_extent_item *ei;
2960         struct btrfs_key key;
2961         u32 item_size;
2962         int ret;
2963
2964         key.objectid = bytenr;
2965         key.offset = (u64)-1;
2966         key.type = BTRFS_EXTENT_ITEM_KEY;
2967
2968         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2969         if (ret < 0)
2970                 goto out;
2971         BUG_ON(ret == 0); /* Corruption */
2972
2973         ret = -ENOENT;
2974         if (path->slots[0] == 0)
2975                 goto out;
2976
2977         path->slots[0]--;
2978         leaf = path->nodes[0];
2979         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2980
2981         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2982                 goto out;
2983
2984         ret = 1;
2985         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2986 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2987         if (item_size < sizeof(*ei)) {
2988                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2989                 goto out;
2990         }
2991 #endif
2992         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2993
2994         if (item_size != sizeof(*ei) +
2995             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2996                 goto out;
2997
2998         if (btrfs_extent_generation(leaf, ei) <=
2999             btrfs_root_last_snapshot(&root->root_item))
3000                 goto out;
3001
3002         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3003         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3004             BTRFS_EXTENT_DATA_REF_KEY)
3005                 goto out;
3006
3007         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3008         if (btrfs_extent_refs(leaf, ei) !=
3009             btrfs_extent_data_ref_count(leaf, ref) ||
3010             btrfs_extent_data_ref_root(leaf, ref) !=
3011             root->root_key.objectid ||
3012             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3013             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3014                 goto out;
3015
3016         ret = 0;
3017 out:
3018         return ret;
3019 }
3020
3021 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3022                           struct btrfs_root *root,
3023                           u64 objectid, u64 offset, u64 bytenr)
3024 {
3025         struct btrfs_path *path;
3026         int ret;
3027         int ret2;
3028
3029         path = btrfs_alloc_path();
3030         if (!path)
3031                 return -ENOENT;
3032
3033         do {
3034                 ret = check_committed_ref(trans, root, path, objectid,
3035                                           offset, bytenr);
3036                 if (ret && ret != -ENOENT)
3037                         goto out;
3038
3039                 ret2 = check_delayed_ref(trans, root, path, objectid,
3040                                          offset, bytenr);
3041         } while (ret2 == -EAGAIN);
3042
3043         if (ret2 && ret2 != -ENOENT) {
3044                 ret = ret2;
3045                 goto out;
3046         }
3047
3048         if (ret != -ENOENT || ret2 != -ENOENT)
3049                 ret = 0;
3050 out:
3051         btrfs_free_path(path);
3052         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3053                 WARN_ON(ret > 0);
3054         return ret;
3055 }
3056
3057 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3058                            struct btrfs_root *root,
3059                            struct extent_buffer *buf,
3060                            int full_backref, int inc, int no_quota)
3061 {
3062         u64 bytenr;
3063         u64 num_bytes;
3064         u64 parent;
3065         u64 ref_root;
3066         u32 nritems;
3067         struct btrfs_key key;
3068         struct btrfs_file_extent_item *fi;
3069         int i;
3070         int level;
3071         int ret = 0;
3072         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3073                             u64, u64, u64, u64, u64, u64, int);
3074
3075 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3076         if (unlikely(test_bit(BTRFS_ROOT_DUMMY_ROOT, &root->state)))
3077                 return 0;
3078 #endif
3079         ref_root = btrfs_header_owner(buf);
3080         nritems = btrfs_header_nritems(buf);
3081         level = btrfs_header_level(buf);
3082
3083         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3084                 return 0;
3085
3086         if (inc)
3087                 process_func = btrfs_inc_extent_ref;
3088         else
3089                 process_func = btrfs_free_extent;
3090
3091         if (full_backref)
3092                 parent = buf->start;
3093         else
3094                 parent = 0;
3095
3096         for (i = 0; i < nritems; i++) {
3097                 if (level == 0) {
3098                         btrfs_item_key_to_cpu(buf, &key, i);
3099                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
3100                                 continue;
3101                         fi = btrfs_item_ptr(buf, i,
3102                                             struct btrfs_file_extent_item);
3103                         if (btrfs_file_extent_type(buf, fi) ==
3104                             BTRFS_FILE_EXTENT_INLINE)
3105                                 continue;
3106                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3107                         if (bytenr == 0)
3108                                 continue;
3109
3110                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3111                         key.offset -= btrfs_file_extent_offset(buf, fi);
3112                         ret = process_func(trans, root, bytenr, num_bytes,
3113                                            parent, ref_root, key.objectid,
3114                                            key.offset, no_quota);
3115                         if (ret)
3116                                 goto fail;
3117                 } else {
3118                         bytenr = btrfs_node_blockptr(buf, i);
3119                         num_bytes = btrfs_level_size(root, level - 1);
3120                         ret = process_func(trans, root, bytenr, num_bytes,
3121                                            parent, ref_root, level - 1, 0,
3122                                            no_quota);
3123                         if (ret)
3124                                 goto fail;
3125                 }
3126         }
3127         return 0;
3128 fail:
3129         return ret;
3130 }
3131
3132 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3133                   struct extent_buffer *buf, int full_backref, int no_quota)
3134 {
3135         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, no_quota);
3136 }
3137
3138 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3139                   struct extent_buffer *buf, int full_backref, int no_quota)
3140 {
3141         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, no_quota);
3142 }
3143
3144 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3145                                  struct btrfs_root *root,
3146                                  struct btrfs_path *path,
3147                                  struct btrfs_block_group_cache *cache)
3148 {
3149         int ret;
3150         struct btrfs_root *extent_root = root->fs_info->extent_root;
3151         unsigned long bi;
3152         struct extent_buffer *leaf;
3153
3154         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3155         if (ret < 0)
3156                 goto fail;
3157         BUG_ON(ret); /* Corruption */
3158
3159         leaf = path->nodes[0];
3160         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3161         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3162         btrfs_mark_buffer_dirty(leaf);
3163         btrfs_release_path(path);
3164 fail:
3165         if (ret) {
3166                 btrfs_abort_transaction(trans, root, ret);
3167                 return ret;
3168         }
3169         return 0;
3170
3171 }
3172
3173 static struct btrfs_block_group_cache *
3174 next_block_group(struct btrfs_root *root,
3175                  struct btrfs_block_group_cache *cache)
3176 {
3177         struct rb_node *node;
3178         spin_lock(&root->fs_info->block_group_cache_lock);
3179         node = rb_next(&cache->cache_node);
3180         btrfs_put_block_group(cache);
3181         if (node) {
3182                 cache = rb_entry(node, struct btrfs_block_group_cache,
3183                                  cache_node);
3184                 btrfs_get_block_group(cache);
3185         } else
3186                 cache = NULL;
3187         spin_unlock(&root->fs_info->block_group_cache_lock);
3188         return cache;
3189 }
3190
3191 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3192                             struct btrfs_trans_handle *trans,
3193                             struct btrfs_path *path)
3194 {
3195         struct btrfs_root *root = block_group->fs_info->tree_root;
3196         struct inode *inode = NULL;
3197         u64 alloc_hint = 0;
3198         int dcs = BTRFS_DC_ERROR;
3199         int num_pages = 0;
3200         int retries = 0;
3201         int ret = 0;
3202
3203         /*
3204          * If this block group is smaller than 100 megs don't bother caching the
3205          * block group.
3206          */
3207         if (block_group->key.offset < (100 * 1024 * 1024)) {
3208                 spin_lock(&block_group->lock);
3209                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3210                 spin_unlock(&block_group->lock);
3211                 return 0;
3212         }
3213
3214 again:
3215         inode = lookup_free_space_inode(root, block_group, path);
3216         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3217                 ret = PTR_ERR(inode);
3218                 btrfs_release_path(path);
3219                 goto out;
3220         }
3221
3222         if (IS_ERR(inode)) {
3223                 BUG_ON(retries);
3224                 retries++;
3225
3226                 if (block_group->ro)
3227                         goto out_free;
3228
3229                 ret = create_free_space_inode(root, trans, block_group, path);
3230                 if (ret)
3231                         goto out_free;
3232                 goto again;
3233         }
3234
3235         /* We've already setup this transaction, go ahead and exit */
3236         if (block_group->cache_generation == trans->transid &&
3237             i_size_read(inode)) {
3238                 dcs = BTRFS_DC_SETUP;
3239                 goto out_put;
3240         }
3241
3242         /*
3243          * We want to set the generation to 0, that way if anything goes wrong
3244          * from here on out we know not to trust this cache when we load up next
3245          * time.
3246          */
3247         BTRFS_I(inode)->generation = 0;
3248         ret = btrfs_update_inode(trans, root, inode);
3249         WARN_ON(ret);
3250
3251         if (i_size_read(inode) > 0) {
3252                 ret = btrfs_check_trunc_cache_free_space(root,
3253                                         &root->fs_info->global_block_rsv);
3254                 if (ret)
3255                         goto out_put;
3256
3257                 ret = btrfs_truncate_free_space_cache(root, trans, inode);
3258                 if (ret)
3259                         goto out_put;
3260         }
3261
3262         spin_lock(&block_group->lock);
3263         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3264             !btrfs_test_opt(root, SPACE_CACHE) ||
3265             block_group->delalloc_bytes) {
3266                 /*
3267                  * don't bother trying to write stuff out _if_
3268                  * a) we're not cached,
3269                  * b) we're with nospace_cache mount option.
3270                  */
3271                 dcs = BTRFS_DC_WRITTEN;
3272                 spin_unlock(&block_group->lock);
3273                 goto out_put;
3274         }
3275         spin_unlock(&block_group->lock);
3276
3277         /*
3278          * Try to preallocate enough space based on how big the block group is.
3279          * Keep in mind this has to include any pinned space which could end up
3280          * taking up quite a bit since it's not folded into the other space
3281          * cache.
3282          */
3283         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3284         if (!num_pages)
3285                 num_pages = 1;
3286
3287         num_pages *= 16;
3288         num_pages *= PAGE_CACHE_SIZE;
3289
3290         ret = btrfs_check_data_free_space(inode, num_pages);
3291         if (ret)
3292                 goto out_put;
3293
3294         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3295                                               num_pages, num_pages,
3296                                               &alloc_hint);
3297         if (!ret)
3298                 dcs = BTRFS_DC_SETUP;
3299         btrfs_free_reserved_data_space(inode, num_pages);
3300
3301 out_put:
3302         iput(inode);
3303 out_free:
3304         btrfs_release_path(path);
3305 out:
3306         spin_lock(&block_group->lock);
3307         if (!ret && dcs == BTRFS_DC_SETUP)
3308                 block_group->cache_generation = trans->transid;
3309         block_group->disk_cache_state = dcs;
3310         spin_unlock(&block_group->lock);
3311
3312         return ret;
3313 }
3314
3315 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3316                                    struct btrfs_root *root)
3317 {
3318         struct btrfs_block_group_cache *cache;
3319         int err = 0;
3320         struct btrfs_path *path;
3321         u64 last = 0;
3322
3323         path = btrfs_alloc_path();
3324         if (!path)
3325                 return -ENOMEM;
3326
3327 again:
3328         while (1) {
3329                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3330                 while (cache) {
3331                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3332                                 break;
3333                         cache = next_block_group(root, cache);
3334                 }
3335                 if (!cache) {
3336                         if (last == 0)
3337                                 break;
3338                         last = 0;
3339                         continue;
3340                 }
3341                 err = cache_save_setup(cache, trans, path);
3342                 last = cache->key.objectid + cache->key.offset;
3343                 btrfs_put_block_group(cache);
3344         }
3345
3346         while (1) {
3347                 if (last == 0) {
3348                         err = btrfs_run_delayed_refs(trans, root,
3349                                                      (unsigned long)-1);
3350                         if (err) /* File system offline */
3351                                 goto out;
3352                 }
3353
3354                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3355                 while (cache) {
3356                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3357                                 btrfs_put_block_group(cache);
3358                                 goto again;
3359                         }
3360
3361                         if (cache->dirty)
3362                                 break;
3363                         cache = next_block_group(root, cache);
3364                 }
3365                 if (!cache) {
3366                         if (last == 0)
3367                                 break;
3368                         last = 0;
3369                         continue;
3370                 }
3371
3372                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3373                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3374                 cache->dirty = 0;
3375                 last = cache->key.objectid + cache->key.offset;
3376
3377                 err = write_one_cache_group(trans, root, path, cache);
3378                 btrfs_put_block_group(cache);
3379                 if (err) /* File system offline */
3380                         goto out;
3381         }
3382
3383         while (1) {
3384                 /*
3385                  * I don't think this is needed since we're just marking our
3386                  * preallocated extent as written, but just in case it can't
3387                  * hurt.
3388                  */
3389                 if (last == 0) {
3390                         err = btrfs_run_delayed_refs(trans, root,
3391                                                      (unsigned long)-1);
3392                         if (err) /* File system offline */
3393                                 goto out;
3394                 }
3395
3396                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3397                 while (cache) {
3398                         /*
3399                          * Really this shouldn't happen, but it could if we
3400                          * couldn't write the entire preallocated extent and
3401                          * splitting the extent resulted in a new block.
3402                          */
3403                         if (cache->dirty) {
3404                                 btrfs_put_block_group(cache);
3405                                 goto again;
3406                         }
3407                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3408                                 break;
3409                         cache = next_block_group(root, cache);
3410                 }
3411                 if (!cache) {
3412                         if (last == 0)
3413                                 break;
3414                         last = 0;
3415                         continue;
3416                 }
3417
3418                 err = btrfs_write_out_cache(root, trans, cache, path);
3419
3420                 /*
3421                  * If we didn't have an error then the cache state is still
3422                  * NEED_WRITE, so we can set it to WRITTEN.
3423                  */
3424                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3425                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3426                 last = cache->key.objectid + cache->key.offset;
3427                 btrfs_put_block_group(cache);
3428         }
3429 out:
3430
3431         btrfs_free_path(path);
3432         return err;
3433 }
3434
3435 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3436 {
3437         struct btrfs_block_group_cache *block_group;
3438         int readonly = 0;
3439
3440         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3441         if (!block_group || block_group->ro)
3442                 readonly = 1;
3443         if (block_group)
3444                 btrfs_put_block_group(block_group);
3445         return readonly;
3446 }
3447
3448 static const char *alloc_name(u64 flags)
3449 {
3450         switch (flags) {
3451         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3452                 return "mixed";
3453         case BTRFS_BLOCK_GROUP_METADATA:
3454                 return "metadata";
3455         case BTRFS_BLOCK_GROUP_DATA:
3456                 return "data";
3457         case BTRFS_BLOCK_GROUP_SYSTEM:
3458                 return "system";
3459         default:
3460                 WARN_ON(1);
3461                 return "invalid-combination";
3462         };
3463 }
3464
3465 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3466                              u64 total_bytes, u64 bytes_used,
3467                              struct btrfs_space_info **space_info)
3468 {
3469         struct btrfs_space_info *found;
3470         int i;
3471         int factor;
3472         int ret;
3473
3474         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3475                      BTRFS_BLOCK_GROUP_RAID10))
3476                 factor = 2;
3477         else
3478                 factor = 1;
3479
3480         found = __find_space_info(info, flags);
3481         if (found) {
3482                 spin_lock(&found->lock);
3483                 found->total_bytes += total_bytes;
3484                 found->disk_total += total_bytes * factor;
3485                 found->bytes_used += bytes_used;
3486                 found->disk_used += bytes_used * factor;
3487                 found->full = 0;
3488                 spin_unlock(&found->lock);
3489                 *space_info = found;
3490                 return 0;
3491         }
3492         found = kzalloc(sizeof(*found), GFP_NOFS);
3493         if (!found)
3494                 return -ENOMEM;
3495
3496         ret = percpu_counter_init(&found->total_bytes_pinned, 0);
3497         if (ret) {
3498                 kfree(found);
3499                 return ret;
3500         }
3501
3502         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3503                 INIT_LIST_HEAD(&found->block_groups[i]);
3504         init_rwsem(&found->groups_sem);
3505         spin_lock_init(&found->lock);
3506         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3507         found->total_bytes = total_bytes;
3508         found->disk_total = total_bytes * factor;
3509         found->bytes_used = bytes_used;
3510         found->disk_used = bytes_used * factor;
3511         found->bytes_pinned = 0;
3512         found->bytes_reserved = 0;
3513         found->bytes_readonly = 0;
3514         found->bytes_may_use = 0;
3515         found->full = 0;
3516         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3517         found->chunk_alloc = 0;
3518         found->flush = 0;
3519         init_waitqueue_head(&found->wait);
3520
3521         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3522                                     info->space_info_kobj, "%s",
3523                                     alloc_name(found->flags));
3524         if (ret) {
3525                 kfree(found);
3526                 return ret;
3527         }
3528
3529         *space_info = found;
3530         list_add_rcu(&found->list, &info->space_info);
3531         if (flags & BTRFS_BLOCK_GROUP_DATA)
3532                 info->data_sinfo = found;
3533
3534         return ret;
3535 }
3536
3537 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3538 {
3539         u64 extra_flags = chunk_to_extended(flags) &
3540                                 BTRFS_EXTENDED_PROFILE_MASK;
3541
3542         write_seqlock(&fs_info->profiles_lock);
3543         if (flags & BTRFS_BLOCK_GROUP_DATA)
3544                 fs_info->avail_data_alloc_bits |= extra_flags;
3545         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3546                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3547         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3548                 fs_info->avail_system_alloc_bits |= extra_flags;
3549         write_sequnlock(&fs_info->profiles_lock);
3550 }
3551
3552 /*
3553  * returns target flags in extended format or 0 if restripe for this
3554  * chunk_type is not in progress
3555  *
3556  * should be called with either volume_mutex or balance_lock held
3557  */
3558 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3559 {
3560         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3561         u64 target = 0;
3562
3563         if (!bctl)
3564                 return 0;
3565
3566         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3567             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3568                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3569         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3570                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3571                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3572         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3573                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3574                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3575         }
3576
3577         return target;
3578 }
3579
3580 /*
3581  * @flags: available profiles in extended format (see ctree.h)
3582  *
3583  * Returns reduced profile in chunk format.  If profile changing is in
3584  * progress (either running or paused) picks the target profile (if it's
3585  * already available), otherwise falls back to plain reducing.
3586  */
3587 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3588 {
3589         /*
3590          * we add in the count of missing devices because we want
3591          * to make sure that any RAID levels on a degraded FS
3592          * continue to be honored.
3593          */
3594         u64 num_devices = root->fs_info->fs_devices->rw_devices +
3595                 root->fs_info->fs_devices->missing_devices;
3596         u64 target;
3597         u64 tmp;
3598
3599         /*
3600          * see if restripe for this chunk_type is in progress, if so
3601          * try to reduce to the target profile
3602          */
3603         spin_lock(&root->fs_info->balance_lock);
3604         target = get_restripe_target(root->fs_info, flags);
3605         if (target) {
3606                 /* pick target profile only if it's already available */
3607                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3608                         spin_unlock(&root->fs_info->balance_lock);
3609                         return extended_to_chunk(target);
3610                 }
3611         }
3612         spin_unlock(&root->fs_info->balance_lock);
3613
3614         /* First, mask out the RAID levels which aren't possible */
3615         if (num_devices == 1)
3616                 flags &= ~(BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID0 |
3617                            BTRFS_BLOCK_GROUP_RAID5);
3618         if (num_devices < 3)
3619                 flags &= ~BTRFS_BLOCK_GROUP_RAID6;
3620         if (num_devices < 4)
3621                 flags &= ~BTRFS_BLOCK_GROUP_RAID10;
3622
3623         tmp = flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID0 |
3624                        BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID5 |
3625                        BTRFS_BLOCK_GROUP_RAID6 | BTRFS_BLOCK_GROUP_RAID10);
3626         flags &= ~tmp;
3627
3628         if (tmp & BTRFS_BLOCK_GROUP_RAID6)
3629                 tmp = BTRFS_BLOCK_GROUP_RAID6;
3630         else if (tmp & BTRFS_BLOCK_GROUP_RAID5)
3631                 tmp = BTRFS_BLOCK_GROUP_RAID5;
3632         else if (tmp & BTRFS_BLOCK_GROUP_RAID10)
3633                 tmp = BTRFS_BLOCK_GROUP_RAID10;
3634         else if (tmp & BTRFS_BLOCK_GROUP_RAID1)
3635                 tmp = BTRFS_BLOCK_GROUP_RAID1;
3636         else if (tmp & BTRFS_BLOCK_GROUP_RAID0)
3637                 tmp = BTRFS_BLOCK_GROUP_RAID0;
3638
3639         return extended_to_chunk(flags | tmp);
3640 }
3641
3642 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
3643 {
3644         unsigned seq;
3645         u64 flags;
3646
3647         do {
3648                 flags = orig_flags;
3649                 seq = read_seqbegin(&root->fs_info->profiles_lock);
3650
3651                 if (flags & BTRFS_BLOCK_GROUP_DATA)
3652                         flags |= root->fs_info->avail_data_alloc_bits;
3653                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3654                         flags |= root->fs_info->avail_system_alloc_bits;
3655                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
3656                         flags |= root->fs_info->avail_metadata_alloc_bits;
3657         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
3658
3659         return btrfs_reduce_alloc_profile(root, flags);
3660 }
3661
3662 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
3663 {
3664         u64 flags;
3665         u64 ret;
3666
3667         if (data)
3668                 flags = BTRFS_BLOCK_GROUP_DATA;
3669         else if (root == root->fs_info->chunk_root)
3670                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
3671         else
3672                 flags = BTRFS_BLOCK_GROUP_METADATA;
3673
3674         ret = get_alloc_profile(root, flags);
3675         return ret;
3676 }
3677
3678 /*
3679  * This will check the space that the inode allocates from to make sure we have
3680  * enough space for bytes.
3681  */
3682 int btrfs_check_data_free_space(struct inode *inode, u64 bytes)
3683 {
3684         struct btrfs_space_info *data_sinfo;
3685         struct btrfs_root *root = BTRFS_I(inode)->root;
3686         struct btrfs_fs_info *fs_info = root->fs_info;
3687         u64 used;
3688         int ret = 0, committed = 0, alloc_chunk = 1;
3689
3690         /* make sure bytes are sectorsize aligned */
3691         bytes = ALIGN(bytes, root->sectorsize);
3692
3693         if (btrfs_is_free_space_inode(inode)) {
3694                 committed = 1;
3695                 ASSERT(current->journal_info);
3696         }
3697
3698         data_sinfo = fs_info->data_sinfo;
3699         if (!data_sinfo)
3700                 goto alloc;
3701
3702 again:
3703         /* make sure we have enough space to handle the data first */
3704         spin_lock(&data_sinfo->lock);
3705         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
3706                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
3707                 data_sinfo->bytes_may_use;
3708
3709         if (used + bytes > data_sinfo->total_bytes) {
3710                 struct btrfs_trans_handle *trans;
3711
3712                 /*
3713                  * if we don't have enough free bytes in this space then we need
3714                  * to alloc a new chunk.
3715                  */
3716                 if (!data_sinfo->full && alloc_chunk) {
3717                         u64 alloc_target;
3718
3719                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
3720                         spin_unlock(&data_sinfo->lock);
3721 alloc:
3722                         alloc_target = btrfs_get_alloc_profile(root, 1);
3723                         /*
3724                          * It is ugly that we don't call nolock join
3725                          * transaction for the free space inode case here.
3726                          * But it is safe because we only do the data space
3727                          * reservation for the free space cache in the
3728                          * transaction context, the common join transaction
3729                          * just increase the counter of the current transaction
3730                          * handler, doesn't try to acquire the trans_lock of
3731                          * the fs.
3732                          */
3733                         trans = btrfs_join_transaction(root);
3734                         if (IS_ERR(trans))
3735                                 return PTR_ERR(trans);
3736
3737                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
3738                                              alloc_target,
3739                                              CHUNK_ALLOC_NO_FORCE);
3740                         btrfs_end_transaction(trans, root);
3741                         if (ret < 0) {
3742                                 if (ret != -ENOSPC)
3743                                         return ret;
3744                                 else
3745                                         goto commit_trans;
3746                         }
3747
3748                         if (!data_sinfo)
3749                                 data_sinfo = fs_info->data_sinfo;
3750
3751                         goto again;
3752                 }
3753
3754                 /*
3755                  * If we don't have enough pinned space to deal with this
3756                  * allocation don't bother committing the transaction.
3757                  */
3758                 if (percpu_counter_compare(&data_sinfo->total_bytes_pinned,
3759                                            bytes) < 0)
3760                         committed = 1;
3761                 spin_unlock(&data_sinfo->lock);
3762
3763                 /* commit the current transaction and try again */
3764 commit_trans:
3765                 if (!committed &&
3766                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
3767                         committed = 1;
3768
3769                         trans = btrfs_join_transaction(root);
3770                         if (IS_ERR(trans))
3771                                 return PTR_ERR(trans);
3772                         ret = btrfs_commit_transaction(trans, root);
3773                         if (ret)
3774                                 return ret;
3775                         goto again;
3776                 }
3777
3778                 trace_btrfs_space_reservation(root->fs_info,
3779                                               "space_info:enospc",
3780                                               data_sinfo->flags, bytes, 1);
3781                 return -ENOSPC;
3782         }
3783         data_sinfo->bytes_may_use += bytes;
3784         trace_btrfs_space_reservation(root->fs_info, "space_info",
3785                                       data_sinfo->flags, bytes, 1);
3786         spin_unlock(&data_sinfo->lock);
3787
3788         return 0;
3789 }
3790
3791 /*
3792  * Called if we need to clear a data reservation for this inode.
