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 "free-space-tree.h"
  37 #include "math.h"
  38 #include "sysfs.h"
  39 #include "qgroup.h"
  40
  41 #undef SCRAMBLE_DELAYED_REFS
  42
  43 /*
  44  * control flags for do_chunk_alloc's force field
  45  * CHUNK_ALLOC_NO_FORCE means to only allocate a chunk
  46  * if we really need one.
  47  *
  48  * CHUNK_ALLOC_LIMITED means to only try and allocate one
  49  * if we have very few chunks already allocated.  This is
  50  * used as part of the clustering code to help make sure
  51  * we have a good pool of storage to cluster in, without
  52  * filling the FS with empty chunks
  53  *
  54  * CHUNK_ALLOC_FORCE means it must try to allocate one
  55  *
  56  */
  57 enum {
  58         CHUNK_ALLOC_NO_FORCE = 0,
  59         CHUNK_ALLOC_LIMITED = 1,
  60         CHUNK_ALLOC_FORCE = 2,
  61 };
  62
  63 /*
  64  * Control how reservations are dealt with.
  65  *
  66  * RESERVE_FREE - freeing a reservation.
  67  * RESERVE_ALLOC - allocating space and we need to update bytes_may_use for
  68  *   ENOSPC accounting
  69  * RESERVE_ALLOC_NO_ACCOUNT - allocating space and we should not update
  70  *   bytes_may_use as the ENOSPC accounting is done elsewhere
  71  */
  72 enum {
  73         RESERVE_FREE = 0,
  74         RESERVE_ALLOC = 1,
  75         RESERVE_ALLOC_NO_ACCOUNT = 2,
  76 };
  77
  78 static int update_block_group(struct btrfs_trans_handle *trans,
  79                               struct btrfs_root *root, u64 bytenr,
  80                               u64 num_bytes, int alloc);
  81 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
  82                                 struct btrfs_root *root,
  83                                 struct btrfs_delayed_ref_node *node, u64 parent,
  84                                 u64 root_objectid, u64 owner_objectid,
  85                                 u64 owner_offset, int refs_to_drop,
  86                                 struct btrfs_delayed_extent_op *extra_op);
  87 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
  88                                     struct extent_buffer *leaf,
  89                                     struct btrfs_extent_item *ei);
  90 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
  91                                       struct btrfs_root *root,
  92                                       u64 parent, u64 root_objectid,
  93                                       u64 flags, u64 owner, u64 offset,
  94                                       struct btrfs_key *ins, int ref_mod);
  95 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
  96                                      struct btrfs_root *root,
  97                                      u64 parent, u64 root_objectid,
  98                                      u64 flags, struct btrfs_disk_key *key,
  99                                      int level, struct btrfs_key *ins);
 100 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
 101                           struct btrfs_root *extent_root, u64 flags,
 102                           int force);
 103 static int find_next_key(struct btrfs_path *path, int level,
 104                          struct btrfs_key *key);
 105 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
 106                             int dump_block_groups);
 107 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
 108                                        u64 num_bytes, int reserve,
 109                                        int delalloc);
 110 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
 111                                u64 num_bytes);
 112 int btrfs_pin_extent(struct btrfs_root *root,
 113                      u64 bytenr, u64 num_bytes, int reserved);
 114
 115 static noinline int
 116 block_group_cache_done(struct btrfs_block_group_cache *cache)
 117 {
 118         smp_mb();
 119         return cache->cached == BTRFS_CACHE_FINISHED ||
 120                 cache->cached == BTRFS_CACHE_ERROR;
 121 }
 122
 123 static int block_group_bits(struct btrfs_block_group_cache *cache, u64 bits)
 124 {
 125         return (cache->flags & bits) == bits;
 126 }
 127
 128 void btrfs_get_block_group(struct btrfs_block_group_cache *cache)
 129 {
 130         atomic_inc(&cache->count);
 131 }
 132
 133 void btrfs_put_block_group(struct btrfs_block_group_cache *cache)
 134 {
 135         if (atomic_dec_and_test(&cache->count)) {
 136                 WARN_ON(cache->pinned > 0);
 137                 WARN_ON(cache->reserved > 0);
 138                 kfree(cache->free_space_ctl);
 139                 kfree(cache);
 140         }
 141 }
 142
 143 /*
 144  * this adds the block group to the fs_info rb tree for the block group
 145  * cache
 146  */
 147 static int btrfs_add_block_group_cache(struct btrfs_fs_info *info,
 148                                 struct btrfs_block_group_cache *block_group)
 149 {
 150         struct rb_node **p;
 151         struct rb_node *parent = NULL;
 152         struct btrfs_block_group_cache *cache;
 153
 154         spin_lock(&info->block_group_cache_lock);
 155         p = &info->block_group_cache_tree.rb_node;
 156
 157         while (*p) {
 158                 parent = *p;
 159                 cache = rb_entry(parent, struct btrfs_block_group_cache,
 160                                  cache_node);
 161                 if (block_group->key.objectid < cache->key.objectid) {
 162                         p = &(*p)->rb_left;
 163                 } else if (block_group->key.objectid > cache->key.objectid) {
 164                         p = &(*p)->rb_right;
 165                 } else {
 166                         spin_unlock(&info->block_group_cache_lock);
 167                         return -EEXIST;
 168                 }
 169         }
 170
 171         rb_link_node(&block_group->cache_node, parent, p);
 172         rb_insert_color(&block_group->cache_node,
 173                         &info->block_group_cache_tree);
 174
 175         if (info->first_logical_byte > block_group->key.objectid)
 176                 info->first_logical_byte = block_group->key.objectid;
 177
 178         spin_unlock(&info->block_group_cache_lock);
 179
 180         return 0;
 181 }
 182
 183 /*
 184  * This will return the block group at or after bytenr if contains is 0, else
 185  * it will return the block group that contains the bytenr
 186  */
 187 static struct btrfs_block_group_cache *
 188 block_group_cache_tree_search(struct btrfs_fs_info *info, u64 bytenr,
 189                               int contains)
 190 {
 191         struct btrfs_block_group_cache *cache, *ret = NULL;
 192         struct rb_node *n;
 193         u64 end, start;
 194
 195         spin_lock(&info->block_group_cache_lock);
 196         n = info->block_group_cache_tree.rb_node;
 197
 198         while (n) {
 199                 cache = rb_entry(n, struct btrfs_block_group_cache,
 200                                  cache_node);
 201                 end = cache->key.objectid + cache->key.offset - 1;
 202                 start = cache->key.objectid;
 203
 204                 if (bytenr < start) {
 205                         if (!contains && (!ret || start < ret->key.objectid))
 206                                 ret = cache;
 207                         n = n->rb_left;
 208                 } else if (bytenr > start) {
 209                         if (contains && bytenr <= end) {
 210                                 ret = cache;
 211                                 break;
 212                         }
 213                         n = n->rb_right;
 214                 } else {
 215                         ret = cache;
 216                         break;
 217                 }
 218         }
 219         if (ret) {
 220                 btrfs_get_block_group(ret);
 221                 if (bytenr == 0 && info->first_logical_byte > ret->key.objectid)
 222                         info->first_logical_byte = ret->key.objectid;
 223         }
 224         spin_unlock(&info->block_group_cache_lock);
 225
 226         return ret;
 227 }
 228
 229 static int add_excluded_extent(struct btrfs_root *root,
 230                                u64 start, u64 num_bytes)
 231 {
 232         u64 end = start + num_bytes - 1;
 233         set_extent_bits(&root->fs_info->freed_extents[0],
 234                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 235         set_extent_bits(&root->fs_info->freed_extents[1],
 236                         start, end, EXTENT_UPTODATE, GFP_NOFS);
 237         return 0;
 238 }
 239
 240 static void free_excluded_extents(struct btrfs_root *root,
 241                                   struct btrfs_block_group_cache *cache)
 242 {
 243         u64 start, end;
 244
 245         start = cache->key.objectid;
 246         end = start + cache->key.offset - 1;
 247
 248         clear_extent_bits(&root->fs_info->freed_extents[0],
 249                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 250         clear_extent_bits(&root->fs_info->freed_extents[1],
 251                           start, end, EXTENT_UPTODATE, GFP_NOFS);
 252 }
 253
 254 static int exclude_super_stripes(struct btrfs_root *root,
 255                                  struct btrfs_block_group_cache *cache)
 256 {
 257         u64 bytenr;
 258         u64 *logical;
 259         int stripe_len;
 260         int i, nr, ret;
 261
 262         if (cache->key.objectid < BTRFS_SUPER_INFO_OFFSET) {
 263                 stripe_len = BTRFS_SUPER_INFO_OFFSET - cache->key.objectid;
 264                 cache->bytes_super += stripe_len;
 265                 ret = add_excluded_extent(root, cache->key.objectid,
 266                                           stripe_len);
 267                 if (ret)
 268                         return ret;
 269         }
 270
 271         for (i = 0; i < BTRFS_SUPER_MIRROR_MAX; i++) {
 272                 bytenr = btrfs_sb_offset(i);
 273                 ret = btrfs_rmap_block(&root->fs_info->mapping_tree,
 274                                        cache->key.objectid, bytenr,
 275                                        0, &logical, &nr, &stripe_len);
 276                 if (ret)
 277                         return ret;
 278
 279                 while (nr--) {
 280                         u64 start, len;
 281
 282                         if (logical[nr] > cache->key.objectid +
 283                             cache->key.offset)
 284                                 continue;
 285
 286                         if (logical[nr] + stripe_len <= cache->key.objectid)
 287                                 continue;
 288
 289                         start = logical[nr];
 290                         if (start < cache->key.objectid) {
 291                                 start = cache->key.objectid;
 292                                 len = (logical[nr] + stripe_len) - start;
 293                         } else {
 294                                 len = min_t(u64, stripe_len,
 295                                             cache->key.objectid +
 296                                             cache->key.offset - start);
 297                         }
 298
 299                         cache->bytes_super += len;
 300                         ret = add_excluded_extent(root, start, len);
 301                         if (ret) {
 302                                 kfree(logical);
 303                                 return ret;
 304                         }
 305                 }
 306
 307                 kfree(logical);
 308         }
 309         return 0;
 310 }
 311
 312 static struct btrfs_caching_control *
 313 get_caching_control(struct btrfs_block_group_cache *cache)
 314 {
 315         struct btrfs_caching_control *ctl;
 316
 317         spin_lock(&cache->lock);
 318         if (!cache->caching_ctl) {
 319                 spin_unlock(&cache->lock);
 320                 return NULL;
 321         }
 322
 323         ctl = cache->caching_ctl;
 324         atomic_inc(&ctl->count);
 325         spin_unlock(&cache->lock);
 326         return ctl;
 327 }
 328
 329 static void put_caching_control(struct btrfs_caching_control *ctl)
 330 {
 331         if (atomic_dec_and_test(&ctl->count))
 332                 kfree(ctl);
 333 }
 334
 335 #ifdef CONFIG_BTRFS_DEBUG
 336 static void fragment_free_space(struct btrfs_root *root,
 337                                 struct btrfs_block_group_cache *block_group)
 338 {
 339         u64 start = block_group->key.objectid;
 340         u64 len = block_group->key.offset;
 341         u64 chunk = block_group->flags & BTRFS_BLOCK_GROUP_METADATA ?
 342                 root->nodesize : root->sectorsize;
 343         u64 step = chunk << 1;
 344
 345         while (len > chunk) {
 346                 btrfs_remove_free_space(block_group, start, chunk);
 347                 start += step;
 348                 if (len < step)
 349                         len = 0;
 350                 else
 351                         len -= step;
 352         }
 353 }
 354 #endif
 355
 356 /*
 357  * this is only called by cache_block_group, since we could have freed extents
 358  * we need to check the pinned_extents for any extents that can't be used yet
 359  * since their free space will be released as soon as the transaction commits.
 360  */
 361 u64 add_new_free_space(struct btrfs_block_group_cache *block_group,
 362                        struct btrfs_fs_info *info, u64 start, u64 end)
 363 {
 364         u64 extent_start, extent_end, size, total_added = 0;
 365         int ret;
 366
 367         while (start < end) {
 368                 ret = find_first_extent_bit(info->pinned_extents, start,
 369                                             &extent_start, &extent_end,
 370                                             EXTENT_DIRTY | EXTENT_UPTODATE,
 371                                             NULL);
 372                 if (ret)
 373                         break;
 374
 375                 if (extent_start <= start) {
 376                         start = extent_end + 1;
 377                 } else if (extent_start > start && extent_start < end) {
 378                         size = extent_start - start;
 379                         total_added += size;
 380                         ret = btrfs_add_free_space(block_group, start,
 381                                                    size);
 382                         BUG_ON(ret); /* -ENOMEM or logic error */
 383                         start = extent_end + 1;
 384                 } else {
 385                         break;
 386                 }
 387         }
 388
 389         if (start < end) {
 390                 size = end - start;
 391                 total_added += size;
 392                 ret = btrfs_add_free_space(block_group, start, size);
 393                 BUG_ON(ret); /* -ENOMEM or logic error */
 394         }
 395
 396         return total_added;
 397 }
 398
 399 static int load_extent_tree_free(struct btrfs_caching_control *caching_ctl)
 400 {
 401         struct btrfs_block_group_cache *block_group;
 402         struct btrfs_fs_info *fs_info;
 403         struct btrfs_root *extent_root;
 404         struct btrfs_path *path;
 405         struct extent_buffer *leaf;
 406         struct btrfs_key key;
 407         u64 total_found = 0;
 408         u64 last = 0;
 409         u32 nritems;
 410         int ret;
 411         bool wakeup = true;
 412
 413         block_group = caching_ctl->block_group;
 414         fs_info = block_group->fs_info;
 415         extent_root = fs_info->extent_root;
 416
 417         path = btrfs_alloc_path();
 418         if (!path)
 419                 return -ENOMEM;
 420
 421         last = max_t(u64, block_group->key.objectid, BTRFS_SUPER_INFO_OFFSET);
 422
 423 #ifdef CONFIG_BTRFS_DEBUG
 424         /*
 425          * If we're fragmenting we don't want to make anybody think we can
 426          * allocate from this block group until we've had a chance to fragment
 427          * the free space.
 428          */
 429         if (btrfs_should_fragment_free_space(extent_root, block_group))
 430                 wakeup = false;
 431 #endif
 432         /*
 433          * We don't want to deadlock with somebody trying to allocate a new
 434          * extent for the extent root while also trying to search the extent
 435          * root to add free space.  So we skip locking and search the commit
 436          * root, since its read-only
 437          */
 438         path->skip_locking = 1;
 439         path->search_commit_root = 1;
 440         path->reada = 1;
 441
 442         key.objectid = last;
 443         key.offset = 0;
 444         key.type = BTRFS_EXTENT_ITEM_KEY;
 445
 446 next:
 447         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
 448         if (ret < 0)
 449                 goto out;
 450
 451         leaf = path->nodes[0];
 452         nritems = btrfs_header_nritems(leaf);
 453
 454         while (1) {
 455                 if (btrfs_fs_closing(fs_info) > 1) {
 456                         last = (u64)-1;
 457                         break;
 458                 }
 459
 460                 if (path->slots[0] < nritems) {
 461                         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
 462                 } else {
 463                         ret = find_next_key(path, 0, &key);
 464                         if (ret)
 465                                 break;
 466
 467                         if (need_resched() ||
 468                             rwsem_is_contended(&fs_info->commit_root_sem)) {
 469                                 if (wakeup)
 470                                         caching_ctl->progress = last;
 471                                 btrfs_release_path(path);
 472                                 up_read(&fs_info->commit_root_sem);
 473                                 mutex_unlock(&caching_ctl->mutex);
 474                                 cond_resched();
 475                                 mutex_lock(&caching_ctl->mutex);
 476                                 down_read(&fs_info->commit_root_sem);
 477                                 goto next;
 478                         }
 479
 480                         ret = btrfs_next_leaf(extent_root, path);
 481                         if (ret < 0)
 482                                 goto out;
 483                         if (ret)
 484                                 break;
 485                         leaf = path->nodes[0];
 486                         nritems = btrfs_header_nritems(leaf);
 487                         continue;
 488                 }
 489
 490                 if (key.objectid < last) {
 491                         key.objectid = last;
 492                         key.offset = 0;
 493                         key.type = BTRFS_EXTENT_ITEM_KEY;
 494
 495                         if (wakeup)
 496                                 caching_ctl->progress = last;
 497                         btrfs_release_path(path);
 498                         goto next;
 499                 }
 500
 501                 if (key.objectid < block_group->key.objectid) {
 502                         path->slots[0]++;
 503                         continue;
 504                 }
 505
 506                 if (key.objectid >= block_group->key.objectid +
 507                     block_group->key.offset)
 508                         break;
 509
 510                 if (key.type == BTRFS_EXTENT_ITEM_KEY ||
 511                     key.type == BTRFS_METADATA_ITEM_KEY) {
 512                         total_found += add_new_free_space(block_group,
 513                                                           fs_info, last,
 514                                                           key.objectid);
 515                         if (key.type == BTRFS_METADATA_ITEM_KEY)
 516                                 last = key.objectid +
 517                                         fs_info->tree_root->nodesize;
 518                         else
 519                                 last = key.objectid + key.offset;
 520
 521                         if (total_found > CACHING_CTL_WAKE_UP) {
 522                                 total_found = 0;
 523                                 if (wakeup)
 524                                         wake_up(&caching_ctl->wait);
 525                         }
 526                 }
 527                 path->slots[0]++;
 528         }
 529         ret = 0;
 530
 531         total_found += add_new_free_space(block_group, fs_info, last,
 532                                           block_group->key.objectid +
 533                                           block_group->key.offset);
 534         caching_ctl->progress = (u64)-1;
 535
 536 out:
 537         btrfs_free_path(path);
 538         return ret;
 539 }
 540
 541 static noinline void caching_thread(struct btrfs_work *work)
 542 {
 543         struct btrfs_block_group_cache *block_group;
 544         struct btrfs_fs_info *fs_info;
 545         struct btrfs_caching_control *caching_ctl;
 546         struct btrfs_root *extent_root;
 547         int ret;
 548
 549         caching_ctl = container_of(work, struct btrfs_caching_control, work);
 550         block_group = caching_ctl->block_group;
 551         fs_info = block_group->fs_info;
 552         extent_root = fs_info->extent_root;
 553
 554         mutex_lock(&caching_ctl->mutex);
 555         down_read(&fs_info->commit_root_sem);
 556
 557         if (btrfs_fs_compat_ro(fs_info, FREE_SPACE_TREE))
 558                 ret = load_free_space_tree(caching_ctl);
 559         else
 560                 ret = load_extent_tree_free(caching_ctl);
 561
 562         spin_lock(&block_group->lock);
 563         block_group->caching_ctl = NULL;
 564         block_group->cached = ret ? BTRFS_CACHE_ERROR : BTRFS_CACHE_FINISHED;
 565         spin_unlock(&block_group->lock);
 566
 567 #ifdef CONFIG_BTRFS_DEBUG
 568         if (btrfs_should_fragment_free_space(extent_root, block_group)) {
 569                 u64 bytes_used;
 570
 571                 spin_lock(&block_group->space_info->lock);
 572                 spin_lock(&block_group->lock);
 573                 bytes_used = block_group->key.offset -
 574                         btrfs_block_group_used(&block_group->item);
 575                 block_group->space_info->bytes_used += bytes_used >> 1;
 576                 spin_unlock(&block_group->lock);
 577                 spin_unlock(&block_group->space_info->lock);
 578                 fragment_free_space(extent_root, block_group);
 579         }
 580 #endif
 581
 582         caching_ctl->progress = (u64)-1;
 583
 584         up_read(&fs_info->commit_root_sem);
 585         free_excluded_extents(fs_info->extent_root, block_group);
 586         mutex_unlock(&caching_ctl->mutex);
 587
 588         wake_up(&caching_ctl->wait);
 589
 590         put_caching_control(caching_ctl);
 591         btrfs_put_block_group(block_group);
 592 }
 593
 594 static int cache_block_group(struct btrfs_block_group_cache *cache,
 595                              int load_cache_only)
 596 {
 597         DEFINE_WAIT(wait);
 598         struct btrfs_fs_info *fs_info = cache->fs_info;
 599         struct btrfs_caching_control *caching_ctl;
 600         int ret = 0;
 601
 602         caching_ctl = kzalloc(sizeof(*caching_ctl), GFP_NOFS);
 603         if (!caching_ctl)
 604                 return -ENOMEM;
 605
 606         INIT_LIST_HEAD(&caching_ctl->list);
 607         mutex_init(&caching_ctl->mutex);
 608         init_waitqueue_head(&caching_ctl->wait);
 609         caching_ctl->block_group = cache;
 610         caching_ctl->progress = cache->key.objectid;
 611         atomic_set(&caching_ctl->count, 1);
 612         btrfs_init_work(&caching_ctl->work, btrfs_cache_helper,
 613                         caching_thread, NULL, NULL);
 614
 615         spin_lock(&cache->lock);
 616         /*
 617          * This should be a rare occasion, but this could happen I think in the
 618          * case where one thread starts to load the space cache info, and then
 619          * some other thread starts a transaction commit which tries to do an
 620          * allocation while the other thread is still loading the space cache
 621          * info.  The previous loop should have kept us from choosing this block
 622          * group, but if we've moved to the state where we will wait on caching
 623          * block groups we need to first check if we're doing a fast load here,
 624          * so we can wait for it to finish, otherwise we could end up allocating
 625          * from a block group who's cache gets evicted for one reason or
 626          * another.
 627          */
 628         while (cache->cached == BTRFS_CACHE_FAST) {
 629                 struct btrfs_caching_control *ctl;
 630
 631                 ctl = cache->caching_ctl;
 632                 atomic_inc(&ctl->count);
 633                 prepare_to_wait(&ctl->wait, &wait, TASK_UNINTERRUPTIBLE);
 634                 spin_unlock(&cache->lock);
 635
 636                 schedule();
 637
 638                 finish_wait(&ctl->wait, &wait);
 639                 put_caching_control(ctl);
 640                 spin_lock(&cache->lock);
 641         }
 642
 643         if (cache->cached != BTRFS_CACHE_NO) {
 644                 spin_unlock(&cache->lock);
 645                 kfree(caching_ctl);
 646                 return 0;
 647         }
 648         WARN_ON(cache->caching_ctl);
 649         cache->caching_ctl = caching_ctl;
 650         cache->cached = BTRFS_CACHE_FAST;
 651         spin_unlock(&cache->lock);
 652
 653         if (fs_info->mount_opt & BTRFS_MOUNT_SPACE_CACHE) {
 654                 mutex_lock(&caching_ctl->mutex);
 655                 ret = load_free_space_cache(fs_info, cache);
 656
 657                 spin_lock(&cache->lock);
 658                 if (ret == 1) {
 659                         cache->caching_ctl = NULL;
 660                         cache->cached = BTRFS_CACHE_FINISHED;
 661                         cache->last_byte_to_unpin = (u64)-1;
 662                         caching_ctl->progress = (u64)-1;
 663                 } else {
 664                         if (load_cache_only) {
 665                                 cache->caching_ctl = NULL;
 666                                 cache->cached = BTRFS_CACHE_NO;
 667                         } else {
 668                                 cache->cached = BTRFS_CACHE_STARTED;
 669                                 cache->has_caching_ctl = 1;
 670                         }
 671                 }
 672                 spin_unlock(&cache->lock);
 673 #ifdef CONFIG_BTRFS_DEBUG
 674                 if (ret == 1 &&
 675                     btrfs_should_fragment_free_space(fs_info->extent_root,
 676                                                      cache)) {
 677                         u64 bytes_used;
 678
 679                         spin_lock(&cache->space_info->lock);
 680                         spin_lock(&cache->lock);
 681                         bytes_used = cache->key.offset -
 682                                 btrfs_block_group_used(&cache->item);
 683                         cache->space_info->bytes_used += bytes_used >> 1;
 684                         spin_unlock(&cache->lock);
 685                         spin_unlock(&cache->space_info->lock);
 686                         fragment_free_space(fs_info->extent_root, cache);
 687                 }
 688 #endif
 689                 mutex_unlock(&caching_ctl->mutex);
 690
 691                 wake_up(&caching_ctl->wait);
 692                 if (ret == 1) {
 693                         put_caching_control(caching_ctl);
 694                         free_excluded_extents(fs_info->extent_root, cache);
 695                         return 0;
 696                 }
 697         } else {
 698                 /*
 699                  * We're either using the free space tree or no caching at all.
 700                  * Set cached to the appropriate value and wakeup any waiters.
 701                  */
 702                 spin_lock(&cache->lock);
 703                 if (load_cache_only) {
 704                         cache->caching_ctl = NULL;
 705                         cache->cached = BTRFS_CACHE_NO;
 706                 } else {
 707                         cache->cached = BTRFS_CACHE_STARTED;
 708                         cache->has_caching_ctl = 1;
 709                 }
 710                 spin_unlock(&cache->lock);
 711                 wake_up(&caching_ctl->wait);
 712         }
 713
 714         if (load_cache_only) {
 715                 put_caching_control(caching_ctl);
 716                 return 0;
 717         }
 718
 719         down_write(&fs_info->commit_root_sem);
 720         atomic_inc(&caching_ctl->count);
 721         list_add_tail(&caching_ctl->list, &fs_info->caching_block_groups);
 722         up_write(&fs_info->commit_root_sem);
 723
 724         btrfs_get_block_group(cache);
 725
 726         btrfs_queue_work(fs_info->caching_workers, &caching_ctl->work);
 727
 728         return ret;
 729 }
 730
 731 /*
 732  * return the block group that starts at or after bytenr
 733  */
 734 static struct btrfs_block_group_cache *
 735 btrfs_lookup_first_block_group(struct btrfs_fs_info *info, u64 bytenr)
 736 {
 737         struct btrfs_block_group_cache *cache;
 738
 739         cache = block_group_cache_tree_search(info, bytenr, 0);
 740
 741         return cache;
 742 }
 743
 744 /*
 745  * return the block group that contains the given bytenr
 746  */
 747 struct btrfs_block_group_cache *btrfs_lookup_block_group(
 748                                                  struct btrfs_fs_info *info,
 749                                                  u64 bytenr)
 750 {
 751         struct btrfs_block_group_cache *cache;
 752
 753         cache = block_group_cache_tree_search(info, bytenr, 1);
 754
 755         return cache;
 756 }
 757
 758 static struct btrfs_space_info *__find_space_info(struct btrfs_fs_info *info,
 759                                                   u64 flags)
 760 {
 761         struct list_head *head = &info->space_info;
 762         struct btrfs_space_info *found;
 763
 764         flags &= BTRFS_BLOCK_GROUP_TYPE_MASK;
 765
 766         rcu_read_lock();
 767         list_for_each_entry_rcu(found, head, list) {
 768                 if (found->flags & flags) {
 769                         rcu_read_unlock();
 770                         return found;
 771                 }
 772         }
 773         rcu_read_unlock();
 774         return NULL;
 775 }
 776
 777 /*
 778  * after adding space to the filesystem, we need to clear the full flags
 779  * on all the space infos.
 780  */
 781 void btrfs_clear_space_info_full(struct btrfs_fs_info *info)
 782 {
 783         struct list_head *head = &info->space_info;
 784         struct btrfs_space_info *found;
 785
 786         rcu_read_lock();
 787         list_for_each_entry_rcu(found, head, list)
 788                 found->full = 0;
 789         rcu_read_unlock();
 790 }
 791
 792 /* simple helper to search for an existing data extent at a given offset */
 793 int btrfs_lookup_data_extent(struct btrfs_root *root, u64 start, u64 len)
 794 {
 795         int ret;
 796         struct btrfs_key key;
 797         struct btrfs_path *path;
 798
 799         path = btrfs_alloc_path();
 800         if (!path)
 801                 return -ENOMEM;
 802
 803         key.objectid = start;
 804         key.offset = len;
 805         key.type = BTRFS_EXTENT_ITEM_KEY;
 806         ret = btrfs_search_slot(NULL, root->fs_info->extent_root, &key, path,
 807                                 0, 0);
 808         btrfs_free_path(path);
 809         return ret;
 810 }
 811
 812 /*
 813  * helper function to lookup reference count and flags of a tree block.
 814  *
 815  * the head node for delayed ref is used to store the sum of all the
 816  * reference count modifications queued up in the rbtree. the head
 817  * node may also store the extent flags to set. This way you can check
 818  * to see what the reference count and extent flags would be if all of
 819  * the delayed refs are not processed.
 820  */
 821 int btrfs_lookup_extent_info(struct btrfs_trans_handle *trans,
 822                              struct btrfs_root *root, u64 bytenr,
 823                              u64 offset, int metadata, u64 *refs, u64 *flags)
 824 {
 825         struct btrfs_delayed_ref_head *head;
 826         struct btrfs_delayed_ref_root *delayed_refs;
 827         struct btrfs_path *path;
 828         struct btrfs_extent_item *ei;
 829         struct extent_buffer *leaf;
 830         struct btrfs_key key;
 831         u32 item_size;
 832         u64 num_refs;
 833         u64 extent_flags;
 834         int ret;
 835
 836         /*
 837          * If we don't have skinny metadata, don't bother doing anything
 838          * different
 839          */
 840         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA)) {
 841                 offset = root->nodesize;
 842                 metadata = 0;
 843         }
 844
 845         path = btrfs_alloc_path();
 846         if (!path)
 847                 return -ENOMEM;
 848
 849         if (!trans) {
 850                 path->skip_locking = 1;
 851                 path->search_commit_root = 1;
 852         }
 853
 854 search_again:
 855         key.objectid = bytenr;
 856         key.offset = offset;
 857         if (metadata)
 858                 key.type = BTRFS_METADATA_ITEM_KEY;
 859         else
 860                 key.type = BTRFS_EXTENT_ITEM_KEY;
 861
 862         ret = btrfs_search_slot(trans, root->fs_info->extent_root,
 863                                 &key, path, 0, 0);
 864         if (ret < 0)
 865                 goto out_free;
 866
 867         if (ret > 0 && metadata && key.type == BTRFS_METADATA_ITEM_KEY) {
 868                 if (path->slots[0]) {
 869                         path->slots[0]--;
 870                         btrfs_item_key_to_cpu(path->nodes[0], &key,
 871                                               path->slots[0]);
 872                         if (key.objectid == bytenr &&
 873                             key.type == BTRFS_EXTENT_ITEM_KEY &&
 874                             key.offset == root->nodesize)
 875                                 ret = 0;
 876                 }
 877         }
 878
 879         if (ret == 0) {
 880                 leaf = path->nodes[0];
 881                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
 882                 if (item_size >= sizeof(*ei)) {
 883                         ei = btrfs_item_ptr(leaf, path->slots[0],
 884                                             struct btrfs_extent_item);
 885                         num_refs = btrfs_extent_refs(leaf, ei);
 886                         extent_flags = btrfs_extent_flags(leaf, ei);
 887                 } else {
 888 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
 889                         struct btrfs_extent_item_v0 *ei0;
 890                         BUG_ON(item_size != sizeof(*ei0));
 891                         ei0 = btrfs_item_ptr(leaf, path->slots[0],
 892                                              struct btrfs_extent_item_v0);
 893                         num_refs = btrfs_extent_refs_v0(leaf, ei0);
 894                         /* FIXME: this isn't correct for data */
 895                         extent_flags = BTRFS_BLOCK_FLAG_FULL_BACKREF;
 896 #else
 897                         BUG();
 898 #endif
 899                 }
 900                 BUG_ON(num_refs == 0);
 901         } else {
 902                 num_refs = 0;
 903                 extent_flags = 0;
 904                 ret = 0;
 905         }
 906
 907         if (!trans)
 908                 goto out;
 909
 910         delayed_refs = &trans->transaction->delayed_refs;
 911         spin_lock(&delayed_refs->lock);
 912         head = btrfs_find_delayed_ref_head(trans, bytenr);
 913         if (head) {
 914                 if (!mutex_trylock(&head->mutex)) {
 915                         atomic_inc(&head->node.refs);
 916                         spin_unlock(&delayed_refs->lock);
 917
 918                         btrfs_release_path(path);
 919
 920                         /*
 921                          * Mutex was contended, block until it's released and try
 922                          * again
 923                          */
 924                         mutex_lock(&head->mutex);
 925                         mutex_unlock(&head->mutex);
 926                         btrfs_put_delayed_ref(&head->node);
 927                         goto search_again;
 928                 }
 929                 spin_lock(&head->lock);
 930                 if (head->extent_op && head->extent_op->update_flags)
 931                         extent_flags |= head->extent_op->flags_to_set;
 932                 else
 933                         BUG_ON(num_refs == 0);
 934
 935                 num_refs += head->node.ref_mod;
 936                 spin_unlock(&head->lock);
 937                 mutex_unlock(&head->mutex);
 938         }
 939         spin_unlock(&delayed_refs->lock);
 940 out:
 941         WARN_ON(num_refs == 0);
 942         if (refs)
 943                 *refs = num_refs;
 944         if (flags)
 945                 *flags = extent_flags;
 946 out_free:
 947         btrfs_free_path(path);
 948         return ret;
 949 }
 950
 951 /*
 952  * Back reference rules.  Back refs have three main goals:
 953  *
 954  * 1) differentiate between all holders of references to an extent so that
 955  *    when a reference is dropped we can make sure it was a valid reference
 956  *    before freeing the extent.
 957  *
 958  * 2) Provide enough information to quickly find the holders of an extent
 959  *    if we notice a given block is corrupted or bad.
 960  *
 961  * 3) Make it easy to migrate blocks for FS shrinking or storage pool
 962  *    maintenance.  This is actually the same as #2, but with a slightly
 963  *    different use case.
 964  *
 965  * There are two kinds of back refs. The implicit back refs is optimized
 966  * for pointers in non-shared tree blocks. For a given pointer in a block,
 967  * back refs of this kind provide information about the block's owner tree
 968  * and the pointer's key. These information allow us to find the block by
 969  * b-tree searching. The full back refs is for pointers in tree blocks not
 970  * referenced by their owner trees. The location of tree block is recorded
 971  * in the back refs. Actually the full back refs is generic, and can be
 972  * used in all cases the implicit back refs is used. The major shortcoming
 973  * of the full back refs is its overhead. Every time a tree block gets
 974  * COWed, we have to update back refs entry for all pointers in it.
 975  *
 976  * For a newly allocated tree block, we use implicit back refs for
 977  * pointers in it. This means most tree related operations only involve
 978  * implicit back refs. For a tree block created in old transaction, the
 979  * only way to drop a reference to it is COW it. So we can detect the
 980  * event that tree block loses its owner tree's reference and do the
 981  * back refs conversion.
 982  *
 983  * When a tree block is COW'd through a tree, there are four cases:
 984  *
 985  * The reference count of the block is one and the tree is the block's
 986  * owner tree. Nothing to do in this case.
 987  *
 988  * The reference count of the block is one and the tree is not the
 989  * block's owner tree. In this case, full back refs is used for pointers
 990  * in the block. Remove these full back refs, add implicit back refs for
 991  * every pointers in the new block.
 992  *
 993  * The reference count of the block is greater than one and the tree is
 994  * the block's owner tree. In this case, implicit back refs is used for
 995  * pointers in the block. Add full back refs for every pointers in the
 996  * block, increase lower level extents' reference counts. The original
 997  * implicit back refs are entailed to the new block.
 998  *
 999  * The reference count of the block is greater than one and the tree is
1000  * not the block's owner tree. Add implicit back refs for every pointer in
1001  * the new block, increase lower level extents' reference count.
1002  *
1003  * Back Reference Key composing:
1004  *
1005  * The key objectid corresponds to the first byte in the extent,
1006  * The key type is used to differentiate between types of back refs.
1007  * There are different meanings of the key offset for different types
1008  * of back refs.
1009  *
1010  * File extents can be referenced by:
1011  *
1012  * - multiple snapshots, subvolumes, or different generations in one subvol
1013  * - different files inside a single subvolume
1014  * - different offsets inside a file (bookend extents in file.c)
1015  *
1016  * The extent ref structure for the implicit back refs has fields for:
1017  *
1018  * - Objectid of the subvolume root
1019  * - objectid of the file holding the reference
1020  * - original offset in the file
1021  * - how many bookend extents
1022  *
1023  * The key offset for the implicit back refs is hash of the first
1024  * three fields.
1025  *
1026  * The extent ref structure for the full back refs has field for:
1027  *
1028  * - number of pointers in the tree leaf
1029  *
1030  * The key offset for the implicit back refs is the first byte of
1031  * the tree leaf
1032  *
1033  * When a file extent is allocated, The implicit back refs is used.
1034  * the fields are filled in:
1035  *
1036  *     (root_key.objectid, inode objectid, offset in file, 1)
1037  *
1038  * When a file extent is removed file truncation, we find the
1039  * corresponding implicit back refs and check the following fields:
1040  *
1041  *     (btrfs_header_owner(leaf), inode objectid, offset in file)
1042  *
1043  * Btree extents can be referenced by:
1044  *
1045  * - Different subvolumes
1046  *
1047  * Both the implicit back refs and the full back refs for tree blocks
1048  * only consist of key. The key offset for the implicit back refs is
1049  * objectid of block's owner tree. The key offset for the full back refs
1050  * is the first byte of parent block.
1051  *
1052  * When implicit back refs is used, information about the lowest key and
1053  * level of the tree block are required. These information are stored in
1054  * tree block info structure.
1055  */
1056
1057 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1058 static int convert_extent_item_v0(struct btrfs_trans_handle *trans,
1059                                   struct btrfs_root *root,
1060                                   struct btrfs_path *path,
1061                                   u64 owner, u32 extra_size)
1062 {
1063         struct btrfs_extent_item *item;
1064         struct btrfs_extent_item_v0 *ei0;
1065         struct btrfs_extent_ref_v0 *ref0;
1066         struct btrfs_tree_block_info *bi;
1067         struct extent_buffer *leaf;
1068         struct btrfs_key key;
1069         struct btrfs_key found_key;
1070         u32 new_size = sizeof(*item);
1071         u64 refs;
1072         int ret;
1073
1074         leaf = path->nodes[0];
1075         BUG_ON(btrfs_item_size_nr(leaf, path->slots[0]) != sizeof(*ei0));
1076
1077         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1078         ei0 = btrfs_item_ptr(leaf, path->slots[0],
1079                              struct btrfs_extent_item_v0);
1080         refs = btrfs_extent_refs_v0(leaf, ei0);
1081
1082         if (owner == (u64)-1) {
1083                 while (1) {
1084                         if (path->slots[0] >= btrfs_header_nritems(leaf)) {
1085                                 ret = btrfs_next_leaf(root, path);
1086                                 if (ret < 0)
1087                                         return ret;
1088                                 BUG_ON(ret > 0); /* Corruption */
1089                                 leaf = path->nodes[0];
1090                         }
1091                         btrfs_item_key_to_cpu(leaf, &found_key,
1092                                               path->slots[0]);
1093                         BUG_ON(key.objectid != found_key.objectid);
1094                         if (found_key.type != BTRFS_EXTENT_REF_V0_KEY) {
1095                                 path->slots[0]++;
1096                                 continue;
1097                         }
1098                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1099                                               struct btrfs_extent_ref_v0);
1100                         owner = btrfs_ref_objectid_v0(leaf, ref0);
1101                         break;
1102                 }
1103         }
1104         btrfs_release_path(path);
1105
1106         if (owner < BTRFS_FIRST_FREE_OBJECTID)
1107                 new_size += sizeof(*bi);
1108
1109         new_size -= sizeof(*ei0);
1110         ret = btrfs_search_slot(trans, root, &key, path,
1111                                 new_size + extra_size, 1);
1112         if (ret < 0)
1113                 return ret;
1114         BUG_ON(ret); /* Corruption */
1115
1116         btrfs_extend_item(root, path, new_size);
1117
1118         leaf = path->nodes[0];
1119         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1120         btrfs_set_extent_refs(leaf, item, refs);
1121         /* FIXME: get real generation */
1122         btrfs_set_extent_generation(leaf, item, 0);
1123         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1124                 btrfs_set_extent_flags(leaf, item,
1125                                        BTRFS_EXTENT_FLAG_TREE_BLOCK |
1126                                        BTRFS_BLOCK_FLAG_FULL_BACKREF);
1127                 bi = (struct btrfs_tree_block_info *)(item + 1);
1128                 /* FIXME: get first key of the block */
1129                 memset_extent_buffer(leaf, 0, (unsigned long)bi, sizeof(*bi));
1130                 btrfs_set_tree_block_level(leaf, bi, (int)owner);
1131         } else {
1132                 btrfs_set_extent_flags(leaf, item, BTRFS_EXTENT_FLAG_DATA);
1133         }
1134         btrfs_mark_buffer_dirty(leaf);
1135         return 0;
1136 }
1137 #endif
1138
1139 static u64 hash_extent_data_ref(u64 root_objectid, u64 owner, u64 offset)
1140 {
1141         u32 high_crc = ~(u32)0;
1142         u32 low_crc = ~(u32)0;
1143         __le64 lenum;
1144
1145         lenum = cpu_to_le64(root_objectid);
1146         high_crc = btrfs_crc32c(high_crc, &lenum, sizeof(lenum));
1147         lenum = cpu_to_le64(owner);
1148         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1149         lenum = cpu_to_le64(offset);
1150         low_crc = btrfs_crc32c(low_crc, &lenum, sizeof(lenum));
1151
1152         return ((u64)high_crc << 31) ^ (u64)low_crc;
1153 }
1154
1155 static u64 hash_extent_data_ref_item(struct extent_buffer *leaf,
1156                                      struct btrfs_extent_data_ref *ref)
1157 {
1158         return hash_extent_data_ref(btrfs_extent_data_ref_root(leaf, ref),
1159                                     btrfs_extent_data_ref_objectid(leaf, ref),
1160                                     btrfs_extent_data_ref_offset(leaf, ref));
1161 }
1162
1163 static int match_extent_data_ref(struct extent_buffer *leaf,
1164                                  struct btrfs_extent_data_ref *ref,
1165                                  u64 root_objectid, u64 owner, u64 offset)
1166 {
1167         if (btrfs_extent_data_ref_root(leaf, ref) != root_objectid ||
1168             btrfs_extent_data_ref_objectid(leaf, ref) != owner ||
1169             btrfs_extent_data_ref_offset(leaf, ref) != offset)
1170                 return 0;
1171         return 1;
1172 }
1173
1174 static noinline int lookup_extent_data_ref(struct btrfs_trans_handle *trans,
1175                                            struct btrfs_root *root,
1176                                            struct btrfs_path *path,
1177                                            u64 bytenr, u64 parent,
1178                                            u64 root_objectid,
1179                                            u64 owner, u64 offset)
1180 {
1181         struct btrfs_key key;
1182         struct btrfs_extent_data_ref *ref;
1183         struct extent_buffer *leaf;
1184         u32 nritems;
1185         int ret;
1186         int recow;
1187         int err = -ENOENT;
1188
1189         key.objectid = bytenr;
1190         if (parent) {
1191                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1192                 key.offset = parent;
1193         } else {
1194                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1195                 key.offset = hash_extent_data_ref(root_objectid,
1196                                                   owner, offset);
1197         }
1198 again:
1199         recow = 0;
1200         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1201         if (ret < 0) {
1202                 err = ret;
1203                 goto fail;
1204         }
1205
1206         if (parent) {
1207                 if (!ret)
1208                         return 0;
1209 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1210                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1211                 btrfs_release_path(path);
1212                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1213                 if (ret < 0) {
1214                         err = ret;
1215                         goto fail;
1216                 }
1217                 if (!ret)
1218                         return 0;
1219 #endif
1220                 goto fail;
1221         }
1222
1223         leaf = path->nodes[0];
1224         nritems = btrfs_header_nritems(leaf);
1225         while (1) {
1226                 if (path->slots[0] >= nritems) {
1227                         ret = btrfs_next_leaf(root, path);
1228                         if (ret < 0)
1229                                 err = ret;
1230                         if (ret)
1231                                 goto fail;
1232
1233                         leaf = path->nodes[0];
1234                         nritems = btrfs_header_nritems(leaf);
1235                         recow = 1;
1236                 }
1237
1238                 btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1239                 if (key.objectid != bytenr ||
1240                     key.type != BTRFS_EXTENT_DATA_REF_KEY)
1241                         goto fail;
1242
1243                 ref = btrfs_item_ptr(leaf, path->slots[0],
1244                                      struct btrfs_extent_data_ref);
1245
1246                 if (match_extent_data_ref(leaf, ref, root_objectid,
1247                                           owner, offset)) {
1248                         if (recow) {
1249                                 btrfs_release_path(path);
1250                                 goto again;
1251                         }
1252                         err = 0;
1253                         break;
1254                 }
1255                 path->slots[0]++;
1256         }
1257 fail:
1258         return err;
1259 }
1260
1261 static noinline int insert_extent_data_ref(struct btrfs_trans_handle *trans,
1262                                            struct btrfs_root *root,
1263                                            struct btrfs_path *path,
1264                                            u64 bytenr, u64 parent,
1265                                            u64 root_objectid, u64 owner,
1266                                            u64 offset, int refs_to_add)
1267 {
1268         struct btrfs_key key;
1269         struct extent_buffer *leaf;
1270         u32 size;
1271         u32 num_refs;
1272         int ret;
1273
1274         key.objectid = bytenr;
1275         if (parent) {
1276                 key.type = BTRFS_SHARED_DATA_REF_KEY;
1277                 key.offset = parent;
1278                 size = sizeof(struct btrfs_shared_data_ref);
1279         } else {
1280                 key.type = BTRFS_EXTENT_DATA_REF_KEY;
1281                 key.offset = hash_extent_data_ref(root_objectid,
1282                                                   owner, offset);
1283                 size = sizeof(struct btrfs_extent_data_ref);
1284         }
1285
1286         ret = btrfs_insert_empty_item(trans, root, path, &key, size);
1287         if (ret && ret != -EEXIST)
1288                 goto fail;
1289
1290         leaf = path->nodes[0];
1291         if (parent) {
1292                 struct btrfs_shared_data_ref *ref;
1293                 ref = btrfs_item_ptr(leaf, path->slots[0],
1294                                      struct btrfs_shared_data_ref);
1295                 if (ret == 0) {
1296                         btrfs_set_shared_data_ref_count(leaf, ref, refs_to_add);
1297                 } else {
1298                         num_refs = btrfs_shared_data_ref_count(leaf, ref);
1299                         num_refs += refs_to_add;
1300                         btrfs_set_shared_data_ref_count(leaf, ref, num_refs);
1301                 }
1302         } else {
1303                 struct btrfs_extent_data_ref *ref;
1304                 while (ret == -EEXIST) {
1305                         ref = btrfs_item_ptr(leaf, path->slots[0],
1306                                              struct btrfs_extent_data_ref);
1307                         if (match_extent_data_ref(leaf, ref, root_objectid,
1308                                                   owner, offset))
1309                                 break;
1310                         btrfs_release_path(path);
1311                         key.offset++;
1312                         ret = btrfs_insert_empty_item(trans, root, path, &key,
1313                                                       size);
1314                         if (ret && ret != -EEXIST)
1315                                 goto fail;
1316
1317                         leaf = path->nodes[0];
1318                 }
1319                 ref = btrfs_item_ptr(leaf, path->slots[0],
1320                                      struct btrfs_extent_data_ref);
1321                 if (ret == 0) {
1322                         btrfs_set_extent_data_ref_root(leaf, ref,
1323                                                        root_objectid);
1324                         btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
1325                         btrfs_set_extent_data_ref_offset(leaf, ref, offset);
1326                         btrfs_set_extent_data_ref_count(leaf, ref, refs_to_add);
1327                 } else {
1328                         num_refs = btrfs_extent_data_ref_count(leaf, ref);
1329                         num_refs += refs_to_add;
1330                         btrfs_set_extent_data_ref_count(leaf, ref, num_refs);
1331                 }
1332         }
1333         btrfs_mark_buffer_dirty(leaf);
1334         ret = 0;
1335 fail:
1336         btrfs_release_path(path);
1337         return ret;
1338 }
1339
1340 static noinline int remove_extent_data_ref(struct btrfs_trans_handle *trans,
1341                                            struct btrfs_root *root,
1342                                            struct btrfs_path *path,
1343                                            int refs_to_drop, int *last_ref)
1344 {
1345         struct btrfs_key key;
1346         struct btrfs_extent_data_ref *ref1 = NULL;
1347         struct btrfs_shared_data_ref *ref2 = NULL;
1348         struct extent_buffer *leaf;
1349         u32 num_refs = 0;
1350         int ret = 0;
1351
1352         leaf = path->nodes[0];
1353         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1354
1355         if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1356                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1357                                       struct btrfs_extent_data_ref);
1358                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1359         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1360                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1361                                       struct btrfs_shared_data_ref);
1362                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1363 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1364         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1365                 struct btrfs_extent_ref_v0 *ref0;
1366                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1367                                       struct btrfs_extent_ref_v0);
1368                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1369 #endif
1370         } else {
1371                 BUG();
1372         }
1373
1374         BUG_ON(num_refs < refs_to_drop);
1375         num_refs -= refs_to_drop;
1376
1377         if (num_refs == 0) {
1378                 ret = btrfs_del_item(trans, root, path);
1379                 *last_ref = 1;
1380         } else {
1381                 if (key.type == BTRFS_EXTENT_DATA_REF_KEY)
1382                         btrfs_set_extent_data_ref_count(leaf, ref1, num_refs);
1383                 else if (key.type == BTRFS_SHARED_DATA_REF_KEY)
1384                         btrfs_set_shared_data_ref_count(leaf, ref2, num_refs);
1385 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1386                 else {
1387                         struct btrfs_extent_ref_v0 *ref0;
1388                         ref0 = btrfs_item_ptr(leaf, path->slots[0],
1389                                         struct btrfs_extent_ref_v0);
1390                         btrfs_set_ref_count_v0(leaf, ref0, num_refs);
1391                 }
1392 #endif
1393                 btrfs_mark_buffer_dirty(leaf);
1394         }
1395         return ret;
1396 }
1397
1398 static noinline u32 extent_data_ref_count(struct btrfs_path *path,
1399                                           struct btrfs_extent_inline_ref *iref)
1400 {
1401         struct btrfs_key key;
1402         struct extent_buffer *leaf;
1403         struct btrfs_extent_data_ref *ref1;
1404         struct btrfs_shared_data_ref *ref2;
1405         u32 num_refs = 0;
1406
1407         leaf = path->nodes[0];
1408         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
1409         if (iref) {
1410                 if (btrfs_extent_inline_ref_type(leaf, iref) ==
1411                     BTRFS_EXTENT_DATA_REF_KEY) {
1412                         ref1 = (struct btrfs_extent_data_ref *)(&iref->offset);
1413                         num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1414                 } else {
1415                         ref2 = (struct btrfs_shared_data_ref *)(iref + 1);
1416                         num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1417                 }
1418         } else if (key.type == BTRFS_EXTENT_DATA_REF_KEY) {
1419                 ref1 = btrfs_item_ptr(leaf, path->slots[0],
1420                                       struct btrfs_extent_data_ref);
1421                 num_refs = btrfs_extent_data_ref_count(leaf, ref1);
1422         } else if (key.type == BTRFS_SHARED_DATA_REF_KEY) {
1423                 ref2 = btrfs_item_ptr(leaf, path->slots[0],
1424                                       struct btrfs_shared_data_ref);
1425                 num_refs = btrfs_shared_data_ref_count(leaf, ref2);
1426 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1427         } else if (key.type == BTRFS_EXTENT_REF_V0_KEY) {
1428                 struct btrfs_extent_ref_v0 *ref0;
1429                 ref0 = btrfs_item_ptr(leaf, path->slots[0],
1430                                       struct btrfs_extent_ref_v0);
1431                 num_refs = btrfs_ref_count_v0(leaf, ref0);
1432 #endif
1433         } else {
1434                 WARN_ON(1);
1435         }
1436         return num_refs;
1437 }
1438
1439 static noinline int lookup_tree_block_ref(struct btrfs_trans_handle *trans,
1440                                           struct btrfs_root *root,
1441                                           struct btrfs_path *path,
1442                                           u64 bytenr, u64 parent,
1443                                           u64 root_objectid)
1444 {
1445         struct btrfs_key key;
1446         int ret;
1447
1448         key.objectid = bytenr;
1449         if (parent) {
1450                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1451                 key.offset = parent;
1452         } else {
1453                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1454                 key.offset = root_objectid;
1455         }
1456
1457         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1458         if (ret > 0)
1459                 ret = -ENOENT;
1460 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1461         if (ret == -ENOENT && parent) {
1462                 btrfs_release_path(path);
1463                 key.type = BTRFS_EXTENT_REF_V0_KEY;
1464                 ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
1465                 if (ret > 0)
1466                         ret = -ENOENT;
1467         }
1468 #endif
1469         return ret;
1470 }
1471
1472 static noinline int insert_tree_block_ref(struct btrfs_trans_handle *trans,
1473                                           struct btrfs_root *root,
1474                                           struct btrfs_path *path,
1475                                           u64 bytenr, u64 parent,
1476                                           u64 root_objectid)
1477 {
1478         struct btrfs_key key;
1479         int ret;
1480
1481         key.objectid = bytenr;
1482         if (parent) {
1483                 key.type = BTRFS_SHARED_BLOCK_REF_KEY;
1484                 key.offset = parent;
1485         } else {
1486                 key.type = BTRFS_TREE_BLOCK_REF_KEY;
1487                 key.offset = root_objectid;
1488         }
1489
1490         ret = btrfs_insert_empty_item(trans, root, path, &key, 0);
1491         btrfs_release_path(path);
1492         return ret;
1493 }
1494
1495 static inline int extent_ref_type(u64 parent, u64 owner)
1496 {
1497         int type;
1498         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1499                 if (parent > 0)
1500                         type = BTRFS_SHARED_BLOCK_REF_KEY;
1501                 else
1502                         type = BTRFS_TREE_BLOCK_REF_KEY;
1503         } else {
1504                 if (parent > 0)
1505                         type = BTRFS_SHARED_DATA_REF_KEY;
1506                 else
1507                         type = BTRFS_EXTENT_DATA_REF_KEY;
1508         }
1509         return type;
1510 }
1511
1512 static int find_next_key(struct btrfs_path *path, int level,
1513                          struct btrfs_key *key)
1514
1515 {
1516         for (; level < BTRFS_MAX_LEVEL; level++) {
1517                 if (!path->nodes[level])
1518                         break;
1519                 if (path->slots[level] + 1 >=
1520                     btrfs_header_nritems(path->nodes[level]))
1521                         continue;
1522                 if (level == 0)
1523                         btrfs_item_key_to_cpu(path->nodes[level], key,
1524                                               path->slots[level] + 1);
1525                 else
1526                         btrfs_node_key_to_cpu(path->nodes[level], key,
1527                                               path->slots[level] + 1);
1528                 return 0;
1529         }
1530         return 1;
1531 }
1532
1533 /*
1534  * look for inline back ref. if back ref is found, *ref_ret is set
1535  * to the address of inline back ref, and 0 is returned.
