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 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 110                                u64 num_bytes);
 111 int btrfs_pin_extent(struct btrfs_root *root,
 112                      u64 bytenr, u64 num_bytes, int reserved);
 113
 114 static noinline int
 115 block_group_cache_done(struct btrfs_block_group_cache *cache)
 116 {
 117         smp_mb();
 118         return cache->cached == BTRFS_CACHE_FINISHED ||
 119                 cache->cached == BTRFS_CACHE_ERROR;
 120 }
 121
 122 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 123 {
 124         return (cache->flags & bits) == bits;
 125 }
 126
 127 static void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 128 {
 129         atomic_inc(&cache->count);
 130 }
 131
 132 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 133 {
 134         if (atomic_dec_and_test(&cache->count)) {
 135                 WARN_ON(cache->pinned > 0);
 136                 WARN_ON(cache->reserved > 0);
 137                 kfree(cache->free_space_ctl);
 138                 kfree(cache);
 139         }
 140 }
 141
 142 /*
 143  * this adds the block group to the fs_info rb tree for the block group
 144  * cache
 145  */
 146 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 147                                 struct btrfs_block_group_cache *block_group)
 148 {
 149         struct rb_node **p;
 150         struct rb_node *parent = NULL;
 151         struct btrfs_block_group_cache *cache;
 152
 153         spin_lock(&info->block_group_cache_lock);
 154         p = &info->block_group_cache_tree.rb_node;
 155
 156         while (*p) {
 157                 parent = *p;
 158                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 159                                  cache_node);
 160                 if (block_group->key.objectid < cache->key.objectid) {
 161                         p = &(*p)->rb_left;
 162                 } else if (block_group->key.objectid > cache->key.objectid) {
 163                         p = &(*p)->rb_right;
 164                 } else {
 165                         spin_unlock(&info->block_group_cache_lock);
 166                         return -EEXIST;
 167                 }
 168         }
 169
 170         rb_link_node(&block_group->cache_node, parent, p);
 171         rb_insert_color(&block_group->cache_node,
 172                         &info->block_group_cache_tree);
 173
 174         if (info->first_logical_byte > block_group->key.objectid)
 175                 info->first_logical_byte = block_group->key.objectid;
 176
 177         spin_unlock(&info->block_group_cache_lock);
 178
 179         return 0;
 180 }
 181
 182 /*
 183  * This will return the block group at or after bytenr if contains is 0, else
 184  * it will return the block group that contains the bytenr
 185  */
 186 static struct btrfs_block_group_cache *
 187 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 188                               int contains)
 189 {
 190         struct btrfs_block_group_cache *cache, *ret = NULL;
 191         struct rb_node *n;
 192         u64 end, start;
 193
 194         spin_lock(&info->block_group_cache_lock);
 195         n = info->block_group_cache_tree.rb_node;
 196
 197         while (n) {
 198                 cache = rb_entry(n, struct btrfs_block_group_cache,
 199                                  cache_node);
 200                 end = cache->key.objectid + cache->key.offset - 1;
 201                 start = cache->key.objectid;
 202
 203                 if (bytenr < start) {
 204                         if (!contains && (!ret || start < ret->key.objectid))
 205                                 ret = cache;
 206                         n = n->rb_left;
 207                 } else if (bytenr > start) {
 208                         if (contains && bytenr <= end) {
 209                                 ret = cache;
 210                                 break;
 211                         }
 212                         n = n->rb_right;
 213                 } else {
 214                         ret = cache;
 215                         break;
 216                 }
 217         }
 218         if (ret) {
 219                 btrfs_get_block_group(ret);
 220                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 221                         info->first_logical_byte = ret->key.objectid;
 222         }
 223         spin_unlock(&info->block_group_cache_lock);
 224
 225         return ret;
 226 }
 227
 228 static int add_excluded_extent(struct btrfs_root *root,
 229                                u64 start, u64 num_bytes)
 230 {
 231         u64 end = start + num_bytes - 1;
 232         set_extent_bits(&root->fs_info->freed_extents[0],
 233                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 234         set_extent_bits(&root->fs_info->freed_extents[1],
 235                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 236         return 0;
 237 }
 238
 239 static void free_excluded_extents(struct btrfs_root *root,
 240                                   struct btrfs_block_group_cache *cache)
 241 {
 242         u64 start, end;
 243
 244         start = cache->key.objectid;
 245         end = start + cache->key.offset - 1;
 246
 247         clear_extent_bits(&root->fs_info->freed_extents[0],
 248                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 249         clear_extent_bits(&root->fs_info->freed_extents[1],
 250                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 251 }
 252
 253 static int exclude_super_stripes(struct btrfs_root *root,
 254                                  struct btrfs_block_group_cache *cache)
 255 {
 256         u64 bytenr;
 257         u64 *logical;
 258         int stripe_len;
 259         int i, nr, ret;
 260
 261         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 262                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 263                 cache->bytes_super += stripe_len;
 264                 ret = add_excluded_extent(root, cache->key.objectid,
 265                                           stripe_len);
 266                 if (ret)
 267                         return ret;
 268         }
 269
 270         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 271                 bytenr = btrfs_sb_offset(i);
 272                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 273                                        cache->key.objectid, bytenr,
 274                                        0, &logical, &nr, &stripe_len);
 275                 if (ret)
 276                         return ret;
 277
 278                 while (nr--) {
 279                         u64 start, len;
 280
 281                         if (logical[nr] > cache->key.objectid +
 282                             cache->key.offset)
 283                                 continue;
 284
 285                         if (logical[nr] + stripe_len <= cache->key.objectid)
 286                                 continue;
 287
 288                         start = logical[nr];
 289                         if (start < cache->key.objectid) {
 290                                 start = cache->key.objectid;
 291                                 len = (logical[nr] + stripe_len) - start;
 292                         } else {
 293                                 len = min_t(u64, stripe_len,
 294                                             cache->key.objectid +
 295                                             cache->key.offset - start);
 296                         }
 297
 298                         cache->bytes_super += len;
 299                         ret = add_excluded_extent(root, start, len);
 300                         if (ret) {
 301                                 kfree(logical);
 302                                 return ret;
 303                         }
 304                 }
 305
 306                 kfree(logical);
 307         }
 308         return 0;
 309 }
 310
 311 static struct btrfs_caching_control *
 312 get_caching_control(struct btrfs_block_group_cache *cache)
 313 {
 314         struct btrfs_caching_control *ctl;
 315
 316         spin_lock(&cache->lock);
 317         if (cache->cached != BTRFS_CACHE_STARTED) {
 318                 spin_unlock(&cache->lock);
 319                 return NULL;
 320         }
 321
 322         /* We're loading it the fast way, so we don't have a caching_ctl. */
 323         if (!cache->caching_ctl) {
 324                 spin_unlock(&cache->lock);
 325                 return NULL;
 326         }
 327
 328         ctl = cache->caching_ctl;
 329         atomic_inc(&ctl->count);
 330         spin_unlock(&cache->lock);
 331         return ctl;
 332 }
 333
 334 static void put_caching_control(struct btrfs_caching_control *ctl)
 335 {
 336         if (atomic_dec_and_test(&ctl->count))
 337                 kfree(ctl);
 338 }
 339
 340 /*
 341  * this is only called by cache_block_group, since we could have freed extents
 342  * we need to check the pinned_extents for any extents that can't be used yet
 343  * since their free space will be released as soon as the transaction commits.
 344  */
 345 static u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 346                               struct btrfs_fs_info *info, u64 start, u64 end)
 347 {
 348         u64 extent_start, extent_end, size, total_added = 0;
 349         int ret;
 350
 351         while (start < end) {
 352                 ret = find_first_extent_bit(info->pinned_extents, start,
 353                                             &extent_start, &extent_end,
 354                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 355                                             NULL);
 356                 if (ret)
 357                         break;
 358
 359                 if (extent_start <= start) {
 360                         start = extent_end + 1;
 361                 } else if (extent_start > start && extent_start < end) {
 362                         size = extent_start - start;
 363                         total_added += size;
 364                         ret = btrfs_add_free_space(block_group, start,
 365                                                    size);
 366                         BUG_ON(ret); /* -ENOMEM or logic error */
 367                         start = extent_end + 1;
 368                 } else {
 369                         break;
 370                 }
 371         }
 372
 373         if (start < end) {
 374                 size = end - start;
 375                 total_added += size;
 376                 ret = btrfs_add_free_space(block_group, start, size);
 377                 BUG_ON(ret); /* -ENOMEM or logic error */
 378         }
 379
 380         return total_added;
 381 }
 382
 383 static noinline void caching_thread(struct btrfs_work *work)
 384 {
 385         struct btrfs_block_group_cache *block_group;
 386         struct btrfs_fs_info *fs_info;
 387         struct btrfs_caching_control *caching_ctl;
 388         struct btrfs_root *extent_root;
 389         struct btrfs_path *path;
 390         struct extent_buffer *leaf;
 391         struct btrfs_key key;
 392         u64 total_found = 0;
 393         u64 last = 0;
 394         u32 nritems;
 395         int ret = -ENOMEM;
 396
 397         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 398         block_group = caching_ctl->block_group;
 399         fs_info = block_group->fs_info;
 400         extent_root = fs_info->extent_root;
 401
 402         path = btrfs_alloc_path();
 403         if (!path)
 404                 goto out;
 405
 406         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 407
 408         /*
 409          * We don't want to deadlock with somebody trying to allocate a new
 410          * extent for the extent root while also trying to search the extent
 411          * root to add free space.  So we skip locking and search the commit
 412          * root, since its read-only
 413          */
 414         path->skip_locking = 1;
 415         path->search_commit_root = 1;
 416         path->reada = 1;
 417
 418         key.objectid = last;
 419         key.offset = 0;
 420         key.type = BTRFS_EXTENT_ITEM_KEY;
 421 again:
 422         mutex_lock(&caching_ctl->mutex);
 423         /* need to make sure the commit_root doesn't disappear */
 424         down_read(&fs_info->commit_root_sem);
 425
 426 next:
 427         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 428         if (ret < 0)
 429                 goto err;
 430
 431         leaf = path->nodes[0];
 432         nritems = btrfs_header_nritems(leaf);
 433
 434         while (1) {
 435                 if (btrfs_fs_closing(fs_info) > 1) {
 436                         last = (u64)-1;
 437                         break;
 438                 }
 439
 440                 if (path->slots[0] < nritems) {
 441                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 442                 } else {
 443                         ret = find_next_key(path, 0, &key);
 444                         if (ret)
 445                                 break;
 446
 447                         if (need_resched() ||
 448                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 449                                 caching_ctl->progress = last;
 450                                 btrfs_release_path(path);
 451                                 up_read(&fs_info->commit_root_sem);
 452                                 mutex_unlock(&caching_ctl->mutex);
 453                                 cond_resched();
 454                                 goto again;
 455                         }
 456
 457                         ret = btrfs_next_leaf(extent_root, path);
 458                         if (ret < 0)
 459                                 goto err;
 460                         if (ret)
 461                                 break;
 462                         leaf = path->nodes[0];
 463                         nritems = btrfs_header_nritems(leaf);
 464                         continue;
 465                 }
 466
 467                 if (key.objectid < last) {
 468                         key.objectid = last;
 469                         key.offset = 0;
 470                         key.type = BTRFS_EXTENT_ITEM_KEY;
 471
 472                         caching_ctl->progress = last;
 473                         btrfs_release_path(path);
 474                         goto next;
 475                 }
 476
 477                 if (key.objectid < block_group->key.objectid) {
 478                         path->slots[0]++;
 479                         continue;
 480                 }
 481
 482                 if (key.objectid >= block_group->key.objectid +
 483                     block_group->key.offset)
 484                         break;
 485
 486                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 487                     key.type == BTRFS_METADATA_ITEM_KEY) {
 488                         total_found += add_new_free_space(block_group,
 489                                                           fs_info, last,
 490                                                           key.objectid);
 491                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 492                                 last = key.objectid +
 493                                         fs_info->tree_root->leafsize;
 494                         else
 495                                 last = key.objectid + key.offset;
 496
 497                         if (total_found > (1024 * 1024 * 2)) {
 498                                 total_found = 0;
 499                                 wake_up(&caching_ctl->wait);
 500                         }
 501                 }
 502                 path->slots[0]++;
 503         }
 504         ret = 0;
 505
 506         total_found += add_new_free_space(block_group, fs_info, last,
 507                                           block_group->key.objectid +
 508                                           block_group->key.offset);
 509         caching_ctl->progress = (u64)-1;
 510
 511         spin_lock(&block_group->lock);
 512         block_group->caching_ctl = NULL;
 513         block_group->cached = BTRFS_CACHE_FINISHED;
 514         spin_unlock(&block_group->lock);
 515
 516 err:
 517         btrfs_free_path(path);
 518         up_read(&fs_info->commit_root_sem);
 519
 520         free_excluded_extents(extent_root, block_group);
 521
 522         mutex_unlock(&caching_ctl->mutex);
 523 out:
 524         if (ret) {
 525                 spin_lock(&block_group->lock);
 526                 block_group->caching_ctl = NULL;
 527                 block_group->cached = BTRFS_CACHE_ERROR;
 528                 spin_unlock(&block_group->lock);
 529         }
 530         wake_up(&caching_ctl->wait);
 531
 532         put_caching_control(caching_ctl);
 533         btrfs_put_block_group(block_group);
 534 }
 535
 536 static int cache_block_group(struct btrfs_block_group_cache *cache,
 537                              int load_cache_only)
 538 {
 539         DEFINE_WAIT(wait);
 540         struct btrfs_fs_info *fs_info = cache->fs_info;
 541         struct btrfs_caching_control *caching_ctl;
 542         int ret = 0;
 543
 544         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 545         if (!caching_ctl)
 546                 return -ENOMEM;
 547
 548         INIT_LIST_HEAD(&caching_ctl->list);
 549         mutex_init(&caching_ctl->mutex);
 550         init_waitqueue_head(&caching_ctl->wait);
 551         caching_ctl->block_group = cache;
 552         caching_ctl->progress = cache->key.objectid;
 553         atomic_set(&caching_ctl->count, 1);
 554         btrfs_init_work(&caching_ctl->work, caching_thread, NULL, NULL);
 555
 556         spin_lock(&cache->lock);
 557         /*
 558          * This should be a rare occasion, but this could happen I think in the
 559          * case where one thread starts to load the space cache info, and then
 560          * some other thread starts a transaction commit which tries to do an
 561          * allocation while the other thread is still loading the space cache
 562          * info.  The previous loop should have kept us from choosing this block
 563          * group, but if we've moved to the state where we will wait on caching
 564          * block groups we need to first check if we're doing a fast load here,
 565          * so we can wait for it to finish, otherwise we could end up allocating
 566          * from a block group who's cache gets evicted for one reason or
 567          * another.
 568          */
 569         while (cache->cached == BTRFS_CACHE_FAST) {
 570                 struct btrfs_caching_control *ctl;
 571
 572                 ctl = cache->caching_ctl;
 573                 atomic_inc(&ctl->count);
 574                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 575                 spin_unlock(&cache->lock);
 576
 577                 schedule();
 578
 579                 finish_wait(&ctl->wait, &wait);
 580                 put_caching_control(ctl);
 581                 spin_lock(&cache->lock);
 582         }
 583
 584         if (cache->cached != BTRFS_CACHE_NO) {
 585                 spin_unlock(&cache->lock);
 586                 kfree(caching_ctl);
 587                 return 0;
 588         }
 589         WARN_ON(cache->caching_ctl);
 590         cache->caching_ctl = caching_ctl;
 591         cache->cached = BTRFS_CACHE_FAST;
 592         spin_unlock(&cache->lock);
 593
 594         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 595                 ret = load_free_space_cache(fs_info, cache);
 596
 597                 spin_lock(&cache->lock);
 598                 if (ret == 1) {
 599                         cache->caching_ctl = NULL;
 600                         cache->cached = BTRFS_CACHE_FINISHED;
 601                         cache->last_byte_to_unpin = (u64)-1;
 602                 } else {
 603                         if (load_cache_only) {
 604                                 cache->caching_ctl = NULL;
 605                                 cache->cached = BTRFS_CACHE_NO;
 606                         } else {
 607                                 cache->cached = BTRFS_CACHE_STARTED;
 608                         }
 609                 }
 610                 spin_unlock(&cache->lock);
 611                 wake_up(&caching_ctl->wait);
 612                 if (ret == 1) {
 613                         put_caching_control(caching_ctl);
 614                         free_excluded_extents(fs_info->extent_root, cache);
 615                         return 0;
 616                 }
 617         } else {
 618                 /*
 619                  * We are not going to do the fast caching, set cached to the
 620                  * appropriate value and wakeup any waiters.
 621                  */
 622                 spin_lock(&cache->lock);
 623                 if (load_cache_only) {
 624                         cache->caching_ctl = NULL;
 625                         cache->cached = BTRFS_CACHE_NO;
 626                 } else {
 627                         cache->cached = BTRFS_CACHE_STARTED;
 628                 }
 629                 spin_unlock(&cache->lock);
 630                 wake_up(&caching_ctl->wait);
 631         }
 632
 633         if (load_cache_only) {
 634                 put_caching_control(caching_ctl);
 635                 return 0;
 636         }
 637
 638         down_write(&fs_info->commit_root_sem);
 639         atomic_inc(&caching_ctl->count);
 640         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 641         up_write(&fs_info->commit_root_sem);
 642
 643         btrfs_get_block_group(cache);
 644
 645         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 646
 647         return ret;
 648 }
 649
 650 /*
 651  * return the block group that starts at or after bytenr
 652  */
 653 static struct btrfs_block_group_cache *
 654 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 655 {
 656         struct btrfs_block_group_cache *cache;
 657
 658         cache = block_group_cache_tree_search(info, bytenr, 0);
 659
 660         return cache;
 661 }
 662
 663 /*
 664  * return the block group that contains the given bytenr
 665  */
 666 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 667                                                  struct btrfs_fs_info *info,
 668                                                  u64 bytenr)
 669 {
 670         struct btrfs_block_group_cache *cache;
 671
 672         cache = block_group_cache_tree_search(info, bytenr, 1);
 673
 674         return cache;
 675 }
 676
 677 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 678                                                   u64 flags)
 679 {
 680         struct list_head *head = &info->space_info;
 681         struct btrfs_space_info *found;
 682
 683         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 684
 685         rcu_read_lock();
 686         list_for_each_entry_rcu(found, head, list) {
 687                 if (found->flags & flags) {
 688                         rcu_read_unlock();
 689                         return found;
 690                 }
 691         }
 692         rcu_read_unlock();
 693         return NULL;
 694 }
 695
 696 /*
 697  * after adding space to the filesystem, we need to clear the full flags
 698  * on all the space infos.
 699  */
 700 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 701 {
 702         struct list_head *head = &info->space_info;
 703         struct btrfs_space_info *found;
 704
 705         rcu_read_lock();
 706         list_for_each_entry_rcu(found, head, list)
 707                 found->full = 0;
 708         rcu_read_unlock();
 709 }
 710
 711 /* simple helper to search for an existing extent at a given offset */
 712 int btrfs_lookup_extent(struct btrfs_root *root, u64 start, u64 len)
 713 {
 714         int ret;
 715         struct btrfs_key key;
 716         struct btrfs_path *path;
 717
 718         path = btrfs_alloc_path();
 719         if (!path)
 720                 return -ENOMEM;
 721
 722         key.objectid = start;
 723         key.offset = len;
 724         key.type = BTRFS_EXTENT_ITEM_KEY;
 725         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 726                                 0, 0);
 727         if (ret > 0) {
 728                 btrfs_item_key_to_cpu(path->nodes[0], &key, path->slots[0]);
 729                 if (key.objectid == start &&
 730                     key.type == BTRFS_METADATA_ITEM_KEY)
 731                         ret = 0;
 732         }
 733         btrfs_free_path(path);
 734         return ret;
 735 }
 736
 737 /*
 738  * helper function to lookup reference count and flags of a tree block.
 739  *
 740  * the head node for delayed ref is used to store the sum of all the
 741  * reference count modifications queued up in the rbtree. the head
 742  * node may also store the extent flags to set. This way you can check
 743  * to see what the reference count and extent flags would be if all of
 744  * the delayed refs are not processed.
 745  */
 746 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 747                              struct btrfs_root *root, u64 bytenr,
 748                              u64 offset, int metadata, u64 *refs, u64 *flags)
 749 {
 750         struct btrfs_delayed_ref_head *head;
 751         struct btrfs_delayed_ref_root *delayed_refs;
 752         struct btrfs_path *path;
 753         struct btrfs_extent_item *ei;
 754         struct extent_buffer *leaf;
 755         struct btrfs_key key;
 756         u32 item_size;
 757         u64 num_refs;
 758         u64 extent_flags;
 759         int ret;
 760
 761         /*
 762          * If we don't have skinny metadata, don't bother doing anything
 763          * different
 764          */
 765         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 766                 offset = root->leafsize;
 767                 metadata = 0;
 768         }
 769
 770         path = btrfs_alloc_path();
 771         if (!path)
 772                 return -ENOMEM;
 773
 774         if (!trans) {
 775                 path->skip_locking = 1;
 776                 path->search_commit_root = 1;
 777         }
 778
 779 search_again:
 780         key.objectid = bytenr;
 781         key.offset = offset;
 782         if (metadata)
 783                 key.type = BTRFS_METADATA_ITEM_KEY;
 784         else
 785                 key.type = BTRFS_EXTENT_ITEM_KEY;
 786
 787 again:
 788         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 789                                 &key, path, 0, 0);
 790         if (ret < 0)
 791                 goto out_free;
 792
 793         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 794                 if (path->slots[0]) {
 795                         path->slots[0]--;
 796                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 797                                               path->slots[0]);
 798                         if (key.objectid == bytenr &&
 799                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 800                             key.offset == root->leafsize)
 801                                 ret = 0;
 802                 }
 803                 if (ret) {
 804                         key.objectid = bytenr;
 805                         key.type = BTRFS_EXTENT_ITEM_KEY;
 806                         key.offset = root->leafsize;
 807                         btrfs_release_path(path);
 808                         goto again;
 809                 }
 810         }
 811
 812         if (ret == 0) {
 813                 leaf = path->nodes[0];
 814                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 815                 if (item_size >= sizeof(*ei)) {
 816                         ei = btrfs_item_ptr(leaf, path->slots[0],
 817                                             struct btrfs_extent_item);
 818                         num_refs = btrfs_extent_refs(leaf, ei);
 819                         extent_flags = btrfs_extent_flags(leaf, ei);
 820                 } else {
 821 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 822                         struct btrfs_extent_item_v0 *ei0;
 823                         BUG_ON(item_size != sizeof(*ei0));
 824                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 825                                              struct btrfs_extent_item_v0);
 826                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 827                         /* FIXME: this isn't correct for data */
 828                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 829 #else
 830                         BUG();
 831 #endif
 832                 }
 833                 BUG_ON(num_refs == 0);
 834         } else {
 835                 num_refs = 0;
 836                 extent_flags = 0;
 837                 ret = 0;
 838         }
 839
 840         if (!trans)
 841                 goto out;
 842
 843         delayed_refs = &trans->transaction->delayed_refs;
 844         spin_lock(&delayed_refs->lock);
 845         head = btrfs_find_delayed_ref_head(trans, bytenr);
 846         if (head) {
 847                 if (!mutex_trylock(&head->mutex)) {
 848                         atomic_inc(&head->node.refs);
 849                         spin_unlock(&delayed_refs->lock);
 850
 851                         btrfs_release_path(path);
 852
 853                         /*
 854                          * Mutex was contended, block until it's released and try
 855                          * again
 856                          */
 857                         mutex_lock(&head->mutex);
 858                         mutex_unlock(&head->mutex);
 859                         btrfs_put_delayed_ref(&head->node);
 860                         goto search_again;
 861                 }
 862                 spin_lock(&head->lock);
 863                 if (head->extent_op && head->extent_op->update_flags)
 864                         extent_flags |= head->extent_op->flags_to_set;
 865                 else
 866                         BUG_ON(num_refs == 0);
 867
 868                 num_refs += head->node.ref_mod;
 869                 spin_unlock(&head->lock);
 870                 mutex_unlock(&head->mutex);
 871         }
 872         spin_unlock(&delayed_refs->lock);
 873 out:
 874         WARN_ON(num_refs == 0);
 875         if (refs)
 876                 *refs = num_refs;
 877         if (flags)
 878                 *flags = extent_flags;
 879 out_free:
 880         btrfs_free_path(path);
 881         return ret;
 882 }
 883
 884 /*
 885  * Back reference rules.  Back refs have three main goals:
 886  *
 887  * 1) differentiate between all holders of references to an extent so that
 888  *    when a reference is dropped we can make sure it was a valid reference
 889  *    before freeing the extent.
 890  *
 891  * 2) Provide enough information to quickly find the holders of an extent
 892  *    if we notice a given block is corrupted or bad.
 893  *
 894  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 895  *    maintenance.  This is actually the same as #2, but with a slightly
 896  *    different use case.
 897  *
 898  * There are two kinds of back refs. The implicit back refs is optimized
 899  * for pointers in non-shared tree blocks. For a given pointer in a block,
 900  * back refs of this kind provide information about the block's owner tree
 901  * and the pointer's key. These information allow us to find the block by
 902  * b-tree searching. The full back refs is for pointers in tree blocks not
 903  * referenced by their owner trees. The location of tree block is recorded
 904  * in the back refs. Actually the full back refs is generic, and can be
 905  * used in all cases the implicit back refs is used. The major shortcoming
 906  * of the full back refs is its overhead. Every time a tree block gets
 907  * COWed, we have to update back refs entry for all pointers in it.
 908  *
 909  * For a newly allocated tree block, we use implicit back refs for
 910  * pointers in it. This means most tree related operations only involve
 911  * implicit back refs. For a tree block created in old transaction, the
 912  * only way to drop a reference to it is COW it. So we can detect the
 913  * event that tree block loses its owner tree's reference and do the
 914  * back refs conversion.
 915  *
 916  * When a tree block is COW'd through a tree, there are four cases:
 917  *
 918  * The reference count of the block is one and the tree is the block's
 919  * owner tree. Nothing to do in this case.
 920  *
 921  * The reference count of the block is one and the tree is not the
 922  * block's owner tree. In this case, full back refs is used for pointers
 923  * in the block. Remove these full back refs, add implicit back refs for
 924  * every pointers in the new block.
 925  *
 926  * The reference count of the block is greater than one and the tree is
 927  * the block's owner tree. In this case, implicit back refs is used for
 928  * pointers in the block. Add full back refs for every pointers in the
 929  * block, increase lower level extents' reference counts. The original
 930  * implicit back refs are entailed to the new block.
 931  *
 932  * The reference count of the block is greater than one and the tree is
 933  * not the block's owner tree. Add implicit back refs for every pointer in
 934  * the new block, increase lower level extents' reference count.
 935  *
 936  * Back Reference Key composing:
 937  *
 938  * The key objectid corresponds to the first byte in the extent,
 939  * The key type is used to differentiate between types of back refs.
 940  * There are different meanings of the key offset for different types
 941  * of back refs.
 942  *
 943  * File extents can be referenced by:
 944  *
 945  * - multiple snapshots, subvolumes, or different generations in one subvol
 946  * - different files inside a single subvolume
 947  * - different offsets inside a file (bookend extents in file.c)
 948  *
 949  * The extent ref structure for the implicit back refs has fields for:
 950  *
 951  * - Objectid of the subvolume root
 952  * - objectid of the file holding the reference
 953  * - original offset in the file
 954  * - how many bookend extents
 955  *
 956  * The key offset for the implicit back refs is hash of the first
 957  * three fields.
 958  *
 959  * The extent ref structure for the full back refs has field for:
 960  *
 961  * - number of pointers in the tree leaf
 962  *
 963  * The key offset for the implicit back refs is the first byte of
 964  * the tree leaf
 965  *
 966  * When a file extent is allocated, The implicit back refs is used.