3793  */
3794 void btrfs_free_reserved_data_space(struct inode *inode, u64 bytes)
3795 {
3796         struct btrfs_root *root = BTRFS_I(inode)->root;
3797         struct btrfs_space_info *data_sinfo;
3798
3799         /* make sure bytes are sectorsize aligned */
3800         bytes = ALIGN(bytes, root->sectorsize);
3801
3802         data_sinfo = root->fs_info->data_sinfo;
3803         spin_lock(&data_sinfo->lock);
3804         WARN_ON(data_sinfo->bytes_may_use < bytes);
3805         data_sinfo->bytes_may_use -= bytes;
3806         trace_btrfs_space_reservation(root->fs_info, "space_info",
3807                                       data_sinfo->flags, bytes, 0);
3808         spin_unlock(&data_sinfo->lock);
3809 }
3810
3811 static void force_metadata_allocation(struct btrfs_fs_info *info)
3812 {
3813         struct list_head *head = &info->space_info;
3814         struct btrfs_space_info *found;
3815
3816         rcu_read_lock();
3817         list_for_each_entry_rcu(found, head, list) {
3818                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
3819                         found->force_alloc = CHUNK_ALLOC_FORCE;
3820         }
3821         rcu_read_unlock();
3822 }
3823
3824 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
3825 {
3826         return (global->size << 1);
3827 }
3828
3829 static int should_alloc_chunk(struct btrfs_root *root,
3830                               struct btrfs_space_info *sinfo, int force)
3831 {
3832         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
3833         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
3834         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
3835         u64 thresh;
3836
3837         if (force == CHUNK_ALLOC_FORCE)
3838                 return 1;
3839
3840         /*
3841          * We need to take into account the global rsv because for all intents
3842          * and purposes it's used space.  Don't worry about locking the
3843          * global_rsv, it doesn't change except when the transaction commits.
3844          */
3845         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
3846                 num_allocated += calc_global_rsv_need_space(global_rsv);
3847
3848         /*
3849          * in limited mode, we want to have some free space up to
3850          * about 1% of the FS size.
3851          */
3852         if (force == CHUNK_ALLOC_LIMITED) {
3853                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
3854                 thresh = max_t(u64, 64 * 1024 * 1024,
3855                                div_factor_fine(thresh, 1));
3856
3857                 if (num_bytes - num_allocated < thresh)
3858                         return 1;
3859         }
3860
3861         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
3862                 return 0;
3863         return 1;
3864 }
3865
3866 static u64 get_system_chunk_thresh(struct btrfs_root *root, u64 type)
3867 {
3868         u64 num_dev;
3869
3870         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
3871                     BTRFS_BLOCK_GROUP_RAID0 |
3872                     BTRFS_BLOCK_GROUP_RAID5 |
3873                     BTRFS_BLOCK_GROUP_RAID6))
3874                 num_dev = root->fs_info->fs_devices->rw_devices;
3875         else if (type & BTRFS_BLOCK_GROUP_RAID1)
3876                 num_dev = 2;
3877         else
3878                 num_dev = 1;    /* DUP or single */
3879
3880         /* metadata for updaing devices and chunk tree */
3881         return btrfs_calc_trans_metadata_size(root, num_dev + 1);
3882 }
3883
3884 static void check_system_chunk(struct btrfs_trans_handle *trans,
3885                                struct btrfs_root *root, u64 type)
3886 {
3887         struct btrfs_space_info *info;
3888         u64 left;
3889         u64 thresh;
3890
3891         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
3892         spin_lock(&info->lock);
3893         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
3894                 info->bytes_reserved - info->bytes_readonly;
3895         spin_unlock(&info->lock);
3896
3897         thresh = get_system_chunk_thresh(root, type);
3898         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
3899                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
3900                         left, thresh, type);
3901                 dump_space_info(info, 0, 0);
3902         }
3903
3904         if (left < thresh) {
3905                 u64 flags;
3906
3907                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
3908                 btrfs_alloc_chunk(trans, root, flags);
3909         }
3910 }
3911
3912 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
3913                           struct btrfs_root *extent_root, u64 flags, int force)
3914 {
3915         struct btrfs_space_info *space_info;
3916         struct btrfs_fs_info *fs_info = extent_root->fs_info;
3917         int wait_for_alloc = 0;
3918         int ret = 0;
3919
3920         /* Don't re-enter if we're already allocating a chunk */
3921         if (trans->allocating_chunk)
3922                 return -ENOSPC;
3923
3924         space_info = __find_space_info(extent_root->fs_info, flags);
3925         if (!space_info) {
3926                 ret = update_space_info(extent_root->fs_info, flags,
3927                                         0, 0, &space_info);
3928                 BUG_ON(ret); /* -ENOMEM */
3929         }
3930         BUG_ON(!space_info); /* Logic error */
3931
3932 again:
3933         spin_lock(&space_info->lock);
3934         if (force < space_info->force_alloc)
3935                 force = space_info->force_alloc;
3936         if (space_info->full) {
3937                 if (should_alloc_chunk(extent_root, space_info, force))
3938                         ret = -ENOSPC;
3939                 else
3940                         ret = 0;
3941                 spin_unlock(&space_info->lock);
3942                 return ret;
3943         }
3944
3945         if (!should_alloc_chunk(extent_root, space_info, force)) {
3946                 spin_unlock(&space_info->lock);
3947                 return 0;
3948         } else if (space_info->chunk_alloc) {
3949                 wait_for_alloc = 1;
3950         } else {
3951                 space_info->chunk_alloc = 1;
3952         }
3953
3954         spin_unlock(&space_info->lock);
3955
3956         mutex_lock(&fs_info->chunk_mutex);
3957
3958         /*
3959          * The chunk_mutex is held throughout the entirety of a chunk
3960          * allocation, so once we've acquired the chunk_mutex we know that the
3961          * other guy is done and we need to recheck and see if we should
3962          * allocate.
3963          */
3964         if (wait_for_alloc) {
3965                 mutex_unlock(&fs_info->chunk_mutex);
3966                 wait_for_alloc = 0;
3967                 goto again;
3968         }
3969
3970         trans->allocating_chunk = true;
3971
3972         /*
3973          * If we have mixed data/metadata chunks we want to make sure we keep
3974          * allocating mixed chunks instead of individual chunks.
3975          */
3976         if (btrfs_mixed_space_info(space_info))
3977                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
3978
3979         /*
3980          * if we're doing a data chunk, go ahead and make sure that
3981          * we keep a reasonable number of metadata chunks allocated in the
3982          * FS as well.
3983          */
3984         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
3985                 fs_info->data_chunk_allocations++;
3986                 if (!(fs_info->data_chunk_allocations %
3987                       fs_info->metadata_ratio))
3988                         force_metadata_allocation(fs_info);
3989         }
3990
3991         /*
3992          * Check if we have enough space in SYSTEM chunk because we may need
3993          * to update devices.
3994          */
3995         check_system_chunk(trans, extent_root, flags);
3996
3997         ret = btrfs_alloc_chunk(trans, extent_root, flags);
3998         trans->allocating_chunk = false;
3999
4000         spin_lock(&space_info->lock);
4001         if (ret < 0 && ret != -ENOSPC)
4002                 goto out;
4003         if (ret)
4004                 space_info->full = 1;
4005         else
4006                 ret = 1;
4007
4008         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4009 out:
4010         space_info->chunk_alloc = 0;
4011         spin_unlock(&space_info->lock);
4012         mutex_unlock(&fs_info->chunk_mutex);
4013         return ret;
4014 }
4015
4016 static int can_overcommit(struct btrfs_root *root,
4017                           struct btrfs_space_info *space_info, u64 bytes,
4018                           enum btrfs_reserve_flush_enum flush)
4019 {
4020         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4021         u64 profile = btrfs_get_alloc_profile(root, 0);
4022         u64 space_size;
4023         u64 avail;
4024         u64 used;
4025
4026         used = space_info->bytes_used + space_info->bytes_reserved +
4027                 space_info->bytes_pinned + space_info->bytes_readonly;
4028
4029         /*
4030          * We only want to allow over committing if we have lots of actual space
4031          * free, but if we don't have enough space to handle the global reserve
4032          * space then we could end up having a real enospc problem when trying
4033          * to allocate a chunk or some other such important allocation.
4034          */
4035         spin_lock(&global_rsv->lock);
4036         space_size = calc_global_rsv_need_space(global_rsv);
4037         spin_unlock(&global_rsv->lock);
4038         if (used + space_size >= space_info->total_bytes)
4039                 return 0;
4040
4041         used += space_info->bytes_may_use;
4042
4043         spin_lock(&root->fs_info->free_chunk_lock);
4044         avail = root->fs_info->free_chunk_space;
4045         spin_unlock(&root->fs_info->free_chunk_lock);
4046
4047         /*
4048          * If we have dup, raid1 or raid10 then only half of the free
4049          * space is actually useable.  For raid56, the space info used
4050          * doesn't include the parity drive, so we don't have to
4051          * change the math
4052          */
4053         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4054                        BTRFS_BLOCK_GROUP_RAID1 |
4055                        BTRFS_BLOCK_GROUP_RAID10))
4056                 avail >>= 1;
4057
4058         /*
4059          * If we aren't flushing all things, let us overcommit up to
4060          * 1/2th of the space. If we can flush, don't let us overcommit
4061          * too much, let it overcommit up to 1/8 of the space.
4062          */
4063         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4064                 avail >>= 3;
4065         else
4066                 avail >>= 1;
4067
4068         if (used + bytes < space_info->total_bytes + avail)
4069                 return 1;
4070         return 0;
4071 }
4072
4073 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4074                                          unsigned long nr_pages, int nr_items)
4075 {
4076         struct super_block *sb = root->fs_info->sb;
4077
4078         if (down_read_trylock(&sb->s_umount)) {
4079                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4080                 up_read(&sb->s_umount);
4081         } else {
4082                 /*
4083                  * We needn't worry the filesystem going from r/w to r/o though
4084                  * we don't acquire ->s_umount mutex, because the filesystem
4085                  * should guarantee the delalloc inodes list be empty after
4086                  * the filesystem is readonly(all dirty pages are written to
4087                  * the disk).
4088                  */
4089                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4090                 if (!current->journal_info)
4091                         btrfs_wait_ordered_roots(root->fs_info, nr_items);
4092         }
4093 }
4094
4095 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4096 {
4097         u64 bytes;
4098         int nr;
4099
4100         bytes = btrfs_calc_trans_metadata_size(root, 1);
4101         nr = (int)div64_u64(to_reclaim, bytes);
4102         if (!nr)
4103                 nr = 1;
4104         return nr;
4105 }
4106
4107 #define EXTENT_SIZE_PER_ITEM    (256 * 1024)
4108
4109 /*
4110  * shrink metadata reservation for delalloc
4111  */
4112 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4113                             bool wait_ordered)
4114 {
4115         struct btrfs_block_rsv *block_rsv;
4116         struct btrfs_space_info *space_info;
4117         struct btrfs_trans_handle *trans;
4118         u64 delalloc_bytes;
4119         u64 max_reclaim;
4120         long time_left;
4121         unsigned long nr_pages;
4122         int loops;
4123         int items;
4124         enum btrfs_reserve_flush_enum flush;
4125
4126         /* Calc the number of the pages we need flush for space reservation */
4127         items = calc_reclaim_items_nr(root, to_reclaim);
4128         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4129
4130         trans = (struct btrfs_trans_handle *)current->journal_info;
4131         block_rsv = &root->fs_info->delalloc_block_rsv;
4132         space_info = block_rsv->space_info;
4133
4134         delalloc_bytes = percpu_counter_sum_positive(
4135                                                 &root->fs_info->delalloc_bytes);
4136         if (delalloc_bytes == 0) {
4137                 if (trans)
4138                         return;
4139                 if (wait_ordered)
4140                         btrfs_wait_ordered_roots(root->fs_info, items);
4141                 return;
4142         }
4143
4144         loops = 0;
4145         while (delalloc_bytes && loops < 3) {
4146                 max_reclaim = min(delalloc_bytes, to_reclaim);
4147                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4148                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4149                 /*
4150                  * We need to wait for the async pages to actually start before
4151                  * we do anything.
4152                  */
4153                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4154                 if (!max_reclaim)
4155                         goto skip_async;
4156
4157                 if (max_reclaim <= nr_pages)
4158                         max_reclaim = 0;
4159                 else
4160                         max_reclaim -= nr_pages;
4161
4162                 wait_event(root->fs_info->async_submit_wait,
4163                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4164                            (int)max_reclaim);
4165 skip_async:
4166                 if (!trans)
4167                         flush = BTRFS_RESERVE_FLUSH_ALL;
4168                 else
4169                         flush = BTRFS_RESERVE_NO_FLUSH;
4170                 spin_lock(&space_info->lock);
4171                 if (can_overcommit(root, space_info, orig, flush)) {
4172                         spin_unlock(&space_info->lock);
4173                         break;
4174                 }
4175                 spin_unlock(&space_info->lock);
4176
4177                 loops++;
4178                 if (wait_ordered && !trans) {
4179                         btrfs_wait_ordered_roots(root->fs_info, items);
4180                 } else {
4181                         time_left = schedule_timeout_killable(1);
4182                         if (time_left)
4183                                 break;
4184                 }
4185                 delalloc_bytes = percpu_counter_sum_positive(
4186                                                 &root->fs_info->delalloc_bytes);
4187         }
4188 }
4189
4190 /**
4191  * maybe_commit_transaction - possibly commit the transaction if its ok to
4192  * @root - the root we're allocating for
4193  * @bytes - the number of bytes we want to reserve
4194  * @force - force the commit
4195  *
4196  * This will check to make sure that committing the transaction will actually
4197  * get us somewhere and then commit the transaction if it does.  Otherwise it
4198  * will return -ENOSPC.
4199  */
4200 static int may_commit_transaction(struct btrfs_root *root,
4201                                   struct btrfs_space_info *space_info,
4202                                   u64 bytes, int force)
4203 {
4204         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4205         struct btrfs_trans_handle *trans;
4206
4207         trans = (struct btrfs_trans_handle *)current->journal_info;
4208         if (trans)
4209                 return -EAGAIN;
4210
4211         if (force)
4212                 goto commit;
4213
4214         /* See if there is enough pinned space to make this reservation */
4215         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4216                                    bytes) >= 0)
4217                 goto commit;
4218
4219         /*
4220          * See if there is some space in the delayed insertion reservation for
4221          * this reservation.
4222          */
4223         if (space_info != delayed_rsv->space_info)
4224                 return -ENOSPC;
4225
4226         spin_lock(&delayed_rsv->lock);
4227         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4228                                    bytes - delayed_rsv->size) >= 0) {
4229                 spin_unlock(&delayed_rsv->lock);
4230                 return -ENOSPC;
4231         }
4232         spin_unlock(&delayed_rsv->lock);
4233
4234 commit:
4235         trans = btrfs_join_transaction(root);
4236         if (IS_ERR(trans))
4237                 return -ENOSPC;
4238
4239         return btrfs_commit_transaction(trans, root);
4240 }
4241
4242 enum flush_state {
4243         FLUSH_DELAYED_ITEMS_NR  =       1,
4244         FLUSH_DELAYED_ITEMS     =       2,
4245         FLUSH_DELALLOC          =       3,
4246         FLUSH_DELALLOC_WAIT     =       4,
4247         ALLOC_CHUNK             =       5,
4248         COMMIT_TRANS            =       6,
4249 };
4250
4251 static int flush_space(struct btrfs_root *root,
4252                        struct btrfs_space_info *space_info, u64 num_bytes,
4253                        u64 orig_bytes, int state)
4254 {
4255         struct btrfs_trans_handle *trans;
4256         int nr;
4257         int ret = 0;
4258
4259         switch (state) {
4260         case FLUSH_DELAYED_ITEMS_NR:
4261         case FLUSH_DELAYED_ITEMS:
4262                 if (state == FLUSH_DELAYED_ITEMS_NR)
4263                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4264                 else
4265                         nr = -1;
4266
4267                 trans = btrfs_join_transaction(root);
4268                 if (IS_ERR(trans)) {
4269                         ret = PTR_ERR(trans);
4270                         break;
4271                 }
4272                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4273                 btrfs_end_transaction(trans, root);
4274                 break;
4275         case FLUSH_DELALLOC:
4276         case FLUSH_DELALLOC_WAIT:
4277                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4278                                 state == FLUSH_DELALLOC_WAIT);
4279                 break;
4280         case ALLOC_CHUNK:
4281                 trans = btrfs_join_transaction(root);
4282                 if (IS_ERR(trans)) {
4283                         ret = PTR_ERR(trans);
4284                         break;
4285                 }
4286                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4287                                      btrfs_get_alloc_profile(root, 0),
4288                                      CHUNK_ALLOC_NO_FORCE);
4289                 btrfs_end_transaction(trans, root);
4290                 if (ret == -ENOSPC)
4291                         ret = 0;
4292                 break;
4293         case COMMIT_TRANS:
4294                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4295                 break;
4296         default:
4297                 ret = -ENOSPC;
4298                 break;
4299         }
4300
4301         return ret;
4302 }
4303
4304 static inline u64
4305 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4306                                  struct btrfs_space_info *space_info)
4307 {
4308         u64 used;
4309         u64 expected;
4310         u64 to_reclaim;
4311
4312         to_reclaim = min_t(u64, num_online_cpus() * 1024 * 1024,
4313                                 16 * 1024 * 1024);
4314         spin_lock(&space_info->lock);
4315         if (can_overcommit(root, space_info, to_reclaim,
4316                            BTRFS_RESERVE_FLUSH_ALL)) {
4317                 to_reclaim = 0;
4318                 goto out;
4319         }
4320
4321         used = space_info->bytes_used + space_info->bytes_reserved +
4322                space_info->bytes_pinned + space_info->bytes_readonly +
4323                space_info->bytes_may_use;
4324         if (can_overcommit(root, space_info, 1024 * 1024,
4325                            BTRFS_RESERVE_FLUSH_ALL))
4326                 expected = div_factor_fine(space_info->total_bytes, 95);
4327         else
4328                 expected = div_factor_fine(space_info->total_bytes, 90);
4329
4330         if (used > expected)
4331                 to_reclaim = used - expected;
4332         else
4333                 to_reclaim = 0;
4334         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4335                                      space_info->bytes_reserved);
4336 out:
4337         spin_unlock(&space_info->lock);
4338
4339         return to_reclaim;
4340 }
4341
4342 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4343                                         struct btrfs_fs_info *fs_info, u64 used)
4344 {
4345         return (used >= div_factor_fine(space_info->total_bytes, 98) &&
4346                 !btrfs_fs_closing(fs_info) &&
4347                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4348 }
4349
4350 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4351                                        struct btrfs_fs_info *fs_info)
4352 {
4353         u64 used;
4354
4355         spin_lock(&space_info->lock);
4356         used = space_info->bytes_used + space_info->bytes_reserved +
4357                space_info->bytes_pinned + space_info->bytes_readonly +
4358                space_info->bytes_may_use;
4359         if (need_do_async_reclaim(space_info, fs_info, used)) {
4360                 spin_unlock(&space_info->lock);
4361                 return 1;
4362         }
4363         spin_unlock(&space_info->lock);
4364
4365         return 0;
4366 }
4367
4368 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4369 {
4370         struct btrfs_fs_info *fs_info;
4371         struct btrfs_space_info *space_info;
4372         u64 to_reclaim;
4373         int flush_state;
4374
4375         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4376         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4377
4378         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4379                                                       space_info);
4380         if (!to_reclaim)
4381                 return;
4382
4383         flush_state = FLUSH_DELAYED_ITEMS_NR;
4384         do {
4385                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4386                             to_reclaim, flush_state);
4387                 flush_state++;
4388                 if (!btrfs_need_do_async_reclaim(space_info, fs_info))
4389                         return;
4390         } while (flush_state <= COMMIT_TRANS);
4391
4392         if (btrfs_need_do_async_reclaim(space_info, fs_info))
4393                 queue_work(system_unbound_wq, work);
4394 }
4395
4396 void btrfs_init_async_reclaim_work(struct work_struct *work)
4397 {
4398         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4399 }
4400
4401 /**
4402  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4403  * @root - the root we're allocating for
4404  * @block_rsv - the block_rsv we're allocating for
4405  * @orig_bytes - the number of bytes we want
4406  * @flush - whether or not we can flush to make our reservation
4407  *
4408  * This will reserve orgi_bytes number of bytes from the space info associated
4409  * with the block_rsv.  If there is not enough space it will make an attempt to
4410  * flush out space to make room.  It will do this by flushing delalloc if
4411  * possible or committing the transaction.  If flush is 0 then no attempts to
4412  * regain reservations will be made and this will fail if there is not enough
4413  * space already.
4414  */
4415 static int reserve_metadata_bytes(struct btrfs_root *root,
4416                                   struct btrfs_block_rsv *block_rsv,
4417                                   u64 orig_bytes,
4418                                   enum btrfs_reserve_flush_enum flush)
4419 {
4420         struct btrfs_space_info *space_info = block_rsv->space_info;
4421         u64 used;
4422         u64 num_bytes = orig_bytes;
4423         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4424         int ret = 0;
4425         bool flushing = false;
4426
4427 again:
4428         ret = 0;
4429         spin_lock(&space_info->lock);
4430         /*
4431          * We only want to wait if somebody other than us is flushing and we
4432          * are actually allowed to flush all things.
4433          */
4434         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4435                space_info->flush) {
4436                 spin_unlock(&space_info->lock);
4437                 /*
4438                  * If we have a trans handle we can't wait because the flusher
4439                  * may have to commit the transaction, which would mean we would
4440                  * deadlock since we are waiting for the flusher to finish, but
4441                  * hold the current transaction open.
4442                  */
4443                 if (current->journal_info)
4444                         return -EAGAIN;
4445                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4446                 /* Must have been killed, return */
4447                 if (ret)
4448                         return -EINTR;
4449
4450                 spin_lock(&space_info->lock);
4451         }
4452
4453         ret = -ENOSPC;
4454         used = space_info->bytes_used + space_info->bytes_reserved +
4455                 space_info->bytes_pinned + space_info->bytes_readonly +
4456                 space_info->bytes_may_use;
4457
4458         /*
4459          * The idea here is that we've not already over-reserved the block group
4460          * then we can go ahead and save our reservation first and then start
4461          * flushing if we need to.  Otherwise if we've already overcommitted
4462          * lets start flushing stuff first and then come back and try to make
4463          * our reservation.
4464          */
4465         if (used <= space_info->total_bytes) {
4466                 if (used + orig_bytes <= space_info->total_bytes) {
4467                         space_info->bytes_may_use += orig_bytes;
4468                         trace_btrfs_space_reservation(root->fs_info,
4469                                 "space_info", space_info->flags, orig_bytes, 1);
4470                         ret = 0;
4471                 } else {
4472                         /*
4473                          * Ok set num_bytes to orig_bytes since we aren't
4474                          * overocmmitted, this way we only try and reclaim what
4475                          * we need.
4476                          */
4477                         num_bytes = orig_bytes;
4478                 }
4479         } else {
4480                 /*
4481                  * Ok we're over committed, set num_bytes to the overcommitted
4482                  * amount plus the amount of bytes that we need for this
4483                  * reservation.
4484                  */
4485                 num_bytes = used - space_info->total_bytes +
4486                         (orig_bytes * 2);
4487         }
4488
4489         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4490                 space_info->bytes_may_use += orig_bytes;
4491                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4492                                               space_info->flags, orig_bytes,
4493                                               1);
4494                 ret = 0;
4495         }
4496
4497         /*
4498          * Couldn't make our reservation, save our place so while we're trying
4499          * to reclaim space we can actually use it instead of somebody else
4500          * stealing it from us.
4501          *
4502          * We make the other tasks wait for the flush only when we can flush
4503          * all things.
4504          */
4505         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4506                 flushing = true;
4507                 space_info->flush = 1;
4508         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
4509                 used += orig_bytes;
4510                 if (need_do_async_reclaim(space_info, root->fs_info, used) &&
4511                     !work_busy(&root->fs_info->async_reclaim_work))
4512                         queue_work(system_unbound_wq,
4513                                    &root->fs_info->async_reclaim_work);
4514         }
4515         spin_unlock(&space_info->lock);
4516
4517         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
4518                 goto out;
4519
4520         ret = flush_space(root, space_info, num_bytes, orig_bytes,
4521                           flush_state);
4522         flush_state++;
4523
4524         /*
4525          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
4526          * would happen. So skip delalloc flush.