1536  *
1537  * if back ref isn't found, *ref_ret is set to the address where it
1538  * should be inserted, and -ENOENT is returned.
1539  *
1540  * if insert is true and there are too many inline back refs, the path
1541  * points to the extent item, and -EAGAIN is returned.
1542  *
1543  * NOTE: inline back refs are ordered in the same way that back ref
1544  *       items in the tree are ordered.
1545  */
1546 static noinline_for_stack
1547 int lookup_inline_extent_backref(struct btrfs_trans_handle *trans,
1548                                  struct btrfs_root *root,
1549                                  struct btrfs_path *path,
1550                                  struct btrfs_extent_inline_ref **ref_ret,
1551                                  u64 bytenr, u64 num_bytes,
1552                                  u64 parent, u64 root_objectid,
1553                                  u64 owner, u64 offset, int insert)
1554 {
1555         struct btrfs_key key;
1556         struct extent_buffer *leaf;
1557         struct btrfs_extent_item *ei;
1558         struct btrfs_extent_inline_ref *iref;
1559         u64 flags;
1560         u64 item_size;
1561         unsigned long ptr;
1562         unsigned long end;
1563         int extra_size;
1564         int type;
1565         int want;
1566         int ret;
1567         int err = 0;
1568         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
1569                                                  SKINNY_METADATA);
1570
1571         key.objectid = bytenr;
1572         key.type = BTRFS_EXTENT_ITEM_KEY;
1573         key.offset = num_bytes;
1574
1575         want = extent_ref_type(parent, owner);
1576         if (insert) {
1577                 extra_size = btrfs_extent_inline_ref_size(want);
1578                 path->keep_locks = 1;
1579         } else
1580                 extra_size = -1;
1581
1582         /*
1583          * Owner is our parent level, so we can just add one to get the level
1584          * for the block we are interested in.
1585          */
1586         if (skinny_metadata && owner < BTRFS_FIRST_FREE_OBJECTID) {
1587                 key.type = BTRFS_METADATA_ITEM_KEY;
1588                 key.offset = owner;
1589         }
1590
1591 again:
1592         ret = btrfs_search_slot(trans, root, &key, path, extra_size, 1);
1593         if (ret < 0) {
1594                 err = ret;
1595                 goto out;
1596         }
1597
1598         /*
1599          * We may be a newly converted file system which still has the old fat
1600          * extent entries for metadata, so try and see if we have one of those.
1601          */
1602         if (ret > 0 && skinny_metadata) {
1603                 skinny_metadata = false;
1604                 if (path->slots[0]) {
1605                         path->slots[0]--;
1606                         btrfs_item_key_to_cpu(path->nodes[0], &key,
1607                                               path->slots[0]);
1608                         if (key.objectid == bytenr &&
1609                             key.type == BTRFS_EXTENT_ITEM_KEY &&
1610                             key.offset == num_bytes)
1611                                 ret = 0;
1612                 }
1613                 if (ret) {
1614                         key.objectid = bytenr;
1615                         key.type = BTRFS_EXTENT_ITEM_KEY;
1616                         key.offset = num_bytes;
1617                         btrfs_release_path(path);
1618                         goto again;
1619                 }
1620         }
1621
1622         if (ret && !insert) {
1623                 err = -ENOENT;
1624                 goto out;
1625         } else if (WARN_ON(ret)) {
1626                 err = -EIO;
1627                 goto out;
1628         }
1629
1630         leaf = path->nodes[0];
1631         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1632 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
1633         if (item_size < sizeof(*ei)) {
1634                 if (!insert) {
1635                         err = -ENOENT;
1636                         goto out;
1637                 }
1638                 ret = convert_extent_item_v0(trans, root, path, owner,
1639                                              extra_size);
1640                 if (ret < 0) {
1641                         err = ret;
1642                         goto out;
1643                 }
1644                 leaf = path->nodes[0];
1645                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1646         }
1647 #endif
1648         BUG_ON(item_size < sizeof(*ei));
1649
1650         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1651         flags = btrfs_extent_flags(leaf, ei);
1652
1653         ptr = (unsigned long)(ei + 1);
1654         end = (unsigned long)ei + item_size;
1655
1656         if (flags & BTRFS_EXTENT_FLAG_TREE_BLOCK && !skinny_metadata) {
1657                 ptr += sizeof(struct btrfs_tree_block_info);
1658                 BUG_ON(ptr > end);
1659         }
1660
1661         err = -ENOENT;
1662         while (1) {
1663                 if (ptr >= end) {
1664                         WARN_ON(ptr > end);
1665                         break;
1666                 }
1667                 iref = (struct btrfs_extent_inline_ref *)ptr;
1668                 type = btrfs_extent_inline_ref_type(leaf, iref);
1669                 if (want < type)
1670                         break;
1671                 if (want > type) {
1672                         ptr += btrfs_extent_inline_ref_size(type);
1673                         continue;
1674                 }
1675
1676                 if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1677                         struct btrfs_extent_data_ref *dref;
1678                         dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1679                         if (match_extent_data_ref(leaf, dref, root_objectid,
1680                                                   owner, offset)) {
1681                                 err = 0;
1682                                 break;
1683                         }
1684                         if (hash_extent_data_ref_item(leaf, dref) <
1685                             hash_extent_data_ref(root_objectid, owner, offset))
1686                                 break;
1687                 } else {
1688                         u64 ref_offset;
1689                         ref_offset = btrfs_extent_inline_ref_offset(leaf, iref);
1690                         if (parent > 0) {
1691                                 if (parent == ref_offset) {
1692                                         err = 0;
1693                                         break;
1694                                 }
1695                                 if (ref_offset < parent)
1696                                         break;
1697                         } else {
1698                                 if (root_objectid == ref_offset) {
1699                                         err = 0;
1700                                         break;
1701                                 }
1702                                 if (ref_offset < root_objectid)
1703                                         break;
1704                         }
1705                 }
1706                 ptr += btrfs_extent_inline_ref_size(type);
1707         }
1708         if (err == -ENOENT && insert) {
1709                 if (item_size + extra_size >=
1710                     BTRFS_MAX_EXTENT_ITEM_SIZE(root)) {
1711                         err = -EAGAIN;
1712                         goto out;
1713                 }
1714                 /*
1715                  * To add new inline back ref, we have to make sure
1716                  * there is no corresponding back ref item.
1717                  * For simplicity, we just do not add new inline back
1718                  * ref if there is any kind of item for this block
1719                  */
1720                 if (find_next_key(path, 0, &key) == 0 &&
1721                     key.objectid == bytenr &&
1722                     key.type < BTRFS_BLOCK_GROUP_ITEM_KEY) {
1723                         err = -EAGAIN;
1724                         goto out;
1725                 }
1726         }
1727         *ref_ret = (struct btrfs_extent_inline_ref *)ptr;
1728 out:
1729         if (insert) {
1730                 path->keep_locks = 0;
1731                 btrfs_unlock_up_safe(path, 1);
1732         }
1733         return err;
1734 }
1735
1736 /*
1737  * helper to add new inline back ref
1738  */
1739 static noinline_for_stack
1740 void setup_inline_extent_backref(struct btrfs_root *root,
1741                                  struct btrfs_path *path,
1742                                  struct btrfs_extent_inline_ref *iref,
1743                                  u64 parent, u64 root_objectid,
1744                                  u64 owner, u64 offset, int refs_to_add,
1745                                  struct btrfs_delayed_extent_op *extent_op)
1746 {
1747         struct extent_buffer *leaf;
1748         struct btrfs_extent_item *ei;
1749         unsigned long ptr;
1750         unsigned long end;
1751         unsigned long item_offset;
1752         u64 refs;
1753         int size;
1754         int type;
1755
1756         leaf = path->nodes[0];
1757         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1758         item_offset = (unsigned long)iref - (unsigned long)ei;
1759
1760         type = extent_ref_type(parent, owner);
1761         size = btrfs_extent_inline_ref_size(type);
1762
1763         btrfs_extend_item(root, path, size);
1764
1765         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1766         refs = btrfs_extent_refs(leaf, ei);
1767         refs += refs_to_add;
1768         btrfs_set_extent_refs(leaf, ei, refs);
1769         if (extent_op)
1770                 __run_delayed_extent_op(extent_op, leaf, ei);
1771
1772         ptr = (unsigned long)ei + item_offset;
1773         end = (unsigned long)ei + btrfs_item_size_nr(leaf, path->slots[0]);
1774         if (ptr < end - size)
1775                 memmove_extent_buffer(leaf, ptr + size, ptr,
1776                                       end - size - ptr);
1777
1778         iref = (struct btrfs_extent_inline_ref *)ptr;
1779         btrfs_set_extent_inline_ref_type(leaf, iref, type);
1780         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1781                 struct btrfs_extent_data_ref *dref;
1782                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1783                 btrfs_set_extent_data_ref_root(leaf, dref, root_objectid);
1784                 btrfs_set_extent_data_ref_objectid(leaf, dref, owner);
1785                 btrfs_set_extent_data_ref_offset(leaf, dref, offset);
1786                 btrfs_set_extent_data_ref_count(leaf, dref, refs_to_add);
1787         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1788                 struct btrfs_shared_data_ref *sref;
1789                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1790                 btrfs_set_shared_data_ref_count(leaf, sref, refs_to_add);
1791                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1792         } else if (type == BTRFS_SHARED_BLOCK_REF_KEY) {
1793                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
1794         } else {
1795                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
1796         }
1797         btrfs_mark_buffer_dirty(leaf);
1798 }
1799
1800 static int lookup_extent_backref(struct btrfs_trans_handle *trans,
1801                                  struct btrfs_root *root,
1802                                  struct btrfs_path *path,
1803                                  struct btrfs_extent_inline_ref **ref_ret,
1804                                  u64 bytenr, u64 num_bytes, u64 parent,
1805                                  u64 root_objectid, u64 owner, u64 offset)
1806 {
1807         int ret;
1808
1809         ret = lookup_inline_extent_backref(trans, root, path, ref_ret,
1810                                            bytenr, num_bytes, parent,
1811                                            root_objectid, owner, offset, 0);
1812         if (ret != -ENOENT)
1813                 return ret;
1814
1815         btrfs_release_path(path);
1816         *ref_ret = NULL;
1817
1818         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1819                 ret = lookup_tree_block_ref(trans, root, path, bytenr, parent,
1820                                             root_objectid);
1821         } else {
1822                 ret = lookup_extent_data_ref(trans, root, path, bytenr, parent,
1823                                              root_objectid, owner, offset);
1824         }
1825         return ret;
1826 }
1827
1828 /*
1829  * helper to update/remove inline back ref
1830  */
1831 static noinline_for_stack
1832 void update_inline_extent_backref(struct btrfs_root *root,
1833                                   struct btrfs_path *path,
1834                                   struct btrfs_extent_inline_ref *iref,
1835                                   int refs_to_mod,
1836                                   struct btrfs_delayed_extent_op *extent_op,
1837                                   int *last_ref)
1838 {
1839         struct extent_buffer *leaf;
1840         struct btrfs_extent_item *ei;
1841         struct btrfs_extent_data_ref *dref = NULL;
1842         struct btrfs_shared_data_ref *sref = NULL;
1843         unsigned long ptr;
1844         unsigned long end;
1845         u32 item_size;
1846         int size;
1847         int type;
1848         u64 refs;
1849
1850         leaf = path->nodes[0];
1851         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
1852         refs = btrfs_extent_refs(leaf, ei);
1853         WARN_ON(refs_to_mod < 0 && refs + refs_to_mod <= 0);
1854         refs += refs_to_mod;
1855         btrfs_set_extent_refs(leaf, ei, refs);
1856         if (extent_op)
1857                 __run_delayed_extent_op(extent_op, leaf, ei);
1858
1859         type = btrfs_extent_inline_ref_type(leaf, iref);
1860
1861         if (type == BTRFS_EXTENT_DATA_REF_KEY) {
1862                 dref = (struct btrfs_extent_data_ref *)(&iref->offset);
1863                 refs = btrfs_extent_data_ref_count(leaf, dref);
1864         } else if (type == BTRFS_SHARED_DATA_REF_KEY) {
1865                 sref = (struct btrfs_shared_data_ref *)(iref + 1);
1866                 refs = btrfs_shared_data_ref_count(leaf, sref);
1867         } else {
1868                 refs = 1;
1869                 BUG_ON(refs_to_mod != -1);
1870         }
1871
1872         BUG_ON(refs_to_mod < 0 && refs < -refs_to_mod);
1873         refs += refs_to_mod;
1874
1875         if (refs > 0) {
1876                 if (type == BTRFS_EXTENT_DATA_REF_KEY)
1877                         btrfs_set_extent_data_ref_count(leaf, dref, refs);
1878                 else
1879                         btrfs_set_shared_data_ref_count(leaf, sref, refs);
1880         } else {
1881                 *last_ref = 1;
1882                 size =  btrfs_extent_inline_ref_size(type);
1883                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
1884                 ptr = (unsigned long)iref;
1885                 end = (unsigned long)ei + item_size;
1886                 if (ptr + size < end)
1887                         memmove_extent_buffer(leaf, ptr, ptr + size,
1888                                               end - ptr - size);
1889                 item_size -= size;
1890                 btrfs_truncate_item(root, path, item_size, 1);
1891         }
1892         btrfs_mark_buffer_dirty(leaf);
1893 }
1894
1895 static noinline_for_stack
1896 int insert_inline_extent_backref(struct btrfs_trans_handle *trans,
1897                                  struct btrfs_root *root,
1898                                  struct btrfs_path *path,
1899                                  u64 bytenr, u64 num_bytes, u64 parent,
1900                                  u64 root_objectid, u64 owner,
1901                                  u64 offset, int refs_to_add,
1902                                  struct btrfs_delayed_extent_op *extent_op)
1903 {
1904         struct btrfs_extent_inline_ref *iref;
1905         int ret;
1906
1907         ret = lookup_inline_extent_backref(trans, root, path, &iref,
1908                                            bytenr, num_bytes, parent,
1909                                            root_objectid, owner, offset, 1);
1910         if (ret == 0) {
1911                 BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID);
1912                 update_inline_extent_backref(root, path, iref,
1913                                              refs_to_add, extent_op, NULL);
1914         } else if (ret == -ENOENT) {
1915                 setup_inline_extent_backref(root, path, iref, parent,
1916                                             root_objectid, owner, offset,
1917                                             refs_to_add, extent_op);
1918                 ret = 0;
1919         }
1920         return ret;
1921 }
1922
1923 static int insert_extent_backref(struct btrfs_trans_handle *trans,
1924                                  struct btrfs_root *root,
1925                                  struct btrfs_path *path,
1926                                  u64 bytenr, u64 parent, u64 root_objectid,
1927                                  u64 owner, u64 offset, int refs_to_add)
1928 {
1929         int ret;
1930         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
1931                 BUG_ON(refs_to_add != 1);
1932                 ret = insert_tree_block_ref(trans, root, path, bytenr,
1933                                             parent, root_objectid);
1934         } else {
1935                 ret = insert_extent_data_ref(trans, root, path, bytenr,
1936                                              parent, root_objectid,
1937                                              owner, offset, refs_to_add);
1938         }
1939         return ret;
1940 }
1941
1942 static int remove_extent_backref(struct btrfs_trans_handle *trans,
1943                                  struct btrfs_root *root,
1944                                  struct btrfs_path *path,
1945                                  struct btrfs_extent_inline_ref *iref,
1946                                  int refs_to_drop, int is_data, int *last_ref)
1947 {
1948         int ret = 0;
1949
1950         BUG_ON(!is_data && refs_to_drop != 1);
1951         if (iref) {
1952                 update_inline_extent_backref(root, path, iref,
1953                                              -refs_to_drop, NULL, last_ref);
1954         } else if (is_data) {
1955                 ret = remove_extent_data_ref(trans, root, path, refs_to_drop,
1956                                              last_ref);
1957         } else {
1958                 *last_ref = 1;
1959                 ret = btrfs_del_item(trans, root, path);
1960         }
1961         return ret;
1962 }
1963
1964 #define in_range(b, first, len)        ((b) >= (first) && (b) < (first) + (len))
1965 static int btrfs_issue_discard(struct block_device *bdev, u64 start, u64 len,
1966                                u64 *discarded_bytes)
1967 {
1968         int j, ret = 0;
1969         u64 bytes_left, end;
1970         u64 aligned_start = ALIGN(start, 1 << 9);
1971
1972         if (WARN_ON(start != aligned_start)) {
1973                 len -= aligned_start - start;
1974                 len = round_down(len, 1 << 9);
1975                 start = aligned_start;
1976         }
1977
1978         *discarded_bytes = 0;
1979
1980         if (!len)
1981                 return 0;
1982
1983         end = start + len;
1984         bytes_left = len;
1985
1986         /* Skip any superblocks on this device. */
1987         for (j = 0; j < BTRFS_SUPER_MIRROR_MAX; j++) {
1988                 u64 sb_start = btrfs_sb_offset(j);
1989                 u64 sb_end = sb_start + BTRFS_SUPER_INFO_SIZE;
1990                 u64 size = sb_start - start;
1991
1992                 if (!in_range(sb_start, start, bytes_left) &&
1993                     !in_range(sb_end, start, bytes_left) &&
1994                     !in_range(start, sb_start, BTRFS_SUPER_INFO_SIZE))
1995                         continue;
1996
1997                 /*
1998                  * Superblock spans beginning of range.  Adjust start and
1999                  * try again.
2000                  */
2001                 if (sb_start <= start) {
2002                         start += sb_end - start;
2003                         if (start > end) {
2004                                 bytes_left = 0;
2005                                 break;
2006                         }
2007                         bytes_left = end - start;
2008                         continue;
2009                 }
2010
2011                 if (size) {
2012                         ret = blkdev_issue_discard(bdev, start >> 9, size >> 9,
2013                                                    GFP_NOFS, 0);
2014                         if (!ret)
2015                                 *discarded_bytes += size;
2016                         else if (ret != -EOPNOTSUPP)
2017                                 return ret;
2018                 }
2019
2020                 start = sb_end;
2021                 if (start > end) {
2022                         bytes_left = 0;
2023                         break;
2024                 }
2025                 bytes_left = end - start;
2026         }
2027
2028         if (bytes_left) {
2029                 ret = blkdev_issue_discard(bdev, start >> 9, bytes_left >> 9,
2030                                            GFP_NOFS, 0);
2031                 if (!ret)
2032                         *discarded_bytes += bytes_left;
2033         }
2034         return ret;
2035 }
2036
2037 int btrfs_discard_extent(struct btrfs_root *root, u64 bytenr,
2038                          u64 num_bytes, u64 *actual_bytes)
2039 {
2040         int ret;
2041         u64 discarded_bytes = 0;
2042         struct btrfs_bio *bbio = NULL;
2043
2044
2045         /* Tell the block device(s) that the sectors can be discarded */
2046         ret = btrfs_map_block(root->fs_info, REQ_DISCARD,
2047                               bytenr, &num_bytes, &bbio, 0);
2048         /* Error condition is -ENOMEM */
2049         if (!ret) {
2050                 struct btrfs_bio_stripe *stripe = bbio->stripes;
2051                 int i;
2052
2053
2054                 for (i = 0; i < bbio->num_stripes; i++, stripe++) {
2055                         u64 bytes;
2056                         if (!stripe->dev->can_discard)
2057                                 continue;
2058
2059                         ret = btrfs_issue_discard(stripe->dev->bdev,
2060                                                   stripe->physical,
2061                                                   stripe->length,
2062                                                   &bytes);
2063                         if (!ret)
2064                                 discarded_bytes += bytes;
2065                         else if (ret != -EOPNOTSUPP)
2066                                 break; /* Logic errors or -ENOMEM, or -EIO but I don't know how that could happen JDM */
2067
2068                         /*
2069                          * Just in case we get back EOPNOTSUPP for some reason,
2070                          * just ignore the return value so we don't screw up
2071                          * people calling discard_extent.
2072                          */
2073                         ret = 0;
2074                 }
2075                 btrfs_put_bbio(bbio);
2076         }
2077
2078         if (actual_bytes)
2079                 *actual_bytes = discarded_bytes;
2080
2081
2082         if (ret == -EOPNOTSUPP)
2083                 ret = 0;
2084         return ret;
2085 }
2086
2087 /* Can return -ENOMEM */
2088 int btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2089                          struct btrfs_root *root,
2090                          u64 bytenr, u64 num_bytes, u64 parent,
2091                          u64 root_objectid, u64 owner, u64 offset)
2092 {
2093         int ret;
2094         struct btrfs_fs_info *fs_info = root->fs_info;
2095
2096         BUG_ON(owner < BTRFS_FIRST_FREE_OBJECTID &&
2097                root_objectid == BTRFS_TREE_LOG_OBJECTID);
2098
2099         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
2100                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
2101                                         num_bytes,
2102                                         parent, root_objectid, (int)owner,
2103                                         BTRFS_ADD_DELAYED_REF, NULL);
2104         } else {
2105                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
2106                                         num_bytes, parent, root_objectid,
2107                                         owner, offset, 0,
2108                                         BTRFS_ADD_DELAYED_REF, NULL);
2109         }
2110         return ret;
2111 }
2112
2113 static int __btrfs_inc_extent_ref(struct btrfs_trans_handle *trans,
2114                                   struct btrfs_root *root,
2115                                   struct btrfs_delayed_ref_node *node,
2116                                   u64 parent, u64 root_objectid,
2117                                   u64 owner, u64 offset, int refs_to_add,
2118                                   struct btrfs_delayed_extent_op *extent_op)
2119 {
2120         struct btrfs_fs_info *fs_info = root->fs_info;
2121         struct btrfs_path *path;
2122         struct extent_buffer *leaf;
2123         struct btrfs_extent_item *item;
2124         struct btrfs_key key;
2125         u64 bytenr = node->bytenr;
2126         u64 num_bytes = node->num_bytes;
2127         u64 refs;
2128         int ret;
2129
2130         path = btrfs_alloc_path();
2131         if (!path)
2132                 return -ENOMEM;
2133
2134         path->reada = 1;
2135         path->leave_spinning = 1;
2136         /* this will setup the path even if it fails to insert the back ref */
2137         ret = insert_inline_extent_backref(trans, fs_info->extent_root, path,
2138                                            bytenr, num_bytes, parent,
2139                                            root_objectid, owner, offset,
2140                                            refs_to_add, extent_op);
2141         if ((ret < 0 && ret != -EAGAIN) || !ret)
2142                 goto out;
2143
2144         /*
2145          * Ok we had -EAGAIN which means we didn't have space to insert and
2146          * inline extent ref, so just update the reference count and add a
2147          * normal backref.
2148          */
2149         leaf = path->nodes[0];
2150         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
2151         item = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2152         refs = btrfs_extent_refs(leaf, item);
2153         btrfs_set_extent_refs(leaf, item, refs + refs_to_add);
2154         if (extent_op)
2155                 __run_delayed_extent_op(extent_op, leaf, item);
2156
2157         btrfs_mark_buffer_dirty(leaf);
2158         btrfs_release_path(path);
2159
2160         path->reada = 1;
2161         path->leave_spinning = 1;
2162         /* now insert the actual backref */
2163         ret = insert_extent_backref(trans, root->fs_info->extent_root,
2164                                     path, bytenr, parent, root_objectid,
2165                                     owner, offset, refs_to_add);
2166         if (ret)
2167                 btrfs_abort_transaction(trans, root, ret);
2168 out:
2169         btrfs_free_path(path);
2170         return ret;
2171 }
2172
2173 static int run_delayed_data_ref(struct btrfs_trans_handle *trans,
2174                                 struct btrfs_root *root,
2175                                 struct btrfs_delayed_ref_node *node,
2176                                 struct btrfs_delayed_extent_op *extent_op,
2177                                 int insert_reserved)
2178 {
2179         int ret = 0;
2180         struct btrfs_delayed_data_ref *ref;
2181         struct btrfs_key ins;
2182         u64 parent = 0;
2183         u64 ref_root = 0;
2184         u64 flags = 0;
2185
2186         ins.objectid = node->bytenr;
2187         ins.offset = node->num_bytes;
2188         ins.type = BTRFS_EXTENT_ITEM_KEY;
2189
2190         ref = btrfs_delayed_node_to_data_ref(node);
2191         trace_run_delayed_data_ref(node, ref, node->action);
2192
2193         if (node->type == BTRFS_SHARED_DATA_REF_KEY)
2194                 parent = ref->parent;
2195         ref_root = ref->root;
2196
2197         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2198                 if (extent_op)
2199                         flags |= extent_op->flags_to_set;
2200                 ret = alloc_reserved_file_extent(trans, root,
2201                                                  parent, ref_root, flags,
2202                                                  ref->objectid, ref->offset,
2203                                                  &ins, node->ref_mod);
2204         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2205                 ret = __btrfs_inc_extent_ref(trans, root, node, parent,
2206                                              ref_root, ref->objectid,
2207                                              ref->offset, node->ref_mod,
2208                                              extent_op);
2209         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2210                 ret = __btrfs_free_extent(trans, root, node, parent,
2211                                           ref_root, ref->objectid,
2212                                           ref->offset, node->ref_mod,
2213                                           extent_op);
2214         } else {
2215                 BUG();
2216         }
2217         return ret;
2218 }
2219
2220 static void __run_delayed_extent_op(struct btrfs_delayed_extent_op *extent_op,
2221                                     struct extent_buffer *leaf,
2222                                     struct btrfs_extent_item *ei)
2223 {
2224         u64 flags = btrfs_extent_flags(leaf, ei);
2225         if (extent_op->update_flags) {
2226                 flags |= extent_op->flags_to_set;
2227                 btrfs_set_extent_flags(leaf, ei, flags);
2228         }
2229
2230         if (extent_op->update_key) {
2231                 struct btrfs_tree_block_info *bi;
2232                 BUG_ON(!(flags & BTRFS_EXTENT_FLAG_TREE_BLOCK));
2233                 bi = (struct btrfs_tree_block_info *)(ei + 1);
2234                 btrfs_set_tree_block_key(leaf, bi, &extent_op->key);
2235         }
2236 }
2237
2238 static int run_delayed_extent_op(struct btrfs_trans_handle *trans,
2239                                  struct btrfs_root *root,
2240                                  struct btrfs_delayed_ref_node *node,
2241                                  struct btrfs_delayed_extent_op *extent_op)
2242 {
2243         struct btrfs_key key;
2244         struct btrfs_path *path;
2245         struct btrfs_extent_item *ei;
2246         struct extent_buffer *leaf;
2247         u32 item_size;
2248         int ret;
2249         int err = 0;
2250         int metadata = !extent_op->is_data;
2251
2252         if (trans->aborted)
2253                 return 0;
2254
2255         if (metadata && !btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2256                 metadata = 0;
2257
2258         path = btrfs_alloc_path();
2259         if (!path)
2260                 return -ENOMEM;
2261
2262         key.objectid = node->bytenr;
2263
2264         if (metadata) {
2265                 key.type = BTRFS_METADATA_ITEM_KEY;
2266                 key.offset = extent_op->level;
2267         } else {
2268                 key.type = BTRFS_EXTENT_ITEM_KEY;
2269                 key.offset = node->num_bytes;
2270         }
2271
2272 again:
2273         path->reada = 1;
2274         path->leave_spinning = 1;
2275         ret = btrfs_search_slot(trans, root->fs_info->extent_root, &key,
2276                                 path, 0, 1);
2277         if (ret < 0) {
2278                 err = ret;
2279                 goto out;
2280         }
2281         if (ret > 0) {
2282                 if (metadata) {
2283                         if (path->slots[0] > 0) {
2284                                 path->slots[0]--;
2285                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
2286                                                       path->slots[0]);
2287                                 if (key.objectid == node->bytenr &&
2288                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
2289                                     key.offset == node->num_bytes)
2290                                         ret = 0;
2291                         }
2292                         if (ret > 0) {
2293                                 btrfs_release_path(path);
2294                                 metadata = 0;
2295
2296                                 key.objectid = node->bytenr;
2297                                 key.offset = node->num_bytes;
2298                                 key.type = BTRFS_EXTENT_ITEM_KEY;
2299                                 goto again;
2300                         }
2301                 } else {
2302                         err = -EIO;
2303                         goto out;
2304                 }
2305         }
2306
2307         leaf = path->nodes[0];
2308         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2309 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
2310         if (item_size < sizeof(*ei)) {
2311                 ret = convert_extent_item_v0(trans, root->fs_info->extent_root,
2312                                              path, (u64)-1, 0);
2313                 if (ret < 0) {
2314                         err = ret;
2315                         goto out;
2316                 }
2317                 leaf = path->nodes[0];
2318                 item_size = btrfs_item_size_nr(leaf, path->slots[0]);
2319         }
2320 #endif
2321         BUG_ON(item_size < sizeof(*ei));
2322         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
2323         __run_delayed_extent_op(extent_op, leaf, ei);
2324
2325         btrfs_mark_buffer_dirty(leaf);
2326 out:
2327         btrfs_free_path(path);
2328         return err;
2329 }
2330
2331 static int run_delayed_tree_ref(struct btrfs_trans_handle *trans,
2332                                 struct btrfs_root *root,
2333                                 struct btrfs_delayed_ref_node *node,
2334                                 struct btrfs_delayed_extent_op *extent_op,
2335                                 int insert_reserved)
2336 {
2337         int ret = 0;
2338         struct btrfs_delayed_tree_ref *ref;
2339         struct btrfs_key ins;
2340         u64 parent = 0;
2341         u64 ref_root = 0;
2342         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
2343                                                  SKINNY_METADATA);
2344
2345         ref = btrfs_delayed_node_to_tree_ref(node);
2346         trace_run_delayed_tree_ref(node, ref, node->action);
2347
2348         if (node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2349                 parent = ref->parent;
2350         ref_root = ref->root;
2351
2352         ins.objectid = node->bytenr;
2353         if (skinny_metadata) {
2354                 ins.offset = ref->level;
2355                 ins.type = BTRFS_METADATA_ITEM_KEY;
2356         } else {
2357                 ins.offset = node->num_bytes;
2358                 ins.type = BTRFS_EXTENT_ITEM_KEY;
2359         }
2360
2361         BUG_ON(node->ref_mod != 1);
2362         if (node->action == BTRFS_ADD_DELAYED_REF && insert_reserved) {
2363                 BUG_ON(!extent_op || !extent_op->update_flags);
2364                 ret = alloc_reserved_tree_block(trans, root,
2365                                                 parent, ref_root,
2366                                                 extent_op->flags_to_set,
2367                                                 &extent_op->key,
2368                                                 ref->level, &ins);
2369         } else if (node->action == BTRFS_ADD_DELAYED_REF) {
2370                 ret = __btrfs_inc_extent_ref(trans, root, node,
2371                                              parent, ref_root,
2372                                              ref->level, 0, 1,
2373                                              extent_op);
2374         } else if (node->action == BTRFS_DROP_DELAYED_REF) {
2375                 ret = __btrfs_free_extent(trans, root, node,
2376                                           parent, ref_root,
2377                                           ref->level, 0, 1, extent_op);
2378         } else {
2379                 BUG();
2380         }
2381         return ret;
2382 }
2383
2384 /* helper function to actually process a single delayed ref entry */
2385 static int run_one_delayed_ref(struct btrfs_trans_handle *trans,
2386                                struct btrfs_root *root,
2387                                struct btrfs_delayed_ref_node *node,
2388                                struct btrfs_delayed_extent_op *extent_op,
2389                                int insert_reserved)
2390 {
2391         int ret = 0;
2392
2393         if (trans->aborted) {
2394                 if (insert_reserved)
2395                         btrfs_pin_extent(root, node->bytenr,
2396                                          node->num_bytes, 1);
2397                 return 0;
2398         }
2399
2400         if (btrfs_delayed_ref_is_head(node)) {
2401                 struct btrfs_delayed_ref_head *head;
2402                 /*
2403                  * we've hit the end of the chain and we were supposed
2404                  * to insert this extent into the tree.  But, it got
2405                  * deleted before we ever needed to insert it, so all
2406                  * we have to do is clean up the accounting
2407                  */
2408                 BUG_ON(extent_op);
2409                 head = btrfs_delayed_node_to_head(node);
2410                 trace_run_delayed_ref_head(node, head, node->action);
2411
2412                 if (insert_reserved) {
2413                         btrfs_pin_extent(root, node->bytenr,
2414                                          node->num_bytes, 1);
2415                         if (head->is_data) {
2416                                 ret = btrfs_del_csums(trans, root,
2417                                                       node->bytenr,
2418                                                       node->num_bytes);
2419                         }
2420                 }
2421
2422                 /* Also free its reserved qgroup space */
2423                 btrfs_qgroup_free_delayed_ref(root->fs_info,
2424                                               head->qgroup_ref_root,
2425                                               head->qgroup_reserved);
2426                 return ret;
2427         }
2428
2429         if (node->type == BTRFS_TREE_BLOCK_REF_KEY ||
2430             node->type == BTRFS_SHARED_BLOCK_REF_KEY)
2431                 ret = run_delayed_tree_ref(trans, root, node, extent_op,
2432                                            insert_reserved);
2433         else if (node->type == BTRFS_EXTENT_DATA_REF_KEY ||
2434                  node->type == BTRFS_SHARED_DATA_REF_KEY)
2435                 ret = run_delayed_data_ref(trans, root, node, extent_op,
2436                                            insert_reserved);
2437         else
2438                 BUG();
2439         return ret;
2440 }
2441
2442 static inline struct btrfs_delayed_ref_node *
2443 select_delayed_ref(struct btrfs_delayed_ref_head *head)
2444 {
2445         struct btrfs_delayed_ref_node *ref;
2446
2447         if (list_empty(&head->ref_list))
2448                 return NULL;
2449
2450         /*
2451          * Select a delayed ref of type BTRFS_ADD_DELAYED_REF first.
2452          * This is to prevent a ref count from going down to zero, which deletes
2453          * the extent item from the extent tree, when there still are references
2454          * to add, which would fail because they would not find the extent item.
2455          */
2456         list_for_each_entry(ref, &head->ref_list, list) {
2457                 if (ref->action == BTRFS_ADD_DELAYED_REF)
2458                         return ref;
2459         }
2460
2461         return list_entry(head->ref_list.next, struct btrfs_delayed_ref_node,
2462                           list);
2463 }
2464
2465 /*
2466  * Returns 0 on success or if called with an already aborted transaction.
2467  * Returns -ENOMEM or -EIO on failure and will abort the transaction.
2468  */
2469 static noinline int __btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2470                                              struct btrfs_root *root,
2471                                              unsigned long nr)
2472 {
2473         struct btrfs_delayed_ref_root *delayed_refs;
2474         struct btrfs_delayed_ref_node *ref;
2475         struct btrfs_delayed_ref_head *locked_ref = NULL;
2476         struct btrfs_delayed_extent_op *extent_op;
2477         struct btrfs_fs_info *fs_info = root->fs_info;
2478         ktime_t start = ktime_get();
2479         int ret;
2480         unsigned long count = 0;
2481         unsigned long actual_count = 0;
2482         int must_insert_reserved = 0;
2483
2484         delayed_refs = &trans->transaction->delayed_refs;
2485         while (1) {
2486                 if (!locked_ref) {
2487                         if (count >= nr)
2488                                 break;
2489
2490                         spin_lock(&delayed_refs->lock);
2491                         locked_ref = btrfs_select_ref_head(trans);
2492                         if (!locked_ref) {
2493                                 spin_unlock(&delayed_refs->lock);
2494                                 break;
2495                         }
2496
2497                         /* grab the lock that says we are going to process
2498                          * all the refs for this head */
2499                         ret = btrfs_delayed_ref_lock(trans, locked_ref);
2500                         spin_unlock(&delayed_refs->lock);
2501                         /*
2502                          * we may have dropped the spin lock to get the head
2503                          * mutex lock, and that might have given someone else
2504                          * time to free the head.  If that's true, it has been
2505                          * removed from our list and we can move on.
2506                          */
2507                         if (ret == -EAGAIN) {
2508                                 locked_ref = NULL;
2509                                 count++;
2510                                 continue;
2511                         }
2512                 }
2513
2514                 /*
2515                  * We need to try and merge add/drops of the same ref since we
2516                  * can run into issues with relocate dropping the implicit ref
2517                  * and then it being added back again before the drop can
2518                  * finish.  If we merged anything we need to re-loop so we can
2519                  * get a good ref.
2520                  * Or we can get node references of the same type that weren't
2521                  * merged when created due to bumps in the tree mod seq, and
2522                  * we need to merge them to prevent adding an inline extent
2523                  * backref before dropping it (triggering a BUG_ON at
2524                  * insert_inline_extent_backref()).
2525                  */
2526                 spin_lock(&locked_ref->lock);
2527                 btrfs_merge_delayed_refs(trans, fs_info, delayed_refs,
2528                                          locked_ref);
2529
2530                 /*
2531                  * locked_ref is the head node, so we have to go one
2532                  * node back for any delayed ref updates
2533                  */
2534                 ref = select_delayed_ref(locked_ref);
2535
2536                 if (ref && ref->seq &&
2537                     btrfs_check_delayed_seq(fs_info, delayed_refs, ref->seq)) {
2538                         spin_unlock(&locked_ref->lock);
2539                         btrfs_delayed_ref_unlock(locked_ref);
2540                         spin_lock(&delayed_refs->lock);
2541                         locked_ref->processing = 0;
2542                         delayed_refs->num_heads_ready++;
2543                         spin_unlock(&delayed_refs->lock);
2544                         locked_ref = NULL;
2545                         cond_resched();
2546                         count++;
2547                         continue;
2548                 }
2549
2550                 /*
2551                  * record the must insert reserved flag before we
2552                  * drop the spin lock.
2553                  */
2554                 must_insert_reserved = locked_ref->must_insert_reserved;
2555                 locked_ref->must_insert_reserved = 0;
2556
2557                 extent_op = locked_ref->extent_op;
2558                 locked_ref->extent_op = NULL;
2559
2560                 if (!ref) {
2561
2562
2563                         /* All delayed refs have been processed, Go ahead
2564                          * and send the head node to run_one_delayed_ref,
2565                          * so that any accounting fixes can happen
2566                          */
2567                         ref = &locked_ref->node;
2568
2569                         if (extent_op && must_insert_reserved) {
2570                                 btrfs_free_delayed_extent_op(extent_op);
2571                                 extent_op = NULL;
2572                         }
2573
2574                         if (extent_op) {
2575                                 spin_unlock(&locked_ref->lock);
2576                                 ret = run_delayed_extent_op(trans, root,
2577                                                             ref, extent_op);
2578                                 btrfs_free_delayed_extent_op(extent_op);
2579
2580                                 if (ret) {
2581                                         /*
2582                                          * Need to reset must_insert_reserved if
2583                                          * there was an error so the abort stuff
2584                                          * can cleanup the reserved space
2585                                          * properly.
2586                                          */
2587                                         if (must_insert_reserved)
2588                                                 locked_ref->must_insert_reserved = 1;
2589                                         locked_ref->processing = 0;
2590                                         btrfs_debug(fs_info, "run_delayed_extent_op returned %d", ret);
2591                                         btrfs_delayed_ref_unlock(locked_ref);
2592                                         return ret;
2593                                 }
2594                                 continue;
2595                         }
2596
2597                         /*
2598                          * Need to drop our head ref lock and re-aqcuire the
2599                          * delayed ref lock and then re-check to make sure
2600                          * nobody got added.
2601                          */
2602                         spin_unlock(&locked_ref->lock);
2603                         spin_lock(&delayed_refs->lock);
2604                         spin_lock(&locked_ref->lock);
2605                         if (!list_empty(&locked_ref->ref_list) ||
2606                             locked_ref->extent_op) {
2607                                 spin_unlock(&locked_ref->lock);
2608                                 spin_unlock(&delayed_refs->lock);
2609                                 continue;
2610                         }
2611                         ref->in_tree = 0;
2612                         delayed_refs->num_heads--;
2613                         rb_erase(&locked_ref->href_node,
2614                                  &delayed_refs->href_root);
2615                         spin_unlock(&delayed_refs->lock);
2616                 } else {
2617                         actual_count++;
2618                         ref->in_tree = 0;
2619                         list_del(&ref->list);
2620                 }
2621                 atomic_dec(&delayed_refs->num_entries);
2622
2623                 if (!btrfs_delayed_ref_is_head(ref)) {
2624                         /*
2625                          * when we play the delayed ref, also correct the
2626                          * ref_mod on head
2627                          */
2628                         switch (ref->action) {
2629                         case BTRFS_ADD_DELAYED_REF:
2630                         case BTRFS_ADD_DELAYED_EXTENT:
2631                                 locked_ref->node.ref_mod -= ref->ref_mod;
2632                                 break;
2633                         case BTRFS_DROP_DELAYED_REF:
2634                                 locked_ref->node.ref_mod += ref->ref_mod;
2635                                 break;
2636                         default:
2637                                 WARN_ON(1);
2638                         }
2639                 }
2640                 spin_unlock(&locked_ref->lock);
2641
2642                 ret = run_one_delayed_ref(trans, root, ref, extent_op,
2643                                           must_insert_reserved);
2644
2645                 btrfs_free_delayed_extent_op(extent_op);
2646                 if (ret) {
2647                         locked_ref->processing = 0;
2648                         btrfs_delayed_ref_unlock(locked_ref);
2649                         btrfs_put_delayed_ref(ref);
2650                         btrfs_debug(fs_info, "run_one_delayed_ref returned %d", ret);
2651                         return ret;
2652                 }
2653
2654                 /*
2655                  * If this node is a head, that means all the refs in this head
2656                  * have been dealt with, and we will pick the next head to deal
2657                  * with, so we must unlock the head and drop it from the cluster
2658                  * list before we release it.
2659                  */
2660                 if (btrfs_delayed_ref_is_head(ref)) {
2661                         if (locked_ref->is_data &&
2662                             locked_ref->total_ref_mod < 0) {
2663                                 spin_lock(&delayed_refs->lock);
2664                                 delayed_refs->pending_csums -= ref->num_bytes;
2665                                 spin_unlock(&delayed_refs->lock);
2666                         }
2667                         btrfs_delayed_ref_unlock(locked_ref);
2668                         locked_ref = NULL;
2669                 }
2670                 btrfs_put_delayed_ref(ref);
2671                 count++;
2672                 cond_resched();
2673         }
2674
2675         /*
2676          * We don't want to include ref heads since we can have empty ref heads
2677          * and those will drastically skew our runtime down since we just do
2678          * accounting, no actual extent tree updates.
2679          */
2680         if (actual_count > 0) {
2681                 u64 runtime = ktime_to_ns(ktime_sub(ktime_get(), start));
2682                 u64 avg;
2683
2684                 /*
2685                  * We weigh the current average higher than our current runtime
2686                  * to avoid large swings in the average.
2687                  */
2688                 spin_lock(&delayed_refs->lock);
2689                 avg = fs_info->avg_delayed_ref_runtime * 3 + runtime;
2690                 fs_info->avg_delayed_ref_runtime = avg >> 2;    /* div by 4 */
2691                 spin_unlock(&delayed_refs->lock);
2692         }
2693         return 0;
2694 }
2695
2696 #ifdef SCRAMBLE_DELAYED_REFS
2697 /*
2698  * Normally delayed refs get processed in ascending bytenr order. This
2699  * correlates in most cases to the order added. To expose dependencies on this
2700  * order, we start to process the tree in the middle instead of the beginning
2701  */
2702 static u64 find_middle(struct rb_root *root)
2703 {
2704         struct rb_node *n = root->rb_node;
2705         struct btrfs_delayed_ref_node *entry;
2706         int alt = 1;
2707         u64 middle;
2708         u64 first = 0, last = 0;
2709
2710         n = rb_first(root);
2711         if (n) {
2712                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2713                 first = entry->bytenr;
2714         }
2715         n = rb_last(root);
2716         if (n) {
2717                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2718                 last = entry->bytenr;
2719         }
2720         n = root->rb_node;
2721
2722         while (n) {
2723                 entry = rb_entry(n, struct btrfs_delayed_ref_node, rb_node);
2724                 WARN_ON(!entry->in_tree);
2725
2726                 middle = entry->bytenr;
2727
2728                 if (alt)
2729                         n = n->rb_left;
2730                 else
2731                         n = n->rb_right;
2732
2733                 alt = 1 - alt;
2734         }
2735         return middle;
2736 }
2737 #endif
2738
2739 static inline u64 heads_to_leaves(struct btrfs_root *root, u64 heads)
2740 {
2741         u64 num_bytes;
2742
2743         num_bytes = heads * (sizeof(struct btrfs_extent_item) +
2744                              sizeof(struct btrfs_extent_inline_ref));
2745         if (!btrfs_fs_incompat(root->fs_info, SKINNY_METADATA))
2746                 num_bytes += heads * sizeof(struct btrfs_tree_block_info);
2747
2748         /*
2749          * We don't ever fill up leaves all the way so multiply by 2 just to be
2750          * closer to what we're really going to want to ouse.
2751          */
2752         return div_u64(num_bytes, BTRFS_LEAF_DATA_SIZE(root));
2753 }
2754
2755 /*
2756  * Takes the number of bytes to be csumm'ed and figures out how many leaves it
2757  * would require to store the csums for that many bytes.
2758  */
2759 u64 btrfs_csum_bytes_to_leaves(struct btrfs_root *root, u64 csum_bytes)
2760 {
2761         u64 csum_size;
2762         u64 num_csums_per_leaf;
2763         u64 num_csums;
2764
2765         csum_size = BTRFS_LEAF_DATA_SIZE(root) - sizeof(struct btrfs_item);
2766         num_csums_per_leaf = div64_u64(csum_size,
2767                         (u64)btrfs_super_csum_size(root->fs_info->super_copy));
2768         num_csums = div64_u64(csum_bytes, root->sectorsize);
2769         num_csums += num_csums_per_leaf - 1;
2770         num_csums = div64_u64(num_csums, num_csums_per_leaf);
2771         return num_csums;
2772 }
2773
2774 int btrfs_check_space_for_delayed_refs(struct btrfs_trans_handle *trans,
2775                                        struct btrfs_root *root)
2776 {
2777         struct btrfs_block_rsv *global_rsv;
2778         u64 num_heads = trans->transaction->delayed_refs.num_heads_ready;
2779         u64 csum_bytes = trans->transaction->delayed_refs.pending_csums;
2780         u64 num_dirty_bgs = trans->transaction->num_dirty_bgs;
2781         u64 num_bytes, num_dirty_bgs_bytes;
2782         int ret = 0;
2783
2784         num_bytes = btrfs_calc_trans_metadata_size(root, 1);
2785         num_heads = heads_to_leaves(root, num_heads);
2786         if (num_heads > 1)
2787                 num_bytes += (num_heads - 1) * root->nodesize;
2788         num_bytes <<= 1;
2789         num_bytes += btrfs_csum_bytes_to_leaves(root, csum_bytes) * root->nodesize;
2790         num_dirty_bgs_bytes = btrfs_calc_trans_metadata_size(root,
2791                                                              num_dirty_bgs);
2792         global_rsv = &root->fs_info->global_block_rsv;
2793
2794         /*
2795          * If we can't allocate any more chunks lets make sure we have _lots_ of
2796          * wiggle room since running delayed refs can create more delayed refs.