 967  * the fields are filled in:
 968  *
 969  *     (root_key.objectid, inode objectid, offset in file, 1)
 970  *
 971  * When a file extent is removed file truncation, we find the
 972  * corresponding implicit back refs and check the following fields:
 973  *
 974  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
 975  *
 976  * Btree extents can be referenced by:
 977  *
 978  * - Different subvolumes
 979  *
 980  * Both the implicit back refs and the full back refs for tree blocks
 981  * only consist of key. The key offset for the implicit back refs is
 982  * objectid of block's owner tree. The key offset for the full back refs
 983  * is the first byte of parent block.
 984  *
 985  * When implicit back refs is used, information about the lowest key and
 986  * level of the tree block are required. These information are stored in
 987  * tree block info structure.
 988  */
 989
 990 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 991 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
 992                                   struct btrfs_root *root,
 993                                   struct btrfs_path *path,
 994                                   u64 owner, u32 extra_size)
 995 {
 996         struct btrfs_extent_item *item;
 997         struct btrfs_extent_item_v0 *ei0;
 998         struct btrfs_extent_ref_v0 *ref0;
 999         struct btrfs_tree_block_info *bi;
1000         struct extent_buffer *leaf;
1001         struct btrfs_key key;
1002         struct btrfs_key found_key;
1003         u32 new_size = sizeof(*item);
1004         u64 refs;
1005         int ret;
1006
1007         leaf = path->nodes[0];
1008         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1009
1010         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1011         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1012                              struct btrfs_extent_item_v0);
1013         refs = btrfs_extent_refs_v0(leaf, ei0);
1014
1015         if (owner == (u64)-1) {
1016                 while (1) {
1017                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1018                                 ret = btrfs_next_leaf(root, path);
1019                                 if (ret < 0)
1020                                         return ret;
1021                                 BUG_ON(ret > 0); /* Corruption */
1022                                 leaf = path->nodes[0];
1023                         }
1024                         btrfs_item_key_to_cpu(leaf, &found_key,
1025                                               path->slots[0]);
1026                         BUG_ON(key.objectid != found_key.objectid);
1027                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1028                                 path->slots[0]++;
1029                                 continue;
1030                         }
1031                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1032                                               struct btrfs_extent_ref_v0);
1033                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1034                         break;
1035                 }
1036         }
1037         btrfs_release_path(path);
1038
1039         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1040                 new_size += sizeof(*bi);
1041
1042         new_size -= sizeof(*ei0);
1043         ret = btrfs_search_slot(trans, root, &key, path,
1044                                 new_size + extra_size, 1);
1045         if (ret < 0)
1046                 return ret;
1047         BUG_ON(ret); /* Corruption */
1048
1049         btrfs_extend_item(root, path, new_size);
1050
1051         leaf = path->nodes[0];
1052         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1053         btrfs_set_extent_refs(leaf, item, refs);
1054         /* FIXME: get real generation */
1055         btrfs_set_extent_generation(leaf, item, 0);
1056         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1057                 btrfs_set_extent_flags(leaf, item,
1058                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1059                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1060                 bi = (struct btrfs_tree_block_info *)(item + 1);
1061                 /* FIXME: get first key of the block */
1062                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1063                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1064         } else {
1065                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1066         }
1067         btrfs_mark_buffer_dirty(leaf);
1068         return 0;
1069 }
1070 #endif
1071
1072 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1073 {
1074         u32 high_crc = ~(u32)0;
1075         u32 low_crc = ~(u32)0;
1076         __le64 lenum;
1077
1078         lenum = cpu_to_le64(root_objectid);
1079         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1080         lenum = cpu_to_le64(owner);
1081         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1082         lenum = cpu_to_le64(offset);
1083         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1084
1085         return ((u64)high_crc << 31) ^ (u64)low_crc;
1086 }
1087
1088 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1089                                      struct btrfs_extent_data_ref *ref)
1090 {
1091         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1092                                     btrfs_extent_data_ref_objectid(leaf, ref),
1093                                     btrfs_extent_data_ref_offset(leaf, ref));
1094 }
1095
1096 static int match_extent_data_ref(struct extent_buffer *leaf,
1097                                  struct btrfs_extent_data_ref *ref,
1098                                  u64 root_objectid, u64 owner, u64 offset)
1099 {
1100         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1101             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1102             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1103                 return 0;
1104         return 1;
1105 }
1106
1107 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1108                                            struct btrfs_root *root,
1109                                            struct btrfs_path *path,
1110                                            u64 bytenr, u64 parent,
1111                                            u64 root_objectid,
1112                                            u64 owner, u64 offset)
1113 {
1114         struct btrfs_key key;
1115         struct btrfs_extent_data_ref *ref;
1116         struct extent_buffer *leaf;
1117         u32 nritems;
1118         int ret;
1119         int recow;
1120         int err = -ENOENT;
1121
1122         key.objectid = bytenr;
1123         if (parent) {
1124                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1125                 key.offset = parent;
1126         } else {
1127                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1128                 key.offset = hash_extent_data_ref(root_objectid,
1129                                                   owner, offset);
1130         }
1131 again:
1132         recow = 0;
1133         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1134         if (ret < 0) {
1135                 err = ret;
1136                 goto fail;
1137         }
1138
1139         if (parent) {
1140                 if (!ret)
1141                         return 0;
1142 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1143                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1144                 btrfs_release_path(path);
1145                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1146                 if (ret < 0) {
1147                         err = ret;
1148                         goto fail;
1149                 }
1150                 if (!ret)
1151                         return 0;
1152 #endif
1153                 goto fail;
1154         }
1155
1156         leaf = path->nodes[0];
1157         nritems = btrfs_header_nritems(leaf);
1158         while (1) {
1159                 if (path->slots[0] >= nritems) {
1160                         ret = btrfs_next_leaf(root, path);
1161                         if (ret < 0)
1162                                 err = ret;
1163                         if (ret)
1164                                 goto fail;
1165
1166                         leaf = path->nodes[0];
1167                         nritems = btrfs_header_nritems(leaf);
1168                         recow = 1;
1169                 }
1170
1171                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1172                 if (key.objectid != bytenr ||
1173                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1174                         goto fail;
1175
1176                 ref = btrfs_item_ptr(leaf, path->slots[0],
1177                                      struct btrfs_extent_data_ref);
1178
1179                 if (match_extent_data_ref(leaf, ref, root_objectid,
1180                                           owner, offset)) {
1181                         if (recow) {
1182                                 btrfs_release_path(path);
1183                                 goto again;
1184                         }
1185                         err = 0;
1186                         break;
1187                 }
1188                 path->slots[0]++;
1189         }
1190 fail:
1191         return err;
1192 }
1193
1194 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1195                                            struct btrfs_root *root,
1196                                            struct btrfs_path *path,
1197                                            u64 bytenr, u64 parent,
1198                                            u64 root_objectid, u64 owner,
1199                                            u64 offset, int refs_to_add)
1200 {
1201         struct btrfs_key key;
1202         struct extent_buffer *leaf;
1203         u32 size;
1204         u32 num_refs;
1205         int ret;
1206
1207         key.objectid = bytenr;
1208         if (parent) {
1209                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1210                 key.offset = parent;
1211                 size = sizeof(struct btrfs_shared_data_ref);
1212         } else {
1213                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1214                 key.offset = hash_extent_data_ref(root_objectid,
1215                                                   owner, offset);
1216                 size = sizeof(struct btrfs_extent_data_ref);
1217         }
1218
1219         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1220         if (ret && ret != -EEXIST)
1221                 goto fail;
1222
1223         leaf = path->nodes[0];
1224         if (parent) {
1225                 struct btrfs_shared_data_ref *ref;
1226                 ref = btrfs_item_ptr(leaf, path->slots[0],
1227                                      struct btrfs_shared_data_ref);
1228                 if (ret == 0) {
1229                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1230                 } else {
1231                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1232                         num_refs += refs_to_add;
1233                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1234                 }
1235         } else {
1236                 struct btrfs_extent_data_ref *ref;
1237                 while (ret == -EEXIST) {
1238                         ref = btrfs_item_ptr(leaf, path->slots[0],
1239                                              struct btrfs_extent_data_ref);
1240                         if (match_extent_data_ref(leaf, ref, root_objectid,
1241                                                   owner, offset))
1242                                 break;
1243                         btrfs_release_path(path);
1244                         key.offset++;
1245                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1246                                                       size);
1247                         if (ret && ret != -EEXIST)
1248                                 goto fail;
1249
1250                         leaf = path->nodes[0];
1251                 }
1252                 ref = btrfs_item_ptr(leaf, path->slots[0],
1253                                      struct btrfs_extent_data_ref);
1254                 if (ret == 0) {
1255                         btrfs_set_extent_data_ref_root(leaf, ref,
1256                                                        root_objectid);
1257                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1258                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1259                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1260                 } else {
1261                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1262                         num_refs += refs_to_add;
1263                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1264                 }
1265         }
1266         btrfs_mark_buffer_dirty(leaf);
1267         ret = 0;
1268 fail:
1269         btrfs_release_path(path);
1270         return ret;
1271 }
1272
1273 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1274                                            struct btrfs_root *root,
1275                                            struct btrfs_path *path,
1276                                            int refs_to_drop, int *last_ref)
1277 {
1278         struct btrfs_key key;
1279         struct btrfs_extent_data_ref *ref1 = NULL;
1280         struct btrfs_shared_data_ref *ref2 = NULL;
1281         struct extent_buffer *leaf;
1282         u32 num_refs = 0;
1283         int ret = 0;
1284
1285         leaf = path->nodes[0];
1286         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1287
1288         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1289                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1290                                       struct btrfs_extent_data_ref);
1291                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1292         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1293                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1294                                       struct btrfs_shared_data_ref);
1295                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1296 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1297         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1298                 struct btrfs_extent_ref_v0 *ref0;
1299                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1300                                       struct btrfs_extent_ref_v0);
1301                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1302 #endif
1303         } else {
1304                 BUG();
1305         }
1306
1307         BUG_ON(num_refs < refs_to_drop);
1308         num_refs -= refs_to_drop;
1309
1310         if (num_refs == 0) {
1311                 ret = btrfs_del_item(trans, root, path);
1312                 *last_ref = 1;
1313         } else {
1314                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1315                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1316                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1317                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1318 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1319                 else {
1320                         struct btrfs_extent_ref_v0 *ref0;
1321                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1322                                         struct btrfs_extent_ref_v0);
1323                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1324                 }
1325 #endif
1326                 btrfs_mark_buffer_dirty(leaf);
1327         }
1328         return ret;
1329 }
1330
1331 static noinline u32 extent_data_ref_count(struct btrfs_root *root,
1332                                           struct btrfs_path *path,
1333                                           struct btrfs_extent_inline_ref *iref)
1334 {
1335         struct btrfs_key key;
1336         struct extent_buffer *leaf;
1337         struct btrfs_extent_data_ref *ref1;
1338         struct btrfs_shared_data_ref *ref2;
1339         u32 num_refs = 0;
1340
1341         leaf = path->nodes[0];
1342         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1343         if (iref) {
1344                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1345                     BTRFS_EXTENT_DATA_REF_KEY) {
1346                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1347                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1348                 } else {
1349                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1350                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1351                 }
1352         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1353                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1354                                       struct btrfs_extent_data_ref);
1355                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1356         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1357                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1358                                       struct btrfs_shared_data_ref);
1359                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1360 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1361         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1362                 struct btrfs_extent_ref_v0 *ref0;
1363                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1364                                       struct btrfs_extent_ref_v0);
1365                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1366 #endif
1367         } else {
1368                 WARN_ON(1);
1369         }
1370         return num_refs;
1371 }
1372
1373 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1374                                           struct btrfs_root *root,
1375                                           struct btrfs_path *path,
1376                                           u64 bytenr, u64 parent,
1377                                           u64 root_objectid)
1378 {
1379         struct btrfs_key key;
1380         int ret;
1381
1382         key.objectid = bytenr;
1383         if (parent) {
1384                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1385                 key.offset = parent;
1386         } else {
1387                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1388                 key.offset = root_objectid;
1389         }
1390
1391         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1392         if (ret > 0)
1393                 ret = -ENOENT;
1394 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1395         if (ret == -ENOENT && parent) {
1396                 btrfs_release_path(path);
1397                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1398                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1399                 if (ret > 0)
1400                         ret = -ENOENT;
1401         }
1402 #endif
1403         return ret;
1404 }
1405
1406 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1407                                           struct btrfs_root *root,
1408                                           struct btrfs_path *path,
1409                                           u64 bytenr, u64 parent,
1410                                           u64 root_objectid)
1411 {
1412         struct btrfs_key key;
1413         int ret;
1414
1415         key.objectid = bytenr;
1416         if (parent) {
1417                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1418                 key.offset = parent;
1419         } else {
1420                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1421                 key.offset = root_objectid;
1422         }
1423
1424         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1425         btrfs_release_path(path);
1426         return ret;
1427 }
1428
1429 static inline int extent_ref_type(u64 parent, u64 owner)
1430 {
1431         int type;
1432         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1433                 if (parent > 0)
1434                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1435                 else
1436                         type = BTRFS_TREE_BLOCK_REF_KEY;
1437         } else {
1438                 if (parent > 0)
1439                         type = BTRFS_SHARED_DATA_REF_KEY;
1440                 else
1441                         type = BTRFS_EXTENT_DATA_REF_KEY;
1442         }
1443         return type;
1444 }
1445
1446 static int find_next_key(struct btrfs_path *path, int level,
1447                          struct btrfs_key *key)
1448
1449 {
1450         for (; level < BTRFS_MAX_LEVEL; level++) {
1451                 if (!path->nodes[level])
1452                         break;
1453                 if (path->slots[level] + 1 >=
1454                     btrfs_header_nritems(path->nodes[level]))
1455                         continue;
1456                 if (level == 0)
1457                         btrfs_item_key_to_cpu(path->nodes[level], key,
1458                                               path->slots[level] + 1);
1459                 else
1460                         btrfs_node_key_to_cpu(path->nodes[level], key,
1461                                               path->slots[level] + 1);
1462                 return 0;
1463         }
1464         return 1;
1465 }
1466
1467 /*
1468  * look for inline back ref. if back ref is found, *ref_ret is set
1469  * to the address of inline back ref, and 0 is returned.
1470  *
1471  * if back ref isn't found, *ref_ret is set to the address where it
1472  * should be inserted, and -ENOENT is returned.
1473  *
1474  * if insert is true and there are too many inline back refs, the path
1475  * points to the extent item, and -EAGAIN is returned.
1476  *
1477  * NOTE: inline back refs are ordered in the same way that back ref
1478  *       items in the tree are ordered.
1479  */
1480 static noinline_for_stack
1481 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1482                                  struct btrfs_root *root,
1483                                  struct btrfs_path *path,
1484                                  struct btrfs_extent_inline_ref **ref_ret,
1485                                  u64 bytenr, u64 num_bytes,
1486                                  u64 parent, u64 root_objectid,
1487                                  u64 owner, u64 offset, int insert)
1488 {
1489         struct btrfs_key key;
1490         struct extent_buffer *leaf;
1491         struct btrfs_extent_item *ei;
1492         struct btrfs_extent_inline_ref *iref;
1493         u64 flags;
1494         u64 item_size;
1495         unsigned long ptr;
1496         unsigned long end;
1497         int extra_size;
1498         int type;
1499         int want;
1500         int ret;
1501         int err = 0;
1502         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1503                                                  SKINNY_METADATA);
1504
1505         key.objectid = bytenr;
1506         key.type = BTRFS_EXTENT_ITEM_KEY;
1507         key.offset = num_bytes;
1508
1509         want = extent_ref_type(parent, owner);
1510         if (insert) {
1511                 extra_size = btrfs_extent_inline_ref_size(want);
1512                 path->keep_locks = 1;
1513         } else
1514                 extra_size = -1;
1515
1516         /*
1517          * Owner is our parent level, so we can just add one to get the level
1518          * for the block we are interested in.
1519          */
1520         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1521                 key.type = BTRFS_METADATA_ITEM_KEY;
1522                 key.offset = owner;
1523         }
1524
1525 again:
1526         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1527         if (ret < 0) {
1528                 err = ret;
1529                 goto out;
1530         }
1531
1532         /*
1533          * We may be a newly converted file system which still has the old fat
1534          * extent entries for metadata, so try and see if we have one of those.
1535          */
1536         if (ret > 0 && skinny_metadata) {
1537                 skinny_metadata = false;
1538                 if (path->slots[0]) {
1539                         path->slots[0]--;
1540                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1541                                               path->slots[0]);
1542                         if (key.objectid == bytenr &&
1543                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1544                             key.offset == num_bytes)
1545                                 ret = 0;
1546                 }
1547                 if (ret) {
1548                         key.objectid = bytenr;
1549                         key.type = BTRFS_EXTENT_ITEM_KEY;
1550                         key.offset = num_bytes;
1551                         btrfs_release_path(path);
1552                         goto again;
1553                 }
1554         }
1555
1556         if (ret && !insert) {
1557                 err = -ENOENT;
1558                 goto out;
1559         } else if (WARN_ON(ret)) {
1560                 err = -EIO;
1561                 goto out;
1562         }
1563
1564         leaf = path->nodes[0];
1565         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1566 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1567         if (item_size < sizeof(*ei)) {
1568                 if (!insert) {
1569                         err = -ENOENT;
1570                         goto out;
1571                 }
1572                 ret = convert_extent_item_v0(trans, root, path, owner,
1573                                              extra_size);
1574                 if (ret < 0) {
1575                         err = ret;
1576                         goto out;
1577                 }
1578                 leaf = path->nodes[0];
1579                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1580         }
1581 #endif
1582         BUG_ON(item_size < sizeof(*ei));
1583
1584         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1585         flags = btrfs_extent_flags(leaf, ei);
1586
1587         ptr = (unsigned long)(ei + 1);
1588         end = (unsigned long)ei + item_size;
1589
1590         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1591                 ptr += sizeof(struct btrfs_tree_block_info);
1592                 BUG_ON(ptr > end);
1593         }
1594
1595         err = -ENOENT;
1596         while (1) {
1597                 if (ptr >= end) {
1598                         WARN_ON(ptr > end);
1599                         break;
1600                 }
1601                 iref = (struct btrfs_extent_inline_ref *)ptr;
1602                 type = btrfs_extent_inline_ref_type(leaf, iref);
1603                 if (want < type)
1604                         break;
1605                 if (want > type) {
1606                         ptr += btrfs_extent_inline_ref_size(type);
1607                         continue;
1608                 }
1609
1610                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1611                         struct btrfs_extent_data_ref *dref;
1612                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1613                         if (match_extent_data_ref(leaf, dref, root_objectid,
1614                                                   owner, offset)) {
1615                                 err = 0;
1616                                 break;
1617                         }
1618                         if (hash_extent_data_ref_item(leaf, dref) <
1619                             hash_extent_data_ref(root_objectid, owner, offset))
1620                                 break;
1621                 } else {
1622                         u64 ref_offset;
1623                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1624                         if (parent > 0) {
1625                                 if (parent == ref_offset) {
1626                                         err = 0;
1627                                         break;
1628                                 }
1629                                 if (ref_offset < parent)
1630                                         break;
1631                         } else {
1632                                 if (root_objectid == ref_offset) {
1633                                         err = 0;
1634                                         break;
1635                                 }
1636                                 if (ref_offset < root_objectid)
1637                                         break;
1638                         }
1639                 }
1640                 ptr += btrfs_extent_inline_ref_size(type);
1641         }
1642         if (err == -ENOENT && insert) {
1643                 if (item_size + extra_size >=
1644                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1645                         err = -EAGAIN;
1646                         goto out;
1647                 }
1648                 /*
1649                  * To add new inline back ref, we have to make sure
1650                  * there is no corresponding back ref item.
1651                  * For simplicity, we just do not add new inline back
1652                  * ref if there is any kind of item for this block
1653                  */
1654                 if (find_next_key(path, 0, &key) == 0 &&
1655                     key.objectid == bytenr &&
1656                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1657                         err = -EAGAIN;
1658                         goto out;
1659                 }
1660         }
1661         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1662 out:
1663         if (insert) {
1664                 path->keep_locks = 0;
1665                 btrfs_unlock_up_safe(path, 1);
1666         }
1667         return err;
1668 }
1669
1670 /*
1671  * helper to add new inline back ref
1672  */
1673 static noinline_for_stack
1674 void setup_inline_extent_backref(struct btrfs_root *root,
1675                                  struct btrfs_path *path,
1676                                  struct btrfs_extent_inline_ref *iref,
1677                                  u64 parent, u64 root_objectid,
1678                                  u64 owner, u64 offset, int refs_to_add,
1679                                  struct btrfs_delayed_extent_op *extent_op)
1680 {
1681         struct extent_buffer *leaf;
1682         struct btrfs_extent_item *ei;
1683         unsigned long ptr;
1684         unsigned long end;
1685         unsigned long item_offset;
1686         u64 refs;
1687         int size;
1688         int type;
1689
1690         leaf = path->nodes[0];
1691         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1692         item_offset = (unsigned long)iref - (unsigned long)ei;
1693
1694         type = extent_ref_type(parent, owner);
1695         size = btrfs_extent_inline_ref_size(type);
1696
1697         btrfs_extend_item(root, path, size);
1698
1699         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1700         refs = btrfs_extent_refs(leaf, ei);
1701         refs += refs_to_add;
1702         btrfs_set_extent_refs(leaf, ei, refs);
1703         if (extent_op)
1704                 __run_delayed_extent_op(extent_op, leaf, ei);
1705
1706         ptr = (unsigned long)ei + item_offset;
1707         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1708         if (ptr < end - size)
1709                 memmove_extent_buffer(leaf, ptr + size, ptr,
1710                                       end - size - ptr);
1711
1712         iref = (struct btrfs_extent_inline_ref *)ptr;
1713         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1714         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1715                 struct btrfs_extent_data_ref *dref;
1716                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1717                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1718                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1719                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1720                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1721         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1722                 struct btrfs_shared_data_ref *sref;
1723                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1724                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1725                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1726         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1727                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1728         } else {
1729                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1730         }
1731         btrfs_mark_buffer_dirty(leaf);
1732 }
1733
1734 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1735                                  struct btrfs_root *root,
1736                                  struct btrfs_path *path,
1737                                  struct btrfs_extent_inline_ref **ref_ret,
1738                                  u64 bytenr, u64 num_bytes, u64 parent,
1739                                  u64 root_objectid, u64 owner, u64 offset)
1740 {
1741         int ret;
1742
1743         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1744                                            bytenr, num_bytes, parent,
1745                                            root_objectid, owner, offset, 0);
1746         if (ret != -ENOENT)
1747                 return ret;
1748
1749         btrfs_release_path(path);
1750         *ref_ret = NULL;
1751
1752         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1753                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1754                                             root_objectid);
1755         } else {
1756                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1757                                              root_objectid, owner, offset);
1758         }
1759         return ret;
1760 }
1761
1762 /*
1763  * helper to update/remove inline back ref
1764  */
1765 static noinline_for_stack
1766 void update_inline_extent_backref(struct btrfs_root *root,
1767                                   struct btrfs_path *path,
1768                                   struct btrfs_extent_inline_ref *iref,
1769                                   int refs_to_mod,
1770                                   struct btrfs_delayed_extent_op *extent_op,
1771                                   int *last_ref)
1772 {
1773         struct extent_buffer *leaf;
1774         struct btrfs_extent_item *ei;
1775         struct btrfs_extent_data_ref *dref = NULL;
1776         struct btrfs_shared_data_ref *sref = NULL;
1777         unsigned long ptr;
1778         unsigned long end;
1779         u32 item_size;
1780         int size;
1781         int type;
1782         u64 refs;
1783
1784         leaf = path->nodes[0];
1785         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1786         refs = btrfs_extent_refs(leaf, ei);
1787         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1788         refs += refs_to_mod;
1789         btrfs_set_extent_refs(leaf, ei, refs);
1790         if (extent_op)
1791                 __run_delayed_extent_op(extent_op, leaf, ei);
1792
1793         type = btrfs_extent_inline_ref_type(leaf, iref);
1794
1795         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1796                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1797                 refs = btrfs_extent_data_ref_count(leaf, dref);
1798         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1799                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1800                 refs = btrfs_shared_data_ref_count(leaf, sref);
1801         } else {
1802                 refs = 1;
1803                 BUG_ON(refs_to_mod != -1);
1804         }
1805
1806         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1807         refs += refs_to_mod;
1808
1809         if (refs > 0) {
1810                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1811                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1812                 else
1813                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1814         } else {
1815                 *last_ref = 1;
1816                 size =  btrfs_extent_inline_ref_size(type);
1817                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1818                 ptr = (unsigned long)iref;
1819                 end = (unsigned long)ei + item_size;
1820                 if (ptr + size < end)
1821                         memmove_extent_buffer(leaf, ptr, ptr + size,
1822                                               end - ptr - size);
1823                 item_size -= size;
1824                 btrfs_truncate_item(root, path, item_size, 1);
1825         }
1826         btrfs_mark_buffer_dirty(leaf);
1827 }
1828
1829 static noinline_for_stack
1830 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1831                                  struct btrfs_root *root,
1832                                  struct btrfs_path *path,
1833                                  u64 bytenr, u64 num_bytes, u64 parent,
1834                                  u64 root_objectid, u64 owner,
1835                                  u64 offset, int refs_to_add,
1836                                  struct btrfs_delayed_extent_op *extent_op)
1837 {
1838         struct btrfs_extent_inline_ref *iref;
1839         int ret;
1840
1841         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1842                                            bytenr, num_bytes, parent,
1843                                            root_objectid, owner, offset, 1);
1844         if (ret == 0) {
1845                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1846                 update_inline_extent_backref(root, path, iref,
1847                                              refs_to_add, extent_op, NULL);
1848         } else if (ret == -ENOENT) {
1849                 setup_inline_extent_backref(root, path, iref, parent,
1850                                             root_objectid, owner, offset,
1851                                             refs_to_add, extent_op);
1852                 ret = 0;
1853         }
1854         return ret;
1855 }
1856
1857 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1858                                  struct btrfs_root *root,
1859                                  struct btrfs_path *path,
1860                                  u64 bytenr, u64 parent, u64 root_objectid,
1861                                  u64 owner, u64 offset, int refs_to_add)
1862 {
1863         int ret;
1864         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1865                 BUG_ON(refs_to_add != 1);
1866                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1867                                             parent, root_objectid);
1868         } else {
1869                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1870                                              parent, root_objectid,
1871                                              owner, offset, refs_to_add);
1872         }
1873         return ret;
1874 }
1875
1876 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1877                                  struct btrfs_root *root,
1878                                  struct btrfs_path *path,
1879                                  struct btrfs_extent_inline_ref *iref,
1880                                  int refs_to_drop, int is_data, int *last_ref)
1881 {
1882         int ret = 0;
1883
1884         BUG_ON(!is_data && refs_to_drop != 1);
1885         if (iref) {
1886                 update_inline_extent_backref(root, path, iref,
1887                                              -refs_to_drop, NULL, last_ref);
1888         } else if (is_data) {
1889                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop,
1890                                              last_ref);
1891         } else {
1892                 *last_ref = 1;
1893                 ret = btrfs_del_item(trans, root, path);
1894         }
1895         return ret;
1896 }
1897
1898 static int btrfs_issue_discard(struct block_device *bdev,
1899                                 u64 start, u64 len)
1900 {
1901         return blkdev_issue_discard(bdev, start >> 9, len >> 9, GFP_NOFS, 0);
1902 }
1903
1904 static int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
1905                                 u64 num_bytes, u64 *actual_bytes)
1906 {
1907         int ret;
1908         u64 discarded_bytes = 0;
1909         struct btrfs_bio *bbio = NULL;
1910
1911
1912         /* Tell the block device(s) that the sectors can be discarded */
1913         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
1914                               bytenr, &num_bytes, &bbio, 0);
1915         /* Error condition is -ENOMEM */
1916         if (!ret) {
1917                 struct btrfs_bio_stripe *stripe = bbio->stripes;
1918                 int i;
1919
1920
1921                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
1922                         if (!stripe->dev->can_discard)
1923                                 continue;
1924
1925                         ret = btrfs_issue_discard(stripe->dev->bdev,
1926                                                   stripe->physical,
1927                                                   stripe->length);
1928                         if (!ret)
1929                                 discarded_bytes += stripe->length;
1930                         else if (ret != -EOPNOTSUPP)
1931                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
1932
1933                         /*
1934                          * Just in case we get back EOPNOTSUPP for some reason,
1935                          * just ignore the return value so we don't screw up
1936                          * people calling discard_extent.