4527          */
4528         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4529             (flush_state == FLUSH_DELALLOC ||
4530              flush_state == FLUSH_DELALLOC_WAIT))
4531                 flush_state = ALLOC_CHUNK;
4532
4533         if (!ret)
4534                 goto again;
4535         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
4536                  flush_state < COMMIT_TRANS)
4537                 goto again;
4538         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
4539                  flush_state <= COMMIT_TRANS)
4540                 goto again;
4541
4542 out:
4543         if (ret == -ENOSPC &&
4544             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
4545                 struct btrfs_block_rsv *global_rsv =
4546                         &root->fs_info->global_block_rsv;
4547
4548                 if (block_rsv != global_rsv &&
4549                     !block_rsv_use_bytes(global_rsv, orig_bytes))
4550                         ret = 0;
4551         }
4552         if (ret == -ENOSPC)
4553                 trace_btrfs_space_reservation(root->fs_info,
4554                                               "space_info:enospc",
4555                                               space_info->flags, orig_bytes, 1);
4556         if (flushing) {
4557                 spin_lock(&space_info->lock);
4558                 space_info->flush = 0;
4559                 wake_up_all(&space_info->wait);
4560                 spin_unlock(&space_info->lock);
4561         }
4562         return ret;
4563 }
4564
4565 static struct btrfs_block_rsv *get_block_rsv(
4566                                         const struct btrfs_trans_handle *trans,
4567                                         const struct btrfs_root *root)
4568 {
4569         struct btrfs_block_rsv *block_rsv = NULL;
4570
4571         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state))
4572                 block_rsv = trans->block_rsv;
4573
4574         if (root == root->fs_info->csum_root && trans->adding_csums)
4575                 block_rsv = trans->block_rsv;
4576
4577         if (root == root->fs_info->uuid_root)
4578                 block_rsv = trans->block_rsv;
4579
4580         if (!block_rsv)
4581                 block_rsv = root->block_rsv;
4582
4583         if (!block_rsv)
4584                 block_rsv = &root->fs_info->empty_block_rsv;
4585
4586         return block_rsv;
4587 }
4588
4589 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
4590                                u64 num_bytes)
4591 {
4592         int ret = -ENOSPC;
4593         spin_lock(&block_rsv->lock);
4594         if (block_rsv->reserved >= num_bytes) {
4595                 block_rsv->reserved -= num_bytes;
4596                 if (block_rsv->reserved < block_rsv->size)
4597                         block_rsv->full = 0;
4598                 ret = 0;
4599         }
4600         spin_unlock(&block_rsv->lock);
4601         return ret;
4602 }
4603
4604 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
4605                                 u64 num_bytes, int update_size)
4606 {
4607         spin_lock(&block_rsv->lock);
4608         block_rsv->reserved += num_bytes;
4609         if (update_size)
4610                 block_rsv->size += num_bytes;
4611         else if (block_rsv->reserved >= block_rsv->size)
4612                 block_rsv->full = 1;
4613         spin_unlock(&block_rsv->lock);
4614 }
4615
4616 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
4617                              struct btrfs_block_rsv *dest, u64 num_bytes,
4618                              int min_factor)
4619 {
4620         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
4621         u64 min_bytes;
4622
4623         if (global_rsv->space_info != dest->space_info)
4624                 return -ENOSPC;
4625
4626         spin_lock(&global_rsv->lock);
4627         min_bytes = div_factor(global_rsv->size, min_factor);
4628         if (global_rsv->reserved < min_bytes + num_bytes) {
4629                 spin_unlock(&global_rsv->lock);
4630                 return -ENOSPC;
4631         }
4632         global_rsv->reserved -= num_bytes;
4633         if (global_rsv->reserved < global_rsv->size)
4634                 global_rsv->full = 0;
4635         spin_unlock(&global_rsv->lock);
4636
4637         block_rsv_add_bytes(dest, num_bytes, 1);
4638         return 0;
4639 }
4640
4641 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
4642                                     struct btrfs_block_rsv *block_rsv,
4643                                     struct btrfs_block_rsv *dest, u64 num_bytes)
4644 {
4645         struct btrfs_space_info *space_info = block_rsv->space_info;
4646
4647         spin_lock(&block_rsv->lock);
4648         if (num_bytes == (u64)-1)
4649                 num_bytes = block_rsv->size;
4650         block_rsv->size -= num_bytes;
4651         if (block_rsv->reserved >= block_rsv->size) {
4652                 num_bytes = block_rsv->reserved - block_rsv->size;
4653                 block_rsv->reserved = block_rsv->size;
4654                 block_rsv->full = 1;
4655         } else {
4656                 num_bytes = 0;
4657         }
4658         spin_unlock(&block_rsv->lock);
4659
4660         if (num_bytes > 0) {
4661                 if (dest) {
4662                         spin_lock(&dest->lock);
4663                         if (!dest->full) {
4664                                 u64 bytes_to_add;
4665
4666                                 bytes_to_add = dest->size - dest->reserved;
4667                                 bytes_to_add = min(num_bytes, bytes_to_add);
4668                                 dest->reserved += bytes_to_add;
4669                                 if (dest->reserved >= dest->size)
4670                                         dest->full = 1;
4671                                 num_bytes -= bytes_to_add;
4672                         }
4673                         spin_unlock(&dest->lock);
4674                 }
4675                 if (num_bytes) {
4676                         spin_lock(&space_info->lock);
4677                         space_info->bytes_may_use -= num_bytes;
4678                         trace_btrfs_space_reservation(fs_info, "space_info",
4679                                         space_info->flags, num_bytes, 0);
4680                         spin_unlock(&space_info->lock);
4681                 }
4682         }
4683 }
4684
4685 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
4686                                    struct btrfs_block_rsv *dst, u64 num_bytes)
4687 {
4688         int ret;
4689
4690         ret = block_rsv_use_bytes(src, num_bytes);
4691         if (ret)
4692                 return ret;
4693
4694         block_rsv_add_bytes(dst, num_bytes, 1);
4695         return 0;
4696 }
4697
4698 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
4699 {
4700         memset(rsv, 0, sizeof(*rsv));
4701         spin_lock_init(&rsv->lock);
4702         rsv->type = type;
4703 }
4704
4705 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
4706                                               unsigned short type)
4707 {
4708         struct btrfs_block_rsv *block_rsv;
4709         struct btrfs_fs_info *fs_info = root->fs_info;
4710
4711         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
4712         if (!block_rsv)
4713                 return NULL;
4714
4715         btrfs_init_block_rsv(block_rsv, type);
4716         block_rsv->space_info = __find_space_info(fs_info,
4717                                                   BTRFS_BLOCK_GROUP_METADATA);
4718         return block_rsv;
4719 }
4720
4721 void btrfs_free_block_rsv(struct btrfs_root *root,
4722                           struct btrfs_block_rsv *rsv)
4723 {
4724         if (!rsv)
4725                 return;
4726         btrfs_block_rsv_release(root, rsv, (u64)-1);
4727         kfree(rsv);
4728 }
4729
4730 int btrfs_block_rsv_add(struct btrfs_root *root,
4731                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
4732                         enum btrfs_reserve_flush_enum flush)
4733 {
4734         int ret;
4735
4736         if (num_bytes == 0)
4737                 return 0;
4738
4739         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4740         if (!ret) {
4741                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
4742                 return 0;
4743         }
4744
4745         return ret;
4746 }
4747
4748 int btrfs_block_rsv_check(struct btrfs_root *root,
4749                           struct btrfs_block_rsv *block_rsv, int min_factor)
4750 {
4751         u64 num_bytes = 0;
4752         int ret = -ENOSPC;
4753
4754         if (!block_rsv)
4755                 return 0;
4756
4757         spin_lock(&block_rsv->lock);
4758         num_bytes = div_factor(block_rsv->size, min_factor);
4759         if (block_rsv->reserved >= num_bytes)
4760                 ret = 0;
4761         spin_unlock(&block_rsv->lock);
4762
4763         return ret;
4764 }
4765
4766 int btrfs_block_rsv_refill(struct btrfs_root *root,
4767                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
4768                            enum btrfs_reserve_flush_enum flush)
4769 {
4770         u64 num_bytes = 0;
4771         int ret = -ENOSPC;
4772
4773         if (!block_rsv)
4774                 return 0;
4775
4776         spin_lock(&block_rsv->lock);
4777         num_bytes = min_reserved;
4778         if (block_rsv->reserved >= num_bytes)
4779                 ret = 0;
4780         else
4781                 num_bytes -= block_rsv->reserved;
4782         spin_unlock(&block_rsv->lock);
4783
4784         if (!ret)
4785                 return 0;
4786
4787         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
4788         if (!ret) {
4789                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
4790                 return 0;
4791         }
4792
4793         return ret;
4794 }
4795
4796 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
4797                             struct btrfs_block_rsv *dst_rsv,
4798                             u64 num_bytes)
4799 {
4800         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4801 }
4802
4803 void btrfs_block_rsv_release(struct btrfs_root *root,
4804                              struct btrfs_block_rsv *block_rsv,
4805                              u64 num_bytes)
4806 {
4807         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4808         if (global_rsv == block_rsv ||
4809             block_rsv->space_info != global_rsv->space_info)
4810                 global_rsv = NULL;
4811         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
4812                                 num_bytes);
4813 }
4814
4815 /*
4816  * helper to calculate size of global block reservation.
4817  * the desired value is sum of space used by extent tree,
4818  * checksum tree and root tree
4819  */
4820 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
4821 {
4822         struct btrfs_space_info *sinfo;
4823         u64 num_bytes;
4824         u64 meta_used;
4825         u64 data_used;
4826         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
4827
4828         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
4829         spin_lock(&sinfo->lock);
4830         data_used = sinfo->bytes_used;
4831         spin_unlock(&sinfo->lock);
4832
4833         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4834         spin_lock(&sinfo->lock);
4835         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
4836                 data_used = 0;
4837         meta_used = sinfo->bytes_used;
4838         spin_unlock(&sinfo->lock);
4839
4840         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
4841                     csum_size * 2;
4842         num_bytes += div64_u64(data_used + meta_used, 50);
4843
4844         if (num_bytes * 3 > meta_used)
4845                 num_bytes = div64_u64(meta_used, 3);
4846
4847         return ALIGN(num_bytes, fs_info->extent_root->leafsize << 10);
4848 }
4849
4850 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
4851 {
4852         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
4853         struct btrfs_space_info *sinfo = block_rsv->space_info;
4854         u64 num_bytes;
4855
4856         num_bytes = calc_global_metadata_size(fs_info);
4857
4858         spin_lock(&sinfo->lock);
4859         spin_lock(&block_rsv->lock);
4860
4861         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
4862
4863         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
4864                     sinfo->bytes_reserved + sinfo->bytes_readonly +
4865                     sinfo->bytes_may_use;
4866
4867         if (sinfo->total_bytes > num_bytes) {
4868                 num_bytes = sinfo->total_bytes - num_bytes;
4869                 block_rsv->reserved += num_bytes;
4870                 sinfo->bytes_may_use += num_bytes;
4871                 trace_btrfs_space_reservation(fs_info, "space_info",
4872                                       sinfo->flags, num_bytes, 1);
4873         }
4874
4875         if (block_rsv->reserved >= block_rsv->size) {
4876                 num_bytes = block_rsv->reserved - block_rsv->size;
4877                 sinfo->bytes_may_use -= num_bytes;
4878                 trace_btrfs_space_reservation(fs_info, "space_info",
4879                                       sinfo->flags, num_bytes, 0);
4880                 block_rsv->reserved = block_rsv->size;
4881                 block_rsv->full = 1;
4882         }
4883
4884         spin_unlock(&block_rsv->lock);
4885         spin_unlock(&sinfo->lock);
4886 }
4887
4888 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
4889 {
4890         struct btrfs_space_info *space_info;
4891
4892         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4893         fs_info->chunk_block_rsv.space_info = space_info;
4894
4895         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4896         fs_info->global_block_rsv.space_info = space_info;
4897         fs_info->delalloc_block_rsv.space_info = space_info;
4898         fs_info->trans_block_rsv.space_info = space_info;
4899         fs_info->empty_block_rsv.space_info = space_info;
4900         fs_info->delayed_block_rsv.space_info = space_info;
4901
4902         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
4903         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
4904         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
4905         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
4906         if (fs_info->quota_root)
4907                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
4908         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
4909
4910         update_global_block_rsv(fs_info);
4911 }
4912
4913 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
4914 {
4915         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
4916                                 (u64)-1);
4917         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
4918         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
4919         WARN_ON(fs_info->trans_block_rsv.size > 0);
4920         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
4921         WARN_ON(fs_info->chunk_block_rsv.size > 0);
4922         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
4923         WARN_ON(fs_info->delayed_block_rsv.size > 0);
4924         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
4925 }
4926
4927 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
4928                                   struct btrfs_root *root)
4929 {
4930         if (!trans->block_rsv)
4931                 return;
4932
4933         if (!trans->bytes_reserved)
4934                 return;
4935
4936         trace_btrfs_space_reservation(root->fs_info, "transaction",
4937                                       trans->transid, trans->bytes_reserved, 0);
4938         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
4939         trans->bytes_reserved = 0;
4940 }
4941
4942 /* Can only return 0 or -ENOSPC */
4943 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
4944                                   struct inode *inode)
4945 {
4946         struct btrfs_root *root = BTRFS_I(inode)->root;
4947         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
4948         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
4949
4950         /*
4951          * We need to hold space in order to delete our orphan item once we've
4952          * added it, so this takes the reservation so we can release it later
4953          * when we are truly done with the orphan item.
4954          */
4955         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4956         trace_btrfs_space_reservation(root->fs_info, "orphan",
4957                                       btrfs_ino(inode), num_bytes, 1);
4958         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
4959 }
4960
4961 void btrfs_orphan_release_metadata(struct inode *inode)
4962 {
4963         struct btrfs_root *root = BTRFS_I(inode)->root;
4964         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
4965         trace_btrfs_space_reservation(root->fs_info, "orphan",
4966                                       btrfs_ino(inode), num_bytes, 0);
4967         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
4968 }
4969
4970 /*
4971  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
4972  * root: the root of the parent directory
4973  * rsv: block reservation
4974  * items: the number of items that we need do reservation
4975  * qgroup_reserved: used to return the reserved size in qgroup
4976  *
4977  * This function is used to reserve the space for snapshot/subvolume
4978  * creation and deletion. Those operations are different with the
4979  * common file/directory operations, they change two fs/file trees
4980  * and root tree, the number of items that the qgroup reserves is
4981  * different with the free space reservation. So we can not use
4982  * the space reseravtion mechanism in start_transaction().
4983  */
4984 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
4985                                      struct btrfs_block_rsv *rsv,
4986                                      int items,
4987                                      u64 *qgroup_reserved,
4988                                      bool use_global_rsv)
4989 {
4990         u64 num_bytes;
4991         int ret;
4992         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4993
4994         if (root->fs_info->quota_enabled) {
4995                 /* One for parent inode, two for dir entries */
4996                 num_bytes = 3 * root->leafsize;
4997                 ret = btrfs_qgroup_reserve(root, num_bytes);
4998                 if (ret)
4999                         return ret;
5000         } else {
5001                 num_bytes = 0;
5002         }
5003
5004         *qgroup_reserved = num_bytes;
5005
5006         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5007         rsv->space_info = __find_space_info(root->fs_info,
5008                                             BTRFS_BLOCK_GROUP_METADATA);
5009         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5010                                   BTRFS_RESERVE_FLUSH_ALL);
5011
5012         if (ret == -ENOSPC && use_global_rsv)
5013                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
5014
5015         if (ret) {
5016                 if (*qgroup_reserved)
5017                         btrfs_qgroup_free(root, *qgroup_reserved);
5018         }
5019
5020         return ret;
5021 }
5022
5023 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5024                                       struct btrfs_block_rsv *rsv,
5025                                       u64 qgroup_reserved)
5026 {
5027         btrfs_block_rsv_release(root, rsv, (u64)-1);
5028         if (qgroup_reserved)
5029                 btrfs_qgroup_free(root, qgroup_reserved);
5030 }
5031
5032 /**
5033  * drop_outstanding_extent - drop an outstanding extent
5034  * @inode: the inode we're dropping the extent for
5035  *
5036  * This is called when we are freeing up an outstanding extent, either called
5037  * after an error or after an extent is written.  This will return the number of
5038  * reserved extents that need to be freed.  This must be called with
5039  * BTRFS_I(inode)->lock held.
5040  */
5041 static unsigned drop_outstanding_extent(struct inode *inode)
5042 {
5043         unsigned drop_inode_space = 0;
5044         unsigned dropped_extents = 0;
5045
5046         BUG_ON(!BTRFS_I(inode)->outstanding_extents);
5047         BTRFS_I(inode)->outstanding_extents--;
5048
5049         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5050             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5051                                &BTRFS_I(inode)->runtime_flags))
5052                 drop_inode_space = 1;
5053
5054         /*
5055          * If we have more or the same amount of outsanding extents than we have
5056          * reserved then we need to leave the reserved extents count alone.
5057          */
5058         if (BTRFS_I(inode)->outstanding_extents >=
5059             BTRFS_I(inode)->reserved_extents)
5060                 return drop_inode_space;
5061
5062         dropped_extents = BTRFS_I(inode)->reserved_extents -
5063                 BTRFS_I(inode)->outstanding_extents;
5064         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5065         return dropped_extents + drop_inode_space;
5066 }
5067
5068 /**
5069  * calc_csum_metadata_size - return the amount of metada space that must be
5070  *      reserved/free'd for the given bytes.
5071  * @inode: the inode we're manipulating
5072  * @num_bytes: the number of bytes in question
5073  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5074  *
5075  * This adjusts the number of csum_bytes in the inode and then returns the
5076  * correct amount of metadata that must either be reserved or freed.  We
5077  * calculate how many checksums we can fit into one leaf and then divide the
5078  * number of bytes that will need to be checksumed by this value to figure out
5079  * how many checksums will be required.  If we are adding bytes then the number
5080  * may go up and we will return the number of additional bytes that must be
5081  * reserved.  If it is going down we will return the number of bytes that must
5082  * be freed.
5083  *
5084  * This must be called with BTRFS_I(inode)->lock held.
5085  */
5086 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5087                                    int reserve)
5088 {
5089         struct btrfs_root *root = BTRFS_I(inode)->root;
5090         u64 csum_size;
5091         int num_csums_per_leaf;
5092         int num_csums;
5093         int old_csums;
5094
5095         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5096             BTRFS_I(inode)->csum_bytes == 0)
5097                 return 0;
5098
5099         old_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
5100         if (reserve)
5101                 BTRFS_I(inode)->csum_bytes += num_bytes;
5102         else
5103                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5104         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
5105         num_csums_per_leaf = (int)div64_u64(csum_size,
5106                                             sizeof(struct btrfs_csum_item) +
5107                                             sizeof(struct btrfs_disk_key));
5108         num_csums = (int)div64_u64(BTRFS_I(inode)->csum_bytes, root->sectorsize);
5109         num_csums = num_csums + num_csums_per_leaf - 1;
5110         num_csums = num_csums / num_csums_per_leaf;
5111
5112         old_csums = old_csums + num_csums_per_leaf - 1;
5113         old_csums = old_csums / num_csums_per_leaf;
5114
5115         /* No change, no need to reserve more */
5116         if (old_csums == num_csums)
5117                 return 0;
5118
5119         if (reserve)
5120                 return btrfs_calc_trans_metadata_size(root,
5121                                                       num_csums - old_csums);
5122
5123         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5124 }
5125
5126 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5127 {
5128         struct btrfs_root *root = BTRFS_I(inode)->root;
5129         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5130         u64 to_reserve = 0;
5131         u64 csum_bytes;
5132         unsigned nr_extents = 0;
5133         int extra_reserve = 0;
5134         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5135         int ret = 0;
5136         bool delalloc_lock = true;
5137         u64 to_free = 0;
5138         unsigned dropped;
5139
5140         /* If we are a free space inode we need to not flush since we will be in
5141          * the middle of a transaction commit.  We also don't need the delalloc
5142          * mutex since we won't race with anybody.  We need this mostly to make
5143          * lockdep shut its filthy mouth.
5144          */
5145         if (btrfs_is_free_space_inode(inode)) {
5146                 flush = BTRFS_RESERVE_NO_FLUSH;
5147                 delalloc_lock = false;
5148         }
5149
5150         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5151             btrfs_transaction_in_commit(root->fs_info))
5152                 schedule_timeout(1);
5153
5154         if (delalloc_lock)
5155                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5156
5157         num_bytes = ALIGN(num_bytes, root->sectorsize);
5158
5159         spin_lock(&BTRFS_I(inode)->lock);
5160         BTRFS_I(inode)->outstanding_extents++;
5161
5162         if (BTRFS_I(inode)->outstanding_extents >
5163             BTRFS_I(inode)->reserved_extents)
5164                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5165                         BTRFS_I(inode)->reserved_extents;
5166
5167         /*
5168          * Add an item to reserve for updating the inode when we complete the
5169          * delalloc io.
5170          */
5171         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5172                       &BTRFS_I(inode)->runtime_flags)) {
5173                 nr_extents++;
5174                 extra_reserve = 1;
5175         }
5176
5177         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5178         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5179         csum_bytes = BTRFS_I(inode)->csum_bytes;
5180         spin_unlock(&BTRFS_I(inode)->lock);
5181
5182         if (root->fs_info->quota_enabled) {
5183                 ret = btrfs_qgroup_reserve(root, num_bytes +
5184                                            nr_extents * root->leafsize);
5185                 if (ret)
5186                         goto out_fail;
5187         }
5188
5189         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5190         if (unlikely(ret)) {
5191                 if (root->fs_info->quota_enabled)
5192                         btrfs_qgroup_free(root, num_bytes +
5193                                                 nr_extents * root->leafsize);
5194                 goto out_fail;
5195         }
5196
5197         spin_lock(&BTRFS_I(inode)->lock);
5198         if (extra_reserve) {
5199                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5200                         &BTRFS_I(inode)->runtime_flags);
5201                 nr_extents--;
5202         }
5203         BTRFS_I(inode)->reserved_extents += nr_extents;
5204         spin_unlock(&BTRFS_I(inode)->lock);
5205
5206         if (delalloc_lock)
5207                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5208
5209         if (to_reserve)
5210                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5211                                               btrfs_ino(inode), to_reserve, 1);
5212         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5213
5214         return 0;
5215
5216 out_fail:
5217         spin_lock(&BTRFS_I(inode)->lock);
5218         dropped = drop_outstanding_extent(inode);
5219         /*
5220          * If the inodes csum_bytes is the same as the original
5221          * csum_bytes then we know we haven't raced with any free()ers
5222          * so we can just reduce our inodes csum bytes and carry on.
5223          */
5224         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5225                 calc_csum_metadata_size(inode, num_bytes, 0);
5226         } else {
5227                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5228                 u64 bytes;
5229
5230                 /*
5231                  * This is tricky, but first we need to figure out how much we
5232                  * free'd from any free-ers that occured during this
5233                  * reservation, so we reset ->csum_bytes to the csum_bytes
5234                  * before we dropped our lock, and then call the free for the
5235                  * number of bytes that were freed while we were trying our
5236                  * reservation.
5237                  */
5238                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5239                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5240                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5241
5242
5243                 /*
5244                  * Now we need to see how much we would have freed had we not
5245                  * been making this reservation and our ->csum_bytes were not
5246                  * artificially inflated.
5247                  */
5248                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5249                 bytes = csum_bytes - orig_csum_bytes;
5250                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5251
5252                 /*
5253                  * Now reset ->csum_bytes to what it should be.  If bytes is
5254                  * more than to_free then we would have free'd more space had we
5255                  * not had an artificially high ->csum_bytes, so we need to free
5256                  * the remainder.  If bytes is the same or less then we don't
5257                  * need to do anything, the other free-ers did the correct
5258                  * thing.
5259                  */
5260                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5261                 if (bytes > to_free)
5262                         to_free = bytes - to_free;
5263                 else
5264                         to_free = 0;
5265         }
5266         spin_unlock(&BTRFS_I(inode)->lock);
5267         if (dropped)
5268                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5269
5270         if (to_free) {
5271                 btrfs_block_rsv_release(root, block_rsv, to_free);
5272                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5273                                               btrfs_ino(inode), to_free, 0);
5274         }
5275         if (delalloc_lock)
5276                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5277         return ret;
5278 }
5279
5280 /**
5281  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5282  * @inode: the inode to release the reservation for
5283  * @num_bytes: the number of bytes we're releasing
5284  *
5285  * This will release the metadata reservation for an inode.  This can be called
5286  * once we complete IO for a given set of bytes to release their metadata
5287  * reservations.
5288  */
5289 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5290 {
5291         struct btrfs_root *root = BTRFS_I(inode)->root;
5292         u64 to_free = 0;
5293         unsigned dropped;
5294
5295         num_bytes = ALIGN(num_bytes, root->sectorsize);
5296         spin_lock(&BTRFS_I(inode)->lock);
5297         dropped = drop_outstanding_extent(inode);
5298
5299         if (num_bytes)
5300                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5301         spin_unlock(&BTRFS_I(inode)->lock);
5302         if (dropped > 0)
5303                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5304
5305         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5306                                       btrfs_ino(inode), to_free, 0);
5307         if (root->fs_info->quota_enabled) {
5308                 btrfs_qgroup_free(root, num_bytes +
5309                                         dropped * root->leafsize);
5310         }
5311
5312         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5313                                 to_free);
5314 }
5315
5316 /**
5317  * btrfs_delalloc_reserve_space - reserve data and metadata space for delalloc
5318  * @inode: inode we're writing to
5319  * @num_bytes: the number of bytes we want to allocate
5320  *
5321  * This will do the following things
5322  *
5323  * o reserve space in the data space info for num_bytes
5324  * o reserve space in the metadata space info based on number of outstanding
5325  *   extents and how much csums will be needed
5326  * o add to the inodes ->delalloc_bytes
5327  * o add it to the fs_info's delalloc inodes list.
5328  *
5329  * This will return 0 for success and -ENOSPC if there is no space left.
5330  */
5331 int btrfs_delalloc_reserve_space(struct inode *inode, u64 num_bytes)
5332 {
5333         int ret;
5334
5335         ret = btrfs_check_data_free_space(inode, num_bytes);
5336         if (ret)
5337                 return ret;
5338
5339         ret = btrfs_delalloc_reserve_metadata(inode, num_bytes);
5340         if (ret) {
5341                 btrfs_free_reserved_data_space(inode, num_bytes);
5342                 return ret;
5343         }
5344
5345         return 0;
5346 }
5347
5348 /**
5349  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5350  * @inode: inode we're releasing space for
5351  * @num_bytes: the number of bytes we want to free up
5352  *
5353  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5354  * called in the case that we don't need the metadata AND data reservations
5355  * anymore.  So if there is an error or we insert an inline extent.