2797          */
2798         if (global_rsv->space_info->full) {
2799                 num_dirty_bgs_bytes <<= 1;
2800                 num_bytes <<= 1;
2801         }
2802
2803         spin_lock(&global_rsv->lock);
2804         if (global_rsv->reserved <= num_bytes + num_dirty_bgs_bytes)
2805                 ret = 1;
2806         spin_unlock(&global_rsv->lock);
2807         return ret;
2808 }
2809
2810 int btrfs_should_throttle_delayed_refs(struct btrfs_trans_handle *trans,
2811                                        struct btrfs_root *root)
2812 {
2813         struct btrfs_fs_info *fs_info = root->fs_info;
2814         u64 num_entries =
2815                 atomic_read(&trans->transaction->delayed_refs.num_entries);
2816         u64 avg_runtime;
2817         u64 val;
2818
2819         smp_mb();
2820         avg_runtime = fs_info->avg_delayed_ref_runtime;
2821         val = num_entries * avg_runtime;
2822         if (num_entries * avg_runtime >= NSEC_PER_SEC)
2823                 return 1;
2824         if (val >= NSEC_PER_SEC / 2)
2825                 return 2;
2826
2827         return btrfs_check_space_for_delayed_refs(trans, root);
2828 }
2829
2830 struct async_delayed_refs {
2831         struct btrfs_root *root;
2832         int count;
2833         int error;
2834         int sync;
2835         struct completion wait;
2836         struct btrfs_work work;
2837 };
2838
2839 static void delayed_ref_async_start(struct btrfs_work *work)
2840 {
2841         struct async_delayed_refs *async;
2842         struct btrfs_trans_handle *trans;
2843         int ret;
2844
2845         async = container_of(work, struct async_delayed_refs, work);
2846
2847         trans = btrfs_join_transaction(async->root);
2848         if (IS_ERR(trans)) {
2849                 async->error = PTR_ERR(trans);
2850                 goto done;
2851         }
2852
2853         /*
2854          * trans->sync means that when we call end_transaciton, we won't
2855          * wait on delayed refs
2856          */
2857         trans->sync = true;
2858         ret = btrfs_run_delayed_refs(trans, async->root, async->count);
2859         if (ret)
2860                 async->error = ret;
2861
2862         ret = btrfs_end_transaction(trans, async->root);
2863         if (ret && !async->error)
2864                 async->error = ret;
2865 done:
2866         if (async->sync)
2867                 complete(&async->wait);
2868         else
2869                 kfree(async);
2870 }
2871
2872 int btrfs_async_run_delayed_refs(struct btrfs_root *root,
2873                                  unsigned long count, int wait)
2874 {
2875         struct async_delayed_refs *async;
2876         int ret;
2877
2878         async = kmalloc(sizeof(*async), GFP_NOFS);
2879         if (!async)
2880                 return -ENOMEM;
2881
2882         async->root = root->fs_info->tree_root;
2883         async->count = count;
2884         async->error = 0;
2885         if (wait)
2886                 async->sync = 1;
2887         else
2888                 async->sync = 0;
2889         init_completion(&async->wait);
2890
2891         btrfs_init_work(&async->work, btrfs_extent_refs_helper,
2892                         delayed_ref_async_start, NULL, NULL);
2893
2894         btrfs_queue_work(root->fs_info->extent_workers, &async->work);
2895
2896         if (wait) {
2897                 wait_for_completion(&async->wait);
2898                 ret = async->error;
2899                 kfree(async);
2900                 return ret;
2901         }
2902         return 0;
2903 }
2904
2905 /*
2906  * this starts processing the delayed reference count updates and
2907  * extent insertions we have queued up so far.  count can be
2908  * 0, which means to process everything in the tree at the start
2909  * of the run (but not newly added entries), or it can be some target
2910  * number you'd like to process.
2911  *
2912  * Returns 0 on success or if called with an aborted transaction
2913  * Returns <0 on error and aborts the transaction
2914  */
2915 int btrfs_run_delayed_refs(struct btrfs_trans_handle *trans,
2916                            struct btrfs_root *root, unsigned long count)
2917 {
2918         struct rb_node *node;
2919         struct btrfs_delayed_ref_root *delayed_refs;
2920         struct btrfs_delayed_ref_head *head;
2921         int ret;
2922         int run_all = count == (unsigned long)-1;
2923         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
2924
2925         /* We'll clean this up in btrfs_cleanup_transaction */
2926         if (trans->aborted)
2927                 return 0;
2928
2929         if (root == root->fs_info->extent_root)
2930                 root = root->fs_info->tree_root;
2931
2932         delayed_refs = &trans->transaction->delayed_refs;
2933         if (count == 0)
2934                 count = atomic_read(&delayed_refs->num_entries) * 2;
2935
2936 again:
2937 #ifdef SCRAMBLE_DELAYED_REFS
2938         delayed_refs->run_delayed_start = find_middle(&delayed_refs->root);
2939 #endif
2940         trans->can_flush_pending_bgs = false;
2941         ret = __btrfs_run_delayed_refs(trans, root, count);
2942         if (ret < 0) {
2943                 btrfs_abort_transaction(trans, root, ret);
2944                 return ret;
2945         }
2946
2947         if (run_all) {
2948                 if (!list_empty(&trans->new_bgs))
2949                         btrfs_create_pending_block_groups(trans, root);
2950
2951                 spin_lock(&delayed_refs->lock);
2952                 node = rb_first(&delayed_refs->href_root);
2953                 if (!node) {
2954                         spin_unlock(&delayed_refs->lock);
2955                         goto out;
2956                 }
2957                 count = (unsigned long)-1;
2958
2959                 while (node) {
2960                         head = rb_entry(node, struct btrfs_delayed_ref_head,
2961                                         href_node);
2962                         if (btrfs_delayed_ref_is_head(&head->node)) {
2963                                 struct btrfs_delayed_ref_node *ref;
2964
2965                                 ref = &head->node;
2966                                 atomic_inc(&ref->refs);
2967
2968                                 spin_unlock(&delayed_refs->lock);
2969                                 /*
2970                                  * Mutex was contended, block until it's
2971                                  * released and try again
2972                                  */
2973                                 mutex_lock(&head->mutex);
2974                                 mutex_unlock(&head->mutex);
2975
2976                                 btrfs_put_delayed_ref(ref);
2977                                 cond_resched();
2978                                 goto again;
2979                         } else {
2980                                 WARN_ON(1);
2981                         }
2982                         node = rb_next(node);
2983                 }
2984                 spin_unlock(&delayed_refs->lock);
2985                 cond_resched();
2986                 goto again;
2987         }
2988 out:
2989         assert_qgroups_uptodate(trans);
2990         trans->can_flush_pending_bgs = can_flush_pending_bgs;
2991         return 0;
2992 }
2993
2994 int btrfs_set_disk_extent_flags(struct btrfs_trans_handle *trans,
2995                                 struct btrfs_root *root,
2996                                 u64 bytenr, u64 num_bytes, u64 flags,
2997                                 int level, int is_data)
2998 {
2999         struct btrfs_delayed_extent_op *extent_op;
3000         int ret;
3001
3002         extent_op = btrfs_alloc_delayed_extent_op();
3003         if (!extent_op)
3004                 return -ENOMEM;
3005
3006         extent_op->flags_to_set = flags;
3007         extent_op->update_flags = 1;
3008         extent_op->update_key = 0;
3009         extent_op->is_data = is_data ? 1 : 0;
3010         extent_op->level = level;
3011
3012         ret = btrfs_add_delayed_extent_op(root->fs_info, trans, bytenr,
3013                                           num_bytes, extent_op);
3014         if (ret)
3015                 btrfs_free_delayed_extent_op(extent_op);
3016         return ret;
3017 }
3018
3019 static noinline int check_delayed_ref(struct btrfs_trans_handle *trans,
3020                                       struct btrfs_root *root,
3021                                       struct btrfs_path *path,
3022                                       u64 objectid, u64 offset, u64 bytenr)
3023 {
3024         struct btrfs_delayed_ref_head *head;
3025         struct btrfs_delayed_ref_node *ref;
3026         struct btrfs_delayed_data_ref *data_ref;
3027         struct btrfs_delayed_ref_root *delayed_refs;
3028         int ret = 0;
3029
3030         delayed_refs = &trans->transaction->delayed_refs;
3031         spin_lock(&delayed_refs->lock);
3032         head = btrfs_find_delayed_ref_head(trans, bytenr);
3033         if (!head) {
3034                 spin_unlock(&delayed_refs->lock);
3035                 return 0;
3036         }
3037
3038         if (!mutex_trylock(&head->mutex)) {
3039                 atomic_inc(&head->node.refs);
3040                 spin_unlock(&delayed_refs->lock);
3041
3042                 btrfs_release_path(path);
3043
3044                 /*
3045                  * Mutex was contended, block until it's released and let
3046                  * caller try again
3047                  */
3048                 mutex_lock(&head->mutex);
3049                 mutex_unlock(&head->mutex);
3050                 btrfs_put_delayed_ref(&head->node);
3051                 return -EAGAIN;
3052         }
3053         spin_unlock(&delayed_refs->lock);
3054
3055         spin_lock(&head->lock);
3056         list_for_each_entry(ref, &head->ref_list, list) {
3057                 /* If it's a shared ref we know a cross reference exists */
3058                 if (ref->type != BTRFS_EXTENT_DATA_REF_KEY) {
3059                         ret = 1;
3060                         break;
3061                 }
3062
3063                 data_ref = btrfs_delayed_node_to_data_ref(ref);
3064
3065                 /*
3066                  * If our ref doesn't match the one we're currently looking at
3067                  * then we have a cross reference.
3068                  */
3069                 if (data_ref->root != root->root_key.objectid ||
3070                     data_ref->objectid != objectid ||
3071                     data_ref->offset != offset) {
3072                         ret = 1;
3073                         break;
3074                 }
3075         }
3076         spin_unlock(&head->lock);
3077         mutex_unlock(&head->mutex);
3078         return ret;
3079 }
3080
3081 static noinline int check_committed_ref(struct btrfs_trans_handle *trans,
3082                                         struct btrfs_root *root,
3083                                         struct btrfs_path *path,
3084                                         u64 objectid, u64 offset, u64 bytenr)
3085 {
3086         struct btrfs_root *extent_root = root->fs_info->extent_root;
3087         struct extent_buffer *leaf;
3088         struct btrfs_extent_data_ref *ref;
3089         struct btrfs_extent_inline_ref *iref;
3090         struct btrfs_extent_item *ei;
3091         struct btrfs_key key;
3092         u32 item_size;
3093         int ret;
3094
3095         key.objectid = bytenr;
3096         key.offset = (u64)-1;
3097         key.type = BTRFS_EXTENT_ITEM_KEY;
3098
3099         ret = btrfs_search_slot(NULL, extent_root, &key, path, 0, 0);
3100         if (ret < 0)
3101                 goto out;
3102         BUG_ON(ret == 0); /* Corruption */
3103
3104         ret = -ENOENT;
3105         if (path->slots[0] == 0)
3106                 goto out;
3107
3108         path->slots[0]--;
3109         leaf = path->nodes[0];
3110         btrfs_item_key_to_cpu(leaf, &key, path->slots[0]);
3111
3112         if (key.objectid != bytenr || key.type != BTRFS_EXTENT_ITEM_KEY)
3113                 goto out;
3114
3115         ret = 1;
3116         item_size = btrfs_item_size_nr(leaf, path->slots[0]);
3117 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
3118         if (item_size < sizeof(*ei)) {
3119                 WARN_ON(item_size != sizeof(struct btrfs_extent_item_v0));
3120                 goto out;
3121         }
3122 #endif
3123         ei = btrfs_item_ptr(leaf, path->slots[0], struct btrfs_extent_item);
3124
3125         if (item_size != sizeof(*ei) +
3126             btrfs_extent_inline_ref_size(BTRFS_EXTENT_DATA_REF_KEY))
3127                 goto out;
3128
3129         if (btrfs_extent_generation(leaf, ei) <=
3130             btrfs_root_last_snapshot(&root->root_item))
3131                 goto out;
3132
3133         iref = (struct btrfs_extent_inline_ref *)(ei + 1);
3134         if (btrfs_extent_inline_ref_type(leaf, iref) !=
3135             BTRFS_EXTENT_DATA_REF_KEY)
3136                 goto out;
3137
3138         ref = (struct btrfs_extent_data_ref *)(&iref->offset);
3139         if (btrfs_extent_refs(leaf, ei) !=
3140             btrfs_extent_data_ref_count(leaf, ref) ||
3141             btrfs_extent_data_ref_root(leaf, ref) !=
3142             root->root_key.objectid ||
3143             btrfs_extent_data_ref_objectid(leaf, ref) != objectid ||
3144             btrfs_extent_data_ref_offset(leaf, ref) != offset)
3145                 goto out;
3146
3147         ret = 0;
3148 out:
3149         return ret;
3150 }
3151
3152 int btrfs_cross_ref_exist(struct btrfs_trans_handle *trans,
3153                           struct btrfs_root *root,
3154                           u64 objectid, u64 offset, u64 bytenr)
3155 {
3156         struct btrfs_path *path;
3157         int ret;
3158         int ret2;
3159
3160         path = btrfs_alloc_path();
3161         if (!path)
3162                 return -ENOENT;
3163
3164         do {
3165                 ret = check_committed_ref(trans, root, path, objectid,
3166                                           offset, bytenr);
3167                 if (ret && ret != -ENOENT)
3168                         goto out;
3169
3170                 ret2 = check_delayed_ref(trans, root, path, objectid,
3171                                          offset, bytenr);
3172         } while (ret2 == -EAGAIN);
3173
3174         if (ret2 && ret2 != -ENOENT) {
3175                 ret = ret2;
3176                 goto out;
3177         }
3178
3179         if (ret != -ENOENT || ret2 != -ENOENT)
3180                 ret = 0;
3181 out:
3182         btrfs_free_path(path);
3183         if (root->root_key.objectid == BTRFS_DATA_RELOC_TREE_OBJECTID)
3184                 WARN_ON(ret > 0);
3185         return ret;
3186 }
3187
3188 static int __btrfs_mod_ref(struct btrfs_trans_handle *trans,
3189                            struct btrfs_root *root,
3190                            struct extent_buffer *buf,
3191                            int full_backref, int inc)
3192 {
3193         u64 bytenr;
3194         u64 num_bytes;
3195         u64 parent;
3196         u64 ref_root;
3197         u32 nritems;
3198         struct btrfs_key key;
3199         struct btrfs_file_extent_item *fi;
3200         int i;
3201         int level;
3202         int ret = 0;
3203         int (*process_func)(struct btrfs_trans_handle *, struct btrfs_root *,
3204                             u64, u64, u64, u64, u64, u64);
3205
3206
3207         if (btrfs_test_is_dummy_root(root))
3208                 return 0;
3209
3210         ref_root = btrfs_header_owner(buf);
3211         nritems = btrfs_header_nritems(buf);
3212         level = btrfs_header_level(buf);
3213
3214         if (!test_bit(BTRFS_ROOT_REF_COWS, &root->state) && level == 0)
3215                 return 0;
3216
3217         if (inc)
3218                 process_func = btrfs_inc_extent_ref;
3219         else
3220                 process_func = btrfs_free_extent;
3221
3222         if (full_backref)
3223                 parent = buf->start;
3224         else
3225                 parent = 0;
3226
3227         for (i = 0; i < nritems; i++) {
3228                 if (level == 0) {
3229                         btrfs_item_key_to_cpu(buf, &key, i);
3230                         if (key.type != BTRFS_EXTENT_DATA_KEY)
3231                                 continue;
3232                         fi = btrfs_item_ptr(buf, i,
3233                                             struct btrfs_file_extent_item);
3234                         if (btrfs_file_extent_type(buf, fi) ==
3235                             BTRFS_FILE_EXTENT_INLINE)
3236                                 continue;
3237                         bytenr = btrfs_file_extent_disk_bytenr(buf, fi);
3238                         if (bytenr == 0)
3239                                 continue;
3240
3241                         num_bytes = btrfs_file_extent_disk_num_bytes(buf, fi);
3242                         key.offset -= btrfs_file_extent_offset(buf, fi);
3243                         ret = process_func(trans, root, bytenr, num_bytes,
3244                                            parent, ref_root, key.objectid,
3245                                            key.offset);
3246                         if (ret)
3247                                 goto fail;
3248                 } else {
3249                         bytenr = btrfs_node_blockptr(buf, i);
3250                         num_bytes = root->nodesize;
3251                         ret = process_func(trans, root, bytenr, num_bytes,
3252                                            parent, ref_root, level - 1, 0);
3253                         if (ret)
3254                                 goto fail;
3255                 }
3256         }
3257         return 0;
3258 fail:
3259         return ret;
3260 }
3261
3262 int btrfs_inc_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3263                   struct extent_buffer *buf, int full_backref)
3264 {
3265         return __btrfs_mod_ref(trans, root, buf, full_backref, 1);
3266 }
3267
3268 int btrfs_dec_ref(struct btrfs_trans_handle *trans, struct btrfs_root *root,
3269                   struct extent_buffer *buf, int full_backref)
3270 {
3271         return __btrfs_mod_ref(trans, root, buf, full_backref, 0);
3272 }
3273
3274 static int write_one_cache_group(struct btrfs_trans_handle *trans,
3275                                  struct btrfs_root *root,
3276                                  struct btrfs_path *path,
3277                                  struct btrfs_block_group_cache *cache)
3278 {
3279         int ret;
3280         struct btrfs_root *extent_root = root->fs_info->extent_root;
3281         unsigned long bi;
3282         struct extent_buffer *leaf;
3283
3284         ret = btrfs_search_slot(trans, extent_root, &cache->key, path, 0, 1);
3285         if (ret) {
3286                 if (ret > 0)
3287                         ret = -ENOENT;
3288                 goto fail;
3289         }
3290
3291         leaf = path->nodes[0];
3292         bi = btrfs_item_ptr_offset(leaf, path->slots[0]);
3293         write_extent_buffer(leaf, &cache->item, bi, sizeof(cache->item));
3294         btrfs_mark_buffer_dirty(leaf);
3295 fail:
3296         btrfs_release_path(path);
3297         return ret;
3298
3299 }
3300
3301 static struct btrfs_block_group_cache *
3302 next_block_group(struct btrfs_root *root,
3303                  struct btrfs_block_group_cache *cache)
3304 {
3305         struct rb_node *node;
3306
3307         spin_lock(&root->fs_info->block_group_cache_lock);
3308
3309         /* If our block group was removed, we need a full search. */
3310         if (RB_EMPTY_NODE(&cache->cache_node)) {
3311                 const u64 next_bytenr = cache->key.objectid + cache->key.offset;
3312
3313                 spin_unlock(&root->fs_info->block_group_cache_lock);
3314                 btrfs_put_block_group(cache);
3315                 cache = btrfs_lookup_first_block_group(root->fs_info,
3316                                                        next_bytenr);
3317                 return cache;
3318         }
3319         node = rb_next(&cache->cache_node);
3320         btrfs_put_block_group(cache);
3321         if (node) {
3322                 cache = rb_entry(node, struct btrfs_block_group_cache,
3323                                  cache_node);
3324                 btrfs_get_block_group(cache);
3325         } else
3326                 cache = NULL;
3327         spin_unlock(&root->fs_info->block_group_cache_lock);
3328         return cache;
3329 }
3330
3331 static int cache_save_setup(struct btrfs_block_group_cache *block_group,
3332                             struct btrfs_trans_handle *trans,
3333                             struct btrfs_path *path)
3334 {
3335         struct btrfs_root *root = block_group->fs_info->tree_root;
3336         struct inode *inode = NULL;
3337         u64 alloc_hint = 0;
3338         int dcs = BTRFS_DC_ERROR;
3339         u64 num_pages = 0;
3340         int retries = 0;
3341         int ret = 0;
3342
3343         /*
3344          * If this block group is smaller than 100 megs don't bother caching the
3345          * block group.
3346          */
3347         if (block_group->key.offset < (100 * 1024 * 1024)) {
3348                 spin_lock(&block_group->lock);
3349                 block_group->disk_cache_state = BTRFS_DC_WRITTEN;
3350                 spin_unlock(&block_group->lock);
3351                 return 0;
3352         }
3353
3354         if (trans->aborted)
3355                 return 0;
3356 again:
3357         inode = lookup_free_space_inode(root, block_group, path);
3358         if (IS_ERR(inode) && PTR_ERR(inode) != -ENOENT) {
3359                 ret = PTR_ERR(inode);
3360                 btrfs_release_path(path);
3361                 goto out;
3362         }
3363
3364         if (IS_ERR(inode)) {
3365                 BUG_ON(retries);
3366                 retries++;
3367
3368                 if (block_group->ro)
3369                         goto out_free;
3370
3371                 ret = create_free_space_inode(root, trans, block_group, path);
3372                 if (ret)
3373                         goto out_free;
3374                 goto again;
3375         }
3376
3377         /* We've already setup this transaction, go ahead and exit */
3378         if (block_group->cache_generation == trans->transid &&
3379             i_size_read(inode)) {
3380                 dcs = BTRFS_DC_SETUP;
3381                 goto out_put;
3382         }
3383
3384         /*
3385          * We want to set the generation to 0, that way if anything goes wrong
3386          * from here on out we know not to trust this cache when we load up next
3387          * time.
3388          */
3389         BTRFS_I(inode)->generation = 0;
3390         ret = btrfs_update_inode(trans, root, inode);
3391         if (ret) {
3392                 /*
3393                  * So theoretically we could recover from this, simply set the
3394                  * super cache generation to 0 so we know to invalidate the
3395                  * cache, but then we'd have to keep track of the block groups
3396                  * that fail this way so we know we _have_ to reset this cache
3397                  * before the next commit or risk reading stale cache.  So to
3398                  * limit our exposure to horrible edge cases lets just abort the
3399                  * transaction, this only happens in really bad situations
3400                  * anyway.
3401                  */
3402                 btrfs_abort_transaction(trans, root, ret);
3403                 goto out_put;
3404         }
3405         WARN_ON(ret);
3406
3407         if (i_size_read(inode) > 0) {
3408                 ret = btrfs_check_trunc_cache_free_space(root,
3409                                         &root->fs_info->global_block_rsv);
3410                 if (ret)
3411                         goto out_put;
3412
3413                 ret = btrfs_truncate_free_space_cache(root, trans, NULL, inode);
3414                 if (ret)
3415                         goto out_put;
3416         }
3417
3418         spin_lock(&block_group->lock);
3419         if (block_group->cached != BTRFS_CACHE_FINISHED ||
3420             !btrfs_test_opt(root, SPACE_CACHE)) {
3421                 /*
3422                  * don't bother trying to write stuff out _if_
3423                  * a) we're not cached,
3424                  * b) we're with nospace_cache mount option.
3425                  */
3426                 dcs = BTRFS_DC_WRITTEN;
3427                 spin_unlock(&block_group->lock);
3428                 goto out_put;
3429         }
3430         spin_unlock(&block_group->lock);
3431
3432         /*
3433          * We hit an ENOSPC when setting up the cache in this transaction, just
3434          * skip doing the setup, we've already cleared the cache so we're safe.
3435          */
3436         if (test_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags)) {
3437                 ret = -ENOSPC;
3438                 goto out_put;
3439         }
3440
3441         /*
3442          * Try to preallocate enough space based on how big the block group is.
3443          * Keep in mind this has to include any pinned space which could end up
3444          * taking up quite a bit since it's not folded into the other space
3445          * cache.
3446          */
3447         num_pages = div_u64(block_group->key.offset, 256 * 1024 * 1024);
3448         if (!num_pages)
3449                 num_pages = 1;
3450
3451         num_pages *= 16;
3452         num_pages *= PAGE_CACHE_SIZE;
3453
3454         ret = btrfs_check_data_free_space(inode, 0, num_pages);
3455         if (ret)
3456                 goto out_put;
3457
3458         ret = btrfs_prealloc_file_range_trans(inode, trans, 0, 0, num_pages,
3459                                               num_pages, num_pages,
3460                                               &alloc_hint);
3461         /*
3462          * Our cache requires contiguous chunks so that we don't modify a bunch
3463          * of metadata or split extents when writing the cache out, which means
3464          * we can enospc if we are heavily fragmented in addition to just normal
3465          * out of space conditions.  So if we hit this just skip setting up any
3466          * other block groups for this transaction, maybe we'll unpin enough
3467          * space the next time around.
3468          */
3469         if (!ret)
3470                 dcs = BTRFS_DC_SETUP;
3471         else if (ret == -ENOSPC)
3472                 set_bit(BTRFS_TRANS_CACHE_ENOSPC, &trans->transaction->flags);
3473         btrfs_free_reserved_data_space(inode, 0, num_pages);
3474
3475 out_put:
3476         iput(inode);
3477 out_free:
3478         btrfs_release_path(path);
3479 out:
3480         spin_lock(&block_group->lock);
3481         if (!ret && dcs == BTRFS_DC_SETUP)
3482                 block_group->cache_generation = trans->transid;
3483         block_group->disk_cache_state = dcs;
3484         spin_unlock(&block_group->lock);
3485
3486         return ret;
3487 }
3488
3489 int btrfs_setup_space_cache(struct btrfs_trans_handle *trans,
3490                             struct btrfs_root *root)
3491 {
3492         struct btrfs_block_group_cache *cache, *tmp;
3493         struct btrfs_transaction *cur_trans = trans->transaction;
3494         struct btrfs_path *path;
3495
3496         if (list_empty(&cur_trans->dirty_bgs) ||
3497             !btrfs_test_opt(root, SPACE_CACHE))
3498                 return 0;
3499
3500         path = btrfs_alloc_path();
3501         if (!path)
3502                 return -ENOMEM;
3503
3504         /* Could add new block groups, use _safe just in case */
3505         list_for_each_entry_safe(cache, tmp, &cur_trans->dirty_bgs,
3506                                  dirty_list) {
3507                 if (cache->disk_cache_state == BTRFS_DC_CLEAR)
3508                         cache_save_setup(cache, trans, path);
3509         }
3510
3511         btrfs_free_path(path);
3512         return 0;
3513 }
3514
3515 /*
3516  * transaction commit does final block group cache writeback during a
3517  * critical section where nothing is allowed to change the FS.  This is
3518  * required in order for the cache to actually match the block group,
3519  * but can introduce a lot of latency into the commit.
3520  *
3521  * So, btrfs_start_dirty_block_groups is here to kick off block group
3522  * cache IO.  There's a chance we'll have to redo some of it if the
3523  * block group changes again during the commit, but it greatly reduces
3524  * the commit latency by getting rid of the easy block groups while
3525  * we're still allowing others to join the commit.
3526  */
3527 int btrfs_start_dirty_block_groups(struct btrfs_trans_handle *trans,
3528                                    struct btrfs_root *root)
3529 {
3530         struct btrfs_block_group_cache *cache;
3531         struct btrfs_transaction *cur_trans = trans->transaction;
3532         int ret = 0;
3533         int should_put;
3534         struct btrfs_path *path = NULL;
3535         LIST_HEAD(dirty);
3536         struct list_head *io = &cur_trans->io_bgs;
3537         int num_started = 0;
3538         int loops = 0;
3539
3540         spin_lock(&cur_trans->dirty_bgs_lock);
3541         if (list_empty(&cur_trans->dirty_bgs)) {
3542                 spin_unlock(&cur_trans->dirty_bgs_lock);
3543                 return 0;
3544         }
3545         list_splice_init(&cur_trans->dirty_bgs, &dirty);
3546         spin_unlock(&cur_trans->dirty_bgs_lock);
3547
3548 again:
3549         /*
3550          * make sure all the block groups on our dirty list actually
3551          * exist
3552          */
3553         btrfs_create_pending_block_groups(trans, root);
3554
3555         if (!path) {
3556                 path = btrfs_alloc_path();
3557                 if (!path)
3558                         return -ENOMEM;
3559         }
3560
3561         /*
3562          * cache_write_mutex is here only to save us from balance or automatic
3563          * removal of empty block groups deleting this block group while we are
3564          * writing out the cache
3565          */
3566         mutex_lock(&trans->transaction->cache_write_mutex);
3567         while (!list_empty(&dirty)) {
3568                 cache = list_first_entry(&dirty,
3569                                          struct btrfs_block_group_cache,
3570                                          dirty_list);
3571                 /*
3572                  * this can happen if something re-dirties a block
3573                  * group that is already under IO.  Just wait for it to
3574                  * finish and then do it all again
3575                  */
3576                 if (!list_empty(&cache->io_list)) {
3577                         list_del_init(&cache->io_list);
3578                         btrfs_wait_cache_io(root, trans, cache,
3579                                             &cache->io_ctl, path,
3580                                             cache->key.objectid);
3581                         btrfs_put_block_group(cache);
3582                 }
3583
3584
3585                 /*
3586                  * btrfs_wait_cache_io uses the cache->dirty_list to decide
3587                  * if it should update the cache_state.  Don't delete
3588                  * until after we wait.
3589                  *
3590                  * Since we're not running in the commit critical section
3591                  * we need the dirty_bgs_lock to protect from update_block_group
3592                  */
3593                 spin_lock(&cur_trans->dirty_bgs_lock);
3594                 list_del_init(&cache->dirty_list);
3595                 spin_unlock(&cur_trans->dirty_bgs_lock);
3596
3597                 should_put = 1;
3598
3599                 cache_save_setup(cache, trans, path);
3600
3601                 if (cache->disk_cache_state == BTRFS_DC_SETUP) {
3602                         cache->io_ctl.inode = NULL;
3603                         ret = btrfs_write_out_cache(root, trans, cache, path);
3604                         if (ret == 0 && cache->io_ctl.inode) {
3605                                 num_started++;
3606                                 should_put = 0;
3607
3608                                 /*
3609                                  * the cache_write_mutex is protecting
3610                                  * the io_list
3611                                  */
3612                                 list_add_tail(&cache->io_list, io);
3613                         } else {
3614                                 /*
3615                                  * if we failed to write the cache, the
3616                                  * generation will be bad and life goes on
3617                                  */
3618                                 ret = 0;
3619                         }
3620                 }
3621                 if (!ret) {
3622                         ret = write_one_cache_group(trans, root, path, cache);
3623                         /*
3624                          * Our block group might still be attached to the list
3625                          * of new block groups in the transaction handle of some
3626                          * other task (struct btrfs_trans_handle->new_bgs). This
3627                          * means its block group item isn't yet in the extent
3628                          * tree. If this happens ignore the error, as we will
3629                          * try again later in the critical section of the
3630                          * transaction commit.
3631                          */
3632                         if (ret == -ENOENT) {
3633                                 ret = 0;
3634                                 spin_lock(&cur_trans->dirty_bgs_lock);
3635                                 if (list_empty(&cache->dirty_list)) {
3636                                         list_add_tail(&cache->dirty_list,
3637                                                       &cur_trans->dirty_bgs);
3638                                         btrfs_get_block_group(cache);
3639                                 }
3640                                 spin_unlock(&cur_trans->dirty_bgs_lock);
3641                         } else if (ret) {
3642                                 btrfs_abort_transaction(trans, root, ret);
3643                         }
3644                 }
3645
3646                 /* if its not on the io list, we need to put the block group */
3647                 if (should_put)
3648                         btrfs_put_block_group(cache);
3649
3650                 if (ret)
3651                         break;
3652
3653                 /*
3654                  * Avoid blocking other tasks for too long. It might even save
3655                  * us from writing caches for block groups that are going to be
3656                  * removed.
3657                  */
3658                 mutex_unlock(&trans->transaction->cache_write_mutex);
3659                 mutex_lock(&trans->transaction->cache_write_mutex);
3660         }
3661         mutex_unlock(&trans->transaction->cache_write_mutex);
3662
3663         /*
3664          * go through delayed refs for all the stuff we've just kicked off
3665          * and then loop back (just once)
3666          */
3667         ret = btrfs_run_delayed_refs(trans, root, 0);
3668         if (!ret && loops == 0) {
3669                 loops++;
3670                 spin_lock(&cur_trans->dirty_bgs_lock);
3671                 list_splice_init(&cur_trans->dirty_bgs, &dirty);
3672                 /*
3673                  * dirty_bgs_lock protects us from concurrent block group
3674                  * deletes too (not just cache_write_mutex).
3675                  */
3676                 if (!list_empty(&dirty)) {
3677                         spin_unlock(&cur_trans->dirty_bgs_lock);
3678                         goto again;
3679                 }
3680                 spin_unlock(&cur_trans->dirty_bgs_lock);
3681         }
3682
3683         btrfs_free_path(path);
3684         return ret;
3685 }
3686
3687 int btrfs_write_dirty_block_groups(struct btrfs_trans_handle *trans,
3688                                    struct btrfs_root *root)
3689 {
3690         struct btrfs_block_group_cache *cache;
3691         struct btrfs_transaction *cur_trans = trans->transaction;
3692         int ret = 0;
3693         int should_put;
3694         struct btrfs_path *path;
3695         struct list_head *io = &cur_trans->io_bgs;
3696         int num_started = 0;
3697
3698         path = btrfs_alloc_path();
3699         if (!path)
3700                 return -ENOMEM;
3701
3702         /*
3703          * Even though we are in the critical section of the transaction commit,
3704          * we can still have concurrent tasks adding elements to this
3705          * transaction's list of dirty block groups. These tasks correspond to
3706          * endio free space workers started when writeback finishes for a
3707          * space cache, which run inode.c:btrfs_finish_ordered_io(), and can
3708          * allocate new block groups as a result of COWing nodes of the root
3709          * tree when updating the free space inode. The writeback for the space
3710          * caches is triggered by an earlier call to
3711          * btrfs_start_dirty_block_groups() and iterations of the following
3712          * loop.
3713          * Also we want to do the cache_save_setup first and then run the
3714          * delayed refs to make sure we have the best chance at doing this all
3715          * in one shot.
3716          */
3717         spin_lock(&cur_trans->dirty_bgs_lock);
3718         while (!list_empty(&cur_trans->dirty_bgs)) {
3719                 cache = list_first_entry(&cur_trans->dirty_bgs,
3720                                          struct btrfs_block_group_cache,
3721                                          dirty_list);
3722
3723                 /*
3724                  * this can happen if cache_save_setup re-dirties a block
3725                  * group that is already under IO.  Just wait for it to
3726                  * finish and then do it all again
3727                  */
3728                 if (!list_empty(&cache->io_list)) {
3729                         spin_unlock(&cur_trans->dirty_bgs_lock);
3730                         list_del_init(&cache->io_list);
3731                         btrfs_wait_cache_io(root, trans, cache,
3732                                             &cache->io_ctl, path,
3733                                             cache->key.objectid);
3734                         btrfs_put_block_group(cache);
3735                         spin_lock(&cur_trans->dirty_bgs_lock);
3736                 }
3737
3738                 /*
3739                  * don't remove from the dirty list until after we've waited
3740                  * on any pending IO
3741                  */
3742                 list_del_init(&cache->dirty_list);
3743                 spin_unlock(&cur_trans->dirty_bgs_lock);
3744                 should_put = 1;
3745
3746                 cache_save_setup(cache, trans, path);
3747
3748                 if (!ret)
3749                         ret = btrfs_run_delayed_refs(trans, root, (unsigned long) -1);
3750
3751                 if (!ret && cache->disk_cache_state == BTRFS_DC_SETUP) {
3752                         cache->io_ctl.inode = NULL;
3753                         ret = btrfs_write_out_cache(root, trans, cache, path);
3754                         if (ret == 0 && cache->io_ctl.inode) {
3755                                 num_started++;
3756                                 should_put = 0;
3757                                 list_add_tail(&cache->io_list, io);
3758                         } else {
3759                                 /*
3760                                  * if we failed to write the cache, the
3761                                  * generation will be bad and life goes on
3762                                  */
3763                                 ret = 0;
3764                         }
3765                 }
3766                 if (!ret) {
3767                         ret = write_one_cache_group(trans, root, path, cache);
3768                         if (ret)
3769                                 btrfs_abort_transaction(trans, root, ret);
3770                 }
3771
3772                 /* if its not on the io list, we need to put the block group */
3773                 if (should_put)
3774                         btrfs_put_block_group(cache);
3775                 spin_lock(&cur_trans->dirty_bgs_lock);
3776         }
3777         spin_unlock(&cur_trans->dirty_bgs_lock);
3778
3779         while (!list_empty(io)) {
3780                 cache = list_first_entry(io, struct btrfs_block_group_cache,
3781                                          io_list);
3782                 list_del_init(&cache->io_list);
3783                 btrfs_wait_cache_io(root, trans, cache,
3784                                     &cache->io_ctl, path, cache->key.objectid);
3785                 btrfs_put_block_group(cache);
3786         }
3787
3788         btrfs_free_path(path);
3789         return ret;
3790 }
3791
3792 int btrfs_extent_readonly(struct btrfs_root *root, u64 bytenr)
3793 {
3794         struct btrfs_block_group_cache *block_group;
3795         int readonly = 0;
3796
3797         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
3798         if (!block_group || block_group->ro)
3799                 readonly = 1;
3800         if (block_group)
3801                 btrfs_put_block_group(block_group);
3802         return readonly;
3803 }
3804
3805 static const char *alloc_name(u64 flags)
3806 {
3807         switch (flags) {
3808         case BTRFS_BLOCK_GROUP_METADATA|BTRFS_BLOCK_GROUP_DATA:
3809                 return "mixed";
3810         case BTRFS_BLOCK_GROUP_METADATA:
3811                 return "metadata";
3812         case BTRFS_BLOCK_GROUP_DATA:
3813                 return "data";
3814         case BTRFS_BLOCK_GROUP_SYSTEM:
3815                 return "system";
3816         default:
3817                 WARN_ON(1);
3818                 return "invalid-combination";
3819         };
3820 }
3821
3822 static int update_space_info(struct btrfs_fs_info *info, u64 flags,
3823                              u64 total_bytes, u64 bytes_used,
3824                              struct btrfs_space_info **space_info)
3825 {
3826         struct btrfs_space_info *found;
3827         int i;
3828         int factor;
3829         int ret;
3830
3831         if (flags & (BTRFS_BLOCK_GROUP_DUP | BTRFS_BLOCK_GROUP_RAID1 |
3832                      BTRFS_BLOCK_GROUP_RAID10))
3833                 factor = 2;
3834         else
3835                 factor = 1;
3836
3837         found = __find_space_info(info, flags);
3838         if (found) {
3839                 spin_lock(&found->lock);
3840                 found->total_bytes += total_bytes;
3841                 found->disk_total += total_bytes * factor;
3842                 found->bytes_used += bytes_used;
3843                 found->disk_used += bytes_used * factor;
3844                 if (total_bytes > 0)
3845                         found->full = 0;
3846                 spin_unlock(&found->lock);
3847                 *space_info = found;
3848                 return 0;
3849         }
3850         found = kzalloc(sizeof(*found), GFP_NOFS);
3851         if (!found)
3852                 return -ENOMEM;
3853
3854         ret = percpu_counter_init(&found->total_bytes_pinned, 0, GFP_KERNEL);
3855         if (ret) {
3856                 kfree(found);
3857                 return ret;
3858         }
3859
3860         for (i = 0; i < BTRFS_NR_RAID_TYPES; i++)
3861                 INIT_LIST_HEAD(&found->block_groups[i]);
3862         init_rwsem(&found->groups_sem);
3863         spin_lock_init(&found->lock);
3864         found->flags = flags & BTRFS_BLOCK_GROUP_TYPE_MASK;
3865         found->total_bytes = total_bytes;
3866         found->disk_total = total_bytes * factor;
3867         found->bytes_used = bytes_used;
3868         found->disk_used = bytes_used * factor;
3869         found->bytes_pinned = 0;
3870         found->bytes_reserved = 0;
3871         found->bytes_readonly = 0;
3872         found->bytes_may_use = 0;
3873         found->full = 0;
3874         found->max_extent_size = 0;
3875         found->force_alloc = CHUNK_ALLOC_NO_FORCE;
3876         found->chunk_alloc = 0;
3877         found->flush = 0;
3878         init_waitqueue_head(&found->wait);
3879         INIT_LIST_HEAD(&found->ro_bgs);
3880
3881         ret = kobject_init_and_add(&found->kobj, &space_info_ktype,
3882                                     info->space_info_kobj, "%s",
3883                                     alloc_name(found->flags));
3884         if (ret) {
3885                 kfree(found);
3886                 return ret;
3887         }
3888
3889         *space_info = found;
3890         list_add_rcu(&found->list, &info->space_info);
3891         if (flags & BTRFS_BLOCK_GROUP_DATA)
3892                 info->data_sinfo = found;
3893
3894         return ret;
3895 }
3896
3897 static void set_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
3898 {
3899         u64 extra_flags = chunk_to_extended(flags) &
3900                                 BTRFS_EXTENDED_PROFILE_MASK;
3901
3902         write_seqlock(&fs_info->profiles_lock);
3903         if (flags & BTRFS_BLOCK_GROUP_DATA)
3904                 fs_info->avail_data_alloc_bits |= extra_flags;
3905         if (flags & BTRFS_BLOCK_GROUP_METADATA)
3906                 fs_info->avail_metadata_alloc_bits |= extra_flags;
3907         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
3908                 fs_info->avail_system_alloc_bits |= extra_flags;
3909         write_sequnlock(&fs_info->profiles_lock);
3910 }
3911
3912 /*
3913  * returns target flags in extended format or 0 if restripe for this
3914  * chunk_type is not in progress
3915  *
3916  * should be called with either volume_mutex or balance_lock held
3917  */
3918 static u64 get_restripe_target(struct btrfs_fs_info *fs_info, u64 flags)
3919 {
3920         struct btrfs_balance_control *bctl = fs_info->balance_ctl;
3921         u64 target = 0;
3922
3923         if (!bctl)
3924                 return 0;
3925
3926         if (flags & BTRFS_BLOCK_GROUP_DATA &&
3927             bctl->data.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3928                 target = BTRFS_BLOCK_GROUP_DATA | bctl->data.target;
3929         } else if (flags & BTRFS_BLOCK_GROUP_SYSTEM &&
3930                    bctl->sys.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3931                 target = BTRFS_BLOCK_GROUP_SYSTEM | bctl->sys.target;
3932         } else if (flags & BTRFS_BLOCK_GROUP_METADATA &&
3933                    bctl->meta.flags & BTRFS_BALANCE_ARGS_CONVERT) {
3934                 target = BTRFS_BLOCK_GROUP_METADATA | bctl->meta.target;
3935         }
3936
3937         return target;
3938 }
3939
3940 /*
3941  * @flags: available profiles in extended format (see ctree.h)
3942  *
3943  * Returns reduced profile in chunk format.  If profile changing is in
3944  * progress (either running or paused) picks the target profile (if it's
3945  * already available), otherwise falls back to plain reducing.
3946  */
3947 static u64 btrfs_reduce_alloc_profile(struct btrfs_root *root, u64 flags)
3948 {
3949         u64 num_devices = root->fs_info->fs_devices->rw_devices;
3950         u64 target;
3951         u64 raid_type;
3952         u64 allowed = 0;
3953
3954         /*
3955          * see if restripe for this chunk_type is in progress, if so
3956          * try to reduce to the target profile
3957          */
3958         spin_lock(&root->fs_info->balance_lock);
3959         target = get_restripe_target(root->fs_info, flags);
3960         if (target) {
3961                 /* pick target profile only if it's already available */
3962                 if ((flags & target) & BTRFS_EXTENDED_PROFILE_MASK) {
3963                         spin_unlock(&root->fs_info->balance_lock);
3964                         return extended_to_chunk(target);
3965                 }
3966         }
3967         spin_unlock(&root->fs_info->balance_lock);
3968
3969         /* First, mask out the RAID levels which aren't possible */
3970         for (raid_type = 0; raid_type < BTRFS_NR_RAID_TYPES; raid_type++) {
3971                 if (num_devices >= btrfs_raid_array[raid_type].devs_min)
3972                         allowed |= btrfs_raid_group[raid_type];
3973         }
3974         allowed &= flags;
3975
3976         if (allowed & BTRFS_BLOCK_GROUP_RAID6)
3977                 allowed = BTRFS_BLOCK_GROUP_RAID6;
3978         else if (allowed & BTRFS_BLOCK_GROUP_RAID5)
3979                 allowed = BTRFS_BLOCK_GROUP_RAID5;
3980         else if (allowed & BTRFS_BLOCK_GROUP_RAID10)
3981                 allowed = BTRFS_BLOCK_GROUP_RAID10;
3982         else if (allowed & BTRFS_BLOCK_GROUP_RAID1)
3983                 allowed = BTRFS_BLOCK_GROUP_RAID1;
3984         else if (allowed & BTRFS_BLOCK_GROUP_RAID0)
3985                 allowed = BTRFS_BLOCK_GROUP_RAID0;
3986
3987         flags &= ~BTRFS_BLOCK_GROUP_PROFILE_MASK;
3988
3989         return extended_to_chunk(flags | allowed);
3990 }
3991
3992 static u64 get_alloc_profile(struct btrfs_root *root, u64 orig_flags)
3993 {
3994         unsigned seq;
3995         u64 flags;
3996
3997         do {
3998                 flags = orig_flags;
3999                 seq = read_seqbegin(&root->fs_info->profiles_lock);
4000
4001                 if (flags & BTRFS_BLOCK_GROUP_DATA)
4002                         flags |= root->fs_info->avail_data_alloc_bits;
4003                 else if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
4004                         flags |= root->fs_info->avail_system_alloc_bits;
4005                 else if (flags & BTRFS_BLOCK_GROUP_METADATA)
4006                         flags |= root->fs_info->avail_metadata_alloc_bits;
4007         } while (read_seqretry(&root->fs_info->profiles_lock, seq));
4008
4009         return btrfs_reduce_alloc_profile(root, flags);
4010 }
4011
4012 u64 btrfs_get_alloc_profile(struct btrfs_root *root, int data)
4013 {
4014         u64 flags;
4015         u64 ret;
4016
4017         if (data)
4018                 flags = BTRFS_BLOCK_GROUP_DATA;
4019         else if (root == root->fs_info->chunk_root)
4020                 flags = BTRFS_BLOCK_GROUP_SYSTEM;
4021         else
4022                 flags = BTRFS_BLOCK_GROUP_METADATA;
4023
4024         ret = get_alloc_profile(root, flags);
4025         return ret;
4026 }
4027
4028 int btrfs_alloc_data_chunk_ondemand(struct inode *inode, u64 bytes)
4029 {
4030         struct btrfs_space_info *data_sinfo;
4031         struct btrfs_root *root = BTRFS_I(inode)->root;
4032         struct btrfs_fs_info *fs_info = root->fs_info;
4033         u64 used;
4034         int ret = 0;
4035         int need_commit = 2;
4036         int have_pinned_space;
4037
4038         /* make sure bytes are sectorsize aligned */
4039         bytes = ALIGN(bytes, root->sectorsize);
4040
4041         if (btrfs_is_free_space_inode(inode)) {
4042                 need_commit = 0;
4043                 ASSERT(current->journal_info);
4044         }
4045
4046         data_sinfo = fs_info->data_sinfo;
4047         if (!data_sinfo)
4048                 goto alloc;
4049
4050 again:
4051         /* make sure we have enough space to handle the data first */
4052         spin_lock(&data_sinfo->lock);
4053         used = data_sinfo->bytes_used + data_sinfo->bytes_reserved +
4054                 data_sinfo->bytes_pinned + data_sinfo->bytes_readonly +
4055                 data_sinfo->bytes_may_use;
4056
4057         if (used + bytes > data_sinfo->total_bytes) {
4058                 struct btrfs_trans_handle *trans;
4059
4060                 /*
4061                  * if we don't have enough free bytes in this space then we need
4062                  * to alloc a new chunk.
4063                  */
4064                 if (!data_sinfo->full) {
4065                         u64 alloc_target;
4066
4067                         data_sinfo->force_alloc = CHUNK_ALLOC_FORCE;
4068                         spin_unlock(&data_sinfo->lock);
4069 alloc:
4070                         alloc_target = btrfs_get_alloc_profile(root, 1);
4071                         /*
4072                          * It is ugly that we don't call nolock join
4073                          * transaction for the free space inode case here.
4074                          * But it is safe because we only do the data space
4075                          * reservation for the free space cache in the
4076                          * transaction context, the common join transaction
4077                          * just increase the counter of the current transaction
4078                          * handler, doesn't try to acquire the trans_lock of
4079                          * the fs.
4080                          */
4081                         trans = btrfs_join_transaction(root);
4082                         if (IS_ERR(trans))
4083                                 return PTR_ERR(trans);
4084
4085                         ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4086                                              alloc_target,
4087                                              CHUNK_ALLOC_NO_FORCE);
4088                         btrfs_end_transaction(trans, root);
4089                         if (ret < 0) {
4090                                 if (ret != -ENOSPC)
4091                                         return ret;
4092                                 else {
4093                                         have_pinned_space = 1;
4094                                         goto commit_trans;
4095                                 }
4096                         }
4097
4098                         if (!data_sinfo)
4099                                 data_sinfo = fs_info->data_sinfo;
4100
4101                         goto again;
4102                 }
4103
4104                 /*
4105                  * If we don't have enough pinned space to deal with this
4106                  * allocation, and no removed chunk in current transaction,
4107                  * don't bother committing the transaction.
4108                  */
4109                 have_pinned_space = percpu_counter_compare(
4110                         &data_sinfo->total_bytes_pinned,
4111                         used + bytes - data_sinfo->total_bytes);
4112                 spin_unlock(&data_sinfo->lock);
4113
4114                 /* commit the current transaction and try again */
4115 commit_trans:
4116                 if (need_commit &&
4117                     !atomic_read(&root->fs_info->open_ioctl_trans)) {
4118                         need_commit--;
4119
4120                         if (need_commit > 0)
4121                                 btrfs_wait_ordered_roots(fs_info, -1);
4122
4123                         trans = btrfs_join_transaction(root);
4124                         if (IS_ERR(trans))
4125                                 return PTR_ERR(trans);
4126                         if (have_pinned_space >= 0 ||
4127                             test_bit(BTRFS_TRANS_HAVE_FREE_BGS,
4128                                      &trans->transaction->flags) ||
4129                             need_commit > 0) {
4130                                 ret = btrfs_commit_transaction(trans, root);
4131                                 if (ret)
4132                                         return ret;
4133                                 /*
4134                                  * make sure that all running delayed iput are
4135                                  * done
4136                                  */
4137                                 down_write(&root->fs_info->delayed_iput_sem);
4138                                 up_write(&root->fs_info->delayed_iput_sem);
4139                                 goto again;
4140                         } else {
4141                                 btrfs_end_transaction(trans, root);
4142                         }
4143                 }
4144
4145                 trace_btrfs_space_reservation(root->fs_info,
4146                                               "space_info:enospc",
4147                                               data_sinfo->flags, bytes, 1);
4148                 return -ENOSPC;
4149         }
4150         data_sinfo->bytes_may_use += bytes;
4151         trace_btrfs_space_reservation(root->fs_info, "space_info",
4152                                       data_sinfo->flags, bytes, 1);
4153         spin_unlock(&data_sinfo->lock);
4154
4155         return ret;
4156 }
4157
4158 /*
4159  * New check_data_free_space() with ability for precious data reservation
4160  * Will replace old btrfs_check_data_free_space(), but for patch split,
4161  * add a new function first and then replace it.