1937                          */
1938                         ret = 0;
1939                 }
1940                 kfree(bbio);
1941         }
1942
1943         if (actual_bytes)
1944                 *actual_bytes = discarded_bytes;
1945
1946
1947         if (ret == -EOPNOTSUPP)
1948                 ret = 0;
1949         return ret;
1950 }
1951
1952 /* Can return -ENOMEM */
1953 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1954                          struct btrfs_root *root,
1955                          u64 bytenr, u64 num_bytes, u64 parent,
1956                          u64 root_objectid, u64 owner, u64 offset,
1957                          int no_quota)
1958 {
1959         int ret;
1960         struct btrfs_fs_info *fs_info = root->fs_info;
1961
1962         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
1963                root_objectid == BTRFS_TREE_LOG_OBJECTID);
1964
1965         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1966                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
1967                                         num_bytes,
1968                                         parent, root_objectid, (int)owner,
1969                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
1970         } else {
1971                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
1972                                         num_bytes,
1973                                         parent, root_objectid, owner, offset,
1974                                         BTRFS_ADD_DELAYED_REF, NULL, no_quota);
1975         }
1976         return ret;
1977 }
1978
1979 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
1980                                   struct btrfs_root *root,
1981                                   u64 bytenr, u64 num_bytes,
1982                                   u64 parent, u64 root_objectid,
1983                                   u64 owner, u64 offset, int refs_to_add,
1984                                   int no_quota,
1985                                   struct btrfs_delayed_extent_op *extent_op)
1986 {
1987         struct btrfs_fs_info *fs_info = root->fs_info;
1988         struct btrfs_path *path;
1989         struct extent_buffer *leaf;
1990         struct btrfs_extent_item *item;
1991         struct btrfs_key key;
1992         u64 refs;
1993         int ret;
1994         enum btrfs_qgroup_operation_type type = BTRFS_QGROUP_OPER_ADD_EXCL;
1995
1996         path = btrfs_alloc_path();
1997         if (!path)
1998                 return -ENOMEM;
1999
2000         if (!is_fstree(root_objectid) || !root->fs_info->quota_enabled)
2001                 no_quota = 1;
2002
2003         path->reada = 1;
2004         path->leave_spinning = 1;
2005         /* this will setup the path even if it fails to insert the back ref */
2006         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
2007                                            bytenr, num_bytes, parent,
2008                                            root_objectid, owner, offset,
2009                                            refs_to_add, extent_op);
2010         if ((ret < 0 && ret != -EAGAIN) || (!ret && no_quota))
2011                 goto out;
2012         /*
2013          * Ok we were able to insert an inline extent and it appears to be a new
2014          * reference, deal with the qgroup accounting.
2015          */
2016         if (!ret && !no_quota) {
2017                 ASSERT(root->fs_info->quota_enabled);
2018                 leaf = path->nodes[0];
2019                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2020                 item = btrfs_item_ptr(leaf, path->slots[0],
2021                                       struct btrfs_extent_item);
2022                 if (btrfs_extent_refs(leaf, item) > (u64)refs_to_add)
2023                         type = BTRFS_QGROUP_OPER_ADD_SHARED;
2024                 btrfs_release_path(path);
2025
2026                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
2027                                               bytenr, num_bytes, type, 0);
2028                 goto out;
2029         }
2030
2031         /*
2032          * Ok we had -EAGAIN which means we didn't have space to insert and
2033          * inline extent ref, so just update the reference count and add a
2034          * normal backref.
2035          */
2036         leaf = path->nodes[0];
2037         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2038         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2039         refs = btrfs_extent_refs(leaf, item);
2040         if (refs)
2041                 type = BTRFS_QGROUP_OPER_ADD_SHARED;
2042         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2043         if (extent_op)
2044                 __run_delayed_extent_op(extent_op, leaf, item);
2045
2046         btrfs_mark_buffer_dirty(leaf);
2047         btrfs_release_path(path);
2048
2049         if (!no_quota) {
2050                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
2051                                               bytenr, num_bytes, type, 0);
2052                 if (ret)
2053                         goto out;
2054         }
2055
2056         path->reada = 1;
2057         path->leave_spinning = 1;
2058         /* now insert the actual backref */
2059         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2060                                     path, bytenr, parent, root_objectid,
2061                                     owner, offset, refs_to_add);
2062         if (ret)
2063                 btrfs_abort_transaction(trans, root, ret);
2064 out:
2065         btrfs_free_path(path);
2066         return ret;
2067 }
2068
2069 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2070                                 struct btrfs_root *root,
2071                                 struct btrfs_delayed_ref_node *node,
2072                                 struct btrfs_delayed_extent_op *extent_op,
2073                                 int insert_reserved)
2074 {
2075         int ret = 0;
2076         struct btrfs_delayed_data_ref *ref;
2077         struct btrfs_key ins;
2078         u64 parent = 0;
2079         u64 ref_root = 0;
2080         u64 flags = 0;
2081
2082         ins.objectid = node->bytenr;
2083         ins.offset = node->num_bytes;
2084         ins.type = BTRFS_EXTENT_ITEM_KEY;
2085
2086         ref = btrfs_delayed_node_to_data_ref(node);
2087         trace_run_delayed_data_ref(node, ref, node->action);
2088
2089         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2090                 parent = ref->parent;
2091         ref_root = ref->root;
2092
2093         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2094                 if (extent_op)
2095                         flags |= extent_op->flags_to_set;
2096                 ret = alloc_reserved_file_extent(trans, root,
2097                                                  parent, ref_root, flags,
2098                                                  ref->objectid, ref->offset,
2099                                                  &ins, node->ref_mod);
2100         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2101                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2102                                              node->num_bytes, parent,
2103                                              ref_root, ref->objectid,
2104                                              ref->offset, node->ref_mod,
2105                                              node->no_quota, extent_op);
2106         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2107                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2108                                           node->num_bytes, parent,
2109                                           ref_root, ref->objectid,
2110                                           ref->offset, node->ref_mod,
2111                                           extent_op, node->no_quota);
2112         } else {
2113                 BUG();
2114         }
2115         return ret;
2116 }
2117
2118 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2119                                     struct extent_buffer *leaf,
2120                                     struct btrfs_extent_item *ei)
2121 {
2122         u64 flags = btrfs_extent_flags(leaf, ei);
2123         if (extent_op->update_flags) {
2124                 flags |= extent_op->flags_to_set;
2125                 btrfs_set_extent_flags(leaf, ei, flags);
2126         }
2127
2128         if (extent_op->update_key) {
2129                 struct btrfs_tree_block_info *bi;
2130                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2131                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2132                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2133         }
2134 }
2135
2136 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2137                                  struct btrfs_root *root,
2138                                  struct btrfs_delayed_ref_node *node,
2139                                  struct btrfs_delayed_extent_op *extent_op)
2140 {
2141         struct btrfs_key key;
2142         struct btrfs_path *path;
2143         struct btrfs_extent_item *ei;
2144         struct extent_buffer *leaf;
2145         u32 item_size;
2146         int ret;
2147         int err = 0;
2148         int metadata = !extent_op->is_data;
2149
2150         if (trans->aborted)
2151                 return 0;
2152
2153         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2154                 metadata = 0;
2155
2156         path = btrfs_alloc_path();
2157         if (!path)
2158                 return -ENOMEM;
2159
2160         key.objectid = node->bytenr;
2161
2162         if (metadata) {
2163                 key.type = BTRFS_METADATA_ITEM_KEY;
2164                 key.offset = extent_op->level;
2165         } else {
2166                 key.type = BTRFS_EXTENT_ITEM_KEY;
2167                 key.offset = node->num_bytes;
2168         }
2169
2170 again:
2171         path->reada = 1;
2172         path->leave_spinning = 1;
2173         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2174                                 path, 0, 1);
2175         if (ret < 0) {
2176                 err = ret;
2177                 goto out;
2178         }
2179         if (ret > 0) {
2180                 if (metadata) {
2181                         if (path->slots[0] > 0) {
2182                                 path->slots[0]--;
2183                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2184                                                       path->slots[0]);
2185                                 if (key.objectid == node->bytenr &&
2186                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2187                                     key.offset == node->num_bytes)
2188                                         ret = 0;
2189                         }
2190                         if (ret > 0) {
2191                                 btrfs_release_path(path);
2192                                 metadata = 0;
2193
2194                                 key.objectid = node->bytenr;
2195                                 key.offset = node->num_bytes;
2196                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2197                                 goto again;
2198                         }
2199                 } else {
2200                         err = -EIO;
2201                         goto out;
2202                 }
2203         }
2204
2205         leaf = path->nodes[0];
2206         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2207 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2208         if (item_size < sizeof(*ei)) {
2209                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2210                                              path, (u64)-1, 0);
2211                 if (ret < 0) {
2212                         err = ret;
2213                         goto out;
2214                 }
2215                 leaf = path->nodes[0];
2216                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2217         }
2218 #endif
2219         BUG_ON(item_size < sizeof(*ei));
2220         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2221         __run_delayed_extent_op(extent_op, leaf, ei);
2222
2223         btrfs_mark_buffer_dirty(leaf);
2224 out:
2225         btrfs_free_path(path);
2226         return err;
2227 }
2228
2229 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2230                                 struct btrfs_root *root,
2231                                 struct btrfs_delayed_ref_node *node,
2232                                 struct btrfs_delayed_extent_op *extent_op,
2233                                 int insert_reserved)
2234 {
2235         int ret = 0;
2236         struct btrfs_delayed_tree_ref *ref;
2237         struct btrfs_key ins;
2238         u64 parent = 0;
2239         u64 ref_root = 0;
2240         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2241                                                  SKINNY_METADATA);
2242
2243         ref = btrfs_delayed_node_to_tree_ref(node);
2244         trace_run_delayed_tree_ref(node, ref, node->action);
2245
2246         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2247                 parent = ref->parent;
2248         ref_root = ref->root;
2249
2250         ins.objectid = node->bytenr;
2251         if (skinny_metadata) {
2252                 ins.offset = ref->level;
2253                 ins.type = BTRFS_METADATA_ITEM_KEY;
2254         } else {
2255                 ins.offset = node->num_bytes;
2256                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2257         }
2258
2259         BUG_ON(node->ref_mod != 1);
2260         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2261                 BUG_ON(!extent_op || !extent_op->update_flags);
2262                 ret = alloc_reserved_tree_block(trans, root,
2263                                                 parent, ref_root,
2264                                                 extent_op->flags_to_set,
2265                                                 &extent_op->key,
2266                                                 ref->level, &ins,
2267                                                 node->no_quota);
2268         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2269                 ret = __btrfs_inc_extent_ref(trans, root, node->bytenr,
2270                                              node->num_bytes, parent, ref_root,
2271                                              ref->level, 0, 1, node->no_quota,
2272                                              extent_op);
2273         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2274                 ret = __btrfs_free_extent(trans, root, node->bytenr,
2275                                           node->num_bytes, parent, ref_root,
2276                                           ref->level, 0, 1, extent_op,
2277                                           node->no_quota);
2278         } else {
2279                 BUG();
2280         }
2281         return ret;
2282 }
2283
2284 /* helper function to actually process a single delayed ref entry */
2285 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2286                                struct btrfs_root *root,
2287                                struct btrfs_delayed_ref_node *node,
2288                                struct btrfs_delayed_extent_op *extent_op,
2289                                int insert_reserved)
2290 {
2291         int ret = 0;
2292
2293         if (trans->aborted) {
2294                 if (insert_reserved)
2295                         btrfs_pin_extent(root, node->bytenr,
2296                                          node->num_bytes, 1);
2297                 return 0;
2298         }
2299
2300         if (btrfs_delayed_ref_is_head(node)) {
2301                 struct btrfs_delayed_ref_head *head;
2302                 /*
2303                  * we've hit the end of the chain and we were supposed
2304                  * to insert this extent into the tree.  But, it got
2305                  * deleted before we ever needed to insert it, so all
2306                  * we have to do is clean up the accounting
2307                  */
2308                 BUG_ON(extent_op);
2309                 head = btrfs_delayed_node_to_head(node);
2310                 trace_run_delayed_ref_head(node, head, node->action);
2311
2312                 if (insert_reserved) {
2313                         btrfs_pin_extent(root, node->bytenr,
2314                                          node->num_bytes, 1);
2315                         if (head->is_data) {
2316                                 ret = btrfs_del_csums(trans, root,
2317                                                       node->bytenr,
2318                                                       node->num_bytes);
2319                         }
2320                 }
2321                 return ret;
2322         }
2323
2324         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2325             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2326                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2327                                            insert_reserved);
2328         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2329                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2330                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2331                                            insert_reserved);
2332         else
2333                 BUG();
2334         return ret;
2335 }
2336
2337 static noinline struct btrfs_delayed_ref_node *
2338 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2339 {
2340         struct rb_node *node;
2341         struct btrfs_delayed_ref_node *ref, *last = NULL;;
2342
2343         /*
2344          * select delayed ref of type BTRFS_ADD_DELAYED_REF first.
2345          * this prevents ref count from going down to zero when
2346          * there still are pending delayed ref.
2347          */
2348         node = rb_first(&head->ref_root);
2349         while (node) {
2350                 ref = rb_entry(node, struct btrfs_delayed_ref_node,
2351                                 rb_node);
2352                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2353                         return ref;
2354                 else if (last == NULL)
2355                         last = ref;
2356                 node = rb_next(node);
2357         }
2358         return last;
2359 }
2360
2361 /*
2362  * Returns 0 on success or if called with an already aborted transaction.
2363  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2364  */
2365 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2366                                              struct btrfs_root *root,
2367                                              unsigned long nr)
2368 {
2369         struct btrfs_delayed_ref_root *delayed_refs;
2370         struct btrfs_delayed_ref_node *ref;
2371         struct btrfs_delayed_ref_head *locked_ref = NULL;
2372         struct btrfs_delayed_extent_op *extent_op;
2373         struct btrfs_fs_info *fs_info = root->fs_info;
2374         ktime_t start = ktime_get();
2375         int ret;
2376         unsigned long count = 0;
2377         unsigned long actual_count = 0;
2378         int must_insert_reserved = 0;
2379
2380         delayed_refs = &trans->transaction->delayed_refs;
2381         while (1) {
2382                 if (!locked_ref) {
2383                         if (count >= nr)
2384                                 break;
2385
2386                         spin_lock(&delayed_refs->lock);
2387                         locked_ref = btrfs_select_ref_head(trans);
2388                         if (!locked_ref) {
2389                                 spin_unlock(&delayed_refs->lock);
2390                                 break;
2391                         }
2392
2393                         /* grab the lock that says we are going to process
2394                          * all the refs for this head */
2395                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2396                         spin_unlock(&delayed_refs->lock);
2397                         /*
2398                          * we may have dropped the spin lock to get the head
2399                          * mutex lock, and that might have given someone else
2400                          * time to free the head.  If that's true, it has been
2401                          * removed from our list and we can move on.
2402                          */
2403                         if (ret == -EAGAIN) {
2404                                 locked_ref = NULL;
2405                                 count++;
2406                                 continue;
2407                         }
2408                 }
2409
2410                 /*
2411                  * We need to try and merge add/drops of the same ref since we
2412                  * can run into issues with relocate dropping the implicit ref
2413                  * and then it being added back again before the drop can
2414                  * finish.  If we merged anything we need to re-loop so we can
2415                  * get a good ref.
2416                  */
2417                 spin_lock(&locked_ref->lock);
2418                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2419                                          locked_ref);
2420
2421                 /*
2422                  * locked_ref is the head node, so we have to go one
2423                  * node back for any delayed ref updates
2424                  */
2425                 ref = select_delayed_ref(locked_ref);
2426
2427                 if (ref && ref->seq &&
2428                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2429                         spin_unlock(&locked_ref->lock);
2430                         btrfs_delayed_ref_unlock(locked_ref);
2431                         spin_lock(&delayed_refs->lock);
2432                         locked_ref->processing = 0;
2433                         delayed_refs->num_heads_ready++;
2434                         spin_unlock(&delayed_refs->lock);
2435                         locked_ref = NULL;
2436                         cond_resched();
2437                         count++;
2438                         continue;
2439                 }
2440
2441                 /*
2442                  * record the must insert reserved flag before we
2443                  * drop the spin lock.
2444                  */
2445                 must_insert_reserved = locked_ref->must_insert_reserved;
2446                 locked_ref->must_insert_reserved = 0;
2447
2448                 extent_op = locked_ref->extent_op;
2449                 locked_ref->extent_op = NULL;
2450
2451                 if (!ref) {
2452
2453
2454                         /* All delayed refs have been processed, Go ahead
2455                          * and send the head node to run_one_delayed_ref,
2456                          * so that any accounting fixes can happen
2457                          */
2458                         ref = &locked_ref->node;
2459
2460                         if (extent_op && must_insert_reserved) {
2461                                 btrfs_free_delayed_extent_op(extent_op);
2462                                 extent_op = NULL;
2463                         }
2464
2465                         if (extent_op) {
2466                                 spin_unlock(&locked_ref->lock);
2467                                 ret = run_delayed_extent_op(trans, root,
2468                                                             ref, extent_op);
2469                                 btrfs_free_delayed_extent_op(extent_op);
2470
2471                                 if (ret) {
2472                                         /*
2473                                          * Need to reset must_insert_reserved if
2474                                          * there was an error so the abort stuff
2475                                          * can cleanup the reserved space
2476                                          * properly.
2477                                          */
2478                                         if (must_insert_reserved)
2479                                                 locked_ref->must_insert_reserved = 1;
2480                                         locked_ref->processing = 0;
2481                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2482                                         btrfs_delayed_ref_unlock(locked_ref);
2483                                         return ret;
2484                                 }
2485                                 continue;
2486                         }
2487
2488                         /*
2489                          * Need to drop our head ref lock and re-aqcuire the
2490                          * delayed ref lock and then re-check to make sure
2491                          * nobody got added.
2492                          */
2493                         spin_unlock(&locked_ref->lock);
2494                         spin_lock(&delayed_refs->lock);
2495                         spin_lock(&locked_ref->lock);
2496                         if (rb_first(&locked_ref->ref_root) ||
2497                             locked_ref->extent_op) {
2498                                 spin_unlock(&locked_ref->lock);
2499                                 spin_unlock(&delayed_refs->lock);
2500                                 continue;
2501                         }
2502                         ref->in_tree = 0;
2503                         delayed_refs->num_heads--;
2504                         rb_erase(&locked_ref->href_node,
2505                                  &delayed_refs->href_root);
2506                         spin_unlock(&delayed_refs->lock);
2507                 } else {
2508                         actual_count++;
2509                         ref->in_tree = 0;
2510                         rb_erase(&ref->rb_node, &locked_ref->ref_root);
2511                 }
2512                 atomic_dec(&delayed_refs->num_entries);
2513
2514                 if (!btrfs_delayed_ref_is_head(ref)) {
2515                         /*
2516                          * when we play the delayed ref, also correct the
2517                          * ref_mod on head
2518                          */
2519                         switch (ref->action) {
2520                         case BTRFS_ADD_DELAYED_REF:
2521                         case BTRFS_ADD_DELAYED_EXTENT:
2522                                 locked_ref->node.ref_mod -= ref->ref_mod;
2523                                 break;
2524                         case BTRFS_DROP_DELAYED_REF:
2525                                 locked_ref->node.ref_mod += ref->ref_mod;
2526                                 break;
2527                         default:
2528                                 WARN_ON(1);
2529                         }
2530                 }
2531                 spin_unlock(&locked_ref->lock);
2532
2533                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2534                                           must_insert_reserved);
2535
2536                 btrfs_free_delayed_extent_op(extent_op);
2537                 if (ret) {
2538                         locked_ref->processing = 0;
2539                         btrfs_delayed_ref_unlock(locked_ref);
2540                         btrfs_put_delayed_ref(ref);
2541                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2542                         return ret;
2543                 }
2544
2545                 /*
2546                  * If this node is a head, that means all the refs in this head
2547                  * have been dealt with, and we will pick the next head to deal
2548                  * with, so we must unlock the head and drop it from the cluster
2549                  * list before we release it.
2550                  */
2551                 if (btrfs_delayed_ref_is_head(ref)) {
2552                         btrfs_delayed_ref_unlock(locked_ref);
2553                         locked_ref = NULL;
2554                 }
2555                 btrfs_put_delayed_ref(ref);
2556                 count++;
2557                 cond_resched();
2558         }
2559
2560         /*
2561          * We don't want to include ref heads since we can have empty ref heads
2562          * and those will drastically skew our runtime down since we just do
2563          * accounting, no actual extent tree updates.
2564          */
2565         if (actual_count > 0) {
2566                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2567                 u64 avg;
2568
2569                 /*
2570                  * We weigh the current average higher than our current runtime
2571                  * to avoid large swings in the average.
2572                  */
2573                 spin_lock(&delayed_refs->lock);
2574                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2575                 avg = div64_u64(avg, 4);
2576                 fs_info->avg_delayed_ref_runtime = avg;
2577                 spin_unlock(&delayed_refs->lock);
2578         }
2579         return 0;
2580 }
2581
2582 #ifdef SCRAMBLE_DELAYED_REFS
2583 /*
2584  * Normally delayed refs get processed in ascending bytenr order. This
2585  * correlates in most cases to the order added. To expose dependencies on this
2586  * order, we start to process the tree in the middle instead of the beginning
2587  */
2588 static u64 find_middle(struct rb_root *root)
2589 {
2590         struct rb_node *n = root->rb_node;
2591         struct btrfs_delayed_ref_node *entry;
2592         int alt = 1;
2593         u64 middle;
2594         u64 first = 0, last = 0;
2595
2596         n = rb_first(root);
2597         if (n) {
2598                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2599                 first = entry->bytenr;
2600         }
2601         n = rb_last(root);
2602         if (n) {
2603                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2604                 last = entry->bytenr;
2605         }
2606         n = root->rb_node;
2607
2608         while (n) {
2609                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2610                 WARN_ON(!entry->in_tree);
2611
2612                 middle = entry->bytenr;
2613
2614                 if (alt)
2615                         n = n->rb_left;
2616                 else
2617                         n = n->rb_right;
2618
2619                 alt = 1 - alt;
2620         }
2621         return middle;
2622 }
2623 #endif
2624
2625 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2626 {
2627         u64 num_bytes;
2628
2629         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2630                              sizeof(struct btrfs_extent_inline_ref));
2631         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2632                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2633
2634         /*
2635          * We don't ever fill up leaves all the way so multiply by 2 just to be
2636          * closer to what we're really going to want to ouse.
2637          */
2638         return div64_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2639 }
2640
2641 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2642                                        struct btrfs_root *root)
2643 {
2644         struct btrfs_block_rsv *global_rsv;
2645         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2646         u64 num_bytes;
2647         int ret = 0;
2648
2649         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2650         num_heads = heads_to_leaves(root, num_heads);
2651         if (num_heads > 1)
2652                 num_bytes += (num_heads - 1) * root->leafsize;
2653         num_bytes <<= 1;
2654         global_rsv = &root->fs_info->global_block_rsv;
2655
2656         /*
2657          * If we can't allocate any more chunks lets make sure we have _lots_ of
2658          * wiggle room since running delayed refs can create more delayed refs.
2659          */
2660         if (global_rsv->space_info->full)
2661                 num_bytes <<= 1;
2662
2663         spin_lock(&global_rsv->lock);
2664         if (global_rsv->reserved <= num_bytes)
2665                 ret = 1;
2666         spin_unlock(&global_rsv->lock);
2667         return ret;
2668 }
2669
2670 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2671                                        struct btrfs_root *root)
2672 {
2673         struct btrfs_fs_info *fs_info = root->fs_info;
2674         u64 num_entries =
2675                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2676         u64 avg_runtime;
2677         u64 val;
2678
2679         smp_mb();
2680         avg_runtime = fs_info->avg_delayed_ref_runtime;
2681         val = num_entries * avg_runtime;
2682         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2683                 return 1;
2684         if (val >= NSEC_PER_SEC / 2)
2685                 return 2;
2686
2687         return btrfs_check_space_for_delayed_refs(trans, root);
2688 }
2689
2690 struct async_delayed_refs {
2691         struct btrfs_root *root;
2692         int count;
2693         int error;
2694         int sync;
2695         struct completion wait;
2696         struct btrfs_work work;
2697 };
2698
2699 static void delayed_ref_async_start(struct btrfs_work *work)
2700 {
2701         struct async_delayed_refs *async;
2702         struct btrfs_trans_handle *trans;
2703         int ret;
2704
2705         async = container_of(work, struct async_delayed_refs, work);
2706
2707         trans = btrfs_join_transaction(async->root);
2708         if (IS_ERR(trans)) {
2709                 async->error = PTR_ERR(trans);
2710                 goto done;
2711         }
2712
2713         /*
2714          * trans->sync means that when we call end_transaciton, we won't
2715          * wait on delayed refs
2716          */
2717         trans->sync = true;
2718         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2719         if (ret)
2720                 async->error = ret;
2721
2722         ret = btrfs_end_transaction(trans, async->root);
2723         if (ret && !async->error)
2724                 async->error = ret;
2725 done:
2726         if (async->sync)
2727                 complete(&async->wait);
2728         else
2729                 kfree(async);
2730 }
2731
2732 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2733                                  unsigned long count, int wait)
2734 {
2735         struct async_delayed_refs *async;
2736         int ret;
2737
2738         async = kmalloc(sizeof(*async), GFP_NOFS);
2739         if (!async)
2740                 return -ENOMEM;
2741
2742         async->root = root->fs_info->tree_root;
2743         async->count = count;
2744         async->error = 0;
2745         if (wait)
2746                 async->sync = 1;
2747         else
2748                 async->sync = 0;
2749         init_completion(&async->wait);
2750
2751         btrfs_init_work(&async->work, delayed_ref_async_start,
2752                         NULL, NULL);
2753
2754         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2755
2756         if (wait) {
2757                 wait_for_completion(&async->wait);
2758                 ret = async->error;
2759                 kfree(async);
2760                 return ret;
2761         }
2762         return 0;
2763 }
2764
2765 /*
2766  * this starts processing the delayed reference count updates and
2767  * extent insertions we have queued up so far.  count can be
2768  * 0, which means to process everything in the tree at the start
2769  * of the run (but not newly added entries), or it can be some target
2770  * number you'd like to process.