5356  *
5357  * This function will release the metadata space that was not used and will
5358  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5359  * list if there are no delalloc bytes left.
5360  */
5361 void btrfs_delalloc_release_space(struct inode *inode, u64 num_bytes)
5362 {
5363         btrfs_delalloc_release_metadata(inode, num_bytes);
5364         btrfs_free_reserved_data_space(inode, num_bytes);
5365 }
5366
5367 static int update_block_group(struct btrfs_root *root,
5368                               u64 bytenr, u64 num_bytes, int alloc)
5369 {
5370         struct btrfs_block_group_cache *cache = NULL;
5371         struct btrfs_fs_info *info = root->fs_info;
5372         u64 total = num_bytes;
5373         u64 old_val;
5374         u64 byte_in_group;
5375         int factor;
5376
5377         /* block accounting for super block */
5378         spin_lock(&info->delalloc_root_lock);
5379         old_val = btrfs_super_bytes_used(info->super_copy);
5380         if (alloc)
5381                 old_val += num_bytes;
5382         else
5383                 old_val -= num_bytes;
5384         btrfs_set_super_bytes_used(info->super_copy, old_val);
5385         spin_unlock(&info->delalloc_root_lock);
5386
5387         while (total) {
5388                 cache = btrfs_lookup_block_group(info, bytenr);
5389                 if (!cache)
5390                         return -ENOENT;
5391                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5392                                     BTRFS_BLOCK_GROUP_RAID1 |
5393                                     BTRFS_BLOCK_GROUP_RAID10))
5394                         factor = 2;
5395                 else
5396                         factor = 1;
5397                 /*
5398                  * If this block group has free space cache written out, we
5399                  * need to make sure to load it if we are removing space.  This
5400                  * is because we need the unpinning stage to actually add the
5401                  * space back to the block group, otherwise we will leak space.
5402                  */
5403                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5404                         cache_block_group(cache, 1);
5405
5406                 byte_in_group = bytenr - cache->key.objectid;
5407                 WARN_ON(byte_in_group > cache->key.offset);
5408
5409                 spin_lock(&cache->space_info->lock);
5410                 spin_lock(&cache->lock);
5411
5412                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5413                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5414                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5415
5416                 cache->dirty = 1;
5417                 old_val = btrfs_block_group_used(&cache->item);
5418                 num_bytes = min(total, cache->key.offset - byte_in_group);
5419                 if (alloc) {
5420                         old_val += num_bytes;
5421                         btrfs_set_block_group_used(&cache->item, old_val);
5422                         cache->reserved -= num_bytes;
5423                         cache->space_info->bytes_reserved -= num_bytes;
5424                         cache->space_info->bytes_used += num_bytes;
5425                         cache->space_info->disk_used += num_bytes * factor;
5426                         spin_unlock(&cache->lock);
5427                         spin_unlock(&cache->space_info->lock);
5428                 } else {
5429                         old_val -= num_bytes;
5430                         btrfs_set_block_group_used(&cache->item, old_val);
5431                         cache->pinned += num_bytes;
5432                         cache->space_info->bytes_pinned += num_bytes;
5433                         cache->space_info->bytes_used -= num_bytes;
5434                         cache->space_info->disk_used -= num_bytes * factor;
5435                         spin_unlock(&cache->lock);
5436                         spin_unlock(&cache->space_info->lock);
5437
5438                         set_extent_dirty(info->pinned_extents,
5439                                          bytenr, bytenr + num_bytes - 1,
5440                                          GFP_NOFS | __GFP_NOFAIL);
5441                 }
5442                 btrfs_put_block_group(cache);
5443                 total -= num_bytes;
5444                 bytenr += num_bytes;
5445         }
5446         return 0;
5447 }
5448
5449 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5450 {
5451         struct btrfs_block_group_cache *cache;
5452         u64 bytenr;
5453
5454         spin_lock(&root->fs_info->block_group_cache_lock);
5455         bytenr = root->fs_info->first_logical_byte;
5456         spin_unlock(&root->fs_info->block_group_cache_lock);
5457
5458         if (bytenr < (u64)-1)
5459                 return bytenr;
5460
5461         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5462         if (!cache)
5463                 return 0;
5464
5465         bytenr = cache->key.objectid;
5466         btrfs_put_block_group(cache);
5467
5468         return bytenr;
5469 }
5470
5471 static int pin_down_extent(struct btrfs_root *root,
5472                            struct btrfs_block_group_cache *cache,
5473                            u64 bytenr, u64 num_bytes, int reserved)
5474 {
5475         spin_lock(&cache->space_info->lock);
5476         spin_lock(&cache->lock);
5477         cache->pinned += num_bytes;
5478         cache->space_info->bytes_pinned += num_bytes;
5479         if (reserved) {
5480                 cache->reserved -= num_bytes;
5481                 cache->space_info->bytes_reserved -= num_bytes;
5482         }
5483         spin_unlock(&cache->lock);
5484         spin_unlock(&cache->space_info->lock);
5485
5486         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
5487                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
5488         if (reserved)
5489                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
5490         return 0;
5491 }
5492
5493 /*
5494  * this function must be called within transaction
5495  */
5496 int btrfs_pin_extent(struct btrfs_root *root,
5497                      u64 bytenr, u64 num_bytes, int reserved)
5498 {
5499         struct btrfs_block_group_cache *cache;
5500
5501         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5502         BUG_ON(!cache); /* Logic error */
5503
5504         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
5505
5506         btrfs_put_block_group(cache);
5507         return 0;
5508 }
5509
5510 /*
5511  * this function must be called within transaction
5512  */
5513 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
5514                                     u64 bytenr, u64 num_bytes)
5515 {
5516         struct btrfs_block_group_cache *cache;
5517         int ret;
5518
5519         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
5520         if (!cache)
5521                 return -EINVAL;
5522
5523         /*
5524          * pull in the free space cache (if any) so that our pin
5525          * removes the free space from the cache.  We have load_only set
5526          * to one because the slow code to read in the free extents does check
5527          * the pinned extents.
5528          */
5529         cache_block_group(cache, 1);
5530
5531         pin_down_extent(root, cache, bytenr, num_bytes, 0);
5532
5533         /* remove us from the free space cache (if we're there at all) */
5534         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
5535         btrfs_put_block_group(cache);
5536         return ret;
5537 }
5538
5539 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
5540 {
5541         int ret;
5542         struct btrfs_block_group_cache *block_group;
5543         struct btrfs_caching_control *caching_ctl;
5544
5545         block_group = btrfs_lookup_block_group(root->fs_info, start);
5546         if (!block_group)
5547                 return -EINVAL;
5548
5549         cache_block_group(block_group, 0);
5550         caching_ctl = get_caching_control(block_group);
5551
5552         if (!caching_ctl) {
5553                 /* Logic error */
5554                 BUG_ON(!block_group_cache_done(block_group));
5555                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
5556         } else {
5557                 mutex_lock(&caching_ctl->mutex);
5558
5559                 if (start >= caching_ctl->progress) {
5560                         ret = add_excluded_extent(root, start, num_bytes);
5561                 } else if (start + num_bytes <= caching_ctl->progress) {
5562                         ret = btrfs_remove_free_space(block_group,
5563                                                       start, num_bytes);
5564                 } else {
5565                         num_bytes = caching_ctl->progress - start;
5566                         ret = btrfs_remove_free_space(block_group,
5567                                                       start, num_bytes);
5568                         if (ret)
5569                                 goto out_lock;
5570
5571                         num_bytes = (start + num_bytes) -
5572                                 caching_ctl->progress;
5573                         start = caching_ctl->progress;
5574                         ret = add_excluded_extent(root, start, num_bytes);
5575                 }
5576 out_lock:
5577                 mutex_unlock(&caching_ctl->mutex);
5578                 put_caching_control(caching_ctl);
5579         }
5580         btrfs_put_block_group(block_group);
5581         return ret;
5582 }
5583
5584 int btrfs_exclude_logged_extents(struct btrfs_root *log,
5585                                  struct extent_buffer *eb)
5586 {
5587         struct btrfs_file_extent_item *item;
5588         struct btrfs_key key;
5589         int found_type;
5590         int i;
5591
5592         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
5593                 return 0;
5594
5595         for (i = 0; i < btrfs_header_nritems(eb); i++) {
5596                 btrfs_item_key_to_cpu(eb, &key, i);
5597                 if (key.type != BTRFS_EXTENT_DATA_KEY)
5598                         continue;
5599                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
5600                 found_type = btrfs_file_extent_type(eb, item);
5601                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
5602                         continue;
5603                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
5604                         continue;
5605                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
5606                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
5607                 __exclude_logged_extent(log, key.objectid, key.offset);
5608         }
5609
5610         return 0;
5611 }
5612
5613 /**
5614  * btrfs_update_reserved_bytes - update the block_group and space info counters
5615  * @cache:      The cache we are manipulating
5616  * @num_bytes:  The number of bytes in question
5617  * @reserve:    One of the reservation enums
5618  * @delalloc:   The blocks are allocated for the delalloc write
5619  *
5620  * This is called by the allocator when it reserves space, or by somebody who is
5621  * freeing space that was never actually used on disk.  For example if you
5622  * reserve some space for a new leaf in transaction A and before transaction A
5623  * commits you free that leaf, you call this with reserve set to 0 in order to
5624  * clear the reservation.
5625  *
5626  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5627  * ENOSPC accounting.  For data we handle the reservation through clearing the
5628  * delalloc bits in the io_tree.  We have to do this since we could end up
5629  * allocating less disk space for the amount of data we have reserved in the
5630  * case of compression.
5631  *
5632  * If this is a reservation and the block group has become read only we cannot
5633  * make the reservation and return -EAGAIN, otherwise this function always
5634  * succeeds.
5635  */
5636 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5637                                        u64 num_bytes, int reserve, int delalloc)
5638 {
5639         struct btrfs_space_info *space_info = cache->space_info;
5640         int ret = 0;
5641
5642         spin_lock(&space_info->lock);
5643         spin_lock(&cache->lock);
5644         if (reserve != RESERVE_FREE) {
5645                 if (cache->ro) {
5646                         ret = -EAGAIN;
5647                 } else {
5648                         cache->reserved += num_bytes;
5649                         space_info->bytes_reserved += num_bytes;
5650                         if (reserve == RESERVE_ALLOC) {
5651                                 trace_btrfs_space_reservation(cache->fs_info,
5652                                                 "space_info", space_info->flags,
5653                                                 num_bytes, 0);
5654                                 space_info->bytes_may_use -= num_bytes;
5655                         }
5656
5657                         if (delalloc)
5658                                 cache->delalloc_bytes += num_bytes;
5659                 }
5660         } else {
5661                 if (cache->ro)
5662                         space_info->bytes_readonly += num_bytes;
5663                 cache->reserved -= num_bytes;
5664                 space_info->bytes_reserved -= num_bytes;
5665
5666                 if (delalloc)
5667                         cache->delalloc_bytes -= num_bytes;
5668         }
5669         spin_unlock(&cache->lock);
5670         spin_unlock(&space_info->lock);
5671         return ret;
5672 }
5673
5674 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5675                                 struct btrfs_root *root)
5676 {
5677         struct btrfs_fs_info *fs_info = root->fs_info;
5678         struct btrfs_caching_control *next;
5679         struct btrfs_caching_control *caching_ctl;
5680         struct btrfs_block_group_cache *cache;
5681         struct btrfs_space_info *space_info;
5682
5683         down_write(&fs_info->commit_root_sem);
5684
5685         list_for_each_entry_safe(caching_ctl, next,
5686                                  &fs_info->caching_block_groups, list) {
5687                 cache = caching_ctl->block_group;
5688                 if (block_group_cache_done(cache)) {
5689                         cache->last_byte_to_unpin = (u64)-1;
5690                         list_del_init(&caching_ctl->list);
5691                         put_caching_control(caching_ctl);
5692                 } else {
5693                         cache->last_byte_to_unpin = caching_ctl->progress;
5694                 }
5695         }
5696
5697         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5698                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5699         else
5700                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5701
5702         up_write(&fs_info->commit_root_sem);
5703
5704         list_for_each_entry_rcu(space_info, &fs_info->space_info, list)
5705                 percpu_counter_set(&space_info->total_bytes_pinned, 0);
5706
5707         update_global_block_rsv(fs_info);
5708 }
5709
5710 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5711 {
5712         struct btrfs_fs_info *fs_info = root->fs_info;
5713         struct btrfs_block_group_cache *cache = NULL;
5714         struct btrfs_space_info *space_info;
5715         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5716         u64 len;
5717         bool readonly;
5718
5719         while (start <= end) {
5720                 readonly = false;
5721                 if (!cache ||
5722                     start >= cache->key.objectid + cache->key.offset) {
5723                         if (cache)
5724                                 btrfs_put_block_group(cache);
5725                         cache = btrfs_lookup_block_group(fs_info, start);
5726                         BUG_ON(!cache); /* Logic error */
5727                 }
5728
5729                 len = cache->key.objectid + cache->key.offset - start;
5730                 len = min(len, end + 1 - start);
5731
5732                 if (start < cache->last_byte_to_unpin) {
5733                         len = min(len, cache->last_byte_to_unpin - start);
5734                         btrfs_add_free_space(cache, start, len);
5735                 }
5736
5737                 start += len;
5738                 space_info = cache->space_info;
5739
5740                 spin_lock(&space_info->lock);
5741                 spin_lock(&cache->lock);
5742                 cache->pinned -= len;
5743                 space_info->bytes_pinned -= len;
5744                 if (cache->ro) {
5745                         space_info->bytes_readonly += len;
5746                         readonly = true;
5747                 }
5748                 spin_unlock(&cache->lock);
5749                 if (!readonly && global_rsv->space_info == space_info) {
5750                         spin_lock(&global_rsv->lock);
5751                         if (!global_rsv->full) {
5752                                 len = min(len, global_rsv->size -
5753                                           global_rsv->reserved);
5754                                 global_rsv->reserved += len;
5755                                 space_info->bytes_may_use += len;
5756                                 if (global_rsv->reserved >= global_rsv->size)
5757                                         global_rsv->full = 1;
5758                         }
5759                         spin_unlock(&global_rsv->lock);
5760                 }
5761                 spin_unlock(&space_info->lock);
5762         }
5763
5764         if (cache)
5765                 btrfs_put_block_group(cache);
5766         return 0;
5767 }
5768
5769 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5770                                struct btrfs_root *root)
5771 {
5772         struct btrfs_fs_info *fs_info = root->fs_info;
5773         struct extent_io_tree *unpin;
5774         u64 start;
5775         u64 end;
5776         int ret;
5777
5778         if (trans->aborted)
5779                 return 0;
5780
5781         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5782                 unpin = &fs_info->freed_extents[1];
5783         else
5784                 unpin = &fs_info->freed_extents[0];
5785
5786         while (1) {
5787                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5788                                             EXTENT_DIRTY, NULL);
5789                 if (ret)
5790                         break;
5791
5792                 if (btrfs_test_opt(root, DISCARD))
5793                         ret = btrfs_discard_extent(root, start,
5794                                                    end + 1 - start, NULL);
5795
5796                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5797                 unpin_extent_range(root, start, end);
5798                 cond_resched();
5799         }
5800
5801         return 0;
5802 }
5803
5804 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
5805                              u64 owner, u64 root_objectid)
5806 {
5807         struct btrfs_space_info *space_info;
5808         u64 flags;
5809
5810         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5811                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
5812                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
5813                 else
5814                         flags = BTRFS_BLOCK_GROUP_METADATA;
5815         } else {
5816                 flags = BTRFS_BLOCK_GROUP_DATA;
5817         }
5818
5819         space_info = __find_space_info(fs_info, flags);
5820         BUG_ON(!space_info); /* Logic bug */
5821         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
5822 }
5823
5824
5825 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5826                                 struct btrfs_root *root,
5827                                 u64 bytenr, u64 num_bytes, u64 parent,
5828                                 u64 root_objectid, u64 owner_objectid,
5829                                 u64 owner_offset, int refs_to_drop,
5830                                 struct btrfs_delayed_extent_op *extent_op,
5831                                 int no_quota)
5832 {
5833         struct btrfs_key key;
5834         struct btrfs_path *path;
5835         struct btrfs_fs_info *info = root->fs_info;
5836         struct btrfs_root *extent_root = info->extent_root;
5837         struct extent_buffer *leaf;
5838         struct btrfs_extent_item *ei;
5839         struct btrfs_extent_inline_ref *iref;
5840         int ret;
5841         int is_data;
5842         int extent_slot = 0;
5843         int found_extent = 0;
5844         int num_to_del = 1;
5845         u32 item_size;
5846         u64 refs;
5847         int last_ref = 0;
5848         enum btrfs_qgroup_operation_type type = BTRFS_QGROUP_OPER_SUB_EXCL;
5849         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
5850                                                  SKINNY_METADATA);
5851
5852         if (!info->quota_enabled || !is_fstree(root_objectid))
5853                 no_quota = 1;
5854
5855         path = btrfs_alloc_path();
5856         if (!path)
5857                 return -ENOMEM;
5858
5859         path->reada = 1;
5860         path->leave_spinning = 1;
5861
5862         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5863         BUG_ON(!is_data && refs_to_drop != 1);
5864
5865         if (is_data)
5866                 skinny_metadata = 0;
5867
5868         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5869                                     bytenr, num_bytes, parent,
5870                                     root_objectid, owner_objectid,
5871                                     owner_offset);
5872         if (ret == 0) {
5873                 extent_slot = path->slots[0];
5874                 while (extent_slot >= 0) {
5875                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5876                                               extent_slot);
5877                         if (key.objectid != bytenr)
5878                                 break;
5879                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5880                             key.offset == num_bytes) {
5881                                 found_extent = 1;
5882                                 break;
5883                         }
5884                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
5885                             key.offset == owner_objectid) {
5886                                 found_extent = 1;
5887                                 break;
5888                         }
5889                         if (path->slots[0] - extent_slot > 5)
5890                                 break;
5891                         extent_slot--;
5892                 }
5893 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5894                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5895                 if (found_extent && item_size < sizeof(*ei))
5896                         found_extent = 0;
5897 #endif
5898                 if (!found_extent) {
5899                         BUG_ON(iref);
5900                         ret = remove_extent_backref(trans, extent_root, path,
5901                                                     NULL, refs_to_drop,
5902                                                     is_data, &last_ref);
5903                         if (ret) {
5904                                 btrfs_abort_transaction(trans, extent_root, ret);
5905                                 goto out;
5906                         }
5907                         btrfs_release_path(path);
5908                         path->leave_spinning = 1;
5909
5910                         key.objectid = bytenr;
5911                         key.type = BTRFS_EXTENT_ITEM_KEY;
5912                         key.offset = num_bytes;
5913
5914                         if (!is_data && skinny_metadata) {
5915                                 key.type = BTRFS_METADATA_ITEM_KEY;
5916                                 key.offset = owner_objectid;
5917                         }
5918
5919                         ret = btrfs_search_slot(trans, extent_root,
5920                                                 &key, path, -1, 1);
5921                         if (ret > 0 && skinny_metadata && path->slots[0]) {
5922                                 /*
5923                                  * Couldn't find our skinny metadata item,
5924                                  * see if we have ye olde extent item.
5925                                  */
5926                                 path->slots[0]--;
5927                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
5928                                                       path->slots[0]);
5929                                 if (key.objectid == bytenr &&
5930                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
5931                                     key.offset == num_bytes)
5932                                         ret = 0;
5933                         }
5934
5935                         if (ret > 0 && skinny_metadata) {
5936                                 skinny_metadata = false;
5937                                 key.objectid = bytenr;
5938                                 key.type = BTRFS_EXTENT_ITEM_KEY;
5939                                 key.offset = num_bytes;
5940                                 btrfs_release_path(path);
5941                                 ret = btrfs_search_slot(trans, extent_root,
5942                                                         &key, path, -1, 1);
5943                         }
5944
5945                         if (ret) {
5946                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5947                                         ret, bytenr);
5948                                 if (ret > 0)
5949                                         btrfs_print_leaf(extent_root,
5950                                                          path->nodes[0]);
5951                         }
5952                         if (ret < 0) {
5953                                 btrfs_abort_transaction(trans, extent_root, ret);
5954                                 goto out;
5955                         }
5956                         extent_slot = path->slots[0];
5957                 }
5958         } else if (WARN_ON(ret == -ENOENT)) {
5959                 btrfs_print_leaf(extent_root, path->nodes[0]);
5960                 btrfs_err(info,
5961                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
5962                         bytenr, parent, root_objectid, owner_objectid,
5963                         owner_offset);
5964                 btrfs_abort_transaction(trans, extent_root, ret);
5965                 goto out;
5966         } else {
5967                 btrfs_abort_transaction(trans, extent_root, ret);
5968                 goto out;
5969         }
5970
5971         leaf = path->nodes[0];
5972         item_size = btrfs_item_size_nr(leaf, extent_slot);
5973 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5974         if (item_size < sizeof(*ei)) {
5975                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5976                 ret = convert_extent_item_v0(trans, extent_root, path,
5977                                              owner_objectid, 0);
5978                 if (ret < 0) {
5979                         btrfs_abort_transaction(trans, extent_root, ret);
5980                         goto out;
5981                 }
5982
5983                 btrfs_release_path(path);
5984                 path->leave_spinning = 1;
5985
5986                 key.objectid = bytenr;
5987                 key.type = BTRFS_EXTENT_ITEM_KEY;
5988                 key.offset = num_bytes;
5989
5990                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5991                                         -1, 1);
5992                 if (ret) {
5993                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5994                                 ret, bytenr);
5995                         btrfs_print_leaf(extent_root, path->nodes[0]);
5996                 }
5997                 if (ret < 0) {
5998                         btrfs_abort_transaction(trans, extent_root, ret);
5999                         goto out;
6000                 }
6001
6002                 extent_slot = path->slots[0];
6003                 leaf = path->nodes[0];
6004                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6005         }
6006 #endif
6007         BUG_ON(item_size < sizeof(*ei));
6008         ei = btrfs_item_ptr(leaf, extent_slot,
6009                             struct btrfs_extent_item);
6010         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6011             key.type == BTRFS_EXTENT_ITEM_KEY) {
6012                 struct btrfs_tree_block_info *bi;
6013                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6014                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6015                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6016         }
6017
6018         refs = btrfs_extent_refs(leaf, ei);
6019         if (refs < refs_to_drop) {
6020                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6021                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6022                 ret = -EINVAL;
6023                 btrfs_abort_transaction(trans, extent_root, ret);
6024                 goto out;
6025         }
6026         refs -= refs_to_drop;
6027
6028         if (refs > 0) {
6029                 type = BTRFS_QGROUP_OPER_SUB_SHARED;
6030                 if (extent_op)
6031                         __run_delayed_extent_op(extent_op, leaf, ei);
6032                 /*
6033                  * In the case of inline back ref, reference count will
6034                  * be updated by remove_extent_backref
6035                  */
6036                 if (iref) {
6037                         BUG_ON(!found_extent);
6038                 } else {
6039                         btrfs_set_extent_refs(leaf, ei, refs);
6040                         btrfs_mark_buffer_dirty(leaf);
6041                 }
6042                 if (found_extent) {
6043                         ret = remove_extent_backref(trans, extent_root, path,
6044                                                     iref, refs_to_drop,
6045                                                     is_data, &last_ref);
6046                         if (ret) {
6047                                 btrfs_abort_transaction(trans, extent_root, ret);
6048                                 goto out;
6049                         }
6050                 }
6051                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6052                                  root_objectid);
6053         } else {
6054                 if (found_extent) {
6055                         BUG_ON(is_data && refs_to_drop !=
6056                                extent_data_ref_count(root, path, iref));
6057                         if (iref) {
6058                                 BUG_ON(path->slots[0] != extent_slot);
6059                         } else {
6060                                 BUG_ON(path->slots[0] != extent_slot + 1);
6061                                 path->slots[0] = extent_slot;
6062                                 num_to_del = 2;
6063                         }
6064                 }
6065
6066                 last_ref = 1;
6067                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6068                                       num_to_del);
6069                 if (ret) {
6070                         btrfs_abort_transaction(trans, extent_root, ret);
6071                         goto out;
6072                 }
6073                 btrfs_release_path(path);
6074
6075                 if (is_data) {
6076                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6077                         if (ret) {
6078                                 btrfs_abort_transaction(trans, extent_root, ret);
6079                                 goto out;
6080                         }
6081                 }
6082
6083                 ret = update_block_group(root, bytenr, num_bytes, 0);
6084                 if (ret) {
6085                         btrfs_abort_transaction(trans, extent_root, ret);
6086                         goto out;
6087                 }
6088         }
6089         btrfs_release_path(path);
6090
6091         /* Deal with the quota accounting */
6092         if (!ret && last_ref && !no_quota) {
6093                 int mod_seq = 0;
6094
6095                 if (owner_objectid >= BTRFS_FIRST_FREE_OBJECTID &&
6096                     type == BTRFS_QGROUP_OPER_SUB_SHARED)
6097                         mod_seq = 1;
6098
6099                 ret = btrfs_qgroup_record_ref(trans, info, root_objectid,
6100                                               bytenr, num_bytes, type,
6101                                               mod_seq);
6102         }
6103 out:
6104         btrfs_free_path(path);
6105         return ret;
6106 }
6107
6108 /*
6109  * when we free an block, it is possible (and likely) that we free the last
6110  * delayed ref for that extent as well.  This searches the delayed ref tree for
6111  * a given extent, and if there are no other delayed refs to be processed, it
6112  * removes it from the tree.