4162  */
4163 int btrfs_check_data_free_space(struct inode *inode, u64 start, u64 len)
4164 {
4165         struct btrfs_root *root = BTRFS_I(inode)->root;
4166         int ret;
4167
4168         /* align the range */
4169         len = round_up(start + len, root->sectorsize) -
4170               round_down(start, root->sectorsize);
4171         start = round_down(start, root->sectorsize);
4172
4173         ret = btrfs_alloc_data_chunk_ondemand(inode, len);
4174         if (ret < 0)
4175                 return ret;
4176
4177         /*
4178          * Use new btrfs_qgroup_reserve_data to reserve precious data space
4179          *
4180          * TODO: Find a good method to avoid reserve data space for NOCOW
4181          * range, but don't impact performance on quota disable case.
4182          */
4183         ret = btrfs_qgroup_reserve_data(inode, start, len);
4184         return ret;
4185 }
4186
4187 /*
4188  * Called if we need to clear a data reservation for this inode
4189  * Normally in a error case.
4190  *
4191  * This one will *NOT* use accurate qgroup reserved space API, just for case
4192  * which we can't sleep and is sure it won't affect qgroup reserved space.
4193  * Like clear_bit_hook().
4194  */
4195 void btrfs_free_reserved_data_space_noquota(struct inode *inode, u64 start,
4196                                             u64 len)
4197 {
4198         struct btrfs_root *root = BTRFS_I(inode)->root;
4199         struct btrfs_space_info *data_sinfo;
4200
4201         /* Make sure the range is aligned to sectorsize */
4202         len = round_up(start + len, root->sectorsize) -
4203               round_down(start, root->sectorsize);
4204         start = round_down(start, root->sectorsize);
4205
4206         data_sinfo = root->fs_info->data_sinfo;
4207         spin_lock(&data_sinfo->lock);
4208         if (WARN_ON(data_sinfo->bytes_may_use < len))
4209                 data_sinfo->bytes_may_use = 0;
4210         else
4211                 data_sinfo->bytes_may_use -= len;
4212         trace_btrfs_space_reservation(root->fs_info, "space_info",
4213                                       data_sinfo->flags, len, 0);
4214         spin_unlock(&data_sinfo->lock);
4215 }
4216
4217 /*
4218  * Called if we need to clear a data reservation for this inode
4219  * Normally in a error case.
4220  *
4221  * This one will handle the per-indoe data rsv map for accurate reserved
4222  * space framework.
4223  */
4224 void btrfs_free_reserved_data_space(struct inode *inode, u64 start, u64 len)
4225 {
4226         btrfs_free_reserved_data_space_noquota(inode, start, len);
4227         btrfs_qgroup_free_data(inode, start, len);
4228 }
4229
4230 static void force_metadata_allocation(struct btrfs_fs_info *info)
4231 {
4232         struct list_head *head = &info->space_info;
4233         struct btrfs_space_info *found;
4234
4235         rcu_read_lock();
4236         list_for_each_entry_rcu(found, head, list) {
4237                 if (found->flags & BTRFS_BLOCK_GROUP_METADATA)
4238                         found->force_alloc = CHUNK_ALLOC_FORCE;
4239         }
4240         rcu_read_unlock();
4241 }
4242
4243 static inline u64 calc_global_rsv_need_space(struct btrfs_block_rsv *global)
4244 {
4245         return (global->size << 1);
4246 }
4247
4248 static int should_alloc_chunk(struct btrfs_root *root,
4249                               struct btrfs_space_info *sinfo, int force)
4250 {
4251         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4252         u64 num_bytes = sinfo->total_bytes - sinfo->bytes_readonly;
4253         u64 num_allocated = sinfo->bytes_used + sinfo->bytes_reserved;
4254         u64 thresh;
4255
4256         if (force == CHUNK_ALLOC_FORCE)
4257                 return 1;
4258
4259         /*
4260          * We need to take into account the global rsv because for all intents
4261          * and purposes it's used space.  Don't worry about locking the
4262          * global_rsv, it doesn't change except when the transaction commits.
4263          */
4264         if (sinfo->flags & BTRFS_BLOCK_GROUP_METADATA)
4265                 num_allocated += calc_global_rsv_need_space(global_rsv);
4266
4267         /*
4268          * in limited mode, we want to have some free space up to
4269          * about 1% of the FS size.
4270          */
4271         if (force == CHUNK_ALLOC_LIMITED) {
4272                 thresh = btrfs_super_total_bytes(root->fs_info->super_copy);
4273                 thresh = max_t(u64, 64 * 1024 * 1024,
4274                                div_factor_fine(thresh, 1));
4275
4276                 if (num_bytes - num_allocated < thresh)
4277                         return 1;
4278         }
4279
4280         if (num_allocated + 2 * 1024 * 1024 < div_factor(num_bytes, 8))
4281                 return 0;
4282         return 1;
4283 }
4284
4285 static u64 get_profile_num_devs(struct btrfs_root *root, u64 type)
4286 {
4287         u64 num_dev;
4288
4289         if (type & (BTRFS_BLOCK_GROUP_RAID10 |
4290                     BTRFS_BLOCK_GROUP_RAID0 |
4291                     BTRFS_BLOCK_GROUP_RAID5 |
4292                     BTRFS_BLOCK_GROUP_RAID6))
4293                 num_dev = root->fs_info->fs_devices->rw_devices;
4294         else if (type & BTRFS_BLOCK_GROUP_RAID1)
4295                 num_dev = 2;
4296         else
4297                 num_dev = 1;    /* DUP or single */
4298
4299         return num_dev;
4300 }
4301
4302 /*
4303  * If @is_allocation is true, reserve space in the system space info necessary
4304  * for allocating a chunk, otherwise if it's false, reserve space necessary for
4305  * removing a chunk.
4306  */
4307 void check_system_chunk(struct btrfs_trans_handle *trans,
4308                         struct btrfs_root *root,
4309                         u64 type)
4310 {
4311         struct btrfs_space_info *info;
4312         u64 left;
4313         u64 thresh;
4314         int ret = 0;
4315         u64 num_devs;
4316
4317         /*
4318          * Needed because we can end up allocating a system chunk and for an
4319          * atomic and race free space reservation in the chunk block reserve.
4320          */
4321         ASSERT(mutex_is_locked(&root->fs_info->chunk_mutex));
4322
4323         info = __find_space_info(root->fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
4324         spin_lock(&info->lock);
4325         left = info->total_bytes - info->bytes_used - info->bytes_pinned -
4326                 info->bytes_reserved - info->bytes_readonly -
4327                 info->bytes_may_use;
4328         spin_unlock(&info->lock);
4329
4330         num_devs = get_profile_num_devs(root, type);
4331
4332         /* num_devs device items to update and 1 chunk item to add or remove */
4333         thresh = btrfs_calc_trunc_metadata_size(root, num_devs) +
4334                 btrfs_calc_trans_metadata_size(root, 1);
4335
4336         if (left < thresh && btrfs_test_opt(root, ENOSPC_DEBUG)) {
4337                 btrfs_info(root->fs_info, "left=%llu, need=%llu, flags=%llu",
4338                         left, thresh, type);
4339                 dump_space_info(info, 0, 0);
4340         }
4341
4342         if (left < thresh) {
4343                 u64 flags;
4344
4345                 flags = btrfs_get_alloc_profile(root->fs_info->chunk_root, 0);
4346                 /*
4347                  * Ignore failure to create system chunk. We might end up not
4348                  * needing it, as we might not need to COW all nodes/leafs from
4349                  * the paths we visit in the chunk tree (they were already COWed
4350                  * or created in the current transaction for example).
4351                  */
4352                 ret = btrfs_alloc_chunk(trans, root, flags);
4353         }
4354
4355         if (!ret) {
4356                 ret = btrfs_block_rsv_add(root->fs_info->chunk_root,
4357                                           &root->fs_info->chunk_block_rsv,
4358                                           thresh, BTRFS_RESERVE_NO_FLUSH);
4359                 if (!ret)
4360                         trans->chunk_bytes_reserved += thresh;
4361         }
4362 }
4363
4364 static int do_chunk_alloc(struct btrfs_trans_handle *trans,
4365                           struct btrfs_root *extent_root, u64 flags, int force)
4366 {
4367         struct btrfs_space_info *space_info;
4368         struct btrfs_fs_info *fs_info = extent_root->fs_info;
4369         int wait_for_alloc = 0;
4370         int ret = 0;
4371
4372         /* Don't re-enter if we're already allocating a chunk */
4373         if (trans->allocating_chunk)
4374                 return -ENOSPC;
4375
4376         space_info = __find_space_info(extent_root->fs_info, flags);
4377         if (!space_info) {
4378                 ret = update_space_info(extent_root->fs_info, flags,
4379                                         0, 0, &space_info);
4380                 BUG_ON(ret); /* -ENOMEM */
4381         }
4382         BUG_ON(!space_info); /* Logic error */
4383
4384 again:
4385         spin_lock(&space_info->lock);
4386         if (force < space_info->force_alloc)
4387                 force = space_info->force_alloc;
4388         if (space_info->full) {
4389                 if (should_alloc_chunk(extent_root, space_info, force))
4390                         ret = -ENOSPC;
4391                 else
4392                         ret = 0;
4393                 spin_unlock(&space_info->lock);
4394                 return ret;
4395         }
4396
4397         if (!should_alloc_chunk(extent_root, space_info, force)) {
4398                 spin_unlock(&space_info->lock);
4399                 return 0;
4400         } else if (space_info->chunk_alloc) {
4401                 wait_for_alloc = 1;
4402         } else {
4403                 space_info->chunk_alloc = 1;
4404         }
4405
4406         spin_unlock(&space_info->lock);
4407
4408         mutex_lock(&fs_info->chunk_mutex);
4409
4410         /*
4411          * The chunk_mutex is held throughout the entirety of a chunk
4412          * allocation, so once we've acquired the chunk_mutex we know that the
4413          * other guy is done and we need to recheck and see if we should
4414          * allocate.
4415          */
4416         if (wait_for_alloc) {
4417                 mutex_unlock(&fs_info->chunk_mutex);
4418                 wait_for_alloc = 0;
4419                 goto again;
4420         }
4421
4422         trans->allocating_chunk = true;
4423
4424         /*
4425          * If we have mixed data/metadata chunks we want to make sure we keep
4426          * allocating mixed chunks instead of individual chunks.
4427          */
4428         if (btrfs_mixed_space_info(space_info))
4429                 flags |= (BTRFS_BLOCK_GROUP_DATA | BTRFS_BLOCK_GROUP_METADATA);
4430
4431         /*
4432          * if we're doing a data chunk, go ahead and make sure that
4433          * we keep a reasonable number of metadata chunks allocated in the
4434          * FS as well.
4435          */
4436         if (flags & BTRFS_BLOCK_GROUP_DATA && fs_info->metadata_ratio) {
4437                 fs_info->data_chunk_allocations++;
4438                 if (!(fs_info->data_chunk_allocations %
4439                       fs_info->metadata_ratio))
4440                         force_metadata_allocation(fs_info);
4441         }
4442
4443         /*
4444          * Check if we have enough space in SYSTEM chunk because we may need
4445          * to update devices.
4446          */
4447         check_system_chunk(trans, extent_root, flags);
4448
4449         ret = btrfs_alloc_chunk(trans, extent_root, flags);
4450         trans->allocating_chunk = false;
4451
4452         spin_lock(&space_info->lock);
4453         if (ret < 0 && ret != -ENOSPC)
4454                 goto out;
4455         if (ret)
4456                 space_info->full = 1;
4457         else
4458                 ret = 1;
4459
4460         space_info->force_alloc = CHUNK_ALLOC_NO_FORCE;
4461 out:
4462         space_info->chunk_alloc = 0;
4463         spin_unlock(&space_info->lock);
4464         mutex_unlock(&fs_info->chunk_mutex);
4465         /*
4466          * When we allocate a new chunk we reserve space in the chunk block
4467          * reserve to make sure we can COW nodes/leafs in the chunk tree or
4468          * add new nodes/leafs to it if we end up needing to do it when
4469          * inserting the chunk item and updating device items as part of the
4470          * second phase of chunk allocation, performed by
4471          * btrfs_finish_chunk_alloc(). So make sure we don't accumulate a
4472          * large number of new block groups to create in our transaction
4473          * handle's new_bgs list to avoid exhausting the chunk block reserve
4474          * in extreme cases - like having a single transaction create many new
4475          * block groups when starting to write out the free space caches of all
4476          * the block groups that were made dirty during the lifetime of the
4477          * transaction.
4478          */
4479         if (trans->can_flush_pending_bgs &&
4480             trans->chunk_bytes_reserved >= (2 * 1024 * 1024ull)) {
4481                 btrfs_create_pending_block_groups(trans, trans->root);
4482                 btrfs_trans_release_chunk_metadata(trans);
4483         }
4484         return ret;
4485 }
4486
4487 static int can_overcommit(struct btrfs_root *root,
4488                           struct btrfs_space_info *space_info, u64 bytes,
4489                           enum btrfs_reserve_flush_enum flush)
4490 {
4491         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
4492         u64 profile = btrfs_get_alloc_profile(root, 0);
4493         u64 space_size;
4494         u64 avail;
4495         u64 used;
4496
4497         used = space_info->bytes_used + space_info->bytes_reserved +
4498                 space_info->bytes_pinned + space_info->bytes_readonly;
4499
4500         /*
4501          * We only want to allow over committing if we have lots of actual space
4502          * free, but if we don't have enough space to handle the global reserve
4503          * space then we could end up having a real enospc problem when trying
4504          * to allocate a chunk or some other such important allocation.
4505          */
4506         spin_lock(&global_rsv->lock);
4507         space_size = calc_global_rsv_need_space(global_rsv);
4508         spin_unlock(&global_rsv->lock);
4509         if (used + space_size >= space_info->total_bytes)
4510                 return 0;
4511
4512         used += space_info->bytes_may_use;
4513
4514         spin_lock(&root->fs_info->free_chunk_lock);
4515         avail = root->fs_info->free_chunk_space;
4516         spin_unlock(&root->fs_info->free_chunk_lock);
4517
4518         /*
4519          * If we have dup, raid1 or raid10 then only half of the free
4520          * space is actually useable.  For raid56, the space info used
4521          * doesn't include the parity drive, so we don't have to
4522          * change the math
4523          */
4524         if (profile & (BTRFS_BLOCK_GROUP_DUP |
4525                        BTRFS_BLOCK_GROUP_RAID1 |
4526                        BTRFS_BLOCK_GROUP_RAID10))
4527                 avail >>= 1;
4528
4529         /*
4530          * If we aren't flushing all things, let us overcommit up to
4531          * 1/2th of the space. If we can flush, don't let us overcommit
4532          * too much, let it overcommit up to 1/8 of the space.
4533          */
4534         if (flush == BTRFS_RESERVE_FLUSH_ALL)
4535                 avail >>= 3;
4536         else
4537                 avail >>= 1;
4538
4539         if (used + bytes < space_info->total_bytes + avail)
4540                 return 1;
4541         return 0;
4542 }
4543
4544 static void btrfs_writeback_inodes_sb_nr(struct btrfs_root *root,
4545                                          unsigned long nr_pages, int nr_items)
4546 {
4547         struct super_block *sb = root->fs_info->sb;
4548
4549         if (down_read_trylock(&sb->s_umount)) {
4550                 writeback_inodes_sb_nr(sb, nr_pages, WB_REASON_FS_FREE_SPACE);
4551                 up_read(&sb->s_umount);
4552         } else {
4553                 /*
4554                  * We needn't worry the filesystem going from r/w to r/o though
4555                  * we don't acquire ->s_umount mutex, because the filesystem
4556                  * should guarantee the delalloc inodes list be empty after
4557                  * the filesystem is readonly(all dirty pages are written to
4558                  * the disk).
4559                  */
4560                 btrfs_start_delalloc_roots(root->fs_info, 0, nr_items);
4561                 if (!current->journal_info)
4562                         btrfs_wait_ordered_roots(root->fs_info, nr_items);
4563         }
4564 }
4565
4566 static inline int calc_reclaim_items_nr(struct btrfs_root *root, u64 to_reclaim)
4567 {
4568         u64 bytes;
4569         int nr;
4570
4571         bytes = btrfs_calc_trans_metadata_size(root, 1);
4572         nr = (int)div64_u64(to_reclaim, bytes);
4573         if (!nr)
4574                 nr = 1;
4575         return nr;
4576 }
4577
4578 #define EXTENT_SIZE_PER_ITEM    (256 * 1024)
4579
4580 /*
4581  * shrink metadata reservation for delalloc
4582  */
4583 static void shrink_delalloc(struct btrfs_root *root, u64 to_reclaim, u64 orig,
4584                             bool wait_ordered)
4585 {
4586         struct btrfs_block_rsv *block_rsv;
4587         struct btrfs_space_info *space_info;
4588         struct btrfs_trans_handle *trans;
4589         u64 delalloc_bytes;
4590         u64 max_reclaim;
4591         long time_left;
4592         unsigned long nr_pages;
4593         int loops;
4594         int items;
4595         enum btrfs_reserve_flush_enum flush;
4596
4597         /* Calc the number of the pages we need flush for space reservation */
4598         items = calc_reclaim_items_nr(root, to_reclaim);
4599         to_reclaim = items * EXTENT_SIZE_PER_ITEM;
4600
4601         trans = (struct btrfs_trans_handle *)current->journal_info;
4602         block_rsv = &root->fs_info->delalloc_block_rsv;
4603         space_info = block_rsv->space_info;
4604
4605         delalloc_bytes = percpu_counter_sum_positive(
4606                                                 &root->fs_info->delalloc_bytes);
4607         if (delalloc_bytes == 0) {
4608                 if (trans)
4609                         return;
4610                 if (wait_ordered)
4611                         btrfs_wait_ordered_roots(root->fs_info, items);
4612                 return;
4613         }
4614
4615         loops = 0;
4616         while (delalloc_bytes && loops < 3) {
4617                 max_reclaim = min(delalloc_bytes, to_reclaim);
4618                 nr_pages = max_reclaim >> PAGE_CACHE_SHIFT;
4619                 btrfs_writeback_inodes_sb_nr(root, nr_pages, items);
4620                 /*
4621                  * We need to wait for the async pages to actually start before
4622                  * we do anything.
4623                  */
4624                 max_reclaim = atomic_read(&root->fs_info->async_delalloc_pages);
4625                 if (!max_reclaim)
4626                         goto skip_async;
4627
4628                 if (max_reclaim <= nr_pages)
4629                         max_reclaim = 0;
4630                 else
4631                         max_reclaim -= nr_pages;
4632
4633                 wait_event(root->fs_info->async_submit_wait,
4634                            atomic_read(&root->fs_info->async_delalloc_pages) <=
4635                            (int)max_reclaim);
4636 skip_async:
4637                 if (!trans)
4638                         flush = BTRFS_RESERVE_FLUSH_ALL;
4639                 else
4640                         flush = BTRFS_RESERVE_NO_FLUSH;
4641                 spin_lock(&space_info->lock);
4642                 if (can_overcommit(root, space_info, orig, flush)) {
4643                         spin_unlock(&space_info->lock);
4644                         break;
4645                 }
4646                 spin_unlock(&space_info->lock);
4647
4648                 loops++;
4649                 if (wait_ordered && !trans) {
4650                         btrfs_wait_ordered_roots(root->fs_info, items);
4651                 } else {
4652                         time_left = schedule_timeout_killable(1);
4653                         if (time_left)
4654                                 break;
4655                 }
4656                 delalloc_bytes = percpu_counter_sum_positive(
4657                                                 &root->fs_info->delalloc_bytes);
4658         }
4659 }
4660
4661 /**
4662  * maybe_commit_transaction - possibly commit the transaction if its ok to
4663  * @root - the root we're allocating for
4664  * @bytes - the number of bytes we want to reserve
4665  * @force - force the commit
4666  *
4667  * This will check to make sure that committing the transaction will actually
4668  * get us somewhere and then commit the transaction if it does.  Otherwise it
4669  * will return -ENOSPC.
4670  */
4671 static int may_commit_transaction(struct btrfs_root *root,
4672                                   struct btrfs_space_info *space_info,
4673                                   u64 bytes, int force)
4674 {
4675         struct btrfs_block_rsv *delayed_rsv = &root->fs_info->delayed_block_rsv;
4676         struct btrfs_trans_handle *trans;
4677
4678         trans = (struct btrfs_trans_handle *)current->journal_info;
4679         if (trans)
4680                 return -EAGAIN;
4681
4682         if (force)
4683                 goto commit;
4684
4685         /* See if there is enough pinned space to make this reservation */
4686         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4687                                    bytes) >= 0)
4688                 goto commit;
4689
4690         /*
4691          * See if there is some space in the delayed insertion reservation for
4692          * this reservation.
4693          */
4694         if (space_info != delayed_rsv->space_info)
4695                 return -ENOSPC;
4696
4697         spin_lock(&delayed_rsv->lock);
4698         if (percpu_counter_compare(&space_info->total_bytes_pinned,
4699                                    bytes - delayed_rsv->size) >= 0) {
4700                 spin_unlock(&delayed_rsv->lock);
4701                 return -ENOSPC;
4702         }
4703         spin_unlock(&delayed_rsv->lock);
4704
4705 commit:
4706         trans = btrfs_join_transaction(root);
4707         if (IS_ERR(trans))
4708                 return -ENOSPC;
4709
4710         return btrfs_commit_transaction(trans, root);
4711 }
4712
4713 enum flush_state {
4714         FLUSH_DELAYED_ITEMS_NR  =       1,
4715         FLUSH_DELAYED_ITEMS     =       2,
4716         FLUSH_DELALLOC          =       3,
4717         FLUSH_DELALLOC_WAIT     =       4,
4718         ALLOC_CHUNK             =       5,
4719         COMMIT_TRANS            =       6,
4720 };
4721
4722 static int flush_space(struct btrfs_root *root,
4723                        struct btrfs_space_info *space_info, u64 num_bytes,
4724                        u64 orig_bytes, int state)
4725 {
4726         struct btrfs_trans_handle *trans;
4727         int nr;
4728         int ret = 0;
4729
4730         switch (state) {
4731         case FLUSH_DELAYED_ITEMS_NR:
4732         case FLUSH_DELAYED_ITEMS:
4733                 if (state == FLUSH_DELAYED_ITEMS_NR)
4734                         nr = calc_reclaim_items_nr(root, num_bytes) * 2;
4735                 else
4736                         nr = -1;
4737
4738                 trans = btrfs_join_transaction(root);
4739                 if (IS_ERR(trans)) {
4740                         ret = PTR_ERR(trans);
4741                         break;
4742                 }
4743                 ret = btrfs_run_delayed_items_nr(trans, root, nr);
4744                 btrfs_end_transaction(trans, root);
4745                 break;
4746         case FLUSH_DELALLOC:
4747         case FLUSH_DELALLOC_WAIT:
4748                 shrink_delalloc(root, num_bytes * 2, orig_bytes,
4749                                 state == FLUSH_DELALLOC_WAIT);
4750                 break;
4751         case ALLOC_CHUNK:
4752                 trans = btrfs_join_transaction(root);
4753                 if (IS_ERR(trans)) {
4754                         ret = PTR_ERR(trans);
4755                         break;
4756                 }
4757                 ret = do_chunk_alloc(trans, root->fs_info->extent_root,
4758                                      btrfs_get_alloc_profile(root, 0),
4759                                      CHUNK_ALLOC_NO_FORCE);
4760                 btrfs_end_transaction(trans, root);
4761                 if (ret == -ENOSPC)
4762                         ret = 0;
4763                 break;
4764         case COMMIT_TRANS:
4765                 ret = may_commit_transaction(root, space_info, orig_bytes, 0);
4766                 break;
4767         default:
4768                 ret = -ENOSPC;
4769                 break;
4770         }
4771
4772         return ret;
4773 }
4774
4775 static inline u64
4776 btrfs_calc_reclaim_metadata_size(struct btrfs_root *root,
4777                                  struct btrfs_space_info *space_info)
4778 {
4779         u64 used;
4780         u64 expected;
4781         u64 to_reclaim;
4782
4783         to_reclaim = min_t(u64, num_online_cpus() * 1024 * 1024,
4784                                 16 * 1024 * 1024);
4785         spin_lock(&space_info->lock);
4786         if (can_overcommit(root, space_info, to_reclaim,
4787                            BTRFS_RESERVE_FLUSH_ALL)) {
4788                 to_reclaim = 0;
4789                 goto out;
4790         }
4791
4792         used = space_info->bytes_used + space_info->bytes_reserved +
4793                space_info->bytes_pinned + space_info->bytes_readonly +
4794                space_info->bytes_may_use;
4795         if (can_overcommit(root, space_info, 1024 * 1024,
4796                            BTRFS_RESERVE_FLUSH_ALL))
4797                 expected = div_factor_fine(space_info->total_bytes, 95);
4798         else
4799                 expected = div_factor_fine(space_info->total_bytes, 90);
4800
4801         if (used > expected)
4802                 to_reclaim = used - expected;
4803         else
4804                 to_reclaim = 0;
4805         to_reclaim = min(to_reclaim, space_info->bytes_may_use +
4806                                      space_info->bytes_reserved);
4807 out:
4808         spin_unlock(&space_info->lock);
4809
4810         return to_reclaim;
4811 }
4812
4813 static inline int need_do_async_reclaim(struct btrfs_space_info *space_info,
4814                                         struct btrfs_fs_info *fs_info, u64 used)
4815 {
4816         u64 thresh = div_factor_fine(space_info->total_bytes, 98);
4817
4818         /* If we're just plain full then async reclaim just slows us down. */
4819         if (space_info->bytes_used >= thresh)
4820                 return 0;
4821
4822         return (used >= thresh && !btrfs_fs_closing(fs_info) &&
4823                 !test_bit(BTRFS_FS_STATE_REMOUNTING, &fs_info->fs_state));
4824 }
4825
4826 static int btrfs_need_do_async_reclaim(struct btrfs_space_info *space_info,
4827                                        struct btrfs_fs_info *fs_info,
4828                                        int flush_state)
4829 {
4830         u64 used;
4831
4832         spin_lock(&space_info->lock);
4833         /*
4834          * We run out of space and have not got any free space via flush_space,
4835          * so don't bother doing async reclaim.
4836          */
4837         if (flush_state > COMMIT_TRANS && space_info->full) {
4838                 spin_unlock(&space_info->lock);
4839                 return 0;
4840         }
4841
4842         used = space_info->bytes_used + space_info->bytes_reserved +
4843                space_info->bytes_pinned + space_info->bytes_readonly +
4844                space_info->bytes_may_use;
4845         if (need_do_async_reclaim(space_info, fs_info, used)) {
4846                 spin_unlock(&space_info->lock);
4847                 return 1;
4848         }
4849         spin_unlock(&space_info->lock);
4850
4851         return 0;
4852 }
4853
4854 static void btrfs_async_reclaim_metadata_space(struct work_struct *work)
4855 {
4856         struct btrfs_fs_info *fs_info;
4857         struct btrfs_space_info *space_info;
4858         u64 to_reclaim;
4859         int flush_state;
4860
4861         fs_info = container_of(work, struct btrfs_fs_info, async_reclaim_work);
4862         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
4863
4864         to_reclaim = btrfs_calc_reclaim_metadata_size(fs_info->fs_root,
4865                                                       space_info);
4866         if (!to_reclaim)
4867                 return;
4868
4869         flush_state = FLUSH_DELAYED_ITEMS_NR;
4870         do {
4871                 flush_space(fs_info->fs_root, space_info, to_reclaim,
4872                             to_reclaim, flush_state);
4873                 flush_state++;
4874                 if (!btrfs_need_do_async_reclaim(space_info, fs_info,
4875                                                  flush_state))
4876                         return;
4877         } while (flush_state < COMMIT_TRANS);
4878 }
4879
4880 void btrfs_init_async_reclaim_work(struct work_struct *work)
4881 {
4882         INIT_WORK(work, btrfs_async_reclaim_metadata_space);
4883 }
4884
4885 /**
4886  * reserve_metadata_bytes - try to reserve bytes from the block_rsv's space
4887  * @root - the root we're allocating for
4888  * @block_rsv - the block_rsv we're allocating for
4889  * @orig_bytes - the number of bytes we want
4890  * @flush - whether or not we can flush to make our reservation
4891  *
4892  * This will reserve orgi_bytes number of bytes from the space info associated
4893  * with the block_rsv.  If there is not enough space it will make an attempt to
4894  * flush out space to make room.  It will do this by flushing delalloc if
4895  * possible or committing the transaction.  If flush is 0 then no attempts to
4896  * regain reservations will be made and this will fail if there is not enough
4897  * space already.
4898  */
4899 static int reserve_metadata_bytes(struct btrfs_root *root,
4900                                   struct btrfs_block_rsv *block_rsv,
4901                                   u64 orig_bytes,
4902                                   enum btrfs_reserve_flush_enum flush)
4903 {
4904         struct btrfs_space_info *space_info = block_rsv->space_info;
4905         u64 used;
4906         u64 num_bytes = orig_bytes;
4907         int flush_state = FLUSH_DELAYED_ITEMS_NR;
4908         int ret = 0;
4909         bool flushing = false;
4910
4911 again:
4912         ret = 0;
4913         spin_lock(&space_info->lock);
4914         /*
4915          * We only want to wait if somebody other than us is flushing and we
4916          * are actually allowed to flush all things.
4917          */
4918         while (flush == BTRFS_RESERVE_FLUSH_ALL && !flushing &&
4919                space_info->flush) {
4920                 spin_unlock(&space_info->lock);
4921                 /*
4922                  * If we have a trans handle we can't wait because the flusher
4923                  * may have to commit the transaction, which would mean we would
4924                  * deadlock since we are waiting for the flusher to finish, but
4925                  * hold the current transaction open.
4926                  */
4927                 if (current->journal_info)
4928                         return -EAGAIN;
4929                 ret = wait_event_killable(space_info->wait, !space_info->flush);
4930                 /* Must have been killed, return */
4931                 if (ret)
4932                         return -EINTR;
4933
4934                 spin_lock(&space_info->lock);
4935         }
4936
4937         ret = -ENOSPC;
4938         used = space_info->bytes_used + space_info->bytes_reserved +
4939                 space_info->bytes_pinned + space_info->bytes_readonly +
4940                 space_info->bytes_may_use;
4941
4942         /*
4943          * The idea here is that we've not already over-reserved the block group
4944          * then we can go ahead and save our reservation first and then start
4945          * flushing if we need to.  Otherwise if we've already overcommitted
4946          * lets start flushing stuff first and then come back and try to make
4947          * our reservation.
4948          */
4949         if (used <= space_info->total_bytes) {
4950                 if (used + orig_bytes <= space_info->total_bytes) {
4951                         space_info->bytes_may_use += orig_bytes;
4952                         trace_btrfs_space_reservation(root->fs_info,
4953                                 "space_info", space_info->flags, orig_bytes, 1);
4954                         ret = 0;
4955                 } else {
4956                         /*
4957                          * Ok set num_bytes to orig_bytes since we aren't
4958                          * overocmmitted, this way we only try and reclaim what
4959                          * we need.
4960                          */
4961                         num_bytes = orig_bytes;
4962                 }
4963         } else {
4964                 /*
4965                  * Ok we're over committed, set num_bytes to the overcommitted
4966                  * amount plus the amount of bytes that we need for this
4967                  * reservation.
4968                  */
4969                 num_bytes = used - space_info->total_bytes +
4970                         (orig_bytes * 2);
4971         }
4972
4973         if (ret && can_overcommit(root, space_info, orig_bytes, flush)) {
4974                 space_info->bytes_may_use += orig_bytes;
4975                 trace_btrfs_space_reservation(root->fs_info, "space_info",
4976                                               space_info->flags, orig_bytes,
4977                                               1);
4978                 ret = 0;
4979         }
4980
4981         /*
4982          * Couldn't make our reservation, save our place so while we're trying
4983          * to reclaim space we can actually use it instead of somebody else
4984          * stealing it from us.
4985          *
4986          * We make the other tasks wait for the flush only when we can flush
4987          * all things.
4988          */
4989         if (ret && flush != BTRFS_RESERVE_NO_FLUSH) {
4990                 flushing = true;
4991                 space_info->flush = 1;
4992         } else if (!ret && space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
4993                 used += orig_bytes;
4994                 /*
4995                  * We will do the space reservation dance during log replay,
4996                  * which means we won't have fs_info->fs_root set, so don't do
4997                  * the async reclaim as we will panic.
4998                  */
4999                 if (!root->fs_info->log_root_recovering &&
5000                     need_do_async_reclaim(space_info, root->fs_info, used) &&
5001                     !work_busy(&root->fs_info->async_reclaim_work))
5002                         queue_work(system_unbound_wq,
5003                                    &root->fs_info->async_reclaim_work);
5004         }
5005         spin_unlock(&space_info->lock);
5006
5007         if (!ret || flush == BTRFS_RESERVE_NO_FLUSH)
5008                 goto out;
5009
5010         ret = flush_space(root, space_info, num_bytes, orig_bytes,
5011                           flush_state);
5012         flush_state++;
5013
5014         /*
5015          * If we are FLUSH_LIMIT, we can not flush delalloc, or the deadlock
5016          * would happen. So skip delalloc flush.
5017          */
5018         if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
5019             (flush_state == FLUSH_DELALLOC ||
5020              flush_state == FLUSH_DELALLOC_WAIT))
5021                 flush_state = ALLOC_CHUNK;
5022
5023         if (!ret)
5024                 goto again;
5025         else if (flush == BTRFS_RESERVE_FLUSH_LIMIT &&
5026                  flush_state < COMMIT_TRANS)
5027                 goto again;
5028         else if (flush == BTRFS_RESERVE_FLUSH_ALL &&
5029                  flush_state <= COMMIT_TRANS)
5030                 goto again;
5031
5032 out:
5033         if (ret == -ENOSPC &&
5034             unlikely(root->orphan_cleanup_state == ORPHAN_CLEANUP_STARTED)) {
5035                 struct btrfs_block_rsv *global_rsv =
5036                         &root->fs_info->global_block_rsv;
5037
5038                 if (block_rsv != global_rsv &&
5039                     !block_rsv_use_bytes(global_rsv, orig_bytes))
5040                         ret = 0;
5041         }
5042         if (ret == -ENOSPC)
5043                 trace_btrfs_space_reservation(root->fs_info,
5044                                               "space_info:enospc",
5045                                               space_info->flags, orig_bytes, 1);
5046         if (flushing) {
5047                 spin_lock(&space_info->lock);
5048                 space_info->flush = 0;
5049                 wake_up_all(&space_info->wait);
5050                 spin_unlock(&space_info->lock);
5051         }
5052         return ret;
5053 }
5054
5055 static struct btrfs_block_rsv *get_block_rsv(
5056                                         const struct btrfs_trans_handle *trans,
5057                                         const struct btrfs_root *root)
5058 {
5059         struct btrfs_block_rsv *block_rsv = NULL;
5060
5061         if (test_bit(BTRFS_ROOT_REF_COWS, &root->state) ||
5062             (root == root->fs_info->csum_root && trans->adding_csums) ||
5063              (root == root->fs_info->uuid_root))
5064                 block_rsv = trans->block_rsv;
5065
5066         if (!block_rsv)
5067                 block_rsv = root->block_rsv;
5068
5069         if (!block_rsv)
5070                 block_rsv = &root->fs_info->empty_block_rsv;
5071
5072         return block_rsv;
5073 }
5074
5075 static int block_rsv_use_bytes(struct btrfs_block_rsv *block_rsv,
5076                                u64 num_bytes)
5077 {
5078         int ret = -ENOSPC;
5079         spin_lock(&block_rsv->lock);
5080         if (block_rsv->reserved >= num_bytes) {
5081                 block_rsv->reserved -= num_bytes;
5082                 if (block_rsv->reserved < block_rsv->size)
5083                         block_rsv->full = 0;
5084                 ret = 0;
5085         }
5086         spin_unlock(&block_rsv->lock);
5087         return ret;
5088 }
5089
5090 static void block_rsv_add_bytes(struct btrfs_block_rsv *block_rsv,
5091                                 u64 num_bytes, int update_size)
5092 {
5093         spin_lock(&block_rsv->lock);
5094         block_rsv->reserved += num_bytes;
5095         if (update_size)
5096                 block_rsv->size += num_bytes;
5097         else if (block_rsv->reserved >= block_rsv->size)
5098                 block_rsv->full = 1;
5099         spin_unlock(&block_rsv->lock);
5100 }
5101
5102 int btrfs_cond_migrate_bytes(struct btrfs_fs_info *fs_info,
5103                              struct btrfs_block_rsv *dest, u64 num_bytes,
5104                              int min_factor)
5105 {
5106         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
5107         u64 min_bytes;
5108
5109         if (global_rsv->space_info != dest->space_info)
5110                 return -ENOSPC;
5111
5112         spin_lock(&global_rsv->lock);
5113         min_bytes = div_factor(global_rsv->size, min_factor);
5114         if (global_rsv->reserved < min_bytes + num_bytes) {
5115                 spin_unlock(&global_rsv->lock);
5116                 return -ENOSPC;
5117         }
5118         global_rsv->reserved -= num_bytes;
5119         if (global_rsv->reserved < global_rsv->size)
5120                 global_rsv->full = 0;
5121         spin_unlock(&global_rsv->lock);
5122
5123         block_rsv_add_bytes(dest, num_bytes, 1);
5124         return 0;
5125 }
5126
5127 static void block_rsv_release_bytes(struct btrfs_fs_info *fs_info,
5128                                     struct btrfs_block_rsv *block_rsv,
5129                                     struct btrfs_block_rsv *dest, u64 num_bytes)
5130 {
5131         struct btrfs_space_info *space_info = block_rsv->space_info;
5132
5133         spin_lock(&block_rsv->lock);
5134         if (num_bytes == (u64)-1)
5135                 num_bytes = block_rsv->size;
5136         block_rsv->size -= num_bytes;
5137         if (block_rsv->reserved >= block_rsv->size) {
5138                 num_bytes = block_rsv->reserved - block_rsv->size;
5139                 block_rsv->reserved = block_rsv->size;
5140                 block_rsv->full = 1;
5141         } else {
5142                 num_bytes = 0;
5143         }
5144         spin_unlock(&block_rsv->lock);
5145
5146         if (num_bytes > 0) {
5147                 if (dest) {
5148                         spin_lock(&dest->lock);
5149                         if (!dest->full) {
5150                                 u64 bytes_to_add;
5151
5152                                 bytes_to_add = dest->size - dest->reserved;
5153                                 bytes_to_add = min(num_bytes, bytes_to_add);
5154                                 dest->reserved += bytes_to_add;
5155                                 if (dest->reserved >= dest->size)
5156                                         dest->full = 1;
5157                                 num_bytes -= bytes_to_add;
5158                         }
5159                         spin_unlock(&dest->lock);
5160                 }
5161                 if (num_bytes) {
5162                         spin_lock(&space_info->lock);
5163                         space_info->bytes_may_use -= num_bytes;
5164                         trace_btrfs_space_reservation(fs_info, "space_info",
5165                                         space_info->flags, num_bytes, 0);
5166                         spin_unlock(&space_info->lock);
5167                 }
5168         }
5169 }
5170
5171 static int block_rsv_migrate_bytes(struct btrfs_block_rsv *src,
5172                                    struct btrfs_block_rsv *dst, u64 num_bytes)
5173 {
5174         int ret;
5175
5176         ret = block_rsv_use_bytes(src, num_bytes);
5177         if (ret)
5178                 return ret;
5179
5180         block_rsv_add_bytes(dst, num_bytes, 1);
5181         return 0;
5182 }
5183
5184 void btrfs_init_block_rsv(struct btrfs_block_rsv *rsv, unsigned short type)
5185 {
5186         memset(rsv, 0, sizeof(*rsv));
5187         spin_lock_init(&rsv->lock);
5188         rsv->type = type;
5189 }
5190
5191 struct btrfs_block_rsv *btrfs_alloc_block_rsv(struct btrfs_root *root,
5192                                               unsigned short type)
5193 {
5194         struct btrfs_block_rsv *block_rsv;
5195         struct btrfs_fs_info *fs_info = root->fs_info;
5196
5197         block_rsv = kmalloc(sizeof(*block_rsv), GFP_NOFS);
5198         if (!block_rsv)
5199                 return NULL;
5200
5201         btrfs_init_block_rsv(block_rsv, type);
5202         block_rsv->space_info = __find_space_info(fs_info,
5203                                                   BTRFS_BLOCK_GROUP_METADATA);
5204         return block_rsv;
5205 }
5206
5207 void btrfs_free_block_rsv(struct btrfs_root *root,
5208                           struct btrfs_block_rsv *rsv)
5209 {
5210         if (!rsv)
5211                 return;
5212         btrfs_block_rsv_release(root, rsv, (u64)-1);
5213         kfree(rsv);
5214 }
5215
5216 void __btrfs_free_block_rsv(struct btrfs_block_rsv *rsv)
5217 {
5218         kfree(rsv);
5219 }
5220
5221 int btrfs_block_rsv_add(struct btrfs_root *root,
5222                         struct btrfs_block_rsv *block_rsv, u64 num_bytes,
5223                         enum btrfs_reserve_flush_enum flush)
5224 {
5225         int ret;
5226
5227         if (num_bytes == 0)
5228                 return 0;
5229
5230         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5231         if (!ret) {
5232                 block_rsv_add_bytes(block_rsv, num_bytes, 1);
5233                 return 0;
5234         }
5235
5236         return ret;
5237 }
5238
5239 int btrfs_block_rsv_check(struct btrfs_root *root,
5240                           struct btrfs_block_rsv *block_rsv, int min_factor)
5241 {
5242         u64 num_bytes = 0;
5243         int ret = -ENOSPC;
5244
5245         if (!block_rsv)
5246                 return 0;
5247
5248         spin_lock(&block_rsv->lock);
5249         num_bytes = div_factor(block_rsv->size, min_factor);
5250         if (block_rsv->reserved >= num_bytes)
5251                 ret = 0;
5252         spin_unlock(&block_rsv->lock);
5253
5254         return ret;
5255 }
5256
5257 int btrfs_block_rsv_refill(struct btrfs_root *root,
5258                            struct btrfs_block_rsv *block_rsv, u64 min_reserved,
5259                            enum btrfs_reserve_flush_enum flush)
5260 {
5261         u64 num_bytes = 0;
5262         int ret = -ENOSPC;
5263
5264         if (!block_rsv)
5265                 return 0;
5266
5267         spin_lock(&block_rsv->lock);
5268         num_bytes = min_reserved;
5269         if (block_rsv->reserved >= num_bytes)
5270                 ret = 0;
5271         else
5272                 num_bytes -= block_rsv->reserved;
5273         spin_unlock(&block_rsv->lock);
5274
5275         if (!ret)
5276                 return 0;
5277
5278         ret = reserve_metadata_bytes(root, block_rsv, num_bytes, flush);
5279         if (!ret) {
5280                 block_rsv_add_bytes(block_rsv, num_bytes, 0);
5281                 return 0;
5282         }
5283
5284         return ret;
5285 }
5286
5287 int btrfs_block_rsv_migrate(struct btrfs_block_rsv *src_rsv,
5288                             struct btrfs_block_rsv *dst_rsv,
5289                             u64 num_bytes)
5290 {
5291         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5292 }
5293
5294 void btrfs_block_rsv_release(struct btrfs_root *root,
5295                              struct btrfs_block_rsv *block_rsv,
5296                              u64 num_bytes)
5297 {
5298         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5299         if (global_rsv == block_rsv ||
5300             block_rsv->space_info != global_rsv->space_info)
5301                 global_rsv = NULL;
5302         block_rsv_release_bytes(root->fs_info, block_rsv, global_rsv,
5303                                 num_bytes);
5304 }
5305
5306 /*
5307  * helper to calculate size of global block reservation.
5308  * the desired value is sum of space used by extent tree,
5309  * checksum tree and root tree
5310  */
5311 static u64 calc_global_metadata_size(struct btrfs_fs_info *fs_info)
5312 {
5313         struct btrfs_space_info *sinfo;
5314         u64 num_bytes;
5315         u64 meta_used;
5316         u64 data_used;
5317         int csum_size = btrfs_super_csum_size(fs_info->super_copy);
5318
5319         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_DATA);
5320         spin_lock(&sinfo->lock);
5321         data_used = sinfo->bytes_used;
5322         spin_unlock(&sinfo->lock);
5323
5324         sinfo = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5325         spin_lock(&sinfo->lock);
5326         if (sinfo->flags & BTRFS_BLOCK_GROUP_DATA)
5327                 data_used = 0;
5328         meta_used = sinfo->bytes_used;
5329         spin_unlock(&sinfo->lock);
5330
5331         num_bytes = (data_used >> fs_info->sb->s_blocksize_bits) *
5332                     csum_size * 2;
5333         num_bytes += div_u64(data_used + meta_used, 50);
5334
5335         if (num_bytes * 3 > meta_used)
5336                 num_bytes = div_u64(meta_used, 3);
5337
5338         return ALIGN(num_bytes, fs_info->extent_root->nodesize << 10);
5339 }
5340
5341 static void update_global_block_rsv(struct btrfs_fs_info *fs_info)
5342 {
5343         struct btrfs_block_rsv *block_rsv = &fs_info->global_block_rsv;
5344         struct btrfs_space_info *sinfo = block_rsv->space_info;
5345         u64 num_bytes;
5346
5347         num_bytes = calc_global_metadata_size(fs_info);
5348
5349         spin_lock(&sinfo->lock);
5350         spin_lock(&block_rsv->lock);
5351
5352         block_rsv->size = min_t(u64, num_bytes, 512 * 1024 * 1024);
5353
5354         num_bytes = sinfo->bytes_used + sinfo->bytes_pinned +
5355                     sinfo->bytes_reserved + sinfo->bytes_readonly +
5356                     sinfo->bytes_may_use;
5357
5358         if (sinfo->total_bytes > num_bytes) {
5359                 num_bytes = sinfo->total_bytes - num_bytes;
5360                 block_rsv->reserved += num_bytes;
5361                 sinfo->bytes_may_use += num_bytes;
5362                 trace_btrfs_space_reservation(fs_info, "space_info",
5363                                       sinfo->flags, num_bytes, 1);
5364         }
5365
5366         if (block_rsv->reserved >= block_rsv->size) {
5367                 num_bytes = block_rsv->reserved - block_rsv->size;
5368                 sinfo->bytes_may_use -= num_bytes;
5369                 trace_btrfs_space_reservation(fs_info, "space_info",
5370                                       sinfo->flags, num_bytes, 0);
5371                 block_rsv->reserved = block_rsv->size;
5372                 block_rsv->full = 1;
5373         }
5374
5375         spin_unlock(&block_rsv->lock);
5376         spin_unlock(&sinfo->lock);
5377 }
5378
5379 static void init_global_block_rsv(struct btrfs_fs_info *fs_info)
5380 {
5381         struct btrfs_space_info *space_info;
5382
5383         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_SYSTEM);
5384         fs_info->chunk_block_rsv.space_info = space_info;
5385
5386         space_info = __find_space_info(fs_info, BTRFS_BLOCK_GROUP_METADATA);
5387         fs_info->global_block_rsv.space_info = space_info;
5388         fs_info->delalloc_block_rsv.space_info = space_info;
5389         fs_info->trans_block_rsv.space_info = space_info;
5390         fs_info->empty_block_rsv.space_info = space_info;
5391         fs_info->delayed_block_rsv.space_info = space_info;
5392
5393         fs_info->extent_root->block_rsv = &fs_info->global_block_rsv;
5394         fs_info->csum_root->block_rsv = &fs_info->global_block_rsv;
5395         fs_info->dev_root->block_rsv = &fs_info->global_block_rsv;
5396         fs_info->tree_root->block_rsv = &fs_info->global_block_rsv;
5397         if (fs_info->quota_root)
5398                 fs_info->quota_root->block_rsv = &fs_info->global_block_rsv;
5399         fs_info->chunk_root->block_rsv = &fs_info->chunk_block_rsv;
5400
5401         update_global_block_rsv(fs_info);
5402 }
5403
5404 static void release_global_block_rsv(struct btrfs_fs_info *fs_info)
5405 {
5406         block_rsv_release_bytes(fs_info, &fs_info->global_block_rsv, NULL,
5407                                 (u64)-1);
5408         WARN_ON(fs_info->delalloc_block_rsv.size > 0);
5409         WARN_ON(fs_info->delalloc_block_rsv.reserved > 0);
5410         WARN_ON(fs_info->trans_block_rsv.size > 0);
5411         WARN_ON(fs_info->trans_block_rsv.reserved > 0);
5412         WARN_ON(fs_info->chunk_block_rsv.size > 0);
5413         WARN_ON(fs_info->chunk_block_rsv.reserved > 0);
5414         WARN_ON(fs_info->delayed_block_rsv.size > 0);
5415         WARN_ON(fs_info->delayed_block_rsv.reserved > 0);
5416 }
5417
5418 void btrfs_trans_release_metadata(struct btrfs_trans_handle *trans,
5419                                   struct btrfs_root *root)
5420 {
5421         if (!trans->block_rsv)
5422                 return;
5423
5424         if (!trans->bytes_reserved)
5425                 return;
5426
5427         trace_btrfs_space_reservation(root->fs_info, "transaction",
5428                                       trans->transid, trans->bytes_reserved, 0);
5429         btrfs_block_rsv_release(root, trans->block_rsv, trans->bytes_reserved);
5430         trans->bytes_reserved = 0;
5431 }
5432
5433 /*
5434  * To be called after all the new block groups attached to the transaction
5435  * handle have been created (btrfs_create_pending_block_groups()).
5436  */
5437 void btrfs_trans_release_chunk_metadata(struct btrfs_trans_handle *trans)
5438 {
5439         struct btrfs_fs_info *fs_info = trans->root->fs_info;
5440
5441         if (!trans->chunk_bytes_reserved)
5442                 return;
5443
5444         WARN_ON_ONCE(!list_empty(&trans->new_bgs));
5445
5446         block_rsv_release_bytes(fs_info, &fs_info->chunk_block_rsv, NULL,
5447                                 trans->chunk_bytes_reserved);
5448         trans->chunk_bytes_reserved = 0;
5449 }
5450
5451 /* Can only return 0 or -ENOSPC */
5452 int btrfs_orphan_reserve_metadata(struct btrfs_trans_handle *trans,
5453                                   struct inode *inode)
5454 {
5455         struct btrfs_root *root = BTRFS_I(inode)->root;
5456         struct btrfs_block_rsv *src_rsv = get_block_rsv(trans, root);
5457         struct btrfs_block_rsv *dst_rsv = root->orphan_block_rsv;
5458
5459         /*
5460          * We need to hold space in order to delete our orphan item once we've
5461          * added it, so this takes the reservation so we can release it later
5462          * when we are truly done with the orphan item.