2771  *
2772  * Returns 0 on success or if called with an aborted transaction
2773  * Returns <0 on error and aborts the transaction
2774  */
2775 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2776                            struct btrfs_root *root, unsigned long count)
2777 {
2778         struct rb_node *node;
2779         struct btrfs_delayed_ref_root *delayed_refs;
2780         struct btrfs_delayed_ref_head *head;
2781         int ret;
2782         int run_all = count == (unsigned long)-1;
2783         int run_most = 0;
2784
2785         /* We'll clean this up in btrfs_cleanup_transaction */
2786         if (trans->aborted)
2787                 return 0;
2788
2789         if (root == root->fs_info->extent_root)
2790                 root = root->fs_info->tree_root;
2791
2792         delayed_refs = &trans->transaction->delayed_refs;
2793         if (count == 0) {
2794                 count = atomic_read(&delayed_refs->num_entries) * 2;
2795                 run_most = 1;
2796         }
2797
2798 again:
2799 #ifdef SCRAMBLE_DELAYED_REFS
2800         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2801 #endif
2802         ret = __btrfs_run_delayed_refs(trans, root, count);
2803         if (ret < 0) {
2804                 btrfs_abort_transaction(trans, root, ret);
2805                 return ret;
2806         }
2807
2808         if (run_all) {
2809                 if (!list_empty(&trans->new_bgs))
2810                         btrfs_create_pending_block_groups(trans, root);
2811
2812                 spin_lock(&delayed_refs->lock);
2813                 node = rb_first(&delayed_refs->href_root);
2814                 if (!node) {
2815                         spin_unlock(&delayed_refs->lock);
2816                         goto out;
2817                 }
2818                 count = (unsigned long)-1;
2819
2820                 while (node) {
2821                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2822                                         href_node);
2823                         if (btrfs_delayed_ref_is_head(&head->node)) {
2824                                 struct btrfs_delayed_ref_node *ref;
2825
2826                                 ref = &head->node;
2827                                 atomic_inc(&ref->refs);
2828
2829                                 spin_unlock(&delayed_refs->lock);
2830                                 /*
2831                                  * Mutex was contended, block until it's
2832                                  * released and try again
2833                                  */
2834                                 mutex_lock(&head->mutex);
2835                                 mutex_unlock(&head->mutex);
2836
2837                                 btrfs_put_delayed_ref(ref);
2838                                 cond_resched();
2839                                 goto again;
2840                         } else {
2841                                 WARN_ON(1);
2842                         }
2843                         node = rb_next(node);
2844                 }
2845                 spin_unlock(&delayed_refs->lock);
2846                 cond_resched();
2847                 goto again;
2848         }
2849 out:
2850         ret = btrfs_delayed_qgroup_accounting(trans, root->fs_info);
2851         if (ret)
2852                 return ret;
2853         assert_qgroups_uptodate(trans);
2854         return 0;
2855 }
2856
2857 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2858                                 struct btrfs_root *root,
2859                                 u64 bytenr, u64 num_bytes, u64 flags,
2860                                 int level, int is_data)
2861 {
2862         struct btrfs_delayed_extent_op *extent_op;
2863         int ret;
2864
2865         extent_op = btrfs_alloc_delayed_extent_op();
2866         if (!extent_op)
2867                 return -ENOMEM;
2868
2869         extent_op->flags_to_set = flags;
2870         extent_op->update_flags = 1;
2871         extent_op->update_key = 0;
2872         extent_op->is_data = is_data ? 1 : 0;
2873         extent_op->level = level;
2874
2875         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
2876                                           num_bytes, extent_op);
2877         if (ret)
2878                 btrfs_free_delayed_extent_op(extent_op);
2879         return ret;
2880 }
2881
2882 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
2883                                       struct btrfs_root *root,
2884                                       struct btrfs_path *path,
2885                                       u64 objectid, u64 offset, u64 bytenr)
2886 {
2887         struct btrfs_delayed_ref_head *head;
2888         struct btrfs_delayed_ref_node *ref;
2889         struct btrfs_delayed_data_ref *data_ref;
2890         struct btrfs_delayed_ref_root *delayed_refs;
2891         struct rb_node *node;
2892         int ret = 0;
2893
2894         delayed_refs = &trans->transaction->delayed_refs;
2895         spin_lock(&delayed_refs->lock);
2896         head = btrfs_find_delayed_ref_head(trans, bytenr);
2897         if (!head) {
2898                 spin_unlock(&delayed_refs->lock);
2899                 return 0;
2900         }
2901
2902         if (!mutex_trylock(&head->mutex)) {
2903                 atomic_inc(&head->node.refs);
2904                 spin_unlock(&delayed_refs->lock);
2905
2906                 btrfs_release_path(path);
2907
2908                 /*
2909                  * Mutex was contended, block until it's released and let
2910                  * caller try again
2911                  */
2912                 mutex_lock(&head->mutex);
2913                 mutex_unlock(&head->mutex);
2914                 btrfs_put_delayed_ref(&head->node);
2915                 return -EAGAIN;
2916         }
2917         spin_unlock(&delayed_refs->lock);
2918
2919         spin_lock(&head->lock);
2920         node = rb_first(&head->ref_root);
2921         while (node) {
2922                 ref = rb_entry(node, struct btrfs_delayed_ref_node, rb_node);
2923                 node = rb_next(node);
2924
2925                 /* If it's a shared ref we know a cross reference exists */
2926                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
2927                         ret = 1;
2928                         break;
2929                 }
2930
2931                 data_ref = btrfs_delayed_node_to_data_ref(ref);
2932
2933                 /*
2934                  * If our ref doesn't match the one we're currently looking at
2935                  * then we have a cross reference.
2936                  */
2937                 if (data_ref->root != root->root_key.objectid ||
2938                     data_ref->objectid != objectid ||
2939                     data_ref->offset != offset) {
2940                         ret = 1;
2941                         break;
2942                 }
2943         }
2944         spin_unlock(&head->lock);
2945         mutex_unlock(&head->mutex);
2946         return ret;
2947 }
2948
2949 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
2950                                         struct btrfs_root *root,
2951                                         struct btrfs_path *path,
2952                                         u64 objectid, u64 offset, u64 bytenr)
2953 {
2954         struct btrfs_root *extent_root = root->fs_info->extent_root;
2955         struct extent_buffer *leaf;
2956         struct btrfs_extent_data_ref *ref;
2957         struct btrfs_extent_inline_ref *iref;
2958         struct btrfs_extent_item *ei;
2959         struct btrfs_key key;
2960         u32 item_size;
2961         int ret;
2962
2963         key.objectid = bytenr;
2964         key.offset = (u64)-1;
2965         key.type = BTRFS_EXTENT_ITEM_KEY;
2966
2967         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
2968         if (ret < 0)
2969                 goto out;
2970         BUG_ON(ret == 0); /* Corruption */
2971
2972         ret = -ENOENT;
2973         if (path->slots[0] == 0)
2974                 goto out;
2975
2976         path->slots[0]--;
2977         leaf = path->nodes[0];
2978         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2979
2980         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
2981                 goto out;
2982
2983         ret = 1;
2984         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2985 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2986         if (item_size < sizeof(*ei)) {
2987                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
2988                 goto out;
2989         }
2990 #endif
2991         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2992
2993         if (item_size != sizeof(*ei) +
2994             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
2995                 goto out;
2996
2997         if (btrfs_extent_generation(leaf, ei) <=
2998             btrfs_root_last_snapshot(&root->root_item))
2999                 goto out;
3000
3001         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3002         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3003             BTRFS_EXTENT_DATA_REF_KEY)
3004                 goto out;
3005
3006         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3007         if (btrfs_extent_refs(leaf, ei) !=
3008             btrfs_extent_data_ref_count(leaf, ref) ||
3009             btrfs_extent_data_ref_root(leaf, ref) !=
3010             root->root_key.objectid ||
3011             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3012             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3013                 goto out;
3014
3015         ret = 0;
3016 out:
3017         return ret;
3018 }
3019
3020 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3021                           struct btrfs_root *root,
3022                           u64 objectid, u64 offset, u64 bytenr)
3023 {
3024         struct btrfs_path *path;
3025         int ret;
3026         int ret2;
3027
3028         path = btrfs_alloc_path();
3029         if (!path)
3030                 return -ENOENT;
3031
3032         do {
3033                 ret = check_committed_ref(trans, root, path, objectid,
3034                                           offset, bytenr);
3035                 if (ret && ret != -ENOENT)
3036                         goto out;
3037
3038                 ret2 = check_delayed_ref(trans, root, path, objectid,
3039                                          offset, bytenr);
3040         } while (ret2 == -EAGAIN);
3041
3042         if (ret2 && ret2 != -ENOENT) {
3043                 ret = ret2;
3044                 goto out;
3045         }
3046
3047         if (ret != -ENOENT || ret2 != -ENOENT)
3048                 ret = 0;
3049 out:
3050         btrfs_free_path(path);
3051         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3052                 WARN_ON(ret > 0);
3053         return ret;
3054 }
3055
3056 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3057                            struct btrfs_root *root,
3058                            struct extent_buffer *buf,
3059                            int full_backref, int inc, int no_quota)
3060 {
3061         u64 bytenr;
3062         u64 num_bytes;
3063         u64 parent;
3064         u64 ref_root;
3065         u32 nritems;
3066         struct btrfs_key key;
3067         struct btrfs_file_extent_item *fi;
3068         int i;
3069         int level;
3070         int ret = 0;
3071         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3072                             u64, u64, u64, u64, u64, u64, int);
3073
3074 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
3075         if (unlikely(test_bit(BTRFS_ROOT_DUMMY_ROOT, &root->state)))
3076                 return 0;
3077 #endif
3078         ref_root = btrfs_header_owner(buf);
3079         nritems = btrfs_header_nritems(buf);
3080         level = btrfs_header_level(buf);
3081
3082         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3083                 return 0;
3084
3085         if (inc)
3086                 process_func = btrfs_inc_extent_ref;
3087         else
3088                 process_func = btrfs_free_extent;
3089
3090         if (full_backref)
3091                 parent = buf->start;
3092         else
3093                 parent = 0;
3094
3095         for (i = 0; i < nritems; i++) {
3096                 if (level == 0) {
3097                         btrfs_item_key_to_cpu(buf, &key, i);
3098                         if (btrfs_key_type(&key) != BTRFS_EXTENT_DATA_KEY)
3099                                 continue;
3100                         fi = btrfs_item_ptr(buf, i,
3101                                             struct btrfs_file_extent_item);
3102                         if (btrfs_file_extent_type(buf, fi) ==
3103                             BTRFS_FILE_EXTENT_INLINE)
3104                                 continue;
3105                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3106                         if (bytenr == 0)
3107                                 continue;
3108
3109                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3110                         key.offset -= btrfs_file_extent_offset(buf, fi);
3111                         ret = process_func(trans, root, bytenr, num_bytes,
3112                                            parent, ref_root, key.objectid,
3113                                            key.offset, no_quota);
3114                         if (ret)
3115                                 goto fail;
3116                 } else {
3117                         bytenr = btrfs_node_blockptr(buf, i);
3118                         num_bytes = btrfs_level_size(root, level - 1);
3119                         ret = process_func(trans, root, bytenr, num_bytes,
3120                                            parent, ref_root, level - 1, 0,
3121                                            no_quota);
3122                         if (ret)
3123                                 goto fail;
3124                 }
3125         }
3126         return 0;
3127 fail:
3128         return ret;
3129 }
3130
3131 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3132                   struct extent_buffer *buf, int full_backref, int no_quota)
3133 {
3134         return __btrfs_mod_ref(trans, root, buf, full_backref, 1, no_quota);
3135 }
3136
3137 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3138                   struct extent_buffer *buf, int full_backref, int no_quota)
3139 {
3140         return __btrfs_mod_ref(trans, root, buf, full_backref, 0, no_quota);
3141 }
3142
3143 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3144                                  struct btrfs_root *root,
3145                                  struct btrfs_path *path,
3146                                  struct btrfs_block_group_cache *cache)
3147 {
3148         int ret;
3149         struct btrfs_root *extent_root = root->fs_info->extent_root;
3150         unsigned long bi;
3151         struct extent_buffer *leaf;
3152
3153         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3154         if (ret < 0)
3155                 goto fail;
3156         BUG_ON(ret); /* Corruption */
3157
3158         leaf = path->nodes[0];
3159         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3160         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3161         btrfs_mark_buffer_dirty(leaf);
3162         btrfs_release_path(path);
3163 fail:
3164         if (ret) {
3165                 btrfs_abort_transaction(trans, root, ret);
3166                 return ret;
3167         }
3168         return 0;
3169
3170 }
3171
3172 static struct btrfs_block_group_cache *
3173 next_block_group(struct btrfs_root *root,
3174                  struct btrfs_block_group_cache *cache)
3175 {
3176         struct rb_node *node;
3177         spin_lock(&root->fs_info->block_group_cache_lock);
3178         node = rb_next(&cache->cache_node);
3179         btrfs_put_block_group(cache);
3180         if (node) {
3181                 cache = rb_entry(node, struct btrfs_block_group_cache,
3182                                  cache_node);
3183                 btrfs_get_block_group(cache);
3184         } else
3185                 cache = NULL;
3186         spin_unlock(&root->fs_info->block_group_cache_lock);
3187         return cache;
3188 }
3189
3190 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3191                             struct btrfs_trans_handle *trans,
3192                             struct btrfs_path *path)
3193 {
3194         struct btrfs_root *root = block_group->fs_info->tree_root;
3195         struct inode *inode = NULL;
3196         u64 alloc_hint = 0;
3197         int dcs = BTRFS_DC_ERROR;
3198         int num_pages = 0;
3199         int retries = 0;
3200         int ret = 0;
3201
3202         /*
3203          * If this block group is smaller than 100 megs don't bother caching the
3204          * block group.
3205          */
3206         if (block_group->key.offset < (100 * 1024 * 1024)) {
3207                 spin_lock(&block_group->lock);
3208                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3209                 spin_unlock(&block_group->lock);
3210                 return 0;
3211         }
3212
3213 again:
3214         inode = lookup_free_space_inode(root, block_group, path);
3215         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3216                 ret = PTR_ERR(inode);
3217                 btrfs_release_path(path);
3218                 goto out;
3219         }
3220
3221         if (IS_ERR(inode)) {
3222                 BUG_ON(retries);
3223                 retries++;
3224
3225                 if (block_group->ro)
3226                         goto out_free;
3227
3228                 ret = create_free_space_inode(root, trans, block_group, path);
3229                 if (ret)
3230                         goto out_free;
3231                 goto again;
3232         }
3233
3234         /* We've already setup this transaction, go ahead and exit */
3235         if (block_group->cache_generation == trans->transid &&
3236             i_size_read(inode)) {
3237                 dcs = BTRFS_DC_SETUP;
3238                 goto out_put;
3239         }
3240
3241         /*
3242          * We want to set the generation to 0, that way if anything goes wrong
3243          * from here on out we know not to trust this cache when we load up next
3244          * time.
3245          */
3246         BTRFS_I(inode)->generation = 0;
3247         ret = btrfs_update_inode(trans, root, inode);
3248         WARN_ON(ret);
3249
3250         if (i_size_read(inode) > 0) {
3251                 ret = btrfs_check_trunc_cache_free_space(root,
3252                                         &root->fs_info->global_block_rsv);
3253                 if (ret)
3254                         goto out_put;
3255
3256                 ret = btrfs_truncate_free_space_cache(root, trans, inode);
3257                 if (ret)
3258                         goto out_put;
3259         }
3260
3261         spin_lock(&block_group->lock);
3262         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3263             !btrfs_test_opt(root, SPACE_CACHE)) {
3264                 /*
3265                  * don't bother trying to write stuff out _if_
3266                  * a) we're not cached,
3267                  * b) we're with nospace_cache mount option.
3268                  */
3269                 dcs = BTRFS_DC_WRITTEN;
3270                 spin_unlock(&block_group->lock);
3271                 goto out_put;
3272         }
3273         spin_unlock(&block_group->lock);
3274
3275         /*
3276          * Try to preallocate enough space based on how big the block group is.
3277          * Keep in mind this has to include any pinned space which could end up
3278          * taking up quite a bit since it's not folded into the other space
3279          * cache.
3280          */
3281         num_pages = (int)div64_u64(block_group->key.offset, 256 * 1024 * 1024);
3282         if (!num_pages)
3283                 num_pages = 1;
3284
3285         num_pages *= 16;
3286         num_pages *= PAGE_CACHE_SIZE;
3287
3288         ret = btrfs_check_data_free_space(inode, num_pages);
3289         if (ret)
3290                 goto out_put;
3291
3292         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3293                                               num_pages, num_pages,
3294                                               &alloc_hint);
3295         if (!ret)
3296                 dcs = BTRFS_DC_SETUP;
3297         btrfs_free_reserved_data_space(inode, num_pages);
3298
3299 out_put:
3300         iput(inode);
3301 out_free:
3302         btrfs_release_path(path);
3303 out:
3304         spin_lock(&block_group->lock);
3305         if (!ret && dcs == BTRFS_DC_SETUP)
3306                 block_group->cache_generation = trans->transid;
3307         block_group->disk_cache_state = dcs;
3308         spin_unlock(&block_group->lock);
3309
3310         return ret;
3311 }
3312
3313 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3314                                    struct btrfs_root *root)
3315 {
3316         struct btrfs_block_group_cache *cache;
3317         int err = 0;
3318         struct btrfs_path *path;
3319         u64 last = 0;
3320
3321         path = btrfs_alloc_path();
3322         if (!path)
3323                 return -ENOMEM;
3324
3325 again:
3326         while (1) {
3327                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3328                 while (cache) {
3329                         if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3330                                 break;
3331                         cache = next_block_group(root, cache);
3332                 }
3333                 if (!cache) {
3334                         if (last == 0)
3335                                 break;
3336                         last = 0;
3337                         continue;
3338                 }
3339                 err = cache_save_setup(cache, trans, path);
3340                 last = cache->key.objectid + cache->key.offset;
3341                 btrfs_put_block_group(cache);
3342         }
3343
3344         while (1) {
3345                 if (last == 0) {
3346                         err = btrfs_run_delayed_refs(trans, root,
3347                                                      (unsigned long)-1);
3348                         if (err) /* File system offline */
3349                                 goto out;
3350                 }
3351
3352                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3353                 while (cache) {
3354                         if (cache->disk_cache_state == BTRFS_DC_CLEAR) {
3355                                 btrfs_put_block_group(cache);
3356                                 goto again;
3357                         }
3358
3359                         if (cache->dirty)
3360                                 break;
3361                         cache = next_block_group(root, cache);
3362                 }
3363                 if (!cache) {
3364                         if (last == 0)
3365                                 break;
3366                         last = 0;
3367                         continue;
3368                 }
3369
3370                 if (cache->disk_cache_state == BTRFS_DC_SETUP)
3371                         cache->disk_cache_state = BTRFS_DC_NEED_WRITE;
3372                 cache->dirty = 0;
3373                 last = cache->key.objectid + cache->key.offset;
3374
3375                 err = write_one_cache_group(trans, root, path, cache);
3376                 btrfs_put_block_group(cache);
3377                 if (err) /* File system offline */
3378                         goto out;
3379         }
3380
3381         while (1) {
3382                 /*
3383                  * I don't think this is needed since we're just marking our
3384                  * preallocated extent as written, but just in case it can't
3385                  * hurt.
3386                  */
3387                 if (last == 0) {
3388                         err = btrfs_run_delayed_refs(trans, root,
3389                                                      (unsigned long)-1);
3390                         if (err) /* File system offline */
3391                                 goto out;
3392                 }
3393
3394                 cache = btrfs_lookup_first_block_group(root->fs_info, last);
3395                 while (cache) {
3396                         /*
3397                          * Really this shouldn't happen, but it could if we
3398                          * couldn't write the entire preallocated extent and
3399                          * splitting the extent resulted in a new block.
3400                          */
3401                         if (cache->dirty) {
3402                                 btrfs_put_block_group(cache);
3403                                 goto again;
3404                         }
3405                         if (cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3406                                 break;
3407                         cache = next_block_group(root, cache);
3408                 }
3409                 if (!cache) {
3410                         if (last == 0)
3411                                 break;
3412                         last = 0;
3413                         continue;
3414                 }
3415
3416                 err = btrfs_write_out_cache(root, trans, cache, path);
3417
3418                 /*
3419                  * If we didn't have an error then the cache state is still
3420                  * NEED_WRITE, so we can set it to WRITTEN.
3421                  */
3422                 if (!err && cache->disk_cache_state == BTRFS_DC_NEED_WRITE)
3423                         cache->disk_cache_state = BTRFS_DC_WRITTEN;
3424                 last = cache->key.objectid + cache->key.offset;
3425                 btrfs_put_block_group(cache);
3426         }
3427 out:
3428
3429         btrfs_free_path(path);
3430         return err;
3431 }
3432
3433 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3434 {
3435         struct btrfs_block_group_cache *block_group;
3436         int readonly = 0;
3437
3438         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3439         if (!block_group || block_group->ro)
3440                 readonly = 1;
3441         if (block_group)
3442                 btrfs_put_block_group(block_group);
3443         return readonly;
3444 }
3445
3446 static const char *alloc_name(u64 flags)
3447 {
3448         switch (flags) {
3449         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3450                 return "mixed";
3451         case BTRFS_BLOCK_GROUP_METADATA:
3452                 return "metadata";
3453         case BTRFS_BLOCK_GROUP_DATA:
3454                 return "data";
3455         case BTRFS_BLOCK_GROUP_SYSTEM:
3456                 return "system";
3457         default:
3458                 WARN_ON(1);
3459                 return "invalid-combination";
3460         };
3461 }
3462
3463 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3464                              u64 total_bytes, u64 bytes_used,
3465                              struct btrfs_space_info **space_info)
3466 {
3467         struct btrfs_space_info *found;
3468         int i;
3469         int factor;
3470         int ret;
3471
3472         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3473                      BTRFS_BLOCK_GROUP_RAID10))
3474                 factor = 2;
3475         else
3476                 factor = 1;
3477
3478         found = __find_space_info(info, flags);
3479         if (found) {
3480                 spin_lock(&found->lock);
3481                 found->total_bytes += total_bytes;
3482                 found->disk_total += total_bytes * factor;
3483                 found->bytes_used += bytes_used;
3484                 found->disk_used += bytes_used * factor;
3485                 found->full = 0;
3486                 spin_unlock(&found->lock);
3487                 *space_info = found;
3488                 return 0;
3489         }
3490         found = kzalloc(sizeof(*found), GFP_NOFS);
3491         if (!found)
3492                 return -ENOMEM;
3493
3494         ret = percpu_counter_init(&found->total_bytes_pinned, 0);
3495         if (ret) {
3496                 kfree(found);
3497                 return ret;
3498         }
3499
3500         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
3501                 INIT_LIST_HEAD(&found->block_groups[i]);
3502                 kobject_init(&found->block_group_kobjs[i], &btrfs_raid_ktype);
3503         }
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  *
5619  * This is called by the allocator when it reserves space, or by somebody who is
5620  * freeing space that was never actually used on disk.  For example if you
5621  * reserve some space for a new leaf in transaction A and before transaction A
5622  * commits you free that leaf, you call this with reserve set to 0 in order to
5623  * clear the reservation.
5624  *
5625  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
5626  * ENOSPC accounting.  For data we handle the reservation through clearing the
5627  * delalloc bits in the io_tree.  We have to do this since we could end up
5628  * allocating less disk space for the amount of data we have reserved in the
5629  * case of compression.
5630  *
5631  * If this is a reservation and the block group has become read only we cannot
5632  * make the reservation and return -EAGAIN, otherwise this function always
5633  * succeeds.
5634  */
5635 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
5636                                        u64 num_bytes, int reserve)
5637 {
5638         struct btrfs_space_info *space_info = cache->space_info;
5639         int ret = 0;
5640
5641         spin_lock(&space_info->lock);
5642         spin_lock(&cache->lock);
5643         if (reserve != RESERVE_FREE) {
5644                 if (cache->ro) {
5645                         ret = -EAGAIN;
5646                 } else {
5647                         cache->reserved += num_bytes;
5648                         space_info->bytes_reserved += num_bytes;
5649                         if (reserve == RESERVE_ALLOC) {
5650                                 trace_btrfs_space_reservation(cache->fs_info,
5651                                                 "space_info", space_info->flags,
5652                                                 num_bytes, 0);
5653                                 space_info->bytes_may_use -= num_bytes;
5654                         }
5655                 }
5656         } else {
5657                 if (cache->ro)
5658                         space_info->bytes_readonly += num_bytes;
5659                 cache->reserved -= num_bytes;
5660                 space_info->bytes_reserved -= num_bytes;
5661         }
5662         spin_unlock(&cache->lock);
5663         spin_unlock(&space_info->lock);
5664         return ret;
5665 }
5666
5667 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
5668                                 struct btrfs_root *root)
5669 {
5670         struct btrfs_fs_info *fs_info = root->fs_info;
5671         struct btrfs_caching_control *next;
5672         struct btrfs_caching_control *caching_ctl;
5673         struct btrfs_block_group_cache *cache;
5674         struct btrfs_space_info *space_info;
5675
5676         down_write(&fs_info->commit_root_sem);
5677
5678         list_for_each_entry_safe(caching_ctl, next,
5679                                  &fs_info->caching_block_groups, list) {
5680                 cache = caching_ctl->block_group;
5681                 if (block_group_cache_done(cache)) {
5682                         cache->last_byte_to_unpin = (u64)-1;
5683                         list_del_init(&caching_ctl->list);
5684                         put_caching_control(caching_ctl);
5685                 } else {
5686                         cache->last_byte_to_unpin = caching_ctl->progress;
5687                 }
5688         }
5689
5690         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5691                 fs_info->pinned_extents = &fs_info->freed_extents[1];
5692         else
5693                 fs_info->pinned_extents = &fs_info->freed_extents[0];
5694
5695         up_write(&fs_info->commit_root_sem);
5696
5697         list_for_each_entry_rcu(space_info, &fs_info->space_info, list)
5698                 percpu_counter_set(&space_info->total_bytes_pinned, 0);
5699
5700         update_global_block_rsv(fs_info);
5701 }
5702
5703 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
5704 {
5705         struct btrfs_fs_info *fs_info = root->fs_info;
5706         struct btrfs_block_group_cache *cache = NULL;
5707         struct btrfs_space_info *space_info;
5708         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5709         u64 len;
5710         bool readonly;
5711
5712         while (start <= end) {
5713                 readonly = false;
5714                 if (!cache ||
5715                     start >= cache->key.objectid + cache->key.offset) {
5716                         if (cache)
5717                                 btrfs_put_block_group(cache);
5718                         cache = btrfs_lookup_block_group(fs_info, start);
5719                         BUG_ON(!cache); /* Logic error */
5720                 }
5721
5722                 len = cache->key.objectid + cache->key.offset - start;
5723                 len = min(len, end + 1 - start);
5724
5725                 if (start < cache->last_byte_to_unpin) {
5726                         len = min(len, cache->last_byte_to_unpin - start);
5727                         btrfs_add_free_space(cache, start, len);
5728                 }
5729
5730                 start += len;
5731                 space_info = cache->space_info;
5732
5733                 spin_lock(&space_info->lock);
5734                 spin_lock(&cache->lock);
5735                 cache->pinned -= len;
5736                 space_info->bytes_pinned -= len;
5737                 if (cache->ro) {
5738                         space_info->bytes_readonly += len;
5739                         readonly = true;
5740                 }
5741                 spin_unlock(&cache->lock);
5742                 if (!readonly && global_rsv->space_info == space_info) {
5743                         spin_lock(&global_rsv->lock);
5744                         if (!global_rsv->full) {
5745                                 len = min(len, global_rsv->size -
5746                                           global_rsv->reserved);
5747                                 global_rsv->reserved += len;
5748                                 space_info->bytes_may_use += len;
5749                                 if (global_rsv->reserved >= global_rsv->size)
5750                                         global_rsv->full = 1;
5751                         }
5752                         spin_unlock(&global_rsv->lock);
5753                 }
5754                 spin_unlock(&space_info->lock);
5755         }
5756
5757         if (cache)
5758                 btrfs_put_block_group(cache);
5759         return 0;
5760 }
5761
5762 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
5763                                struct btrfs_root *root)
5764 {
5765         struct btrfs_fs_info *fs_info = root->fs_info;
5766         struct extent_io_tree *unpin;
5767         u64 start;
5768         u64 end;
5769         int ret;
5770
5771         if (trans->aborted)
5772                 return 0;
5773
5774         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
5775                 unpin = &fs_info->freed_extents[1];
5776         else
5777                 unpin = &fs_info->freed_extents[0];
5778
5779         while (1) {
5780                 ret = find_first_extent_bit(unpin, 0, &start, &end,
5781                                             EXTENT_DIRTY, NULL);
5782                 if (ret)
5783                         break;
5784
5785                 if (btrfs_test_opt(root, DISCARD))
5786                         ret = btrfs_discard_extent(root, start,
5787                                                    end + 1 - start, NULL);
5788
5789                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
5790                 unpin_extent_range(root, start, end);
5791                 cond_resched();
5792         }
5793
5794         return 0;
5795 }
5796
5797 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
5798                              u64 owner, u64 root_objectid)
5799 {
5800         struct btrfs_space_info *space_info;
5801         u64 flags;
5802
5803         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
5804                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
5805                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
5806                 else
5807                         flags = BTRFS_BLOCK_GROUP_METADATA;
5808         } else {
5809                 flags = BTRFS_BLOCK_GROUP_DATA;
5810         }
5811
5812         space_info = __find_space_info(fs_info, flags);
5813         BUG_ON(!space_info); /* Logic bug */
5814         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
5815 }
5816
5817
5818 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
5819                                 struct btrfs_root *root,
5820                                 u64 bytenr, u64 num_bytes, u64 parent,
5821                                 u64 root_objectid, u64 owner_objectid,
5822                                 u64 owner_offset, int refs_to_drop,
5823                                 struct btrfs_delayed_extent_op *extent_op,
5824                                 int no_quota)
5825 {
5826         struct btrfs_key key;
5827         struct btrfs_path *path;
5828         struct btrfs_fs_info *info = root->fs_info;
5829         struct btrfs_root *extent_root = info->extent_root;
5830         struct extent_buffer *leaf;
5831         struct btrfs_extent_item *ei;
5832         struct btrfs_extent_inline_ref *iref;
5833         int ret;
5834         int is_data;
5835         int extent_slot = 0;
5836         int found_extent = 0;
5837         int num_to_del = 1;
5838         u32 item_size;
5839         u64 refs;
5840         int last_ref = 0;
5841         enum btrfs_qgroup_operation_type type = BTRFS_QGROUP_OPER_SUB_EXCL;
5842         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
5843                                                  SKINNY_METADATA);
5844
5845         if (!info->quota_enabled || !is_fstree(root_objectid))
5846                 no_quota = 1;
5847
5848         path = btrfs_alloc_path();
5849         if (!path)
5850                 return -ENOMEM;
5851
5852         path->reada = 1;
5853         path->leave_spinning = 1;
5854
5855         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
5856         BUG_ON(!is_data && refs_to_drop != 1);
5857
5858         if (is_data)
5859                 skinny_metadata = 0;
5860
5861         ret = lookup_extent_backref(trans, extent_root, path, &iref,
5862                                     bytenr, num_bytes, parent,
5863                                     root_objectid, owner_objectid,
5864                                     owner_offset);
5865         if (ret == 0) {
5866                 extent_slot = path->slots[0];
5867                 while (extent_slot >= 0) {
5868                         btrfs_item_key_to_cpu(path->nodes[0], &key,
5869                                               extent_slot);
5870                         if (key.objectid != bytenr)
5871                                 break;
5872                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
5873                             key.offset == num_bytes) {
5874                                 found_extent = 1;
5875                                 break;
5876                         }
5877                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
5878                             key.offset == owner_objectid) {
5879                                 found_extent = 1;
5880                                 break;
5881                         }
5882                         if (path->slots[0] - extent_slot > 5)
5883                                 break;
5884                         extent_slot--;
5885                 }
5886 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5887                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
5888                 if (found_extent && item_size < sizeof(*ei))
5889                         found_extent = 0;
5890 #endif
5891                 if (!found_extent) {
5892                         BUG_ON(iref);
5893                         ret = remove_extent_backref(trans, extent_root, path,
5894                                                     NULL, refs_to_drop,
5895                                                     is_data, &last_ref);
5896                         if (ret) {
5897                                 btrfs_abort_transaction(trans, extent_root, ret);
5898                                 goto out;
5899                         }
5900                         btrfs_release_path(path);
5901                         path->leave_spinning = 1;
5902
5903                         key.objectid = bytenr;
5904                         key.type = BTRFS_EXTENT_ITEM_KEY;
5905                         key.offset = num_bytes;
5906
5907                         if (!is_data && skinny_metadata) {
5908                                 key.type = BTRFS_METADATA_ITEM_KEY;
5909                                 key.offset = owner_objectid;
5910                         }
5911
5912                         ret = btrfs_search_slot(trans, extent_root,
5913                                                 &key, path, -1, 1);
5914                         if (ret > 0 && skinny_metadata && path->slots[0]) {
5915                                 /*
5916                                  * Couldn't find our skinny metadata item,
5917                                  * see if we have ye olde extent item.