6113  */
6114 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6115                                       struct btrfs_root *root, u64 bytenr)
6116 {
6117         struct btrfs_delayed_ref_head *head;
6118         struct btrfs_delayed_ref_root *delayed_refs;
6119         int ret = 0;
6120
6121         delayed_refs = &trans->transaction->delayed_refs;
6122         spin_lock(&delayed_refs->lock);
6123         head = btrfs_find_delayed_ref_head(trans, bytenr);
6124         if (!head)
6125                 goto out_delayed_unlock;
6126
6127         spin_lock(&head->lock);
6128         if (rb_first(&head->ref_root))
6129                 goto out;
6130
6131         if (head->extent_op) {
6132                 if (!head->must_insert_reserved)
6133                         goto out;
6134                 btrfs_free_delayed_extent_op(head->extent_op);
6135                 head->extent_op = NULL;
6136         }
6137
6138         /*
6139          * waiting for the lock here would deadlock.  If someone else has it
6140          * locked they are already in the process of dropping it anyway
6141          */
6142         if (!mutex_trylock(&head->mutex))
6143                 goto out;
6144
6145         /*
6146          * at this point we have a head with no other entries.  Go
6147          * ahead and process it.
6148          */
6149         head->node.in_tree = 0;
6150         rb_erase(&head->href_node, &delayed_refs->href_root);
6151
6152         atomic_dec(&delayed_refs->num_entries);
6153
6154         /*
6155          * we don't take a ref on the node because we're removing it from the
6156          * tree, so we just steal the ref the tree was holding.
6157          */
6158         delayed_refs->num_heads--;
6159         if (head->processing == 0)
6160                 delayed_refs->num_heads_ready--;
6161         head->processing = 0;
6162         spin_unlock(&head->lock);
6163         spin_unlock(&delayed_refs->lock);
6164
6165         BUG_ON(head->extent_op);
6166         if (head->must_insert_reserved)
6167                 ret = 1;
6168
6169         mutex_unlock(&head->mutex);
6170         btrfs_put_delayed_ref(&head->node);
6171         return ret;
6172 out:
6173         spin_unlock(&head->lock);
6174
6175 out_delayed_unlock:
6176         spin_unlock(&delayed_refs->lock);
6177         return 0;
6178 }
6179
6180 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6181                            struct btrfs_root *root,
6182                            struct extent_buffer *buf,
6183                            u64 parent, int last_ref)
6184 {
6185         struct btrfs_block_group_cache *cache = NULL;
6186         int pin = 1;
6187         int ret;
6188
6189         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6190                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6191                                         buf->start, buf->len,
6192                                         parent, root->root_key.objectid,
6193                                         btrfs_header_level(buf),
6194                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
6195                 BUG_ON(ret); /* -ENOMEM */
6196         }
6197
6198         if (!last_ref)
6199                 return;
6200
6201         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6202
6203         if (btrfs_header_generation(buf) == trans->transid) {
6204                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6205                         ret = check_ref_cleanup(trans, root, buf->start);
6206                         if (!ret)
6207                                 goto out;
6208                 }
6209
6210                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6211                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6212                         goto out;
6213                 }
6214
6215                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6216
6217                 btrfs_add_free_space(cache, buf->start, buf->len);
6218                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
6219                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6220                 pin = 0;
6221         }
6222 out:
6223         if (pin)
6224                 add_pinned_bytes(root->fs_info, buf->len,
6225                                  btrfs_header_level(buf),
6226                                  root->root_key.objectid);
6227
6228         /*
6229          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6230          * anymore.
6231          */
6232         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6233         btrfs_put_block_group(cache);
6234 }
6235
6236 /* Can return -ENOMEM */
6237 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6238                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6239                       u64 owner, u64 offset, int no_quota)
6240 {
6241         int ret;
6242         struct btrfs_fs_info *fs_info = root->fs_info;
6243
6244 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
6245         if (unlikely(test_bit(BTRFS_ROOT_DUMMY_ROOT, &root->state)))
6246                 return 0;
6247 #endif
6248         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6249
6250         /*
6251          * tree log blocks never actually go into the extent allocation
6252          * tree, just update pinning info and exit early.
6253          */
6254         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6255                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6256                 /* unlocks the pinned mutex */
6257                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6258                 ret = 0;
6259         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6260                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6261                                         num_bytes,
6262                                         parent, root_objectid, (int)owner,
6263                                         BTRFS_DROP_DELAYED_REF, NULL, no_quota);
6264         } else {
6265                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6266                                                 num_bytes,
6267                                                 parent, root_objectid, owner,
6268                                                 offset, BTRFS_DROP_DELAYED_REF,
6269                                                 NULL, no_quota);
6270         }
6271         return ret;
6272 }
6273
6274 static u64 stripe_align(struct btrfs_root *root,
6275                         struct btrfs_block_group_cache *cache,
6276                         u64 val, u64 num_bytes)
6277 {
6278         u64 ret = ALIGN(val, root->stripesize);
6279         return ret;
6280 }
6281
6282 /*
6283  * when we wait for progress in the block group caching, its because
6284  * our allocation attempt failed at least once.  So, we must sleep
6285  * and let some progress happen before we try again.
6286  *
6287  * This function will sleep at least once waiting for new free space to
6288  * show up, and then it will check the block group free space numbers
6289  * for our min num_bytes.  Another option is to have it go ahead
6290  * and look in the rbtree for a free extent of a given size, but this
6291  * is a good start.
6292  *
6293  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6294  * any of the information in this block group.
6295  */
6296 static noinline void
6297 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6298                                 u64 num_bytes)
6299 {
6300         struct btrfs_caching_control *caching_ctl;
6301
6302         caching_ctl = get_caching_control(cache);
6303         if (!caching_ctl)
6304                 return;
6305
6306         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6307                    (cache->free_space_ctl->free_space >= num_bytes));
6308
6309         put_caching_control(caching_ctl);
6310 }
6311
6312 static noinline int
6313 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6314 {
6315         struct btrfs_caching_control *caching_ctl;
6316         int ret = 0;
6317
6318         caching_ctl = get_caching_control(cache);
6319         if (!caching_ctl)
6320                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6321
6322         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6323         if (cache->cached == BTRFS_CACHE_ERROR)
6324                 ret = -EIO;
6325         put_caching_control(caching_ctl);
6326         return ret;
6327 }
6328
6329 int __get_raid_index(u64 flags)
6330 {
6331         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6332                 return BTRFS_RAID_RAID10;
6333         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6334                 return BTRFS_RAID_RAID1;
6335         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6336                 return BTRFS_RAID_DUP;
6337         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6338                 return BTRFS_RAID_RAID0;
6339         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6340                 return BTRFS_RAID_RAID5;
6341         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6342                 return BTRFS_RAID_RAID6;
6343
6344         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6345 }
6346
6347 int get_block_group_index(struct btrfs_block_group_cache *cache)
6348 {
6349         return __get_raid_index(cache->flags);
6350 }
6351
6352 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6353         [BTRFS_RAID_RAID10]     = "raid10",
6354         [BTRFS_RAID_RAID1]      = "raid1",
6355         [BTRFS_RAID_DUP]        = "dup",
6356         [BTRFS_RAID_RAID0]      = "raid0",
6357         [BTRFS_RAID_SINGLE]     = "single",
6358         [BTRFS_RAID_RAID5]      = "raid5",
6359         [BTRFS_RAID_RAID6]      = "raid6",
6360 };
6361
6362 static const char *get_raid_name(enum btrfs_raid_types type)
6363 {
6364         if (type >= BTRFS_NR_RAID_TYPES)
6365                 return NULL;
6366
6367         return btrfs_raid_type_names[type];
6368 }
6369
6370 enum btrfs_loop_type {
6371         LOOP_CACHING_NOWAIT = 0,
6372         LOOP_CACHING_WAIT = 1,
6373         LOOP_ALLOC_CHUNK = 2,
6374         LOOP_NO_EMPTY_SIZE = 3,
6375 };
6376
6377 static inline void
6378 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
6379                        int delalloc)
6380 {
6381         if (delalloc)
6382                 down_read(&cache->data_rwsem);
6383 }
6384
6385 static inline void
6386 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
6387                        int delalloc)
6388 {
6389         btrfs_get_block_group(cache);
6390         if (delalloc)
6391                 down_read(&cache->data_rwsem);
6392 }
6393
6394 static struct btrfs_block_group_cache *
6395 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
6396                    struct btrfs_free_cluster *cluster,
6397                    int delalloc)
6398 {
6399         struct btrfs_block_group_cache *used_bg;
6400         bool locked = false;
6401 again:
6402         spin_lock(&cluster->refill_lock);
6403         if (locked) {
6404                 if (used_bg == cluster->block_group)
6405                         return used_bg;
6406
6407                 up_read(&used_bg->data_rwsem);
6408                 btrfs_put_block_group(used_bg);
6409         }
6410
6411         used_bg = cluster->block_group;
6412         if (!used_bg)
6413                 return NULL;
6414
6415         if (used_bg == block_group)
6416                 return used_bg;
6417
6418         btrfs_get_block_group(used_bg);
6419
6420         if (!delalloc)
6421                 return used_bg;
6422
6423         if (down_read_trylock(&used_bg->data_rwsem))
6424                 return used_bg;
6425
6426         spin_unlock(&cluster->refill_lock);
6427         down_read(&used_bg->data_rwsem);
6428         locked = true;
6429         goto again;
6430 }
6431
6432 static inline void
6433 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
6434                          int delalloc)
6435 {
6436         if (delalloc)
6437                 up_read(&cache->data_rwsem);
6438         btrfs_put_block_group(cache);
6439 }
6440
6441 /*
6442  * walks the btree of allocated extents and find a hole of a given size.
6443  * The key ins is changed to record the hole:
6444  * ins->objectid == start position
6445  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6446  * ins->offset == the size of the hole.
6447  * Any available blocks before search_start are skipped.
6448  *
6449  * If there is no suitable free space, we will record the max size of
6450  * the free space extent currently.
6451  */
6452 static noinline int find_free_extent(struct btrfs_root *orig_root,
6453                                      u64 num_bytes, u64 empty_size,
6454                                      u64 hint_byte, struct btrfs_key *ins,
6455                                      u64 flags, int delalloc)
6456 {
6457         int ret = 0;
6458         struct btrfs_root *root = orig_root->fs_info->extent_root;
6459         struct btrfs_free_cluster *last_ptr = NULL;
6460         struct btrfs_block_group_cache *block_group = NULL;
6461         u64 search_start = 0;
6462         u64 max_extent_size = 0;
6463         int empty_cluster = 2 * 1024 * 1024;
6464         struct btrfs_space_info *space_info;
6465         int loop = 0;
6466         int index = __get_raid_index(flags);
6467         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6468                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6469         bool failed_cluster_refill = false;
6470         bool failed_alloc = false;
6471         bool use_cluster = true;
6472         bool have_caching_bg = false;
6473
6474         WARN_ON(num_bytes < root->sectorsize);
6475         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
6476         ins->objectid = 0;
6477         ins->offset = 0;
6478
6479         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6480
6481         space_info = __find_space_info(root->fs_info, flags);
6482         if (!space_info) {
6483                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6484                 return -ENOSPC;
6485         }
6486
6487         /*
6488          * If the space info is for both data and metadata it means we have a
6489          * small filesystem and we can't use the clustering stuff.
6490          */
6491         if (btrfs_mixed_space_info(space_info))
6492                 use_cluster = false;
6493
6494         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6495                 last_ptr = &root->fs_info->meta_alloc_cluster;
6496                 if (!btrfs_test_opt(root, SSD))
6497                         empty_cluster = 64 * 1024;
6498         }
6499
6500         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6501             btrfs_test_opt(root, SSD)) {
6502                 last_ptr = &root->fs_info->data_alloc_cluster;
6503         }
6504
6505         if (last_ptr) {
6506                 spin_lock(&last_ptr->lock);
6507                 if (last_ptr->block_group)
6508                         hint_byte = last_ptr->window_start;
6509                 spin_unlock(&last_ptr->lock);
6510         }
6511
6512         search_start = max(search_start, first_logical_byte(root, 0));
6513         search_start = max(search_start, hint_byte);
6514
6515         if (!last_ptr)
6516                 empty_cluster = 0;
6517
6518         if (search_start == hint_byte) {
6519                 block_group = btrfs_lookup_block_group(root->fs_info,
6520                                                        search_start);
6521                 /*
6522                  * we don't want to use the block group if it doesn't match our
6523                  * allocation bits, or if its not cached.
6524                  *
6525                  * However if we are re-searching with an ideal block group
6526                  * picked out then we don't care that the block group is cached.
6527                  */
6528                 if (block_group && block_group_bits(block_group, flags) &&
6529                     block_group->cached != BTRFS_CACHE_NO) {
6530                         down_read(&space_info->groups_sem);
6531                         if (list_empty(&block_group->list) ||
6532                             block_group->ro) {
6533                                 /*
6534                                  * someone is removing this block group,
6535                                  * we can't jump into the have_block_group
6536                                  * target because our list pointers are not
6537                                  * valid
6538                                  */
6539                                 btrfs_put_block_group(block_group);
6540                                 up_read(&space_info->groups_sem);
6541                         } else {
6542                                 index = get_block_group_index(block_group);
6543                                 btrfs_lock_block_group(block_group, delalloc);
6544                                 goto have_block_group;
6545                         }
6546                 } else if (block_group) {
6547                         btrfs_put_block_group(block_group);
6548                 }
6549         }
6550 search:
6551         have_caching_bg = false;
6552         down_read(&space_info->groups_sem);
6553         list_for_each_entry(block_group, &space_info->block_groups[index],
6554                             list) {
6555                 u64 offset;
6556                 int cached;
6557
6558                 btrfs_grab_block_group(block_group, delalloc);
6559                 search_start = block_group->key.objectid;
6560
6561                 /*
6562                  * this can happen if we end up cycling through all the
6563                  * raid types, but we want to make sure we only allocate
6564                  * for the proper type.
6565                  */
6566                 if (!block_group_bits(block_group, flags)) {
6567                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6568                                 BTRFS_BLOCK_GROUP_RAID1 |
6569                                 BTRFS_BLOCK_GROUP_RAID5 |
6570                                 BTRFS_BLOCK_GROUP_RAID6 |
6571                                 BTRFS_BLOCK_GROUP_RAID10;
6572
6573                         /*
6574                          * if they asked for extra copies and this block group
6575                          * doesn't provide them, bail.  This does allow us to
6576                          * fill raid0 from raid1.
6577                          */
6578                         if ((flags & extra) && !(block_group->flags & extra))
6579                                 goto loop;
6580                 }
6581
6582 have_block_group:
6583                 cached = block_group_cache_done(block_group);
6584                 if (unlikely(!cached)) {
6585                         ret = cache_block_group(block_group, 0);
6586                         BUG_ON(ret < 0);
6587                         ret = 0;
6588                 }
6589
6590                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
6591                         goto loop;
6592                 if (unlikely(block_group->ro))
6593                         goto loop;
6594
6595                 /*
6596                  * Ok we want to try and use the cluster allocator, so
6597                  * lets look there
6598                  */
6599                 if (last_ptr) {
6600                         struct btrfs_block_group_cache *used_block_group;
6601                         unsigned long aligned_cluster;
6602                         /*
6603                          * the refill lock keeps out other
6604                          * people trying to start a new cluster
6605                          */
6606                         used_block_group = btrfs_lock_cluster(block_group,
6607                                                               last_ptr,
6608                                                               delalloc);
6609                         if (!used_block_group)
6610                                 goto refill_cluster;
6611
6612                         if (used_block_group != block_group &&
6613                             (used_block_group->ro ||
6614                              !block_group_bits(used_block_group, flags)))
6615                                 goto release_cluster;
6616
6617                         offset = btrfs_alloc_from_cluster(used_block_group,
6618                                                 last_ptr,
6619                                                 num_bytes,
6620                                                 used_block_group->key.objectid,
6621                                                 &max_extent_size);
6622                         if (offset) {
6623                                 /* we have a block, we're done */
6624                                 spin_unlock(&last_ptr->refill_lock);
6625                                 trace_btrfs_reserve_extent_cluster(root,
6626                                                 used_block_group,
6627                                                 search_start, num_bytes);
6628                                 if (used_block_group != block_group) {
6629                                         btrfs_release_block_group(block_group,
6630                                                                   delalloc);
6631                                         block_group = used_block_group;
6632                                 }
6633                                 goto checks;
6634                         }
6635
6636                         WARN_ON(last_ptr->block_group != used_block_group);
6637 release_cluster:
6638                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6639                          * set up a new clusters, so lets just skip it
6640                          * and let the allocator find whatever block
6641                          * it can find.  If we reach this point, we
6642                          * will have tried the cluster allocator
6643                          * plenty of times and not have found
6644                          * anything, so we are likely way too
6645                          * fragmented for the clustering stuff to find
6646                          * anything.
6647                          *
6648                          * However, if the cluster is taken from the
6649                          * current block group, release the cluster
6650                          * first, so that we stand a better chance of
6651                          * succeeding in the unclustered
6652                          * allocation.  */
6653                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6654                             used_block_group != block_group) {
6655                                 spin_unlock(&last_ptr->refill_lock);
6656                                 btrfs_release_block_group(used_block_group,
6657                                                           delalloc);
6658                                 goto unclustered_alloc;
6659                         }
6660
6661                         /*
6662                          * this cluster didn't work out, free it and
6663                          * start over
6664                          */
6665                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6666
6667                         if (used_block_group != block_group)
6668                                 btrfs_release_block_group(used_block_group,
6669                                                           delalloc);
6670 refill_cluster:
6671                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6672                                 spin_unlock(&last_ptr->refill_lock);
6673                                 goto unclustered_alloc;
6674                         }
6675
6676                         aligned_cluster = max_t(unsigned long,
6677                                                 empty_cluster + empty_size,
6678                                               block_group->full_stripe_len);
6679
6680                         /* allocate a cluster in this block group */
6681                         ret = btrfs_find_space_cluster(root, block_group,
6682                                                        last_ptr, search_start,
6683                                                        num_bytes,
6684                                                        aligned_cluster);
6685                         if (ret == 0) {
6686                                 /*
6687                                  * now pull our allocation out of this
6688                                  * cluster
6689                                  */
6690                                 offset = btrfs_alloc_from_cluster(block_group,
6691                                                         last_ptr,
6692                                                         num_bytes,
6693                                                         search_start,
6694                                                         &max_extent_size);
6695                                 if (offset) {
6696                                         /* we found one, proceed */
6697                                         spin_unlock(&last_ptr->refill_lock);
6698                                         trace_btrfs_reserve_extent_cluster(root,
6699                                                 block_group, search_start,
6700                                                 num_bytes);
6701                                         goto checks;
6702                                 }
6703                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6704                                    && !failed_cluster_refill) {
6705                                 spin_unlock(&last_ptr->refill_lock);
6706
6707                                 failed_cluster_refill = true;
6708                                 wait_block_group_cache_progress(block_group,
6709                                        num_bytes + empty_cluster + empty_size);
6710                                 goto have_block_group;
6711                         }
6712
6713                         /*
6714                          * at this point we either didn't find a cluster
6715                          * or we weren't able to allocate a block from our
6716                          * cluster.  Free the cluster we've been trying
6717                          * to use, and go to the next block group
6718                          */
6719                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6720                         spin_unlock(&last_ptr->refill_lock);
6721                         goto loop;
6722                 }
6723
6724 unclustered_alloc:
6725                 spin_lock(&block_group->free_space_ctl->tree_lock);
6726                 if (cached &&
6727                     block_group->free_space_ctl->free_space <
6728                     num_bytes + empty_cluster + empty_size) {
6729                         if (block_group->free_space_ctl->free_space >
6730                             max_extent_size)
6731                                 max_extent_size =
6732                                         block_group->free_space_ctl->free_space;
6733                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6734                         goto loop;
6735                 }
6736                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6737
6738                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6739                                                     num_bytes, empty_size,
6740                                                     &max_extent_size);
6741                 /*
6742                  * If we didn't find a chunk, and we haven't failed on this
6743                  * block group before, and this block group is in the middle of
6744                  * caching and we are ok with waiting, then go ahead and wait
6745                  * for progress to be made, and set failed_alloc to true.
6746                  *
6747                  * If failed_alloc is true then we've already waited on this
6748                  * block group once and should move on to the next block group.
6749                  */
6750                 if (!offset && !failed_alloc && !cached &&
6751                     loop > LOOP_CACHING_NOWAIT) {
6752                         wait_block_group_cache_progress(block_group,
6753                                                 num_bytes + empty_size);
6754                         failed_alloc = true;
6755                         goto have_block_group;
6756                 } else if (!offset) {
6757                         if (!cached)
6758                                 have_caching_bg = true;
6759                         goto loop;
6760                 }
6761 checks:
6762                 search_start = stripe_align(root, block_group,
6763                                             offset, num_bytes);
6764
6765                 /* move on to the next group */
6766                 if (search_start + num_bytes >
6767                     block_group->key.objectid + block_group->key.offset) {
6768                         btrfs_add_free_space(block_group, offset, num_bytes);
6769                         goto loop;
6770                 }
6771
6772                 if (offset < search_start)
6773                         btrfs_add_free_space(block_group, offset,
6774                                              search_start - offset);
6775                 BUG_ON(offset > search_start);
6776
6777                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
6778                                                   alloc_type, delalloc);
6779                 if (ret == -EAGAIN) {
6780                         btrfs_add_free_space(block_group, offset, num_bytes);
6781                         goto loop;
6782                 }
6783
6784                 /* we are all good, lets return */
6785                 ins->objectid = search_start;
6786                 ins->offset = num_bytes;
6787
6788                 trace_btrfs_reserve_extent(orig_root, block_group,
6789                                            search_start, num_bytes);
6790                 btrfs_release_block_group(block_group, delalloc);
6791                 break;
6792 loop:
6793                 failed_cluster_refill = false;
6794                 failed_alloc = false;
6795                 BUG_ON(index != get_block_group_index(block_group));
6796                 btrfs_release_block_group(block_group, delalloc);
6797         }
6798         up_read(&space_info->groups_sem);
6799
6800         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6801                 goto search;
6802
6803         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6804                 goto search;
6805
6806         /*
6807          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6808          *                      caching kthreads as we move along
6809          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6810          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6811          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6812          *                      again
6813          */
6814         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6815                 index = 0;
6816                 loop++;
6817                 if (loop == LOOP_ALLOC_CHUNK) {
6818                         struct btrfs_trans_handle *trans;
6819                         int exist = 0;
6820
6821                         trans = current->journal_info;
6822                         if (trans)
6823                                 exist = 1;
6824                         else
6825                                 trans = btrfs_join_transaction(root);
6826
6827                         if (IS_ERR(trans)) {
6828                                 ret = PTR_ERR(trans);
6829                                 goto out;
6830                         }
6831
6832                         ret = do_chunk_alloc(trans, root, flags,
6833                                              CHUNK_ALLOC_FORCE);
6834                         /*
6835                          * Do not bail out on ENOSPC since we
6836                          * can do more things.