5463          */
5464         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5465         trace_btrfs_space_reservation(root->fs_info, "orphan",
5466                                       btrfs_ino(inode), num_bytes, 1);
5467         return block_rsv_migrate_bytes(src_rsv, dst_rsv, num_bytes);
5468 }
5469
5470 void btrfs_orphan_release_metadata(struct inode *inode)
5471 {
5472         struct btrfs_root *root = BTRFS_I(inode)->root;
5473         u64 num_bytes = btrfs_calc_trans_metadata_size(root, 1);
5474         trace_btrfs_space_reservation(root->fs_info, "orphan",
5475                                       btrfs_ino(inode), num_bytes, 0);
5476         btrfs_block_rsv_release(root, root->orphan_block_rsv, num_bytes);
5477 }
5478
5479 /*
5480  * btrfs_subvolume_reserve_metadata() - reserve space for subvolume operation
5481  * root: the root of the parent directory
5482  * rsv: block reservation
5483  * items: the number of items that we need do reservation
5484  * qgroup_reserved: used to return the reserved size in qgroup
5485  *
5486  * This function is used to reserve the space for snapshot/subvolume
5487  * creation and deletion. Those operations are different with the
5488  * common file/directory operations, they change two fs/file trees
5489  * and root tree, the number of items that the qgroup reserves is
5490  * different with the free space reservation. So we can not use
5491  * the space reseravtion mechanism in start_transaction().
5492  */
5493 int btrfs_subvolume_reserve_metadata(struct btrfs_root *root,
5494                                      struct btrfs_block_rsv *rsv,
5495                                      int items,
5496                                      u64 *qgroup_reserved,
5497                                      bool use_global_rsv)
5498 {
5499         u64 num_bytes;
5500         int ret;
5501         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
5502
5503         if (root->fs_info->quota_enabled) {
5504                 /* One for parent inode, two for dir entries */
5505                 num_bytes = 3 * root->nodesize;
5506                 ret = btrfs_qgroup_reserve_meta(root, num_bytes);
5507                 if (ret)
5508                         return ret;
5509         } else {
5510                 num_bytes = 0;
5511         }
5512
5513         *qgroup_reserved = num_bytes;
5514
5515         num_bytes = btrfs_calc_trans_metadata_size(root, items);
5516         rsv->space_info = __find_space_info(root->fs_info,
5517                                             BTRFS_BLOCK_GROUP_METADATA);
5518         ret = btrfs_block_rsv_add(root, rsv, num_bytes,
5519                                   BTRFS_RESERVE_FLUSH_ALL);
5520
5521         if (ret == -ENOSPC && use_global_rsv)
5522                 ret = btrfs_block_rsv_migrate(global_rsv, rsv, num_bytes);
5523
5524         if (ret && *qgroup_reserved)
5525                 btrfs_qgroup_free_meta(root, *qgroup_reserved);
5526
5527         return ret;
5528 }
5529
5530 void btrfs_subvolume_release_metadata(struct btrfs_root *root,
5531                                       struct btrfs_block_rsv *rsv,
5532                                       u64 qgroup_reserved)
5533 {
5534         btrfs_block_rsv_release(root, rsv, (u64)-1);
5535 }
5536
5537 /**
5538  * drop_outstanding_extent - drop an outstanding extent
5539  * @inode: the inode we're dropping the extent for
5540  * @num_bytes: the number of bytes we're relaseing.
5541  *
5542  * This is called when we are freeing up an outstanding extent, either called
5543  * after an error or after an extent is written.  This will return the number of
5544  * reserved extents that need to be freed.  This must be called with
5545  * BTRFS_I(inode)->lock held.
5546  */
5547 static unsigned drop_outstanding_extent(struct inode *inode, u64 num_bytes)
5548 {
5549         unsigned drop_inode_space = 0;
5550         unsigned dropped_extents = 0;
5551         unsigned num_extents = 0;
5552
5553         num_extents = (unsigned)div64_u64(num_bytes +
5554                                           BTRFS_MAX_EXTENT_SIZE - 1,
5555                                           BTRFS_MAX_EXTENT_SIZE);
5556         ASSERT(num_extents);
5557         ASSERT(BTRFS_I(inode)->outstanding_extents >= num_extents);
5558         BTRFS_I(inode)->outstanding_extents -= num_extents;
5559
5560         if (BTRFS_I(inode)->outstanding_extents == 0 &&
5561             test_and_clear_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5562                                &BTRFS_I(inode)->runtime_flags))
5563                 drop_inode_space = 1;
5564
5565         /*
5566          * If we have more or the same amount of outsanding extents than we have
5567          * reserved then we need to leave the reserved extents count alone.
5568          */
5569         if (BTRFS_I(inode)->outstanding_extents >=
5570             BTRFS_I(inode)->reserved_extents)
5571                 return drop_inode_space;
5572
5573         dropped_extents = BTRFS_I(inode)->reserved_extents -
5574                 BTRFS_I(inode)->outstanding_extents;
5575         BTRFS_I(inode)->reserved_extents -= dropped_extents;
5576         return dropped_extents + drop_inode_space;
5577 }
5578
5579 /**
5580  * calc_csum_metadata_size - return the amount of metada space that must be
5581  *      reserved/free'd for the given bytes.
5582  * @inode: the inode we're manipulating
5583  * @num_bytes: the number of bytes in question
5584  * @reserve: 1 if we are reserving space, 0 if we are freeing space
5585  *
5586  * This adjusts the number of csum_bytes in the inode and then returns the
5587  * correct amount of metadata that must either be reserved or freed.  We
5588  * calculate how many checksums we can fit into one leaf and then divide the
5589  * number of bytes that will need to be checksumed by this value to figure out
5590  * how many checksums will be required.  If we are adding bytes then the number
5591  * may go up and we will return the number of additional bytes that must be
5592  * reserved.  If it is going down we will return the number of bytes that must
5593  * be freed.
5594  *
5595  * This must be called with BTRFS_I(inode)->lock held.
5596  */
5597 static u64 calc_csum_metadata_size(struct inode *inode, u64 num_bytes,
5598                                    int reserve)
5599 {
5600         struct btrfs_root *root = BTRFS_I(inode)->root;
5601         u64 old_csums, num_csums;
5602
5603         if (BTRFS_I(inode)->flags & BTRFS_INODE_NODATASUM &&
5604             BTRFS_I(inode)->csum_bytes == 0)
5605                 return 0;
5606
5607         old_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5608         if (reserve)
5609                 BTRFS_I(inode)->csum_bytes += num_bytes;
5610         else
5611                 BTRFS_I(inode)->csum_bytes -= num_bytes;
5612         num_csums = btrfs_csum_bytes_to_leaves(root, BTRFS_I(inode)->csum_bytes);
5613
5614         /* No change, no need to reserve more */
5615         if (old_csums == num_csums)
5616                 return 0;
5617
5618         if (reserve)
5619                 return btrfs_calc_trans_metadata_size(root,
5620                                                       num_csums - old_csums);
5621
5622         return btrfs_calc_trans_metadata_size(root, old_csums - num_csums);
5623 }
5624
5625 int btrfs_delalloc_reserve_metadata(struct inode *inode, u64 num_bytes)
5626 {
5627         struct btrfs_root *root = BTRFS_I(inode)->root;
5628         struct btrfs_block_rsv *block_rsv = &root->fs_info->delalloc_block_rsv;
5629         u64 to_reserve = 0;
5630         u64 csum_bytes;
5631         unsigned nr_extents = 0;
5632         int extra_reserve = 0;
5633         enum btrfs_reserve_flush_enum flush = BTRFS_RESERVE_FLUSH_ALL;
5634         int ret = 0;
5635         bool delalloc_lock = true;
5636         u64 to_free = 0;
5637         unsigned dropped;
5638
5639         /* If we are a free space inode we need to not flush since we will be in
5640          * the middle of a transaction commit.  We also don't need the delalloc
5641          * mutex since we won't race with anybody.  We need this mostly to make
5642          * lockdep shut its filthy mouth.
5643          */
5644         if (btrfs_is_free_space_inode(inode)) {
5645                 flush = BTRFS_RESERVE_NO_FLUSH;
5646                 delalloc_lock = false;
5647         }
5648
5649         if (flush != BTRFS_RESERVE_NO_FLUSH &&
5650             btrfs_transaction_in_commit(root->fs_info))
5651                 schedule_timeout(1);
5652
5653         if (delalloc_lock)
5654                 mutex_lock(&BTRFS_I(inode)->delalloc_mutex);
5655
5656         num_bytes = ALIGN(num_bytes, root->sectorsize);
5657
5658         spin_lock(&BTRFS_I(inode)->lock);
5659         nr_extents = (unsigned)div64_u64(num_bytes +
5660                                          BTRFS_MAX_EXTENT_SIZE - 1,
5661                                          BTRFS_MAX_EXTENT_SIZE);
5662         BTRFS_I(inode)->outstanding_extents += nr_extents;
5663         nr_extents = 0;
5664
5665         if (BTRFS_I(inode)->outstanding_extents >
5666             BTRFS_I(inode)->reserved_extents)
5667                 nr_extents = BTRFS_I(inode)->outstanding_extents -
5668                         BTRFS_I(inode)->reserved_extents;
5669
5670         /*
5671          * Add an item to reserve for updating the inode when we complete the
5672          * delalloc io.
5673          */
5674         if (!test_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5675                       &BTRFS_I(inode)->runtime_flags)) {
5676                 nr_extents++;
5677                 extra_reserve = 1;
5678         }
5679
5680         to_reserve = btrfs_calc_trans_metadata_size(root, nr_extents);
5681         to_reserve += calc_csum_metadata_size(inode, num_bytes, 1);
5682         csum_bytes = BTRFS_I(inode)->csum_bytes;
5683         spin_unlock(&BTRFS_I(inode)->lock);
5684
5685         if (root->fs_info->quota_enabled) {
5686                 ret = btrfs_qgroup_reserve_meta(root,
5687                                 nr_extents * root->nodesize);
5688                 if (ret)
5689                         goto out_fail;
5690         }
5691
5692         ret = reserve_metadata_bytes(root, block_rsv, to_reserve, flush);
5693         if (unlikely(ret)) {
5694                 btrfs_qgroup_free_meta(root, nr_extents * root->nodesize);
5695                 goto out_fail;
5696         }
5697
5698         spin_lock(&BTRFS_I(inode)->lock);
5699         if (extra_reserve) {
5700                 set_bit(BTRFS_INODE_DELALLOC_META_RESERVED,
5701                         &BTRFS_I(inode)->runtime_flags);
5702                 nr_extents--;
5703         }
5704         BTRFS_I(inode)->reserved_extents += nr_extents;
5705         spin_unlock(&BTRFS_I(inode)->lock);
5706
5707         if (delalloc_lock)
5708                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5709
5710         if (to_reserve)
5711                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5712                                               btrfs_ino(inode), to_reserve, 1);
5713         block_rsv_add_bytes(block_rsv, to_reserve, 1);
5714
5715         return 0;
5716
5717 out_fail:
5718         spin_lock(&BTRFS_I(inode)->lock);
5719         dropped = drop_outstanding_extent(inode, num_bytes);
5720         /*
5721          * If the inodes csum_bytes is the same as the original
5722          * csum_bytes then we know we haven't raced with any free()ers
5723          * so we can just reduce our inodes csum bytes and carry on.
5724          */
5725         if (BTRFS_I(inode)->csum_bytes == csum_bytes) {
5726                 calc_csum_metadata_size(inode, num_bytes, 0);
5727         } else {
5728                 u64 orig_csum_bytes = BTRFS_I(inode)->csum_bytes;
5729                 u64 bytes;
5730
5731                 /*
5732                  * This is tricky, but first we need to figure out how much we
5733                  * free'd from any free-ers that occured during this
5734                  * reservation, so we reset ->csum_bytes to the csum_bytes
5735                  * before we dropped our lock, and then call the free for the
5736                  * number of bytes that were freed while we were trying our
5737                  * reservation.
5738                  */
5739                 bytes = csum_bytes - BTRFS_I(inode)->csum_bytes;
5740                 BTRFS_I(inode)->csum_bytes = csum_bytes;
5741                 to_free = calc_csum_metadata_size(inode, bytes, 0);
5742
5743
5744                 /*
5745                  * Now we need to see how much we would have freed had we not
5746                  * been making this reservation and our ->csum_bytes were not
5747                  * artificially inflated.
5748                  */
5749                 BTRFS_I(inode)->csum_bytes = csum_bytes - num_bytes;
5750                 bytes = csum_bytes - orig_csum_bytes;
5751                 bytes = calc_csum_metadata_size(inode, bytes, 0);
5752
5753                 /*
5754                  * Now reset ->csum_bytes to what it should be.  If bytes is
5755                  * more than to_free then we would have free'd more space had we
5756                  * not had an artificially high ->csum_bytes, so we need to free
5757                  * the remainder.  If bytes is the same or less then we don't
5758                  * need to do anything, the other free-ers did the correct
5759                  * thing.
5760                  */
5761                 BTRFS_I(inode)->csum_bytes = orig_csum_bytes - num_bytes;
5762                 if (bytes > to_free)
5763                         to_free = bytes - to_free;
5764                 else
5765                         to_free = 0;
5766         }
5767         spin_unlock(&BTRFS_I(inode)->lock);
5768         if (dropped)
5769                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5770
5771         if (to_free) {
5772                 btrfs_block_rsv_release(root, block_rsv, to_free);
5773                 trace_btrfs_space_reservation(root->fs_info, "delalloc",
5774                                               btrfs_ino(inode), to_free, 0);
5775         }
5776         if (delalloc_lock)
5777                 mutex_unlock(&BTRFS_I(inode)->delalloc_mutex);
5778         return ret;
5779 }
5780
5781 /**
5782  * btrfs_delalloc_release_metadata - release a metadata reservation for an inode
5783  * @inode: the inode to release the reservation for
5784  * @num_bytes: the number of bytes we're releasing
5785  *
5786  * This will release the metadata reservation for an inode.  This can be called
5787  * once we complete IO for a given set of bytes to release their metadata
5788  * reservations.
5789  */
5790 void btrfs_delalloc_release_metadata(struct inode *inode, u64 num_bytes)
5791 {
5792         struct btrfs_root *root = BTRFS_I(inode)->root;
5793         u64 to_free = 0;
5794         unsigned dropped;
5795
5796         num_bytes = ALIGN(num_bytes, root->sectorsize);
5797         spin_lock(&BTRFS_I(inode)->lock);
5798         dropped = drop_outstanding_extent(inode, num_bytes);
5799
5800         if (num_bytes)
5801                 to_free = calc_csum_metadata_size(inode, num_bytes, 0);
5802         spin_unlock(&BTRFS_I(inode)->lock);
5803         if (dropped > 0)
5804                 to_free += btrfs_calc_trans_metadata_size(root, dropped);
5805
5806         if (btrfs_test_is_dummy_root(root))
5807                 return;
5808
5809         trace_btrfs_space_reservation(root->fs_info, "delalloc",
5810                                       btrfs_ino(inode), to_free, 0);
5811
5812         btrfs_block_rsv_release(root, &root->fs_info->delalloc_block_rsv,
5813                                 to_free);
5814 }
5815
5816 /**
5817  * btrfs_delalloc_reserve_space - reserve data and metadata space for
5818  * delalloc
5819  * @inode: inode we're writing to
5820  * @start: start range we are writing to
5821  * @len: how long the range we are writing to
5822  *
5823  * TODO: This function will finally replace old btrfs_delalloc_reserve_space()
5824  *
5825  * This will do the following things
5826  *
5827  * o reserve space in data space info for num bytes
5828  *   and reserve precious corresponding qgroup space
5829  *   (Done in check_data_free_space)
5830  *
5831  * o reserve space for metadata space, based on the number of outstanding
5832  *   extents and how much csums will be needed
5833  *   also reserve metadata space in a per root over-reserve method.
5834  * o add to the inodes->delalloc_bytes
5835  * o add it to the fs_info's delalloc inodes list.
5836  *   (Above 3 all done in delalloc_reserve_metadata)
5837  *
5838  * Return 0 for success
5839  * Return <0 for error(-ENOSPC or -EQUOT)
5840  */
5841 int btrfs_delalloc_reserve_space(struct inode *inode, u64 start, u64 len)
5842 {
5843         int ret;
5844
5845         ret = btrfs_check_data_free_space(inode, start, len);
5846         if (ret < 0)
5847                 return ret;
5848         ret = btrfs_delalloc_reserve_metadata(inode, len);
5849         if (ret < 0)
5850                 btrfs_free_reserved_data_space(inode, start, len);
5851         return ret;
5852 }
5853
5854 /**
5855  * btrfs_delalloc_release_space - release data and metadata space for delalloc
5856  * @inode: inode we're releasing space for
5857  * @start: start position of the space already reserved
5858  * @len: the len of the space already reserved
5859  *
5860  * This must be matched with a call to btrfs_delalloc_reserve_space.  This is
5861  * called in the case that we don't need the metadata AND data reservations
5862  * anymore.  So if there is an error or we insert an inline extent.
5863  *
5864  * This function will release the metadata space that was not used and will
5865  * decrement ->delalloc_bytes and remove it from the fs_info delalloc_inodes
5866  * list if there are no delalloc bytes left.
5867  * Also it will handle the qgroup reserved space.
5868  */
5869 void btrfs_delalloc_release_space(struct inode *inode, u64 start, u64 len)
5870 {
5871         btrfs_delalloc_release_metadata(inode, len);
5872         btrfs_free_reserved_data_space(inode, start, len);
5873 }
5874
5875 static int update_block_group(struct btrfs_trans_handle *trans,
5876                               struct btrfs_root *root, u64 bytenr,
5877                               u64 num_bytes, int alloc)
5878 {
5879         struct btrfs_block_group_cache *cache = NULL;
5880         struct btrfs_fs_info *info = root->fs_info;
5881         u64 total = num_bytes;
5882         u64 old_val;
5883         u64 byte_in_group;
5884         int factor;
5885
5886         /* block accounting for super block */
5887         spin_lock(&info->delalloc_root_lock);
5888         old_val = btrfs_super_bytes_used(info->super_copy);
5889         if (alloc)
5890                 old_val += num_bytes;
5891         else
5892                 old_val -= num_bytes;
5893         btrfs_set_super_bytes_used(info->super_copy, old_val);
5894         spin_unlock(&info->delalloc_root_lock);
5895
5896         while (total) {
5897                 cache = btrfs_lookup_block_group(info, bytenr);
5898                 if (!cache)
5899                         return -ENOENT;
5900                 if (cache->flags & (BTRFS_BLOCK_GROUP_DUP |
5901                                     BTRFS_BLOCK_GROUP_RAID1 |
5902                                     BTRFS_BLOCK_GROUP_RAID10))
5903                         factor = 2;
5904                 else
5905                         factor = 1;
5906                 /*
5907                  * If this block group has free space cache written out, we
5908                  * need to make sure to load it if we are removing space.  This
5909                  * is because we need the unpinning stage to actually add the
5910                  * space back to the block group, otherwise we will leak space.
5911                  */
5912                 if (!alloc && cache->cached == BTRFS_CACHE_NO)
5913                         cache_block_group(cache, 1);
5914
5915                 byte_in_group = bytenr - cache->key.objectid;
5916                 WARN_ON(byte_in_group > cache->key.offset);
5917
5918                 spin_lock(&cache->space_info->lock);
5919                 spin_lock(&cache->lock);
5920
5921                 if (btrfs_test_opt(root, SPACE_CACHE) &&
5922                     cache->disk_cache_state < BTRFS_DC_CLEAR)
5923                         cache->disk_cache_state = BTRFS_DC_CLEAR;
5924
5925                 old_val = btrfs_block_group_used(&cache->item);
5926                 num_bytes = min(total, cache->key.offset - byte_in_group);
5927                 if (alloc) {
5928                         old_val += num_bytes;
5929                         btrfs_set_block_group_used(&cache->item, old_val);
5930                         cache->reserved -= num_bytes;
5931                         cache->space_info->bytes_reserved -= num_bytes;
5932                         cache->space_info->bytes_used += num_bytes;
5933                         cache->space_info->disk_used += num_bytes * factor;
5934                         spin_unlock(&cache->lock);
5935                         spin_unlock(&cache->space_info->lock);
5936                 } else {
5937                         old_val -= num_bytes;
5938                         btrfs_set_block_group_used(&cache->item, old_val);
5939                         cache->pinned += num_bytes;
5940                         cache->space_info->bytes_pinned += num_bytes;
5941                         cache->space_info->bytes_used -= num_bytes;
5942                         cache->space_info->disk_used -= num_bytes * factor;
5943                         spin_unlock(&cache->lock);
5944                         spin_unlock(&cache->space_info->lock);
5945
5946                         set_extent_dirty(info->pinned_extents,
5947                                          bytenr, bytenr + num_bytes - 1,
5948                                          GFP_NOFS | __GFP_NOFAIL);
5949                 }
5950
5951                 spin_lock(&trans->transaction->dirty_bgs_lock);
5952                 if (list_empty(&cache->dirty_list)) {
5953                         list_add_tail(&cache->dirty_list,
5954                                       &trans->transaction->dirty_bgs);
5955                                 trans->transaction->num_dirty_bgs++;
5956                         btrfs_get_block_group(cache);
5957                 }
5958                 spin_unlock(&trans->transaction->dirty_bgs_lock);
5959
5960                 /*
5961                  * No longer have used bytes in this block group, queue it for
5962                  * deletion. We do this after adding the block group to the
5963                  * dirty list to avoid races between cleaner kthread and space
5964                  * cache writeout.
5965                  */
5966                 if (!alloc && old_val == 0) {
5967                         spin_lock(&info->unused_bgs_lock);
5968                         if (list_empty(&cache->bg_list)) {
5969                                 btrfs_get_block_group(cache);
5970                                 list_add_tail(&cache->bg_list,
5971                                               &info->unused_bgs);
5972                         }
5973                         spin_unlock(&info->unused_bgs_lock);
5974                 }
5975
5976                 btrfs_put_block_group(cache);
5977                 total -= num_bytes;
5978                 bytenr += num_bytes;
5979         }
5980         return 0;
5981 }
5982
5983 static u64 first_logical_byte(struct btrfs_root *root, u64 search_start)
5984 {
5985         struct btrfs_block_group_cache *cache;
5986         u64 bytenr;
5987
5988         spin_lock(&root->fs_info->block_group_cache_lock);
5989         bytenr = root->fs_info->first_logical_byte;
5990         spin_unlock(&root->fs_info->block_group_cache_lock);
5991
5992         if (bytenr < (u64)-1)
5993                 return bytenr;
5994
5995         cache = btrfs_lookup_first_block_group(root->fs_info, search_start);
5996         if (!cache)
5997                 return 0;
5998
5999         bytenr = cache->key.objectid;
6000         btrfs_put_block_group(cache);
6001
6002         return bytenr;
6003 }
6004
6005 static int pin_down_extent(struct btrfs_root *root,
6006                            struct btrfs_block_group_cache *cache,
6007                            u64 bytenr, u64 num_bytes, int reserved)
6008 {
6009         spin_lock(&cache->space_info->lock);
6010         spin_lock(&cache->lock);
6011         cache->pinned += num_bytes;
6012         cache->space_info->bytes_pinned += num_bytes;
6013         if (reserved) {
6014                 cache->reserved -= num_bytes;
6015                 cache->space_info->bytes_reserved -= num_bytes;
6016         }
6017         spin_unlock(&cache->lock);
6018         spin_unlock(&cache->space_info->lock);
6019
6020         set_extent_dirty(root->fs_info->pinned_extents, bytenr,
6021                          bytenr + num_bytes - 1, GFP_NOFS | __GFP_NOFAIL);
6022         if (reserved)
6023                 trace_btrfs_reserved_extent_free(root, bytenr, num_bytes);
6024         return 0;
6025 }
6026
6027 /*
6028  * this function must be called within transaction
6029  */
6030 int btrfs_pin_extent(struct btrfs_root *root,
6031                      u64 bytenr, u64 num_bytes, int reserved)
6032 {
6033         struct btrfs_block_group_cache *cache;
6034
6035         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6036         BUG_ON(!cache); /* Logic error */
6037
6038         pin_down_extent(root, cache, bytenr, num_bytes, reserved);
6039
6040         btrfs_put_block_group(cache);
6041         return 0;
6042 }
6043
6044 /*
6045  * this function must be called within transaction
6046  */
6047 int btrfs_pin_extent_for_log_replay(struct btrfs_root *root,
6048                                     u64 bytenr, u64 num_bytes)
6049 {
6050         struct btrfs_block_group_cache *cache;
6051         int ret;
6052
6053         cache = btrfs_lookup_block_group(root->fs_info, bytenr);
6054         if (!cache)
6055                 return -EINVAL;
6056
6057         /*
6058          * pull in the free space cache (if any) so that our pin
6059          * removes the free space from the cache.  We have load_only set
6060          * to one because the slow code to read in the free extents does check
6061          * the pinned extents.
6062          */
6063         cache_block_group(cache, 1);
6064
6065         pin_down_extent(root, cache, bytenr, num_bytes, 0);
6066
6067         /* remove us from the free space cache (if we're there at all) */
6068         ret = btrfs_remove_free_space(cache, bytenr, num_bytes);
6069         btrfs_put_block_group(cache);
6070         return ret;
6071 }
6072
6073 static int __exclude_logged_extent(struct btrfs_root *root, u64 start, u64 num_bytes)
6074 {
6075         int ret;
6076         struct btrfs_block_group_cache *block_group;
6077         struct btrfs_caching_control *caching_ctl;
6078
6079         block_group = btrfs_lookup_block_group(root->fs_info, start);
6080         if (!block_group)
6081                 return -EINVAL;
6082
6083         cache_block_group(block_group, 0);
6084         caching_ctl = get_caching_control(block_group);
6085
6086         if (!caching_ctl) {
6087                 /* Logic error */
6088                 BUG_ON(!block_group_cache_done(block_group));
6089                 ret = btrfs_remove_free_space(block_group, start, num_bytes);
6090         } else {
6091                 mutex_lock(&caching_ctl->mutex);
6092
6093                 if (start >= caching_ctl->progress) {
6094                         ret = add_excluded_extent(root, start, num_bytes);
6095                 } else if (start + num_bytes <= caching_ctl->progress) {
6096                         ret = btrfs_remove_free_space(block_group,
6097                                                       start, num_bytes);
6098                 } else {
6099                         num_bytes = caching_ctl->progress - start;
6100                         ret = btrfs_remove_free_space(block_group,
6101                                                       start, num_bytes);
6102                         if (ret)
6103                                 goto out_lock;
6104
6105                         num_bytes = (start + num_bytes) -
6106                                 caching_ctl->progress;
6107                         start = caching_ctl->progress;
6108                         ret = add_excluded_extent(root, start, num_bytes);
6109                 }
6110 out_lock:
6111                 mutex_unlock(&caching_ctl->mutex);
6112                 put_caching_control(caching_ctl);
6113         }
6114         btrfs_put_block_group(block_group);
6115         return ret;
6116 }
6117
6118 int btrfs_exclude_logged_extents(struct btrfs_root *log,
6119                                  struct extent_buffer *eb)
6120 {
6121         struct btrfs_file_extent_item *item;
6122         struct btrfs_key key;
6123         int found_type;
6124         int i;
6125
6126         if (!btrfs_fs_incompat(log->fs_info, MIXED_GROUPS))
6127                 return 0;
6128
6129         for (i = 0; i < btrfs_header_nritems(eb); i++) {
6130                 btrfs_item_key_to_cpu(eb, &key, i);
6131                 if (key.type != BTRFS_EXTENT_DATA_KEY)
6132                         continue;
6133                 item = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
6134                 found_type = btrfs_file_extent_type(eb, item);
6135                 if (found_type == BTRFS_FILE_EXTENT_INLINE)
6136                         continue;
6137                 if (btrfs_file_extent_disk_bytenr(eb, item) == 0)
6138                         continue;
6139                 key.objectid = btrfs_file_extent_disk_bytenr(eb, item);
6140                 key.offset = btrfs_file_extent_disk_num_bytes(eb, item);
6141                 __exclude_logged_extent(log, key.objectid, key.offset);
6142         }
6143
6144         return 0;
6145 }
6146
6147 /**
6148  * btrfs_update_reserved_bytes - update the block_group and space info counters
6149  * @cache:      The cache we are manipulating
6150  * @num_bytes:  The number of bytes in question
6151  * @reserve:    One of the reservation enums
6152  * @delalloc:   The blocks are allocated for the delalloc write
6153  *
6154  * This is called by the allocator when it reserves space, or by somebody who is
6155  * freeing space that was never actually used on disk.  For example if you
6156  * reserve some space for a new leaf in transaction A and before transaction A
6157  * commits you free that leaf, you call this with reserve set to 0 in order to
6158  * clear the reservation.
6159  *
6160  * Metadata reservations should be called with RESERVE_ALLOC so we do the proper
6161  * ENOSPC accounting.  For data we handle the reservation through clearing the
6162  * delalloc bits in the io_tree.  We have to do this since we could end up
6163  * allocating less disk space for the amount of data we have reserved in the
6164  * case of compression.
6165  *
6166  * If this is a reservation and the block group has become read only we cannot
6167  * make the reservation and return -EAGAIN, otherwise this function always
6168  * succeeds.
6169  */
6170 static int btrfs_update_reserved_bytes(struct btrfs_block_group_cache *cache,
6171                                        u64 num_bytes, int reserve, int delalloc)
6172 {
6173         struct btrfs_space_info *space_info = cache->space_info;
6174         int ret = 0;
6175
6176         spin_lock(&space_info->lock);
6177         spin_lock(&cache->lock);
6178         if (reserve != RESERVE_FREE) {
6179                 if (cache->ro) {
6180                         ret = -EAGAIN;
6181                 } else {
6182                         cache->reserved += num_bytes;
6183                         space_info->bytes_reserved += num_bytes;
6184                         if (reserve == RESERVE_ALLOC) {
6185                                 trace_btrfs_space_reservation(cache->fs_info,
6186                                                 "space_info", space_info->flags,
6187                                                 num_bytes, 0);
6188                                 space_info->bytes_may_use -= num_bytes;
6189                         }
6190
6191                         if (delalloc)
6192                                 cache->delalloc_bytes += num_bytes;
6193                 }
6194         } else {
6195                 if (cache->ro)
6196                         space_info->bytes_readonly += num_bytes;
6197                 cache->reserved -= num_bytes;
6198                 space_info->bytes_reserved -= num_bytes;
6199
6200                 if (delalloc)
6201                         cache->delalloc_bytes -= num_bytes;
6202         }
6203         spin_unlock(&cache->lock);
6204         spin_unlock(&space_info->lock);
6205         return ret;
6206 }
6207
6208 void btrfs_prepare_extent_commit(struct btrfs_trans_handle *trans,
6209                                 struct btrfs_root *root)
6210 {
6211         struct btrfs_fs_info *fs_info = root->fs_info;
6212         struct btrfs_caching_control *next;
6213         struct btrfs_caching_control *caching_ctl;
6214         struct btrfs_block_group_cache *cache;
6215
6216         down_write(&fs_info->commit_root_sem);
6217
6218         list_for_each_entry_safe(caching_ctl, next,
6219                                  &fs_info->caching_block_groups, list) {
6220                 cache = caching_ctl->block_group;
6221                 if (block_group_cache_done(cache)) {
6222                         cache->last_byte_to_unpin = (u64)-1;
6223                         list_del_init(&caching_ctl->list);
6224                         put_caching_control(caching_ctl);
6225                 } else {
6226                         cache->last_byte_to_unpin = caching_ctl->progress;
6227                 }
6228         }
6229
6230         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6231                 fs_info->pinned_extents = &fs_info->freed_extents[1];
6232         else
6233                 fs_info->pinned_extents = &fs_info->freed_extents[0];
6234
6235         up_write(&fs_info->commit_root_sem);
6236
6237         update_global_block_rsv(fs_info);
6238 }
6239
6240 /*
6241  * Returns the free cluster for the given space info and sets empty_cluster to
6242  * what it should be based on the mount options.
6243  */
6244 static struct btrfs_free_cluster *
6245 fetch_cluster_info(struct btrfs_root *root, struct btrfs_space_info *space_info,
6246                    u64 *empty_cluster)
6247 {
6248         struct btrfs_free_cluster *ret = NULL;
6249         bool ssd = btrfs_test_opt(root, SSD);
6250
6251         *empty_cluster = 0;
6252         if (btrfs_mixed_space_info(space_info))
6253                 return ret;
6254
6255         if (ssd)
6256                 *empty_cluster = 2 * 1024 * 1024;
6257         if (space_info->flags & BTRFS_BLOCK_GROUP_METADATA) {
6258                 ret = &root->fs_info->meta_alloc_cluster;
6259                 if (!ssd)
6260                         *empty_cluster = 64 * 1024;
6261         } else if ((space_info->flags & BTRFS_BLOCK_GROUP_DATA) && ssd) {
6262                 ret = &root->fs_info->data_alloc_cluster;
6263         }
6264
6265         return ret;
6266 }
6267
6268 static int unpin_extent_range(struct btrfs_root *root, u64 start, u64 end,
6269                               const bool return_free_space)
6270 {
6271         struct btrfs_fs_info *fs_info = root->fs_info;
6272         struct btrfs_block_group_cache *cache = NULL;
6273         struct btrfs_space_info *space_info;
6274         struct btrfs_block_rsv *global_rsv = &fs_info->global_block_rsv;
6275         struct btrfs_free_cluster *cluster = NULL;
6276         u64 len;
6277         u64 total_unpinned = 0;
6278         u64 empty_cluster = 0;
6279         bool readonly;
6280
6281         while (start <= end) {
6282                 readonly = false;
6283                 if (!cache ||
6284                     start >= cache->key.objectid + cache->key.offset) {
6285                         if (cache)
6286                                 btrfs_put_block_group(cache);
6287                         total_unpinned = 0;
6288                         cache = btrfs_lookup_block_group(fs_info, start);
6289                         BUG_ON(!cache); /* Logic error */
6290
6291                         cluster = fetch_cluster_info(root,
6292                                                      cache->space_info,
6293                                                      &empty_cluster);
6294                         empty_cluster <<= 1;
6295                 }
6296
6297                 len = cache->key.objectid + cache->key.offset - start;
6298                 len = min(len, end + 1 - start);
6299
6300                 if (start < cache->last_byte_to_unpin) {
6301                         len = min(len, cache->last_byte_to_unpin - start);
6302                         if (return_free_space)
6303                                 btrfs_add_free_space(cache, start, len);
6304                 }
6305
6306                 start += len;
6307                 total_unpinned += len;
6308                 space_info = cache->space_info;
6309
6310                 /*
6311                  * If this space cluster has been marked as fragmented and we've
6312                  * unpinned enough in this block group to potentially allow a
6313                  * cluster to be created inside of it go ahead and clear the
6314                  * fragmented check.
6315                  */
6316                 if (cluster && cluster->fragmented &&
6317                     total_unpinned > empty_cluster) {
6318                         spin_lock(&cluster->lock);
6319                         cluster->fragmented = 0;
6320                         spin_unlock(&cluster->lock);
6321                 }
6322
6323                 spin_lock(&space_info->lock);
6324                 spin_lock(&cache->lock);
6325                 cache->pinned -= len;
6326                 space_info->bytes_pinned -= len;
6327                 space_info->max_extent_size = 0;
6328                 percpu_counter_add(&space_info->total_bytes_pinned, -len);
6329                 if (cache->ro) {
6330                         space_info->bytes_readonly += len;
6331                         readonly = true;
6332                 }
6333                 spin_unlock(&cache->lock);
6334                 if (!readonly && global_rsv->space_info == space_info) {
6335                         spin_lock(&global_rsv->lock);
6336                         if (!global_rsv->full) {
6337                                 len = min(len, global_rsv->size -
6338                                           global_rsv->reserved);
6339                                 global_rsv->reserved += len;
6340                                 space_info->bytes_may_use += len;
6341                                 if (global_rsv->reserved >= global_rsv->size)
6342                                         global_rsv->full = 1;
6343                         }
6344                         spin_unlock(&global_rsv->lock);
6345                 }
6346                 spin_unlock(&space_info->lock);
6347         }
6348
6349         if (cache)
6350                 btrfs_put_block_group(cache);
6351         return 0;
6352 }
6353
6354 int btrfs_finish_extent_commit(struct btrfs_trans_handle *trans,
6355                                struct btrfs_root *root)
6356 {
6357         struct btrfs_fs_info *fs_info = root->fs_info;
6358         struct btrfs_block_group_cache *block_group, *tmp;
6359         struct list_head *deleted_bgs;
6360         struct extent_io_tree *unpin;
6361         u64 start;
6362         u64 end;
6363         int ret;
6364
6365         if (fs_info->pinned_extents == &fs_info->freed_extents[0])
6366                 unpin = &fs_info->freed_extents[1];
6367         else
6368                 unpin = &fs_info->freed_extents[0];
6369
6370         while (!trans->aborted) {
6371                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
6372                 ret = find_first_extent_bit(unpin, 0, &start, &end,
6373                                             EXTENT_DIRTY, NULL);
6374                 if (ret) {
6375                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6376                         break;
6377                 }
6378
6379                 if (btrfs_test_opt(root, DISCARD))
6380                         ret = btrfs_discard_extent(root, start,
6381                                                    end + 1 - start, NULL);
6382
6383                 clear_extent_dirty(unpin, start, end, GFP_NOFS);
6384                 unpin_extent_range(root, start, end, true);
6385                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
6386                 cond_resched();
6387         }
6388
6389         /*
6390          * Transaction is finished.  We don't need the lock anymore.  We
6391          * do need to clean up the block groups in case of a transaction
6392          * abort.
6393          */
6394         deleted_bgs = &trans->transaction->deleted_bgs;
6395         list_for_each_entry_safe(block_group, tmp, deleted_bgs, bg_list) {
6396                 u64 trimmed = 0;
6397
6398                 ret = -EROFS;
6399                 if (!trans->aborted)
6400                         ret = btrfs_discard_extent(root,
6401                                                    block_group->key.objectid,
6402                                                    block_group->key.offset,
6403                                                    &trimmed);
6404
6405                 list_del_init(&block_group->bg_list);
6406                 btrfs_put_block_group_trimming(block_group);
6407                 btrfs_put_block_group(block_group);
6408
6409                 if (ret) {
6410                         const char *errstr = btrfs_decode_error(ret);
6411                         btrfs_warn(fs_info,
6412                                    "Discard failed while removing blockgroup: errno=%d %s\n",
6413                                    ret, errstr);
6414                 }
6415         }
6416
6417         return 0;
6418 }
6419
6420 static void add_pinned_bytes(struct btrfs_fs_info *fs_info, u64 num_bytes,
6421                              u64 owner, u64 root_objectid)
6422 {
6423         struct btrfs_space_info *space_info;
6424         u64 flags;
6425
6426         if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6427                 if (root_objectid == BTRFS_CHUNK_TREE_OBJECTID)
6428                         flags = BTRFS_BLOCK_GROUP_SYSTEM;
6429                 else
6430                         flags = BTRFS_BLOCK_GROUP_METADATA;
6431         } else {
6432                 flags = BTRFS_BLOCK_GROUP_DATA;
6433         }
6434
6435         space_info = __find_space_info(fs_info, flags);
6436         BUG_ON(!space_info); /* Logic bug */
6437         percpu_counter_add(&space_info->total_bytes_pinned, num_bytes);
6438 }
6439
6440
6441 static int __btrfs_free_extent(struct btrfs_trans_handle *trans,
6442                                 struct btrfs_root *root,
6443                                 struct btrfs_delayed_ref_node *node, u64 parent,
6444                                 u64 root_objectid, u64 owner_objectid,
6445                                 u64 owner_offset, int refs_to_drop,
6446                                 struct btrfs_delayed_extent_op *extent_op)
6447 {
6448         struct btrfs_key key;
6449         struct btrfs_path *path;
6450         struct btrfs_fs_info *info = root->fs_info;
6451         struct btrfs_root *extent_root = info->extent_root;
6452         struct extent_buffer *leaf;
6453         struct btrfs_extent_item *ei;
6454         struct btrfs_extent_inline_ref *iref;
6455         int ret;
6456         int is_data;
6457         int extent_slot = 0;
6458         int found_extent = 0;
6459         int num_to_del = 1;
6460         u32 item_size;
6461         u64 refs;
6462         u64 bytenr = node->bytenr;
6463         u64 num_bytes = node->num_bytes;
6464         int last_ref = 0;
6465         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
6466                                                  SKINNY_METADATA);
6467
6468         path = btrfs_alloc_path();
6469         if (!path)
6470                 return -ENOMEM;
6471
6472         path->reada = 1;
6473         path->leave_spinning = 1;
6474
6475         is_data = owner_objectid >= BTRFS_FIRST_FREE_OBJECTID;
6476         BUG_ON(!is_data && refs_to_drop != 1);
6477
6478         if (is_data)
6479                 skinny_metadata = 0;
6480
6481         ret = lookup_extent_backref(trans, extent_root, path, &iref,
6482                                     bytenr, num_bytes, parent,
6483                                     root_objectid, owner_objectid,
6484                                     owner_offset);
6485         if (ret == 0) {
6486                 extent_slot = path->slots[0];
6487                 while (extent_slot >= 0) {
6488                         btrfs_item_key_to_cpu(path->nodes[0], &key,
6489                                               extent_slot);
6490                         if (key.objectid != bytenr)
6491                                 break;
6492                         if (key.type == BTRFS_EXTENT_ITEM_KEY &&
6493                             key.offset == num_bytes) {
6494                                 found_extent = 1;
6495                                 break;
6496                         }
6497                         if (key.type == BTRFS_METADATA_ITEM_KEY &&
6498                             key.offset == owner_objectid) {
6499                                 found_extent = 1;
6500                                 break;
6501                         }
6502                         if (path->slots[0] - extent_slot > 5)
6503                                 break;
6504                         extent_slot--;
6505                 }
6506 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6507                 item_size = btrfs_item_size_nr(path->nodes[0], extent_slot);
6508                 if (found_extent && item_size < sizeof(*ei))
6509                         found_extent = 0;
6510 #endif
6511                 if (!found_extent) {
6512                         BUG_ON(iref);
6513                         ret = remove_extent_backref(trans, extent_root, path,
6514                                                     NULL, refs_to_drop,
6515                                                     is_data, &last_ref);
6516                         if (ret) {
6517                                 btrfs_abort_transaction(trans, extent_root, ret);
6518                                 goto out;
6519                         }
6520                         btrfs_release_path(path);
6521                         path->leave_spinning = 1;
6522
6523                         key.objectid = bytenr;
6524                         key.type = BTRFS_EXTENT_ITEM_KEY;
6525                         key.offset = num_bytes;
6526
6527                         if (!is_data && skinny_metadata) {
6528                                 key.type = BTRFS_METADATA_ITEM_KEY;
6529                                 key.offset = owner_objectid;
6530                         }
6531
6532                         ret = btrfs_search_slot(trans, extent_root,
6533                                                 &key, path, -1, 1);
6534                         if (ret > 0 && skinny_metadata && path->slots[0]) {
6535                                 /*
6536                                  * Couldn't find our skinny metadata item,
6537                                  * see if we have ye olde extent item.
6538                                  */
6539                                 path->slots[0]--;
6540                                 btrfs_item_key_to_cpu(path->nodes[0], &key,
6541                                                       path->slots[0]);
6542                                 if (key.objectid == bytenr &&
6543                                     key.type == BTRFS_EXTENT_ITEM_KEY &&
6544                                     key.offset == num_bytes)
6545                                         ret = 0;
6546                         }
6547
6548                         if (ret > 0 && skinny_metadata) {
6549                                 skinny_metadata = false;
6550                                 key.objectid = bytenr;
6551                                 key.type = BTRFS_EXTENT_ITEM_KEY;
6552                                 key.offset = num_bytes;
6553                                 btrfs_release_path(path);
6554                                 ret = btrfs_search_slot(trans, extent_root,
6555                                                         &key, path, -1, 1);
6556                         }
6557
6558                         if (ret) {
6559                                 btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6560                                         ret, bytenr);
6561                                 if (ret > 0)
6562                                         btrfs_print_leaf(extent_root,
6563                                                          path->nodes[0]);
6564                         }
6565                         if (ret < 0) {
6566                                 btrfs_abort_transaction(trans, extent_root, ret);
6567                                 goto out;
6568                         }
6569                         extent_slot = path->slots[0];
6570                 }
6571         } else if (WARN_ON(ret == -ENOENT)) {
6572                 btrfs_print_leaf(extent_root, path->nodes[0]);
6573                 btrfs_err(info,
6574                         "unable to find ref byte nr %llu parent %llu root %llu  owner %llu offset %llu",
6575                         bytenr, parent, root_objectid, owner_objectid,
6576                         owner_offset);
6577                 btrfs_abort_transaction(trans, extent_root, ret);
6578                 goto out;
6579         } else {
6580                 btrfs_abort_transaction(trans, extent_root, ret);
6581                 goto out;
6582         }
6583
6584         leaf = path->nodes[0];
6585         item_size = btrfs_item_size_nr(leaf, extent_slot);
6586 #ifdef BTRFS_COMPAT_EXTENT_TREE_V0
6587         if (item_size < sizeof(*ei)) {
6588                 BUG_ON(found_extent || extent_slot != path->slots[0]);
6589                 ret = convert_extent_item_v0(trans, extent_root, path,
6590                                              owner_objectid, 0);
6591                 if (ret < 0) {
6592                         btrfs_abort_transaction(trans, extent_root, ret);
6593                         goto out;
6594                 }
6595
6596                 btrfs_release_path(path);
6597                 path->leave_spinning = 1;
6598
6599                 key.objectid = bytenr;
6600                 key.type = BTRFS_EXTENT_ITEM_KEY;
6601                 key.offset = num_bytes;
6602
6603                 ret = btrfs_search_slot(trans, extent_root, &key, path,
6604                                         -1, 1);
6605                 if (ret) {
6606                         btrfs_err(info, "umm, got %d back from search, was looking for %llu",
6607                                 ret, bytenr);
6608                         btrfs_print_leaf(extent_root, path->nodes[0]);
6609                 }
6610                 if (ret < 0) {
6611                         btrfs_abort_transaction(trans, extent_root, ret);
6612                         goto out;
6613                 }
6614
6615                 extent_slot = path->slots[0];
6616                 leaf = path->nodes[0];
6617                 item_size = btrfs_item_size_nr(leaf, extent_slot);
6618         }
6619 #endif
6620         BUG_ON(item_size < sizeof(*ei));
6621         ei = btrfs_item_ptr(leaf, extent_slot,
6622                             struct btrfs_extent_item);
6623         if (owner_objectid < BTRFS_FIRST_FREE_OBJECTID &&
6624             key.type == BTRFS_EXTENT_ITEM_KEY) {
6625                 struct btrfs_tree_block_info *bi;
6626                 BUG_ON(item_size < sizeof(*ei) + sizeof(*bi));
6627                 bi = (struct btrfs_tree_block_info *)(ei + 1);
6628                 WARN_ON(owner_objectid != btrfs_tree_block_level(leaf, bi));
6629         }
6630
6631         refs = btrfs_extent_refs(leaf, ei);
6632         if (refs < refs_to_drop) {
6633                 btrfs_err(info, "trying to drop %d refs but we only have %Lu "
6634                           "for bytenr %Lu", refs_to_drop, refs, bytenr);
6635                 ret = -EINVAL;
6636                 btrfs_abort_transaction(trans, extent_root, ret);
6637                 goto out;
6638         }
6639         refs -= refs_to_drop;
6640
6641         if (refs > 0) {
6642                 if (extent_op)
6643                         __run_delayed_extent_op(extent_op, leaf, ei);
6644                 /*
6645                  * In the case of inline back ref, reference count will
6646                  * be updated by remove_extent_backref
6647                  */
6648                 if (iref) {
6649                         BUG_ON(!found_extent);
6650                 } else {
6651                         btrfs_set_extent_refs(leaf, ei, refs);
6652                         btrfs_mark_buffer_dirty(leaf);
6653                 }
6654                 if (found_extent) {
6655                         ret = remove_extent_backref(trans, extent_root, path,
6656                                                     iref, refs_to_drop,
6657                                                     is_data, &last_ref);
6658                         if (ret) {
6659                                 btrfs_abort_transaction(trans, extent_root, ret);
6660                                 goto out;
6661                         }
6662                 }
6663                 add_pinned_bytes(root->fs_info, -num_bytes, owner_objectid,
6664                                  root_objectid);
6665         } else {
6666                 if (found_extent) {
6667                         BUG_ON(is_data && refs_to_drop !=
6668                                extent_data_ref_count(path, iref));
6669                         if (iref) {
6670                                 BUG_ON(path->slots[0] != extent_slot);
6671                         } else {
6672                                 BUG_ON(path->slots[0] != extent_slot + 1);
6673                                 path->slots[0] = extent_slot;
6674                                 num_to_del = 2;
6675                         }
6676                 }
6677
6678                 last_ref = 1;
6679                 ret = btrfs_del_items(trans, extent_root, path, path->slots[0],
6680                                       num_to_del);
6681                 if (ret) {
6682                         btrfs_abort_transaction(trans, extent_root, ret);
6683                         goto out;
6684                 }
6685                 btrfs_release_path(path);
6686
6687                 if (is_data) {
6688                         ret = btrfs_del_csums(trans, root, bytenr, num_bytes);
6689                         if (ret) {
6690                                 btrfs_abort_transaction(trans, extent_root, ret);
6691                                 goto out;
6692                         }
6693                 }
6694
6695                 ret = add_to_free_space_tree(trans, root->fs_info, bytenr,
6696                                              num_bytes);
6697                 if (ret) {
6698                         btrfs_abort_transaction(trans, extent_root, ret);
6699                         goto out;
6700                 }
6701
6702                 ret = update_block_group(trans, root, bytenr, num_bytes, 0);
6703                 if (ret) {
6704                         btrfs_abort_transaction(trans, extent_root, ret);
6705                         goto out;
6706                 }
6707         }
6708         btrfs_release_path(path);
6709
6710 out:
6711         btrfs_free_path(path);
6712         return ret;
6713 }
6714
6715 /*
6716  * when we free an block, it is possible (and likely) that we free the last
6717  * delayed ref for that extent as well.  This searches the delayed ref tree for
6718  * a given extent, and if there are no other delayed refs to be processed, it
6719  * removes it from the tree.