5918                                  */
5919                                 path->slots[0]--;
5920                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
5921                                                       path->slots[0]);
5922                                 if (key.objectid == bytenr &&
5923                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
5924                                     key.offset == num_bytes)
5925                                         ret = 0;
5926                         }
5927
5928                         if (ret > 0 && skinny_metadata) {
5929                                 skinny_metadata = false;
5930                                 key.objectid = bytenr;
5931                                 key.type = BTRFS_EXTENT_ITEM_KEY;
5932                                 key.offset = num_bytes;
5933                                 btrfs_release_path(path);
5934                                 ret = btrfs_search_slot(trans, extent_root,
5935                                                         &key, path, -1, 1);
5936                         }
5937
5938                         if (ret) {
5939                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5940                                         ret, bytenr);
5941                                 if (ret > 0)
5942                                         btrfs_print_leaf(extent_root,
5943                                                          path->nodes[0]);
5944                         }
5945                         if (ret < 0) {
5946                                 btrfs_abort_transaction(trans, extent_root, ret);
5947                                 goto out;
5948                         }
5949                         extent_slot = path->slots[0];
5950                 }
5951         } else if (WARN_ON(ret == -ENOENT)) {
5952                 btrfs_print_leaf(extent_root, path->nodes[0]);
5953                 btrfs_err(info,
5954                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
5955                         bytenr, parent, root_objectid, owner_objectid,
5956                         owner_offset);
5957                 btrfs_abort_transaction(trans, extent_root, ret);
5958                 goto out;
5959         } else {
5960                 btrfs_abort_transaction(trans, extent_root, ret);
5961                 goto out;
5962         }
5963
5964         leaf = path->nodes[0];
5965         item_size = btrfs_item_size_nr(leaf, extent_slot);
5966 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
5967         if (item_size < sizeof(*ei)) {
5968                 BUG_ON(found_extent || extent_slot != path->slots[0]);
5969                 ret = convert_extent_item_v0(trans, extent_root, path,
5970                                              owner_objectid, 0);
5971                 if (ret < 0) {
5972                         btrfs_abort_transaction(trans, extent_root, ret);
5973                         goto out;
5974                 }
5975
5976                 btrfs_release_path(path);
5977                 path->leave_spinning = 1;
5978
5979                 key.objectid = bytenr;
5980                 key.type = BTRFS_EXTENT_ITEM_KEY;
5981                 key.offset = num_bytes;
5982
5983                 ret = btrfs_search_slot(trans, extent_root, &key, path,
5984                                         -1, 1);
5985                 if (ret) {
5986                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
5987                                 ret, bytenr);
5988                         btrfs_print_leaf(extent_root, path->nodes[0]);
5989                 }
5990                 if (ret < 0) {
5991                         btrfs_abort_transaction(trans, extent_root, ret);
5992                         goto out;
5993                 }
5994
5995                 extent_slot = path->slots[0];
5996                 leaf = path->nodes[0];
5997                 item_size = btrfs_item_size_nr(leaf, extent_slot);
5998         }
5999 #endif
6000         BUG_ON(item_size < sizeof(*ei));
6001         ei = btrfs_item_ptr(leaf, extent_slot,
6002                             struct btrfs_extent_item);
6003         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6004             key.type == BTRFS_EXTENT_ITEM_KEY) {
6005                 struct btrfs_tree_block_info *bi;
6006                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6007                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6008                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6009         }
6010
6011         refs = btrfs_extent_refs(leaf, ei);
6012         if (refs < refs_to_drop) {
6013                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6014                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6015                 ret = -EINVAL;
6016                 btrfs_abort_transaction(trans, extent_root, ret);
6017                 goto out;
6018         }
6019         refs -= refs_to_drop;
6020
6021         if (refs > 0) {
6022                 type = BTRFS_QGROUP_OPER_SUB_SHARED;
6023                 if (extent_op)
6024                         __run_delayed_extent_op(extent_op, leaf, ei);
6025                 /*
6026                  * In the case of inline back ref, reference count will
6027                  * be updated by remove_extent_backref
6028                  */
6029                 if (iref) {
6030                         BUG_ON(!found_extent);
6031                 } else {
6032                         btrfs_set_extent_refs(leaf, ei, refs);
6033                         btrfs_mark_buffer_dirty(leaf);
6034                 }
6035                 if (found_extent) {
6036                         ret = remove_extent_backref(trans, extent_root, path,
6037                                                     iref, refs_to_drop,
6038                                                     is_data, &last_ref);
6039                         if (ret) {
6040                                 btrfs_abort_transaction(trans, extent_root, ret);
6041                                 goto out;
6042                         }
6043                 }
6044                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6045                                  root_objectid);
6046         } else {
6047                 if (found_extent) {
6048                         BUG_ON(is_data && refs_to_drop !=
6049                                extent_data_ref_count(root, path, iref));
6050                         if (iref) {
6051                                 BUG_ON(path->slots[0] != extent_slot);
6052                         } else {
6053                                 BUG_ON(path->slots[0] != extent_slot + 1);
6054                                 path->slots[0] = extent_slot;
6055                                 num_to_del = 2;
6056                         }
6057                 }
6058
6059                 last_ref = 1;
6060                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6061                                       num_to_del);
6062                 if (ret) {
6063                         btrfs_abort_transaction(trans, extent_root, ret);
6064                         goto out;
6065                 }
6066                 btrfs_release_path(path);
6067
6068                 if (is_data) {
6069                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6070                         if (ret) {
6071                                 btrfs_abort_transaction(trans, extent_root, ret);
6072                                 goto out;
6073                         }
6074                 }
6075
6076                 ret = update_block_group(root, bytenr, num_bytes, 0);
6077                 if (ret) {
6078                         btrfs_abort_transaction(trans, extent_root, ret);
6079                         goto out;
6080                 }
6081         }
6082         btrfs_release_path(path);
6083
6084         /* Deal with the quota accounting */
6085         if (!ret && last_ref && !no_quota) {
6086                 int mod_seq = 0;
6087
6088                 if (owner_objectid >= BTRFS_FIRST_FREE_OBJECTID &&
6089                     type == BTRFS_QGROUP_OPER_SUB_SHARED)
6090                         mod_seq = 1;
6091
6092                 ret = btrfs_qgroup_record_ref(trans, info, root_objectid,
6093                                               bytenr, num_bytes, type,
6094                                               mod_seq);
6095         }
6096 out:
6097         btrfs_free_path(path);
6098         return ret;
6099 }
6100
6101 /*
6102  * when we free an block, it is possible (and likely) that we free the last
6103  * delayed ref for that extent as well.  This searches the delayed ref tree for
6104  * a given extent, and if there are no other delayed refs to be processed, it
6105  * removes it from the tree.
6106  */
6107 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6108                                       struct btrfs_root *root, u64 bytenr)
6109 {
6110         struct btrfs_delayed_ref_head *head;
6111         struct btrfs_delayed_ref_root *delayed_refs;
6112         int ret = 0;
6113
6114         delayed_refs = &trans->transaction->delayed_refs;
6115         spin_lock(&delayed_refs->lock);
6116         head = btrfs_find_delayed_ref_head(trans, bytenr);
6117         if (!head)
6118                 goto out_delayed_unlock;
6119
6120         spin_lock(&head->lock);
6121         if (rb_first(&head->ref_root))
6122                 goto out;
6123
6124         if (head->extent_op) {
6125                 if (!head->must_insert_reserved)
6126                         goto out;
6127                 btrfs_free_delayed_extent_op(head->extent_op);
6128                 head->extent_op = NULL;
6129         }
6130
6131         /*
6132          * waiting for the lock here would deadlock.  If someone else has it
6133          * locked they are already in the process of dropping it anyway
6134          */
6135         if (!mutex_trylock(&head->mutex))
6136                 goto out;
6137
6138         /*
6139          * at this point we have a head with no other entries.  Go
6140          * ahead and process it.
6141          */
6142         head->node.in_tree = 0;
6143         rb_erase(&head->href_node, &delayed_refs->href_root);
6144
6145         atomic_dec(&delayed_refs->num_entries);
6146
6147         /*
6148          * we don't take a ref on the node because we're removing it from the
6149          * tree, so we just steal the ref the tree was holding.
6150          */
6151         delayed_refs->num_heads--;
6152         if (head->processing == 0)
6153                 delayed_refs->num_heads_ready--;
6154         head->processing = 0;
6155         spin_unlock(&head->lock);
6156         spin_unlock(&delayed_refs->lock);
6157
6158         BUG_ON(head->extent_op);
6159         if (head->must_insert_reserved)
6160                 ret = 1;
6161
6162         mutex_unlock(&head->mutex);
6163         btrfs_put_delayed_ref(&head->node);
6164         return ret;
6165 out:
6166         spin_unlock(&head->lock);
6167
6168 out_delayed_unlock:
6169         spin_unlock(&delayed_refs->lock);
6170         return 0;
6171 }
6172
6173 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6174                            struct btrfs_root *root,
6175                            struct extent_buffer *buf,
6176                            u64 parent, int last_ref)
6177 {
6178         struct btrfs_block_group_cache *cache = NULL;
6179         int pin = 1;
6180         int ret;
6181
6182         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6183                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6184                                         buf->start, buf->len,
6185                                         parent, root->root_key.objectid,
6186                                         btrfs_header_level(buf),
6187                                         BTRFS_DROP_DELAYED_REF, NULL, 0);
6188                 BUG_ON(ret); /* -ENOMEM */
6189         }
6190
6191         if (!last_ref)
6192                 return;
6193
6194         cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6195
6196         if (btrfs_header_generation(buf) == trans->transid) {
6197                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6198                         ret = check_ref_cleanup(trans, root, buf->start);
6199                         if (!ret)
6200                                 goto out;
6201                 }
6202
6203                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6204                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6205                         goto out;
6206                 }
6207
6208                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6209
6210                 btrfs_add_free_space(cache, buf->start, buf->len);
6211                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE);
6212                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6213                 pin = 0;
6214         }
6215 out:
6216         if (pin)
6217                 add_pinned_bytes(root->fs_info, buf->len,
6218                                  btrfs_header_level(buf),
6219                                  root->root_key.objectid);
6220
6221         /*
6222          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6223          * anymore.
6224          */
6225         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6226         btrfs_put_block_group(cache);
6227 }
6228
6229 /* Can return -ENOMEM */
6230 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6231                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6232                       u64 owner, u64 offset, int no_quota)
6233 {
6234         int ret;
6235         struct btrfs_fs_info *fs_info = root->fs_info;
6236
6237 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
6238         if (unlikely(test_bit(BTRFS_ROOT_DUMMY_ROOT, &root->state)))
6239                 return 0;
6240 #endif
6241         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6242
6243         /*
6244          * tree log blocks never actually go into the extent allocation
6245          * tree, just update pinning info and exit early.
6246          */
6247         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6248                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6249                 /* unlocks the pinned mutex */
6250                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6251                 ret = 0;
6252         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6253                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6254                                         num_bytes,
6255                                         parent, root_objectid, (int)owner,
6256                                         BTRFS_DROP_DELAYED_REF, NULL, no_quota);
6257         } else {
6258                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6259                                                 num_bytes,
6260                                                 parent, root_objectid, owner,
6261                                                 offset, BTRFS_DROP_DELAYED_REF,
6262                                                 NULL, no_quota);
6263         }
6264         return ret;
6265 }
6266
6267 static u64 stripe_align(struct btrfs_root *root,
6268                         struct btrfs_block_group_cache *cache,
6269                         u64 val, u64 num_bytes)
6270 {
6271         u64 ret = ALIGN(val, root->stripesize);
6272         return ret;
6273 }
6274
6275 /*
6276  * when we wait for progress in the block group caching, its because
6277  * our allocation attempt failed at least once.  So, we must sleep
6278  * and let some progress happen before we try again.
6279  *
6280  * This function will sleep at least once waiting for new free space to
6281  * show up, and then it will check the block group free space numbers
6282  * for our min num_bytes.  Another option is to have it go ahead
6283  * and look in the rbtree for a free extent of a given size, but this
6284  * is a good start.
6285  *
6286  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6287  * any of the information in this block group.
6288  */
6289 static noinline void
6290 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6291                                 u64 num_bytes)
6292 {
6293         struct btrfs_caching_control *caching_ctl;
6294
6295         caching_ctl = get_caching_control(cache);
6296         if (!caching_ctl)
6297                 return;
6298
6299         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6300                    (cache->free_space_ctl->free_space >= num_bytes));
6301
6302         put_caching_control(caching_ctl);
6303 }
6304
6305 static noinline int
6306 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6307 {
6308         struct btrfs_caching_control *caching_ctl;
6309         int ret = 0;
6310
6311         caching_ctl = get_caching_control(cache);
6312         if (!caching_ctl)
6313                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6314
6315         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6316         if (cache->cached == BTRFS_CACHE_ERROR)
6317                 ret = -EIO;
6318         put_caching_control(caching_ctl);
6319         return ret;
6320 }
6321
6322 int __get_raid_index(u64 flags)
6323 {
6324         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6325                 return BTRFS_RAID_RAID10;
6326         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6327                 return BTRFS_RAID_RAID1;
6328         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6329                 return BTRFS_RAID_DUP;
6330         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6331                 return BTRFS_RAID_RAID0;
6332         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6333                 return BTRFS_RAID_RAID5;
6334         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6335                 return BTRFS_RAID_RAID6;
6336
6337         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6338 }
6339
6340 int get_block_group_index(struct btrfs_block_group_cache *cache)
6341 {
6342         return __get_raid_index(cache->flags);
6343 }
6344
6345 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6346         [BTRFS_RAID_RAID10]     = "raid10",
6347         [BTRFS_RAID_RAID1]      = "raid1",
6348         [BTRFS_RAID_DUP]        = "dup",
6349         [BTRFS_RAID_RAID0]      = "raid0",
6350         [BTRFS_RAID_SINGLE]     = "single",
6351         [BTRFS_RAID_RAID5]      = "raid5",
6352         [BTRFS_RAID_RAID6]      = "raid6",
6353 };
6354
6355 static const char *get_raid_name(enum btrfs_raid_types type)
6356 {
6357         if (type >= BTRFS_NR_RAID_TYPES)
6358                 return NULL;
6359
6360         return btrfs_raid_type_names[type];
6361 }
6362
6363 enum btrfs_loop_type {
6364         LOOP_CACHING_NOWAIT = 0,
6365         LOOP_CACHING_WAIT = 1,
6366         LOOP_ALLOC_CHUNK = 2,
6367         LOOP_NO_EMPTY_SIZE = 3,
6368 };
6369
6370 /*
6371  * walks the btree of allocated extents and find a hole of a given size.
6372  * The key ins is changed to record the hole:
6373  * ins->objectid == start position
6374  * ins->flags = BTRFS_EXTENT_ITEM_KEY
6375  * ins->offset == the size of the hole.
6376  * Any available blocks before search_start are skipped.
6377  *
6378  * If there is no suitable free space, we will record the max size of
6379  * the free space extent currently.
6380  */
6381 static noinline int find_free_extent(struct btrfs_root *orig_root,
6382                                      u64 num_bytes, u64 empty_size,
6383                                      u64 hint_byte, struct btrfs_key *ins,
6384                                      u64 flags)
6385 {
6386         int ret = 0;
6387         struct btrfs_root *root = orig_root->fs_info->extent_root;
6388         struct btrfs_free_cluster *last_ptr = NULL;
6389         struct btrfs_block_group_cache *block_group = NULL;
6390         u64 search_start = 0;
6391         u64 max_extent_size = 0;
6392         int empty_cluster = 2 * 1024 * 1024;
6393         struct btrfs_space_info *space_info;
6394         int loop = 0;
6395         int index = __get_raid_index(flags);
6396         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
6397                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
6398         bool failed_cluster_refill = false;
6399         bool failed_alloc = false;
6400         bool use_cluster = true;
6401         bool have_caching_bg = false;
6402
6403         WARN_ON(num_bytes < root->sectorsize);
6404         btrfs_set_key_type(ins, BTRFS_EXTENT_ITEM_KEY);
6405         ins->objectid = 0;
6406         ins->offset = 0;
6407
6408         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
6409
6410         space_info = __find_space_info(root->fs_info, flags);
6411         if (!space_info) {
6412                 btrfs_err(root->fs_info, "No space info for %llu", flags);
6413                 return -ENOSPC;
6414         }
6415
6416         /*
6417          * If the space info is for both data and metadata it means we have a
6418          * small filesystem and we can't use the clustering stuff.
6419          */
6420         if (btrfs_mixed_space_info(space_info))
6421                 use_cluster = false;
6422
6423         if (flags & BTRFS_BLOCK_GROUP_METADATA && use_cluster) {
6424                 last_ptr = &root->fs_info->meta_alloc_cluster;
6425                 if (!btrfs_test_opt(root, SSD))
6426                         empty_cluster = 64 * 1024;
6427         }
6428
6429         if ((flags & BTRFS_BLOCK_GROUP_DATA) && use_cluster &&
6430             btrfs_test_opt(root, SSD)) {
6431                 last_ptr = &root->fs_info->data_alloc_cluster;
6432         }
6433
6434         if (last_ptr) {
6435                 spin_lock(&last_ptr->lock);
6436                 if (last_ptr->block_group)
6437                         hint_byte = last_ptr->window_start;
6438                 spin_unlock(&last_ptr->lock);
6439         }
6440
6441         search_start = max(search_start, first_logical_byte(root, 0));
6442         search_start = max(search_start, hint_byte);
6443
6444         if (!last_ptr)
6445                 empty_cluster = 0;
6446
6447         if (search_start == hint_byte) {
6448                 block_group = btrfs_lookup_block_group(root->fs_info,
6449                                                        search_start);
6450                 /*
6451                  * we don't want to use the block group if it doesn't match our
6452                  * allocation bits, or if its not cached.
6453                  *
6454                  * However if we are re-searching with an ideal block group
6455                  * picked out then we don't care that the block group is cached.
6456                  */
6457                 if (block_group && block_group_bits(block_group, flags) &&
6458                     block_group->cached != BTRFS_CACHE_NO) {
6459                         down_read(&space_info->groups_sem);
6460                         if (list_empty(&block_group->list) ||
6461                             block_group->ro) {
6462                                 /*
6463                                  * someone is removing this block group,
6464                                  * we can't jump into the have_block_group
6465                                  * target because our list pointers are not
6466                                  * valid
6467                                  */
6468                                 btrfs_put_block_group(block_group);
6469                                 up_read(&space_info->groups_sem);
6470                         } else {
6471                                 index = get_block_group_index(block_group);
6472                                 goto have_block_group;
6473                         }
6474                 } else if (block_group) {
6475                         btrfs_put_block_group(block_group);
6476                 }
6477         }
6478 search:
6479         have_caching_bg = false;
6480         down_read(&space_info->groups_sem);
6481         list_for_each_entry(block_group, &space_info->block_groups[index],
6482                             list) {
6483                 u64 offset;
6484                 int cached;
6485
6486                 btrfs_get_block_group(block_group);
6487                 search_start = block_group->key.objectid;
6488
6489                 /*
6490                  * this can happen if we end up cycling through all the
6491                  * raid types, but we want to make sure we only allocate
6492                  * for the proper type.
6493                  */
6494                 if (!block_group_bits(block_group, flags)) {
6495                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
6496                                 BTRFS_BLOCK_GROUP_RAID1 |
6497                                 BTRFS_BLOCK_GROUP_RAID5 |
6498                                 BTRFS_BLOCK_GROUP_RAID6 |
6499                                 BTRFS_BLOCK_GROUP_RAID10;
6500
6501                         /*
6502                          * if they asked for extra copies and this block group
6503                          * doesn't provide them, bail.  This does allow us to
6504                          * fill raid0 from raid1.
6505                          */
6506                         if ((flags & extra) && !(block_group->flags & extra))
6507                                 goto loop;
6508                 }
6509
6510 have_block_group:
6511                 cached = block_group_cache_done(block_group);
6512                 if (unlikely(!cached)) {
6513                         ret = cache_block_group(block_group, 0);
6514                         BUG_ON(ret < 0);
6515                         ret = 0;
6516                 }
6517
6518                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
6519                         goto loop;
6520                 if (unlikely(block_group->ro))
6521                         goto loop;
6522
6523                 /*
6524                  * Ok we want to try and use the cluster allocator, so
6525                  * lets look there
6526                  */
6527                 if (last_ptr) {
6528                         struct btrfs_block_group_cache *used_block_group;
6529                         unsigned long aligned_cluster;
6530                         /*
6531                          * the refill lock keeps out other
6532                          * people trying to start a new cluster
6533                          */
6534                         spin_lock(&last_ptr->refill_lock);
6535                         used_block_group = last_ptr->block_group;
6536                         if (used_block_group != block_group &&
6537                             (!used_block_group ||
6538                              used_block_group->ro ||
6539                              !block_group_bits(used_block_group, flags)))
6540                                 goto refill_cluster;
6541
6542                         if (used_block_group != block_group)
6543                                 btrfs_get_block_group(used_block_group);
6544
6545                         offset = btrfs_alloc_from_cluster(used_block_group,
6546                                                 last_ptr,
6547                                                 num_bytes,
6548                                                 used_block_group->key.objectid,
6549                                                 &max_extent_size);
6550                         if (offset) {
6551                                 /* we have a block, we're done */
6552                                 spin_unlock(&last_ptr->refill_lock);
6553                                 trace_btrfs_reserve_extent_cluster(root,
6554                                                 used_block_group,
6555                                                 search_start, num_bytes);
6556                                 if (used_block_group != block_group) {
6557                                         btrfs_put_block_group(block_group);
6558                                         block_group = used_block_group;
6559                                 }
6560                                 goto checks;
6561                         }
6562
6563                         WARN_ON(last_ptr->block_group != used_block_group);
6564                         if (used_block_group != block_group)
6565                                 btrfs_put_block_group(used_block_group);
6566 refill_cluster:
6567                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
6568                          * set up a new clusters, so lets just skip it
6569                          * and let the allocator find whatever block
6570                          * it can find.  If we reach this point, we
6571                          * will have tried the cluster allocator
6572                          * plenty of times and not have found
6573                          * anything, so we are likely way too
6574                          * fragmented for the clustering stuff to find
6575                          * anything.
6576                          *
6577                          * However, if the cluster is taken from the
6578                          * current block group, release the cluster
6579                          * first, so that we stand a better chance of
6580                          * succeeding in the unclustered
6581                          * allocation.  */
6582                         if (loop >= LOOP_NO_EMPTY_SIZE &&
6583                             last_ptr->block_group != block_group) {
6584                                 spin_unlock(&last_ptr->refill_lock);
6585                                 goto unclustered_alloc;
6586                         }
6587
6588                         /*
6589                          * this cluster didn't work out, free it and
6590                          * start over
6591                          */
6592                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6593
6594                         if (loop >= LOOP_NO_EMPTY_SIZE) {
6595                                 spin_unlock(&last_ptr->refill_lock);
6596                                 goto unclustered_alloc;
6597                         }
6598
6599                         aligned_cluster = max_t(unsigned long,
6600                                                 empty_cluster + empty_size,
6601                                               block_group->full_stripe_len);
6602
6603                         /* allocate a cluster in this block group */
6604                         ret = btrfs_find_space_cluster(root, block_group,
6605                                                        last_ptr, search_start,
6606                                                        num_bytes,
6607                                                        aligned_cluster);
6608                         if (ret == 0) {
6609                                 /*
6610                                  * now pull our allocation out of this
6611                                  * cluster
6612                                  */
6613                                 offset = btrfs_alloc_from_cluster(block_group,
6614                                                         last_ptr,
6615                                                         num_bytes,
6616                                                         search_start,
6617                                                         &max_extent_size);
6618                                 if (offset) {
6619                                         /* we found one, proceed */
6620                                         spin_unlock(&last_ptr->refill_lock);
6621                                         trace_btrfs_reserve_extent_cluster(root,
6622                                                 block_group, search_start,
6623                                                 num_bytes);
6624                                         goto checks;
6625                                 }
6626                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
6627                                    && !failed_cluster_refill) {
6628                                 spin_unlock(&last_ptr->refill_lock);
6629
6630                                 failed_cluster_refill = true;
6631                                 wait_block_group_cache_progress(block_group,
6632                                        num_bytes + empty_cluster + empty_size);
6633                                 goto have_block_group;
6634                         }
6635
6636                         /*
6637                          * at this point we either didn't find a cluster
6638                          * or we weren't able to allocate a block from our
6639                          * cluster.  Free the cluster we've been trying
6640                          * to use, and go to the next block group
6641                          */
6642                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
6643                         spin_unlock(&last_ptr->refill_lock);
6644                         goto loop;
6645                 }
6646
6647 unclustered_alloc:
6648                 spin_lock(&block_group->free_space_ctl->tree_lock);
6649                 if (cached &&
6650                     block_group->free_space_ctl->free_space <
6651                     num_bytes + empty_cluster + empty_size) {
6652                         if (block_group->free_space_ctl->free_space >
6653                             max_extent_size)
6654                                 max_extent_size =
6655                                         block_group->free_space_ctl->free_space;
6656                         spin_unlock(&block_group->free_space_ctl->tree_lock);
6657                         goto loop;
6658                 }
6659                 spin_unlock(&block_group->free_space_ctl->tree_lock);
6660
6661                 offset = btrfs_find_space_for_alloc(block_group, search_start,
6662                                                     num_bytes, empty_size,
6663                                                     &max_extent_size);
6664                 /*
6665                  * If we didn't find a chunk, and we haven't failed on this
6666                  * block group before, and this block group is in the middle of
6667                  * caching and we are ok with waiting, then go ahead and wait
6668                  * for progress to be made, and set failed_alloc to true.