6837                          */
6838                         if (ret < 0 && ret != -ENOSPC)
6839                                 btrfs_abort_transaction(trans,
6840                                                         root, ret);
6841                         else
6842                                 ret = 0;
6843                         if (!exist)
6844                                 btrfs_end_transaction(trans, root);
6845                         if (ret)
6846                                 goto out;
6847                 }
6848
6849                 if (loop == LOOP_NO_EMPTY_SIZE) {
6850                         empty_size = 0;
6851                         empty_cluster = 0;
6852                 }
6853
6854                 goto search;
6855         } else if (!ins->objectid) {
6856                 ret = -ENOSPC;
6857         } else if (ins->objectid) {
6858                 ret = 0;
6859         }
6860 out:
6861         if (ret == -ENOSPC)
6862                 ins->offset = max_extent_size;
6863         return ret;
6864 }
6865
6866 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6867                             int dump_block_groups)
6868 {
6869         struct btrfs_block_group_cache *cache;
6870         int index = 0;
6871
6872         spin_lock(&info->lock);
6873         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
6874                info->flags,
6875                info->total_bytes - info->bytes_used - info->bytes_pinned -
6876                info->bytes_reserved - info->bytes_readonly,
6877                (info->full) ? "" : "not ");
6878         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
6879                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6880                info->total_bytes, info->bytes_used, info->bytes_pinned,
6881                info->bytes_reserved, info->bytes_may_use,
6882                info->bytes_readonly);
6883         spin_unlock(&info->lock);
6884
6885         if (!dump_block_groups)
6886                 return;
6887
6888         down_read(&info->groups_sem);
6889 again:
6890         list_for_each_entry(cache, &info->block_groups[index], list) {
6891                 spin_lock(&cache->lock);
6892                 printk(KERN_INFO "BTRFS: "
6893                            "block group %llu has %llu bytes, "
6894                            "%llu used %llu pinned %llu reserved %s\n",
6895                        cache->key.objectid, cache->key.offset,
6896                        btrfs_block_group_used(&cache->item), cache->pinned,
6897                        cache->reserved, cache->ro ? "[readonly]" : "");
6898                 btrfs_dump_free_space(cache, bytes);
6899                 spin_unlock(&cache->lock);
6900         }
6901         if (++index < BTRFS_NR_RAID_TYPES)
6902                 goto again;
6903         up_read(&info->groups_sem);
6904 }
6905
6906 int btrfs_reserve_extent(struct btrfs_root *root,
6907                          u64 num_bytes, u64 min_alloc_size,
6908                          u64 empty_size, u64 hint_byte,
6909                          struct btrfs_key *ins, int is_data, int delalloc)
6910 {
6911         bool final_tried = false;
6912         u64 flags;
6913         int ret;
6914
6915         flags = btrfs_get_alloc_profile(root, is_data);
6916 again:
6917         WARN_ON(num_bytes < root->sectorsize);
6918         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
6919                                flags, delalloc);
6920
6921         if (ret == -ENOSPC) {
6922                 if (!final_tried && ins->offset) {
6923                         num_bytes = min(num_bytes >> 1, ins->offset);
6924                         num_bytes = round_down(num_bytes, root->sectorsize);
6925                         num_bytes = max(num_bytes, min_alloc_size);
6926                         if (num_bytes == min_alloc_size)
6927                                 final_tried = true;
6928                         goto again;
6929                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6930                         struct btrfs_space_info *sinfo;
6931
6932                         sinfo = __find_space_info(root->fs_info, flags);
6933                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
6934                                 flags, num_bytes);
6935                         if (sinfo)
6936                                 dump_space_info(sinfo, num_bytes, 1);
6937                 }
6938         }
6939
6940         return ret;
6941 }
6942
6943 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6944                                         u64 start, u64 len,
6945                                         int pin, int delalloc)
6946 {
6947         struct btrfs_block_group_cache *cache;
6948         int ret = 0;
6949
6950         cache = btrfs_lookup_block_group(root->fs_info, start);
6951         if (!cache) {
6952                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
6953                         start);
6954                 return -ENOSPC;
6955         }
6956
6957         if (btrfs_test_opt(root, DISCARD))
6958                 ret = btrfs_discard_extent(root, start, len, NULL);
6959
6960         if (pin)
6961                 pin_down_extent(root, cache, start, len, 1);
6962         else {
6963                 btrfs_add_free_space(cache, start, len);
6964                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
6965         }
6966         btrfs_put_block_group(cache);
6967
6968         trace_btrfs_reserved_extent_free(root, start, len);
6969
6970         return ret;
6971 }
6972
6973 int btrfs_free_reserved_extent(struct btrfs_root *root,
6974                                u64 start, u64 len, int delalloc)
6975 {
6976         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
6977 }
6978
6979 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6980                                        u64 start, u64 len)
6981 {
6982         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
6983 }
6984
6985 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6986                                       struct btrfs_root *root,
6987                                       u64 parent, u64 root_objectid,
6988                                       u64 flags, u64 owner, u64 offset,
6989                                       struct btrfs_key *ins, int ref_mod)
6990 {
6991         int ret;
6992         struct btrfs_fs_info *fs_info = root->fs_info;
6993         struct btrfs_extent_item *extent_item;
6994         struct btrfs_extent_inline_ref *iref;
6995         struct btrfs_path *path;
6996         struct extent_buffer *leaf;
6997         int type;
6998         u32 size;
6999
7000         if (parent > 0)
7001                 type = BTRFS_SHARED_DATA_REF_KEY;
7002         else
7003                 type = BTRFS_EXTENT_DATA_REF_KEY;
7004
7005         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7006
7007         path = btrfs_alloc_path();
7008         if (!path)
7009                 return -ENOMEM;
7010
7011         path->leave_spinning = 1;
7012         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7013                                       ins, size);
7014         if (ret) {
7015                 btrfs_free_path(path);
7016                 return ret;
7017         }
7018
7019         leaf = path->nodes[0];
7020         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7021                                      struct btrfs_extent_item);
7022         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7023         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7024         btrfs_set_extent_flags(leaf, extent_item,
7025                                flags | BTRFS_EXTENT_FLAG_DATA);
7026
7027         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7028         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7029         if (parent > 0) {
7030                 struct btrfs_shared_data_ref *ref;
7031                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7032                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7033                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7034         } else {
7035                 struct btrfs_extent_data_ref *ref;
7036                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7037                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7038                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7039                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7040                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7041         }
7042
7043         btrfs_mark_buffer_dirty(path->nodes[0]);
7044         btrfs_free_path(path);
7045
7046         /* Always set parent to 0 here since its exclusive anyway. */
7047         ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
7048                                       ins->objectid, ins->offset,
7049                                       BTRFS_QGROUP_OPER_ADD_EXCL, 0);
7050         if (ret)
7051                 return ret;
7052
7053         ret = update_block_group(root, ins->objectid, ins->offset, 1);
7054         if (ret) { /* -ENOENT, logic error */
7055                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7056                         ins->objectid, ins->offset);
7057                 BUG();
7058         }
7059         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
7060         return ret;
7061 }
7062
7063 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7064                                      struct btrfs_root *root,
7065                                      u64 parent, u64 root_objectid,
7066                                      u64 flags, struct btrfs_disk_key *key,
7067                                      int level, struct btrfs_key *ins,
7068                                      int no_quota)
7069 {
7070         int ret;
7071         struct btrfs_fs_info *fs_info = root->fs_info;
7072         struct btrfs_extent_item *extent_item;
7073         struct btrfs_tree_block_info *block_info;
7074         struct btrfs_extent_inline_ref *iref;
7075         struct btrfs_path *path;
7076         struct extent_buffer *leaf;
7077         u32 size = sizeof(*extent_item) + sizeof(*iref);
7078         u64 num_bytes = ins->offset;
7079         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7080                                                  SKINNY_METADATA);
7081
7082         if (!skinny_metadata)
7083                 size += sizeof(*block_info);
7084
7085         path = btrfs_alloc_path();
7086         if (!path) {
7087                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7088                                                    root->leafsize);
7089                 return -ENOMEM;
7090         }
7091
7092         path->leave_spinning = 1;
7093         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7094                                       ins, size);
7095         if (ret) {
7096                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7097                                                    root->leafsize);
7098                 btrfs_free_path(path);
7099                 return ret;
7100         }
7101
7102         leaf = path->nodes[0];
7103         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7104                                      struct btrfs_extent_item);
7105         btrfs_set_extent_refs(leaf, extent_item, 1);
7106         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7107         btrfs_set_extent_flags(leaf, extent_item,
7108                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7109
7110         if (skinny_metadata) {
7111                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7112                 num_bytes = root->leafsize;
7113         } else {
7114                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7115                 btrfs_set_tree_block_key(leaf, block_info, key);
7116                 btrfs_set_tree_block_level(leaf, block_info, level);
7117                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7118         }
7119
7120         if (parent > 0) {
7121                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7122                 btrfs_set_extent_inline_ref_type(leaf, iref,
7123                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7124                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7125         } else {
7126                 btrfs_set_extent_inline_ref_type(leaf, iref,
7127                                                  BTRFS_TREE_BLOCK_REF_KEY);
7128                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7129         }
7130
7131         btrfs_mark_buffer_dirty(leaf);
7132         btrfs_free_path(path);
7133
7134         if (!no_quota) {
7135                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
7136                                               ins->objectid, num_bytes,
7137                                               BTRFS_QGROUP_OPER_ADD_EXCL, 0);
7138                 if (ret)
7139                         return ret;
7140         }
7141
7142         ret = update_block_group(root, ins->objectid, root->leafsize, 1);
7143         if (ret) { /* -ENOENT, logic error */
7144                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7145                         ins->objectid, ins->offset);
7146                 BUG();
7147         }
7148
7149         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->leafsize);
7150         return ret;
7151 }
7152
7153 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7154                                      struct btrfs_root *root,
7155                                      u64 root_objectid, u64 owner,
7156                                      u64 offset, struct btrfs_key *ins)
7157 {
7158         int ret;
7159
7160         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7161
7162         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7163                                          ins->offset, 0,
7164                                          root_objectid, owner, offset,
7165                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
7166         return ret;
7167 }
7168
7169 /*
7170  * this is used by the tree logging recovery code.  It records that
7171  * an extent has been allocated and makes sure to clear the free
7172  * space cache bits as well
7173  */
7174 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7175                                    struct btrfs_root *root,
7176                                    u64 root_objectid, u64 owner, u64 offset,
7177                                    struct btrfs_key *ins)
7178 {
7179         int ret;
7180         struct btrfs_block_group_cache *block_group;
7181
7182         /*
7183          * Mixed block groups will exclude before processing the log so we only
7184          * need to do the exlude dance if this fs isn't mixed.
7185          */
7186         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7187                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7188                 if (ret)
7189                         return ret;
7190         }
7191
7192         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
7193         if (!block_group)
7194                 return -EINVAL;
7195
7196         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
7197                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
7198         BUG_ON(ret); /* logic error */
7199         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
7200                                          0, owner, offset, ins, 1);
7201         btrfs_put_block_group(block_group);
7202         return ret;
7203 }
7204
7205 static struct extent_buffer *
7206 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7207                       u64 bytenr, u32 blocksize, int level)
7208 {
7209         struct extent_buffer *buf;
7210
7211         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
7212         if (!buf)
7213                 return ERR_PTR(-ENOMEM);
7214         btrfs_set_header_generation(buf, trans->transid);
7215         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
7216         btrfs_tree_lock(buf);
7217         clean_tree_block(trans, root, buf);
7218         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
7219
7220         btrfs_set_lock_blocking(buf);
7221         btrfs_set_buffer_uptodate(buf);
7222
7223         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
7224                 /*
7225                  * we allow two log transactions at a time, use different
7226                  * EXENT bit to differentiate dirty pages.
7227                  */
7228                 if (root->log_transid % 2 == 0)
7229                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7230                                         buf->start + buf->len - 1, GFP_NOFS);
7231                 else
7232                         set_extent_new(&root->dirty_log_pages, buf->start,
7233                                         buf->start + buf->len - 1, GFP_NOFS);
7234         } else {
7235                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7236                          buf->start + buf->len - 1, GFP_NOFS);
7237         }
7238         trans->blocks_used++;
7239         /* this returns a buffer locked for blocking */
7240         return buf;
7241 }
7242
7243 static struct btrfs_block_rsv *
7244 use_block_rsv(struct btrfs_trans_handle *trans,
7245               struct btrfs_root *root, u32 blocksize)
7246 {
7247         struct btrfs_block_rsv *block_rsv;
7248         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7249         int ret;
7250         bool global_updated = false;
7251
7252         block_rsv = get_block_rsv(trans, root);
7253
7254         if (unlikely(block_rsv->size == 0))
7255                 goto try_reserve;
7256 again:
7257         ret = block_rsv_use_bytes(block_rsv, blocksize);
7258         if (!ret)
7259                 return block_rsv;
7260
7261         if (block_rsv->failfast)
7262                 return ERR_PTR(ret);
7263
7264         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7265                 global_updated = true;
7266                 update_global_block_rsv(root->fs_info);
7267                 goto again;
7268         }
7269
7270         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7271                 static DEFINE_RATELIMIT_STATE(_rs,
7272                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7273                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7274                 if (__ratelimit(&_rs))
7275                         WARN(1, KERN_DEBUG
7276                                 "BTRFS: block rsv returned %d\n", ret);
7277         }
7278 try_reserve:
7279         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7280                                      BTRFS_RESERVE_NO_FLUSH);
7281         if (!ret)
7282                 return block_rsv;
7283         /*
7284          * If we couldn't reserve metadata bytes try and use some from
7285          * the global reserve if its space type is the same as the global
7286          * reservation.
7287          */
7288         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7289             block_rsv->space_info == global_rsv->space_info) {
7290                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7291                 if (!ret)
7292                         return global_rsv;
7293         }
7294         return ERR_PTR(ret);
7295 }
7296
7297 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7298                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7299 {
7300         block_rsv_add_bytes(block_rsv, blocksize, 0);
7301         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7302 }
7303
7304 /*
7305  * finds a free extent and does all the dirty work required for allocation
7306  * returns the key for the extent through ins, and a tree buffer for
7307  * the first block of the extent through buf.
7308  *
7309  * returns the tree buffer or NULL.
7310  */
7311 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
7312                                         struct btrfs_root *root, u32 blocksize,
7313                                         u64 parent, u64 root_objectid,
7314                                         struct btrfs_disk_key *key, int level,
7315                                         u64 hint, u64 empty_size)
7316 {
7317         struct btrfs_key ins;
7318         struct btrfs_block_rsv *block_rsv;
7319         struct extent_buffer *buf;
7320         u64 flags = 0;
7321         int ret;
7322         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7323                                                  SKINNY_METADATA);
7324
7325 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
7326         if (unlikely(test_bit(BTRFS_ROOT_DUMMY_ROOT, &root->state))) {
7327                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
7328                                             blocksize, level);
7329                 if (!IS_ERR(buf))
7330                         root->alloc_bytenr += blocksize;
7331                 return buf;
7332         }
7333 #endif
7334         block_rsv = use_block_rsv(trans, root, blocksize);
7335         if (IS_ERR(block_rsv))
7336                 return ERR_CAST(block_rsv);
7337
7338         ret = btrfs_reserve_extent(root, blocksize, blocksize,
7339                                    empty_size, hint, &ins, 0, 0);
7340         if (ret) {
7341                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
7342                 return ERR_PTR(ret);
7343         }
7344
7345         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
7346                                     blocksize, level);
7347         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
7348
7349         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
7350                 if (parent == 0)
7351                         parent = ins.objectid;
7352                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
7353         } else
7354                 BUG_ON(parent > 0);
7355
7356         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
7357                 struct btrfs_delayed_extent_op *extent_op;
7358                 extent_op = btrfs_alloc_delayed_extent_op();
7359                 BUG_ON(!extent_op); /* -ENOMEM */
7360                 if (key)
7361                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
7362                 else
7363                         memset(&extent_op->key, 0, sizeof(extent_op->key));
7364                 extent_op->flags_to_set = flags;
7365                 if (skinny_metadata)
7366                         extent_op->update_key = 0;
7367                 else
7368                         extent_op->update_key = 1;
7369                 extent_op->update_flags = 1;
7370                 extent_op->is_data = 0;
7371                 extent_op->level = level;
7372
7373                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7374                                         ins.objectid,
7375                                         ins.offset, parent, root_objectid,
7376                                         level, BTRFS_ADD_DELAYED_EXTENT,
7377                                         extent_op, 0);
7378                 BUG_ON(ret); /* -ENOMEM */
7379         }
7380         return buf;
7381 }
7382
7383 struct walk_control {
7384         u64 refs[BTRFS_MAX_LEVEL];
7385         u64 flags[BTRFS_MAX_LEVEL];
7386         struct btrfs_key update_progress;
7387         int stage;
7388         int level;
7389         int shared_level;
7390         int update_ref;
7391         int keep_locks;
7392         int reada_slot;
7393         int reada_count;
7394         int for_reloc;
7395 };
7396
7397 #define DROP_REFERENCE  1
7398 #define UPDATE_BACKREF  2
7399
7400 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
7401                                      struct btrfs_root *root,
7402                                      struct walk_control *wc,
7403                                      struct btrfs_path *path)
7404 {
7405         u64 bytenr;
7406         u64 generation;
7407         u64 refs;
7408         u64 flags;
7409         u32 nritems;
7410         u32 blocksize;
7411         struct btrfs_key key;
7412         struct extent_buffer *eb;
7413         int ret;
7414         int slot;
7415         int nread = 0;
7416
7417         if (path->slots[wc->level] < wc->reada_slot) {
7418                 wc->reada_count = wc->reada_count * 2 / 3;
7419                 wc->reada_count = max(wc->reada_count, 2);
7420         } else {
7421                 wc->reada_count = wc->reada_count * 3 / 2;
7422                 wc->reada_count = min_t(int, wc->reada_count,
7423                                         BTRFS_NODEPTRS_PER_BLOCK(root));
7424         }
7425
7426         eb = path->nodes[wc->level];
7427         nritems = btrfs_header_nritems(eb);
7428         blocksize = btrfs_level_size(root, wc->level - 1);
7429
7430         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
7431                 if (nread >= wc->reada_count)
7432                         break;
7433
7434                 cond_resched();
7435                 bytenr = btrfs_node_blockptr(eb, slot);
7436                 generation = btrfs_node_ptr_generation(eb, slot);
7437
7438                 if (slot == path->slots[wc->level])
7439                         goto reada;
7440
7441                 if (wc->stage == UPDATE_BACKREF &&
7442                     generation <= root->root_key.offset)
7443                         continue;
7444
7445                 /* We don't lock the tree block, it's OK to be racy here */
7446                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7447                                                wc->level - 1, 1, &refs,
7448                                                &flags);
7449                 /* We don't care about errors in readahead. */
7450                 if (ret < 0)
7451                         continue;
7452                 BUG_ON(refs == 0);
7453
7454                 if (wc->stage == DROP_REFERENCE) {
7455                         if (refs == 1)
7456                                 goto reada;
7457
7458                         if (wc->level == 1 &&
7459                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7460                                 continue;
7461                         if (!wc->update_ref ||
7462                             generation <= root->root_key.offset)
7463                                 continue;
7464                         btrfs_node_key_to_cpu(eb, &key, slot);
7465                         ret = btrfs_comp_cpu_keys(&key,
7466                                                   &wc->update_progress);
7467                         if (ret < 0)
7468                                 continue;
7469                 } else {
7470                         if (wc->level == 1 &&
7471                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7472                                 continue;
7473                 }
7474 reada:
7475                 ret = readahead_tree_block(root, bytenr, blocksize,
7476                                            generation);
7477                 if (ret)
7478                         break;
7479                 nread++;
7480         }
7481         wc->reada_slot = slot;
7482 }
7483
7484 /*
7485  * helper to process tree block while walking down the tree.
7486  *
7487  * when wc->stage == UPDATE_BACKREF, this function updates
7488  * back refs for pointers in the block.
7489  *
7490  * NOTE: return value 1 means we should stop walking down.
7491  */
7492 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
7493                                    struct btrfs_root *root,
7494                                    struct btrfs_path *path,
7495                                    struct walk_control *wc, int lookup_info)
7496 {
7497         int level = wc->level;
7498         struct extent_buffer *eb = path->nodes[level];
7499         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7500         int ret;
7501
7502         if (wc->stage == UPDATE_BACKREF &&
7503             btrfs_header_owner(eb) != root->root_key.objectid)
7504                 return 1;
7505
7506         /*
7507          * when reference count of tree block is 1, it won't increase
7508          * again. once full backref flag is set, we never clear it.
7509          */
7510         if (lookup_info &&
7511             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
7512              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
7513                 BUG_ON(!path->locks[level]);
7514                 ret = btrfs_lookup_extent_info(trans, root,
7515                                                eb->start, level, 1,
7516                                                &wc->refs[level],
7517                                                &wc->flags[level]);
7518                 BUG_ON(ret == -ENOMEM);
7519                 if (ret)
7520                         return ret;
7521                 BUG_ON(wc->refs[level] == 0);
7522         }
7523
7524         if (wc->stage == DROP_REFERENCE) {
7525                 if (wc->refs[level] > 1)
7526                         return 1;
7527
7528                 if (path->locks[level] && !wc->keep_locks) {
7529                         btrfs_tree_unlock_rw(eb, path->locks[level]);
7530                         path->locks[level] = 0;
7531                 }
7532                 return 0;
7533         }
7534
7535         /* wc->stage == UPDATE_BACKREF */
7536         if (!(wc->flags[level] & flag)) {
7537                 BUG_ON(!path->locks[level]);
7538                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
7539                 BUG_ON(ret); /* -ENOMEM */
7540                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
7541                 BUG_ON(ret); /* -ENOMEM */
7542                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
7543                                                   eb->len, flag,
7544                                                   btrfs_header_level(eb), 0);
7545                 BUG_ON(ret); /* -ENOMEM */
7546                 wc->flags[level] |= flag;
7547         }
7548
7549         /*
7550          * the block is shared by multiple trees, so it's not good to
7551          * keep the tree lock
7552          */
7553         if (path->locks[level] && level > 0) {
7554                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7555                 path->locks[level] = 0;
7556         }
7557         return 0;
7558 }
7559
7560 /*
7561  * helper to process tree block pointer.
7562  *
7563  * when wc->stage == DROP_REFERENCE, this function checks
7564  * reference count of the block pointed to. if the block
7565  * is shared and we need update back refs for the subtree
7566  * rooted at the block, this function changes wc->stage to
7567  * UPDATE_BACKREF. if the block is shared and there is no
7568  * need to update back, this function drops the reference
7569  * to the block.
7570  *
7571  * NOTE: return value 1 means we should stop walking down.
7572  */
7573 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
7574                                  struct btrfs_root *root,
7575                                  struct btrfs_path *path,
7576                                  struct walk_control *wc, int *lookup_info)
7577 {
7578         u64 bytenr;
7579         u64 generation;
7580         u64 parent;
7581         u32 blocksize;
7582         struct btrfs_key key;
7583         struct extent_buffer *next;
7584         int level = wc->level;
7585         int reada = 0;
7586         int ret = 0;
7587
7588         generation = btrfs_node_ptr_generation(path->nodes[level],
7589                                                path->slots[level]);
7590         /*
7591          * if the lower level block was created before the snapshot
7592          * was created, we know there is no need to update back refs
7593          * for the subtree
7594          */
7595         if (wc->stage == UPDATE_BACKREF &&
7596             generation <= root->root_key.offset) {
7597                 *lookup_info = 1;
7598                 return 1;
7599         }
7600
7601         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
7602         blocksize = btrfs_level_size(root, level - 1);
7603
7604         next = btrfs_find_tree_block(root, bytenr, blocksize);
7605         if (!next) {
7606                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
7607                 if (!next)
7608                         return -ENOMEM;
7609                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
7610                                                level - 1);
7611                 reada = 1;
7612         }
7613         btrfs_tree_lock(next);
7614         btrfs_set_lock_blocking(next);
7615
7616         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
7617                                        &wc->refs[level - 1],
7618                                        &wc->flags[level - 1]);
7619         if (ret < 0) {
7620                 btrfs_tree_unlock(next);
7621                 return ret;
7622         }
7623
7624         if (unlikely(wc->refs[level - 1] == 0)) {
7625                 btrfs_err(root->fs_info, "Missing references.");
7626                 BUG();
7627         }
7628         *lookup_info = 0;
7629
7630         if (wc->stage == DROP_REFERENCE) {
7631                 if (wc->refs[level - 1] > 1) {
7632                         if (level == 1 &&
7633                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7634                                 goto skip;
7635
7636                         if (!wc->update_ref ||
7637                             generation <= root->root_key.offset)
7638                                 goto skip;
7639
7640                         btrfs_node_key_to_cpu(path->nodes[level], &key,
7641                                               path->slots[level]);
7642                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
7643                         if (ret < 0)
7644                                 goto skip;
7645
7646                         wc->stage = UPDATE_BACKREF;
7647                         wc->shared_level = level - 1;
7648                 }
7649         } else {
7650                 if (level == 1 &&
7651                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7652                         goto skip;
7653         }
7654
7655         if (!btrfs_buffer_uptodate(next, generation, 0)) {
7656                 btrfs_tree_unlock(next);
7657                 free_extent_buffer(next);
7658                 next = NULL;
7659                 *lookup_info = 1;
7660         }
7661
7662         if (!next) {
7663                 if (reada && level == 1)
7664                         reada_walk_down(trans, root, wc, path);
7665                 next = read_tree_block(root, bytenr, blocksize, generation);
7666                 if (!next || !extent_buffer_uptodate(next)) {
7667                         free_extent_buffer(next);
7668                         return -EIO;
7669                 }
7670                 btrfs_tree_lock(next);
7671                 btrfs_set_lock_blocking(next);
7672         }
7673
7674         level--;
7675         BUG_ON(level != btrfs_header_level(next));
7676         path->nodes[level] = next;
7677         path->slots[level] = 0;
7678         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7679         wc->level = level;
7680         if (wc->level == 1)
7681                 wc->reada_slot = 0;
7682         return 0;
7683 skip:
7684         wc->refs[level - 1] = 0;
7685         wc->flags[level - 1] = 0;
7686         if (wc->stage == DROP_REFERENCE) {
7687                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
7688                         parent = path->nodes[level]->start;
7689                 } else {
7690                         BUG_ON(root->root_key.objectid !=
7691                                btrfs_header_owner(path->nodes[level]));
7692                         parent = 0;
7693                 }
7694
7695                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
7696                                 root->root_key.objectid, level - 1, 0, 0);
7697                 BUG_ON(ret); /* -ENOMEM */
7698         }
7699         btrfs_tree_unlock(next);
7700         free_extent_buffer(next);
7701         *lookup_info = 1;
7702         return 1;
7703 }
7704
7705 /*
7706  * helper to process tree block while walking up the tree.
7707  *
7708  * when wc->stage == DROP_REFERENCE, this function drops
7709  * reference count on the block.
7710  *
7711  * when wc->stage == UPDATE_BACKREF, this function changes
7712  * wc->stage back to DROP_REFERENCE if we changed wc->stage
7713  * to UPDATE_BACKREF previously while processing the block.
7714  *
7715  * NOTE: return value 1 means we should stop walking up.
7716  */
7717 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
7718                                  struct btrfs_root *root,
7719                                  struct btrfs_path *path,
7720                                  struct walk_control *wc)
7721 {
7722         int ret;
7723         int level = wc->level;
7724         struct extent_buffer *eb = path->nodes[level];
7725         u64 parent = 0;
7726
7727         if (wc->stage == UPDATE_BACKREF) {
7728                 BUG_ON(wc->shared_level < level);
7729                 if (level < wc->shared_level)
7730                         goto out;
7731
7732                 ret = find_next_key(path, level + 1, &wc->update_progress);
7733                 if (ret > 0)
7734                         wc->update_ref = 0;
7735
7736                 wc->stage = DROP_REFERENCE;
7737                 wc->shared_level = -1;
7738                 path->slots[level] = 0;
7739
7740                 /*
7741                  * check reference count again if the block isn't locked.