6720  */
6721 static noinline int check_ref_cleanup(struct btrfs_trans_handle *trans,
6722                                       struct btrfs_root *root, u64 bytenr)
6723 {
6724         struct btrfs_delayed_ref_head *head;
6725         struct btrfs_delayed_ref_root *delayed_refs;
6726         int ret = 0;
6727
6728         delayed_refs = &trans->transaction->delayed_refs;
6729         spin_lock(&delayed_refs->lock);
6730         head = btrfs_find_delayed_ref_head(trans, bytenr);
6731         if (!head)
6732                 goto out_delayed_unlock;
6733
6734         spin_lock(&head->lock);
6735         if (!list_empty(&head->ref_list))
6736                 goto out;
6737
6738         if (head->extent_op) {
6739                 if (!head->must_insert_reserved)
6740                         goto out;
6741                 btrfs_free_delayed_extent_op(head->extent_op);
6742                 head->extent_op = NULL;
6743         }
6744
6745         /*
6746          * waiting for the lock here would deadlock.  If someone else has it
6747          * locked they are already in the process of dropping it anyway
6748          */
6749         if (!mutex_trylock(&head->mutex))
6750                 goto out;
6751
6752         /*
6753          * at this point we have a head with no other entries.  Go
6754          * ahead and process it.
6755          */
6756         head->node.in_tree = 0;
6757         rb_erase(&head->href_node, &delayed_refs->href_root);
6758
6759         atomic_dec(&delayed_refs->num_entries);
6760
6761         /*
6762          * we don't take a ref on the node because we're removing it from the
6763          * tree, so we just steal the ref the tree was holding.
6764          */
6765         delayed_refs->num_heads--;
6766         if (head->processing == 0)
6767                 delayed_refs->num_heads_ready--;
6768         head->processing = 0;
6769         spin_unlock(&head->lock);
6770         spin_unlock(&delayed_refs->lock);
6771
6772         BUG_ON(head->extent_op);
6773         if (head->must_insert_reserved)
6774                 ret = 1;
6775
6776         mutex_unlock(&head->mutex);
6777         btrfs_put_delayed_ref(&head->node);
6778         return ret;
6779 out:
6780         spin_unlock(&head->lock);
6781
6782 out_delayed_unlock:
6783         spin_unlock(&delayed_refs->lock);
6784         return 0;
6785 }
6786
6787 void btrfs_free_tree_block(struct btrfs_trans_handle *trans,
6788                            struct btrfs_root *root,
6789                            struct extent_buffer *buf,
6790                            u64 parent, int last_ref)
6791 {
6792         int pin = 1;
6793         int ret;
6794
6795         if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6796                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
6797                                         buf->start, buf->len,
6798                                         parent, root->root_key.objectid,
6799                                         btrfs_header_level(buf),
6800                                         BTRFS_DROP_DELAYED_REF, NULL);
6801                 BUG_ON(ret); /* -ENOMEM */
6802         }
6803
6804         if (!last_ref)
6805                 return;
6806
6807         if (btrfs_header_generation(buf) == trans->transid) {
6808                 struct btrfs_block_group_cache *cache;
6809
6810                 if (root->root_key.objectid != BTRFS_TREE_LOG_OBJECTID) {
6811                         ret = check_ref_cleanup(trans, root, buf->start);
6812                         if (!ret)
6813                                 goto out;
6814                 }
6815
6816                 cache = btrfs_lookup_block_group(root->fs_info, buf->start);
6817
6818                 if (btrfs_header_flag(buf, BTRFS_HEADER_FLAG_WRITTEN)) {
6819                         pin_down_extent(root, cache, buf->start, buf->len, 1);
6820                         btrfs_put_block_group(cache);
6821                         goto out;
6822                 }
6823
6824                 WARN_ON(test_bit(EXTENT_BUFFER_DIRTY, &buf->bflags));
6825
6826                 btrfs_add_free_space(cache, buf->start, buf->len);
6827                 btrfs_update_reserved_bytes(cache, buf->len, RESERVE_FREE, 0);
6828                 btrfs_put_block_group(cache);
6829                 trace_btrfs_reserved_extent_free(root, buf->start, buf->len);
6830                 pin = 0;
6831         }
6832 out:
6833         if (pin)
6834                 add_pinned_bytes(root->fs_info, buf->len,
6835                                  btrfs_header_level(buf),
6836                                  root->root_key.objectid);
6837
6838         /*
6839          * Deleting the buffer, clear the corrupt flag since it doesn't matter
6840          * anymore.
6841          */
6842         clear_bit(EXTENT_BUFFER_CORRUPT, &buf->bflags);
6843 }
6844
6845 /* Can return -ENOMEM */
6846 int btrfs_free_extent(struct btrfs_trans_handle *trans, struct btrfs_root *root,
6847                       u64 bytenr, u64 num_bytes, u64 parent, u64 root_objectid,
6848                       u64 owner, u64 offset)
6849 {
6850         int ret;
6851         struct btrfs_fs_info *fs_info = root->fs_info;
6852
6853         if (btrfs_test_is_dummy_root(root))
6854                 return 0;
6855
6856         add_pinned_bytes(root->fs_info, num_bytes, owner, root_objectid);
6857
6858         /*
6859          * tree log blocks never actually go into the extent allocation
6860          * tree, just update pinning info and exit early.
6861          */
6862         if (root_objectid == BTRFS_TREE_LOG_OBJECTID) {
6863                 WARN_ON(owner >= BTRFS_FIRST_FREE_OBJECTID);
6864                 /* unlocks the pinned mutex */
6865                 btrfs_pin_extent(root, bytenr, num_bytes, 1);
6866                 ret = 0;
6867         } else if (owner < BTRFS_FIRST_FREE_OBJECTID) {
6868                 ret = btrfs_add_delayed_tree_ref(fs_info, trans, bytenr,
6869                                         num_bytes,
6870                                         parent, root_objectid, (int)owner,
6871                                         BTRFS_DROP_DELAYED_REF, NULL);
6872         } else {
6873                 ret = btrfs_add_delayed_data_ref(fs_info, trans, bytenr,
6874                                                 num_bytes,
6875                                                 parent, root_objectid, owner,
6876                                                 offset, 0,
6877                                                 BTRFS_DROP_DELAYED_REF, NULL);
6878         }
6879         return ret;
6880 }
6881
6882 /*
6883  * when we wait for progress in the block group caching, its because
6884  * our allocation attempt failed at least once.  So, we must sleep
6885  * and let some progress happen before we try again.
6886  *
6887  * This function will sleep at least once waiting for new free space to
6888  * show up, and then it will check the block group free space numbers
6889  * for our min num_bytes.  Another option is to have it go ahead
6890  * and look in the rbtree for a free extent of a given size, but this
6891  * is a good start.
6892  *
6893  * Callers of this must check if cache->cached == BTRFS_CACHE_ERROR before using
6894  * any of the information in this block group.
6895  */
6896 static noinline void
6897 wait_block_group_cache_progress(struct btrfs_block_group_cache *cache,
6898                                 u64 num_bytes)
6899 {
6900         struct btrfs_caching_control *caching_ctl;
6901
6902         caching_ctl = get_caching_control(cache);
6903         if (!caching_ctl)
6904                 return;
6905
6906         wait_event(caching_ctl->wait, block_group_cache_done(cache) ||
6907                    (cache->free_space_ctl->free_space >= num_bytes));
6908
6909         put_caching_control(caching_ctl);
6910 }
6911
6912 static noinline int
6913 wait_block_group_cache_done(struct btrfs_block_group_cache *cache)
6914 {
6915         struct btrfs_caching_control *caching_ctl;
6916         int ret = 0;
6917
6918         caching_ctl = get_caching_control(cache);
6919         if (!caching_ctl)
6920                 return (cache->cached == BTRFS_CACHE_ERROR) ? -EIO : 0;
6921
6922         wait_event(caching_ctl->wait, block_group_cache_done(cache));
6923         if (cache->cached == BTRFS_CACHE_ERROR)
6924                 ret = -EIO;
6925         put_caching_control(caching_ctl);
6926         return ret;
6927 }
6928
6929 int __get_raid_index(u64 flags)
6930 {
6931         if (flags & BTRFS_BLOCK_GROUP_RAID10)
6932                 return BTRFS_RAID_RAID10;
6933         else if (flags & BTRFS_BLOCK_GROUP_RAID1)
6934                 return BTRFS_RAID_RAID1;
6935         else if (flags & BTRFS_BLOCK_GROUP_DUP)
6936                 return BTRFS_RAID_DUP;
6937         else if (flags & BTRFS_BLOCK_GROUP_RAID0)
6938                 return BTRFS_RAID_RAID0;
6939         else if (flags & BTRFS_BLOCK_GROUP_RAID5)
6940                 return BTRFS_RAID_RAID5;
6941         else if (flags & BTRFS_BLOCK_GROUP_RAID6)
6942                 return BTRFS_RAID_RAID6;
6943
6944         return BTRFS_RAID_SINGLE; /* BTRFS_BLOCK_GROUP_SINGLE */
6945 }
6946
6947 int get_block_group_index(struct btrfs_block_group_cache *cache)
6948 {
6949         return __get_raid_index(cache->flags);
6950 }
6951
6952 static const char *btrfs_raid_type_names[BTRFS_NR_RAID_TYPES] = {
6953         [BTRFS_RAID_RAID10]     = "raid10",
6954         [BTRFS_RAID_RAID1]      = "raid1",
6955         [BTRFS_RAID_DUP]        = "dup",
6956         [BTRFS_RAID_RAID0]      = "raid0",
6957         [BTRFS_RAID_SINGLE]     = "single",
6958         [BTRFS_RAID_RAID5]      = "raid5",
6959         [BTRFS_RAID_RAID6]      = "raid6",
6960 };
6961
6962 static const char *get_raid_name(enum btrfs_raid_types type)
6963 {
6964         if (type >= BTRFS_NR_RAID_TYPES)
6965                 return NULL;
6966
6967         return btrfs_raid_type_names[type];
6968 }
6969
6970 enum btrfs_loop_type {
6971         LOOP_CACHING_NOWAIT = 0,
6972         LOOP_CACHING_WAIT = 1,
6973         LOOP_ALLOC_CHUNK = 2,
6974         LOOP_NO_EMPTY_SIZE = 3,
6975 };
6976
6977 static inline void
6978 btrfs_lock_block_group(struct btrfs_block_group_cache *cache,
6979                        int delalloc)
6980 {
6981         if (delalloc)
6982                 down_read(&cache->data_rwsem);
6983 }
6984
6985 static inline void
6986 btrfs_grab_block_group(struct btrfs_block_group_cache *cache,
6987                        int delalloc)
6988 {
6989         btrfs_get_block_group(cache);
6990         if (delalloc)
6991                 down_read(&cache->data_rwsem);
6992 }
6993
6994 static struct btrfs_block_group_cache *
6995 btrfs_lock_cluster(struct btrfs_block_group_cache *block_group,
6996                    struct btrfs_free_cluster *cluster,
6997                    int delalloc)
6998 {
6999         struct btrfs_block_group_cache *used_bg;
7000         bool locked = false;
7001 again:
7002         spin_lock(&cluster->refill_lock);
7003         if (locked) {
7004                 if (used_bg == cluster->block_group)
7005                         return used_bg;
7006
7007                 up_read(&used_bg->data_rwsem);
7008                 btrfs_put_block_group(used_bg);
7009         }
7010
7011         used_bg = cluster->block_group;
7012         if (!used_bg)
7013                 return NULL;
7014
7015         if (used_bg == block_group)
7016                 return used_bg;
7017
7018         btrfs_get_block_group(used_bg);
7019
7020         if (!delalloc)
7021                 return used_bg;
7022
7023         if (down_read_trylock(&used_bg->data_rwsem))
7024                 return used_bg;
7025
7026         spin_unlock(&cluster->refill_lock);
7027         down_read(&used_bg->data_rwsem);
7028         locked = true;
7029         goto again;
7030 }
7031
7032 static inline void
7033 btrfs_release_block_group(struct btrfs_block_group_cache *cache,
7034                          int delalloc)
7035 {
7036         if (delalloc)
7037                 up_read(&cache->data_rwsem);
7038         btrfs_put_block_group(cache);
7039 }
7040
7041 /*
7042  * walks the btree of allocated extents and find a hole of a given size.
7043  * The key ins is changed to record the hole:
7044  * ins->objectid == start position
7045  * ins->flags = BTRFS_EXTENT_ITEM_KEY
7046  * ins->offset == the size of the hole.
7047  * Any available blocks before search_start are skipped.
7048  *
7049  * If there is no suitable free space, we will record the max size of
7050  * the free space extent currently.
7051  */
7052 static noinline int find_free_extent(struct btrfs_root *orig_root,
7053                                      u64 num_bytes, u64 empty_size,
7054                                      u64 hint_byte, struct btrfs_key *ins,
7055                                      u64 flags, int delalloc)
7056 {
7057         int ret = 0;
7058         struct btrfs_root *root = orig_root->fs_info->extent_root;
7059         struct btrfs_free_cluster *last_ptr = NULL;
7060         struct btrfs_block_group_cache *block_group = NULL;
7061         u64 search_start = 0;
7062         u64 max_extent_size = 0;
7063         u64 empty_cluster = 0;
7064         struct btrfs_space_info *space_info;
7065         int loop = 0;
7066         int index = __get_raid_index(flags);
7067         int alloc_type = (flags & BTRFS_BLOCK_GROUP_DATA) ?
7068                 RESERVE_ALLOC_NO_ACCOUNT : RESERVE_ALLOC;
7069         bool failed_cluster_refill = false;
7070         bool failed_alloc = false;
7071         bool use_cluster = true;
7072         bool have_caching_bg = false;
7073         bool orig_have_caching_bg = false;
7074         bool full_search = false;
7075
7076         WARN_ON(num_bytes < root->sectorsize);
7077         ins->type = BTRFS_EXTENT_ITEM_KEY;
7078         ins->objectid = 0;
7079         ins->offset = 0;
7080
7081         trace_find_free_extent(orig_root, num_bytes, empty_size, flags);
7082
7083         space_info = __find_space_info(root->fs_info, flags);
7084         if (!space_info) {
7085                 btrfs_err(root->fs_info, "No space info for %llu", flags);
7086                 return -ENOSPC;
7087         }
7088
7089         /*
7090          * If our free space is heavily fragmented we may not be able to make
7091          * big contiguous allocations, so instead of doing the expensive search
7092          * for free space, simply return ENOSPC with our max_extent_size so we
7093          * can go ahead and search for a more manageable chunk.
7094          *
7095          * If our max_extent_size is large enough for our allocation simply
7096          * disable clustering since we will likely not be able to find enough
7097          * space to create a cluster and induce latency trying.
7098          */
7099         if (unlikely(space_info->max_extent_size)) {
7100                 spin_lock(&space_info->lock);
7101                 if (space_info->max_extent_size &&
7102                     num_bytes > space_info->max_extent_size) {
7103                         ins->offset = space_info->max_extent_size;
7104                         spin_unlock(&space_info->lock);
7105                         return -ENOSPC;
7106                 } else if (space_info->max_extent_size) {
7107                         use_cluster = false;
7108                 }
7109                 spin_unlock(&space_info->lock);
7110         }
7111
7112         last_ptr = fetch_cluster_info(orig_root, space_info, &empty_cluster);
7113         if (last_ptr) {
7114                 spin_lock(&last_ptr->lock);
7115                 if (last_ptr->block_group)
7116                         hint_byte = last_ptr->window_start;
7117                 if (last_ptr->fragmented) {
7118                         /*
7119                          * We still set window_start so we can keep track of the
7120                          * last place we found an allocation to try and save
7121                          * some time.
7122                          */
7123                         hint_byte = last_ptr->window_start;
7124                         use_cluster = false;
7125                 }
7126                 spin_unlock(&last_ptr->lock);
7127         }
7128
7129         search_start = max(search_start, first_logical_byte(root, 0));
7130         search_start = max(search_start, hint_byte);
7131         if (search_start == hint_byte) {
7132                 block_group = btrfs_lookup_block_group(root->fs_info,
7133                                                        search_start);
7134                 /*
7135                  * we don't want to use the block group if it doesn't match our
7136                  * allocation bits, or if its not cached.
7137                  *
7138                  * However if we are re-searching with an ideal block group
7139                  * picked out then we don't care that the block group is cached.
7140                  */
7141                 if (block_group && block_group_bits(block_group, flags) &&
7142                     block_group->cached != BTRFS_CACHE_NO) {
7143                         down_read(&space_info->groups_sem);
7144                         if (list_empty(&block_group->list) ||
7145                             block_group->ro) {
7146                                 /*
7147                                  * someone is removing this block group,
7148                                  * we can't jump into the have_block_group
7149                                  * target because our list pointers are not
7150                                  * valid
7151                                  */
7152                                 btrfs_put_block_group(block_group);
7153                                 up_read(&space_info->groups_sem);
7154                         } else {
7155                                 index = get_block_group_index(block_group);
7156                                 btrfs_lock_block_group(block_group, delalloc);
7157                                 goto have_block_group;
7158                         }
7159                 } else if (block_group) {
7160                         btrfs_put_block_group(block_group);
7161                 }
7162         }
7163 search:
7164         have_caching_bg = false;
7165         if (index == 0 || index == __get_raid_index(flags))
7166                 full_search = true;
7167         down_read(&space_info->groups_sem);
7168         list_for_each_entry(block_group, &space_info->block_groups[index],
7169                             list) {
7170                 u64 offset;
7171                 int cached;
7172
7173                 btrfs_grab_block_group(block_group, delalloc);
7174                 search_start = block_group->key.objectid;
7175
7176                 /*
7177                  * this can happen if we end up cycling through all the
7178                  * raid types, but we want to make sure we only allocate
7179                  * for the proper type.
7180                  */
7181                 if (!block_group_bits(block_group, flags)) {
7182                     u64 extra = BTRFS_BLOCK_GROUP_DUP |
7183                                 BTRFS_BLOCK_GROUP_RAID1 |
7184                                 BTRFS_BLOCK_GROUP_RAID5 |
7185                                 BTRFS_BLOCK_GROUP_RAID6 |
7186                                 BTRFS_BLOCK_GROUP_RAID10;
7187
7188                         /*
7189                          * if they asked for extra copies and this block group
7190                          * doesn't provide them, bail.  This does allow us to
7191                          * fill raid0 from raid1.
7192                          */
7193                         if ((flags & extra) && !(block_group->flags & extra))
7194                                 goto loop;
7195                 }
7196
7197 have_block_group:
7198                 cached = block_group_cache_done(block_group);
7199                 if (unlikely(!cached)) {
7200                         have_caching_bg = true;
7201                         ret = cache_block_group(block_group, 0);
7202                         BUG_ON(ret < 0);
7203                         ret = 0;
7204                 }
7205
7206                 if (unlikely(block_group->cached == BTRFS_CACHE_ERROR))
7207                         goto loop;
7208                 if (unlikely(block_group->ro))
7209                         goto loop;
7210
7211                 /*
7212                  * Ok we want to try and use the cluster allocator, so
7213                  * lets look there
7214                  */
7215                 if (last_ptr && use_cluster) {
7216                         struct btrfs_block_group_cache *used_block_group;
7217                         unsigned long aligned_cluster;
7218                         /*
7219                          * the refill lock keeps out other
7220                          * people trying to start a new cluster
7221                          */
7222                         used_block_group = btrfs_lock_cluster(block_group,
7223                                                               last_ptr,
7224                                                               delalloc);
7225                         if (!used_block_group)
7226                                 goto refill_cluster;
7227
7228                         if (used_block_group != block_group &&
7229                             (used_block_group->ro ||
7230                              !block_group_bits(used_block_group, flags)))
7231                                 goto release_cluster;
7232
7233                         offset = btrfs_alloc_from_cluster(used_block_group,
7234                                                 last_ptr,
7235                                                 num_bytes,
7236                                                 used_block_group->key.objectid,
7237                                                 &max_extent_size);
7238                         if (offset) {
7239                                 /* we have a block, we're done */
7240                                 spin_unlock(&last_ptr->refill_lock);
7241                                 trace_btrfs_reserve_extent_cluster(root,
7242                                                 used_block_group,
7243                                                 search_start, num_bytes);
7244                                 if (used_block_group != block_group) {
7245                                         btrfs_release_block_group(block_group,
7246                                                                   delalloc);
7247                                         block_group = used_block_group;
7248                                 }
7249                                 goto checks;
7250                         }
7251
7252                         WARN_ON(last_ptr->block_group != used_block_group);
7253 release_cluster:
7254                         /* If we are on LOOP_NO_EMPTY_SIZE, we can't
7255                          * set up a new clusters, so lets just skip it
7256                          * and let the allocator find whatever block
7257                          * it can find.  If we reach this point, we
7258                          * will have tried the cluster allocator
7259                          * plenty of times and not have found
7260                          * anything, so we are likely way too
7261                          * fragmented for the clustering stuff to find
7262                          * anything.
7263                          *
7264                          * However, if the cluster is taken from the
7265                          * current block group, release the cluster
7266                          * first, so that we stand a better chance of
7267                          * succeeding in the unclustered
7268                          * allocation.  */
7269                         if (loop >= LOOP_NO_EMPTY_SIZE &&
7270                             used_block_group != block_group) {
7271                                 spin_unlock(&last_ptr->refill_lock);
7272                                 btrfs_release_block_group(used_block_group,
7273                                                           delalloc);
7274                                 goto unclustered_alloc;
7275                         }
7276
7277                         /*
7278                          * this cluster didn't work out, free it and
7279                          * start over
7280                          */
7281                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7282
7283                         if (used_block_group != block_group)
7284                                 btrfs_release_block_group(used_block_group,
7285                                                           delalloc);
7286 refill_cluster:
7287                         if (loop >= LOOP_NO_EMPTY_SIZE) {
7288                                 spin_unlock(&last_ptr->refill_lock);
7289                                 goto unclustered_alloc;
7290                         }
7291
7292                         aligned_cluster = max_t(unsigned long,
7293                                                 empty_cluster + empty_size,
7294                                               block_group->full_stripe_len);
7295
7296                         /* allocate a cluster in this block group */
7297                         ret = btrfs_find_space_cluster(root, block_group,
7298                                                        last_ptr, search_start,
7299                                                        num_bytes,
7300                                                        aligned_cluster);
7301                         if (ret == 0) {
7302                                 /*
7303                                  * now pull our allocation out of this
7304                                  * cluster
7305                                  */
7306                                 offset = btrfs_alloc_from_cluster(block_group,
7307                                                         last_ptr,
7308                                                         num_bytes,
7309                                                         search_start,
7310                                                         &max_extent_size);
7311                                 if (offset) {
7312                                         /* we found one, proceed */
7313                                         spin_unlock(&last_ptr->refill_lock);
7314                                         trace_btrfs_reserve_extent_cluster(root,
7315                                                 block_group, search_start,
7316                                                 num_bytes);
7317                                         goto checks;
7318                                 }
7319                         } else if (!cached && loop > LOOP_CACHING_NOWAIT
7320                                    && !failed_cluster_refill) {
7321                                 spin_unlock(&last_ptr->refill_lock);
7322
7323                                 failed_cluster_refill = true;
7324                                 wait_block_group_cache_progress(block_group,
7325                                        num_bytes + empty_cluster + empty_size);
7326                                 goto have_block_group;
7327                         }
7328
7329                         /*
7330                          * at this point we either didn't find a cluster
7331                          * or we weren't able to allocate a block from our
7332                          * cluster.  Free the cluster we've been trying
7333                          * to use, and go to the next block group
7334                          */
7335                         btrfs_return_cluster_to_free_space(NULL, last_ptr);
7336                         spin_unlock(&last_ptr->refill_lock);
7337                         goto loop;
7338                 }
7339
7340 unclustered_alloc:
7341                 /*
7342                  * We are doing an unclustered alloc, set the fragmented flag so
7343                  * we don't bother trying to setup a cluster again until we get
7344                  * more space.
7345                  */
7346                 if (unlikely(last_ptr)) {
7347                         spin_lock(&last_ptr->lock);
7348                         last_ptr->fragmented = 1;
7349                         spin_unlock(&last_ptr->lock);
7350                 }
7351                 spin_lock(&block_group->free_space_ctl->tree_lock);
7352                 if (cached &&
7353                     block_group->free_space_ctl->free_space <
7354                     num_bytes + empty_cluster + empty_size) {
7355                         if (block_group->free_space_ctl->free_space >
7356                             max_extent_size)
7357                                 max_extent_size =
7358                                         block_group->free_space_ctl->free_space;
7359                         spin_unlock(&block_group->free_space_ctl->tree_lock);
7360                         goto loop;
7361                 }
7362                 spin_unlock(&block_group->free_space_ctl->tree_lock);
7363
7364                 offset = btrfs_find_space_for_alloc(block_group, search_start,
7365                                                     num_bytes, empty_size,
7366                                                     &max_extent_size);
7367                 /*
7368                  * If we didn't find a chunk, and we haven't failed on this
7369                  * block group before, and this block group is in the middle of
7370                  * caching and we are ok with waiting, then go ahead and wait
7371                  * for progress to be made, and set failed_alloc to true.
7372                  *
7373                  * If failed_alloc is true then we've already waited on this
7374                  * block group once and should move on to the next block group.
7375                  */
7376                 if (!offset && !failed_alloc && !cached &&
7377                     loop > LOOP_CACHING_NOWAIT) {
7378                         wait_block_group_cache_progress(block_group,
7379                                                 num_bytes + empty_size);
7380                         failed_alloc = true;
7381                         goto have_block_group;
7382                 } else if (!offset) {
7383                         goto loop;
7384                 }
7385 checks:
7386                 search_start = ALIGN(offset, root->stripesize);
7387
7388                 /* move on to the next group */
7389                 if (search_start + num_bytes >
7390                     block_group->key.objectid + block_group->key.offset) {
7391                         btrfs_add_free_space(block_group, offset, num_bytes);
7392                         goto loop;
7393                 }
7394
7395                 if (offset < search_start)
7396                         btrfs_add_free_space(block_group, offset,
7397                                              search_start - offset);
7398                 BUG_ON(offset > search_start);
7399
7400                 ret = btrfs_update_reserved_bytes(block_group, num_bytes,
7401                                                   alloc_type, delalloc);
7402                 if (ret == -EAGAIN) {
7403                         btrfs_add_free_space(block_group, offset, num_bytes);
7404                         goto loop;
7405                 }
7406
7407                 /* we are all good, lets return */
7408                 ins->objectid = search_start;
7409                 ins->offset = num_bytes;
7410
7411                 trace_btrfs_reserve_extent(orig_root, block_group,
7412                                            search_start, num_bytes);
7413                 btrfs_release_block_group(block_group, delalloc);
7414                 break;
7415 loop:
7416                 failed_cluster_refill = false;
7417                 failed_alloc = false;
7418                 BUG_ON(index != get_block_group_index(block_group));
7419                 btrfs_release_block_group(block_group, delalloc);
7420         }
7421         up_read(&space_info->groups_sem);
7422
7423         if ((loop == LOOP_CACHING_NOWAIT) && have_caching_bg
7424                 && !orig_have_caching_bg)
7425                 orig_have_caching_bg = true;
7426
7427         if (!ins->objectid && loop >= LOOP_CACHING_WAIT && have_caching_bg)
7428                 goto search;
7429
7430         if (!ins->objectid && ++index < BTRFS_NR_RAID_TYPES)
7431                 goto search;
7432
7433         /*
7434          * LOOP_CACHING_NOWAIT, search partially cached block groups, kicking
7435          *                      caching kthreads as we move along
7436          * LOOP_CACHING_WAIT, search everything, and wait if our bg is caching
7437          * LOOP_ALLOC_CHUNK, force a chunk allocation and try again
7438          * LOOP_NO_EMPTY_SIZE, set empty_size and empty_cluster to 0 and try
7439          *                      again
7440          */
7441         if (!ins->objectid && loop < LOOP_NO_EMPTY_SIZE) {
7442                 index = 0;
7443                 if (loop == LOOP_CACHING_NOWAIT) {
7444                         /*
7445                          * We want to skip the LOOP_CACHING_WAIT step if we
7446                          * don't have any unached bgs and we've alrelady done a
7447                          * full search through.
7448                          */
7449                         if (orig_have_caching_bg || !full_search)
7450                                 loop = LOOP_CACHING_WAIT;
7451                         else
7452                                 loop = LOOP_ALLOC_CHUNK;
7453                 } else {
7454                         loop++;
7455                 }
7456
7457                 if (loop == LOOP_ALLOC_CHUNK) {
7458                         struct btrfs_trans_handle *trans;
7459                         int exist = 0;
7460
7461                         trans = current->journal_info;
7462                         if (trans)
7463                                 exist = 1;
7464                         else
7465                                 trans = btrfs_join_transaction(root);
7466
7467                         if (IS_ERR(trans)) {
7468                                 ret = PTR_ERR(trans);
7469                                 goto out;
7470                         }
7471
7472                         ret = do_chunk_alloc(trans, root, flags,
7473                                              CHUNK_ALLOC_FORCE);
7474
7475                         /*
7476                          * If we can't allocate a new chunk we've already looped
7477                          * through at least once, move on to the NO_EMPTY_SIZE
7478                          * case.
7479                          */
7480                         if (ret == -ENOSPC)
7481                                 loop = LOOP_NO_EMPTY_SIZE;
7482
7483                         /*
7484                          * Do not bail out on ENOSPC since we
7485                          * can do more things.
7486                          */
7487                         if (ret < 0 && ret != -ENOSPC)
7488                                 btrfs_abort_transaction(trans,
7489                                                         root, ret);
7490                         else
7491                                 ret = 0;
7492                         if (!exist)
7493                                 btrfs_end_transaction(trans, root);
7494                         if (ret)
7495                                 goto out;
7496                 }
7497
7498                 if (loop == LOOP_NO_EMPTY_SIZE) {
7499                         /*
7500                          * Don't loop again if we already have no empty_size and
7501                          * no empty_cluster.
7502                          */
7503                         if (empty_size == 0 &&
7504                             empty_cluster == 0) {
7505                                 ret = -ENOSPC;
7506                                 goto out;
7507                         }
7508                         empty_size = 0;
7509                         empty_cluster = 0;
7510                 }
7511
7512                 goto search;
7513         } else if (!ins->objectid) {
7514                 ret = -ENOSPC;
7515         } else if (ins->objectid) {
7516                 if (!use_cluster && last_ptr) {
7517                         spin_lock(&last_ptr->lock);
7518                         last_ptr->window_start = ins->objectid;
7519                         spin_unlock(&last_ptr->lock);
7520                 }
7521                 ret = 0;
7522         }
7523 out:
7524         if (ret == -ENOSPC) {
7525                 spin_lock(&space_info->lock);
7526                 space_info->max_extent_size = max_extent_size;
7527                 spin_unlock(&space_info->lock);
7528                 ins->offset = max_extent_size;
7529         }
7530         return ret;
7531 }
7532
7533 static void dump_space_info(struct btrfs_space_info *info, u64 bytes,
7534                             int dump_block_groups)
7535 {
7536         struct btrfs_block_group_cache *cache;
7537         int index = 0;
7538
7539         spin_lock(&info->lock);
7540         printk(KERN_INFO "BTRFS: space_info %llu has %llu free, is %sfull\n",
7541                info->flags,
7542                info->total_bytes - info->bytes_used - info->bytes_pinned -
7543                info->bytes_reserved - info->bytes_readonly,
7544                (info->full) ? "" : "not ");
7545         printk(KERN_INFO "BTRFS: space_info total=%llu, used=%llu, pinned=%llu, "
7546                "reserved=%llu, may_use=%llu, readonly=%llu\n",
7547                info->total_bytes, info->bytes_used, info->bytes_pinned,
7548                info->bytes_reserved, info->bytes_may_use,
7549                info->bytes_readonly);
7550         spin_unlock(&info->lock);
7551
7552         if (!dump_block_groups)
7553                 return;
7554
7555         down_read(&info->groups_sem);
7556 again:
7557         list_for_each_entry(cache, &info->block_groups[index], list) {
7558                 spin_lock(&cache->lock);
7559                 printk(KERN_INFO "BTRFS: "
7560                            "block group %llu has %llu bytes, "
7561                            "%llu used %llu pinned %llu reserved %s\n",
7562                        cache->key.objectid, cache->key.offset,
7563                        btrfs_block_group_used(&cache->item), cache->pinned,
7564                        cache->reserved, cache->ro ? "[readonly]" : "");
7565                 btrfs_dump_free_space(cache, bytes);
7566                 spin_unlock(&cache->lock);
7567         }
7568         if (++index < BTRFS_NR_RAID_TYPES)
7569                 goto again;
7570         up_read(&info->groups_sem);
7571 }
7572
7573 int btrfs_reserve_extent(struct btrfs_root *root,
7574                          u64 num_bytes, u64 min_alloc_size,
7575                          u64 empty_size, u64 hint_byte,
7576                          struct btrfs_key *ins, int is_data, int delalloc)
7577 {
7578         bool final_tried = num_bytes == min_alloc_size;
7579         u64 flags;
7580         int ret;
7581
7582         flags = btrfs_get_alloc_profile(root, is_data);
7583 again:
7584         WARN_ON(num_bytes < root->sectorsize);
7585         ret = find_free_extent(root, num_bytes, empty_size, hint_byte, ins,
7586                                flags, delalloc);
7587
7588         if (ret == -ENOSPC) {
7589                 if (!final_tried && ins->offset) {
7590                         num_bytes = min(num_bytes >> 1, ins->offset);
7591                         num_bytes = round_down(num_bytes, root->sectorsize);
7592                         num_bytes = max(num_bytes, min_alloc_size);
7593                         if (num_bytes == min_alloc_size)
7594                                 final_tried = true;
7595                         goto again;
7596                 } else if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7597                         struct btrfs_space_info *sinfo;
7598
7599                         sinfo = __find_space_info(root->fs_info, flags);
7600                         btrfs_err(root->fs_info, "allocation failed flags %llu, wanted %llu",
7601                                 flags, num_bytes);
7602                         if (sinfo)
7603                                 dump_space_info(sinfo, num_bytes, 1);
7604                 }
7605         }
7606
7607         return ret;
7608 }
7609
7610 static int __btrfs_free_reserved_extent(struct btrfs_root *root,
7611                                         u64 start, u64 len,
7612                                         int pin, int delalloc)
7613 {
7614         struct btrfs_block_group_cache *cache;
7615         int ret = 0;
7616
7617         cache = btrfs_lookup_block_group(root->fs_info, start);
7618         if (!cache) {
7619                 btrfs_err(root->fs_info, "Unable to find block group for %llu",
7620                         start);
7621                 return -ENOSPC;
7622         }
7623
7624         if (pin)
7625                 pin_down_extent(root, cache, start, len, 1);
7626         else {
7627                 if (btrfs_test_opt(root, DISCARD))
7628                         ret = btrfs_discard_extent(root, start, len, NULL);
7629                 btrfs_add_free_space(cache, start, len);
7630                 btrfs_update_reserved_bytes(cache, len, RESERVE_FREE, delalloc);
7631         }
7632
7633         btrfs_put_block_group(cache);
7634
7635         trace_btrfs_reserved_extent_free(root, start, len);
7636
7637         return ret;
7638 }
7639
7640 int btrfs_free_reserved_extent(struct btrfs_root *root,
7641                                u64 start, u64 len, int delalloc)
7642 {
7643         return __btrfs_free_reserved_extent(root, start, len, 0, delalloc);
7644 }
7645
7646 int btrfs_free_and_pin_reserved_extent(struct btrfs_root *root,
7647                                        u64 start, u64 len)
7648 {
7649         return __btrfs_free_reserved_extent(root, start, len, 1, 0);
7650 }
7651
7652 static int alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7653                                       struct btrfs_root *root,
7654                                       u64 parent, u64 root_objectid,
7655                                       u64 flags, u64 owner, u64 offset,
7656                                       struct btrfs_key *ins, int ref_mod)
7657 {
7658         int ret;
7659         struct btrfs_fs_info *fs_info = root->fs_info;
7660         struct btrfs_extent_item *extent_item;
7661         struct btrfs_extent_inline_ref *iref;
7662         struct btrfs_path *path;
7663         struct extent_buffer *leaf;
7664         int type;
7665         u32 size;
7666
7667         if (parent > 0)
7668                 type = BTRFS_SHARED_DATA_REF_KEY;
7669         else
7670                 type = BTRFS_EXTENT_DATA_REF_KEY;
7671
7672         size = sizeof(*extent_item) + btrfs_extent_inline_ref_size(type);
7673
7674         path = btrfs_alloc_path();
7675         if (!path)
7676                 return -ENOMEM;
7677
7678         path->leave_spinning = 1;
7679         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7680                                       ins, size);
7681         if (ret) {
7682                 btrfs_free_path(path);
7683                 return ret;
7684         }
7685
7686         leaf = path->nodes[0];
7687         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7688                                      struct btrfs_extent_item);
7689         btrfs_set_extent_refs(leaf, extent_item, ref_mod);
7690         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7691         btrfs_set_extent_flags(leaf, extent_item,
7692                                flags | BTRFS_EXTENT_FLAG_DATA);
7693
7694         iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7695         btrfs_set_extent_inline_ref_type(leaf, iref, type);
7696         if (parent > 0) {
7697                 struct btrfs_shared_data_ref *ref;
7698                 ref = (struct btrfs_shared_data_ref *)(iref + 1);
7699                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7700                 btrfs_set_shared_data_ref_count(leaf, ref, ref_mod);
7701         } else {
7702                 struct btrfs_extent_data_ref *ref;
7703                 ref = (struct btrfs_extent_data_ref *)(&iref->offset);
7704                 btrfs_set_extent_data_ref_root(leaf, ref, root_objectid);
7705                 btrfs_set_extent_data_ref_objectid(leaf, ref, owner);
7706                 btrfs_set_extent_data_ref_offset(leaf, ref, offset);
7707                 btrfs_set_extent_data_ref_count(leaf, ref, ref_mod);
7708         }
7709
7710         btrfs_mark_buffer_dirty(path->nodes[0]);
7711         btrfs_free_path(path);
7712
7713         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
7714                                           ins->offset);
7715         if (ret)
7716                 return ret;
7717
7718         ret = update_block_group(trans, root, ins->objectid, ins->offset, 1);
7719         if (ret) { /* -ENOENT, logic error */
7720                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7721                         ins->objectid, ins->offset);
7722                 BUG();
7723         }
7724         trace_btrfs_reserved_extent_alloc(root, ins->objectid, ins->offset);
7725         return ret;
7726 }
7727
7728 static int alloc_reserved_tree_block(struct btrfs_trans_handle *trans,
7729                                      struct btrfs_root *root,
7730                                      u64 parent, u64 root_objectid,
7731                                      u64 flags, struct btrfs_disk_key *key,
7732                                      int level, struct btrfs_key *ins)
7733 {
7734         int ret;
7735         struct btrfs_fs_info *fs_info = root->fs_info;
7736         struct btrfs_extent_item *extent_item;
7737         struct btrfs_tree_block_info *block_info;
7738         struct btrfs_extent_inline_ref *iref;
7739         struct btrfs_path *path;
7740         struct extent_buffer *leaf;
7741         u32 size = sizeof(*extent_item) + sizeof(*iref);
7742         u64 num_bytes = ins->offset;
7743         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7744                                                  SKINNY_METADATA);
7745
7746         if (!skinny_metadata)
7747                 size += sizeof(*block_info);
7748
7749         path = btrfs_alloc_path();
7750         if (!path) {
7751                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7752                                                    root->nodesize);
7753                 return -ENOMEM;
7754         }
7755
7756         path->leave_spinning = 1;
7757         ret = btrfs_insert_empty_item(trans, fs_info->extent_root, path,
7758                                       ins, size);
7759         if (ret) {
7760                 btrfs_free_path(path);
7761                 btrfs_free_and_pin_reserved_extent(root, ins->objectid,
7762                                                    root->nodesize);
7763                 return ret;
7764         }
7765
7766         leaf = path->nodes[0];
7767         extent_item = btrfs_item_ptr(leaf, path->slots[0],
7768                                      struct btrfs_extent_item);
7769         btrfs_set_extent_refs(leaf, extent_item, 1);
7770         btrfs_set_extent_generation(leaf, extent_item, trans->transid);
7771         btrfs_set_extent_flags(leaf, extent_item,
7772                                flags | BTRFS_EXTENT_FLAG_TREE_BLOCK);
7773
7774         if (skinny_metadata) {
7775                 iref = (struct btrfs_extent_inline_ref *)(extent_item + 1);
7776                 num_bytes = root->nodesize;
7777         } else {
7778                 block_info = (struct btrfs_tree_block_info *)(extent_item + 1);
7779                 btrfs_set_tree_block_key(leaf, block_info, key);
7780                 btrfs_set_tree_block_level(leaf, block_info, level);
7781                 iref = (struct btrfs_extent_inline_ref *)(block_info + 1);
7782         }
7783
7784         if (parent > 0) {
7785                 BUG_ON(!(flags & BTRFS_BLOCK_FLAG_FULL_BACKREF));
7786                 btrfs_set_extent_inline_ref_type(leaf, iref,
7787                                                  BTRFS_SHARED_BLOCK_REF_KEY);
7788                 btrfs_set_extent_inline_ref_offset(leaf, iref, parent);
7789         } else {
7790                 btrfs_set_extent_inline_ref_type(leaf, iref,
7791                                                  BTRFS_TREE_BLOCK_REF_KEY);
7792                 btrfs_set_extent_inline_ref_offset(leaf, iref, root_objectid);
7793         }
7794
7795         btrfs_mark_buffer_dirty(leaf);
7796         btrfs_free_path(path);
7797
7798         ret = remove_from_free_space_tree(trans, fs_info, ins->objectid,
7799                                           num_bytes);
7800         if (ret)
7801                 return ret;
7802
7803         ret = update_block_group(trans, root, ins->objectid, root->nodesize,
7804                                  1);
7805         if (ret) { /* -ENOENT, logic error */
7806                 btrfs_err(fs_info, "update block group failed for %llu %llu",
7807                         ins->objectid, ins->offset);
7808                 BUG();
7809         }
7810
7811         trace_btrfs_reserved_extent_alloc(root, ins->objectid, root->nodesize);
7812         return ret;
7813 }
7814
7815 int btrfs_alloc_reserved_file_extent(struct btrfs_trans_handle *trans,
7816                                      struct btrfs_root *root,
7817                                      u64 root_objectid, u64 owner,
7818                                      u64 offset, u64 ram_bytes,
7819                                      struct btrfs_key *ins)
7820 {
7821         int ret;
7822
7823         BUG_ON(root_objectid == BTRFS_TREE_LOG_OBJECTID);
7824
7825         ret = btrfs_add_delayed_data_ref(root->fs_info, trans, ins->objectid,
7826                                          ins->offset, 0,
7827                                          root_objectid, owner, offset,
7828                                          ram_bytes, BTRFS_ADD_DELAYED_EXTENT,
7829                                          NULL);
7830         return ret;
7831 }
7832
7833 /*
7834  * this is used by the tree logging recovery code.  It records that
7835  * an extent has been allocated and makes sure to clear the free
7836  * space cache bits as well
7837  */
7838 int btrfs_alloc_logged_file_extent(struct btrfs_trans_handle *trans,
7839                                    struct btrfs_root *root,
7840                                    u64 root_objectid, u64 owner, u64 offset,
7841                                    struct btrfs_key *ins)
7842 {
7843         int ret;
7844         struct btrfs_block_group_cache *block_group;
7845
7846         /*
7847          * Mixed block groups will exclude before processing the log so we only
7848          * need to do the exlude dance if this fs isn't mixed.
7849          */
7850         if (!btrfs_fs_incompat(root->fs_info, MIXED_GROUPS)) {
7851                 ret = __exclude_logged_extent(root, ins->objectid, ins->offset);
7852                 if (ret)
7853                         return ret;
7854         }
7855
7856         block_group = btrfs_lookup_block_group(root->fs_info, ins->objectid);
7857         if (!block_group)
7858                 return -EINVAL;
7859
7860         ret = btrfs_update_reserved_bytes(block_group, ins->offset,
7861                                           RESERVE_ALLOC_NO_ACCOUNT, 0);
7862         BUG_ON(ret); /* logic error */
7863         ret = alloc_reserved_file_extent(trans, root, 0, root_objectid,
7864                                          0, owner, offset, ins, 1);
7865         btrfs_put_block_group(block_group);
7866         return ret;
7867 }
7868
7869 static struct extent_buffer *
7870 btrfs_init_new_buffer(struct btrfs_trans_handle *trans, struct btrfs_root *root,
7871                       u64 bytenr, int level)
7872 {
7873         struct extent_buffer *buf;
7874
7875         buf = btrfs_find_create_tree_block(root, bytenr);
7876         if (!buf)
7877                 return ERR_PTR(-ENOMEM);
7878         btrfs_set_header_generation(buf, trans->transid);
7879         btrfs_set_buffer_lockdep_class(root->root_key.objectid, buf, level);
7880         btrfs_tree_lock(buf);
7881         clean_tree_block(trans, root->fs_info, buf);
7882         clear_bit(EXTENT_BUFFER_STALE, &buf->bflags);
7883
7884         btrfs_set_lock_blocking(buf);
7885         set_extent_buffer_uptodate(buf);
7886
7887         if (root->root_key.objectid == BTRFS_TREE_LOG_OBJECTID) {
7888                 buf->log_index = root->log_transid % 2;
7889                 /*
7890                  * we allow two log transactions at a time, use different
7891                  * EXENT bit to differentiate dirty pages.
7892                  */
7893                 if (buf->log_index == 0)
7894                         set_extent_dirty(&root->dirty_log_pages, buf->start,
7895                                         buf->start + buf->len - 1, GFP_NOFS);
7896                 else
7897                         set_extent_new(&root->dirty_log_pages, buf->start,
7898                                         buf->start + buf->len - 1, GFP_NOFS);
7899         } else {
7900                 buf->log_index = -1;
7901                 set_extent_dirty(&trans->transaction->dirty_pages, buf->start,
7902                          buf->start + buf->len - 1, GFP_NOFS);
7903         }
7904         trans->blocks_used++;
7905         /* this returns a buffer locked for blocking */
7906         return buf;
7907 }
7908
7909 static struct btrfs_block_rsv *
7910 use_block_rsv(struct btrfs_trans_handle *trans,
7911               struct btrfs_root *root, u32 blocksize)
7912 {
7913         struct btrfs_block_rsv *block_rsv;
7914         struct btrfs_block_rsv *global_rsv = &root->fs_info->global_block_rsv;
7915         int ret;
7916         bool global_updated = false;
7917
7918         block_rsv = get_block_rsv(trans, root);
7919
7920         if (unlikely(block_rsv->size == 0))
7921                 goto try_reserve;
7922 again:
7923         ret = block_rsv_use_bytes(block_rsv, blocksize);
7924         if (!ret)
7925                 return block_rsv;
7926
7927         if (block_rsv->failfast)
7928                 return ERR_PTR(ret);
7929
7930         if (block_rsv->type == BTRFS_BLOCK_RSV_GLOBAL && !global_updated) {
7931                 global_updated = true;
7932                 update_global_block_rsv(root->fs_info);
7933                 goto again;
7934         }
7935
7936         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
7937                 static DEFINE_RATELIMIT_STATE(_rs,
7938                                 DEFAULT_RATELIMIT_INTERVAL * 10,
7939                                 /*DEFAULT_RATELIMIT_BURST*/ 1);
7940                 if (__ratelimit(&_rs))
7941                         WARN(1, KERN_DEBUG
7942                                 "BTRFS: block rsv returned %d\n", ret);
7943         }
7944 try_reserve:
7945         ret = reserve_metadata_bytes(root, block_rsv, blocksize,
7946                                      BTRFS_RESERVE_NO_FLUSH);
7947         if (!ret)
7948                 return block_rsv;
7949         /*
7950          * If we couldn't reserve metadata bytes try and use some from
7951          * the global reserve if its space type is the same as the global
7952          * reservation.
7953          */
7954         if (block_rsv->type != BTRFS_BLOCK_RSV_GLOBAL &&
7955             block_rsv->space_info == global_rsv->space_info) {
7956                 ret = block_rsv_use_bytes(global_rsv, blocksize);
7957                 if (!ret)
7958                         return global_rsv;
7959         }
7960         return ERR_PTR(ret);
7961 }
7962
7963 static void unuse_block_rsv(struct btrfs_fs_info *fs_info,
7964                             struct btrfs_block_rsv *block_rsv, u32 blocksize)
7965 {
7966         block_rsv_add_bytes(block_rsv, blocksize, 0);
7967         block_rsv_release_bytes(fs_info, block_rsv, NULL, 0);
7968 }
7969
7970 /*
7971  * finds a free extent and does all the dirty work required for allocation
7972  * returns the tree buffer or an ERR_PTR on error.