6669                  *
6670                  * If failed_alloc is true then we've already waited on this
6671                  * block group once and should move on to the next block group.
6672                  */
6673                 if (!offset && !failed_alloc && !cached &&
6674                     loop > LOOP_CACHING_NOWAIT) {
6675                         wait_block_group_cache_progress(block_group,
6676                                                 num_bytes + empty_size);
6677                         failed_alloc = true;
6678                         goto have_block_group;
6679                 } else if (!offset) {
6680                         if (!cached)
6681                                 have_caching_bg = true;
6682                         goto loop;
6683                 }
6684 checks:
6685                 search_start = stripe_align(root, block_group,
6686                                             offset, num_bytes);
6687
6688                 /* move on to the next group */
6689                 if (search_start + num_bytes >
6690                     block_group->key.objectid + block_group->key.offset) {
6691                         btrfs_add_free_space(block_group, offset, num_bytes);
6692                         goto loop;
6693                 }
6694
6695                 if (offset < search_start)
6696                         btrfs_add_free_space(block_group, offset,
6697                                              search_start - offset);
6698                 BUG_ON(offset > search_start);
6699
6700                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
6701                                                   alloc_type);
6702                 if (ret == -EAGAIN) {
6703                         btrfs_add_free_space(block_group, offset, num_bytes);
6704                         goto loop;
6705                 }
6706
6707                 /* we are all good, lets return */
6708                 ins->objectid = search_start;
6709                 ins->offset = num_bytes;
6710
6711                 trace_btrfs_reserve_extent(orig_root, block_group,
6712                                            search_start, num_bytes);
6713                 btrfs_put_block_group(block_group);
6714                 break;
6715 loop:
6716                 failed_cluster_refill = false;
6717                 failed_alloc = false;
6718                 BUG_ON(index != get_block_group_index(block_group));
6719                 btrfs_put_block_group(block_group);
6720         }
6721         up_read(&space_info->groups_sem);
6722
6723         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
6724                 goto search;
6725
6726         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
6727                 goto search;
6728
6729         /*
6730          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
6731          *                      caching kthreads as we move along
6732          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
6733          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
6734          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
6735          *                      again
6736          */
6737         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
6738                 index = 0;
6739                 loop++;
6740                 if (loop == LOOP_ALLOC_CHUNK) {
6741                         struct btrfs_trans_handle *trans;
6742                         int exist = 0;
6743
6744                         trans = current->journal_info;
6745                         if (trans)
6746                                 exist = 1;
6747                         else
6748                                 trans = btrfs_join_transaction(root);
6749
6750                         if (IS_ERR(trans)) {
6751                                 ret = PTR_ERR(trans);
6752                                 goto out;
6753                         }
6754
6755                         ret = do_chunk_alloc(trans, root, flags,
6756                                              CHUNK_ALLOC_FORCE);
6757                         /*
6758                          * Do not bail out on ENOSPC since we
6759                          * can do more things.
6760                          */
6761                         if (ret < 0 && ret != -ENOSPC)
6762                                 btrfs_abort_transaction(trans,
6763                                                         root, ret);
6764                         else
6765                                 ret = 0;
6766                         if (!exist)
6767                                 btrfs_end_transaction(trans, root);
6768                         if (ret)
6769                                 goto out;
6770                 }
6771
6772                 if (loop == LOOP_NO_EMPTY_SIZE) {
6773                         empty_size = 0;
6774                         empty_cluster = 0;
6775                 }
6776
6777                 goto search;
6778         } else if (!ins->objectid) {
6779                 ret = -ENOSPC;
6780         } else if (ins->objectid) {
6781                 ret = 0;
6782         }
6783 out:
6784         if (ret == -ENOSPC)
6785                 ins->offset = max_extent_size;
6786         return ret;
6787 }
6788
6789 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
6790                             int dump_block_groups)
6791 {
6792         struct btrfs_block_group_cache *cache;
6793         int index = 0;
6794
6795         spin_lock(&info->lock);
6796         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
6797                info->flags,
6798                info->total_bytes - info->bytes_used - info->bytes_pinned -
6799                info->bytes_reserved - info->bytes_readonly,
6800                (info->full) ? "" : "not ");
6801         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
6802                "reserved=%llu, may_use=%llu, readonly=%llu\n",
6803                info->total_bytes, info->bytes_used, info->bytes_pinned,
6804                info->bytes_reserved, info->bytes_may_use,
6805                info->bytes_readonly);
6806         spin_unlock(&info->lock);
6807
6808         if (!dump_block_groups)
6809                 return;
6810
6811         down_read(&info->groups_sem);
6812 again:
6813         list_for_each_entry(cache, &info->block_groups[index], list) {
6814                 spin_lock(&cache->lock);
6815                 printk(KERN_INFO "BTRFS: "
6816                            "block group %llu has %llu bytes, "
6817                            "%llu used %llu pinned %llu reserved %s\n",
6818                        cache->key.objectid, cache->key.offset,
6819                        btrfs_block_group_used(&cache->item), cache->pinned,
6820                        cache->reserved, cache->ro ? "[readonly]" : "");
6821                 btrfs_dump_free_space(cache, bytes);
6822                 spin_unlock(&cache->lock);
6823         }
6824         if (++index < BTRFS_NR_RAID_TYPES)
6825                 goto again;
6826         up_read(&info->groups_sem);
6827 }
6828
6829 int btrfs_reserve_extent(struct btrfs_root *root,
6830                          u64 num_bytes, u64 min_alloc_size,
6831                          u64 empty_size, u64 hint_byte,
6832                          struct btrfs_key *ins, int is_data)
6833 {
6834         bool final_tried = false;
6835         u64 flags;
6836         int ret;
6837
6838         flags = btrfs_get_alloc_profile(root, is_data);
6839 again:
6840         WARN_ON(num_bytes < root->sectorsize);
6841         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
6842                                flags);
6843
6844         if (ret == -ENOSPC) {
6845                 if (!final_tried && ins->offset) {
6846                         num_bytes = min(num_bytes >> 1, ins->offset);
6847                         num_bytes = round_down(num_bytes, root->sectorsize);
6848                         num_bytes = max(num_bytes, min_alloc_size);
6849                         if (num_bytes == min_alloc_size)
6850                                 final_tried = true;
6851                         goto again;
6852                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
6853                         struct btrfs_space_info *sinfo;
6854
6855                         sinfo = __find_space_info(root->fs_info, flags);
6856                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
6857                                 flags, num_bytes);
6858                         if (sinfo)
6859                                 dump_space_info(sinfo, num_bytes, 1);
6860                 }
6861         }
6862
6863         return ret;
6864 }
6865
6866 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
6867                                         u64 start, u64 len, int pin)
6868 {
6869         struct btrfs_block_group_cache *cache;
6870         int ret = 0;
6871
6872         cache = btrfs_lookup_block_group(root->fs_info, start);
6873         if (!cache) {
6874                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
6875                         start);
6876                 return -ENOSPC;
6877         }
6878
6879         if (btrfs_test_opt(root, DISCARD))
6880                 ret = btrfs_discard_extent(root, start, len, NULL);
6881
6882         if (pin)
6883                 pin_down_extent(root, cache, start, len, 1);
6884         else {
6885                 btrfs_add_free_space(cache, start, len);
6886                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE);
6887         }
6888         btrfs_put_block_group(cache);
6889
6890         trace_btrfs_reserved_extent_free(root, start, len);
6891
6892         return ret;
6893 }
6894
6895 int btrfs_free_reserved_extent(struct btrfs_root *root,
6896                                         u64 start, u64 len)
6897 {
6898         return __btrfs_free_reserved_extent(root, start, len, 0);
6899 }
6900
6901 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
6902                                        u64 start, u64 len)
6903 {
6904         return __btrfs_free_reserved_extent(root, start, len, 1);
6905 }
6906
6907 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
6908                                       struct btrfs_root *root,
6909                                       u64 parent, u64 root_objectid,
6910                                       u64 flags, u64 owner, u64 offset,
6911                                       struct btrfs_key *ins, int ref_mod)
6912 {
6913         int ret;
6914         struct btrfs_fs_info *fs_info = root->fs_info;
6915         struct btrfs_extent_item *extent_item;
6916         struct btrfs_extent_inline_ref *iref;
6917         struct btrfs_path *path;
6918         struct extent_buffer *leaf;
6919         int type;
6920         u32 size;
6921
6922         if (parent > 0)
6923                 type = BTRFS_SHARED_DATA_REF_KEY;
6924         else
6925                 type = BTRFS_EXTENT_DATA_REF_KEY;
6926
6927         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
6928
6929         path = btrfs_alloc_path();
6930         if (!path)
6931                 return -ENOMEM;
6932
6933         path->leave_spinning = 1;
6934         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
6935                                       ins, size);
6936         if (ret) {
6937                 btrfs_free_path(path);
6938                 return ret;
6939         }
6940
6941         leaf = path->nodes[0];
6942         extent_item = btrfs_item_ptr(leaf, path->slots[0],
6943                                      struct btrfs_extent_item);
6944         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
6945         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
6946         btrfs_set_extent_flags(leaf, extent_item,
6947                                flags | BTRFS_EXTENT_FLAG_DATA);
6948
6949         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
6950         btrfs_set_extent_inline_ref_type(leaf, iref, type);
6951         if (parent > 0) {
6952                 struct btrfs_shared_data_ref *ref;
6953                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
6954                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
6955                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
6956         } else {
6957                 struct btrfs_extent_data_ref *ref;
6958                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
6959                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
6960                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
6961                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
6962                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
6963         }
6964
6965         btrfs_mark_buffer_dirty(path->nodes[0]);
6966         btrfs_free_path(path);
6967
6968         /* Always set parent to 0 here since its exclusive anyway. */
6969         ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
6970                                       ins->objectid, ins->offset,
6971                                       BTRFS_QGROUP_OPER_ADD_EXCL, 0);
6972         if (ret)
6973                 return ret;
6974
6975         ret = update_block_group(root, ins->objectid, ins->offset, 1);
6976         if (ret) { /* -ENOENT, logic error */
6977                 btrfs_err(fs_info, "update block group failed for %llu %llu",
6978                         ins->objectid, ins->offset);
6979                 BUG();
6980         }
6981         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
6982         return ret;
6983 }
6984
6985 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
6986                                      struct btrfs_root *root,
6987                                      u64 parent, u64 root_objectid,
6988                                      u64 flags, struct btrfs_disk_key *key,
6989                                      int level, struct btrfs_key *ins,
6990                                      int no_quota)
6991 {
6992         int ret;
6993         struct btrfs_fs_info *fs_info = root->fs_info;
6994         struct btrfs_extent_item *extent_item;
6995         struct btrfs_tree_block_info *block_info;
6996         struct btrfs_extent_inline_ref *iref;
6997         struct btrfs_path *path;
6998         struct extent_buffer *leaf;
6999         u32 size = sizeof(*extent_item) + sizeof(*iref);
7000         u64 num_bytes = ins->offset;
7001         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7002                                                  SKINNY_METADATA);
7003
7004         if (!skinny_metadata)
7005                 size += sizeof(*block_info);
7006
7007         path = btrfs_alloc_path();
7008         if (!path) {
7009                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7010                                                    root->leafsize);
7011                 return -ENOMEM;
7012         }
7013
7014         path->leave_spinning = 1;
7015         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7016                                       ins, size);
7017         if (ret) {
7018                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7019                                                    root->leafsize);
7020                 btrfs_free_path(path);
7021                 return ret;
7022         }
7023
7024         leaf = path->nodes[0];
7025         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7026                                      struct btrfs_extent_item);
7027         btrfs_set_extent_refs(leaf, extent_item, 1);
7028         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7029         btrfs_set_extent_flags(leaf, extent_item,
7030                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7031
7032         if (skinny_metadata) {
7033                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7034                 num_bytes = root->leafsize;
7035         } else {
7036                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7037                 btrfs_set_tree_block_key(leaf, block_info, key);
7038                 btrfs_set_tree_block_level(leaf, block_info, level);
7039                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7040         }
7041
7042         if (parent > 0) {
7043                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7044                 btrfs_set_extent_inline_ref_type(leaf, iref,
7045                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7046                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7047         } else {
7048                 btrfs_set_extent_inline_ref_type(leaf, iref,
7049                                                  BTRFS_TREE_BLOCK_REF_KEY);
7050                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7051         }
7052
7053         btrfs_mark_buffer_dirty(leaf);
7054         btrfs_free_path(path);
7055
7056         if (!no_quota) {
7057                 ret = btrfs_qgroup_record_ref(trans, fs_info, root_objectid,
7058                                               ins->objectid, num_bytes,
7059                                               BTRFS_QGROUP_OPER_ADD_EXCL, 0);
7060                 if (ret)
7061                         return ret;
7062         }
7063
7064         ret = update_block_group(root, ins->objectid, root->leafsize, 1);
7065         if (ret) { /* -ENOENT, logic error */
7066                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7067                         ins->objectid, ins->offset);
7068                 BUG();
7069         }
7070
7071         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->leafsize);
7072         return ret;
7073 }
7074
7075 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7076                                      struct btrfs_root *root,
7077                                      u64 root_objectid, u64 owner,
7078                                      u64 offset, struct btrfs_key *ins)
7079 {
7080         int ret;
7081
7082         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7083
7084         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7085                                          ins->offset, 0,
7086                                          root_objectid, owner, offset,
7087                                          BTRFS_ADD_DELAYED_EXTENT, NULL, 0);
7088         return ret;
7089 }
7090
7091 /*
7092  * this is used by the tree logging recovery code.  It records that
7093  * an extent has been allocated and makes sure to clear the free
7094  * space cache bits as well
7095  */
7096 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7097                                    struct btrfs_root *root,
7098                                    u64 root_objectid, u64 owner, u64 offset,
7099                                    struct btrfs_key *ins)
7100 {
7101         int ret;
7102         struct btrfs_block_group_cache *block_group;
7103
7104         /*
7105          * Mixed block groups will exclude before processing the log so we only
7106          * need to do the exlude dance if this fs isn't mixed.
7107          */
7108         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7109                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7110                 if (ret)
7111                         return ret;
7112         }
7113
7114         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
7115         if (!block_group)
7116                 return -EINVAL;
7117
7118         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
7119                                           RESERVE_ALLOC_NO_ACCOUNT);
7120         BUG_ON(ret); /* logic error */
7121         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
7122                                          0, owner, offset, ins, 1);
7123         btrfs_put_block_group(block_group);
7124         return ret;
7125 }
7126
7127 static struct extent_buffer *
7128 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7129                       u64 bytenr, u32 blocksize, int level)
7130 {
7131         struct extent_buffer *buf;
7132
7133         buf = btrfs_find_create_tree_block(root, bytenr, blocksize);
7134         if (!buf)
7135                 return ERR_PTR(-ENOMEM);
7136         btrfs_set_header_generation(buf, trans->transid);
7137         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
7138         btrfs_tree_lock(buf);
7139         clean_tree_block(trans, root, buf);
7140         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
7141
7142         btrfs_set_lock_blocking(buf);
7143         btrfs_set_buffer_uptodate(buf);
7144
7145         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
7146                 /*
7147                  * we allow two log transactions at a time, use different
7148                  * EXENT bit to differentiate dirty pages.
7149                  */
7150                 if (root->log_transid % 2 == 0)
7151                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7152                                         buf->start + buf->len - 1, GFP_NOFS);
7153                 else
7154                         set_extent_new(&root->dirty_log_pages, buf->start,
7155                                         buf->start + buf->len - 1, GFP_NOFS);
7156         } else {
7157                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7158                          buf->start + buf->len - 1, GFP_NOFS);
7159         }
7160         trans->blocks_used++;
7161         /* this returns a buffer locked for blocking */
7162         return buf;
7163 }
7164
7165 static struct btrfs_block_rsv *
7166 use_block_rsv(struct btrfs_trans_handle *trans,
7167               struct btrfs_root *root, u32 blocksize)
7168 {
7169         struct btrfs_block_rsv *block_rsv;
7170         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7171         int ret;
7172         bool global_updated = false;
7173
7174         block_rsv = get_block_rsv(trans, root);
7175
7176         if (unlikely(block_rsv->size == 0))
7177                 goto try_reserve;
7178 again:
7179         ret = block_rsv_use_bytes(block_rsv, blocksize);
7180         if (!ret)
7181                 return block_rsv;
7182
7183         if (block_rsv->failfast)
7184                 return ERR_PTR(ret);
7185
7186         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7187                 global_updated = true;
7188                 update_global_block_rsv(root->fs_info);
7189                 goto again;
7190         }
7191
7192         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7193                 static DEFINE_RATELIMIT_STATE(_rs,
7194                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7195                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7196                 if (__ratelimit(&_rs))
7197                         WARN(1, KERN_DEBUG
7198                                 "BTRFS: block rsv returned %d\n", ret);
7199         }
7200 try_reserve:
7201         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7202                                      BTRFS_RESERVE_NO_FLUSH);
7203         if (!ret)
7204                 return block_rsv;
7205         /*
7206          * If we couldn't reserve metadata bytes try and use some from
7207          * the global reserve if its space type is the same as the global
7208          * reservation.
7209          */
7210         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7211             block_rsv->space_info == global_rsv->space_info) {
7212                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7213                 if (!ret)
7214                         return global_rsv;
7215         }
7216         return ERR_PTR(ret);
7217 }
7218
7219 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7220                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7221 {
7222         block_rsv_add_bytes(block_rsv, blocksize, 0);
7223         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7224 }
7225
7226 /*
7227  * finds a free extent and does all the dirty work required for allocation
7228  * returns the key for the extent through ins, and a tree buffer for
7229  * the first block of the extent through buf.
7230  *
7231  * returns the tree buffer or NULL.
7232  */
7233 struct extent_buffer *btrfs_alloc_free_block(struct btrfs_trans_handle *trans,
7234                                         struct btrfs_root *root, u32 blocksize,
7235                                         u64 parent, u64 root_objectid,
7236                                         struct btrfs_disk_key *key, int level,
7237                                         u64 hint, u64 empty_size)
7238 {
7239         struct btrfs_key ins;
7240         struct btrfs_block_rsv *block_rsv;
7241         struct extent_buffer *buf;
7242         u64 flags = 0;
7243         int ret;
7244         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7245                                                  SKINNY_METADATA);
7246
7247 #ifdef CONFIG_BTRFS_FS_RUN_SANITY_TESTS
7248         if (unlikely(test_bit(BTRFS_ROOT_DUMMY_ROOT, &root->state))) {
7249                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
7250                                             blocksize, level);
7251                 if (!IS_ERR(buf))
7252                         root->alloc_bytenr += blocksize;
7253                 return buf;
7254         }
7255 #endif
7256         block_rsv = use_block_rsv(trans, root, blocksize);
7257         if (IS_ERR(block_rsv))
7258                 return ERR_CAST(block_rsv);
7259
7260         ret = btrfs_reserve_extent(root, blocksize, blocksize,
7261                                    empty_size, hint, &ins, 0);
7262         if (ret) {
7263                 unuse_block_rsv(root->fs_info, block_rsv, blocksize);
7264                 return ERR_PTR(ret);
7265         }
7266
7267         buf = btrfs_init_new_buffer(trans, root, ins.objectid,
7268                                     blocksize, level);
7269         BUG_ON(IS_ERR(buf)); /* -ENOMEM */
7270
7271         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
7272                 if (parent == 0)
7273                         parent = ins.objectid;
7274                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
7275         } else
7276                 BUG_ON(parent > 0);
7277
7278         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
7279                 struct btrfs_delayed_extent_op *extent_op;
7280                 extent_op = btrfs_alloc_delayed_extent_op();
7281                 BUG_ON(!extent_op); /* -ENOMEM */
7282                 if (key)
7283                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
7284                 else
7285                         memset(&extent_op->key, 0, sizeof(extent_op->key));
7286                 extent_op->flags_to_set = flags;
7287                 if (skinny_metadata)
7288                         extent_op->update_key = 0;
7289                 else
7290                         extent_op->update_key = 1;
7291                 extent_op->update_flags = 1;
7292                 extent_op->is_data = 0;
7293                 extent_op->level = level;
7294
7295                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
7296                                         ins.objectid,
7297                                         ins.offset, parent, root_objectid,
7298                                         level, BTRFS_ADD_DELAYED_EXTENT,
7299                                         extent_op, 0);
7300                 BUG_ON(ret); /* -ENOMEM */
7301         }
7302         return buf;
7303 }
7304
7305 struct walk_control {
7306         u64 refs[BTRFS_MAX_LEVEL];
7307         u64 flags[BTRFS_MAX_LEVEL];
7308         struct btrfs_key update_progress;
7309         int stage;
7310         int level;
7311         int shared_level;
7312         int update_ref;
7313         int keep_locks;
7314         int reada_slot;
7315         int reada_count;
7316         int for_reloc;
7317 };
7318
7319 #define DROP_REFERENCE  1
7320 #define UPDATE_BACKREF  2
7321
7322 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
7323                                      struct btrfs_root *root,
7324                                      struct walk_control *wc,
7325                                      struct btrfs_path *path)
7326 {
7327         u64 bytenr;
7328         u64 generation;
7329         u64 refs;
7330         u64 flags;
7331         u32 nritems;
7332         u32 blocksize;
7333         struct btrfs_key key;
7334         struct extent_buffer *eb;
7335         int ret;
7336         int slot;
7337         int nread = 0;
7338
7339         if (path->slots[wc->level] < wc->reada_slot) {
7340                 wc->reada_count = wc->reada_count * 2 / 3;
7341                 wc->reada_count = max(wc->reada_count, 2);
7342         } else {
7343                 wc->reada_count = wc->reada_count * 3 / 2;
7344                 wc->reada_count = min_t(int, wc->reada_count,
7345                                         BTRFS_NODEPTRS_PER_BLOCK(root));
7346         }
7347
7348         eb = path->nodes[wc->level];
7349         nritems = btrfs_header_nritems(eb);
7350         blocksize = btrfs_level_size(root, wc->level - 1);
7351
7352         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
7353                 if (nread >= wc->reada_count)
7354                         break;
7355
7356                 cond_resched();
7357                 bytenr = btrfs_node_blockptr(eb, slot);
7358                 generation = btrfs_node_ptr_generation(eb, slot);
7359
7360                 if (slot == path->slots[wc->level])
7361                         goto reada;
7362
7363                 if (wc->stage == UPDATE_BACKREF &&
7364                     generation <= root->root_key.offset)
7365                         continue;
7366
7367                 /* We don't lock the tree block, it's OK to be racy here */
7368                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
7369                                                wc->level - 1, 1, &refs,
7370                                                &flags);
7371                 /* We don't care about errors in readahead. */
7372                 if (ret < 0)
7373                         continue;
7374                 BUG_ON(refs == 0);
7375
7376                 if (wc->stage == DROP_REFERENCE) {
7377                         if (refs == 1)
7378                                 goto reada;
7379
7380                         if (wc->level == 1 &&
7381                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7382                                 continue;
7383                         if (!wc->update_ref ||
7384                             generation <= root->root_key.offset)
7385                                 continue;
7386                         btrfs_node_key_to_cpu(eb, &key, slot);
7387                         ret = btrfs_comp_cpu_keys(&key,
7388                                                   &wc->update_progress);
7389                         if (ret < 0)
7390                                 continue;
7391                 } else {
7392                         if (wc->level == 1 &&
7393                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7394                                 continue;
7395                 }
7396 reada:
7397                 ret = readahead_tree_block(root, bytenr, blocksize,
7398                                            generation);
7399                 if (ret)
7400                         break;
7401                 nread++;
7402         }
7403         wc->reada_slot = slot;
7404 }
7405
7406 /*
7407  * helper to process tree block while walking down the tree.
7408  *
7409  * when wc->stage == UPDATE_BACKREF, this function updates
7410  * back refs for pointers in the block.
7411  *
7412  * NOTE: return value 1 means we should stop walking down.
7413  */
7414 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
7415                                    struct btrfs_root *root,
7416                                    struct btrfs_path *path,
7417                                    struct walk_control *wc, int lookup_info)
7418 {
7419         int level = wc->level;
7420         struct extent_buffer *eb = path->nodes[level];
7421         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
7422         int ret;
7423
7424         if (wc->stage == UPDATE_BACKREF &&
7425             btrfs_header_owner(eb) != root->root_key.objectid)
7426                 return 1;
7427
7428         /*
7429          * when reference count of tree block is 1, it won't increase
7430          * again. once full backref flag is set, we never clear it.
7431          */
7432         if (lookup_info &&
7433             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
7434              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
7435                 BUG_ON(!path->locks[level]);
7436                 ret = btrfs_lookup_extent_info(trans, root,
7437                                                eb->start, level, 1,
7438                                                &wc->refs[level],
7439                                                &wc->flags[level]);
7440                 BUG_ON(ret == -ENOMEM);
7441                 if (ret)
7442                         return ret;
7443                 BUG_ON(wc->refs[level] == 0);
7444         }
7445
7446         if (wc->stage == DROP_REFERENCE) {
7447                 if (wc->refs[level] > 1)
7448                         return 1;
7449
7450                 if (path->locks[level] && !wc->keep_locks) {
7451                         btrfs_tree_unlock_rw(eb, path->locks[level]);
7452                         path->locks[level] = 0;
7453                 }
7454                 return 0;
7455         }
7456
7457         /* wc->stage == UPDATE_BACKREF */
7458         if (!(wc->flags[level] & flag)) {
7459                 BUG_ON(!path->locks[level]);
7460                 ret = btrfs_inc_ref(trans, root, eb, 1, wc->for_reloc);
7461                 BUG_ON(ret); /* -ENOMEM */
7462                 ret = btrfs_dec_ref(trans, root, eb, 0, wc->for_reloc);
7463                 BUG_ON(ret); /* -ENOMEM */
7464                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
7465                                                   eb->len, flag,
7466                                                   btrfs_header_level(eb), 0);
7467                 BUG_ON(ret); /* -ENOMEM */
7468                 wc->flags[level] |= flag;
7469         }
7470
7471         /*
7472          * the block is shared by multiple trees, so it's not good to
7473          * keep the tree lock
7474          */
7475         if (path->locks[level] && level > 0) {
7476                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7477                 path->locks[level] = 0;
7478         }
7479         return 0;
7480 }
7481
7482 /*
7483  * helper to process tree block pointer.
7484  *
7485  * when wc->stage == DROP_REFERENCE, this function checks
7486  * reference count of the block pointed to. if the block
7487  * is shared and we need update back refs for the subtree
7488  * rooted at the block, this function changes wc->stage to
7489  * UPDATE_BACKREF. if the block is shared and there is no
7490  * need to update back, this function drops the reference
7491  * to the block.
7492  *
7493  * NOTE: return value 1 means we should stop walking down.