7742                  * we should start walking down the tree again if reference
7743                  * count is one.
7744                  */
7745                 if (!path->locks[level]) {
7746                         BUG_ON(level == 0);
7747                         btrfs_tree_lock(eb);
7748                         btrfs_set_lock_blocking(eb);
7749                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7750
7751                         ret = btrfs_lookup_extent_info(trans, root,
7752                                                        eb->start, level, 1,
7753                                                        &wc->refs[level],
7754                                                        &wc->flags[level]);
7755                         if (ret < 0) {
7756                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7757                                 path->locks[level] = 0;
7758                                 return ret;
7759                         }
7760                         BUG_ON(wc->refs[level] == 0);
7761                         if (wc->refs[level] == 1) {
7762                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7763                                 path->locks[level] = 0;
7764                                 return 1;
7765                         }
7766                 }
7767         }
7768
7769         /* wc->stage == DROP_REFERENCE */
7770         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
7771
7772         if (wc->refs[level] == 1) {
7773                 if (level == 0) {
7774                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7775                                 ret = btrfs_dec_ref(trans, root, eb, 1,
7776                                                     wc->for_reloc);
7777                         else
7778                                 ret = btrfs_dec_ref(trans, root, eb, 0,
7779                                                     wc->for_reloc);
7780                         BUG_ON(ret); /* -ENOMEM */
7781                 }
7782                 /* make block locked assertion in clean_tree_block happy */
7783                 if (!path->locks[level] &&
7784                     btrfs_header_generation(eb) == trans->transid) {
7785                         btrfs_tree_lock(eb);
7786                         btrfs_set_lock_blocking(eb);
7787                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7788                 }
7789                 clean_tree_block(trans, root, eb);
7790         }
7791
7792         if (eb == root->node) {
7793                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7794                         parent = eb->start;
7795                 else
7796                         BUG_ON(root->root_key.objectid !=
7797                                btrfs_header_owner(eb));
7798         } else {
7799                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7800                         parent = path->nodes[level + 1]->start;
7801                 else
7802                         BUG_ON(root->root_key.objectid !=
7803                                btrfs_header_owner(path->nodes[level + 1]));
7804         }
7805
7806         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
7807 out:
7808         wc->refs[level] = 0;
7809         wc->flags[level] = 0;
7810         return 0;
7811 }
7812
7813 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
7814                                    struct btrfs_root *root,
7815                                    struct btrfs_path *path,
7816                                    struct walk_control *wc)
7817 {
7818         int level = wc->level;
7819         int lookup_info = 1;
7820         int ret;
7821
7822         while (level >= 0) {
7823                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
7824                 if (ret > 0)
7825                         break;
7826
7827                 if (level == 0)
7828                         break;
7829
7830                 if (path->slots[level] >=
7831                     btrfs_header_nritems(path->nodes[level]))
7832                         break;
7833
7834                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
7835                 if (ret > 0) {
7836                         path->slots[level]++;
7837                         continue;
7838                 } else if (ret < 0)
7839                         return ret;
7840                 level = wc->level;
7841         }
7842         return 0;
7843 }
7844
7845 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
7846                                  struct btrfs_root *root,
7847                                  struct btrfs_path *path,
7848                                  struct walk_control *wc, int max_level)
7849 {
7850         int level = wc->level;
7851         int ret;
7852
7853         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
7854         while (level < max_level && path->nodes[level]) {
7855                 wc->level = level;
7856                 if (path->slots[level] + 1 <
7857                     btrfs_header_nritems(path->nodes[level])) {
7858                         path->slots[level]++;
7859                         return 0;
7860                 } else {
7861                         ret = walk_up_proc(trans, root, path, wc);
7862                         if (ret > 0)
7863                                 return 0;
7864
7865                         if (path->locks[level]) {
7866                                 btrfs_tree_unlock_rw(path->nodes[level],
7867                                                      path->locks[level]);
7868                                 path->locks[level] = 0;
7869                         }
7870                         free_extent_buffer(path->nodes[level]);
7871                         path->nodes[level] = NULL;
7872                         level++;
7873                 }
7874         }
7875         return 1;
7876 }
7877
7878 /*
7879  * drop a subvolume tree.
7880  *
7881  * this function traverses the tree freeing any blocks that only
7882  * referenced by the tree.
7883  *
7884  * when a shared tree block is found. this function decreases its
7885  * reference count by one. if update_ref is true, this function
7886  * also make sure backrefs for the shared block and all lower level
7887  * blocks are properly updated.
7888  *
7889  * If called with for_reloc == 0, may exit early with -EAGAIN
7890  */
7891 int btrfs_drop_snapshot(struct btrfs_root *root,
7892                          struct btrfs_block_rsv *block_rsv, int update_ref,
7893                          int for_reloc)
7894 {
7895         struct btrfs_path *path;
7896         struct btrfs_trans_handle *trans;
7897         struct btrfs_root *tree_root = root->fs_info->tree_root;
7898         struct btrfs_root_item *root_item = &root->root_item;
7899         struct walk_control *wc;
7900         struct btrfs_key key;
7901         int err = 0;
7902         int ret;
7903         int level;
7904         bool root_dropped = false;
7905
7906         path = btrfs_alloc_path();
7907         if (!path) {
7908                 err = -ENOMEM;
7909                 goto out;
7910         }
7911
7912         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7913         if (!wc) {
7914                 btrfs_free_path(path);
7915                 err = -ENOMEM;
7916                 goto out;
7917         }
7918
7919         trans = btrfs_start_transaction(tree_root, 0);
7920         if (IS_ERR(trans)) {
7921                 err = PTR_ERR(trans);
7922                 goto out_free;
7923         }
7924
7925         if (block_rsv)
7926                 trans->block_rsv = block_rsv;
7927
7928         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
7929                 level = btrfs_header_level(root->node);
7930                 path->nodes[level] = btrfs_lock_root_node(root);
7931                 btrfs_set_lock_blocking(path->nodes[level]);
7932                 path->slots[level] = 0;
7933                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7934                 memset(&wc->update_progress, 0,
7935                        sizeof(wc->update_progress));
7936         } else {
7937                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
7938                 memcpy(&wc->update_progress, &key,
7939                        sizeof(wc->update_progress));
7940
7941                 level = root_item->drop_level;
7942                 BUG_ON(level == 0);
7943                 path->lowest_level = level;
7944                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
7945                 path->lowest_level = 0;
7946                 if (ret < 0) {
7947                         err = ret;
7948                         goto out_end_trans;
7949                 }
7950                 WARN_ON(ret > 0);
7951
7952                 /*
7953                  * unlock our path, this is safe because only this
7954                  * function is allowed to delete this snapshot
7955                  */
7956                 btrfs_unlock_up_safe(path, 0);
7957
7958                 level = btrfs_header_level(root->node);
7959                 while (1) {
7960                         btrfs_tree_lock(path->nodes[level]);
7961                         btrfs_set_lock_blocking(path->nodes[level]);
7962                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7963
7964                         ret = btrfs_lookup_extent_info(trans, root,
7965                                                 path->nodes[level]->start,
7966                                                 level, 1, &wc->refs[level],
7967                                                 &wc->flags[level]);
7968                         if (ret < 0) {
7969                                 err = ret;
7970                                 goto out_end_trans;
7971                         }
7972                         BUG_ON(wc->refs[level] == 0);
7973
7974                         if (level == root_item->drop_level)
7975                                 break;
7976
7977                         btrfs_tree_unlock(path->nodes[level]);
7978                         path->locks[level] = 0;
7979                         WARN_ON(wc->refs[level] != 1);
7980                         level--;
7981                 }
7982         }
7983
7984         wc->level = level;
7985         wc->shared_level = -1;
7986         wc->stage = DROP_REFERENCE;
7987         wc->update_ref = update_ref;
7988         wc->keep_locks = 0;
7989         wc->for_reloc = for_reloc;
7990         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7991
7992         while (1) {
7993
7994                 ret = walk_down_tree(trans, root, path, wc);
7995                 if (ret < 0) {
7996                         err = ret;
7997                         break;
7998                 }
7999
8000                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8001                 if (ret < 0) {
8002                         err = ret;
8003                         break;
8004                 }
8005
8006                 if (ret > 0) {
8007                         BUG_ON(wc->stage != DROP_REFERENCE);
8008                         break;
8009                 }
8010
8011                 if (wc->stage == DROP_REFERENCE) {
8012                         level = wc->level;
8013                         btrfs_node_key(path->nodes[level],
8014                                        &root_item->drop_progress,
8015                                        path->slots[level]);
8016                         root_item->drop_level = level;
8017                 }
8018
8019                 BUG_ON(wc->level == 0);
8020                 if (btrfs_should_end_transaction(trans, tree_root) ||
8021                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
8022                         ret = btrfs_update_root(trans, tree_root,
8023                                                 &root->root_key,
8024                                                 root_item);
8025                         if (ret) {
8026                                 btrfs_abort_transaction(trans, tree_root, ret);
8027                                 err = ret;
8028                                 goto out_end_trans;
8029                         }
8030
8031                         btrfs_end_transaction_throttle(trans, tree_root);
8032                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
8033                                 pr_debug("BTRFS: drop snapshot early exit\n");
8034                                 err = -EAGAIN;
8035                                 goto out_free;
8036                         }
8037
8038                         trans = btrfs_start_transaction(tree_root, 0);
8039                         if (IS_ERR(trans)) {
8040                                 err = PTR_ERR(trans);
8041                                 goto out_free;
8042                         }
8043                         if (block_rsv)
8044                                 trans->block_rsv = block_rsv;
8045                 }
8046         }
8047         btrfs_release_path(path);
8048         if (err)
8049                 goto out_end_trans;
8050
8051         ret = btrfs_del_root(trans, tree_root, &root->root_key);
8052         if (ret) {
8053                 btrfs_abort_transaction(trans, tree_root, ret);
8054                 goto out_end_trans;
8055         }
8056
8057         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
8058                 ret = btrfs_find_root(tree_root, &root->root_key, path,
8059                                       NULL, NULL);
8060                 if (ret < 0) {
8061                         btrfs_abort_transaction(trans, tree_root, ret);
8062                         err = ret;
8063                         goto out_end_trans;
8064                 } else if (ret > 0) {
8065                         /* if we fail to delete the orphan item this time
8066                          * around, it'll get picked up the next time.
8067                          *
8068                          * The most common failure here is just -ENOENT.
8069                          */
8070                         btrfs_del_orphan_item(trans, tree_root,
8071                                               root->root_key.objectid);
8072                 }
8073         }
8074
8075         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
8076                 btrfs_drop_and_free_fs_root(tree_root->fs_info, root);
8077         } else {
8078                 free_extent_buffer(root->node);
8079                 free_extent_buffer(root->commit_root);
8080                 btrfs_put_fs_root(root);
8081         }
8082         root_dropped = true;
8083 out_end_trans:
8084         btrfs_end_transaction_throttle(trans, tree_root);
8085 out_free:
8086         kfree(wc);
8087         btrfs_free_path(path);
8088 out:
8089         /*
8090          * So if we need to stop dropping the snapshot for whatever reason we
8091          * need to make sure to add it back to the dead root list so that we
8092          * keep trying to do the work later.  This also cleans up roots if we
8093          * don't have it in the radix (like when we recover after a power fail
8094          * or unmount) so we don't leak memory.
8095          */
8096         if (!for_reloc && root_dropped == false)
8097                 btrfs_add_dead_root(root);
8098         if (err && err != -EAGAIN)
8099                 btrfs_std_error(root->fs_info, err);
8100         return err;
8101 }
8102
8103 /*
8104  * drop subtree rooted at tree block 'node'.
8105  *
8106  * NOTE: this function will unlock and release tree block 'node'
8107  * only used by relocation code
8108  */
8109 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
8110                         struct btrfs_root *root,
8111                         struct extent_buffer *node,
8112                         struct extent_buffer *parent)
8113 {
8114         struct btrfs_path *path;
8115         struct walk_control *wc;
8116         int level;
8117         int parent_level;
8118         int ret = 0;
8119         int wret;
8120
8121         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
8122
8123         path = btrfs_alloc_path();
8124         if (!path)
8125                 return -ENOMEM;
8126
8127         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8128         if (!wc) {
8129                 btrfs_free_path(path);
8130                 return -ENOMEM;
8131         }
8132
8133         btrfs_assert_tree_locked(parent);
8134         parent_level = btrfs_header_level(parent);
8135         extent_buffer_get(parent);
8136         path->nodes[parent_level] = parent;
8137         path->slots[parent_level] = btrfs_header_nritems(parent);
8138
8139         btrfs_assert_tree_locked(node);
8140         level = btrfs_header_level(node);
8141         path->nodes[level] = node;
8142         path->slots[level] = 0;
8143         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8144
8145         wc->refs[parent_level] = 1;
8146         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8147         wc->level = level;
8148         wc->shared_level = -1;
8149         wc->stage = DROP_REFERENCE;
8150         wc->update_ref = 0;
8151         wc->keep_locks = 1;
8152         wc->for_reloc = 1;
8153         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8154
8155         while (1) {
8156                 wret = walk_down_tree(trans, root, path, wc);
8157                 if (wret < 0) {
8158                         ret = wret;
8159                         break;
8160                 }
8161
8162                 wret = walk_up_tree(trans, root, path, wc, parent_level);
8163                 if (wret < 0)
8164                         ret = wret;
8165                 if (wret != 0)
8166                         break;
8167         }
8168
8169         kfree(wc);
8170         btrfs_free_path(path);
8171         return ret;
8172 }
8173
8174 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
8175 {
8176         u64 num_devices;
8177         u64 stripped;
8178
8179         /*
8180          * if restripe for this chunk_type is on pick target profile and
8181          * return, otherwise do the usual balance
8182          */
8183         stripped = get_restripe_target(root->fs_info, flags);
8184         if (stripped)
8185                 return extended_to_chunk(stripped);
8186
8187         /*
8188          * we add in the count of missing devices because we want
8189          * to make sure that any RAID levels on a degraded FS
8190          * continue to be honored.
8191          */
8192         num_devices = root->fs_info->fs_devices->rw_devices +
8193                 root->fs_info->fs_devices->missing_devices;
8194
8195         stripped = BTRFS_BLOCK_GROUP_RAID0 |
8196                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
8197                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
8198
8199         if (num_devices == 1) {
8200                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8201                 stripped = flags & ~stripped;
8202
8203                 /* turn raid0 into single device chunks */
8204                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
8205                         return stripped;
8206
8207                 /* turn mirroring into duplication */
8208                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
8209                              BTRFS_BLOCK_GROUP_RAID10))
8210                         return stripped | BTRFS_BLOCK_GROUP_DUP;
8211         } else {
8212                 /* they already had raid on here, just return */
8213                 if (flags & stripped)
8214                         return flags;
8215
8216                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8217                 stripped = flags & ~stripped;
8218
8219                 /* switch duplicated blocks with raid1 */
8220                 if (flags & BTRFS_BLOCK_GROUP_DUP)
8221                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
8222
8223                 /* this is drive concat, leave it alone */
8224         }
8225
8226         return flags;
8227 }
8228
8229 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
8230 {
8231         struct btrfs_space_info *sinfo = cache->space_info;
8232         u64 num_bytes;
8233         u64 min_allocable_bytes;
8234         int ret = -ENOSPC;
8235
8236
8237         /*
8238          * We need some metadata space and system metadata space for
8239          * allocating chunks in some corner cases until we force to set
8240          * it to be readonly.
8241          */
8242         if ((sinfo->flags &
8243              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
8244             !force)
8245                 min_allocable_bytes = 1 * 1024 * 1024;
8246         else
8247                 min_allocable_bytes = 0;
8248
8249         spin_lock(&sinfo->lock);
8250         spin_lock(&cache->lock);
8251
8252         if (cache->ro) {
8253                 ret = 0;
8254                 goto out;
8255         }
8256
8257         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8258                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8259
8260         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8261             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
8262             min_allocable_bytes <= sinfo->total_bytes) {
8263                 sinfo->bytes_readonly += num_bytes;
8264                 cache->ro = 1;
8265                 ret = 0;
8266         }
8267 out:
8268         spin_unlock(&cache->lock);
8269         spin_unlock(&sinfo->lock);
8270         return ret;
8271 }
8272
8273 int btrfs_set_block_group_ro(struct btrfs_root *root,
8274                              struct btrfs_block_group_cache *cache)
8275
8276 {
8277         struct btrfs_trans_handle *trans;
8278         u64 alloc_flags;
8279         int ret;
8280
8281         BUG_ON(cache->ro);
8282
8283         trans = btrfs_join_transaction(root);
8284         if (IS_ERR(trans))
8285                 return PTR_ERR(trans);
8286
8287         alloc_flags = update_block_group_flags(root, cache->flags);
8288         if (alloc_flags != cache->flags) {
8289                 ret = do_chunk_alloc(trans, root, alloc_flags,
8290                                      CHUNK_ALLOC_FORCE);
8291                 if (ret < 0)
8292                         goto out;
8293         }
8294
8295         ret = set_block_group_ro(cache, 0);
8296         if (!ret)
8297                 goto out;
8298         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
8299         ret = do_chunk_alloc(trans, root, alloc_flags,
8300                              CHUNK_ALLOC_FORCE);
8301         if (ret < 0)
8302                 goto out;
8303         ret = set_block_group_ro(cache, 0);
8304 out:
8305         btrfs_end_transaction(trans, root);
8306         return ret;
8307 }
8308
8309 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
8310                             struct btrfs_root *root, u64 type)
8311 {
8312         u64 alloc_flags = get_alloc_profile(root, type);
8313         return do_chunk_alloc(trans, root, alloc_flags,
8314                               CHUNK_ALLOC_FORCE);
8315 }
8316
8317 /*
8318  * helper to account the unused space of all the readonly block group in the
8319  * list. takes mirrors into account.
8320  */
8321 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
8322 {
8323         struct btrfs_block_group_cache *block_group;
8324         u64 free_bytes = 0;
8325         int factor;
8326
8327         list_for_each_entry(block_group, groups_list, list) {
8328                 spin_lock(&block_group->lock);
8329
8330                 if (!block_group->ro) {
8331                         spin_unlock(&block_group->lock);
8332                         continue;
8333                 }
8334
8335                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
8336                                           BTRFS_BLOCK_GROUP_RAID10 |
8337                                           BTRFS_BLOCK_GROUP_DUP))
8338                         factor = 2;
8339                 else
8340                         factor = 1;
8341
8342                 free_bytes += (block_group->key.offset -
8343                                btrfs_block_group_used(&block_group->item)) *
8344                                factor;
8345
8346                 spin_unlock(&block_group->lock);
8347         }
8348
8349         return free_bytes;
8350 }
8351
8352 /*
8353  * helper to account the unused space of all the readonly block group in the
8354  * space_info. takes mirrors into account.
8355  */
8356 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
8357 {
8358         int i;
8359         u64 free_bytes = 0;
8360
8361         spin_lock(&sinfo->lock);
8362
8363         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
8364                 if (!list_empty(&sinfo->block_groups[i]))
8365                         free_bytes += __btrfs_get_ro_block_group_free_space(
8366                                                 &sinfo->block_groups[i]);
8367
8368         spin_unlock(&sinfo->lock);
8369
8370         return free_bytes;
8371 }
8372
8373 void btrfs_set_block_group_rw(struct btrfs_root *root,
8374                               struct btrfs_block_group_cache *cache)
8375 {
8376         struct btrfs_space_info *sinfo = cache->space_info;
8377         u64 num_bytes;
8378
8379         BUG_ON(!cache->ro);
8380
8381         spin_lock(&sinfo->lock);
8382         spin_lock(&cache->lock);
8383         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8384                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8385         sinfo->bytes_readonly -= num_bytes;
8386         cache->ro = 0;
8387         spin_unlock(&cache->lock);
8388         spin_unlock(&sinfo->lock);
8389 }
8390
8391 /*
8392  * checks to see if its even possible to relocate this block group.
8393  *
8394  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
8395  * ok to go ahead and try.
8396  */
8397 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
8398 {
8399         struct btrfs_block_group_cache *block_group;
8400         struct btrfs_space_info *space_info;
8401         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
8402         struct btrfs_device *device;
8403         struct btrfs_trans_handle *trans;
8404         u64 min_free;
8405         u64 dev_min = 1;
8406         u64 dev_nr = 0;
8407         u64 target;
8408         int index;
8409         int full = 0;
8410         int ret = 0;
8411
8412         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
8413
8414         /* odd, couldn't find the block group, leave it alone */
8415         if (!block_group)
8416                 return -1;
8417
8418         min_free = btrfs_block_group_used(&block_group->item);
8419
8420         /* no bytes used, we're good */
8421         if (!min_free)
8422                 goto out;
8423
8424         space_info = block_group->space_info;
8425         spin_lock(&space_info->lock);
8426
8427         full = space_info->full;
8428
8429         /*
8430          * if this is the last block group we have in this space, we can't
8431          * relocate it unless we're able to allocate a new chunk below.
8432          *
8433          * Otherwise, we need to make sure we have room in the space to handle
8434          * all of the extents from this block group.  If we can, we're good
8435          */
8436         if ((space_info->total_bytes != block_group->key.offset) &&
8437             (space_info->bytes_used + space_info->bytes_reserved +
8438              space_info->bytes_pinned + space_info->bytes_readonly +
8439              min_free < space_info->total_bytes)) {
8440                 spin_unlock(&space_info->lock);
8441                 goto out;
8442         }
8443         spin_unlock(&space_info->lock);
8444
8445         /*
8446          * ok we don't have enough space, but maybe we have free space on our
8447          * devices to allocate new chunks for relocation, so loop through our
8448          * alloc devices and guess if we have enough space.  if this block
8449          * group is going to be restriped, run checks against the target
8450          * profile instead of the current one.
8451          */
8452         ret = -1;
8453
8454         /*
8455          * index:
8456          *      0: raid10
8457          *      1: raid1
8458          *      2: dup
8459          *      3: raid0
8460          *      4: single
8461          */
8462         target = get_restripe_target(root->fs_info, block_group->flags);
8463         if (target) {
8464                 index = __get_raid_index(extended_to_chunk(target));
8465         } else {
8466                 /*
8467                  * this is just a balance, so if we were marked as full
8468                  * we know there is no space for a new chunk
8469                  */
8470                 if (full)
8471                         goto out;
8472
8473                 index = get_block_group_index(block_group);
8474         }
8475
8476         if (index == BTRFS_RAID_RAID10) {
8477                 dev_min = 4;
8478                 /* Divide by 2 */
8479                 min_free >>= 1;
8480         } else if (index == BTRFS_RAID_RAID1) {
8481                 dev_min = 2;
8482         } else if (index == BTRFS_RAID_DUP) {
8483                 /* Multiply by 2 */
8484                 min_free <<= 1;
8485         } else if (index == BTRFS_RAID_RAID0) {
8486                 dev_min = fs_devices->rw_devices;
8487                 do_div(min_free, dev_min);
8488         }
8489
8490         /* We need to do this so that we can look at pending chunks */
8491         trans = btrfs_join_transaction(root);
8492         if (IS_ERR(trans)) {
8493                 ret = PTR_ERR(trans);
8494                 goto out;
8495         }
8496
8497         mutex_lock(&root->fs_info->chunk_mutex);
8498         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
8499                 u64 dev_offset;
8500
8501                 /*
8502                  * check to make sure we can actually find a chunk with enough
8503                  * space to fit our block group in.