7973  */
7974 struct extent_buffer *btrfs_alloc_tree_block(struct btrfs_trans_handle *trans,
7975                                         struct btrfs_root *root,
7976                                         u64 parent, u64 root_objectid,
7977                                         struct btrfs_disk_key *key, int level,
7978                                         u64 hint, u64 empty_size)
7979 {
7980         struct btrfs_key ins;
7981         struct btrfs_block_rsv *block_rsv;
7982         struct extent_buffer *buf;
7983         struct btrfs_delayed_extent_op *extent_op;
7984         u64 flags = 0;
7985         int ret;
7986         u32 blocksize = root->nodesize;
7987         bool skinny_metadata = btrfs_fs_incompat(root->fs_info,
7988                                                  SKINNY_METADATA);
7989
7990         if (btrfs_test_is_dummy_root(root)) {
7991                 buf = btrfs_init_new_buffer(trans, root, root->alloc_bytenr,
7992                                             level);
7993                 if (!IS_ERR(buf))
7994                         root->alloc_bytenr += blocksize;
7995                 return buf;
7996         }
7997
7998         block_rsv = use_block_rsv(trans, root, blocksize);
7999         if (IS_ERR(block_rsv))
8000                 return ERR_CAST(block_rsv);
8001
8002         ret = btrfs_reserve_extent(root, blocksize, blocksize,
8003                                    empty_size, hint, &ins, 0, 0);
8004         if (ret)
8005                 goto out_unuse;
8006
8007         buf = btrfs_init_new_buffer(trans, root, ins.objectid, level);
8008         if (IS_ERR(buf)) {
8009                 ret = PTR_ERR(buf);
8010                 goto out_free_reserved;
8011         }
8012
8013         if (root_objectid == BTRFS_TREE_RELOC_OBJECTID) {
8014                 if (parent == 0)
8015                         parent = ins.objectid;
8016                 flags |= BTRFS_BLOCK_FLAG_FULL_BACKREF;
8017         } else
8018                 BUG_ON(parent > 0);
8019
8020         if (root_objectid != BTRFS_TREE_LOG_OBJECTID) {
8021                 extent_op = btrfs_alloc_delayed_extent_op();
8022                 if (!extent_op) {
8023                         ret = -ENOMEM;
8024                         goto out_free_buf;
8025                 }
8026                 if (key)
8027                         memcpy(&extent_op->key, key, sizeof(extent_op->key));
8028                 else
8029                         memset(&extent_op->key, 0, sizeof(extent_op->key));
8030                 extent_op->flags_to_set = flags;
8031                 if (skinny_metadata)
8032                         extent_op->update_key = 0;
8033                 else
8034                         extent_op->update_key = 1;
8035                 extent_op->update_flags = 1;
8036                 extent_op->is_data = 0;
8037                 extent_op->level = level;
8038
8039                 ret = btrfs_add_delayed_tree_ref(root->fs_info, trans,
8040                                                  ins.objectid, ins.offset,
8041                                                  parent, root_objectid, level,
8042                                                  BTRFS_ADD_DELAYED_EXTENT,
8043                                                  extent_op);
8044                 if (ret)
8045                         goto out_free_delayed;
8046         }
8047         return buf;
8048
8049 out_free_delayed:
8050         btrfs_free_delayed_extent_op(extent_op);
8051 out_free_buf:
8052         free_extent_buffer(buf);
8053 out_free_reserved:
8054         btrfs_free_reserved_extent(root, ins.objectid, ins.offset, 0);
8055 out_unuse:
8056         unuse_block_rsv(root->fs_info, block_rsv, blocksize);
8057         return ERR_PTR(ret);
8058 }
8059
8060 struct walk_control {
8061         u64 refs[BTRFS_MAX_LEVEL];
8062         u64 flags[BTRFS_MAX_LEVEL];
8063         struct btrfs_key update_progress;
8064         int stage;
8065         int level;
8066         int shared_level;
8067         int update_ref;
8068         int keep_locks;
8069         int reada_slot;
8070         int reada_count;
8071         int for_reloc;
8072 };
8073
8074 #define DROP_REFERENCE  1
8075 #define UPDATE_BACKREF  2
8076
8077 static noinline void reada_walk_down(struct btrfs_trans_handle *trans,
8078                                      struct btrfs_root *root,
8079                                      struct walk_control *wc,
8080                                      struct btrfs_path *path)
8081 {
8082         u64 bytenr;
8083         u64 generation;
8084         u64 refs;
8085         u64 flags;
8086         u32 nritems;
8087         u32 blocksize;
8088         struct btrfs_key key;
8089         struct extent_buffer *eb;
8090         int ret;
8091         int slot;
8092         int nread = 0;
8093
8094         if (path->slots[wc->level] < wc->reada_slot) {
8095                 wc->reada_count = wc->reada_count * 2 / 3;
8096                 wc->reada_count = max(wc->reada_count, 2);
8097         } else {
8098                 wc->reada_count = wc->reada_count * 3 / 2;
8099                 wc->reada_count = min_t(int, wc->reada_count,
8100                                         BTRFS_NODEPTRS_PER_BLOCK(root));
8101         }
8102
8103         eb = path->nodes[wc->level];
8104         nritems = btrfs_header_nritems(eb);
8105         blocksize = root->nodesize;
8106
8107         for (slot = path->slots[wc->level]; slot < nritems; slot++) {
8108                 if (nread >= wc->reada_count)
8109                         break;
8110
8111                 cond_resched();
8112                 bytenr = btrfs_node_blockptr(eb, slot);
8113                 generation = btrfs_node_ptr_generation(eb, slot);
8114
8115                 if (slot == path->slots[wc->level])
8116                         goto reada;
8117
8118                 if (wc->stage == UPDATE_BACKREF &&
8119                     generation <= root->root_key.offset)
8120                         continue;
8121
8122                 /* We don't lock the tree block, it's OK to be racy here */
8123                 ret = btrfs_lookup_extent_info(trans, root, bytenr,
8124                                                wc->level - 1, 1, &refs,
8125                                                &flags);
8126                 /* We don't care about errors in readahead. */
8127                 if (ret < 0)
8128                         continue;
8129                 BUG_ON(refs == 0);
8130
8131                 if (wc->stage == DROP_REFERENCE) {
8132                         if (refs == 1)
8133                                 goto reada;
8134
8135                         if (wc->level == 1 &&
8136                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8137                                 continue;
8138                         if (!wc->update_ref ||
8139                             generation <= root->root_key.offset)
8140                                 continue;
8141                         btrfs_node_key_to_cpu(eb, &key, slot);
8142                         ret = btrfs_comp_cpu_keys(&key,
8143                                                   &wc->update_progress);
8144                         if (ret < 0)
8145                                 continue;
8146                 } else {
8147                         if (wc->level == 1 &&
8148                             (flags & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8149                                 continue;
8150                 }
8151 reada:
8152                 readahead_tree_block(root, bytenr);
8153                 nread++;
8154         }
8155         wc->reada_slot = slot;
8156 }
8157
8158 /*
8159  * These may not be seen by the usual inc/dec ref code so we have to
8160  * add them here.
8161  */
8162 static int record_one_subtree_extent(struct btrfs_trans_handle *trans,
8163                                      struct btrfs_root *root, u64 bytenr,
8164                                      u64 num_bytes)
8165 {
8166         struct btrfs_qgroup_extent_record *qrecord;
8167         struct btrfs_delayed_ref_root *delayed_refs;
8168
8169         qrecord = kmalloc(sizeof(*qrecord), GFP_NOFS);
8170         if (!qrecord)
8171                 return -ENOMEM;
8172
8173         qrecord->bytenr = bytenr;
8174         qrecord->num_bytes = num_bytes;
8175         qrecord->old_roots = NULL;
8176
8177         delayed_refs = &trans->transaction->delayed_refs;
8178         spin_lock(&delayed_refs->lock);
8179         if (btrfs_qgroup_insert_dirty_extent(delayed_refs, qrecord))
8180                 kfree(qrecord);
8181         spin_unlock(&delayed_refs->lock);
8182
8183         return 0;
8184 }
8185
8186 static int account_leaf_items(struct btrfs_trans_handle *trans,
8187                               struct btrfs_root *root,
8188                               struct extent_buffer *eb)
8189 {
8190         int nr = btrfs_header_nritems(eb);
8191         int i, extent_type, ret;
8192         struct btrfs_key key;
8193         struct btrfs_file_extent_item *fi;
8194         u64 bytenr, num_bytes;
8195
8196         /* We can be called directly from walk_up_proc() */
8197         if (!root->fs_info->quota_enabled)
8198                 return 0;
8199
8200         for (i = 0; i < nr; i++) {
8201                 btrfs_item_key_to_cpu(eb, &key, i);
8202
8203                 if (key.type != BTRFS_EXTENT_DATA_KEY)
8204                         continue;
8205
8206                 fi = btrfs_item_ptr(eb, i, struct btrfs_file_extent_item);
8207                 /* filter out non qgroup-accountable extents  */
8208                 extent_type = btrfs_file_extent_type(eb, fi);
8209
8210                 if (extent_type == BTRFS_FILE_EXTENT_INLINE)
8211                         continue;
8212
8213                 bytenr = btrfs_file_extent_disk_bytenr(eb, fi);
8214                 if (!bytenr)
8215                         continue;
8216
8217                 num_bytes = btrfs_file_extent_disk_num_bytes(eb, fi);
8218
8219                 ret = record_one_subtree_extent(trans, root, bytenr, num_bytes);
8220                 if (ret)
8221                         return ret;
8222         }
8223         return 0;
8224 }
8225
8226 /*
8227  * Walk up the tree from the bottom, freeing leaves and any interior
8228  * nodes which have had all slots visited. If a node (leaf or
8229  * interior) is freed, the node above it will have it's slot
8230  * incremented. The root node will never be freed.
8231  *
8232  * At the end of this function, we should have a path which has all
8233  * slots incremented to the next position for a search. If we need to
8234  * read a new node it will be NULL and the node above it will have the
8235  * correct slot selected for a later read.
8236  *
8237  * If we increment the root nodes slot counter past the number of
8238  * elements, 1 is returned to signal completion of the search.
8239  */
8240 static int adjust_slots_upwards(struct btrfs_root *root,
8241                                 struct btrfs_path *path, int root_level)
8242 {
8243         int level = 0;
8244         int nr, slot;
8245         struct extent_buffer *eb;
8246
8247         if (root_level == 0)
8248                 return 1;
8249
8250         while (level <= root_level) {
8251                 eb = path->nodes[level];
8252                 nr = btrfs_header_nritems(eb);
8253                 path->slots[level]++;
8254                 slot = path->slots[level];
8255                 if (slot >= nr || level == 0) {
8256                         /*
8257                          * Don't free the root -  we will detect this
8258                          * condition after our loop and return a
8259                          * positive value for caller to stop walking the tree.
8260                          */
8261                         if (level != root_level) {
8262                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8263                                 path->locks[level] = 0;
8264
8265                                 free_extent_buffer(eb);
8266                                 path->nodes[level] = NULL;
8267                                 path->slots[level] = 0;
8268                         }
8269                 } else {
8270                         /*
8271                          * We have a valid slot to walk back down
8272                          * from. Stop here so caller can process these
8273                          * new nodes.
8274                          */
8275                         break;
8276                 }
8277
8278                 level++;
8279         }
8280
8281         eb = path->nodes[root_level];
8282         if (path->slots[root_level] >= btrfs_header_nritems(eb))
8283                 return 1;
8284
8285         return 0;
8286 }
8287
8288 /*
8289  * root_eb is the subtree root and is locked before this function is called.
8290  */
8291 static int account_shared_subtree(struct btrfs_trans_handle *trans,
8292                                   struct btrfs_root *root,
8293                                   struct extent_buffer *root_eb,
8294                                   u64 root_gen,
8295                                   int root_level)
8296 {
8297         int ret = 0;
8298         int level;
8299         struct extent_buffer *eb = root_eb;
8300         struct btrfs_path *path = NULL;
8301
8302         BUG_ON(root_level < 0 || root_level > BTRFS_MAX_LEVEL);
8303         BUG_ON(root_eb == NULL);
8304
8305         if (!root->fs_info->quota_enabled)
8306                 return 0;
8307
8308         if (!extent_buffer_uptodate(root_eb)) {
8309                 ret = btrfs_read_buffer(root_eb, root_gen);
8310                 if (ret)
8311                         goto out;
8312         }
8313
8314         if (root_level == 0) {
8315                 ret = account_leaf_items(trans, root, root_eb);
8316                 goto out;
8317         }
8318
8319         path = btrfs_alloc_path();
8320         if (!path)
8321                 return -ENOMEM;
8322
8323         /*
8324          * Walk down the tree.  Missing extent blocks are filled in as
8325          * we go. Metadata is accounted every time we read a new
8326          * extent block.
8327          *
8328          * When we reach a leaf, we account for file extent items in it,
8329          * walk back up the tree (adjusting slot pointers as we go)
8330          * and restart the search process.
8331          */
8332         extent_buffer_get(root_eb); /* For path */
8333         path->nodes[root_level] = root_eb;
8334         path->slots[root_level] = 0;
8335         path->locks[root_level] = 0; /* so release_path doesn't try to unlock */
8336 walk_down:
8337         level = root_level;
8338         while (level >= 0) {
8339                 if (path->nodes[level] == NULL) {
8340                         int parent_slot;
8341                         u64 child_gen;
8342                         u64 child_bytenr;
8343
8344                         /* We need to get child blockptr/gen from
8345                          * parent before we can read it. */
8346                         eb = path->nodes[level + 1];
8347                         parent_slot = path->slots[level + 1];
8348                         child_bytenr = btrfs_node_blockptr(eb, parent_slot);
8349                         child_gen = btrfs_node_ptr_generation(eb, parent_slot);
8350
8351                         eb = read_tree_block(root, child_bytenr, child_gen);
8352                         if (IS_ERR(eb)) {
8353                                 ret = PTR_ERR(eb);
8354                                 goto out;
8355                         } else if (!extent_buffer_uptodate(eb)) {
8356                                 free_extent_buffer(eb);
8357                                 ret = -EIO;
8358                                 goto out;
8359                         }
8360
8361                         path->nodes[level] = eb;
8362                         path->slots[level] = 0;
8363
8364                         btrfs_tree_read_lock(eb);
8365                         btrfs_set_lock_blocking_rw(eb, BTRFS_READ_LOCK);
8366                         path->locks[level] = BTRFS_READ_LOCK_BLOCKING;
8367
8368                         ret = record_one_subtree_extent(trans, root, child_bytenr,
8369                                                         root->nodesize);
8370                         if (ret)
8371                                 goto out;
8372                 }
8373
8374                 if (level == 0) {
8375                         ret = account_leaf_items(trans, root, path->nodes[level]);
8376                         if (ret)
8377                                 goto out;
8378
8379                         /* Nonzero return here means we completed our search */
8380                         ret = adjust_slots_upwards(root, path, root_level);
8381                         if (ret)
8382                                 break;
8383
8384                         /* Restart search with new slots */
8385                         goto walk_down;
8386                 }
8387
8388                 level--;
8389         }
8390
8391         ret = 0;
8392 out:
8393         btrfs_free_path(path);
8394
8395         return ret;
8396 }
8397
8398 /*
8399  * helper to process tree block while walking down the tree.
8400  *
8401  * when wc->stage == UPDATE_BACKREF, this function updates
8402  * back refs for pointers in the block.
8403  *
8404  * NOTE: return value 1 means we should stop walking down.
8405  */
8406 static noinline int walk_down_proc(struct btrfs_trans_handle *trans,
8407                                    struct btrfs_root *root,
8408                                    struct btrfs_path *path,
8409                                    struct walk_control *wc, int lookup_info)
8410 {
8411         int level = wc->level;
8412         struct extent_buffer *eb = path->nodes[level];
8413         u64 flag = BTRFS_BLOCK_FLAG_FULL_BACKREF;
8414         int ret;
8415
8416         if (wc->stage == UPDATE_BACKREF &&
8417             btrfs_header_owner(eb) != root->root_key.objectid)
8418                 return 1;
8419
8420         /*
8421          * when reference count of tree block is 1, it won't increase
8422          * again. once full backref flag is set, we never clear it.
8423          */
8424         if (lookup_info &&
8425             ((wc->stage == DROP_REFERENCE && wc->refs[level] != 1) ||
8426              (wc->stage == UPDATE_BACKREF && !(wc->flags[level] & flag)))) {
8427                 BUG_ON(!path->locks[level]);
8428                 ret = btrfs_lookup_extent_info(trans, root,
8429                                                eb->start, level, 1,
8430                                                &wc->refs[level],
8431                                                &wc->flags[level]);
8432                 BUG_ON(ret == -ENOMEM);
8433                 if (ret)
8434                         return ret;
8435                 BUG_ON(wc->refs[level] == 0);
8436         }
8437
8438         if (wc->stage == DROP_REFERENCE) {
8439                 if (wc->refs[level] > 1)
8440                         return 1;
8441
8442                 if (path->locks[level] && !wc->keep_locks) {
8443                         btrfs_tree_unlock_rw(eb, path->locks[level]);
8444                         path->locks[level] = 0;
8445                 }
8446                 return 0;
8447         }
8448
8449         /* wc->stage == UPDATE_BACKREF */
8450         if (!(wc->flags[level] & flag)) {
8451                 BUG_ON(!path->locks[level]);
8452                 ret = btrfs_inc_ref(trans, root, eb, 1);
8453                 BUG_ON(ret); /* -ENOMEM */
8454                 ret = btrfs_dec_ref(trans, root, eb, 0);
8455                 BUG_ON(ret); /* -ENOMEM */
8456                 ret = btrfs_set_disk_extent_flags(trans, root, eb->start,
8457                                                   eb->len, flag,
8458                                                   btrfs_header_level(eb), 0);
8459                 BUG_ON(ret); /* -ENOMEM */
8460                 wc->flags[level] |= flag;
8461         }
8462
8463         /*
8464          * the block is shared by multiple trees, so it's not good to
8465          * keep the tree lock
8466          */
8467         if (path->locks[level] && level > 0) {
8468                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8469                 path->locks[level] = 0;
8470         }
8471         return 0;
8472 }
8473
8474 /*
8475  * helper to process tree block pointer.
8476  *
8477  * when wc->stage == DROP_REFERENCE, this function checks
8478  * reference count of the block pointed to. if the block
8479  * is shared and we need update back refs for the subtree
8480  * rooted at the block, this function changes wc->stage to
8481  * UPDATE_BACKREF. if the block is shared and there is no
8482  * need to update back, this function drops the reference
8483  * to the block.
8484  *
8485  * NOTE: return value 1 means we should stop walking down.
8486  */
8487 static noinline int do_walk_down(struct btrfs_trans_handle *trans,
8488                                  struct btrfs_root *root,
8489                                  struct btrfs_path *path,
8490                                  struct walk_control *wc, int *lookup_info)
8491 {
8492         u64 bytenr;
8493         u64 generation;
8494         u64 parent;
8495         u32 blocksize;
8496         struct btrfs_key key;
8497         struct extent_buffer *next;
8498         int level = wc->level;
8499         int reada = 0;
8500         int ret = 0;
8501         bool need_account = false;
8502
8503         generation = btrfs_node_ptr_generation(path->nodes[level],
8504                                                path->slots[level]);
8505         /*
8506          * if the lower level block was created before the snapshot
8507          * was created, we know there is no need to update back refs
8508          * for the subtree
8509          */
8510         if (wc->stage == UPDATE_BACKREF &&
8511             generation <= root->root_key.offset) {
8512                 *lookup_info = 1;
8513                 return 1;
8514         }
8515
8516         bytenr = btrfs_node_blockptr(path->nodes[level], path->slots[level]);
8517         blocksize = root->nodesize;
8518
8519         next = btrfs_find_tree_block(root->fs_info, bytenr);
8520         if (!next) {
8521                 next = btrfs_find_create_tree_block(root, bytenr);
8522                 if (!next)
8523                         return -ENOMEM;
8524                 btrfs_set_buffer_lockdep_class(root->root_key.objectid, next,
8525                                                level - 1);
8526                 reada = 1;
8527         }
8528         btrfs_tree_lock(next);
8529         btrfs_set_lock_blocking(next);
8530
8531         ret = btrfs_lookup_extent_info(trans, root, bytenr, level - 1, 1,
8532                                        &wc->refs[level - 1],
8533                                        &wc->flags[level - 1]);
8534         if (ret < 0) {
8535                 btrfs_tree_unlock(next);
8536                 return ret;
8537         }
8538
8539         if (unlikely(wc->refs[level - 1] == 0)) {
8540                 btrfs_err(root->fs_info, "Missing references.");
8541                 BUG();
8542         }
8543         *lookup_info = 0;
8544
8545         if (wc->stage == DROP_REFERENCE) {
8546                 if (wc->refs[level - 1] > 1) {
8547                         need_account = true;
8548                         if (level == 1 &&
8549                             (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8550                                 goto skip;
8551
8552                         if (!wc->update_ref ||
8553                             generation <= root->root_key.offset)
8554                                 goto skip;
8555
8556                         btrfs_node_key_to_cpu(path->nodes[level], &key,
8557                                               path->slots[level]);
8558                         ret = btrfs_comp_cpu_keys(&key, &wc->update_progress);
8559                         if (ret < 0)
8560                                 goto skip;
8561
8562                         wc->stage = UPDATE_BACKREF;
8563                         wc->shared_level = level - 1;
8564                 }
8565         } else {
8566                 if (level == 1 &&
8567                     (wc->flags[0] & BTRFS_BLOCK_FLAG_FULL_BACKREF))
8568                         goto skip;
8569         }
8570
8571         if (!btrfs_buffer_uptodate(next, generation, 0)) {
8572                 btrfs_tree_unlock(next);
8573                 free_extent_buffer(next);
8574                 next = NULL;
8575                 *lookup_info = 1;
8576         }
8577
8578         if (!next) {
8579                 if (reada && level == 1)
8580                         reada_walk_down(trans, root, wc, path);
8581                 next = read_tree_block(root, bytenr, generation);
8582                 if (IS_ERR(next)) {
8583                         return PTR_ERR(next);
8584                 } else if (!extent_buffer_uptodate(next)) {
8585                         free_extent_buffer(next);
8586                         return -EIO;
8587                 }
8588                 btrfs_tree_lock(next);
8589                 btrfs_set_lock_blocking(next);
8590         }
8591
8592         level--;
8593         BUG_ON(level != btrfs_header_level(next));
8594         path->nodes[level] = next;
8595         path->slots[level] = 0;
8596         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8597         wc->level = level;
8598         if (wc->level == 1)
8599                 wc->reada_slot = 0;
8600         return 0;
8601 skip:
8602         wc->refs[level - 1] = 0;
8603         wc->flags[level - 1] = 0;
8604         if (wc->stage == DROP_REFERENCE) {
8605                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF) {
8606                         parent = path->nodes[level]->start;
8607                 } else {
8608                         BUG_ON(root->root_key.objectid !=
8609                                btrfs_header_owner(path->nodes[level]));
8610                         parent = 0;
8611                 }
8612
8613                 if (need_account) {
8614                         ret = account_shared_subtree(trans, root, next,
8615                                                      generation, level - 1);
8616                         if (ret) {
8617                                 btrfs_err_rl(root->fs_info,
8618                                         "Error "
8619                                         "%d accounting shared subtree. Quota "
8620                                         "is out of sync, rescan required.",
8621                                         ret);
8622                         }
8623                 }
8624                 ret = btrfs_free_extent(trans, root, bytenr, blocksize, parent,
8625                                 root->root_key.objectid, level - 1, 0);
8626                 BUG_ON(ret); /* -ENOMEM */
8627         }
8628         btrfs_tree_unlock(next);
8629         free_extent_buffer(next);
8630         *lookup_info = 1;
8631         return 1;
8632 }
8633
8634 /*
8635  * helper to process tree block while walking up the tree.
8636  *
8637  * when wc->stage == DROP_REFERENCE, this function drops
8638  * reference count on the block.
8639  *
8640  * when wc->stage == UPDATE_BACKREF, this function changes
8641  * wc->stage back to DROP_REFERENCE if we changed wc->stage
8642  * to UPDATE_BACKREF previously while processing the block.
8643  *
8644  * NOTE: return value 1 means we should stop walking up.
8645  */
8646 static noinline int walk_up_proc(struct btrfs_trans_handle *trans,
8647                                  struct btrfs_root *root,
8648                                  struct btrfs_path *path,
8649                                  struct walk_control *wc)
8650 {
8651         int ret;
8652         int level = wc->level;
8653         struct extent_buffer *eb = path->nodes[level];
8654         u64 parent = 0;
8655
8656         if (wc->stage == UPDATE_BACKREF) {
8657                 BUG_ON(wc->shared_level < level);
8658                 if (level < wc->shared_level)
8659                         goto out;
8660
8661                 ret = find_next_key(path, level + 1, &wc->update_progress);
8662                 if (ret > 0)
8663                         wc->update_ref = 0;
8664
8665                 wc->stage = DROP_REFERENCE;
8666                 wc->shared_level = -1;
8667                 path->slots[level] = 0;
8668
8669                 /*
8670                  * check reference count again if the block isn't locked.
8671                  * we should start walking down the tree again if reference
8672                  * count is one.
8673                  */
8674                 if (!path->locks[level]) {
8675                         BUG_ON(level == 0);
8676                         btrfs_tree_lock(eb);
8677                         btrfs_set_lock_blocking(eb);
8678                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8679
8680                         ret = btrfs_lookup_extent_info(trans, root,
8681                                                        eb->start, level, 1,
8682                                                        &wc->refs[level],
8683                                                        &wc->flags[level]);
8684                         if (ret < 0) {
8685                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8686                                 path->locks[level] = 0;
8687                                 return ret;
8688                         }
8689                         BUG_ON(wc->refs[level] == 0);
8690                         if (wc->refs[level] == 1) {
8691                                 btrfs_tree_unlock_rw(eb, path->locks[level]);
8692                                 path->locks[level] = 0;
8693                                 return 1;
8694                         }
8695                 }
8696         }
8697
8698         /* wc->stage == DROP_REFERENCE */
8699         BUG_ON(wc->refs[level] > 1 && !path->locks[level]);
8700
8701         if (wc->refs[level] == 1) {
8702                 if (level == 0) {
8703                         if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8704                                 ret = btrfs_dec_ref(trans, root, eb, 1);
8705                         else
8706                                 ret = btrfs_dec_ref(trans, root, eb, 0);
8707                         BUG_ON(ret); /* -ENOMEM */
8708                         ret = account_leaf_items(trans, root, eb);
8709                         if (ret) {
8710                                 btrfs_err_rl(root->fs_info,
8711                                         "error "
8712                                         "%d accounting leaf items. Quota "
8713                                         "is out of sync, rescan required.",
8714                                         ret);
8715                         }
8716                 }
8717                 /* make block locked assertion in clean_tree_block happy */
8718                 if (!path->locks[level] &&
8719                     btrfs_header_generation(eb) == trans->transid) {
8720                         btrfs_tree_lock(eb);
8721                         btrfs_set_lock_blocking(eb);
8722                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8723                 }
8724                 clean_tree_block(trans, root->fs_info, eb);
8725         }
8726
8727         if (eb == root->node) {
8728                 if (wc->flags[level] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8729                         parent = eb->start;
8730                 else
8731                         BUG_ON(root->root_key.objectid !=
8732                                btrfs_header_owner(eb));
8733         } else {
8734                 if (wc->flags[level + 1] & BTRFS_BLOCK_FLAG_FULL_BACKREF)
8735                         parent = path->nodes[level + 1]->start;
8736                 else
8737                         BUG_ON(root->root_key.objectid !=
8738                                btrfs_header_owner(path->nodes[level + 1]));
8739         }
8740
8741         btrfs_free_tree_block(trans, root, eb, parent, wc->refs[level] == 1);
8742 out:
8743         wc->refs[level] = 0;
8744         wc->flags[level] = 0;
8745         return 0;
8746 }
8747
8748 static noinline int walk_down_tree(struct btrfs_trans_handle *trans,
8749                                    struct btrfs_root *root,
8750                                    struct btrfs_path *path,
8751                                    struct walk_control *wc)
8752 {
8753         int level = wc->level;
8754         int lookup_info = 1;
8755         int ret;
8756
8757         while (level >= 0) {
8758                 ret = walk_down_proc(trans, root, path, wc, lookup_info);
8759                 if (ret > 0)
8760                         break;
8761
8762                 if (level == 0)
8763                         break;
8764
8765                 if (path->slots[level] >=
8766                     btrfs_header_nritems(path->nodes[level]))
8767                         break;
8768
8769                 ret = do_walk_down(trans, root, path, wc, &lookup_info);
8770                 if (ret > 0) {
8771                         path->slots[level]++;
8772                         continue;
8773                 } else if (ret < 0)
8774                         return ret;
8775                 level = wc->level;
8776         }
8777         return 0;
8778 }
8779
8780 static noinline int walk_up_tree(struct btrfs_trans_handle *trans,
8781                                  struct btrfs_root *root,
8782                                  struct btrfs_path *path,
8783                                  struct walk_control *wc, int max_level)
8784 {
8785         int level = wc->level;
8786         int ret;
8787
8788         path->slots[level] = btrfs_header_nritems(path->nodes[level]);
8789         while (level < max_level && path->nodes[level]) {
8790                 wc->level = level;
8791                 if (path->slots[level] + 1 <
8792                     btrfs_header_nritems(path->nodes[level])) {
8793                         path->slots[level]++;
8794                         return 0;
8795                 } else {
8796                         ret = walk_up_proc(trans, root, path, wc);
8797                         if (ret > 0)
8798                                 return 0;
8799
8800                         if (path->locks[level]) {
8801                                 btrfs_tree_unlock_rw(path->nodes[level],
8802                                                      path->locks[level]);
8803                                 path->locks[level] = 0;
8804                         }
8805                         free_extent_buffer(path->nodes[level]);
8806                         path->nodes[level] = NULL;
8807                         level++;
8808                 }
8809         }
8810         return 1;
8811 }
8812
8813 /*
8814  * drop a subvolume tree.
8815  *
8816  * this function traverses the tree freeing any blocks that only
8817  * referenced by the tree.
8818  *
8819  * when a shared tree block is found. this function decreases its
8820  * reference count by one. if update_ref is true, this function
8821  * also make sure backrefs for the shared block and all lower level
8822  * blocks are properly updated.
8823  *
8824  * If called with for_reloc == 0, may exit early with -EAGAIN
8825  */
8826 int btrfs_drop_snapshot(struct btrfs_root *root,
8827                          struct btrfs_block_rsv *block_rsv, int update_ref,
8828                          int for_reloc)
8829 {
8830         struct btrfs_path *path;
8831         struct btrfs_trans_handle *trans;
8832         struct btrfs_root *tree_root = root->fs_info->tree_root;
8833         struct btrfs_root_item *root_item = &root->root_item;
8834         struct walk_control *wc;
8835         struct btrfs_key key;
8836         int err = 0;
8837         int ret;
8838         int level;
8839         bool root_dropped = false;
8840
8841         btrfs_debug(root->fs_info, "Drop subvolume %llu", root->objectid);
8842
8843         path = btrfs_alloc_path();
8844         if (!path) {
8845                 err = -ENOMEM;
8846                 goto out;
8847         }
8848
8849         wc = kzalloc(sizeof(*wc), GFP_NOFS);
8850         if (!wc) {
8851                 btrfs_free_path(path);
8852                 err = -ENOMEM;
8853                 goto out;
8854         }
8855
8856         trans = btrfs_start_transaction(tree_root, 0);
8857         if (IS_ERR(trans)) {
8858                 err = PTR_ERR(trans);
8859                 goto out_free;
8860         }
8861
8862         if (block_rsv)
8863                 trans->block_rsv = block_rsv;
8864
8865         if (btrfs_disk_key_objectid(&root_item->drop_progress) == 0) {
8866                 level = btrfs_header_level(root->node);
8867                 path->nodes[level] = btrfs_lock_root_node(root);
8868                 btrfs_set_lock_blocking(path->nodes[level]);
8869                 path->slots[level] = 0;
8870                 path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8871                 memset(&wc->update_progress, 0,
8872                        sizeof(wc->update_progress));
8873         } else {
8874                 btrfs_disk_key_to_cpu(&key, &root_item->drop_progress);
8875                 memcpy(&wc->update_progress, &key,
8876                        sizeof(wc->update_progress));
8877
8878                 level = root_item->drop_level;
8879                 BUG_ON(level == 0);
8880                 path->lowest_level = level;
8881                 ret = btrfs_search_slot(NULL, root, &key, path, 0, 0);
8882                 path->lowest_level = 0;
8883                 if (ret < 0) {
8884                         err = ret;
8885                         goto out_end_trans;
8886                 }
8887                 WARN_ON(ret > 0);
8888
8889                 /*
8890                  * unlock our path, this is safe because only this
8891                  * function is allowed to delete this snapshot
8892                  */
8893                 btrfs_unlock_up_safe(path, 0);
8894
8895                 level = btrfs_header_level(root->node);
8896                 while (1) {
8897                         btrfs_tree_lock(path->nodes[level]);
8898                         btrfs_set_lock_blocking(path->nodes[level]);
8899                         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
8900
8901                         ret = btrfs_lookup_extent_info(trans, root,
8902                                                 path->nodes[level]->start,
8903                                                 level, 1, &wc->refs[level],
8904                                                 &wc->flags[level]);
8905                         if (ret < 0) {
8906                                 err = ret;
8907                                 goto out_end_trans;
8908                         }
8909                         BUG_ON(wc->refs[level] == 0);
8910
8911                         if (level == root_item->drop_level)
8912                                 break;
8913
8914                         btrfs_tree_unlock(path->nodes[level]);
8915                         path->locks[level] = 0;
8916                         WARN_ON(wc->refs[level] != 1);
8917                         level--;
8918                 }
8919         }
8920
8921         wc->level = level;
8922         wc->shared_level = -1;
8923         wc->stage = DROP_REFERENCE;
8924         wc->update_ref = update_ref;
8925         wc->keep_locks = 0;
8926         wc->for_reloc = for_reloc;
8927         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
8928
8929         while (1) {
8930
8931                 ret = walk_down_tree(trans, root, path, wc);
8932                 if (ret < 0) {
8933                         err = ret;
8934                         break;
8935                 }
8936
8937                 ret = walk_up_tree(trans, root, path, wc, BTRFS_MAX_LEVEL);
8938                 if (ret < 0) {
8939                         err = ret;
8940                         break;
8941                 }
8942
8943                 if (ret > 0) {
8944                         BUG_ON(wc->stage != DROP_REFERENCE);
8945                         break;
8946                 }
8947
8948                 if (wc->stage == DROP_REFERENCE) {
8949                         level = wc->level;
8950                         btrfs_node_key(path->nodes[level],
8951                                        &root_item->drop_progress,
8952                                        path->slots[level]);
8953                         root_item->drop_level = level;
8954                 }
8955
8956                 BUG_ON(wc->level == 0);
8957                 if (btrfs_should_end_transaction(trans, tree_root) ||
8958                     (!for_reloc && btrfs_need_cleaner_sleep(root))) {
8959                         ret = btrfs_update_root(trans, tree_root,
8960                                                 &root->root_key,
8961                                                 root_item);
8962                         if (ret) {
8963                                 btrfs_abort_transaction(trans, tree_root, ret);
8964                                 err = ret;
8965                                 goto out_end_trans;
8966                         }
8967
8968                         btrfs_end_transaction_throttle(trans, tree_root);
8969                         if (!for_reloc && btrfs_need_cleaner_sleep(root)) {
8970                                 pr_debug("BTRFS: drop snapshot early exit\n");
8971                                 err = -EAGAIN;
8972                                 goto out_free;
8973                         }
8974
8975                         trans = btrfs_start_transaction(tree_root, 0);
8976                         if (IS_ERR(trans)) {
8977                                 err = PTR_ERR(trans);
8978                                 goto out_free;
8979                         }
8980                         if (block_rsv)
8981                                 trans->block_rsv = block_rsv;
8982                 }
8983         }
8984         btrfs_release_path(path);
8985         if (err)
8986                 goto out_end_trans;
8987
8988         ret = btrfs_del_root(trans, tree_root, &root->root_key);
8989         if (ret) {
8990                 btrfs_abort_transaction(trans, tree_root, ret);
8991                 goto out_end_trans;
8992         }
8993
8994         if (root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID) {
8995                 ret = btrfs_find_root(tree_root, &root->root_key, path,
8996                                       NULL, NULL);
8997                 if (ret < 0) {
8998                         btrfs_abort_transaction(trans, tree_root, ret);
8999                         err = ret;
9000                         goto out_end_trans;
9001                 } else if (ret > 0) {
9002                         /* if we fail to delete the orphan item this time
9003                          * around, it'll get picked up the next time.
9004                          *
9005                          * The most common failure here is just -ENOENT.
9006                          */
9007                         btrfs_del_orphan_item(trans, tree_root,
9008                                               root->root_key.objectid);
9009                 }
9010         }
9011
9012         if (test_bit(BTRFS_ROOT_IN_RADIX, &root->state)) {
9013                 btrfs_add_dropped_root(trans, root);
9014         } else {
9015                 free_extent_buffer(root->node);
9016                 free_extent_buffer(root->commit_root);
9017                 btrfs_put_fs_root(root);
9018         }
9019         root_dropped = true;
9020 out_end_trans:
9021         btrfs_end_transaction_throttle(trans, tree_root);
9022 out_free:
9023         kfree(wc);
9024         btrfs_free_path(path);
9025 out:
9026         /*
9027          * So if we need to stop dropping the snapshot for whatever reason we
9028          * need to make sure to add it back to the dead root list so that we
9029          * keep trying to do the work later.  This also cleans up roots if we
9030          * don't have it in the radix (like when we recover after a power fail
9031          * or unmount) so we don't leak memory.
9032          */
9033         if (!for_reloc && root_dropped == false)
9034                 btrfs_add_dead_root(root);
9035         if (err && err != -EAGAIN)
9036                 btrfs_std_error(root->fs_info, err, NULL);
9037         return err;
9038 }
9039
9040 /*
9041  * drop subtree rooted at tree block 'node'.
9042  *
9043  * NOTE: this function will unlock and release tree block 'node'
9044  * only used by relocation code
9045  */
9046 int btrfs_drop_subtree(struct btrfs_trans_handle *trans,
9047                         struct btrfs_root *root,
9048                         struct extent_buffer *node,
9049                         struct extent_buffer *parent)
9050 {
9051         struct btrfs_path *path;
9052         struct walk_control *wc;
9053         int level;
9054         int parent_level;
9055         int ret = 0;
9056         int wret;
9057
9058         BUG_ON(root->root_key.objectid != BTRFS_TREE_RELOC_OBJECTID);
9059
9060         path = btrfs_alloc_path();
9061         if (!path)
9062                 return -ENOMEM;
9063
9064         wc = kzalloc(sizeof(*wc), GFP_NOFS);
9065         if (!wc) {
9066                 btrfs_free_path(path);
9067                 return -ENOMEM;
9068         }
9069
9070         btrfs_assert_tree_locked(parent);
9071         parent_level = btrfs_header_level(parent);
9072         extent_buffer_get(parent);
9073         path->nodes[parent_level] = parent;
9074         path->slots[parent_level] = btrfs_header_nritems(parent);
9075
9076         btrfs_assert_tree_locked(node);
9077         level = btrfs_header_level(node);
9078         path->nodes[level] = node;
9079         path->slots[level] = 0;
9080         path->locks[level] = BTRFS_WRITE_LOCK_BLOCKING;
9081
9082         wc->refs[parent_level] = 1;
9083         wc->flags[parent_level] = BTRFS_BLOCK_FLAG_FULL_BACKREF;
9084         wc->level = level;
9085         wc->shared_level = -1;
9086         wc->stage = DROP_REFERENCE;
9087         wc->update_ref = 0;
9088         wc->keep_locks = 1;
9089         wc->for_reloc = 1;
9090         wc->reada_count = BTRFS_NODEPTRS_PER_BLOCK(root);
9091
9092         while (1) {
9093                 wret = walk_down_tree(trans, root, path, wc);
9094                 if (wret < 0) {
9095                         ret = wret;
9096                         break;
9097                 }
9098
9099                 wret = walk_up_tree(trans, root, path, wc, parent_level);
9100                 if (wret < 0)
9101                         ret = wret;
9102                 if (wret != 0)
9103                         break;
9104         }
9105
9106         kfree(wc);
9107         btrfs_free_path(path);
9108         return ret;
9109 }
9110
9111 static u64 update_block_group_flags(struct btrfs_root *root, u64 flags)
9112 {
9113         u64 num_devices;
9114         u64 stripped;
9115
9116         /*
9117          * if restripe for this chunk_type is on pick target profile and
9118          * return, otherwise do the usual balance
9119          */
9120         stripped = get_restripe_target(root->fs_info, flags);
9121         if (stripped)
9122                 return extended_to_chunk(stripped);
9123
9124         num_devices = root->fs_info->fs_devices->rw_devices;
9125
9126         stripped = BTRFS_BLOCK_GROUP_RAID0 |
9127                 BTRFS_BLOCK_GROUP_RAID5 | BTRFS_BLOCK_GROUP_RAID6 |
9128                 BTRFS_BLOCK_GROUP_RAID1 | BTRFS_BLOCK_GROUP_RAID10;
9129
9130         if (num_devices == 1) {
9131                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9132                 stripped = flags & ~stripped;
9133
9134                 /* turn raid0 into single device chunks */
9135                 if (flags & BTRFS_BLOCK_GROUP_RAID0)
9136                         return stripped;
9137
9138                 /* turn mirroring into duplication */
9139                 if (flags & (BTRFS_BLOCK_GROUP_RAID1 |
9140                              BTRFS_BLOCK_GROUP_RAID10))
9141                         return stripped | BTRFS_BLOCK_GROUP_DUP;
9142         } else {
9143                 /* they already had raid on here, just return */
9144                 if (flags & stripped)
9145                         return flags;
9146
9147                 stripped |= BTRFS_BLOCK_GROUP_DUP;
9148                 stripped = flags & ~stripped;
9149
9150                 /* switch duplicated blocks with raid1 */
9151                 if (flags & BTRFS_BLOCK_GROUP_DUP)
9152                         return stripped | BTRFS_BLOCK_GROUP_RAID1;
9153
9154                 /* this is drive concat, leave it alone */
9155         }
9156
9157         return flags;
9158 }
9159
9160 static int inc_block_group_ro(struct btrfs_block_group_cache *cache, int force)
9161 {
9162         struct btrfs_space_info *sinfo = cache->space_info;
9163         u64 num_bytes;
9164         u64 min_allocable_bytes;
9165         int ret = -ENOSPC;
9166
9167         /*
9168          * We need some metadata space and system metadata space for
9169          * allocating chunks in some corner cases until we force to set
9170          * it to be readonly.
9171          */
9172         if ((sinfo->flags &
9173              (BTRFS_BLOCK_GROUP_SYSTEM | BTRFS_BLOCK_GROUP_METADATA)) &&
9174             !force)
9175                 min_allocable_bytes = 1 * 1024 * 1024;
9176         else
9177                 min_allocable_bytes = 0;
9178
9179         spin_lock(&sinfo->lock);
9180         spin_lock(&cache->lock);
9181
9182         if (cache->ro) {
9183                 cache->ro++;
9184                 ret = 0;
9185                 goto out;
9186         }
9187
9188         num_bytes = cache->key.offset - cache->reserved - cache->pinned -
9189                     cache->bytes_super - btrfs_block_group_used(&cache->item);
9190
9191         if (sinfo->bytes_used + sinfo->bytes_reserved + sinfo->bytes_pinned +
9192             sinfo->bytes_may_use + sinfo->bytes_readonly + num_bytes +
9193             min_allocable_bytes <= sinfo->total_bytes) {
9194                 sinfo->bytes_readonly += num_bytes;
9195                 cache->ro++;
9196                 list_add_tail(&cache->ro_list, &sinfo->ro_bgs);
9197                 ret = 0;
9198         }
9199 out:
9200         spin_unlock(&cache->lock);
9201         spin_unlock(&sinfo->lock);
9202         return ret;
9203 }
9204
9205 int btrfs_inc_block_group_ro(struct btrfs_root *root,
9206                              struct btrfs_block_group_cache *cache)
9207
9208 {
9209         struct btrfs_trans_handle *trans;
9210         u64 alloc_flags;
9211         int ret;
9212
9213 again:
9214         trans = btrfs_join_transaction(root);
9215         if (IS_ERR(trans))
9216                 return PTR_ERR(trans);
9217
9218         /*
9219          * we're not allowed to set block groups readonly after the dirty
9220          * block groups cache has started writing.  If it already started,
9221          * back off and let this transaction commit
9222          */
9223         mutex_lock(&root->fs_info->ro_block_group_mutex);
9224         if (test_bit(BTRFS_TRANS_DIRTY_BG_RUN, &trans->transaction->flags)) {
9225                 u64 transid = trans->transid;
9226
9227                 mutex_unlock(&root->fs_info->ro_block_group_mutex);
9228                 btrfs_end_transaction(trans, root);
9229
9230                 ret = btrfs_wait_for_commit(root, transid);
9231                 if (ret)
9232                         return ret;
9233                 goto again;
9234         }
9235
9236         /*
9237          * if we are changing raid levels, try to allocate a corresponding
9238          * block group with the new raid level.
9239          */
9240         alloc_flags = update_block_group_flags(root, cache->flags);
9241         if (alloc_flags != cache->flags) {
9242                 ret = do_chunk_alloc(trans, root, alloc_flags,
9243                                      CHUNK_ALLOC_FORCE);
9244                 /*
9245                  * ENOSPC is allowed here, we may have enough space
9246                  * already allocated at the new raid level to
9247                  * carry on
9248                  */
9249                 if (ret == -ENOSPC)
9250                         ret = 0;
9251                 if (ret < 0)
9252                         goto out;
9253         }
9254
9255         ret = inc_block_group_ro(cache, 0);
9256         if (!ret)
9257                 goto out;
9258         alloc_flags = get_alloc_profile(root, cache->space_info->flags);
9259         ret = do_chunk_alloc(trans, root, alloc_flags,
9260                              CHUNK_ALLOC_FORCE);
9261         if (ret < 0)
9262                 goto out;
9263         ret = inc_block_group_ro(cache, 0);
9264 out:
9265         if (cache->flags & BTRFS_BLOCK_GROUP_SYSTEM) {
9266                 alloc_flags = update_block_group_flags(root, cache->flags);
9267                 lock_chunks(root->fs_info->chunk_root);
9268                 check_system_chunk(trans, root, alloc_flags);
9269                 unlock_chunks(root->fs_info->chunk_root);
9270         }
9271         mutex_unlock(&root->fs_info->ro_block_group_mutex);
9272
9273         btrfs_end_transaction(trans, root);
9274         return ret;
9275 }
9276
9277 int btrfs_force_chunk_alloc(struct btrfs_trans_handle *trans,
9278                             struct btrfs_root *root, u64 type)
9279 {
9280         u64 alloc_flags = get_alloc_profile(root, type);
9281         return do_chunk_alloc(trans, root, alloc_flags,
9282                               CHUNK_ALLOC_FORCE);
9283 }
9284
9285 /*
9286  * helper to account the unused space of all the readonly block group in the
9287  * space_info. takes mirrors into account.
9288  */
9289 u64 btrfs_account_ro_block_groups_free_space(struct btrfs_space_info *sinfo)
9290 {
9291         struct btrfs_block_group_cache *block_group;
9292         u64 free_bytes = 0;
9293         int factor;
9294
9295         /* It's df, we don't care if it's racey */
9296         if (list_empty(&sinfo->ro_bgs))
9297                 return 0;
9298
9299         spin_lock(&sinfo->lock);
9300         list_for_each_entry(block_group, &sinfo->ro_bgs, ro_list) {
9301                 spin_lock(&block_group->lock);
9302
9303                 if (!block_group->ro) {
9304                         spin_unlock(&block_group->lock);
9305                         continue;
9306                 }
9307
9308                 if (block_group->flags & (BTRFS_BLOCK_GROUP_RAID1 |
9309                                           BTRFS_BLOCK_GROUP_RAID10 |
9310                                           BTRFS_BLOCK_GROUP_DUP))
9311                         factor = 2;
9312                 else
9313                         factor = 1;
9314
9315                 free_bytes += (block_group->key.offset -
9316                                btrfs_block_group_used(&block_group->item)) *
9317                                factor;
9318
9319                 spin_unlock(&block_group->lock);
9320         }
9321         spin_unlock(&sinfo->lock);
9322
9323         return free_bytes;
9324 }
9325
9326 void btrfs_dec_block_group_ro(struct btrfs_root *root,
9327                               struct btrfs_block_group_cache *cache)
9328 {
9329         struct btrfs_space_info *sinfo = cache->space_info;
9330         u64 num_bytes;
9331
9332         BUG_ON(!cache->ro);
9333
9334         spin_lock(&sinfo->lock);
9335         spin_lock(&cache->lock);
9336         if (!--cache->ro) {
9337                 num_bytes = cache->key.offset - cache->reserved -
9338                             cache->pinned - cache->bytes_super -
9339                             btrfs_block_group_used(&cache->item);
9340                 sinfo->bytes_readonly -= num_bytes;
9341                 list_del_init(&cache->ro_list);
9342         }
9343         spin_unlock(&cache->lock);
9344         spin_unlock(&sinfo->lock);
9345 }
9346
9347 /*
9348  * checks to see if its even possible to relocate this block group.
9349  *
9350  * @return - -1 if it's not a good idea to relocate this block group, 0 if its
9351  * ok to go ahead and try.
9352  */
9353 int btrfs_can_relocate(struct btrfs_root *root, u64 bytenr)
9354 {
9355         struct btrfs_block_group_cache *block_group;
9356         struct btrfs_space_info *space_info;
9357         struct btrfs_fs_devices *fs_devices = root->fs_info->fs_devices;
9358         struct btrfs_device *device;
9359         struct btrfs_trans_handle *trans;
9360         u64 min_free;
9361         u64 dev_min = 1;
9362         u64 dev_nr = 0;
9363         u64 target;
9364         int index;
9365         int full = 0;
9366         int ret = 0;
9367
9368         block_group = btrfs_lookup_block_group(root->fs_info, bytenr);
9369
9370         /* odd, couldn't find the block group, leave it alone */
9371         if (!block_group)
9372                 return -1;
9373
9374         min_free = btrfs_block_group_used(&block_group->item);
9375
9376         /* no bytes used, we're good */
9377         if (!min_free)
9378                 goto out;
9379
9380         space_info = block_group->space_info;
9381         spin_lock(&space_info->lock);
9382
9383         full = space_info->full;
9384
9385         /*
9386          * if this is the last block group we have in this space, we can't
9387          * relocate it unless we're able to allocate a new chunk below.
9388          *
9389          * Otherwise, we need to make sure we have room in the space to handle
9390          * all of the extents from this block group.  If we can, we're good
9391          */
9392         if ((space_info->total_bytes != block_group->key.offset) &&
9393             (space_info->bytes_used + space_info->bytes_reserved +
9394              space_info->bytes_pinned + space_info->bytes_readonly +
9395              min_free < space_info->total_bytes)) {
9396                 spin_unlock(&space_info->lock);
9397                 goto out;
9398         }
9399         spin_unlock(&space_info->lock);
9400
9401         /*
9402          * ok we don't have enough space, but maybe we have free space on our
9403          * devices to allocate new chunks for relocation, so loop through our
9404          * alloc devices and guess if we have enough space.  if this block
9405          * group is going to be restriped, run checks against the target
9406          * profile instead of the current one.
9407          */
9408         ret = -1;
9409
9410         /*
9411          * index:
9412          *      0: raid10
9413          *      1: raid1
9414          *      2: dup
9415          *      3: raid0
9416          *      4: single
9417          */
9418         target = get_restripe_target(root->fs_info, block_group->flags);
9419         if (target) {
9420                 index = __get_raid_index(extended_to_chunk(target));
9421         } else {
9422                 /*
9423                  * this is just a balance, so if we were marked as full
9424                  * we know there is no space for a new chunk
9425                  */
9426                 if (full)
9427                         goto out;
9428
9429                 index = get_block_group_index(block_group);
9430         }
9431
9432         if (index == BTRFS_RAID_RAID10) {
9433                 dev_min = 4;
9434                 /* Divide by 2 */
9435                 min_free >>= 1;
9436         } else if (index == BTRFS_RAID_RAID1) {
9437                 dev_min = 2;
9438         } else if (index == BTRFS_RAID_DUP) {
9439                 /* Multiply by 2 */
9440                 min_free <<= 1;
9441         } else if (index == BTRFS_RAID_RAID0) {
9442                 dev_min = fs_devices->rw_devices;
9443                 min_free = div64_u64(min_free, dev_min);
9444         }
9445
9446         /* We need to do this so that we can look at pending chunks */
9447         trans = btrfs_join_transaction(root);
9448         if (IS_ERR(trans)) {
9449                 ret = PTR_ERR(trans);
9450                 goto out;
9451         }
9452
9453         mutex_lock(&root->fs_info->chunk_mutex);
9454         list_for_each_entry(device, &fs_devices->alloc_list, dev_alloc_list) {
9455                 u64 dev_offset;
9456
9457                 /*
9458                  * check to make sure we can actually find a chunk with enough
9459                  * space to fit our block group in.