7494  */
7495 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
7496                                  struct btrfs_root *root,
7497                                  struct btrfs_path *path,
7498                                  struct walk_control *wc, int *lookup_info)
7499 {
7500         u64 bytenr;
7501         u64 generation;
7502         u64 parent;
7503         u32 blocksize;
7504         struct btrfs_key key;
7505         struct extent_buffer *next;
7506         int level = wc->level;
7507         int reada = 0;
7508         int ret = 0;
7509
7510         generation = btrfs_node_ptr_generation(path->nodes[level],
7511                                                path->slots[level]);
7512         /*
7513          * if the lower level block was created before the snapshot
7514          * was created, we know there is no need to update back refs
7515          * for the subtree
7516          */
7517         if (wc->stage == UPDATE_BACKREF &&
7518             generation <= root->root_key.offset) {
7519                 *lookup_info = 1;
7520                 return 1;
7521         }
7522
7523         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
7524         blocksize = btrfs_level_size(root, level - 1);
7525
7526         next = btrfs_find_tree_block(root, bytenr, blocksize);
7527         if (!next) {
7528                 next = btrfs_find_create_tree_block(root, bytenr, blocksize);
7529                 if (!next)
7530                         return -ENOMEM;
7531                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
7532                                                level - 1);
7533                 reada = 1;
7534         }
7535         btrfs_tree_lock(next);
7536         btrfs_set_lock_blocking(next);
7537
7538         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
7539                                        &wc->refs[level - 1],
7540                                        &wc->flags[level - 1]);
7541         if (ret < 0) {
7542                 btrfs_tree_unlock(next);
7543                 return ret;
7544         }
7545
7546         if (unlikely(wc->refs[level - 1] == 0)) {
7547                 btrfs_err(root->fs_info, "Missing references.");
7548                 BUG();
7549         }
7550         *lookup_info = 0;
7551
7552         if (wc->stage == DROP_REFERENCE) {
7553                 if (wc->refs[level - 1] > 1) {
7554                         if (level == 1 &&
7555                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7556                                 goto skip;
7557
7558                         if (!wc->update_ref ||
7559                             generation <= root->root_key.offset)
7560                                 goto skip;
7561
7562                         btrfs_node_key_to_cpu(path->nodes[level], &key,
7563                                               path->slots[level]);
7564                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
7565                         if (ret < 0)
7566                                 goto skip;
7567
7568                         wc->stage = UPDATE_BACKREF;
7569                         wc->shared_level = level - 1;
7570                 }
7571         } else {
7572                 if (level == 1 &&
7573                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
7574                         goto skip;
7575         }
7576
7577         if (!btrfs_buffer_uptodate(next, generation, 0)) {
7578                 btrfs_tree_unlock(next);
7579                 free_extent_buffer(next);
7580                 next = NULL;
7581                 *lookup_info = 1;
7582         }
7583
7584         if (!next) {
7585                 if (reada && level == 1)
7586                         reada_walk_down(trans, root, wc, path);
7587                 next = read_tree_block(root, bytenr, blocksize, generation);
7588                 if (!next || !extent_buffer_uptodate(next)) {
7589                         free_extent_buffer(next);
7590                         return -EIO;
7591                 }
7592                 btrfs_tree_lock(next);
7593                 btrfs_set_lock_blocking(next);
7594         }
7595
7596         level--;
7597         BUG_ON(level != btrfs_header_level(next));
7598         path->nodes[level] = next;
7599         path->slots[level] = 0;
7600         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7601         wc->level = level;
7602         if (wc->level == 1)
7603                 wc->reada_slot = 0;
7604         return 0;
7605 skip:
7606         wc->refs[level - 1] = 0;
7607         wc->flags[level - 1] = 0;
7608         if (wc->stage == DROP_REFERENCE) {
7609                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
7610                         parent = path->nodes[level]->start;
7611                 } else {
7612                         BUG_ON(root->root_key.objectid !=
7613                                btrfs_header_owner(path->nodes[level]));
7614                         parent = 0;
7615                 }
7616
7617                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
7618                                 root->root_key.objectid, level - 1, 0, 0);
7619                 BUG_ON(ret); /* -ENOMEM */
7620         }
7621         btrfs_tree_unlock(next);
7622         free_extent_buffer(next);
7623         *lookup_info = 1;
7624         return 1;
7625 }
7626
7627 /*
7628  * helper to process tree block while walking up the tree.
7629  *
7630  * when wc->stage == DROP_REFERENCE, this function drops
7631  * reference count on the block.
7632  *
7633  * when wc->stage == UPDATE_BACKREF, this function changes
7634  * wc->stage back to DROP_REFERENCE if we changed wc->stage
7635  * to UPDATE_BACKREF previously while processing the block.
7636  *
7637  * NOTE: return value 1 means we should stop walking up.
7638  */
7639 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
7640                                  struct btrfs_root *root,
7641                                  struct btrfs_path *path,
7642                                  struct walk_control *wc)
7643 {
7644         int ret;
7645         int level = wc->level;
7646         struct extent_buffer *eb = path->nodes[level];
7647         u64 parent = 0;
7648
7649         if (wc->stage == UPDATE_BACKREF) {
7650                 BUG_ON(wc->shared_level < level);
7651                 if (level < wc->shared_level)
7652                         goto out;
7653
7654                 ret = find_next_key(path, level + 1, &wc->update_progress);
7655                 if (ret > 0)
7656                         wc->update_ref = 0;
7657
7658                 wc->stage = DROP_REFERENCE;
7659                 wc->shared_level = -1;
7660                 path->slots[level] = 0;
7661
7662                 /*
7663                  * check reference count again if the block isn't locked.
7664                  * we should start walking down the tree again if reference
7665                  * count is one.
7666                  */
7667                 if (!path->locks[level]) {
7668                         BUG_ON(level == 0);
7669                         btrfs_tree_lock(eb);
7670                         btrfs_set_lock_blocking(eb);
7671                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7672
7673                         ret = btrfs_lookup_extent_info(trans, root,
7674                                                        eb->start, level, 1,
7675                                                        &wc->refs[level],
7676                                                        &wc->flags[level]);
7677                         if (ret < 0) {
7678                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7679                                 path->locks[level] = 0;
7680                                 return ret;
7681                         }
7682                         BUG_ON(wc->refs[level] == 0);
7683                         if (wc->refs[level] == 1) {
7684                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
7685                                 path->locks[level] = 0;
7686                                 return 1;
7687                         }
7688                 }
7689         }
7690
7691         /* wc->stage == DROP_REFERENCE */
7692         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
7693
7694         if (wc->refs[level] == 1) {
7695                 if (level == 0) {
7696                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7697                                 ret = btrfs_dec_ref(trans, root, eb, 1,
7698                                                     wc->for_reloc);
7699                         else
7700                                 ret = btrfs_dec_ref(trans, root, eb, 0,
7701                                                     wc->for_reloc);
7702                         BUG_ON(ret); /* -ENOMEM */
7703                 }
7704                 /* make block locked assertion in clean_tree_block happy */
7705                 if (!path->locks[level] &&
7706                     btrfs_header_generation(eb) == trans->transid) {
7707                         btrfs_tree_lock(eb);
7708                         btrfs_set_lock_blocking(eb);
7709                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7710                 }
7711                 clean_tree_block(trans, root, eb);
7712         }
7713
7714         if (eb == root->node) {
7715                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7716                         parent = eb->start;
7717                 else
7718                         BUG_ON(root->root_key.objectid !=
7719                                btrfs_header_owner(eb));
7720         } else {
7721                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
7722                         parent = path->nodes[level + 1]->start;
7723                 else
7724                         BUG_ON(root->root_key.objectid !=
7725                                btrfs_header_owner(path->nodes[level + 1]));
7726         }
7727
7728         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
7729 out:
7730         wc->refs[level] = 0;
7731         wc->flags[level] = 0;
7732         return 0;
7733 }
7734
7735 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
7736                                    struct btrfs_root *root,
7737                                    struct btrfs_path *path,
7738                                    struct walk_control *wc)
7739 {
7740         int level = wc->level;
7741         int lookup_info = 1;
7742         int ret;
7743
7744         while (level >= 0) {
7745                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
7746                 if (ret > 0)
7747                         break;
7748
7749                 if (level == 0)
7750                         break;
7751
7752                 if (path->slots[level] >=
7753                     btrfs_header_nritems(path->nodes[level]))
7754                         break;
7755
7756                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
7757                 if (ret > 0) {
7758                         path->slots[level]++;
7759                         continue;
7760                 } else if (ret < 0)
7761                         return ret;
7762                 level = wc->level;
7763         }
7764         return 0;
7765 }
7766
7767 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
7768                                  struct btrfs_root *root,
7769                                  struct btrfs_path *path,
7770                                  struct walk_control *wc, int max_level)
7771 {
7772         int level = wc->level;
7773         int ret;
7774
7775         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
7776         while (level < max_level && path->nodes[level]) {
7777                 wc->level = level;
7778                 if (path->slots[level] + 1 <
7779                     btrfs_header_nritems(path->nodes[level])) {
7780                         path->slots[level]++;
7781                         return 0;
7782                 } else {
7783                         ret = walk_up_proc(trans, root, path, wc);
7784                         if (ret > 0)
7785                                 return 0;
7786
7787                         if (path->locks[level]) {
7788                                 btrfs_tree_unlock_rw(path->nodes[level],
7789                                                      path->locks[level]);
7790                                 path->locks[level] = 0;
7791                         }
7792                         free_extent_buffer(path->nodes[level]);
7793                         path->nodes[level] = NULL;
7794                         level++;
7795                 }
7796         }
7797         return 1;
7798 }
7799
7800 /*
7801  * drop a subvolume tree.
7802  *
7803  * this function traverses the tree freeing any blocks that only
7804  * referenced by the tree.
7805  *
7806  * when a shared tree block is found. this function decreases its
7807  * reference count by one. if update_ref is true, this function
7808  * also make sure backrefs for the shared block and all lower level
7809  * blocks are properly updated.
7810  *
7811  * If called with for_reloc == 0, may exit early with -EAGAIN
7812  */
7813 int btrfs_drop_snapshot(struct btrfs_root *root,
7814                          struct btrfs_block_rsv *block_rsv, int update_ref,
7815                          int for_reloc)
7816 {
7817         struct btrfs_path *path;
7818         struct btrfs_trans_handle *trans;
7819         struct btrfs_root *tree_root = root->fs_info->tree_root;
7820         struct btrfs_root_item *root_item = &root->root_item;
7821         struct walk_control *wc;
7822         struct btrfs_key key;
7823         int err = 0;
7824         int ret;
7825         int level;
7826         bool root_dropped = false;
7827
7828         path = btrfs_alloc_path();
7829         if (!path) {
7830                 err = -ENOMEM;
7831                 goto out;
7832         }
7833
7834         wc = kzalloc(sizeof(*wc), GFP_NOFS);
7835         if (!wc) {
7836                 btrfs_free_path(path);
7837                 err = -ENOMEM;
7838                 goto out;
7839         }
7840
7841         trans = btrfs_start_transaction(tree_root, 0);
7842         if (IS_ERR(trans)) {
7843                 err = PTR_ERR(trans);
7844                 goto out_free;
7845         }
7846
7847         if (block_rsv)
7848                 trans->block_rsv = block_rsv;
7849
7850         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
7851                 level = btrfs_header_level(root->node);
7852                 path->nodes[level] = btrfs_lock_root_node(root);
7853                 btrfs_set_lock_blocking(path->nodes[level]);
7854                 path->slots[level] = 0;
7855                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7856                 memset(&wc->update_progress, 0,
7857                        sizeof(wc->update_progress));
7858         } else {
7859                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
7860                 memcpy(&wc->update_progress, &key,
7861                        sizeof(wc->update_progress));
7862
7863                 level = root_item->drop_level;
7864                 BUG_ON(level == 0);
7865                 path->lowest_level = level;
7866                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
7867                 path->lowest_level = 0;
7868                 if (ret < 0) {
7869                         err = ret;
7870                         goto out_end_trans;
7871                 }
7872                 WARN_ON(ret > 0);
7873
7874                 /*
7875                  * unlock our path, this is safe because only this
7876                  * function is allowed to delete this snapshot
7877                  */
7878                 btrfs_unlock_up_safe(path, 0);
7879
7880                 level = btrfs_header_level(root->node);
7881                 while (1) {
7882                         btrfs_tree_lock(path->nodes[level]);
7883                         btrfs_set_lock_blocking(path->nodes[level]);
7884                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
7885
7886                         ret = btrfs_lookup_extent_info(trans, root,
7887                                                 path->nodes[level]->start,
7888                                                 level, 1, &wc->refs[level],
7889                                                 &wc->flags[level]);
7890                         if (ret < 0) {
7891                                 err = ret;
7892                                 goto out_end_trans;
7893                         }
7894                         BUG_ON(wc->refs[level] == 0);
7895
7896                         if (level == root_item->drop_level)
7897                                 break;
7898
7899                         btrfs_tree_unlock(path->nodes[level]);
7900                         path->locks[level] = 0;
7901                         WARN_ON(wc->refs[level] != 1);
7902                         level--;
7903                 }
7904         }
7905
7906         wc->level = level;
7907         wc->shared_level = -1;
7908         wc->stage = DROP_REFERENCE;
7909         wc->update_ref = update_ref;
7910         wc->keep_locks = 0;
7911         wc->for_reloc = for_reloc;
7912         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
7913
7914         while (1) {
7915
7916                 ret = walk_down_tree(trans, root, path, wc);
7917                 if (ret < 0) {
7918                         err = ret;
7919                         break;
7920                 }
7921
7922                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
7923                 if (ret < 0) {
7924                         err = ret;
7925                         break;
7926                 }
7927
7928                 if (ret > 0) {
7929                         BUG_ON(wc->stage != DROP_REFERENCE);
7930                         break;
7931                 }
7932
7933                 if (wc->stage == DROP_REFERENCE) {
7934                         level = wc->level;
7935                         btrfs_node_key(path->nodes[level],
7936                                        &root_item->drop_progress,
7937                                        path->slots[level]);
7938                         root_item->drop_level = level;
7939                 }
7940
7941                 BUG_ON(wc->level == 0);
7942                 if (btrfs_should_end_transaction(trans, tree_root) ||
7943                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
7944                         ret = btrfs_update_root(trans, tree_root,
7945                                                 &root->root_key,
7946                                                 root_item);
7947                         if (ret) {
7948                                 btrfs_abort_transaction(trans, tree_root, ret);
7949                                 err = ret;
7950                                 goto out_end_trans;
7951                         }
7952
7953                         btrfs_end_transaction_throttle(trans, tree_root);
7954                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
7955                                 pr_debug("BTRFS: drop snapshot early exit\n");
7956                                 err = -EAGAIN;
7957                                 goto out_free;
7958                         }
7959
7960                         trans = btrfs_start_transaction(tree_root, 0);
7961                         if (IS_ERR(trans)) {
7962                                 err = PTR_ERR(trans);
7963                                 goto out_free;
7964                         }
7965                         if (block_rsv)
7966                                 trans->block_rsv = block_rsv;
7967                 }
7968         }
7969         btrfs_release_path(path);
7970         if (err)
7971                 goto out_end_trans;
7972
7973         ret = btrfs_del_root(trans, tree_root, &root->root_key);
7974         if (ret) {
7975                 btrfs_abort_transaction(trans, tree_root, ret);
7976                 goto out_end_trans;
7977         }
7978
7979         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
7980                 ret = btrfs_find_root(tree_root, &root->root_key, path,
7981                                       NULL, NULL);
7982                 if (ret < 0) {
7983                         btrfs_abort_transaction(trans, tree_root, ret);
7984                         err = ret;
7985                         goto out_end_trans;
7986                 } else if (ret > 0) {
7987                         /* if we fail to delete the orphan item this time
7988                          * around, it'll get picked up the next time.
7989                          *
7990                          * The most common failure here is just -ENOENT.
7991                          */
7992                         btrfs_del_orphan_item(trans, tree_root,
7993                                               root->root_key.objectid);
7994                 }
7995         }
7996
7997         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
7998                 btrfs_drop_and_free_fs_root(tree_root->fs_info, root);
7999         } else {
8000                 free_extent_buffer(root->node);
8001                 free_extent_buffer(root->commit_root);
8002                 btrfs_put_fs_root(root);
8003         }
8004         root_dropped = true;
8005 out_end_trans:
8006         btrfs_end_transaction_throttle(trans, tree_root);
8007 out_free:
8008         kfree(wc);
8009         btrfs_free_path(path);
8010 out:
8011         /*
8012          * So if we need to stop dropping the snapshot for whatever reason we
8013          * need to make sure to add it back to the dead root list so that we
8014          * keep trying to do the work later.  This also cleans up roots if we
8015          * don't have it in the radix (like when we recover after a power fail
8016          * or unmount) so we don't leak memory.
8017          */
8018         if (!for_reloc && root_dropped == false)
8019                 btrfs_add_dead_root(root);
8020         if (err && err != -EAGAIN)
8021                 btrfs_std_error(root->fs_info, err);
8022         return err;
8023 }
8024
8025 /*
8026  * drop subtree rooted at tree block 'node'.
8027  *
8028  * NOTE: this function will unlock and release tree block 'node'
8029  * only used by relocation code
8030  */
8031 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
8032                         struct btrfs_root *root,
8033                         struct extent_buffer *node,
8034                         struct extent_buffer *parent)
8035 {
8036         struct btrfs_path *path;
8037         struct walk_control *wc;
8038         int level;
8039         int parent_level;
8040         int ret = 0;
8041         int wret;
8042
8043         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
8044
8045         path = btrfs_alloc_path();
8046         if (!path)
8047                 return -ENOMEM;
8048
8049         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8050         if (!wc) {
8051                 btrfs_free_path(path);
8052                 return -ENOMEM;
8053         }
8054
8055         btrfs_assert_tree_locked(parent);
8056         parent_level = btrfs_header_level(parent);
8057         extent_buffer_get(parent);
8058         path->nodes[parent_level] = parent;
8059         path->slots[parent_level] = btrfs_header_nritems(parent);
8060
8061         btrfs_assert_tree_locked(node);
8062         level = btrfs_header_level(node);
8063         path->nodes[level] = node;
8064         path->slots[level] = 0;
8065         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8066
8067         wc->refs[parent_level] = 1;
8068         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8069         wc->level = level;
8070         wc->shared_level = -1;
8071         wc->stage = DROP_REFERENCE;
8072         wc->update_ref = 0;
8073         wc->keep_locks = 1;
8074         wc->for_reloc = 1;
8075         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8076
8077         while (1) {
8078                 wret = walk_down_tree(trans, root, path, wc);
8079                 if (wret < 0) {
8080                         ret = wret;
8081                         break;
8082                 }
8083
8084                 wret = walk_up_tree(trans, root, path, wc, parent_level);
8085                 if (wret < 0)
8086                         ret = wret;
8087                 if (wret != 0)
8088                         break;
8089         }
8090
8091         kfree(wc);
8092         btrfs_free_path(path);
8093         return ret;
8094 }
8095
8096 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
8097 {
8098         u64 num_devices;
8099         u64 stripped;
8100
8101         /*
8102          * if restripe for this chunk_type is on pick target profile and
8103          * return, otherwise do the usual balance
8104          */
8105         stripped = get_restripe_target(root->fs_info, flags);
8106         if (stripped)
8107                 return extended_to_chunk(stripped);
8108
8109         /*
8110          * we add in the count of missing devices because we want
8111          * to make sure that any RAID levels on a degraded FS
8112          * continue to be honored.
8113          */
8114         num_devices = root->fs_info->fs_devices->rw_devices +
8115                 root->fs_info->fs_devices->missing_devices;
8116
8117         stripped = BTRFS_BLOCK_GROUP_RAID0 |
8118                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
8119                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
8120
8121         if (num_devices == 1) {
8122                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8123                 stripped = flags & ~stripped;
8124
8125                 /* turn raid0 into single device chunks */
8126                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
8127                         return stripped;
8128
8129                 /* turn mirroring into duplication */
8130                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
8131                              BTRFS_BLOCK_GROUP_RAID10))
8132                         return stripped | BTRFS_BLOCK_GROUP_DUP;
8133         } else {
8134                 /* they already had raid on here, just return */
8135                 if (flags & stripped)
8136                         return flags;
8137
8138                 stripped |= BTRFS_BLOCK_GROUP_DUP;
8139                 stripped = flags & ~stripped;
8140
8141                 /* switch duplicated blocks with raid1 */
8142                 if (flags & BTRFS_BLOCK_GROUP_DUP)
8143                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
8144
8145                 /* this is drive concat, leave it alone */
8146         }
8147
8148         return flags;
8149 }
8150
8151 static int set_block_group_ro(struct btrfs_block_group_cache *cache, int force)
8152 {
8153         struct btrfs_space_info *sinfo = cache->space_info;
8154         u64 num_bytes;
8155         u64 min_allocable_bytes;
8156         int ret = -ENOSPC;
8157
8158
8159         /*
8160          * We need some metadata space and system metadata space for
8161          * allocating chunks in some corner cases until we force to set
8162          * it to be readonly.
8163          */
8164         if ((sinfo->flags &
8165              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
8166             !force)
8167                 min_allocable_bytes = 1 * 1024 * 1024;
8168         else
8169                 min_allocable_bytes = 0;
8170
8171         spin_lock(&sinfo->lock);
8172         spin_lock(&cache->lock);
8173
8174         if (cache->ro) {
8175                 ret = 0;
8176                 goto out;
8177         }
8178
8179         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8180                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8181
8182         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
8183             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
8184             min_allocable_bytes <= sinfo->total_bytes) {
8185                 sinfo->bytes_readonly += num_bytes;
8186                 cache->ro = 1;
8187                 ret = 0;
8188         }
8189 out:
8190         spin_unlock(&cache->lock);
8191         spin_unlock(&sinfo->lock);
8192         return ret;
8193 }
8194
8195 int btrfs_set_block_group_ro(struct btrfs_root *root,
8196                              struct btrfs_block_group_cache *cache)
8197
8198 {
8199         struct btrfs_trans_handle *trans;
8200         u64 alloc_flags;
8201         int ret;
8202
8203         BUG_ON(cache->ro);
8204
8205         trans = btrfs_join_transaction(root);
8206         if (IS_ERR(trans))
8207                 return PTR_ERR(trans);
8208
8209         alloc_flags = update_block_group_flags(root, cache->flags);
8210         if (alloc_flags != cache->flags) {
8211                 ret = do_chunk_alloc(trans, root, alloc_flags,
8212                                      CHUNK_ALLOC_FORCE);
8213                 if (ret < 0)
8214                         goto out;
8215         }
8216
8217         ret = set_block_group_ro(cache, 0);
8218         if (!ret)
8219                 goto out;
8220         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
8221         ret = do_chunk_alloc(trans, root, alloc_flags,
8222                              CHUNK_ALLOC_FORCE);
8223         if (ret < 0)
8224                 goto out;
8225         ret = set_block_group_ro(cache, 0);
8226 out:
8227         btrfs_end_transaction(trans, root);
8228         return ret;
8229 }
8230
8231 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
8232                             struct btrfs_root *root, u64 type)
8233 {
8234         u64 alloc_flags = get_alloc_profile(root, type);
8235         return do_chunk_alloc(trans, root, alloc_flags,
8236                               CHUNK_ALLOC_FORCE);
8237 }
8238
8239 /*
8240  * helper to account the unused space of all the readonly block group in the
8241  * list. takes mirrors into account.
8242  */
8243 static u64 __btrfs_get_ro_block_group_free_space(struct list_head *groups_list)
8244 {
8245         struct btrfs_block_group_cache *block_group;
8246         u64 free_bytes = 0;
8247         int factor;
8248
8249         list_for_each_entry(block_group, groups_list, list) {
8250                 spin_lock(&block_group->lock);
8251
8252                 if (!block_group->ro) {
8253                         spin_unlock(&block_group->lock);
8254                         continue;
8255                 }
8256
8257                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
8258                                           BTRFS_BLOCK_GROUP_RAID10 |
8259                                           BTRFS_BLOCK_GROUP_DUP))
8260                         factor = 2;
8261                 else
8262                         factor = 1;
8263
8264                 free_bytes += (block_group->key.offset -
8265                                btrfs_block_group_used(&block_group->item)) *
8266                                factor;
8267
8268                 spin_unlock(&block_group->lock);
8269         }
8270
8271         return free_bytes;
8272 }
8273
8274 /*
8275  * helper to account the unused space of all the readonly block group in the
8276  * space_info. takes mirrors into account.
8277  */
8278 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
8279 {
8280         int i;
8281         u64 free_bytes = 0;
8282
8283         spin_lock(&sinfo->lock);
8284
8285         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
8286                 if (!list_empty(&sinfo->block_groups[i]))
8287                         free_bytes += __btrfs_get_ro_block_group_free_space(
8288                                                 &sinfo->block_groups[i]);
8289
8290         spin_unlock(&sinfo->lock);
8291
8292         return free_bytes;
8293 }
8294
8295 void btrfs_set_block_group_rw(struct btrfs_root *root,
8296                               struct btrfs_block_group_cache *cache)
8297 {
8298         struct btrfs_space_info *sinfo = cache->space_info;
8299         u64 num_bytes;
8300
8301         BUG_ON(!cache->ro);
8302
8303         spin_lock(&sinfo->lock);
8304         spin_lock(&cache->lock);
8305         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
8306                     cache->bytes_super - btrfs_block_group_used(&cache->item);
8307         sinfo->bytes_readonly -= num_bytes;
8308         cache->ro = 0;
8309         spin_unlock(&cache->lock);
8310         spin_unlock(&sinfo->lock);
8311 }
8312
8313 /*
8314  * checks to see if its even possible to relocate this block group.
8315  *
8316  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
8317  * ok to go ahead and try.
8318  */
8319 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
8320 {
8321         struct btrfs_block_group_cache *block_group;
8322         struct btrfs_space_info *space_info;
8323         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
8324         struct btrfs_device *device;
8325         struct btrfs_trans_handle *trans;
8326         u64 min_free;
8327         u64 dev_min = 1;
8328         u64 dev_nr = 0;
8329         u64 target;
8330         int index;
8331         int full = 0;
8332         int ret = 0;
8333
8334         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
8335
8336         /* odd, couldn't find the block group, leave it alone */
8337         if (!block_group)
8338                 return -1;
8339
8340         min_free = btrfs_block_group_used(&block_group->item);
8341
8342         /* no bytes used, we're good */
8343         if (!min_free)
8344                 goto out;
8345
8346         space_info = block_group->space_info;
8347         spin_lock(&space_info->lock);
8348
8349         full = space_info->full;
8350
8351         /*
8352          * if this is the last block group we have in this space, we can't
8353          * relocate it unless we're able to allocate a new chunk below.
8354          *
8355          * Otherwise, we need to make sure we have room in the space to handle
8356          * all of the extents from this block group.  If we can, we're good
8357          */
8358         if ((space_info->total_bytes != block_group->key.offset) &&
8359             (space_info->bytes_used + space_info->bytes_reserved +
8360              space_info->bytes_pinned + space_info->bytes_readonly +
8361              min_free < space_info->total_bytes)) {
8362                 spin_unlock(&space_info->lock);
8363                 goto out;
8364         }
8365         spin_unlock(&space_info->lock);
8366
8367         /*
8368          * ok we don't have enough space, but maybe we have free space on our
8369          * devices to allocate new chunks for relocation, so loop through our
8370          * alloc devices and guess if we have enough space.  if this block
8371          * group is going to be restriped, run checks against the target
8372          * profile instead of the current one.
8373          */
8374         ret = -1;
8375
8376         /*
8377          * index:
8378          *      0: raid10
8379          *      1: raid1
8380          *      2: dup
8381          *      3: raid0
8382          *      4: single
8383          */
8384         target = get_restripe_target(root->fs_info, block_group->flags);
8385         if (target) {
8386                 index = __get_raid_index(extended_to_chunk(target));
8387         } else {
8388                 /*
8389                  * this is just a balance, so if we were marked as full
8390                  * we know there is no space for a new chunk
8391                  */
8392                 if (full)
8393                         goto out;
8394
8395                 index = get_block_group_index(block_group);
8396         }
8397
8398         if (index == BTRFS_RAID_RAID10) {
8399                 dev_min = 4;
8400                 /* Divide by 2 */
8401                 min_free >>= 1;
8402         } else if (index == BTRFS_RAID_RAID1) {
8403                 dev_min = 2;
8404         } else if (index == BTRFS_RAID_DUP) {
8405                 /* Multiply by 2 */
8406                 min_free <<= 1;
8407         } else if (index == BTRFS_RAID_RAID0) {
8408                 dev_min = fs_devices->rw_devices;
8409                 do_div(min_free, dev_min);
8410         }
8411
8412         /* We need to do this so that we can look at pending chunks */
8413         trans = btrfs_join_transaction(root);
8414         if (IS_ERR(trans)) {
8415                 ret = PTR_ERR(trans);
8416                 goto out;
8417         }
8418
8419         mutex_lock(&root->fs_info->chunk_mutex);
8420         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
8421                 u64 dev_offset;
8422
8423                 /*
8424                  * check to make sure we can actually find a chunk with enough
8425                  * space to fit our block group in.