8504                  */
8505                 if (device->total_bytes > device->bytes_used + min_free &&
8506                     !device->is_tgtdev_for_dev_replace) {
8507                         ret = find_free_dev_extent(trans, device, min_free,
8508                                                    &dev_offset, NULL);
8509                         if (!ret)
8510                                 dev_nr++;
8511
8512                         if (dev_nr >= dev_min)
8513                                 break;
8514
8515                         ret = -1;
8516                 }
8517         }
8518         mutex_unlock(&root->fs_info->chunk_mutex);
8519         btrfs_end_transaction(trans, root);
8520 out:
8521         btrfs_put_block_group(block_group);
8522         return ret;
8523 }
8524
8525 static int find_first_block_group(struct btrfs_root *root,
8526                 struct btrfs_path *path, struct btrfs_key *key)
8527 {
8528         int ret = 0;
8529         struct btrfs_key found_key;
8530         struct extent_buffer *leaf;
8531         int slot;
8532
8533         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
8534         if (ret < 0)
8535                 goto out;
8536
8537         while (1) {
8538                 slot = path->slots[0];
8539                 leaf = path->nodes[0];
8540                 if (slot >= btrfs_header_nritems(leaf)) {
8541                         ret = btrfs_next_leaf(root, path);
8542                         if (ret == 0)
8543                                 continue;
8544                         if (ret < 0)
8545                                 goto out;
8546                         break;
8547                 }
8548                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
8549
8550                 if (found_key.objectid >= key->objectid &&
8551                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
8552                         ret = 0;
8553                         goto out;
8554                 }
8555                 path->slots[0]++;
8556         }
8557 out:
8558         return ret;
8559 }
8560
8561 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
8562 {
8563         struct btrfs_block_group_cache *block_group;
8564         u64 last = 0;
8565
8566         while (1) {
8567                 struct inode *inode;
8568
8569                 block_group = btrfs_lookup_first_block_group(info, last);
8570                 while (block_group) {
8571                         spin_lock(&block_group->lock);
8572                         if (block_group->iref)
8573                                 break;
8574                         spin_unlock(&block_group->lock);
8575                         block_group = next_block_group(info->tree_root,
8576                                                        block_group);
8577                 }
8578                 if (!block_group) {
8579                         if (last == 0)
8580                                 break;
8581                         last = 0;
8582                         continue;
8583                 }
8584
8585                 inode = block_group->inode;
8586                 block_group->iref = 0;
8587                 block_group->inode = NULL;
8588                 spin_unlock(&block_group->lock);
8589                 iput(inode);
8590                 last = block_group->key.objectid + block_group->key.offset;
8591                 btrfs_put_block_group(block_group);
8592         }
8593 }
8594
8595 int btrfs_free_block_groups(struct btrfs_fs_info *info)
8596 {
8597         struct btrfs_block_group_cache *block_group;
8598         struct btrfs_space_info *space_info;
8599         struct btrfs_caching_control *caching_ctl;
8600         struct rb_node *n;
8601
8602         down_write(&info->commit_root_sem);
8603         while (!list_empty(&info->caching_block_groups)) {
8604                 caching_ctl = list_entry(info->caching_block_groups.next,
8605                                          struct btrfs_caching_control, list);
8606                 list_del(&caching_ctl->list);
8607                 put_caching_control(caching_ctl);
8608         }
8609         up_write(&info->commit_root_sem);
8610
8611         spin_lock(&info->block_group_cache_lock);
8612         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
8613                 block_group = rb_entry(n, struct btrfs_block_group_cache,
8614                                        cache_node);
8615                 rb_erase(&block_group->cache_node,
8616                          &info->block_group_cache_tree);
8617                 spin_unlock(&info->block_group_cache_lock);
8618
8619                 down_write(&block_group->space_info->groups_sem);
8620                 list_del(&block_group->list);
8621                 up_write(&block_group->space_info->groups_sem);
8622
8623                 if (block_group->cached == BTRFS_CACHE_STARTED)
8624                         wait_block_group_cache_done(block_group);
8625
8626                 /*
8627                  * We haven't cached this block group, which means we could
8628                  * possibly have excluded extents on this block group.
8629                  */
8630                 if (block_group->cached == BTRFS_CACHE_NO ||
8631                     block_group->cached == BTRFS_CACHE_ERROR)
8632                         free_excluded_extents(info->extent_root, block_group);
8633
8634                 btrfs_remove_free_space_cache(block_group);
8635                 btrfs_put_block_group(block_group);
8636
8637                 spin_lock(&info->block_group_cache_lock);
8638         }
8639         spin_unlock(&info->block_group_cache_lock);
8640
8641         /* now that all the block groups are freed, go through and
8642          * free all the space_info structs.  This is only called during
8643          * the final stages of unmount, and so we know nobody is
8644          * using them.  We call synchronize_rcu() once before we start,
8645          * just to be on the safe side.
8646          */
8647         synchronize_rcu();
8648
8649         release_global_block_rsv(info);
8650
8651         while (!list_empty(&info->space_info)) {
8652                 int i;
8653
8654                 space_info = list_entry(info->space_info.next,
8655                                         struct btrfs_space_info,
8656                                         list);
8657                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
8658                         if (WARN_ON(space_info->bytes_pinned > 0 ||
8659                             space_info->bytes_reserved > 0 ||
8660                             space_info->bytes_may_use > 0)) {
8661                                 dump_space_info(space_info, 0, 0);
8662                         }
8663                 }
8664                 list_del(&space_info->list);
8665                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
8666                         struct kobject *kobj;
8667                         kobj = space_info->block_group_kobjs[i];
8668                         space_info->block_group_kobjs[i] = NULL;
8669                         if (kobj) {
8670                                 kobject_del(kobj);
8671                                 kobject_put(kobj);
8672                         }
8673                 }
8674                 kobject_del(&space_info->kobj);
8675                 kobject_put(&space_info->kobj);
8676         }
8677         return 0;
8678 }
8679
8680 static void __link_block_group(struct btrfs_space_info *space_info,
8681                                struct btrfs_block_group_cache *cache)
8682 {
8683         int index = get_block_group_index(cache);
8684         bool first = false;
8685
8686         down_write(&space_info->groups_sem);
8687         if (list_empty(&space_info->block_groups[index]))
8688                 first = true;
8689         list_add_tail(&cache->list, &space_info->block_groups[index]);
8690         up_write(&space_info->groups_sem);
8691
8692         if (first) {
8693                 struct raid_kobject *rkobj;
8694                 int ret;
8695
8696                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
8697                 if (!rkobj)
8698                         goto out_err;
8699                 rkobj->raid_type = index;
8700                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
8701                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
8702                                   "%s", get_raid_name(index));
8703                 if (ret) {
8704                         kobject_put(&rkobj->kobj);
8705                         goto out_err;
8706                 }
8707                 space_info->block_group_kobjs[index] = &rkobj->kobj;
8708         }
8709
8710         return;
8711 out_err:
8712         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
8713 }
8714
8715 static struct btrfs_block_group_cache *
8716 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
8717 {
8718         struct btrfs_block_group_cache *cache;
8719
8720         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8721         if (!cache)
8722                 return NULL;
8723
8724         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8725                                         GFP_NOFS);
8726         if (!cache->free_space_ctl) {
8727                 kfree(cache);
8728                 return NULL;
8729         }
8730
8731         cache->key.objectid = start;
8732         cache->key.offset = size;
8733         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8734
8735         cache->sectorsize = root->sectorsize;
8736         cache->fs_info = root->fs_info;
8737         cache->full_stripe_len = btrfs_full_stripe_len(root,
8738                                                &root->fs_info->mapping_tree,
8739                                                start);
8740         atomic_set(&cache->count, 1);
8741         spin_lock_init(&cache->lock);
8742         init_rwsem(&cache->data_rwsem);
8743         INIT_LIST_HEAD(&cache->list);
8744         INIT_LIST_HEAD(&cache->cluster_list);
8745         INIT_LIST_HEAD(&cache->new_bg_list);
8746         btrfs_init_free_space_ctl(cache);
8747
8748         return cache;
8749 }
8750
8751 int btrfs_read_block_groups(struct btrfs_root *root)
8752 {
8753         struct btrfs_path *path;
8754         int ret;
8755         struct btrfs_block_group_cache *cache;
8756         struct btrfs_fs_info *info = root->fs_info;
8757         struct btrfs_space_info *space_info;
8758         struct btrfs_key key;
8759         struct btrfs_key found_key;
8760         struct extent_buffer *leaf;
8761         int need_clear = 0;
8762         u64 cache_gen;
8763
8764         root = info->extent_root;
8765         key.objectid = 0;
8766         key.offset = 0;
8767         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
8768         path = btrfs_alloc_path();
8769         if (!path)
8770                 return -ENOMEM;
8771         path->reada = 1;
8772
8773         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
8774         if (btrfs_test_opt(root, SPACE_CACHE) &&
8775             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
8776                 need_clear = 1;
8777         if (btrfs_test_opt(root, CLEAR_CACHE))
8778                 need_clear = 1;
8779
8780         while (1) {
8781                 ret = find_first_block_group(root, path, &key);
8782                 if (ret > 0)
8783                         break;
8784                 if (ret != 0)
8785                         goto error;
8786
8787                 leaf = path->nodes[0];
8788                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
8789
8790                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
8791                                                        found_key.offset);
8792                 if (!cache) {
8793                         ret = -ENOMEM;
8794                         goto error;
8795                 }
8796
8797                 if (need_clear) {
8798                         /*
8799                          * When we mount with old space cache, we need to
8800                          * set BTRFS_DC_CLEAR and set dirty flag.
8801                          *
8802                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
8803                          *    truncate the old free space cache inode and
8804                          *    setup a new one.
8805                          * b) Setting 'dirty flag' makes sure that we flush
8806                          *    the new space cache info onto disk.
8807                          */
8808                         cache->disk_cache_state = BTRFS_DC_CLEAR;
8809                         if (btrfs_test_opt(root, SPACE_CACHE))
8810                                 cache->dirty = 1;
8811                 }
8812
8813                 read_extent_buffer(leaf, &cache->item,
8814                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
8815                                    sizeof(cache->item));
8816                 cache->flags = btrfs_block_group_flags(&cache->item);
8817
8818                 key.objectid = found_key.objectid + found_key.offset;
8819                 btrfs_release_path(path);
8820
8821                 /*
8822                  * We need to exclude the super stripes now so that the space
8823                  * info has super bytes accounted for, otherwise we'll think
8824                  * we have more space than we actually do.
8825                  */
8826                 ret = exclude_super_stripes(root, cache);
8827                 if (ret) {
8828                         /*
8829                          * We may have excluded something, so call this just in
8830                          * case.
8831                          */
8832                         free_excluded_extents(root, cache);
8833                         btrfs_put_block_group(cache);
8834                         goto error;
8835                 }
8836
8837                 /*
8838                  * check for two cases, either we are full, and therefore
8839                  * don't need to bother with the caching work since we won't
8840                  * find any space, or we are empty, and we can just add all
8841                  * the space in and be done with it.  This saves us _alot_ of
8842                  * time, particularly in the full case.
8843                  */
8844                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
8845                         cache->last_byte_to_unpin = (u64)-1;
8846                         cache->cached = BTRFS_CACHE_FINISHED;
8847                         free_excluded_extents(root, cache);
8848                 } else if (btrfs_block_group_used(&cache->item) == 0) {
8849                         cache->last_byte_to_unpin = (u64)-1;
8850                         cache->cached = BTRFS_CACHE_FINISHED;
8851                         add_new_free_space(cache, root->fs_info,
8852                                            found_key.objectid,
8853                                            found_key.objectid +
8854                                            found_key.offset);
8855                         free_excluded_extents(root, cache);
8856                 }
8857
8858                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
8859                 if (ret) {
8860                         btrfs_remove_free_space_cache(cache);
8861                         btrfs_put_block_group(cache);
8862                         goto error;
8863                 }
8864
8865                 ret = update_space_info(info, cache->flags, found_key.offset,
8866                                         btrfs_block_group_used(&cache->item),
8867                                         &space_info);
8868                 if (ret) {
8869                         btrfs_remove_free_space_cache(cache);
8870                         spin_lock(&info->block_group_cache_lock);
8871                         rb_erase(&cache->cache_node,
8872                                  &info->block_group_cache_tree);
8873                         spin_unlock(&info->block_group_cache_lock);
8874                         btrfs_put_block_group(cache);
8875                         goto error;
8876                 }
8877
8878                 cache->space_info = space_info;
8879                 spin_lock(&cache->space_info->lock);
8880                 cache->space_info->bytes_readonly += cache->bytes_super;
8881                 spin_unlock(&cache->space_info->lock);
8882
8883                 __link_block_group(space_info, cache);
8884
8885                 set_avail_alloc_bits(root->fs_info, cache->flags);
8886                 if (btrfs_chunk_readonly(root, cache->key.objectid))
8887                         set_block_group_ro(cache, 1);
8888         }
8889
8890         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
8891                 if (!(get_alloc_profile(root, space_info->flags) &
8892                       (BTRFS_BLOCK_GROUP_RAID10 |
8893                        BTRFS_BLOCK_GROUP_RAID1 |
8894                        BTRFS_BLOCK_GROUP_RAID5 |
8895                        BTRFS_BLOCK_GROUP_RAID6 |
8896                        BTRFS_BLOCK_GROUP_DUP)))
8897                         continue;
8898                 /*
8899                  * avoid allocating from un-mirrored block group if there are
8900                  * mirrored block groups.
8901                  */
8902                 list_for_each_entry(cache,
8903                                 &space_info->block_groups[BTRFS_RAID_RAID0],
8904                                 list)
8905                         set_block_group_ro(cache, 1);
8906                 list_for_each_entry(cache,
8907                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
8908                                 list)
8909                         set_block_group_ro(cache, 1);
8910         }
8911
8912         init_global_block_rsv(info);
8913         ret = 0;
8914 error:
8915         btrfs_free_path(path);
8916         return ret;
8917 }
8918
8919 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
8920                                        struct btrfs_root *root)
8921 {
8922         struct btrfs_block_group_cache *block_group, *tmp;
8923         struct btrfs_root *extent_root = root->fs_info->extent_root;
8924         struct btrfs_block_group_item item;
8925         struct btrfs_key key;
8926         int ret = 0;
8927
8928         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs,
8929                                  new_bg_list) {
8930                 list_del_init(&block_group->new_bg_list);
8931
8932                 if (ret)
8933                         continue;
8934
8935                 spin_lock(&block_group->lock);
8936                 memcpy(&item, &block_group->item, sizeof(item));
8937                 memcpy(&key, &block_group->key, sizeof(key));
8938                 spin_unlock(&block_group->lock);
8939
8940                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
8941                                         sizeof(item));
8942                 if (ret)
8943                         btrfs_abort_transaction(trans, extent_root, ret);
8944                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
8945                                                key.objectid, key.offset);
8946                 if (ret)
8947                         btrfs_abort_transaction(trans, extent_root, ret);
8948         }
8949 }
8950
8951 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
8952                            struct btrfs_root *root, u64 bytes_used,
8953                            u64 type, u64 chunk_objectid, u64 chunk_offset,
8954                            u64 size)
8955 {
8956         int ret;
8957         struct btrfs_root *extent_root;
8958         struct btrfs_block_group_cache *cache;
8959
8960         extent_root = root->fs_info->extent_root;
8961
8962         btrfs_set_log_full_commit(root->fs_info, trans);
8963
8964         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
8965         if (!cache)
8966                 return -ENOMEM;
8967
8968         btrfs_set_block_group_used(&cache->item, bytes_used);
8969         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
8970         btrfs_set_block_group_flags(&cache->item, type);
8971
8972         cache->flags = type;
8973         cache->last_byte_to_unpin = (u64)-1;
8974         cache->cached = BTRFS_CACHE_FINISHED;
8975         ret = exclude_super_stripes(root, cache);
8976         if (ret) {
8977                 /*
8978                  * We may have excluded something, so call this just in
8979                  * case.
8980                  */
8981                 free_excluded_extents(root, cache);
8982                 btrfs_put_block_group(cache);
8983                 return ret;
8984         }
8985
8986         add_new_free_space(cache, root->fs_info, chunk_offset,
8987                            chunk_offset + size);
8988
8989         free_excluded_extents(root, cache);
8990
8991         ret = btrfs_add_block_group_cache(root->fs_info, cache);
8992         if (ret) {
8993                 btrfs_remove_free_space_cache(cache);
8994                 btrfs_put_block_group(cache);
8995                 return ret;
8996         }
8997
8998         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
8999                                 &cache->space_info);
9000         if (ret) {
9001                 btrfs_remove_free_space_cache(cache);
9002                 spin_lock(&root->fs_info->block_group_cache_lock);
9003                 rb_erase(&cache->cache_node,
9004                          &root->fs_info->block_group_cache_tree);
9005                 spin_unlock(&root->fs_info->block_group_cache_lock);
9006                 btrfs_put_block_group(cache);
9007                 return ret;
9008         }
9009         update_global_block_rsv(root->fs_info);
9010
9011         spin_lock(&cache->space_info->lock);
9012         cache->space_info->bytes_readonly += cache->bytes_super;
9013         spin_unlock(&cache->space_info->lock);
9014
9015         __link_block_group(cache->space_info, cache);
9016
9017         list_add_tail(&cache->new_bg_list, &trans->new_bgs);
9018
9019         set_avail_alloc_bits(extent_root->fs_info, type);
9020
9021         return 0;
9022 }
9023
9024 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
9025 {
9026         u64 extra_flags = chunk_to_extended(flags) &
9027                                 BTRFS_EXTENDED_PROFILE_MASK;
9028
9029         write_seqlock(&fs_info->profiles_lock);
9030         if (flags & BTRFS_BLOCK_GROUP_DATA)
9031                 fs_info->avail_data_alloc_bits &= ~extra_flags;
9032         if (flags & BTRFS_BLOCK_GROUP_METADATA)
9033                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
9034         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
9035                 fs_info->avail_system_alloc_bits &= ~extra_flags;
9036         write_sequnlock(&fs_info->profiles_lock);
9037 }
9038
9039 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
9040                              struct btrfs_root *root, u64 group_start)
9041 {
9042         struct btrfs_path *path;
9043         struct btrfs_block_group_cache *block_group;
9044         struct btrfs_free_cluster *cluster;
9045         struct btrfs_root *tree_root = root->fs_info->tree_root;
9046         struct btrfs_key key;
9047         struct inode *inode;
9048         struct kobject *kobj = NULL;
9049         int ret;
9050         int index;
9051         int factor;
9052
9053         root = root->fs_info->extent_root;
9054
9055         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
9056         BUG_ON(!block_group);
9057         BUG_ON(!block_group->ro);
9058
9059         /*
9060          * Free the reserved super bytes from this block group before
9061          * remove it.
9062          */
9063         free_excluded_extents(root, block_group);
9064
9065         memcpy(&key, &block_group->key, sizeof(key));
9066         index = get_block_group_index(block_group);
9067         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
9068                                   BTRFS_BLOCK_GROUP_RAID1 |
9069                                   BTRFS_BLOCK_GROUP_RAID10))
9070                 factor = 2;
9071         else
9072                 factor = 1;
9073
9074         /* make sure this block group isn't part of an allocation cluster */
9075         cluster = &root->fs_info->data_alloc_cluster;
9076         spin_lock(&cluster->refill_lock);
9077         btrfs_return_cluster_to_free_space(block_group, cluster);
9078         spin_unlock(&cluster->refill_lock);
9079
9080         /*
9081          * make sure this block group isn't part of a metadata
9082          * allocation cluster
9083          */
9084         cluster = &root->fs_info->meta_alloc_cluster;
9085         spin_lock(&cluster->refill_lock);
9086         btrfs_return_cluster_to_free_space(block_group, cluster);
9087         spin_unlock(&cluster->refill_lock);
9088
9089         path = btrfs_alloc_path();
9090         if (!path) {
9091                 ret = -ENOMEM;
9092                 goto out;
9093         }
9094
9095         inode = lookup_free_space_inode(tree_root, block_group, path);
9096         if (!IS_ERR(inode)) {
9097                 ret = btrfs_orphan_add(trans, inode);
9098                 if (ret) {
9099                         btrfs_add_delayed_iput(inode);
9100                         goto out;
9101                 }
9102                 clear_nlink(inode);
9103                 /* One for the block groups ref */
9104                 spin_lock(&block_group->lock);
9105                 if (block_group->iref) {
9106                         block_group->iref = 0;
9107                         block_group->inode = NULL;
9108                         spin_unlock(&block_group->lock);
9109                         iput(inode);
9110                 } else {
9111                         spin_unlock(&block_group->lock);
9112                 }
9113                 /* One for our lookup ref */
9114                 btrfs_add_delayed_iput(inode);
9115         }
9116
9117         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
9118         key.offset = block_group->key.objectid;
9119         key.type = 0;
9120
9121         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
9122         if (ret < 0)
9123                 goto out;
9124         if (ret > 0)
9125                 btrfs_release_path(path);
9126         if (ret == 0) {
9127                 ret = btrfs_del_item(trans, tree_root, path);
9128                 if (ret)
9129                         goto out;
9130                 btrfs_release_path(path);
9131         }
9132
9133         spin_lock(&root->fs_info->block_group_cache_lock);
9134         rb_erase(&block_group->cache_node,
9135                  &root->fs_info->block_group_cache_tree);
9136
9137         if (root->fs_info->first_logical_byte == block_group->key.objectid)
9138                 root->fs_info->first_logical_byte = (u64)-1;
9139         spin_unlock(&root->fs_info->block_group_cache_lock);
9140
9141         down_write(&block_group->space_info->groups_sem);
9142         /*
9143          * we must use list_del_init so people can check to see if they
9144          * are still on the list after taking the semaphore
9145          */
9146         list_del_init(&block_group->list);
9147         if (list_empty(&block_group->space_info->block_groups[index])) {
9148                 kobj = block_group->space_info->block_group_kobjs[index];
9149                 block_group->space_info->block_group_kobjs[index] = NULL;
9150                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
9151         }
9152         up_write(&block_group->space_info->groups_sem);
9153         if (kobj) {
9154                 kobject_del(kobj);
9155                 kobject_put(kobj);
9156         }
9157
9158         if (block_group->cached == BTRFS_CACHE_STARTED)
9159                 wait_block_group_cache_done(block_group);
9160
9161         btrfs_remove_free_space_cache(block_group);
9162
9163         spin_lock(&block_group->space_info->lock);
9164         block_group->space_info->total_bytes -= block_group->key.offset;
9165         block_group->space_info->bytes_readonly -= block_group->key.offset;
9166         block_group->space_info->disk_total -= block_group->key.offset * factor;
9167         spin_unlock(&block_group->space_info->lock);
9168
9169         memcpy(&key, &block_group->key, sizeof(key));
9170
9171         btrfs_clear_space_info_full(root->fs_info);
9172
9173         btrfs_put_block_group(block_group);
9174         btrfs_put_block_group(block_group);
9175
9176         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
9177         if (ret > 0)
9178                 ret = -EIO;
9179         if (ret < 0)
9180                 goto out;
9181
9182         ret = btrfs_del_item(trans, root, path);
9183 out:
9184         btrfs_free_path(path);
9185         return ret;
9186 }
9187
9188 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
9189 {
9190         struct btrfs_space_info *space_info;
9191         struct btrfs_super_block *disk_super;
9192         u64 features;
9193         u64 flags;
9194         int mixed = 0;
9195         int ret;
9196
9197         disk_super = fs_info->super_copy;
9198         if (!btrfs_super_root(disk_super))
9199                 return 1;
9200
9201         features = btrfs_super_incompat_flags(disk_super);
9202         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
9203                 mixed = 1;
9204
9205         flags = BTRFS_BLOCK_GROUP_SYSTEM;
9206         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9207         if (ret)
9208                 goto out;
9209
9210         if (mixed) {
9211                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
9212                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9213         } else {
9214                 flags = BTRFS_BLOCK_GROUP_METADATA;
9215                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9216                 if (ret)
9217                         goto out;
9218
9219                 flags = BTRFS_BLOCK_GROUP_DATA;
9220                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9221         }
9222 out:
9223         return ret;
9224 }
9225
9226 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
9227 {
9228         return unpin_extent_range(root, start, end);
9229 }
9230
9231 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
9232                                u64 num_bytes, u64 *actual_bytes)
9233 {
9234         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
9235 }
9236
9237 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
9238 {
9239         struct btrfs_fs_info *fs_info = root->fs_info;
9240         struct btrfs_block_group_cache *cache = NULL;
9241         u64 group_trimmed;
9242         u64 start;
9243         u64 end;
9244         u64 trimmed = 0;
9245         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
9246         int ret = 0;
9247
9248         /*
9249          * try to trim all FS space, our block group may start from non-zero.
9250          */
9251         if (range->len == total_bytes)
9252                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
9253         else
9254                 cache = btrfs_lookup_block_group(fs_info, range->start);
9255
9256         while (cache) {
9257                 if (cache->key.objectid >= (range->start + range->len)) {
9258                         btrfs_put_block_group(cache);
9259                         break;
9260                 }
9261
9262                 start = max(range->start, cache->key.objectid);
9263                 end = min(range->start + range->len,
9264                                 cache->key.objectid + cache->key.offset);
9265
9266                 if (end - start >= range->minlen) {
9267                         if (!block_group_cache_done(cache)) {
9268                                 ret = cache_block_group(cache, 0);
9269                                 if (ret) {
9270                                         btrfs_put_block_group(cache);
9271                                         break;
9272                                 }
9273                                 ret = wait_block_group_cache_done(cache);
9274                                 if (ret) {
9275                                         btrfs_put_block_group(cache);
9276                                         break;
9277                                 }
9278                         }
9279                         ret = btrfs_trim_block_group(cache,
9280                                                      &group_trimmed,
9281                                                      start,
9282                                                      end,
9283                                                      range->minlen);
9284
9285                         trimmed += group_trimmed;
9286                         if (ret) {
9287                                 btrfs_put_block_group(cache);
9288                                 break;
9289                         }
9290                 }
9291
9292                 cache = next_block_group(fs_info->tree_root, cache);
9293         }
9294
9295         range->len = trimmed;
9296         return ret;
9297 }
9298
9299 /*
9300  * btrfs_{start,end}_write() is similar to mnt_{want, drop}_write(),
9301  * they are used to prevent the some tasks writing data into the page cache
9302  * by nocow before the subvolume is snapshoted, but flush the data into
9303  * the disk after the snapshot creation.
9304  */
9305 void btrfs_end_nocow_write(struct btrfs_root *root)
9306 {
9307         percpu_counter_dec(&root->subv_writers->counter);
9308         /*
9309          * Make sure counter is updated before we wake up
9310          * waiters.
9311          */
9312         smp_mb();
9313         if (waitqueue_active(&root->subv_writers->wait))
9314                 wake_up(&root->subv_writers->wait);
9315 }
9316
9317 int btrfs_start_nocow_write(struct btrfs_root *root)
9318 {
9319         if (unlikely(atomic_read(&root->will_be_snapshoted)))
9320                 return 0;
9321
9322         percpu_counter_inc(&root->subv_writers->counter);
9323         /*
9324          * Make sure counter is updated before we check for snapshot creation.
9325          */
9326         smp_mb();
9327         if (unlikely(atomic_read(&root->will_be_snapshoted))) {
9328                 btrfs_end_nocow_write(root);
9329                 return 0;
9330         }
9331         return 1;
9332 }