9460                  */
9461                 if (device->total_bytes > device->bytes_used + min_free &&
9462                     !device->is_tgtdev_for_dev_replace) {
9463                         ret = find_free_dev_extent(trans, device, min_free,
9464                                                    &dev_offset, NULL);
9465                         if (!ret)
9466                                 dev_nr++;
9467
9468                         if (dev_nr >= dev_min)
9469                                 break;
9470
9471                         ret = -1;
9472                 }
9473         }
9474         mutex_unlock(&root->fs_info->chunk_mutex);
9475         btrfs_end_transaction(trans, root);
9476 out:
9477         btrfs_put_block_group(block_group);
9478         return ret;
9479 }
9480
9481 static int find_first_block_group(struct btrfs_root *root,
9482                 struct btrfs_path *path, struct btrfs_key *key)
9483 {
9484         int ret = 0;
9485         struct btrfs_key found_key;
9486         struct extent_buffer *leaf;
9487         int slot;
9488
9489         ret = btrfs_search_slot(NULL, root, key, path, 0, 0);
9490         if (ret < 0)
9491                 goto out;
9492
9493         while (1) {
9494                 slot = path->slots[0];
9495                 leaf = path->nodes[0];
9496                 if (slot >= btrfs_header_nritems(leaf)) {
9497                         ret = btrfs_next_leaf(root, path);
9498                         if (ret == 0)
9499                                 continue;
9500                         if (ret < 0)
9501                                 goto out;
9502                         break;
9503                 }
9504                 btrfs_item_key_to_cpu(leaf, &found_key, slot);
9505
9506                 if (found_key.objectid >= key->objectid &&
9507                     found_key.type == BTRFS_BLOCK_GROUP_ITEM_KEY) {
9508                         ret = 0;
9509                         goto out;
9510                 }
9511                 path->slots[0]++;
9512         }
9513 out:
9514         return ret;
9515 }
9516
9517 void btrfs_put_block_group_cache(struct btrfs_fs_info *info)
9518 {
9519         struct btrfs_block_group_cache *block_group;
9520         u64 last = 0;
9521
9522         while (1) {
9523                 struct inode *inode;
9524
9525                 block_group = btrfs_lookup_first_block_group(info, last);
9526                 while (block_group) {
9527                         spin_lock(&block_group->lock);
9528                         if (block_group->iref)
9529                                 break;
9530                         spin_unlock(&block_group->lock);
9531                         block_group = next_block_group(info->tree_root,
9532                                                        block_group);
9533                 }
9534                 if (!block_group) {
9535                         if (last == 0)
9536                                 break;
9537                         last = 0;
9538                         continue;
9539                 }
9540
9541                 inode = block_group->inode;
9542                 block_group->iref = 0;
9543                 block_group->inode = NULL;
9544                 spin_unlock(&block_group->lock);
9545                 iput(inode);
9546                 last = block_group->key.objectid + block_group->key.offset;
9547                 btrfs_put_block_group(block_group);
9548         }
9549 }
9550
9551 int btrfs_free_block_groups(struct btrfs_fs_info *info)
9552 {
9553         struct btrfs_block_group_cache *block_group;
9554         struct btrfs_space_info *space_info;
9555         struct btrfs_caching_control *caching_ctl;
9556         struct rb_node *n;
9557
9558         down_write(&info->commit_root_sem);
9559         while (!list_empty(&info->caching_block_groups)) {
9560                 caching_ctl = list_entry(info->caching_block_groups.next,
9561                                          struct btrfs_caching_control, list);
9562                 list_del(&caching_ctl->list);
9563                 put_caching_control(caching_ctl);
9564         }
9565         up_write(&info->commit_root_sem);
9566
9567         spin_lock(&info->unused_bgs_lock);
9568         while (!list_empty(&info->unused_bgs)) {
9569                 block_group = list_first_entry(&info->unused_bgs,
9570                                                struct btrfs_block_group_cache,
9571                                                bg_list);
9572                 list_del_init(&block_group->bg_list);
9573                 btrfs_put_block_group(block_group);
9574         }
9575         spin_unlock(&info->unused_bgs_lock);
9576
9577         spin_lock(&info->block_group_cache_lock);
9578         while ((n = rb_last(&info->block_group_cache_tree)) != NULL) {
9579                 block_group = rb_entry(n, struct btrfs_block_group_cache,
9580                                        cache_node);
9581                 rb_erase(&block_group->cache_node,
9582                          &info->block_group_cache_tree);
9583                 RB_CLEAR_NODE(&block_group->cache_node);
9584                 spin_unlock(&info->block_group_cache_lock);
9585
9586                 down_write(&block_group->space_info->groups_sem);
9587                 list_del(&block_group->list);
9588                 up_write(&block_group->space_info->groups_sem);
9589
9590                 if (block_group->cached == BTRFS_CACHE_STARTED)
9591                         wait_block_group_cache_done(block_group);
9592
9593                 /*
9594                  * We haven't cached this block group, which means we could
9595                  * possibly have excluded extents on this block group.
9596                  */
9597                 if (block_group->cached == BTRFS_CACHE_NO ||
9598                     block_group->cached == BTRFS_CACHE_ERROR)
9599                         free_excluded_extents(info->extent_root, block_group);
9600
9601                 btrfs_remove_free_space_cache(block_group);
9602                 btrfs_put_block_group(block_group);
9603
9604                 spin_lock(&info->block_group_cache_lock);
9605         }
9606         spin_unlock(&info->block_group_cache_lock);
9607
9608         /* now that all the block groups are freed, go through and
9609          * free all the space_info structs.  This is only called during
9610          * the final stages of unmount, and so we know nobody is
9611          * using them.  We call synchronize_rcu() once before we start,
9612          * just to be on the safe side.
9613          */
9614         synchronize_rcu();
9615
9616         release_global_block_rsv(info);
9617
9618         while (!list_empty(&info->space_info)) {
9619                 int i;
9620
9621                 space_info = list_entry(info->space_info.next,
9622                                         struct btrfs_space_info,
9623                                         list);
9624                 if (btrfs_test_opt(info->tree_root, ENOSPC_DEBUG)) {
9625                         if (WARN_ON(space_info->bytes_pinned > 0 ||
9626                             space_info->bytes_reserved > 0 ||
9627                             space_info->bytes_may_use > 0)) {
9628                                 dump_space_info(space_info, 0, 0);
9629                         }
9630                 }
9631                 list_del(&space_info->list);
9632                 for (i = 0; i < BTRFS_NR_RAID_TYPES; i++) {
9633                         struct kobject *kobj;
9634                         kobj = space_info->block_group_kobjs[i];
9635                         space_info->block_group_kobjs[i] = NULL;
9636                         if (kobj) {
9637                                 kobject_del(kobj);
9638                                 kobject_put(kobj);
9639                         }
9640                 }
9641                 kobject_del(&space_info->kobj);
9642                 kobject_put(&space_info->kobj);
9643         }
9644         return 0;
9645 }
9646
9647 static void __link_block_group(struct btrfs_space_info *space_info,
9648                                struct btrfs_block_group_cache *cache)
9649 {
9650         int index = get_block_group_index(cache);
9651         bool first = false;
9652
9653         down_write(&space_info->groups_sem);
9654         if (list_empty(&space_info->block_groups[index]))
9655                 first = true;
9656         list_add_tail(&cache->list, &space_info->block_groups[index]);
9657         up_write(&space_info->groups_sem);
9658
9659         if (first) {
9660                 struct raid_kobject *rkobj;
9661                 int ret;
9662
9663                 rkobj = kzalloc(sizeof(*rkobj), GFP_NOFS);
9664                 if (!rkobj)
9665                         goto out_err;
9666                 rkobj->raid_type = index;
9667                 kobject_init(&rkobj->kobj, &btrfs_raid_ktype);
9668                 ret = kobject_add(&rkobj->kobj, &space_info->kobj,
9669                                   "%s", get_raid_name(index));
9670                 if (ret) {
9671                         kobject_put(&rkobj->kobj);
9672                         goto out_err;
9673                 }
9674                 space_info->block_group_kobjs[index] = &rkobj->kobj;
9675         }
9676
9677         return;
9678 out_err:
9679         pr_warn("BTRFS: failed to add kobject for block cache. ignoring.\n");
9680 }
9681
9682 static struct btrfs_block_group_cache *
9683 btrfs_create_block_group_cache(struct btrfs_root *root, u64 start, u64 size)
9684 {
9685         struct btrfs_block_group_cache *cache;
9686
9687         cache = kzalloc(sizeof(*cache), GFP_NOFS);
9688         if (!cache)
9689                 return NULL;
9690
9691         cache->free_space_ctl = kzalloc(sizeof(*cache->free_space_ctl),
9692                                         GFP_NOFS);
9693         if (!cache->free_space_ctl) {
9694                 kfree(cache);
9695                 return NULL;
9696         }
9697
9698         cache->key.objectid = start;
9699         cache->key.offset = size;
9700         cache->key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9701
9702         cache->sectorsize = root->sectorsize;
9703         cache->fs_info = root->fs_info;
9704         cache->full_stripe_len = btrfs_full_stripe_len(root,
9705                                                &root->fs_info->mapping_tree,
9706                                                start);
9707         set_free_space_tree_thresholds(cache);
9708
9709         atomic_set(&cache->count, 1);
9710         spin_lock_init(&cache->lock);
9711         init_rwsem(&cache->data_rwsem);
9712         INIT_LIST_HEAD(&cache->list);
9713         INIT_LIST_HEAD(&cache->cluster_list);
9714         INIT_LIST_HEAD(&cache->bg_list);
9715         INIT_LIST_HEAD(&cache->ro_list);
9716         INIT_LIST_HEAD(&cache->dirty_list);
9717         INIT_LIST_HEAD(&cache->io_list);
9718         btrfs_init_free_space_ctl(cache);
9719         atomic_set(&cache->trimming, 0);
9720         mutex_init(&cache->free_space_lock);
9721
9722         return cache;
9723 }
9724
9725 int btrfs_read_block_groups(struct btrfs_root *root)
9726 {
9727         struct btrfs_path *path;
9728         int ret;
9729         struct btrfs_block_group_cache *cache;
9730         struct btrfs_fs_info *info = root->fs_info;
9731         struct btrfs_space_info *space_info;
9732         struct btrfs_key key;
9733         struct btrfs_key found_key;
9734         struct extent_buffer *leaf;
9735         int need_clear = 0;
9736         u64 cache_gen;
9737
9738         root = info->extent_root;
9739         key.objectid = 0;
9740         key.offset = 0;
9741         key.type = BTRFS_BLOCK_GROUP_ITEM_KEY;
9742         path = btrfs_alloc_path();
9743         if (!path)
9744                 return -ENOMEM;
9745         path->reada = 1;
9746
9747         cache_gen = btrfs_super_cache_generation(root->fs_info->super_copy);
9748         if (btrfs_test_opt(root, SPACE_CACHE) &&
9749             btrfs_super_generation(root->fs_info->super_copy) != cache_gen)
9750                 need_clear = 1;
9751         if (btrfs_test_opt(root, CLEAR_CACHE))
9752                 need_clear = 1;
9753
9754         while (1) {
9755                 ret = find_first_block_group(root, path, &key);
9756                 if (ret > 0)
9757                         break;
9758                 if (ret != 0)
9759                         goto error;
9760
9761                 leaf = path->nodes[0];
9762                 btrfs_item_key_to_cpu(leaf, &found_key, path->slots[0]);
9763
9764                 cache = btrfs_create_block_group_cache(root, found_key.objectid,
9765                                                        found_key.offset);
9766                 if (!cache) {
9767                         ret = -ENOMEM;
9768                         goto error;
9769                 }
9770
9771                 if (need_clear) {
9772                         /*
9773                          * When we mount with old space cache, we need to
9774                          * set BTRFS_DC_CLEAR and set dirty flag.
9775                          *
9776                          * a) Setting 'BTRFS_DC_CLEAR' makes sure that we
9777                          *    truncate the old free space cache inode and
9778                          *    setup a new one.
9779                          * b) Setting 'dirty flag' makes sure that we flush
9780                          *    the new space cache info onto disk.
9781                          */
9782                         if (btrfs_test_opt(root, SPACE_CACHE))
9783                                 cache->disk_cache_state = BTRFS_DC_CLEAR;
9784                 }
9785
9786                 read_extent_buffer(leaf, &cache->item,
9787                                    btrfs_item_ptr_offset(leaf, path->slots[0]),
9788                                    sizeof(cache->item));
9789                 cache->flags = btrfs_block_group_flags(&cache->item);
9790
9791                 key.objectid = found_key.objectid + found_key.offset;
9792                 btrfs_release_path(path);
9793
9794                 /*
9795                  * We need to exclude the super stripes now so that the space
9796                  * info has super bytes accounted for, otherwise we'll think
9797                  * we have more space than we actually do.
9798                  */
9799                 ret = exclude_super_stripes(root, cache);
9800                 if (ret) {
9801                         /*
9802                          * We may have excluded something, so call this just in
9803                          * case.
9804                          */
9805                         free_excluded_extents(root, cache);
9806                         btrfs_put_block_group(cache);
9807                         goto error;
9808                 }
9809
9810                 /*
9811                  * check for two cases, either we are full, and therefore
9812                  * don't need to bother with the caching work since we won't
9813                  * find any space, or we are empty, and we can just add all
9814                  * the space in and be done with it.  This saves us _alot_ of
9815                  * time, particularly in the full case.
9816                  */
9817                 if (found_key.offset == btrfs_block_group_used(&cache->item)) {
9818                         cache->last_byte_to_unpin = (u64)-1;
9819                         cache->cached = BTRFS_CACHE_FINISHED;
9820                         free_excluded_extents(root, cache);
9821                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9822                         cache->last_byte_to_unpin = (u64)-1;
9823                         cache->cached = BTRFS_CACHE_FINISHED;
9824                         add_new_free_space(cache, root->fs_info,
9825                                            found_key.objectid,
9826                                            found_key.objectid +
9827                                            found_key.offset);
9828                         free_excluded_extents(root, cache);
9829                 }
9830
9831                 ret = btrfs_add_block_group_cache(root->fs_info, cache);
9832                 if (ret) {
9833                         btrfs_remove_free_space_cache(cache);
9834                         btrfs_put_block_group(cache);
9835                         goto error;
9836                 }
9837
9838                 ret = update_space_info(info, cache->flags, found_key.offset,
9839                                         btrfs_block_group_used(&cache->item),
9840                                         &space_info);
9841                 if (ret) {
9842                         btrfs_remove_free_space_cache(cache);
9843                         spin_lock(&info->block_group_cache_lock);
9844                         rb_erase(&cache->cache_node,
9845                                  &info->block_group_cache_tree);
9846                         RB_CLEAR_NODE(&cache->cache_node);
9847                         spin_unlock(&info->block_group_cache_lock);
9848                         btrfs_put_block_group(cache);
9849                         goto error;
9850                 }
9851
9852                 cache->space_info = space_info;
9853                 spin_lock(&cache->space_info->lock);
9854                 cache->space_info->bytes_readonly += cache->bytes_super;
9855                 spin_unlock(&cache->space_info->lock);
9856
9857                 __link_block_group(space_info, cache);
9858
9859                 set_avail_alloc_bits(root->fs_info, cache->flags);
9860                 if (btrfs_chunk_readonly(root, cache->key.objectid)) {
9861                         inc_block_group_ro(cache, 1);
9862                 } else if (btrfs_block_group_used(&cache->item) == 0) {
9863                         spin_lock(&info->unused_bgs_lock);
9864                         /* Should always be true but just in case. */
9865                         if (list_empty(&cache->bg_list)) {
9866                                 btrfs_get_block_group(cache);
9867                                 list_add_tail(&cache->bg_list,
9868                                               &info->unused_bgs);
9869                         }
9870                         spin_unlock(&info->unused_bgs_lock);
9871                 }
9872         }
9873
9874         list_for_each_entry_rcu(space_info, &root->fs_info->space_info, list) {
9875                 if (!(get_alloc_profile(root, space_info->flags) &
9876                       (BTRFS_BLOCK_GROUP_RAID10 |
9877                        BTRFS_BLOCK_GROUP_RAID1 |
9878                        BTRFS_BLOCK_GROUP_RAID5 |
9879                        BTRFS_BLOCK_GROUP_RAID6 |
9880                        BTRFS_BLOCK_GROUP_DUP)))
9881                         continue;
9882                 /*
9883                  * avoid allocating from un-mirrored block group if there are
9884                  * mirrored block groups.
9885                  */
9886                 list_for_each_entry(cache,
9887                                 &space_info->block_groups[BTRFS_RAID_RAID0],
9888                                 list)
9889                         inc_block_group_ro(cache, 1);
9890                 list_for_each_entry(cache,
9891                                 &space_info->block_groups[BTRFS_RAID_SINGLE],
9892                                 list)
9893                         inc_block_group_ro(cache, 1);
9894         }
9895
9896         init_global_block_rsv(info);
9897         ret = 0;
9898 error:
9899         btrfs_free_path(path);
9900         return ret;
9901 }
9902
9903 void btrfs_create_pending_block_groups(struct btrfs_trans_handle *trans,
9904                                        struct btrfs_root *root)
9905 {
9906         struct btrfs_block_group_cache *block_group, *tmp;
9907         struct btrfs_root *extent_root = root->fs_info->extent_root;
9908         struct btrfs_block_group_item item;
9909         struct btrfs_key key;
9910         int ret = 0;
9911         bool can_flush_pending_bgs = trans->can_flush_pending_bgs;
9912
9913         trans->can_flush_pending_bgs = false;
9914         list_for_each_entry_safe(block_group, tmp, &trans->new_bgs, bg_list) {
9915                 if (ret)
9916                         goto next;
9917
9918                 spin_lock(&block_group->lock);
9919                 memcpy(&item, &block_group->item, sizeof(item));
9920                 memcpy(&key, &block_group->key, sizeof(key));
9921                 spin_unlock(&block_group->lock);
9922
9923                 ret = btrfs_insert_item(trans, extent_root, &key, &item,
9924                                         sizeof(item));
9925                 if (ret)
9926                         btrfs_abort_transaction(trans, extent_root, ret);
9927                 ret = btrfs_finish_chunk_alloc(trans, extent_root,
9928                                                key.objectid, key.offset);
9929                 if (ret)
9930                         btrfs_abort_transaction(trans, extent_root, ret);
9931                 add_block_group_free_space(trans, root->fs_info, block_group);
9932                 /* already aborted the transaction if it failed. */
9933 next:
9934                 list_del_init(&block_group->bg_list);
9935         }
9936         trans->can_flush_pending_bgs = can_flush_pending_bgs;
9937 }
9938
9939 int btrfs_make_block_group(struct btrfs_trans_handle *trans,
9940                            struct btrfs_root *root, u64 bytes_used,
9941                            u64 type, u64 chunk_objectid, u64 chunk_offset,
9942                            u64 size)
9943 {
9944         int ret;
9945         struct btrfs_root *extent_root;
9946         struct btrfs_block_group_cache *cache;
9947
9948         extent_root = root->fs_info->extent_root;
9949
9950         btrfs_set_log_full_commit(root->fs_info, trans);
9951
9952         cache = btrfs_create_block_group_cache(root, chunk_offset, size);
9953         if (!cache)
9954                 return -ENOMEM;
9955
9956         btrfs_set_block_group_used(&cache->item, bytes_used);
9957         btrfs_set_block_group_chunk_objectid(&cache->item, chunk_objectid);
9958         btrfs_set_block_group_flags(&cache->item, type);
9959
9960         cache->flags = type;
9961         cache->last_byte_to_unpin = (u64)-1;
9962         cache->cached = BTRFS_CACHE_FINISHED;
9963         cache->needs_free_space = 1;
9964         ret = exclude_super_stripes(root, cache);
9965         if (ret) {
9966                 /*
9967                  * We may have excluded something, so call this just in
9968                  * case.
9969                  */
9970                 free_excluded_extents(root, cache);
9971                 btrfs_put_block_group(cache);
9972                 return ret;
9973         }
9974
9975         add_new_free_space(cache, root->fs_info, chunk_offset,
9976                            chunk_offset + size);
9977
9978         free_excluded_extents(root, cache);
9979
9980 #ifdef CONFIG_BTRFS_DEBUG
9981         if (btrfs_should_fragment_free_space(root, cache)) {
9982                 u64 new_bytes_used = size - bytes_used;
9983
9984                 bytes_used += new_bytes_used >> 1;
9985                 fragment_free_space(root, cache);
9986         }
9987 #endif
9988         /*
9989          * Call to ensure the corresponding space_info object is created and
9990          * assigned to our block group, but don't update its counters just yet.
9991          * We want our bg to be added to the rbtree with its ->space_info set.
9992          */
9993         ret = update_space_info(root->fs_info, cache->flags, 0, 0,
9994                                 &cache->space_info);
9995         if (ret) {
9996                 btrfs_remove_free_space_cache(cache);
9997                 btrfs_put_block_group(cache);
9998                 return ret;
9999         }
10000
10001         ret = btrfs_add_block_group_cache(root->fs_info, cache);
10002         if (ret) {
10003                 btrfs_remove_free_space_cache(cache);
10004                 btrfs_put_block_group(cache);
10005                 return ret;
10006         }
10007
10008         /*
10009          * Now that our block group has its ->space_info set and is inserted in
10010          * the rbtree, update the space info's counters.
10011          */
10012         ret = update_space_info(root->fs_info, cache->flags, size, bytes_used,
10013                                 &cache->space_info);
10014         if (ret) {
10015                 btrfs_remove_free_space_cache(cache);
10016                 spin_lock(&root->fs_info->block_group_cache_lock);
10017                 rb_erase(&cache->cache_node,
10018                          &root->fs_info->block_group_cache_tree);
10019                 RB_CLEAR_NODE(&cache->cache_node);
10020                 spin_unlock(&root->fs_info->block_group_cache_lock);
10021                 btrfs_put_block_group(cache);
10022                 return ret;
10023         }
10024         update_global_block_rsv(root->fs_info);
10025
10026         spin_lock(&cache->space_info->lock);
10027         cache->space_info->bytes_readonly += cache->bytes_super;
10028         spin_unlock(&cache->space_info->lock);
10029
10030         __link_block_group(cache->space_info, cache);
10031
10032         list_add_tail(&cache->bg_list, &trans->new_bgs);
10033
10034         set_avail_alloc_bits(extent_root->fs_info, type);
10035
10036         return 0;
10037 }
10038
10039 static void clear_avail_alloc_bits(struct btrfs_fs_info *fs_info, u64 flags)
10040 {
10041         u64 extra_flags = chunk_to_extended(flags) &
10042                                 BTRFS_EXTENDED_PROFILE_MASK;
10043
10044         write_seqlock(&fs_info->profiles_lock);
10045         if (flags & BTRFS_BLOCK_GROUP_DATA)
10046                 fs_info->avail_data_alloc_bits &= ~extra_flags;
10047         if (flags & BTRFS_BLOCK_GROUP_METADATA)
10048                 fs_info->avail_metadata_alloc_bits &= ~extra_flags;
10049         if (flags & BTRFS_BLOCK_GROUP_SYSTEM)
10050                 fs_info->avail_system_alloc_bits &= ~extra_flags;
10051         write_sequnlock(&fs_info->profiles_lock);
10052 }
10053
10054 int btrfs_remove_block_group(struct btrfs_trans_handle *trans,
10055                              struct btrfs_root *root, u64 group_start,
10056                              struct extent_map *em)
10057 {
10058         struct btrfs_path *path;
10059         struct btrfs_block_group_cache *block_group;
10060         struct btrfs_free_cluster *cluster;
10061         struct btrfs_root *tree_root = root->fs_info->tree_root;
10062         struct btrfs_key key;
10063         struct inode *inode;
10064         struct kobject *kobj = NULL;
10065         int ret;
10066         int index;
10067         int factor;
10068         struct btrfs_caching_control *caching_ctl = NULL;
10069         bool remove_em;
10070
10071         root = root->fs_info->extent_root;
10072
10073         block_group = btrfs_lookup_block_group(root->fs_info, group_start);
10074         BUG_ON(!block_group);
10075         BUG_ON(!block_group->ro);
10076
10077         /*
10078          * Free the reserved super bytes from this block group before
10079          * remove it.
10080          */
10081         free_excluded_extents(root, block_group);
10082
10083         memcpy(&key, &block_group->key, sizeof(key));
10084         index = get_block_group_index(block_group);
10085         if (block_group->flags & (BTRFS_BLOCK_GROUP_DUP |
10086                                   BTRFS_BLOCK_GROUP_RAID1 |
10087                                   BTRFS_BLOCK_GROUP_RAID10))
10088                 factor = 2;
10089         else
10090                 factor = 1;
10091
10092         /* make sure this block group isn't part of an allocation cluster */
10093         cluster = &root->fs_info->data_alloc_cluster;
10094         spin_lock(&cluster->refill_lock);
10095         btrfs_return_cluster_to_free_space(block_group, cluster);
10096         spin_unlock(&cluster->refill_lock);
10097
10098         /*
10099          * make sure this block group isn't part of a metadata
10100          * allocation cluster
10101          */
10102         cluster = &root->fs_info->meta_alloc_cluster;
10103         spin_lock(&cluster->refill_lock);
10104         btrfs_return_cluster_to_free_space(block_group, cluster);
10105         spin_unlock(&cluster->refill_lock);
10106
10107         path = btrfs_alloc_path();
10108         if (!path) {
10109                 ret = -ENOMEM;
10110                 goto out;
10111         }
10112
10113         /*
10114          * get the inode first so any iput calls done for the io_list
10115          * aren't the final iput (no unlinks allowed now)
10116          */
10117         inode = lookup_free_space_inode(tree_root, block_group, path);
10118
10119         mutex_lock(&trans->transaction->cache_write_mutex);
10120         /*
10121          * make sure our free spache cache IO is done before remove the
10122          * free space inode
10123          */
10124         spin_lock(&trans->transaction->dirty_bgs_lock);
10125         if (!list_empty(&block_group->io_list)) {
10126                 list_del_init(&block_group->io_list);
10127
10128                 WARN_ON(!IS_ERR(inode) && inode != block_group->io_ctl.inode);
10129
10130                 spin_unlock(&trans->transaction->dirty_bgs_lock);
10131                 btrfs_wait_cache_io(root, trans, block_group,
10132                                     &block_group->io_ctl, path,
10133                                     block_group->key.objectid);
10134                 btrfs_put_block_group(block_group);
10135                 spin_lock(&trans->transaction->dirty_bgs_lock);
10136         }
10137
10138         if (!list_empty(&block_group->dirty_list)) {
10139                 list_del_init(&block_group->dirty_list);
10140                 btrfs_put_block_group(block_group);
10141         }
10142         spin_unlock(&trans->transaction->dirty_bgs_lock);
10143         mutex_unlock(&trans->transaction->cache_write_mutex);
10144
10145         if (!IS_ERR(inode)) {
10146                 ret = btrfs_orphan_add(trans, inode);
10147                 if (ret) {
10148                         btrfs_add_delayed_iput(inode);
10149                         goto out;
10150                 }
10151                 clear_nlink(inode);
10152                 /* One for the block groups ref */
10153                 spin_lock(&block_group->lock);
10154                 if (block_group->iref) {
10155                         block_group->iref = 0;
10156                         block_group->inode = NULL;
10157                         spin_unlock(&block_group->lock);
10158                         iput(inode);
10159                 } else {
10160                         spin_unlock(&block_group->lock);
10161                 }
10162                 /* One for our lookup ref */
10163                 btrfs_add_delayed_iput(inode);
10164         }
10165
10166         key.objectid = BTRFS_FREE_SPACE_OBJECTID;
10167         key.offset = block_group->key.objectid;
10168         key.type = 0;
10169
10170         ret = btrfs_search_slot(trans, tree_root, &key, path, -1, 1);
10171         if (ret < 0)
10172                 goto out;
10173         if (ret > 0)
10174                 btrfs_release_path(path);
10175         if (ret == 0) {
10176                 ret = btrfs_del_item(trans, tree_root, path);
10177                 if (ret)
10178                         goto out;
10179                 btrfs_release_path(path);
10180         }
10181
10182         spin_lock(&root->fs_info->block_group_cache_lock);
10183         rb_erase(&block_group->cache_node,
10184                  &root->fs_info->block_group_cache_tree);
10185         RB_CLEAR_NODE(&block_group->cache_node);
10186
10187         if (root->fs_info->first_logical_byte == block_group->key.objectid)
10188                 root->fs_info->first_logical_byte = (u64)-1;
10189         spin_unlock(&root->fs_info->block_group_cache_lock);
10190
10191         down_write(&block_group->space_info->groups_sem);
10192         /*
10193          * we must use list_del_init so people can check to see if they
10194          * are still on the list after taking the semaphore
10195          */
10196         list_del_init(&block_group->list);
10197         if (list_empty(&block_group->space_info->block_groups[index])) {
10198                 kobj = block_group->space_info->block_group_kobjs[index];
10199                 block_group->space_info->block_group_kobjs[index] = NULL;
10200                 clear_avail_alloc_bits(root->fs_info, block_group->flags);
10201         }
10202         up_write(&block_group->space_info->groups_sem);
10203         if (kobj) {
10204                 kobject_del(kobj);
10205                 kobject_put(kobj);
10206         }
10207
10208         if (block_group->has_caching_ctl)
10209                 caching_ctl = get_caching_control(block_group);
10210         if (block_group->cached == BTRFS_CACHE_STARTED)
10211                 wait_block_group_cache_done(block_group);
10212         if (block_group->has_caching_ctl) {
10213                 down_write(&root->fs_info->commit_root_sem);
10214                 if (!caching_ctl) {
10215                         struct btrfs_caching_control *ctl;
10216
10217                         list_for_each_entry(ctl,
10218                                     &root->fs_info->caching_block_groups, list)
10219                                 if (ctl->block_group == block_group) {
10220                                         caching_ctl = ctl;
10221                                         atomic_inc(&caching_ctl->count);
10222                                         break;
10223                                 }
10224                 }
10225                 if (caching_ctl)
10226                         list_del_init(&caching_ctl->list);
10227                 up_write(&root->fs_info->commit_root_sem);
10228                 if (caching_ctl) {
10229                         /* Once for the caching bgs list and once for us. */
10230                         put_caching_control(caching_ctl);
10231                         put_caching_control(caching_ctl);
10232                 }
10233         }
10234
10235         spin_lock(&trans->transaction->dirty_bgs_lock);
10236         if (!list_empty(&block_group->dirty_list)) {
10237                 WARN_ON(1);
10238         }
10239         if (!list_empty(&block_group->io_list)) {
10240                 WARN_ON(1);
10241         }
10242         spin_unlock(&trans->transaction->dirty_bgs_lock);
10243         btrfs_remove_free_space_cache(block_group);
10244
10245         spin_lock(&block_group->space_info->lock);
10246         list_del_init(&block_group->ro_list);
10247
10248         if (btrfs_test_opt(root, ENOSPC_DEBUG)) {
10249                 WARN_ON(block_group->space_info->total_bytes
10250                         < block_group->key.offset);
10251                 WARN_ON(block_group->space_info->bytes_readonly
10252                         < block_group->key.offset);
10253                 WARN_ON(block_group->space_info->disk_total
10254                         < block_group->key.offset * factor);
10255         }
10256         block_group->space_info->total_bytes -= block_group->key.offset;
10257         block_group->space_info->bytes_readonly -= block_group->key.offset;
10258         block_group->space_info->disk_total -= block_group->key.offset * factor;
10259
10260         spin_unlock(&block_group->space_info->lock);
10261
10262         memcpy(&key, &block_group->key, sizeof(key));
10263
10264         lock_chunks(root);
10265         if (!list_empty(&em->list)) {
10266                 /* We're in the transaction->pending_chunks list. */
10267                 free_extent_map(em);
10268         }
10269         spin_lock(&block_group->lock);
10270         block_group->removed = 1;
10271         /*
10272          * At this point trimming can't start on this block group, because we
10273          * removed the block group from the tree fs_info->block_group_cache_tree
10274          * so no one can't find it anymore and even if someone already got this
10275          * block group before we removed it from the rbtree, they have already
10276          * incremented block_group->trimming - if they didn't, they won't find
10277          * any free space entries because we already removed them all when we
10278          * called btrfs_remove_free_space_cache().
10279          *
10280          * And we must not remove the extent map from the fs_info->mapping_tree
10281          * to prevent the same logical address range and physical device space
10282          * ranges from being reused for a new block group. This is because our
10283          * fs trim operation (btrfs_trim_fs() / btrfs_ioctl_fitrim()) is
10284          * completely transactionless, so while it is trimming a range the
10285          * currently running transaction might finish and a new one start,
10286          * allowing for new block groups to be created that can reuse the same
10287          * physical device locations unless we take this special care.
10288          *
10289          * There may also be an implicit trim operation if the file system
10290          * is mounted with -odiscard. The same protections must remain
10291          * in place until the extents have been discarded completely when
10292          * the transaction commit has completed.
10293          */
10294         remove_em = (atomic_read(&block_group->trimming) == 0);
10295         /*
10296          * Make sure a trimmer task always sees the em in the pinned_chunks list
10297          * if it sees block_group->removed == 1 (needs to lock block_group->lock
10298          * before checking block_group->removed).
10299          */
10300         if (!remove_em) {
10301                 /*
10302                  * Our em might be in trans->transaction->pending_chunks which
10303                  * is protected by fs_info->chunk_mutex ([lock|unlock]_chunks),
10304                  * and so is the fs_info->pinned_chunks list.
10305                  *
10306                  * So at this point we must be holding the chunk_mutex to avoid
10307                  * any races with chunk allocation (more specifically at
10308                  * volumes.c:contains_pending_extent()), to ensure it always
10309                  * sees the em, either in the pending_chunks list or in the
10310                  * pinned_chunks list.
10311                  */
10312                 list_move_tail(&em->list, &root->fs_info->pinned_chunks);
10313         }
10314         spin_unlock(&block_group->lock);
10315
10316         if (remove_em) {
10317                 struct extent_map_tree *em_tree;
10318
10319                 em_tree = &root->fs_info->mapping_tree.map_tree;
10320                 write_lock(&em_tree->lock);
10321                 /*
10322                  * The em might be in the pending_chunks list, so make sure the
10323                  * chunk mutex is locked, since remove_extent_mapping() will
10324                  * delete us from that list.
10325                  */
10326                 remove_extent_mapping(em_tree, em);
10327                 write_unlock(&em_tree->lock);
10328                 /* once for the tree */
10329                 free_extent_map(em);
10330         }
10331
10332         unlock_chunks(root);
10333
10334         ret = remove_block_group_free_space(trans, root->fs_info, block_group);
10335         if (ret)
10336                 goto out;
10337
10338         btrfs_put_block_group(block_group);
10339         btrfs_put_block_group(block_group);
10340
10341         ret = btrfs_search_slot(trans, root, &key, path, -1, 1);
10342         if (ret > 0)
10343                 ret = -EIO;
10344         if (ret < 0)
10345                 goto out;
10346
10347         ret = btrfs_del_item(trans, root, path);
10348 out:
10349         btrfs_free_path(path);
10350         return ret;
10351 }
10352
10353 struct btrfs_trans_handle *
10354 btrfs_start_trans_remove_block_group(struct btrfs_fs_info *fs_info,
10355                                      const u64 chunk_offset)
10356 {
10357         struct extent_map_tree *em_tree = &fs_info->mapping_tree.map_tree;
10358         struct extent_map *em;
10359         struct map_lookup *map;
10360         unsigned int num_items;
10361
10362         read_lock(&em_tree->lock);
10363         em = lookup_extent_mapping(em_tree, chunk_offset, 1);
10364         read_unlock(&em_tree->lock);
10365         ASSERT(em && em->start == chunk_offset);
10366
10367         /*
10368          * We need to reserve 3 + N units from the metadata space info in order
10369          * to remove a block group (done at btrfs_remove_chunk() and at
10370          * btrfs_remove_block_group()), which are used for:
10371          *
10372          * 1 unit for adding the free space inode's orphan (located in the tree
10373          * of tree roots).
10374          * 1 unit for deleting the block group item (located in the extent
10375          * tree).
10376          * 1 unit for deleting the free space item (located in tree of tree
10377          * roots).
10378          * N units for deleting N device extent items corresponding to each
10379          * stripe (located in the device tree).
10380          *
10381          * In order to remove a block group we also need to reserve units in the
10382          * system space info in order to update the chunk tree (update one or
10383          * more device items and remove one chunk item), but this is done at
10384          * btrfs_remove_chunk() through a call to check_system_chunk().
10385          */
10386         map = (struct map_lookup *)em->bdev;
10387         num_items = 3 + map->num_stripes;
10388         free_extent_map(em);
10389
10390         return btrfs_start_transaction_fallback_global_rsv(fs_info->extent_root,
10391                                                            num_items, 1);
10392 }
10393
10394 /*
10395  * Process the unused_bgs list and remove any that don't have any allocated
10396  * space inside of them.
10397  */
10398 void btrfs_delete_unused_bgs(struct btrfs_fs_info *fs_info)
10399 {
10400         struct btrfs_block_group_cache *block_group;
10401         struct btrfs_space_info *space_info;
10402         struct btrfs_root *root = fs_info->extent_root;
10403         struct btrfs_trans_handle *trans;
10404         int ret = 0;
10405
10406         if (!fs_info->open)
10407                 return;
10408
10409         spin_lock(&fs_info->unused_bgs_lock);
10410         while (!list_empty(&fs_info->unused_bgs)) {
10411                 u64 start, end;
10412                 int trimming;
10413
10414                 block_group = list_first_entry(&fs_info->unused_bgs,
10415                                                struct btrfs_block_group_cache,
10416                                                bg_list);
10417                 list_del_init(&block_group->bg_list);
10418
10419                 space_info = block_group->space_info;
10420
10421                 if (ret || btrfs_mixed_space_info(space_info)) {
10422                         btrfs_put_block_group(block_group);
10423                         continue;
10424                 }
10425                 spin_unlock(&fs_info->unused_bgs_lock);
10426
10427                 mutex_lock(&fs_info->delete_unused_bgs_mutex);
10428
10429                 /* Don't want to race with allocators so take the groups_sem */
10430                 down_write(&space_info->groups_sem);
10431                 spin_lock(&block_group->lock);
10432                 if (block_group->reserved ||
10433                     btrfs_block_group_used(&block_group->item) ||
10434                     block_group->ro ||
10435                     list_is_singular(&block_group->list)) {
10436                         /*
10437                          * We want to bail if we made new allocations or have
10438                          * outstanding allocations in this block group.  We do
10439                          * the ro check in case balance is currently acting on
10440                          * this block group.
10441                          */
10442                         spin_unlock(&block_group->lock);
10443                         up_write(&space_info->groups_sem);
10444                         goto next;
10445                 }
10446                 spin_unlock(&block_group->lock);
10447
10448                 /* We don't want to force the issue, only flip if it's ok. */
10449                 ret = inc_block_group_ro(block_group, 0);
10450                 up_write(&space_info->groups_sem);
10451                 if (ret < 0) {
10452                         ret = 0;
10453                         goto next;
10454                 }
10455
10456                 /*
10457                  * Want to do this before we do anything else so we can recover
10458                  * properly if we fail to join the transaction.
10459                  */
10460                 trans = btrfs_start_trans_remove_block_group(fs_info,
10461                                                      block_group->key.objectid);
10462                 if (IS_ERR(trans)) {
10463                         btrfs_dec_block_group_ro(root, block_group);
10464                         ret = PTR_ERR(trans);
10465                         goto next;
10466                 }
10467
10468                 /*
10469                  * We could have pending pinned extents for this block group,
10470                  * just delete them, we don't care about them anymore.
10471                  */
10472                 start = block_group->key.objectid;
10473                 end = start + block_group->key.offset - 1;
10474                 /*
10475                  * Hold the unused_bg_unpin_mutex lock to avoid racing with
10476                  * btrfs_finish_extent_commit(). If we are at transaction N,
10477                  * another task might be running finish_extent_commit() for the
10478                  * previous transaction N - 1, and have seen a range belonging
10479                  * to the block group in freed_extents[] before we were able to
10480                  * clear the whole block group range from freed_extents[]. This
10481                  * means that task can lookup for the block group after we
10482                  * unpinned it from freed_extents[] and removed it, leading to
10483                  * a BUG_ON() at btrfs_unpin_extent_range().
10484                  */
10485                 mutex_lock(&fs_info->unused_bg_unpin_mutex);
10486                 ret = clear_extent_bits(&fs_info->freed_extents[0], start, end,
10487                                   EXTENT_DIRTY, GFP_NOFS);
10488                 if (ret) {
10489                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10490                         btrfs_dec_block_group_ro(root, block_group);
10491                         goto end_trans;
10492                 }
10493                 ret = clear_extent_bits(&fs_info->freed_extents[1], start, end,
10494                                   EXTENT_DIRTY, GFP_NOFS);
10495                 if (ret) {
10496                         mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10497                         btrfs_dec_block_group_ro(root, block_group);
10498                         goto end_trans;
10499                 }
10500                 mutex_unlock(&fs_info->unused_bg_unpin_mutex);
10501
10502                 /* Reset pinned so btrfs_put_block_group doesn't complain */
10503                 spin_lock(&space_info->lock);
10504                 spin_lock(&block_group->lock);
10505
10506                 space_info->bytes_pinned -= block_group->pinned;
10507                 space_info->bytes_readonly += block_group->pinned;
10508                 percpu_counter_add(&space_info->total_bytes_pinned,
10509                                    -block_group->pinned);
10510                 block_group->pinned = 0;
10511
10512                 spin_unlock(&block_group->lock);
10513                 spin_unlock(&space_info->lock);
10514
10515                 /* DISCARD can flip during remount */
10516                 trimming = btrfs_test_opt(root, DISCARD);
10517
10518                 /* Implicit trim during transaction commit. */
10519                 if (trimming)
10520                         btrfs_get_block_group_trimming(block_group);
10521
10522                 /*
10523                  * Btrfs_remove_chunk will abort the transaction if things go
10524                  * horribly wrong.
10525                  */
10526                 ret = btrfs_remove_chunk(trans, root,
10527                                          block_group->key.objectid);
10528
10529                 if (ret) {
10530                         if (trimming)
10531                                 btrfs_put_block_group_trimming(block_group);
10532                         goto end_trans;
10533                 }
10534
10535                 /*
10536                  * If we're not mounted with -odiscard, we can just forget
10537                  * about this block group. Otherwise we'll need to wait
10538                  * until transaction commit to do the actual discard.
10539                  */
10540                 if (trimming) {
10541                         spin_lock(&fs_info->unused_bgs_lock);
10542                         /*
10543                          * A concurrent scrub might have added us to the list
10544                          * fs_info->unused_bgs, so use a list_move operation
10545                          * to add the block group to the deleted_bgs list.
10546                          */
10547                         list_move(&block_group->bg_list,
10548                                   &trans->transaction->deleted_bgs);
10549                         spin_unlock(&fs_info->unused_bgs_lock);
10550                         btrfs_get_block_group(block_group);
10551                 }
10552 end_trans:
10553                 btrfs_end_transaction(trans, root);
10554 next:
10555                 mutex_unlock(&fs_info->delete_unused_bgs_mutex);
10556                 btrfs_put_block_group(block_group);
10557                 spin_lock(&fs_info->unused_bgs_lock);
10558         }
10559         spin_unlock(&fs_info->unused_bgs_lock);
10560 }
10561
10562 int btrfs_init_space_info(struct btrfs_fs_info *fs_info)
10563 {
10564         struct btrfs_space_info *space_info;
10565         struct btrfs_super_block *disk_super;
10566         u64 features;
10567         u64 flags;
10568         int mixed = 0;
10569         int ret;
10570
10571         disk_super = fs_info->super_copy;
10572         if (!btrfs_super_root(disk_super))
10573                 return 1;
10574
10575         features = btrfs_super_incompat_flags(disk_super);
10576         if (features & BTRFS_FEATURE_INCOMPAT_MIXED_GROUPS)
10577                 mixed = 1;
10578
10579         flags = BTRFS_BLOCK_GROUP_SYSTEM;
10580         ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10581         if (ret)
10582                 goto out;
10583
10584         if (mixed) {
10585                 flags = BTRFS_BLOCK_GROUP_METADATA | BTRFS_BLOCK_GROUP_DATA;
10586                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10587         } else {
10588                 flags = BTRFS_BLOCK_GROUP_METADATA;
10589                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10590                 if (ret)
10591                         goto out;
10592
10593                 flags = BTRFS_BLOCK_GROUP_DATA;
10594                 ret = update_space_info(fs_info, flags, 0, 0, &space_info);
10595         }
10596 out:
10597         return ret;
10598 }
10599
10600 int btrfs_error_unpin_extent_range(struct btrfs_root *root, u64 start, u64 end)
10601 {
10602         return unpin_extent_range(root, start, end, false);
10603 }
10604
10605 /*
10606  * It used to be that old block groups would be left around forever.
10607  * Iterating over them would be enough to trim unused space.  Since we
10608  * now automatically remove them, we also need to iterate over unallocated
10609  * space.
10610  *
10611  * We don't want a transaction for this since the discard may take a
10612  * substantial amount of time.  We don't require that a transaction be
10613  * running, but we do need to take a running transaction into account
10614  * to ensure that we're not discarding chunks that were released in
10615  * the current transaction.
10616  *
10617  * Holding the chunks lock will prevent other threads from allocating
10618  * or releasing chunks, but it won't prevent a running transaction
10619  * from committing and releasing the memory that the pending chunks
10620  * list head uses.  For that, we need to take a reference to the
10621  * transaction.
10622  */
10623 static int btrfs_trim_free_extents(struct btrfs_device *device,
10624                                    u64 minlen, u64 *trimmed)
10625 {
10626         u64 start = 0, len = 0;
10627         int ret;
10628
10629         *trimmed = 0;
10630
10631         /* Not writeable = nothing to do. */
10632         if (!device->writeable)
10633                 return 0;
10634
10635         /* No free space = nothing to do. */
10636         if (device->total_bytes <= device->bytes_used)
10637                 return 0;
10638
10639         ret = 0;
10640
10641         while (1) {
10642                 struct btrfs_fs_info *fs_info = device->dev_root->fs_info;
10643                 struct btrfs_transaction *trans;
10644                 u64 bytes;
10645
10646                 ret = mutex_lock_interruptible(&fs_info->chunk_mutex);
10647                 if (ret)
10648                         return ret;
10649
10650                 down_read(&fs_info->commit_root_sem);
10651
10652                 spin_lock(&fs_info->trans_lock);
10653                 trans = fs_info->running_transaction;
10654                 if (trans)
10655                         atomic_inc(&trans->use_count);
10656                 spin_unlock(&fs_info->trans_lock);
10657
10658                 ret = find_free_dev_extent_start(trans, device, minlen, start,
10659                                                  &start, &len);
10660                 if (trans)
10661                         btrfs_put_transaction(trans);
10662
10663                 if (ret) {
10664                         up_read(&fs_info->commit_root_sem);
10665                         mutex_unlock(&fs_info->chunk_mutex);
10666                         if (ret == -ENOSPC)
10667                                 ret = 0;
10668                         break;
10669                 }
10670
10671                 ret = btrfs_issue_discard(device->bdev, start, len, &bytes);
10672                 up_read(&fs_info->commit_root_sem);
10673                 mutex_unlock(&fs_info->chunk_mutex);
10674
10675                 if (ret)
10676                         break;
10677
10678                 start += len;
10679                 *trimmed += bytes;
10680
10681                 if (fatal_signal_pending(current)) {
10682                         ret = -ERESTARTSYS;
10683                         break;
10684                 }
10685
10686                 cond_resched();
10687         }
10688
10689         return ret;
10690 }
10691
10692 int btrfs_trim_fs(struct btrfs_root *root, struct fstrim_range *range)
10693 {
10694         struct btrfs_fs_info *fs_info = root->fs_info;
10695         struct btrfs_block_group_cache *cache = NULL;
10696         struct btrfs_device *device;
10697         struct list_head *devices;
10698         u64 group_trimmed;
10699         u64 start;
10700         u64 end;
10701         u64 trimmed = 0;
10702         u64 total_bytes = btrfs_super_total_bytes(fs_info->super_copy);
10703         int ret = 0;
10704
10705         /*
10706          * try to trim all FS space, our block group may start from non-zero.
10707          */
10708         if (range->len == total_bytes)
10709                 cache = btrfs_lookup_first_block_group(fs_info, range->start);
10710         else
10711                 cache = btrfs_lookup_block_group(fs_info, range->start);
10712
10713         while (cache) {
10714                 if (cache->key.objectid >= (range->start + range->len)) {
10715                         btrfs_put_block_group(cache);
10716                         break;
10717                 }
10718
10719                 start = max(range->start, cache->key.objectid);
10720                 end = min(range->start + range->len,
10721                                 cache->key.objectid + cache->key.offset);
10722
10723                 if (end - start >= range->minlen) {
10724                         if (!block_group_cache_done(cache)) {
10725                                 ret = cache_block_group(cache, 0);
10726                                 if (ret) {
10727                                         btrfs_put_block_group(cache);
10728                                         break;
10729                                 }
10730                                 ret = wait_block_group_cache_done(cache);
10731                                 if (ret) {
10732                                         btrfs_put_block_group(cache);
10733                                         break;
10734                                 }
10735                         }
10736                         ret = btrfs_trim_block_group(cache,
10737                                                      &group_trimmed,
10738                                                      start,
10739                                                      end,
10740                                                      range->minlen);
10741
10742                         trimmed += group_trimmed;
10743                         if (ret) {
10744                                 btrfs_put_block_group(cache);
10745                                 break;
10746                         }
10747                 }
10748
10749                 cache = next_block_group(fs_info->tree_root, cache);
10750         }
10751
10752         mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
10753         devices = &root->fs_info->fs_devices->alloc_list;
10754         list_for_each_entry(device, devices, dev_alloc_list) {
10755                 ret = btrfs_trim_free_extents(device, range->minlen,
10756                                               &group_trimmed);
10757                 if (ret)
10758                         break;
10759
10760                 trimmed += group_trimmed;
10761         }
10762         mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
10763
10764         range->len = trimmed;
10765         return ret;
10766 }
10767
10768 /*
10769  * btrfs_{start,end}_write_no_snapshoting() are similar to
10770  * mnt_{want,drop}_write(), they are used to prevent some tasks from writing
10771  * data into the page cache through nocow before the subvolume is snapshoted,
10772  * but flush the data into disk after the snapshot creation, or to prevent
10773  * operations while snapshoting is ongoing and that cause the snapshot to be
10774  * inconsistent (writes followed by expanding truncates for example).
10775  */
10776 void btrfs_end_write_no_snapshoting(struct btrfs_root *root)
10777 {
10778         percpu_counter_dec(&root->subv_writers->counter);
10779         /*
10780          * Make sure counter is updated before we wake up waiters.
10781          */
10782         smp_mb();
10783         if (waitqueue_active(&root->subv_writers->wait))
10784                 wake_up(&root->subv_writers->wait);
10785 }
10786
10787 int btrfs_start_write_no_snapshoting(struct btrfs_root *root)
10788 {
10789         if (atomic_read(&root->will_be_snapshoted))
10790                 return 0;
10791
10792         percpu_counter_inc(&root->subv_writers->counter);
10793         /*
10794          * Make sure counter is updated before we check for snapshot creation.
10795          */
10796         smp_mb();
10797         if (atomic_read(&root->will_be_snapshoted)) {
10798                 btrfs_end_write_no_snapshoting(root);
10799                 return 0;
10800         }
10801         return 1;
10802 }