8426                  */
8427                 if (device->total_bytes > device->bytes_used + min_free &&
8428                     !device->is_tgtdev_for_dev_replace) {
8429                         ret = find_free_dev_extent(trans, device, min_free,
8430                                                    &dev_offset, NULL);
8431                         if (!ret)
8432                                 dev_nr++;
8433
8434                         if (dev_nr >= dev_min)
8435                                 break;
8436
8437                         ret = -1;
8438                 }
8439         }
8440         mutex_unlock(&root->fs_info->chunk_mutex);
8441         btrfs_end_transaction(trans, root);
8442 out:
8443         btrfs_put_block_group(block_group);
8444         return ret;
8445 }
8446
8447 static int find_first_block_group(struct btrfs_root *root,
8448                 struct btrfs_path *path, struct btrfs_key *key)
8449 {
8450         int ret = 0;
8451         struct btrfs_key found_key;
8452         struct extent_buffer *leaf;
8453         int slot;
8454
8455         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
8456         if (ret < 0)
8457                 goto out;
8458
8459         while (1) {
8460                 slot = path->slots[0];
8461                 leaf = path->nodes[0];
8462                 if (slot >= btrfs_header_nritems(leaf)) {
8463                         ret = btrfs_next_leaf(root, path);
8464                         if (ret == 0)
8465                                 continue;
8466                         if (ret < 0)
8467                                 goto out;
8468                         break;
8469                 }
8470                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
8471
8472                 if (found_key.objectid >= key->objectid &&
8473                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
8474                         ret = 0;
8475                         goto out;
8476                 }
8477                 path->slots[0]++;
8478         }
8479 out:
8480         return ret;
8481 }
8482
8483 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
8484 {
8485         struct btrfs_block_group_cache *block_group;
8486         u64 last = 0;
8487
8488         while (1) {
8489                 struct inode *inode;
8490
8491                 block_group = btrfs_lookup_first_block_group(info, last);
8492                 while (block_group) {
8493                         spin_lock(&block_group->lock);
8494                         if (block_group->iref)
8495                                 break;
8496                         spin_unlock(&block_group->lock);
8497                         block_group = next_block_group(info->tree_root,
8498                                                        block_group);
8499                 }
8500                 if (!block_group) {
8501                         if (last == 0)
8502                                 break;
8503                         last = 0;
8504                         continue;
8505                 }
8506
8507                 inode = block_group->inode;
8508                 block_group->iref = 0;
8509                 block_group->inode = NULL;
8510                 spin_unlock(&block_group->lock);
8511                 iput(inode);
8512                 last = block_group->key.objectid + block_group->key.offset;
8513                 btrfs_put_block_group(block_group);
8514         }
8515 }
8516
8517 int btrfs_free_block_groups(struct btrfs_fs_info *info)
8518 {
8519         struct btrfs_block_group_cache *block_group;
8520         struct btrfs_space_info *space_info;
8521         struct btrfs_caching_control *caching_ctl;
8522         struct rb_node *n;
8523
8524         down_write(&info->commit_root_sem);
8525         while (!list_empty(&info->caching_block_groups)) {
8526                 caching_ctl = list_entry(info->caching_block_groups.next,
8527                                          struct btrfs_caching_control, list);
8528                 list_del(&caching_ctl->list);
8529                 put_caching_control(caching_ctl);
8530         }
8531         up_write(&info->commit_root_sem);
8532
8533         spin_lock(&info->block_group_cache_lock);
8534         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
8535                 block_group = rb_entry(n, struct btrfs_block_group_cache,
8536                                        cache_node);
8537                 rb_erase(&block_group->cache_node,
8538                          &info->block_group_cache_tree);
8539                 spin_unlock(&info->block_group_cache_lock);
8540
8541                 down_write(&block_group->space_info->groups_sem);
8542                 list_del(&block_group->list);
8543                 up_write(&block_group->space_info->groups_sem);
8544
8545                 if (block_group->cached == BTRFS_CACHE_STARTED)
8546                         wait_block_group_cache_done(block_group);
8547
8548                 /*
8549                  * We haven't cached this block group, which means we could
8550                  * possibly have excluded extents on this block group.
8551                  */
8552                 if (block_group->cached == BTRFS_CACHE_NO ||
8553                     block_group->cached == BTRFS_CACHE_ERROR)
8554                         free_excluded_extents(info->extent_root, block_group);
8555
8556                 btrfs_remove_free_space_cache(block_group);
8557                 btrfs_put_block_group(block_group);
8558
8559                 spin_lock(&info->block_group_cache_lock);
8560         }
8561         spin_unlock(&info->block_group_cache_lock);
8562
8563         /* now that all the block groups are freed, go through and
8564          * free all the space_info structs.  This is only called during
8565          * the final stages of unmount, and so we know nobody is
8566          * using them.  We call synchronize_rcu() once before we start,
8567          * just to be on the safe side.
8568          */
8569         synchronize_rcu();
8570
8571         release_global_block_rsv(info);
8572
8573         while (!list_empty(&info->space_info)) {
8574                 int i;
8575
8576                 space_info = list_entry(info->space_info.next,
8577                                         struct btrfs_space_info,
8578                                         list);
8579                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
8580                         if (WARN_ON(space_info->bytes_pinned > 0 ||
8581                             space_info->bytes_reserved > 0 ||
8582                             space_info->bytes_may_use > 0)) {
8583                                 dump_space_info(space_info, 0, 0);
8584                         }
8585                 }
8586                 list_del(&space_info->list);
8587                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
8588                         struct kobject *kobj;
8589                         kobj = &space_info->block_group_kobjs[i];
8590                         if (kobj->parent) {
8591                                 kobject_del(kobj);
8592                                 kobject_put(kobj);
8593                         }
8594                 }
8595                 kobject_del(&space_info->kobj);
8596                 kobject_put(&space_info->kobj);
8597         }
8598         return 0;
8599 }
8600
8601 static void __link_block_group(struct btrfs_space_info *space_info,
8602                                struct btrfs_block_group_cache *cache)
8603 {
8604         int index = get_block_group_index(cache);
8605         bool first = false;
8606
8607         down_write(&space_info->groups_sem);
8608         if (list_empty(&space_info->block_groups[index]))
8609                 first = true;
8610         list_add_tail(&cache->list, &space_info->block_groups[index]);
8611         up_write(&space_info->groups_sem);
8612
8613         if (first) {
8614                 struct kobject *kobj = &space_info->block_group_kobjs[index];
8615                 int ret;
8616
8617                 kobject_get(&space_info->kobj); /* put in release */
8618                 ret = kobject_add(kobj, &space_info->kobj, "%s",
8619                                   get_raid_name(index));
8620                 if (ret) {
8621                         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
8622                         kobject_put(&space_info->kobj);
8623                 }
8624         }
8625 }
8626
8627 static struct btrfs_block_group_cache *
8628 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
8629 {
8630         struct btrfs_block_group_cache *cache;
8631
8632         cache = kzalloc(sizeof(*cache), GFP_NOFS);
8633         if (!cache)
8634                 return NULL;
8635
8636         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
8637                                         GFP_NOFS);
8638         if (!cache->free_space_ctl) {
8639                 kfree(cache);
8640                 return NULL;
8641         }
8642
8643         cache->key.objectid = start;
8644         cache->key.offset = size;
8645         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
8646
8647         cache->sectorsize = root->sectorsize;
8648         cache->fs_info = root->fs_info;
8649         cache->full_stripe_len = btrfs_full_stripe_len(root,
8650                                                &root->fs_info->mapping_tree,
8651                                                start);
8652         atomic_set(&cache->count, 1);
8653         spin_lock_init(&cache->lock);
8654         INIT_LIST_HEAD(&cache->list);
8655         INIT_LIST_HEAD(&cache->cluster_list);
8656         INIT_LIST_HEAD(&cache->new_bg_list);
8657         btrfs_init_free_space_ctl(cache);
8658
8659         return cache;
8660 }
8661
8662 int btrfs_read_block_groups(struct btrfs_root *root)
8663 {
8664         struct btrfs_path *path;
8665         int ret;
8666         struct btrfs_block_group_cache *cache;
8667         struct btrfs_fs_info *info = root->fs_info;
8668         struct btrfs_space_info *space_info;
8669         struct btrfs_key key;
8670         struct btrfs_key found_key;
8671         struct extent_buffer *leaf;
8672         int need_clear = 0;
8673         u64 cache_gen;
8674
8675         root = info->extent_root;
8676         key.objectid = 0;
8677         key.offset = 0;
8678         btrfs_set_key_type(&key, BTRFS_BLOCK_GROUP_ITEM_KEY);
8679         path = btrfs_alloc_path();
8680         if (!path)
8681                 return -ENOMEM;
8682         path->reada = 1;
8683
8684         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
8685         if (btrfs_test_opt(root, SPACE_CACHE) &&
8686             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
8687                 need_clear = 1;
8688         if (btrfs_test_opt(root, CLEAR_CACHE))
8689                 need_clear = 1;
8690
8691         while (1) {
8692                 ret = find_first_block_group(root, path, &key);
8693                 if (ret > 0)
8694                         break;
8695                 if (ret != 0)
8696                         goto error;
8697
8698                 leaf = path->nodes[0];
8699                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
8700
8701                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
8702                                                        found_key.offset);
8703                 if (!cache) {
8704                         ret = -ENOMEM;
8705                         goto error;
8706                 }
8707
8708                 if (need_clear) {
8709                         /*
8710                          * When we mount with old space cache, we need to
8711                          * set BTRFS_DC_CLEAR and set dirty flag.
8712                          *
8713                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
8714                          *    truncate the old free space cache inode and
8715                          *    setup a new one.
8716                          * b) Setting 'dirty flag' makes sure that we flush
8717                          *    the new space cache info onto disk.
8718                          */
8719                         cache->disk_cache_state = BTRFS_DC_CLEAR;
8720                         if (btrfs_test_opt(root, SPACE_CACHE))
8721                                 cache->dirty = 1;
8722                 }
8723
8724                 read_extent_buffer(leaf, &cache->item,
8725                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
8726                                    sizeof(cache->item));
8727                 cache->flags = btrfs_block_group_flags(&cache->item);
8728
8729                 key.objectid = found_key.objectid + found_key.offset;
8730                 btrfs_release_path(path);
8731
8732                 /*
8733                  * We need to exclude the super stripes now so that the space
8734                  * info has super bytes accounted for, otherwise we'll think
8735                  * we have more space than we actually do.
8736                  */
8737                 ret = exclude_super_stripes(root, cache);
8738                 if (ret) {
8739                         /*
8740                          * We may have excluded something, so call this just in
8741                          * case.
8742                          */
8743                         free_excluded_extents(root, cache);
8744                         btrfs_put_block_group(cache);
8745                         goto error;
8746                 }
8747
8748                 /*
8749                  * check for two cases, either we are full, and therefore
8750                  * don't need to bother with the caching work since we won't
8751                  * find any space, or we are empty, and we can just add all
8752                  * the space in and be done with it.  This saves us _alot_ of
8753                  * time, particularly in the full case.
8754                  */
8755                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
8756                         cache->last_byte_to_unpin = (u64)-1;
8757                         cache->cached = BTRFS_CACHE_FINISHED;
8758                         free_excluded_extents(root, cache);
8759                 } else if (btrfs_block_group_used(&cache->item) == 0) {
8760                         cache->last_byte_to_unpin = (u64)-1;
8761                         cache->cached = BTRFS_CACHE_FINISHED;
8762                         add_new_free_space(cache, root->fs_info,
8763                                            found_key.objectid,
8764                                            found_key.objectid +
8765                                            found_key.offset);
8766                         free_excluded_extents(root, cache);
8767                 }
8768
8769                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
8770                 if (ret) {
8771                         btrfs_remove_free_space_cache(cache);
8772                         btrfs_put_block_group(cache);
8773                         goto error;
8774                 }
8775
8776                 ret = update_space_info(info, cache->flags, found_key.offset,
8777                                         btrfs_block_group_used(&cache->item),
8778                                         &space_info);
8779                 if (ret) {
8780                         btrfs_remove_free_space_cache(cache);
8781                         spin_lock(&info->block_group_cache_lock);
8782                         rb_erase(&cache->cache_node,
8783                                  &info->block_group_cache_tree);
8784                         spin_unlock(&info->block_group_cache_lock);
8785                         btrfs_put_block_group(cache);
8786                         goto error;
8787                 }
8788
8789                 cache->space_info = space_info;
8790                 spin_lock(&cache->space_info->lock);
8791                 cache->space_info->bytes_readonly += cache->bytes_super;
8792                 spin_unlock(&cache->space_info->lock);
8793
8794                 __link_block_group(space_info, cache);
8795
8796                 set_avail_alloc_bits(root->fs_info, cache->flags);
8797                 if (btrfs_chunk_readonly(root, cache->key.objectid))
8798                         set_block_group_ro(cache, 1);
8799         }
8800
8801         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
8802                 if (!(get_alloc_profile(root, space_info->flags) &
8803                       (BTRFS_BLOCK_GROUP_RAID10 |
8804                        BTRFS_BLOCK_GROUP_RAID1 |
8805                        BTRFS_BLOCK_GROUP_RAID5 |
8806                        BTRFS_BLOCK_GROUP_RAID6 |
8807                        BTRFS_BLOCK_GROUP_DUP)))
8808                         continue;
8809                 /*
8810                  * avoid allocating from un-mirrored block group if there are
8811                  * mirrored block groups.
8812                  */
8813                 list_for_each_entry(cache,
8814                                 &space_info->block_groups[BTRFS_RAID_RAID0],
8815                                 list)
8816                         set_block_group_ro(cache, 1);
8817                 list_for_each_entry(cache,
8818                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
8819                                 list)
8820                         set_block_group_ro(cache, 1);
8821         }
8822
8823         init_global_block_rsv(info);
8824         ret = 0;
8825 error:
8826         btrfs_free_path(path);
8827         return ret;
8828 }
8829
8830 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
8831                                        struct btrfs_root *root)
8832 {
8833         struct btrfs_block_group_cache *block_group, *tmp;
8834         struct btrfs_root *extent_root = root->fs_info->extent_root;
8835         struct btrfs_block_group_item item;
8836         struct btrfs_key key;
8837         int ret = 0;
8838
8839         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs,
8840                                  new_bg_list) {
8841                 list_del_init(&block_group->new_bg_list);
8842
8843                 if (ret)
8844                         continue;
8845
8846                 spin_lock(&block_group->lock);
8847                 memcpy(&item, &block_group->item, sizeof(item));
8848                 memcpy(&key, &block_group->key, sizeof(key));
8849                 spin_unlock(&block_group->lock);
8850
8851                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
8852                                         sizeof(item));
8853                 if (ret)
8854                         btrfs_abort_transaction(trans, extent_root, ret);
8855                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
8856                                                key.objectid, key.offset);
8857                 if (ret)
8858                         btrfs_abort_transaction(trans, extent_root, ret);
8859         }
8860 }
8861
8862 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
8863                            struct btrfs_root *root, u64 bytes_used,
8864                            u64 type, u64 chunk_objectid, u64 chunk_offset,
8865                            u64 size)
8866 {
8867         int ret;
8868         struct btrfs_root *extent_root;
8869         struct btrfs_block_group_cache *cache;
8870
8871         extent_root = root->fs_info->extent_root;
8872
8873         btrfs_set_log_full_commit(root->fs_info, trans);
8874
8875         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
8876         if (!cache)
8877                 return -ENOMEM;
8878
8879         btrfs_set_block_group_used(&cache->item, bytes_used);
8880         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
8881         btrfs_set_block_group_flags(&cache->item, type);
8882
8883         cache->flags = type;
8884         cache->last_byte_to_unpin = (u64)-1;
8885         cache->cached = BTRFS_CACHE_FINISHED;
8886         ret = exclude_super_stripes(root, cache);
8887         if (ret) {
8888                 /*
8889                  * We may have excluded something, so call this just in
8890                  * case.
8891                  */
8892                 free_excluded_extents(root, cache);
8893                 btrfs_put_block_group(cache);
8894                 return ret;
8895         }
8896
8897         add_new_free_space(cache, root->fs_info, chunk_offset,
8898                            chunk_offset + size);
8899
8900         free_excluded_extents(root, cache);
8901
8902         ret = btrfs_add_block_group_cache(root->fs_info, cache);
8903         if (ret) {
8904                 btrfs_remove_free_space_cache(cache);
8905                 btrfs_put_block_group(cache);
8906                 return ret;
8907         }
8908
8909         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
8910                                 &cache->space_info);
8911         if (ret) {
8912                 btrfs_remove_free_space_cache(cache);
8913                 spin_lock(&root->fs_info->block_group_cache_lock);
8914                 rb_erase(&cache->cache_node,
8915                          &root->fs_info->block_group_cache_tree);
8916                 spin_unlock(&root->fs_info->block_group_cache_lock);
8917                 btrfs_put_block_group(cache);
8918                 return ret;
8919         }
8920         update_global_block_rsv(root->fs_info);
8921
8922         spin_lock(&cache->space_info->lock);
8923         cache->space_info->bytes_readonly += cache->bytes_super;
8924         spin_unlock(&cache->space_info->lock);
8925
8926         __link_block_group(cache->space_info, cache);
8927
8928         list_add_tail(&cache->new_bg_list, &trans->new_bgs);
8929
8930         set_avail_alloc_bits(extent_root->fs_info, type);
8931
8932         return 0;
8933 }
8934
8935 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
8936 {
8937         u64 extra_flags = chunk_to_extended(flags) &
8938                                 BTRFS_EXTENDED_PROFILE_MASK;
8939
8940         write_seqlock(&fs_info->profiles_lock);
8941         if (flags & BTRFS_BLOCK_GROUP_DATA)
8942                 fs_info->avail_data_alloc_bits &= ~extra_flags;
8943         if (flags & BTRFS_BLOCK_GROUP_METADATA)
8944                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
8945         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
8946                 fs_info->avail_system_alloc_bits &= ~extra_flags;
8947         write_sequnlock(&fs_info->profiles_lock);
8948 }
8949
8950 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
8951                              struct btrfs_root *root, u64 group_start)
8952 {
8953         struct btrfs_path *path;
8954         struct btrfs_block_group_cache *block_group;
8955         struct btrfs_free_cluster *cluster;
8956         struct btrfs_root *tree_root = root->fs_info->tree_root;
8957         struct btrfs_key key;
8958         struct inode *inode;
8959         int ret;
8960         int index;
8961         int factor;
8962
8963         root = root->fs_info->extent_root;
8964
8965         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
8966         BUG_ON(!block_group);
8967         BUG_ON(!block_group->ro);
8968
8969         /*
8970          * Free the reserved super bytes from this block group before
8971          * remove it.
8972          */
8973         free_excluded_extents(root, block_group);
8974
8975         memcpy(&key, &block_group->key, sizeof(key));
8976         index = get_block_group_index(block_group);
8977         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
8978                                   BTRFS_BLOCK_GROUP_RAID1 |
8979                                   BTRFS_BLOCK_GROUP_RAID10))
8980                 factor = 2;
8981         else
8982                 factor = 1;
8983
8984         /* make sure this block group isn't part of an allocation cluster */
8985         cluster = &root->fs_info->data_alloc_cluster;
8986         spin_lock(&cluster->refill_lock);
8987         btrfs_return_cluster_to_free_space(block_group, cluster);
8988         spin_unlock(&cluster->refill_lock);
8989
8990         /*
8991          * make sure this block group isn't part of a metadata
8992          * allocation cluster
8993          */
8994         cluster = &root->fs_info->meta_alloc_cluster;
8995         spin_lock(&cluster->refill_lock);
8996         btrfs_return_cluster_to_free_space(block_group, cluster);
8997         spin_unlock(&cluster->refill_lock);
8998
8999         path = btrfs_alloc_path();
9000         if (!path) {
9001                 ret = -ENOMEM;
9002                 goto out;
9003         }
9004
9005         inode = lookup_free_space_inode(tree_root, block_group, path);
9006         if (!IS_ERR(inode)) {
9007                 ret = btrfs_orphan_add(trans, inode);
9008                 if (ret) {
9009                         btrfs_add_delayed_iput(inode);
9010                         goto out;
9011                 }
9012                 clear_nlink(inode);
9013                 /* One for the block groups ref */
9014                 spin_lock(&block_group->lock);
9015                 if (block_group->iref) {
9016                         block_group->iref = 0;
9017                         block_group->inode = NULL;
9018                         spin_unlock(&block_group->lock);
9019                         iput(inode);
9020                 } else {
9021                         spin_unlock(&block_group->lock);
9022                 }
9023                 /* One for our lookup ref */
9024                 btrfs_add_delayed_iput(inode);
9025         }
9026
9027         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
9028         key.offset = block_group->key.objectid;
9029         key.type = 0;
9030
9031         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
9032         if (ret < 0)
9033                 goto out;
9034         if (ret > 0)
9035                 btrfs_release_path(path);
9036         if (ret == 0) {
9037                 ret = btrfs_del_item(trans, tree_root, path);
9038                 if (ret)
9039                         goto out;
9040                 btrfs_release_path(path);
9041         }
9042
9043         spin_lock(&root->fs_info->block_group_cache_lock);
9044         rb_erase(&block_group->cache_node,
9045                  &root->fs_info->block_group_cache_tree);
9046
9047         if (root->fs_info->first_logical_byte == block_group->key.objectid)
9048                 root->fs_info->first_logical_byte = (u64)-1;
9049         spin_unlock(&root->fs_info->block_group_cache_lock);
9050
9051         down_write(&block_group->space_info->groups_sem);
9052         /*
9053          * we must use list_del_init so people can check to see if they
9054          * are still on the list after taking the semaphore
9055          */
9056         list_del_init(&block_group->list);
9057         if (list_empty(&block_group->space_info->block_groups[index])) {
9058                 kobject_del(&block_group->space_info->block_group_kobjs[index]);
9059                 kobject_put(&block_group->space_info->block_group_kobjs[index]);
9060                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
9061         }
9062         up_write(&block_group->space_info->groups_sem);
9063
9064         if (block_group->cached == BTRFS_CACHE_STARTED)
9065                 wait_block_group_cache_done(block_group);
9066
9067         btrfs_remove_free_space_cache(block_group);
9068
9069         spin_lock(&block_group->space_info->lock);
9070         block_group->space_info->total_bytes -= block_group->key.offset;
9071         block_group->space_info->bytes_readonly -= block_group->key.offset;
9072         block_group->space_info->disk_total -= block_group->key.offset * factor;
9073         spin_unlock(&block_group->space_info->lock);
9074
9075         memcpy(&key, &block_group->key, sizeof(key));
9076
9077         btrfs_clear_space_info_full(root->fs_info);
9078
9079         btrfs_put_block_group(block_group);
9080         btrfs_put_block_group(block_group);
9081
9082         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
9083         if (ret > 0)
9084                 ret = -EIO;
9085         if (ret < 0)
9086                 goto out;
9087
9088         ret = btrfs_del_item(trans, root, path);
9089 out:
9090         btrfs_free_path(path);
9091         return ret;
9092 }
9093
9094 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
9095 {
9096         struct btrfs_space_info *space_info;
9097         struct btrfs_super_block *disk_super;
9098         u64 features;
9099         u64 flags;
9100         int mixed = 0;
9101         int ret;
9102
9103         disk_super = fs_info->super_copy;
9104         if (!btrfs_super_root(disk_super))
9105                 return 1;
9106
9107         features = btrfs_super_incompat_flags(disk_super);
9108         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
9109                 mixed = 1;
9110
9111         flags = BTRFS_BLOCK_GROUP_SYSTEM;
9112         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9113         if (ret)
9114                 goto out;
9115
9116         if (mixed) {
9117                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
9118                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9119         } else {
9120                 flags = BTRFS_BLOCK_GROUP_METADATA;
9121                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9122                 if (ret)
9123                         goto out;
9124
9125                 flags = BTRFS_BLOCK_GROUP_DATA;
9126                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
9127         }
9128 out:
9129         return ret;
9130 }
9131
9132 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
9133 {
9134         return unpin_extent_range(root, start, end);
9135 }
9136
9137 int btrfs_error_discard_extent(struct btrfs_root *root, u64 bytenr,
9138                                u64 num_bytes, u64 *actual_bytes)
9139 {
9140         return btrfs_discard_extent(root, bytenr, num_bytes, actual_bytes);
9141 }
9142
9143 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
9144 {
9145         struct btrfs_fs_info *fs_info = root->fs_info;
9146         struct btrfs_block_group_cache *cache = NULL;
9147         u64 group_trimmed;
9148         u64 start;
9149         u64 end;
9150         u64 trimmed = 0;
9151         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
9152         int ret = 0;
9153
9154         /*
9155          * try to trim all FS space, our block group may start from non-zero.
9156          */
9157         if (range->len == total_bytes)
9158                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
9159         else
9160                 cache = btrfs_lookup_block_group(fs_info, range->start);
9161
9162         while (cache) {
9163                 if (cache->key.objectid >= (range->start + range->len)) {
9164                         btrfs_put_block_group(cache);
9165                         break;
9166                 }
9167
9168                 start = max(range->start, cache->key.objectid);
9169                 end = min(range->start + range->len,
9170                                 cache->key.objectid + cache->key.offset);
9171
9172                 if (end - start >= range->minlen) {
9173                         if (!block_group_cache_done(cache)) {
9174                                 ret = cache_block_group(cache, 0);
9175                                 if (ret) {
9176                                         btrfs_put_block_group(cache);
9177                                         break;
9178                                 }
9179                                 ret = wait_block_group_cache_done(cache);
9180                                 if (ret) {
9181                                         btrfs_put_block_group(cache);
9182                                         break;
9183                                 }
9184                         }
9185                         ret = btrfs_trim_block_group(cache,
9186                                                      &group_trimmed,
9187                                                      start,
9188                                                      end,
9189                                                      range->minlen);
9190
9191                         trimmed += group_trimmed;
9192                         if (ret) {
9193                                 btrfs_put_block_group(cache);
9194                                 break;
9195                         }
9196                 }
9197
9198                 cache = next_block_group(fs_info->tree_root, cache);
9199         }
9200
9201         range->len = trimmed;
9202         return ret;
9203 }
9204
9205 /*
9206  * btrfs_{start,end}_write() is similar to mnt_{want, drop}_write(),
9207  * they are used to prevent the some tasks writing data into the page cache
9208  * by nocow before the subvolume is snapshoted, but flush the data into
9209  * the disk after the snapshot creation.
9210  */
9211 void btrfs_end_nocow_write(struct btrfs_root *root)
9212 {
9213         percpu_counter_dec(&root->subv_writers->counter);
9214         /*
9215          * Make sure counter is updated before we wake up
9216          * waiters.
9217          */
9218         smp_mb();
9219         if (waitqueue_active(&root->subv_writers->wait))
9220                 wake_up(&root->subv_writers->wait);
9221 }
9222
9223 int btrfs_start_nocow_write(struct btrfs_root *root)
9224 {
9225         if (unlikely(atomic_read(&root->will_be_snapshoted)))
9226                 return 0;
9227
9228         percpu_counter_inc(&root->subv_writers->counter);
9229         /*
9230          * Make sure counter is updated before we check for snapshot creation.
9231          */
9232         smp_mb();
9233         if (unlikely(atomic_read(&root->will_be_snapshoted))) {
9234                 btrfs_end_nocow_write(root);
9235                 return 0;
9236         }
9237         return 1;
9238 }